text
stringlengths 2.85k
2.55M
| label
class label 11
classes |
|---|---|
Space-Efficient Algorithms for Longest Increasing Subsequence∗
Masashi Kiyomi†
Hirotaka Ono‡
Yota Otachi§
Pascal Schweitzer¶
Jun Taruik
December 27, 2017
arXiv:1712.09230v1 [cs.DS] 26 Dec 2017
Abstract
Given a sequence of integers, we want to find a longest increasing subsequence of the sequence.
It is known that this problem can be solved in O(n log n) time and space. Our goal
√ in this paper
is to reduce the space consumption while keeping the time complexity small. For n ≤ s ≤ n, we
present algorithms that use O(s log n) bits and O( 1s · n2 · log n) time for computing the length of a
longest increasing subsequence, and O( 1s · n2 · log2 n) time for finding an actual subsequence. We
also show that the time complexity of our algorithms is optimal up to polylogarithmic factors in the
framework of sequential access algorithms with the prescribed amount of space.
1
Introduction
Given a sequence of integers (possibly with repetitions), the problem of finding a longest increasing
subsequence (LIS, for short) is a classic problem in computer science which has many application areas
including bioinfomatics and physics (see [38] and the references therein). It is known that LIS admits an
O(n log n)-time algorithm that uses O(n log n) bits of working space [37, 17, 2], where n is the length of
the sequence.
A wide-spread algorithm achieving these bounds is Patience Sorting, devised by Mallows [24, 25,
26]. Given a sequence of length n, Patience Sorting partitions the elements of the sequence into
so-called piles. It can be shown that the number of piles coincides with the length of a longest increasing
subsequence (see Section 3 for details). Combinatorial and statistical properties of the piles in Patience
Sorting are well studied (see [2, 8, 33]).
However, with the dramatic increase of the typical data sizes in applications over the last decade,
a main memory consumption in the order of Θ(n log n) bits is excessive in many algorithmic contexts,
especially for basic subroutines such as LIS. We therefore investigate the existence of space-efficient
algorithms for LIS.
Our results In this paper, we present the first space-efficient algorithms for LIS that are exact.
We start by observing that when the input is restricted to permutations, an algorithm using O(n) bits
can be obtained straightforwardly by modifying a previously known algorithm (see Section 3.3). Next,
we observe that a Savitch type algorithm [36] for this problem uses O(log2 n) bits and thus runs in
quasipolynomial time. However, we are mainly interested in space-efficient algorithms that also behave
well with regard to running time. To this end
√ we develop an algorithm that determines the length
of a longest increasing subsequence using O( n log n) bits which runs in O(n1.5 log n) time. Since the
constants hidden in the O-notation are negligible, the algorithm, when executed in the main memory of
a standard computer, may handle a peta-byte input on external storage.
More versatile, in fact, our space-efficient algorithm is memory-adjustable in the following
sense.
√
(See [3] for information on memory-adjustable algorithms.) When a memory bound s with n ≤ s ≤ n
∗ Partially
supported by MEXT KAKENHI grant number 24106004.
City University. Yokohama, Japan. [email protected]
‡ Nagoya University. Nagoya, Japan. [email protected]
§ Kumamoto University. Kumamoto, Japan. [email protected]
¶ TU Kaiserslautern. Kaiserslautern, Germany. [email protected]
k The University of Electro-Communications. Chofu, Japan. [email protected]
† Yokohama
1
is given to the algorithm, it computes with O(s log n) bits of working space in O( 1s · n2 log n) time the
length of a longest increasing subsequence. When
to the previously
√
√ s = n our algorithm is equivalent
known algorithms mentioned above. When s = n it uses, as claimed above, O( n log n) bits and runs
in O(n1.5 log n) time.
The algorithm only determines the length of a longest increasing subsequence. To actually find such
a longest increasing subsequence, one can run the length-determining algorithm n times to successively
construct the sought-after subsequence. This would give us a running time of O( 1s · n3 log n). However,
we show that one can do much better, achieving a running time of O( 1s · n2 log2 n) without any increase
in space complexity, by recursively finding a near-mid element of a longest increasing subsequence.
To design the algorithms, we study the structure of the piles arising in Patience Sorting in depth
and show that maintaining certain information regarding the piles suffices to simulate the algorithm.
Roughly speaking, our algorithm divides the execution of Patience Sorting into O(n/s) phases, and
in each phase it computes in O(n log n) time information on the next O(s) piles, while forgetting previous
information.
Finally, we complement our algorithm with a lower bound in a restricted computational model. In
the sequential access model, an algorithm can access the input only sequentially. We also consider further
restricted algorithms in the multi-pass model, where an algorithm has to read the input sequentially from
left to right and can repeat this multiple (not necessarily a constant number of) times. Our algorithm
for the length works within the multi-pass model, while the one for finding a subsequence is a sequential
access algorithm. Such algorithms are useful when large data is placed in an external storage that
supports efficient sequential access. We show that the time complexity of our algorithms is optimal up
to polylogarithmic factors in these models.
Related work The problem of finding a longest increasing subsequence (LIS) is among the most
basic algorithmic problems on integer arrays and has been studied continuously since the early 1960’s.
It is known that LIS can be solved in O(n log n) time and space [37, 17, 2], and that any comparisonbased algorithm needs Ω(n log n) comparisons even for computing the length of a longest increasing
subsequence [17, 32]. For the special case of LIS where the input is restricted to permutations, there are
O(n log log n)-time algorithms [20, 6, 12]. Patience Sorting, an efficient algorithm for LIS, has been
a research topic in itself, especially in the context of Young tableaux [24, 25, 26, 2, 8, 33].
Recently, LIS has been studied intensively in the data-streaming model, where the input can be read
only once (or a constant number of times) sequentially from left to right. This line of research was
initiated by Liben-Nowell, Vee, and Zhu [22], who presented an exact one-pass algorithm and a lower
bound for such algorithms. Their results were then improved and extended by many other groups [19,
38, 18, 34, 15, 28, 35]. These results give a deep understanding on streaming algorithms with a constant
number of passes even under the settings with randomization and approximation. (For details on these
models, see the very recent paper by Saks and Seshadhri [35] and the references therein.) On the other
hand, multi-pass algorithms with a non-constant number of passes have not been studied for LIS.
While space-limited algorithms on both RAM and multi-pass models for basic problems have been
studied since the early stage of algorithm theory, research in this field has recently intensified. Besides
LIS, other frequently studied problems include sorting and selection [27, 7, 16, 30], graph searching [4,
14, 31, 9], geometric computation [10, 13, 5, 1], and k-SUM [39, 23].
2
Preliminaries
Let τ = hτ (1), τ (2), . . . , τ (n)i be a sequence of n integers possibly with repetitions. For 1 ≤ i1 < . . . <
iℓ ≤ n, the subsequence τ [i1 , . . . , iℓ ] of τ is the sequence hτ (i1 ), . . . , τ (iℓ )i. A subsequence τ [i1 , . . . , iℓ ]
is an increasing subsequence of τ if τ (i1 ) < . . . < τ (iℓ ). If τ (i1 ) ≤ . . . ≤ τ (iℓ ), then the sequence τ
is non-decreasing. We analogously define decreasing subsequences and non-increasing subsequences. By
lis(τ ), we denote the length of a longest increasing subsequence of τ .
For example, consider a sequence τ1 = h2, 8, 4, 9, 5, 1, 7, 6, 3i. It has an increasing subsequence
τ1 [1, 3, 5, 8] = h2, 4, 5, 6i. Since there is no increasing subsequence of τ1 with length 5 or more, we
have lis(τ1 ) = 4.
In the computational model in this paper, we use the RAM model with the following restrictions that
are standard in the context of sublinear space algorithms. The input is in a read-only memory and the
2
output must be produced on a write-only memory. We can use an additional memory that is readable
and writable. Our goal is to minimize the size of the additional memory while keeping the running time
fast. We measure space consumption in the number of bits used (instead of words) within the additional
memory.
3
Patience Sorting
Since our algorithms are based on the classic Patience Sorting, we start by describing it in detail and
recalling some important properties regarding its internal configurations.
Internally, the algorithm maintains a collection of piles. A pile is a stack of integers. It is equipped
with the procedures push and top: the push procedure appends a new element to become the new top
of the pile; and the top procedure simply returns the element on top of the pile, which is always the one
that was added last.
We describe how Patience Sorting computes lis(τ ). See Algorithm 1. The algorithm scans the
input τ from left to right (Line 2). It tries to push each newly read element τ (i) to a pile with a top
element larger than or equal to τ (i). If on the one hand there is no such a pile, Patience Sorting
creates a new pile to which it pushes τ (i) (Line 4). On the other hand, if at least one such pile exists,
Patience Sorting pushes τ (i) to the oldest pile that satisfies the property (Line 6). After the scan,
the number of piles is the output, which happens to be equal to lis(τ ) (Line 8).
Algorithm 1 Patience Sorting
1: set ℓ := 0 and initialize the dummy pile P0 with the single element −∞
2: for i = 1 to n do
3:
if τ (i) > top(Pℓ ) then
4:
increment ℓ, let Pℓ be a new empty pile, and set j := ℓ
5:
else
6:
set j to be the smallest index with τ (i) ≤ top(Pj )
7:
push τ (i) to Pj
8: return ℓ
We return to the sequence τ1 = h2, 8, 4, 9, 5, 1, 7, 6, 3i for an example. The following illustration shows
the execution of Algorithm 1 on τ1 . In each step the bold number is the newly added element. The colored
(and underlined) elements in the final piles form a longest increasing subsequence τ1 [1, 3, 5, 8] = h2, 4, 5, 6i,
which can be extracted as described below.
2 8
P1 P2
2
P1
4
2 8
P1 P2
4
2 8 9
P1 P2 P3
4 5
2 8 9
P1 P2 P3
1 4 5
2 8 9
P1 P2 P3
1 4 5
2 8 9 7
P1 P2 P3 P4
1 4 5 6
2 8 9 7
P1 P2 P3 P4
3
1 4 5 6
2 8 9 7
P1 P2 P3 P4
Proposition 3.1 ([37, 17, 2]). Given a sequence τ of length n, Patience Sorting computes lis(τ ) in
O(n log n) time using O(n log n) bits of working space.
3.1
Correctness of Patience Sorting
It is observed in [8] that when the input is a permutation π, the elements of each pile form a decreasing
subsequence of π. This observation easily generalizes as follows.
Observation 3.2. Given a sequence τ , the elements of each pile constructed by Patience Sorting
form a non-increasing subsequence of τ .
Hence, any increasing subsequence of τ can contain at most one element in each pile. This implies
that lis(τ ) ≤ ℓ.
Now we show that lis(τ ) ≥ ℓ. Using the piles, we can obtain an increasing subsequence of length ℓ,
in reversed order, as follows [2]:
1. Pick an arbitrary element of Pℓ ;
3
2. For 1 ≤ i < ℓ, let τ (h) be the element picked from Pi+1 . Pick the element τ (h′ ) that was the top
element of Pi when τ (h) was pushed to Pi+1 .
Since h′ < h and τ (h′ ) < τ (h) in each iteration, the ℓ elements that are selected form an increasing
subsequence of τ . This completes the correctness proof for Patience Sorting.
The proof above can be generalized to show the following characterization for the piles.
Proposition 3.3 ([8]). τ (i) ∈ Pj if and only if a longest increasing subsequence of τ ending at τ (i) has
length j.
3.2
Time and space complexity of Patience Sorting
Observe that at any point in time, the top elements of the piles are ordered increasingly from left to right.
Namely, top(Pk ) < top(Pk′ ) if k < k ′ . This is observed in [8] for inputs with no repeated elements. We
can see that the statement holds also for inputs with repetitions.
Observation 3.4. At any point in time during the execution of Patience Sorting and for any k and
k ′ with 1 ≤ k < k ′ ≤ ℓ, we have top(Pk ) < top(Pk′ ) if Pk and Pk′ are nonempty.
Proof. We prove the statement by contradiction. Let i be the first index for which Patience Sorting
pushes τ (i) to some pile Pj , so that the statement of the observation becomes false.
First assume that top(Pj ) ≥ top(Pj ′ ) for some j ′ > j. Let τ (i′ ) be the element in Pj pushed to the
pile right before τ (i). By the definition of Patience Sorting, it holds that
τ (i′ ) ≥ τ (i) = top(Pj ) ≥ top(Pj ′ ).
This contradicts the minimality of i because τ (i′ ) was the top element of Pj before τ (i) was pushed to
Pj .
Next assume that top(Pj ′ ) ≥ top(Pj ) for some j ′ < j. This case contradicts the definition of
Patience Sorting since τ (i) = top(Pj ) ≤ top(Pj ′ ) and thus τ (i) actually has to be pushed to a pile
with an index smaller or equal to j ′ .
The observation above implies that Line 6 of Algorithm 1 can be executed in O(log n) time by using
binary search. Hence, Patience Sorting runs in O(n log n) time.
The total number of elements in the piles is O(n) and thus Patience Sorting consumes O(n log n)
bits. If it maintains all elements in the piles, it can compute an actual longest increasing subsequence
in the same time and space complexity as described above. Note that to compute lis(τ ), it suffices
to remember the top elements of the piles. However, the algorithm still uses Ω(n log n) bits when
lis(τ ) ∈ Ω(n).
3.3
A simple O(n)-bits algorithm
Here we observe that, when the input is a permutation π of {1, . . . , n}, lis(π) can be computed in O(n2 )
time with O(n) bits of working space. The algorithm maintains a used/unused flag for each number in
{1, . . . , n}. Hence, this noncomparison-based algorithm cannot be generalized for general inputs directly.
Let τ be a sequence of integers without repetitions. A subsequence τ [i1 , . . . , iℓ ] is the left-to-right
minima subsequence if {i1 , . . . , iℓ } = {i : τ (i) = min{τ (j) : 1 ≤ j ≤ i}}. In other words, the left-to-right
minima subsequence is made by scanning τ from left to right and greedily picking elements to construct
a maximal decreasing subsequence.
Burstein and Lankham [8, Lemma 2.9] showed that the first pile P1 is the left-to-right minima
subsequence of π and that the ith pile Pi is the left-to-right minima subsequence of a sequence obtained
from π by removing all elements in the previous piles P1 , . . . , Pi−1 .
Algorithm 2 below uses this characterization of piles. The correctness follows directly from the
characterization. It uses a constant number of pointers of O(log n) bits and a Boolean table of length n
for maintaining “used” and “unused” flags. Thus it uses n + O(log n) bits working space in total. The
running time is O(n2 ): each for-loop takes O(n) time and the loop is repeated at most n times.
4
Algorithm 2 Computing lis(π) with O(n) bits and in O(n2 ) time
1: set ℓ := 0 and mark all elements in π as “unused”
2: while there is an “unused” element in π do
3:
increment ℓ and set t := ∞
4:
for i = 1 to n do
⊲ this for-loop constructs the next pile implicitly
5:
if π(i) is unused and π(i) < t then
6:
mark π(i) as “used” and set t := π(i)
⊲ t is currently on top of Pℓ
7: return ℓ
4
An algorithm for computing the length
In this section,
we present our main algorithm that computes lis(τ ) with O(s log n) bits in O( 1s · n2 log n)
√
time for n ≤ s ≤ n. Note that the algorithm here outputs the length lis(τ ) only. The next section
discusses efficient solutions to actually compute a longest sequence.
In the following, by Pi for some i we mean the ith pile obtained by (completely) executing Patience
Sorting unless otherwise stated. (We sometimes refer to a pile at some specific point of the execution.)
Also, by Pi (j) for 1 ≤ j ≤ |Pi | we denote the jth element added to Pi . That is, Pi (1) is the first element
added to Pi and Pi (|Pi |) is the top element of Pi .
To avoid mixing up repeated elements, we assume that each element τ (j) of the piles is stored with
its index j. In the following, we mean by “τ (j) is in Pi ” that the jth element of τ is pushed to Pi . Also,
by “τ (j) is Pi (r)” we mean that the jth element of τ is the rth element of Pi .
We start with an overview of our algorithm. It scans over the input O(n/s) times. In each pass,
it assumes that a pile Pi with at most s elements is given, which has been computed in the previous
pass. Using this pile Pi , it filters out the elements in the previous piles P1 , . . . , Pi−1 . It then basically
simulates Patience Sorting but only in order to compute the next 2s piles. As a result of the pass, it
computes a new pile Pj with at most s elements such that j ≥ i + s.
The following observation, that follows directly from the definition of Patience Sorting and Observation 3.4, will be useful for the purpose of filtering out elements in irrelevant piles.
Observation 4.1. Let τ (y) ∈ Pj with j 6= i. If τ (x) was the top element of Pi when τ (y) was pushed to
Pj , then j < i if τ (y) < τ (x), and j > i if τ (y) > τ (x).
Using Observation 4.1, we can obtain the following algorithmic lemma that plays an important role
in the main algorithm.
Lemma 4.2. Having stored Pi explicitly in the additional memory and given an index j > i, the size
|Pk | for all i + 1 ≤ k ≤ min{j, lis(τ )} can be computed in O(n log n) time with O((|Pi | + j − i) log n) bits.
If lis(τ ) < j, then we can compute lis(τ ) in the same time and space complexity.
Proof. Recall that Patience Sorting scans the sequence τ from left to right and puts each element to
the appropriate pile. We process the input in the same way except that we filter out, and thereby ignore,
the elements in the piles Ph for which h < i or h > j.
To this end, we use the following two filters whose correctness follows from Observation 4.1.
(Filtering Ph with h < i.) To filter out the elements that lie in Ph for some h < i, we maintain an
index r that points to the element of Pi read most recently in the scan. Since Pi is given explicitly to
the algorithm, we can maintain such a pointer r.
When we read a new element τ (x), we have three cases.
• If τ (x) is Pi (r + 1), then we increment the index r.
• Else if τ (x) < Pi (r), then τ (x) is ignored since it is in Ph for some h < i.
• Otherwise we have τ (x) > Pi (r). In this case τ (x) is in Ph for some h > i.
(Filtering Ph with h > j.) The elements in Ph for h > j can be filtered without maintaining additional
information as follows. Let again τ (x) be the newly read element.
• If no part of Pj has been constructed yet, then τ (x) is in Ph for some h ≤ j.
5
• Otherwise, we compare τ (x) and the element τ (y) currently on the top of Pj .
– If τ (x) > τ (y), then τ (x) is in Ph for some h > j, and thus ignored.
– Otherwise τ (x) is in Ph for some h ≤ j.
We simulate Patience Sorting only for the elements that pass both filters above. While doing so,
we only maintain the top elements of the piles and additionally store the size of each pile. This requires
at most O((j − i) log n) space, as required by the statement of the lemma. For details see Algorithm 3.
The running time remains the same since we only need constant number of additional steps for each
step in Patience Sorting to filter out irrelevant elements. If Pj is still empty after this process, we
can conclude that lis(τ ) is the index of the newest pile constructed.
Algorithm 3 Computing |Pk | for all k with i + 1 ≤ k ≤ min{j, lis(τ )} when Pi is given
1: set r := 0
⊲ r points to the most recently read element in Pi
2: set ℓ := i
⊲ the largest index of the piles constructed so far
3: initialize pi+1 , . . . , pj to ∞
⊲ pk is the element currently on top of Pk
4: initialize ci+1 , . . . , cj to 0
⊲ ck is the current size of Pk
5: for x = 1 to n do
⊲ filtering out irrelevant elements
6:
if τ (x) is Pi (r + 1) then
7:
increment r and continue the for-loop
8:
else if τ (x) < Pi (r) or (ℓ ≥ j and τ (x) > pj ) then
9:
ignore the element and continue the for-loop
⊲ push τ (x) to the appropriate pile
10:
if τ (x) > pℓ then
11:
increment ℓ and set h := ℓ
12:
else
13:
set h to be the smallest index with τ (i) < ph
14:
set ph := τ (x) and increment ch
The proof of Lemma 4.2 can be easily adapted to also compute the pile Pj explicitly. For this, we
simply additionally store all elements of Pj as they are added to the pile.
Lemma 4.3. Given Pi and an index j such that i < j ≤ lis(τ ), we can compute Pj in O(n log n) time
with O((|Pi | + |Pj | + j − i) log n) bits.
Assembling the lemmas of this section, we now present our first main result. The corresponding
pseudocode of the algorithm can be found in Algorithm 4.
√
Theorem 4.4. There is an algorithm that, given an integer s satisfying n ≤ s ≤ n and a sequence τ
of length n, computes lis(τ ) in O( 1s · n2 log n) time with O(s log n) bits of space.
Proof. To apply Lemmas 4.2 and 4.3 at the beginning, we start with a dummy pile P0 with a single
dummy entry P0 (1) = −∞. In the following, assume that for some i ≥ 0 we computed the pile Pi of size
at most s explicitly. We repeat the following process until we find lis(τ ).
In each iteration, we first compute the size |Pk | for i + 1 ≤ k ≤ i + 2s. During this process, we may
find lis(τ ) < i + 2s. In such a case we output lis(τ ) and
√ find√an index j such
√ terminate. Otherwise, we
that i + s + 1 ≤ j ≤ i + 2s and |Pj | ≤ n/s. Since s ≥ n, it holds that |Pj | ≤ n/ n = n ≤ s. We then
compute Pj itself to replace i with j and repeat.
By Lemmas 4.2 and 4.3, each pass can be executed in O(n log n) time with O(s log n) bits. There are
at most lis(τ )/s iterations, since in each iteration the index i increases by at least s or lis(τ ) is determined.
Since lis(τ ) ≤ n, the total running time is O( 1s · n2 log n).
In the case of the smallest memory consumption we conclude the following corollary.
Corollary
4.5. Given a sequence τ of length n, lis(τ ) can be computed in O(n1.5 log n) time with
√
O( n log n) bits of space.
6
Algorithm 4 Computing lis(τ ) with O(s log n) bits in O( 1s · n2 log n) time
1: set i := 0 and initialize the dummy pile P0 with the single element −∞
2: loop
3:
compute the size of Pk for all k with i + 1 ≤ k ≤ i + 2s
4:
if we find lis(τ ) < i + 2s then
5:
return lis(τ )
6:
let j be the largest index such that |Pj | ≤ s
7:
compute Pj and set i := j
5
⊲ i + s + 1 ≤ j ≤ i + 2s
An algorithm for finding a longest increasing subsequence
It is easy to modify the algorithm in the previous section in such a way that it outputs an element of the
final pile Plis(τ ) , which is the last element of a longest increasing subsequence by Proposition 3.3. Thus
we can repeat the modified algorithm n times (considering only the elements smaller than and appearing
before the last output) and actually find a longest increasing subsequence.1 The running time of this
naı̈ve approach is O( 1s · n3 log n).
As we claimed before, we can do much better. In fact, we need only an additional multiplicative
factor of O(log n) instead of O(n) in the running time, while keeping the space complexity as it is. In
the rest of this section, we prove the following theorem.
√
Theorem 5.1. There is an algorithm that, given an integer s satisfying n ≤ s ≤ n and a sequence τ
of length n, computes a longest increasing subsequence of τ in O( 1s · n2 log2 n) time using O(s log n) bits
of space.
Corollary 5.2. Given a sequence
τ of length n, a longest increasing subsequence of τ can be found in
√
O(n1.5 log2 n) time with O( n log n) bits of space.
We should point out that the algorithm in this section is not a multi-pass algorithm. However, we
can easily transform it without any increase in the time and space complexity so that it works as a
sequential access algorithm.
5.1
High-level idea
We first find an element that is in a longest increasing subsequence roughly in the middle. As we will
argue, this can be done in O( 1s · n2 log n) time with O(s log n) bits by running the algorithm from the
previous section twice, once in the ordinary then once in the reversed way. We then divide the input
into the left and right parts at a near-mid element and recurse.
The space complexity remains the same and the time complexity increases only by an O(log n)
multiplicative factor. The depth of recursion is O(log n) and at each level of recursion the total running
time is O( 1s · n2 log n). To remember the path to the current recursion, we need some additional space,
but it is bounded by O(log2 n) bits.
5.2
A subroutine for short longest increasing sequences
We first solve the base case in which lis(τ ) ∈ O(n/s). In this case, we use the original Patience Sorting
and repeat it O(n/s) times. We present the following general form first.
Lemma 5.3. Let τ be a sequences of length n and lis(τ ) = k. Then a longest increasing subsequence of
τ can be found in O(k · n log k) time with O(k log n) bits.
Proof. Without changing the time and space complexity, we can modify the original Patience Sorting
so that
• it maintains only the top elements of the piles;
1 This algorithm outputs a longest increasing subsequence in the reversed order. One can access the input in the reversed
order and find a longest decreasing subsequence to avoid this issue.
7
• it ignores the elements larger than or equal to a given upper bound; and
• it outputs an element in the final pile.
We run the modified algorithm lis(τ ) times. In the first run, we have no upper bound. In the
succeeding runs, we set the upper bound to be the output of the previous run. In each run the input to
the algorithm is the initial part of the sequence that ends right before the last output. The entire output
forms a longest increasing sequence of τ .2
Since lis(τ ) = k, modified Patience Sorting maintains only k piles. Thus each run takes O(n log k)
time and uses O(k log n) bits. The lemma follows since this is repeated k times and each round only
stores O(log n) bits of information from the previous round.
√
The following special form of the lemma above holds since n/s ≤ s when s ≥ n.
√
Corollary 5.4. Let τ be a sequence of length n and lis(τ ) ∈ O(n/s) for some s with n ≤ s ≤ n. A
longest increasing subsequence of τ can be found in O( 1s · n2 log n) time with O(s log n) bits.
5.3
A key lemma
As mentioned above, we use a reversed version of our algorithm. Reverse Patience Sorting is the
reversed version of Patience Sorting: it reads the input from right to left and uses the reversed inequalities. (See Algorithm 5.) Reverse Patience Sorting computes the length of a longest decreasing
subsequence in the reversed sequence, which is a longest increasing subsequence in the original sequence.
Since the difference between the two algorithms is small, we can easily modify our algorithm in Section 4
for the length so that it simulates Reverse Patience Sorting instead of Patience Sorting.
Algorithm 5 Reverse Patience Sorting
1: set ℓ := 0 and initialize the dummy pile Q0 with the single element +∞
2: for i = n to 1 do
3:
if τ (i) < top(Qℓ ) then
4:
increment ℓ, let Qℓ to be a new empty pile, and set j := ℓ
5:
else
6:
set j to be the smallest index with τ (i) > top(Qj )
7:
push τ (i) to Qj
8: return ℓ
Let Qi be the ith pile constructed by Reverse Patience Sorting as in Algorithm 5. Using
Proposition 3.3, we can show that for each τ (i) in Qj , the longest decreasing subsequence of the reversal
of τ ending at τ (i) has length j. This is equivalent to the following observation.
Observation 5.5. τ (i) ∈ Qj if and only if a longest increasing subsequence of τ starting at τ (i) has
length j.
This observation immediately gives the key lemma below.
Lemma 5.6. Pk ∩ Qlis(τ )−k+1 6= ∅ for all k with 1 ≤ k ≤ lis(τ ).
Proof. Let hτ (i1 ), . . . , τ (iℓ )i be a longest increasing subsequence of τ . Proposition 3.3 implies that τ (ik ) ∈
Pk . The subsequence hτ (ik ), . . . , τ (iℓ )i is a longest increasing subsequence of τ starting at τ (ik ) since
otherwise hτ (i1 ), . . . , τ (iℓ )i is not longest. Since the length of hτ (ik ), . . . , τ (iℓ )i is iℓ −k +1 = lis(τ )−k +1,
we have τ (k) ∈ Qlis(τ )−k+1 .
Note that the elements of Pk and Qlis(τ )−k+1 are not the same in general. For example, by applying
Reverse Patience Sorting to τ1 = h2, 8, 4, 9, 5, 1, 7, 6, 3i, we get Q1 = h3, 6, 7, 9i, Q2 = h1, 5, 8i,
Q3 = h4i, and Q4 = h2i as below. (Recall that P1 = h2, 1i, P2 = h8, 4, 3i, P3 = h9, 5i, and P4 = h7, 6i.)
The following diagram depicts the situation. The elements shared by Pk and Qlis(τ )−k+1 are colored and
underlined.
2 Again
this output is reversed. We can also compute the output in nonreversed order as discussed before.
8
3
Q1
5.4
6
3
Q1
7
6
3
Q1
7
6
3 1
Q1 Q2
7
6 5
3 1
Q1 Q2
9
7
6 5
3 1
Q1 Q2
9
7
6 5
3 1 4
Q1 Q2 Q3
9
7
6
3
Q1
8
5
1 4
Q2 Q3
9
7
6
3
Q1
8
5
1 4 2
Q2 Q3 Q4
3
1 4 5 6
2 8 9 7
P1 P2 P3 P4
The algorithm
We first explain the subroutine for finding a near-mid element in a longest increasing subsequence.
√
Lemma 5.7. Let s be an integer satisfying n ≤ s ≤ n. Given a sequence τ of length n, the kth element
of a longest increasing subsequence of τ for some k with lis(τ )/2 ≤ k < lis(τ )/2 + n/s can be found in
O( 1s · n2 log n) time using O(s log n) bits of space.
Proof. We slightly modify Algorithm 4 so that it finds an index k and outputs Pk such that |Pk | ≤ s and
lis(τ )/2 ≤ k ≤ lis(τ )/2 + n/s. Such a k exists since the average of |Pi | for lis(τ )/2 ≤ i < lis(τ )/2 + n/s is
at most s. The time and space complexity of this phase are as required by the lemma.
We now find an element in Pk ∩ Qlis(τ )−k+1 . Since the size |Qlis(τ )−k+1 | is not bounded by O(s) in
general, we cannot store Qlis(τ )−k+1 itself. Instead use the reversed version of the algorithm in Section 4
to enumerate it. Each time we find an element in Qlis(τ )−k+1 , we check whether it is included in Pk . This
can be done with no loss in the running time since Pk is sorted and the elements of Qlis(τ )−k+1 arrive in
increasing order.
The next technical but easy lemma allows us to split the input into two parts at an element of a
longest increasing subsequence and to solve the smaller parts independently.
Lemma 5.8. Let τ (j) be the kth element of a longest increasing subsequence of a sequence τ . Let
τL be the subsequence of τ [1, . . . , j − 1] formed by the elements smaller than τ (j). Similarly let τR be
the subsequence of τ [j + 1, . . . , |τ |] formed by the elements larger than τ (j). Then, a longest increasing
subsequence of τ can be obtained by concatenating a longest increasing subsequence of τL , τ (j), and a
longest increasing subsequence of τR , in this order.
Proof. Observe that the concatenated sequence is an increasing subsequence of τ . Thus it suffices to
show that lis(τL ) + lis(τR ) + 1 ≥ lis(τ ). Let τ [i1 , . . . , ilis(τ ) ] be a longest increasing subsequence of τ
such that ik = j. From the definition, τ [i1 , . . . , ik−1 ] is a subsequence of τL , and τ [ik+1 , . . . , ilis(τ ) ] is a
subsequence of τR . Hence lis(τL ) ≥ k − 1 and lis(τR ) ≥ lis(τ )− k, and thus lis(τL )+ lis(τR )+ 1 ≥ lis(τ ).
As Lemma 5.8 suggests, after finding a near-mid element τ (k), we recurse into τL and τR . If the
input τ ′ to a recursive call has small lis(τ ′ ), we directly compute a longest increasing subsequence. See
Algorithm 6 for details of the whole algorithm. Correctness follows from Lemma 5.8 and correctness of
the subroutines.
Algorithm 6 Recursively finding a longest increasing subsequence of ρ
1: RecursiveLIS(ρ, −∞, +∞)
2: procedure RecursiveLIS(τ , lb, ub)
3:
τ ′ := the subsequence of τ formed by the elements τ (i) such that lb < τ (i) < ub
⊲ τ ′ is not explicitly computed but provided by ignoring the irrelevant elements
′
4:
compute lis(τ )
5:
if lis(τ ′ ) ≤ 3|τ ′ |/s then
6:
output a longest increasing subsequence of τ ′
⊲ Lemma 5.3
7:
else
8:
find the kth element τ ′ (j) of a longest increasing subsequence of τ ′
for some k with lis(τ ′ )/2 ≤ k < lis(τ ′ )/2 + |τ ′ |/s
9:
RecursiveLIS(τ ′ [1, . . . , j − 1], lb, τ ′ (j))
10:
output τ ′ (j)
11:
RecursiveLIS(τ ′ [j + 1, . . . , |τ ′ |], τ ′ (j), ub)
9
5.5
Time and space complexity
In Theorem 5.1, the claimed running time is O( 1s · n2 log2 n). To prove this, we first show that the
depth of the recursion is O(log n). We then show that the total running time in each recursion level is
O( 1s · n2 log n). The claimed running time is guaranteed by these bounds.
Lemma 5.9. Given a sequence τ , the depth of the recursions invoked by RecursiveLIS of Algorithm 6
is at most log6/5 lis(τ ′ ), where τ ′ is the subsequence of τ computed in Line 3.
Proof. We proceed by induction on lis(τ ′ ). If lis(τ ′ ) ≤ 3|τ ′ |/s, then no recursive call occurs, and hence
the lemma holds. In the following, we assume that lis(τ ′ ) = ℓ > 3|τ ′ |/s and that the statement of the
lemma is true for any sequence τ ′′ with lis(τ ′′ ) < ℓ.
Since ℓ > 3|τ ′ |/s, we recurse into two branches on subsequences of τ ′ . From the definition of k
in Line 8 of Algorithm 6, the length of a longest increasing subsequence is less than ℓ/2 + |τ ′ |/s in
each branch. Since ℓ/2 + |τ ′ |/s < ℓ/2 + ℓ/3 = 5ℓ/6, each branch invokes recursions of depth at most
log6/5 (5ℓ/6) = log6/5 ℓ − 1. Therefore the maximum depth of the recursions invoked by their parent is
at most log6/5 ℓ.
Lemma 5.10. Given a sequence τ of length n, the total running time at each depth of recursion excluding
further recursive calls in Algorithm 6 takes O( 1s n2 log n) time.
Proof. In one recursion level, we have many calls of RecursiveLIS on pairwise non-overlapping sub′
|2 log |τ ′ |). Thus the total
sequences of τ . For each subsequence
time O( 1s |τ ′P
P 1 ′ 2τ , the′ algorithm spends
1 2
running time at a depth is O( τ ′ s |τ | log |τ |), which is O( s n log n) since τ ′ |τ ′ |2 ≤ |τ |2 = n2 .
Finally we consider the space complexity of Algorithm 6.
Lemma 5.11. Algorithm 6 uses O(s log n) bits of working space on sequences of length n.
Proof. We have already shown that each subroutine uses O(s log n) bits. Moreover, this space of working
memory can be discarded before another subroutine call occurs. Only a constant number of O(log n)bit words are passed to the new subroutine call. We additionally need to remember the stack trace of
the recursion. The size of this additional information is bounded by O(log2 n) bits since each recursive
call is specified by a constant number of O(log
√ n)-bit words and the depth of recursion is O(log n) by
Lemma 5.9. Since log2 n ∈ O(s log n) for s ≥ n, the lemma holds.
6
Lower bound for algorithms with sequential access
An algorithm is a sequential access algorithm if it can access elements in the input array only sequentially.
In our situation this means that for a given sequence, accessing the ith element of the sequence directly
after having accessed the jth element of the sequence costs time at least linear in |i−j|. As opposed to the
RAM, any Turing machine in which the input is given on single read-only tape has this property. Note
that any lower bound for sequential access algorithms in an asymptotic form is applicable to multi-pass
algorithms as well since every multi-pass algorithm can be simulated by a sequential access algorithm
with the same asymptotic behavior. Although some of our algorithms are not multi-pass algorithms,
it is straightforward to transform them to sequential access algorithms with the same time and space
complexity.
To show a lower bound on the running time of sequential access algorithms with limited working
space, we need the concept of communication complexity (see [21] for more details). Let f be a function.
Given α ∈ A to the first player Alice and β ∈ B to the second player Bob, the players want to compute
f (α, β) together by sending bits to each other (possibly multiple times). The communication complexity
of f is the maximum number of bits transmitted between Alice and Bob over all inputs by the best
protocol for f .
Consider the following variant of the LIS problem: Alice gets the first half of a permutation π of
{1, . . . , 2n} and Bob gets the second half. They compute lis(π) together. It is known that this problem
has high communication complexity [22, 19, 38].
Proposition 6.1 ([19, 38]). Let π be a permutation of {1, . . . , 2n}. Given the first half of π to Alice and
the second half to Bob, they need Ω(n) bits of communication to compute lis(π) in the worst case (even
with 2-sided error randomization).
10
Now we present our lower bound. Note that the lower bound even holds for the special case where
input is restricted to permutations.
Theorem 6.2. Given a permutation π of {1, . . . , 4n}, any sequential access (possibly randomized) algorithm computing lis(π) using b bits takes Ω(n2 /b) time.
Proof. Given an arbitrary n > 1, let π ′ be a permutation of {1, . . . , 2n}. We construct a permutation π
of {1, . . . , 4n} as follows. Let π1′ = hπ(1), . . . , π(n)i be the first half of π ′ , define π2′ = h4n, 4n−1, . . . , 2n+
2i and let π3′ = hπ(n + 1), π(n + 2), . . . , π(2n)i be the second half of π ′ . Then we define π to be the
concatenation of π1′ , π2′ , π3′ and the one element sequence π4′ = h2n + 1i, in that order.
It is not difficult to see that π is a permutation and that lis(π) = lis(π ′ ) + 1. To see the latter, observe
that the concatenation of π2′ and π4′ is a decreasing subsequence of π. Hence any increasing subsequence
of π can contain at most one element not in π ′ . On the other hand, any increasing subsequence of π ′ of
length ℓ can be extended with the element 2n + 1 of π4′ to an increasing subsequence of π of length ℓ + 1.
We say a sequential access algorithm traverses the middle if it accesses a position in π1′ and then
accesses a position in π3′ or vice versa with possibly accessing elements in π2′ but only such elements in
meantime. Since each traversal of the middle takes Ω(n) time, it suffices to show that the number of
traversals of the middle is Ω(n/b).
Suppose we are given a sequential access algorithm M that computes lis(π) with t traversals of the
middle. Using M , we construct a two-player communication protocol for computing lis(π ′ ) with at most
tb bits of communication. (A similar technique is described for streaming algorithms in [38].)
Recall that the first player Alice gets the first half π1′ of π ′ and the second player Bob gets the second
half π3′ of π ′ . They compute lis(π ′ ) together as follows.
• Before starting computation, Alice computes πA by concatenating π1′ and π2′ in that order, and
Bob computes πB by concatenating π2′ , π3′ , and π4′ in that order.
• They first compute lis(π) using M by repeating the following phases:
– Alice starts the computation by M and continues while M stays in π[1, . . . , 3n − 1] = πA .
When M tries to access π[3n, . . . , 4n], and thus a traversal of the middle occurs, Alice stops
and sends all b bits stored by M to Bob.
– Bob restores the b bits received from Alice to the working memory of M and continues
computation while M stays in π[n + 1, . . . , 4n] = πB . A traversal of the middle is occurred
when M tries to access π[1, . . . , n]. Bob then stops and sends the b bits currently stored by
M back to Alice.
• When M outputs lis(π) and terminates, the currently active player outputs lis(π) − 1 as lis(π ′ ) and
terminates the protocol.
The two players correctly simulate M and, as a result, compute lis(π ′ ) together. Since the algorithm M
invokes t traversals, the total number of bits sent is at most tb. Since tb ∈ Ω(n) holds by Proposition 6.1,
we have t ∈ Ω(n/b) as required.
Recall that our algorithms for the LIS problem use O(s log n) bits and runs in O( 1s n2 log n) time
√
for computing the length and in O( 1s n2 log2 n) time for finding a subsequence, where n ≤ s ≤ n. By
Theorem 6.2, their time complexity is optimal for algorithms with sequential access up to polylogarithmic
factors of log2 n and log3 n, respectively.
7
Concluding remarks
√
Our result raises the following question: “Do o( n)-space polynomial-time algorithms for LIS exist?”
An unconditional ‘no’ answer would be surprising as it implies SC 6= P ∩ PolyL, where SC (Steve’s Class)
is the class of problems that can be solved by an algorithm that simultaneously runs in polynomialtime and polylogarithmic-space [11, 29]. A possibly easier question asks for the existence of a log-space
algorithm. For this question, one might be able to give some evidence for a ‘no’ answer by showing
NL-hardness of (a decision version of) LIS.
11
We would like to mention some known results that have a mysterious coincidence in space complexity
with
p our results. For (1 + ǫ)-approximation of
plis(π) by one-pass streaming algorithms, it is known that
O( n/ǫ · log n) bits are sufficient [19] and Ω( n/ǫ) bits are necessary [15, 18]. We were not able to find
any connection here and do not claim anything concrete about this coincidence.
To make the presentation simple, we used n to bound lis(τ ) in the time complexity analyses of the
algorithms. If we carefully analyze the complexity in terms of lis(τ ) instead of n when possible, we can
obtain the following output-sensitive bounds.
√
Theorem 7.1. Let s be an integer satisfying n ≤ s ≤ n, and let τ be a sequence of length n with
lis(τ ) = k. Using O(s log n) bits of space, lis(τ ) can be computed in O( 1s · kn log k) time and a longest
increasing subsequence of τ can be found in O( 1s · kn log2 k) time.
References
[1] Hee-Kap Ahn, Nicola Baraldo, Eunjin Oh, and Francesco Silvestri.
A time-space tradeoff for triangulations of points in the plane.
In COCOON 2017, pages 3–12, 2017.
doi:10.1007/978-3-319-62389-4_1.
[2] David Aldous and Persi Diaconis. Longest increasing subsequences: from patience sorting to the
Baik-Deift-Johansson theorem. Bulletin of the American Mathematical Society, 36(4):413–432, 1999.
doi:10.1090/S0273-0979-99-00796-X.
[3] Tetsuo Asano, Amr Elmasry, and Jyrki Katajainen. Priority queues and sorting for read-only data.
In TAMC 2013, pages 32–41, 2013. doi:10.1007/978-3-642-38236-9_4.
[4] Tetsuo Asano, Taisuke Izumi, Masashi Kiyomi, Matsuo Konagaya, Hirotaka Ono, Yota Otachi,
Pascal Schweitzer, Jun Tarui, and Ryuhei Uehara. Depth-first search using O(n) bits. In ISAAC
2014, pages 553–564, 2014. doi:10.1007/978-3-319-13075-0_44.
[5] Bahareh Banyassady, Matias Korman, Wolfgang Mulzer, André van Renssen, Marcel Roeloffzen,
Paul Seiferth, and Yannik Stein. Improved time-space trade-offs for computing Voronoi diagrams.
In STACS 2017, volume 66, pages 9:1–9:14, 2017. doi:10.4230/LIPIcs.STACS.2017.9.
[6] Sergei Bespamyatnikh and Michael Segal.
quences and patience sorting.
Information
doi:10.1016/S0020-0190(00)00124-1.
Enumerating longest increasing subseProcessing Letters, 76(1–2):7–11, 2000.
[7] Allan Borodin and Stephen Cook. A time-space tradeoff for sorting on a general sequential model
of computation. SIAM Journal on Computing, 11(2):287–297, 1982. doi:10.1137/0211022.
[8] Alexander Burstein and Isaiah Lankham.
Combinatorics of patience
ing piles.
Séminaire Lotharingien de Combinatoire, 54A:B54Ab, 2006.
http://www.mat.univie.ac.at/~slc/wpapers/s54Aburlank.html.
sortURL:
[9] Sankardeep Chakraborty and Srinivasa Rao Satti. Space-efficient algorithms for maximum cardinality search, stack BFS, queue BFS and applications. In COCOON 2017, pages 87–98, 2017.
doi:10.1007/978-3-319-62389-4_8.
[10] Timothy M. Chan and Eric Y. Chen. Multi-pass geometric algorithms. Discrete & Computational
Geometry, 37(1):79–102, 2007. doi:10.1007/s00454-006-1275-6.
[11] Stephen A. Cook. Deterministic CFL’s are accepted simultaneously in polynomial time and log
squared space. In STOC 1979, pages 338–345, 1979. doi:10.1145/800135.804426.
[12] Maxime Crochemore and Ely Porat. Fast computation of a longest increasing subsequence and application. Information and Computation, 208(9):1054–1059, 2010. doi:10.1016/j.ic.2010.04.003.
[13] Omar Darwish and Amr Elmasry. Optimal time-space tradeoff for the 2D convex-hull problem. In
ESA 2014, pages 284–295, 2014. doi:10.1007/978-3-662-44777-2_24.
12
[14] Amr Elmasry, Torben Hagerup, and Frank Kammer. Space-efficient basic graph algorithms. In
STACS 2015, volume 30, pages 288–301, 2015. doi:10.4230/LIPIcs.STACS.2015.288.
[15] Funda Ergun and Hossein Jowhari. On the monotonicity of a data stream. Combinatorica, 35(6):641–
653, 2015. doi:10.1007/s00493-014-3035-1.
[16] Greg N. Frederickson. Upper bounds for time-space trade-offs in sorting and selection. Journal of
Computer and System Sciences, 34(1):19–26, 1987. doi:10.1016/0022-0000(87)90002-X.
[17] Michael L. Fredman. On computing the length of longest increasing subsequences. Discrete Mathematics, 11(1):29–35, 1975. doi:10.1016/0012-365X(75)90103-X.
[18] Anna Gál and Parikshit Gopalan. Lower bounds on streaming algorithms for approximating the
length of the longest increasing subsequence. SIAM Journal on Computing, 39(8):3463–3479, 2010.
doi:10.1137/090770801.
[19] Parikshit Gopalan, T.S. Jayram, Robert Krauthgamer, and Ravi Kumar.
Estimating the sortedness of a data stream.
In SODA 2007, pages 318–327, 2007.
URL:
http://dl.acm.org/citation.cfm?id=1283417.
[20] James W. Hunt and Thomas G. Szymanski. A fast algorithm for computing longest common
subsequences. Communications of the ACM, 20(5):350–353, 1977. doi:10.1145/359581.359603.
[21] Eyal Kushilevitz and Noam Nisan. Communication Complexity. Cambridge University Press, 1997.
[22] David Liben-Nowell, Erik Vee, and An Zhu. Finding longest increasing and common subsequences in streaming data. Journal of Combinatorial Optimization, 11(2):155–175, 2006.
doi:10.1007/s10878-006-7125-x.
[23] Andrea Lincoln, Virginia Vassilevska Williams, Joshua R. Wang, and R. Ryan Williams.
Deterministic time-space trade-offs for k-SUM.
In ICALP 2016, pages 58:1–58:14, 2016.
doi:10.4230/LIPIcs.ICALP.2016.58.
[24] C. L. Mallows. Problem 62-2, patience sorting.
http://www.jstor.org/stable/2028371.
SIAM Review, 4(2):143–149, 1962.
URL:
[25] C. L. Mallows.
Problem 62-2.
http://www.jstor.org/stable/2028347.
Review,
URL:
SIAM
5(4):375–376,
1963.
[26] C. L. Mallows. Patience sorting. Bulletin of the Institute of Mathematics and its Applications,
9:216–224, 1973.
[27] J. Ian Munro and Mike S. Paterson. Selection and sorting with limited storage. Theoretical Computer
Science, 12(3):315–323, 1980. doi:10.1016/0304-3975(80)90061-4.
[28] Timothy Naumovitz and Michael Saks. A polylogarithmic space deterministic streaming algorithm for approximating distance to monotonicity. In SODA 2015, pages 1252–1262, 2015.
doi:10.1137/1.9781611973730.83.
[29] Noam Nisan. RL ⊆ SC. In STOC 1992, pages 619–623, 1992. doi:10.1145/129712.129772.
[30] Jakob Pagter and Theis Rauhe. Optimal time-space trade-offs for sorting. In FOCS 1998, pages
264–268, 1998. doi:10.1109/SFCS.1998.743455.
[31] Michal Pilipczuk and Marcin Wrochna. On space efficiency of algorithms working on structural decompositions of graphs.
In STACS 2016, volume 47, pages 57:1–57:15, 2016.
doi:10.4230/LIPIcs.STACS.2016.57.
[32] Prakash Ramanan.
Tight Ω(n lg n) lower bound for finding a longest increasing
subsequence.
International Journal of Computer Mathematics, 65(3–4):161–164, 1997.
doi:10.1080/00207169708804607.
13
[33] Dan Romik. The surprising mathematics of longest increasing subsequences. Cambridge University
Press, 2015. doi:10.1017/CBO9781139872003.
[34] Michael Saks and C. Seshadhri.
Space efficient streaming algorithms for the distance
to monotonicity and asymmetric edit distance.
In SODA 2013, pages 1698–1709, 2013.
doi:10.1137/1.9781611973105.122.
[35] Michael Saks and C Seshadhri. Estimating the longest increasing sequence in polylogarithmic time.
SIAM Journal on Computing, 46(2):774–823, 2017. doi:10.1137/130942152.
[36] Walter J. Savitch.
Relationships between nondeterministic and deterministic tape
complexities.
Journal of Computer and System Sciences,
4(2):177–192, 1970.
doi:10.1016/S0022-0000(70)80006-X.
[37] Craige Schensted. Longest increasing and decreasing subsequences. Canadian Journal of Mathematics, 13(2):179–191, 1961. doi:10.4153/CJM-1961-015-3.
[38] Xiaoming Sun and David P. Woodruff. The communication and streaming complexity of computing
the longest common and increasing subsequences. In SODA 2007, pages 336–345, 2007. URL:
http://dl.acm.org/citation.cfm?id=1283383.1283419.
[39] Joshua R. Wang. Space-efficient randomized algorithms for K-SUM. In ESA 2014, pages 810–829,
2014. doi:10.1007/978-3-662-44777-2_67.
14
| 8cs.DS
|
REPRESENTATION GROWTH OF SOME TORSION-FREE
FINITELY GENERATED 2-NILPOTENT GROUPS.
arXiv:1711.03849v1 [math.GR] 10 Nov 2017
MICHELE ZORDAN
Abstract. We devise a new method for computing representation zeta functions of torsion-free finitely generated 2-nilpotent groups whose associated Lie
lattices have an extra smoothness condition. This method is used first to derive intrinsic formulae for the abscissa of convergence of such representation
zeta functions; and secondly, as a practical application, to compute global,
local and topological representation zeta functions of groups within an infinite
family containing the Heisenberg group over rings of integers in number fields.
1. Introduction
Let G be a torsion-free finitely generated nilpotent group (T-group). Let ãi (G)
(i ∈ N) be the number of irreducible n-dimensional complex characters of G up to
tensoring by one-dimesional characters. The representation zeta function of G is
X
ζG (s) =
ãi (G)i−s .
i∈N
The abscissa of convergence of this Dirichlet series (when finite) gives the degree of
polynomial growth of the partial sums of the ãi (G)’s (cf. e.g. [20, Section 1.1]).
1.1. A family of T-groups. The main results of this paper concern representation
zeta functions of torsion-free finitely generated 2-nilpotent groups (T2 -groups). Of
particular interest will be those arising from 2-nilpotent Lie lattices over rings of
integers in number fields. Let henceforth K be a number field with ring of integers
O. Stasinski and Voll [20, Section 2.4] associate a unipotent group scheme GΛ with
any 2-nilpotent O-Lattice Λ. Theorem 2.10 gives a general method to compute
ζGΛ (s) when Λ has some additional smoothness property. As this is rather technical,
we first record some zeta functions obtained with this method. We shall consider
Lie lattices obtained from the following Z-Lie lattices.
Definition 1.1. Let m, n ∈ N. We define the following Z-Lie lattice:
Gm×n = hc1 , . . . , cm+n , zij , 1 ≤ i ≤ m, 1 ≤ j ≤ n | [ci , cm+j ] = zij i.
In this definition, non-specified Lie brackets which do not follow formally from the
given ones are understood to be zero.
Let m, n be as above, we define Λm×n = Gm×n ⊗Z O. We denote by Gm×n the
group scheme GΛm×n .
1.1.1. Representation zeta function of Gm×n (O). Fix throughout m, n ∈ N. Without loss of generality we may assume that m ≤ n. Indeed the Z-Lie lattices Gm×n
and Gn×m are isomorphic, so also the O-group schemes Gm×n and Gn×m are.
Theorem A. Let O be the ring of integers of a number field K, and let ζK be the
Dedekind zeta function of K. Then
ζGm×n (O) (s) =
m−1
Y
i=0
ζK (s − n − i)
.
ζK (s − i)
2
MICHELE ZORDAN
This theorem compares to a number of known results. First, in loc. cit. Stasinski and
Voll also computed representation zeta functions of groups in three infinite families
arising from their construction. The Lie lattice Gm×n coincides with their Gn when
m = n; so Theorem A is a generalization of [20, Theorem B] for the groups Gn (O).
In particular the properties of the Dedekind zeta function imply that the results on
functional equation and abscissa of convergence in [20, Corollary 1.3] hold, mutatis
mutandis, for the representation zeta function of Gm×n (O). Also the statement on
meromorphic continuation holds but it is now a consequence of the more general
[5, Theorem A] by Dung and Voll, which studies the analytic properties (such as
meromorphic continuation and rationality of the abscissa of convergence) of the
representation zeta function of a T-group arising from a unipotent group scheme.
Secondly, Theorem A also compares to [4, Section 1.3]. There Carnevale, Schechter and Voll consider Lie lattices obtained from Gn by adding an extra linear relation
for its generators. Finally, the groups Gm×n (O) are a generalization of the Grenham’s groups G1×n used by Snocken to prove that every rational number may be
attained as abscissa of convergence of the representation zeta function of a T2 -group
(cf. [19, Theorem 4.22]). In particular Theorem A may be used to recover Snocken’s
result. Indeed, if ×kZ Gm×n denotes the k-fold central product of Gm×n , it is known
that
ζ×kZ Gm×n (O) (s) = ζGm×n (O) (ks).
This clearly has abscissa of convergence (m+n)/k. Notice that by [10, Theorem 1.5]
the abscissa of convergence must be rational (cf. also [5, Theorem A]).
1.1.2. Local representation zeta function. For a non-zero prime ideal p of O we
denote by Op the completion of O at p. Let Λ be a 2-nilpotent O-Lie lattice. By
[20, Proposition 2.2] one has the following Euler factorization:
Y
(1.1)
ζGΛ (O) (s) =
ζGΛ (Op ) (s).
p
The (global) representation zeta function in Theorem A is obtained by computing the local factors on the right-hand side of (1.1). We shall need the following
notation.
Notation 1.2. Henceforth X, Y and Z will denote indeterminates in the field
Q(X, Y, Z). We define the following objects.
i. Let N ∈ N, we write
(N )X = (1 − X N )
(0) = 1
(N )X ! = (1)X (2)X · · · (N )X
(0)! = 1.
ii. Gauss polynomial. For a, b ∈ N b ≤ a, we define
a
(a)!
.
=
(a − b)!(b)!
b X
iii. Let j ∈ N . We write [j]0 for the set {0, . . . , j}.
iv. Let ℓ ∈ N. We write {i1 , . . . , il }< for an ordered subset of N, i.e. a subset
such that i1 < i2 < · · · < iℓ .
v. X-multinomial coefficient. Let j ∈ N and let I = {i1 , . . . , iℓ }< ⊆ [j − 1]0 .
We write
j
j
iℓ
i2
=
···
.
I
i1 X
iℓ X iℓ−1 X
vi. Pochhammer symbol. Let k ∈ N. We define
(X; Y )k =
k
Y
(1 − XY i ).
i=0
REPRESENTATION GROWTH OF SOME TORSION-FREE NILPOTENT GROUPS
3
Definition 1.3. Let I = {i1 , . . . , iℓ }< ⊆ [m − 1]0 . We define
n
I
(X)
=
fG
(X i1 +1 ; X)m−i1 .
m×n
I +n−m X
Theorem B. Let p be a non-zero prime ideal of O and q be the cardinality of the
residue field of Op . Then
(1.2)
X
ζGm×n (Op ) (s) =
I
(q −1 )
fG
m×n
X
i∈I
I⊆[m−1]0
q (m−i)(n+i)−(m−i)s
,
1 − q (m−i)(n+i)−(m−i)s
where I under the summation symbol denotes an ordered subset.
1.1.3. Topological zeta functions. In [17] Rossmann defines topological representation zeta functions for T-groups arising from unipotent group schemes. The formula
in Theorem A (or rather its proof as we shall see) entails a corresponding formula
for the topological representation zeta functions.
Corollary C. The topological representation zeta function of Gm×n (Op ) is
m−1
Y
i=0
s−i
.
s−n−i
1.2. Abscissae of convergence. Fix throughout a non-zero prime ideal p of O
and let q be the cardinality of the residue field Fq of Op . Theorem B is obtained by
applying an analogue of [22, Theorem E] to the Op -Lie lattice Gm×n ⊗Z Op . This
gives a general method for computing local factors of representation zeta function of
T2 -groups when the attached Lie lattice has some extra smoothness property called
geometrical smoothness (see Definition 2.3). This result is Theorem 2.10. We
record here two of its consequences that give an intrinsic formula for the abscissa
of convergence of the aforementioned local factors.
Definition 1.4. Let g be an Op -Lie lattice and let d = rkOp g, d′ = rkOp g′ .
Let ḡ be the Fq -Lie algebra obtained from g by reducing it modulo p. For ω ∈
Hom(ḡ, Fq ) r {0} set
Rad(ω) = {x ∈ ḡ | ω([x, y]) = 0 ∀y ∈ ḡ}
ρω (s) = 2(d′ − dimFq (Rad(ω))′ − (d − dimFq Rad(ω))s.
We define
α(g) = max
[
{s ∈ Q | ρω (s) = 0}.
ω∈Hom(ḡ,Fq )r{0}
1.2.1. T2 -groups. Let G be a T-group. As noted by Voll in [12, Chapter III,§ 3.2],
the representation zeta function of G has the following Euler factorization:
Y
ζG (s) =
ζG,p (s)
p prime
P∞
−is
where ζG,p (s) = i=0 ãpi p . In order to state the next main result we need to
recall some facts about the Lie theory of T-groups. Namely, with each T-group G
there is associated a Q-Lie algebra LG (Q) and an injective map log : G → LG (Q)
such that log(G) spans LG (Q) (see [18, Chapter 6] and [6, 7]). This construction is
such that if H is a finite index subgroup of G then LH (Q) = LG (Q). The set log(G)
is not necessarily even an additive subgroup of LG (Q), but by [6, Theorem 4.1],
4
MICHELE ZORDAN
there is f ∈ N such that Gf is LR, i.e. log(Gf ) is a Z-Lie sublattice of LG (Q).
When p ∤ f , by [10, Lemma 8.5],
ζG,p = ζGf ,p .
Moreover f may be chosen so that L = log(Gf ) is such that L′ ⊆ c!L where c is the
nilpotency class of G. In this case Howe’s Kirillov orbit method [9] applies and the
representation zeta function may be computed as the Poincaré series of a matrix of
linear forms (see [21, Proposition 3.1]).
Definition 1.5. Let L be a Lie algebra over K. The commutator matrix of a
basis of L is defined analogously to the commutator matrix for a basis of an Op -Lie
lattice (see Definition 2.1). We say that L has smooth rank-loci when the rankloci of a commutator matrix of L are smooth schemes over K. An analogue of
[22, Lemma 2.1] ensures that this definition is independent of the basis chosen to
define the commutator matrix.
We shall show that, if G is a T-group of nilpotency class 2, LG (Q) having smooth
rank-loci implies that, for almost all p, ζG,p is the Poincaré series of a Zp -Lie lattice
with geometrically smooth rank-loci, to which Theorem 2.10 will apply. The output
will be the following theorem.
Theorem D. Let G be a finitely generated torsion-free 2-nilpotent group. Assume
that LG (Q) has smooth rank-loci. Then, for almost all primes, the abscissa of
convergence of ζG,p is α(log(Gf ) ⊗Z Zp ) where f ∈ N such that Gf is LR and
log(Gf )′ ⊆ 2 log(Gf ).
1.2.2. Unipotent group schemes. The last main result is the analogue of Theorem D
but for the factors in the Euler product (1.1).
Theorem E. Let Λ be a 2-nilpotent O-Lie lattice and let GΛ be the unipotent group
scheme associated with it by [20, Section 2.4]. Assume that Λ ⊗O K has smooth
rank-loci. Then, for almost all non-zero prime ideals p in O, ζGΛ (Op ) has abscissa
of convergence α(Λ ⊗O Op ).
1.3. Context. The first three main results mostly compare, as said, to [5, 17, 20].
Zeta functions analogous to the ones considered here have been introduced and
studied for p-adic analytic groups and arithmetic subgroups of semisimple algebraic
groups over number fields. In that case representations are not counted up to
twisting by one-dimesional characters (see for instance the work of Lubotzky and
Martin [15], Jaikin-Zapirain [11], Larsen and Lubotzky [14], and Avni, Klopsch,
Onn and Voll [2, 3]). Recently Stasinski and Häsä have introduced a representation
zeta function counting, up to twisting, characters of the general linear group over
a compact discrete valuation ring (cf. [8]). Theorem 2.10 applies to analogous
situations whenever a Kirillov orbit method applies.
Theorems D and E compare to the upper and lower bounds obtained by Snocken
and Ezzat in [19, Theorem 4.24]. In a broader sense they compare to the bounds
in the context of p-adic analytic groups obtained in [1, Theorem 1.1] and to the
bounds for subgroup zeta functions featuring in [6, 16].
1.4. Notation. Throughout, Z denotes the integers, N the set of positive integers
and N0 = {0} ∪ N the set of natural numbers. The set of rational numbers is
denoted by Q. If p is a rational prime, Qp and Zp denote the sets of the p-adic
numbers and of the p-adic integers respectively.
Let R be a ring, an R-Lie lattice is a free finitely generated R-module endowed
with a Lie bracket. If g is an R-Lie lattice, we write g′ for its derived Lie sublattice.
REPRESENTATION GROWTH OF SOME TORSION-FREE NILPOTENT GROUPS
5
If not otherwise specified, when R is considered as an R-lie lattice it is always
endowed with the trivial Lie bracket.
Let i, j ∈ N. The ring of i×j matrices with entries in R is denoted by Mati×j (R).
If i = j the identity matrix is denoted by Idj , the diagonal matrix with a1 , . . . , aj ∈
R on the diagonal is denoted by diag(a1 , . . . , aj ).
We have consistently denoted tuples by x, y, z, . . . , their components are denoted
by x1 , x2 , x3 , . . . , y1 , y2 , y3 , . . . , z1 , z2 , z3 , . . . respectively. Often we shall represent
mn-tuples of elements in a ring R as m × n matrices. In this case the components
of the mn-tuple x ∈ Matm×n (R) are denoted by xij for i ∈ {1, . . . , m} and j ∈
{1, . . . , n}.
1.5. Acknowledgements. The formulae for the zeta functions computed here
were first computed with ad-hoc more elementary methods during my doctoral
studies. I wish to thank my PhD supervisor Christopher Voll and I acknowledge
financial support by the School of Mathematics of the University of Southampton
the Faculty of Mathematics of the University of Bielefeld and CRC 701. I am
grateful to Ben Martin for his comments on this work.
I am currently supported by the Research Project G.0939.13N of the Research
Foundation - Flanders (FWO).
2. Poincaré series
The main results of this paper all follow from a variant of [22, Theorem E].
In order to treat Poincaré series arising from T-groups of class 2 and unipotent
group schemes, we shall consider Lie lattices that do not necessarily have finite
abelianization. Let, for the rest of the section, g be as in Definition 1.4. We shall
need the following notational conventions.
2.1. Notation. We briefly recall the definition of commutator matrix and that of
its Poincaré series as they are intended in this context.
Definition 2.1. Let B = {bk }k∈{1,...,d} be an Op -basis of g with the last d′ elements
in g′ . We define the commutator matrix of g with respect to B to be
′
d
X
RB (X1 , . . . , Xd′ ) =
λki,j Xk
k=1
where for i, j ∈ {1, . . . , d}, [bi , bj ] =
Pd′
k=1
i,j∈{1,...,d}
λki,j bd−d′ +k .
Let for the rest of this section B be a basis of g as in Definition 2.1 and R = RB .
We denote by R the matrix obtained from R reducing its entries mod p.
Definition 2.2. Let R be a ring. Let i, j, k ∈ N with k ≤ min{i, j}. We define
!
Idk 0
k
Bi×j (R) =
∈ Mati×j (R) .
0 0
The definition of geometrical smoothness for the rank-loci of R in [22, Definition 1.11] now naturally extends to Lie lattices with infinite abelianization if R
is defined as above. Henceforth we consider Lie lattice structures on (Op )d and
Fdq given by the following identification: tuples in (Op )d (resp. Fdq ) are viewed as
coordinates of elements in g (resp. ḡ) with respect to B (resp. the reduction modulo
p of B).
Definition 2.3. We say that g has geometrically smooth rank-loci when for all µ
2µ ≤ d:
6
MICHELE ZORDAN
′
′
i. every x ∈ Fdq such that rkFq R(x) = 2µ has a lift x ∈ (Op )d such that R(x)
has Smith normal form B2µ
d×d (Op ).
′
ii. For every x ∈ (Op )d such that R(x) has Smith normal form B2µ
d×d (Op ),
′
d′
(ker R(x)) is an isolated Op -submodule of (Op ) .
The fact that this definition is independent of the choice of B is a consequence of
the analogue of [22, Lemma 2.1] in this context.
We define the Poincaré series of g as in [22, Definition 1.6] with the obvious modifications needed to deal with a matrix of linear forms in d′ variables. Let h = ⌊d/2⌋
′
and let r ∈ N. Let w ∈ (Op /pr )d . We define
νR,r (w) = (min{ai , r})i∈{1,...,h} ,
where pa1 , pa1 , . . . , pah , pah are the first 2h elementary divisors of R(w) for a w ∈
′
(Op )d lifting w. Indeed, taking the minimum with r ensures that the this definition
is independent of the choice of w.
Definition 2.4. Let
′
Wr (Op ) = (Op /pr )d r p (Op /pr )d
′
r ∈ N.
Let I = {i1 , . . . , iℓ }< ⊆ [h − 1]0 . Let also i0 = 0 and iℓ+1 = h and
µj = ij+1 − ij
for j ∈ {0, . . . , ℓ};
N=
ℓ
X
rj
for rI = (r1 , . . . , rℓ ) ∈ N|I| .
j=1
The Poincaré series of R is defined as
X
Pℓ
X
O
|NI,rpI (R)| q −s j=1 rj (h−ij ) ,
PR (s) =
I⊆[h−1]0
I={i1 ,...,iℓ }<
rI ∈N|I|
where
O
NI,rpI (R) = {w ∈ WN (Op ) | νR,N (w) = (0, . . . , 0, rℓ , . . . , rℓ ,
| {z } | {z }
µℓ
µℓ−1
rℓ + rℓ−1 , . . . , rℓ + rℓ−1 . . . , N, . . . , N ) ∈ Nh0 }.
| {z }
|
{z
}
µ0
µℓ−2
If B1 is another basis for g as in Definition 2.1, it is known that PR (s) = PRB1 (s),
we shall therefore often call PR (s) the Poincaré series of g.
2.2. Technical result. We recall the following definitions from [22]:
′
Definition 2.5. An R-kernel class c is a subset of Fdq such that for any two
x, w ∈ c
dimFq ker R(x) = dimFq ker R(w)
dimFq (ker R(x))′ = dimFq (ker R(w))′ .
For an R-kernel class c we define
dc = dimFq r
d′c
′
= dimFq r
r = ker R(x) for any x ∈ c
r = ker R(x) for any x ∈ c.
Definition 2.6. A classification by R-kernels is a set of R-kernel classes which
′
are disjoint and cover Fdq . An element of a classification by R-kernels A is called a
kernel A-class.
REPRESENTATION GROWTH OF SOME TORSION-FREE NILPOTENT GROUPS
7
Definition 2.7. Let A be a classification by R-kernels and ℓ ∈ N0 . A set
{c1 , . . . , cℓ }
of kernel A-classes such that dc1 > dc2 > · · · > dcℓ is called a sequence of kernel
A-classes.
Definition 2.8. Let S = {c1 , . . . , cℓ } be a sequence of kernel A-classes
i. If S = ∅ we define FS (R) = {∅}.
ii. If S 6= ∅. We define FS (R) as the set of ℓ-tuples (x1 , . . . , xℓ ) of elements of
Pℓ
′
Fdq such that for j = 1, . . . , ℓ and yj = k=j xk :
(a) yj ∈ cj ,
(b) xj−1 ∈ (ker R (yj ))′ for j > 1.
Definition 2.9. Recall that h = ⌊d/2⌋. Let {i1 , . . . , iℓ }< ⊆ [h − 1]0 . A sequence
of kernel A-classes {c1 , . . . , cℓ } with the property that dcj = d − 2(h − iℓ+1−j ), for
j = 1, . . . , ℓ, is called an I-sequence of kernel A-classes. The set of all I-sequences
of kernel A-classes is denoted by KIA (R).
The following analogue of [22, Theorem E] holds.
Theorem 2.10. Let g be an Op -Lie lattice with d = rkOp g and d′ = rkOp g′ . Let
R be a commutator matrix of g and A be a classification by R-kernels. Assume
that R has geometrically smooth rank-loci. Then the Poincaré series of R is
X
X
′
Y
c∈S
S∈KA (R)
I={i1 ,...,iℓ }< S={c1I,...,c }
ℓ
I⊆[h−1]0
′
|FS (R)| q −(d −dcℓ )
′
′
q d −dc −s
′
′
d−dc
2
1 − q d −dc −s
d−dc
2
.
Proof. The proof closely follows the argument for [22, Theorem E], the only difference is that here the size of the commutator matrix is no longer equal to the number
of its variables and one needs to keep them distinct in the proof. This is straightforward, because loc. cit. is a direct consequence of [22, Proposition 3.3], whose
proof is independent of the size of R: the quantity d there, has to be interpreted
as the number of variables in R, which in the present context is d′ . The current result therefore follows immediately from the analogue of [22, Proposition 3.3] stated
with d′ in the place of d. Indeed, the rest of the argument applies unchanged to
the present situation.
Next result now follows in the same fashion as [22, Theorem A] followed from
[22, Theorem E] (see [22, Section 3.3]).
Theorem 2.11. Assume that the rank-loci of g are geometrically smooth. Then the
Poincaré series of g has abscissa of convergence α(g).
3. T-groups
In this section we explain how to apply Theorem 2.11 to representation zeta
functions of T-groups of nilpotency class 2. From now onwards we assume g to be
nilpotent of nilpotency class 2. This implies that the last d′ rows and columns of
the commutator matrix RB of g with respect to B consist of zeroes. The matrix
obtained by discarding them is a (d − d′ ) × (d − d′ ) antisymmetric matrix of linear
forms in d′ variables. To distinguish it from RB , we call it the trimmed commutator
matrix of g with respect to B and denote it by SB .
Remark 3.1. It is clear from its definition that the Poincaré series is the same for
RB and SB .
8
MICHELE ZORDAN
3.1. Proof of Theorem D. Let f ∈ N be such that H = Gf is LR and L = log(H)
is such that L′ ⊆ 2L. Let B be a Z-basis of log(H) such that its last d′ = rkZ log(H)′
elements span log(H)′ . By [21, Proposition 3.1] and Remark 3.1, for all p ∤ f the
representation zeta function ζG,p is the Poincaré series of gp = log(H) ⊗Z Zp (i.e.
the Poincaré series of RB when B is viewed as a basis of gp ). In order to prove the
theorem it is enough to show that Theorem 2.10 applies, i.e. it suffices to show that,
if L = LG (Q) has smooth rank-loci, then for almost all primes, gp has geometrically
smooth rank-loci. This is immediate as smoothness of the rank-loci of L implies
smoothness of the rank-loci of gp ⊗Zp Qp = L ⊗Q Qp for all p. Which for almost
all primes implies the smoothness of the rank-loci of gp . This implies geometrical
smoothness of the rank-loci as shown in [22, Remark 2.4].
3.2. Proof of Theorem E. Let d = rkO Λ and d′ = rkO Λ′ . Let also B =
{bk }k∈{1,...,d} be an O-basis of Λ such that its last d′ elements are in Λ′ . By
[20, Proposition 2.18], the representation zeta function ζGΛ (Op ) is the Poincaré series of the trimmed commutator matrix SB when B is viewed as an Op basis of
g = Λ ⊗O Op . With the same argument used to prove Theorem D, one shows
that L = Λ ⊗O K having smooth rank-loci implies that Λ ⊗O Op has geometrically smooth rank-loci for almost all prime ideals p. This shows that Theorem 2.10
applies to this context and concludes the proof.
4. An infinite family of groups schemes
As explained in the proof of Theorem E the representation zeta function of
G = Gm×n (Op ) is given by the Poincaré series of a trimmed commutator matrix of
g = Gm×n ⊗Z Op . So d = m + n + mn and d′ = mn. We choose B to be the basis
used in the presentation of Gm×n in Definition 1.1. Accordingly we compute
!
0
M(Xij )
,
(4.1)
SB (X11 , X12 . . . , Xmn ) =
−M(Xij )
0
where M(Xij ) is the generic m × n matrix in the variables Xij , for i ∈ {1, . . . , m}
and j ∈ {1, . . . n}. For convenience, in what follows we represent mn-tuples as
m × n matrices, therefore the labelling of the variables in SB .
We shall now show how to apply Theorem 2.10 to the situation in hand. We
begin by showing that its hypotheses are satisfied for g, i.e. we shall show that
Lemma 4.1. The Lie lattice g has geometrically smooth rank-loci.
In what follows we shall see that this is equivalent to a problem involving lifting
Smith normal forms of m × n matrices with entries in Fq .
First of all we notice that since m ≤ n, for x ∈ Matm×n ( Op ) the matrices RB (x)
and SB (x) can have Frac(Op )-rank at most 2m. Similarly, for x ∈ Matm×n ( Fq ), if
RB and SB are the reductions modulo p of RB and SB respectively, the matrices
RB (x) and SB (x) can have Fq -rank at most 2m.
Secondly, by (4.1), x ∈ Matm×n (Op ) has Smith normal form
diag(a1 , . . . , am ) 0 ,
if and only if RB (x) and SB (x) have Smith normal form
!
diag(a1 , a1 , . . . , am , am ) 0
∈ Mat(m+n+mn)×(m+n+mn) (Op )
0
0
!
diag(a1 , a1 , . . . , am , am ) 0
∈ Mat(m+n)×(m+n) (Op )
0
0
REPRESENTATION GROWTH OF SOME TORSION-FREE NILPOTENT GROUPS
9
respectively. Similarly, x ∈ Matm×n (Fq ) has rank m − k (k ∈ N, k ≤ m) if and only
if both RB (x) and RB (x) have rank 2(m − k).
The discussion above implies that g has geometrically smooth rank-loci if and
only if the two following properties are satisfied for all k ∈ N, k ≤ m:
i. for x ∈ Matm×n (Fq ) of rank m−k it is possible to find a lift b
x ∈ Matm×n (Op )
with Smith normal form Bm−k
(O
).
p
m×n
ii. for x ∈ Matm×n (Op ) with Smith normal form Bm−k
m×n (Op ), ker RB (x) has
isolated derived Lie sublattice.
Both the properties above are easily verified: namely the first one by taking a
matrix whose entries are the Teichmüller lifts of the entries of x, while the second
one because of the following result.
Lemma 4.2. If x ∈ Matm×n (Op ) has Smith normal form Bm−k
m×n (Op ) (k ∈ N, k ≤
m), then ker RB (x) is isolated and isomorphic to
(Gk×(n−m+k) ⊗Z Op ) ⊕ Op mn−k(n−m+k)
as an Op -Lie lattice.
Proof. Let S ∈ GLm (Op ) and T ∈ GLn (Op ) such that SxT = Bm−k
m×n (Op ). Let
!
St 0
,
M=
0 T
and let ϕ̃ be the Op -module automorphism of g/g′ whose matrix in the basis {c1 +
g′ , . . . , cm+n + g′ } is M −1 . Choose one c1i each from ϕ̃(ci + g′ ) for i = 1, . . . , m + n.
Since ϕ̃ is invertible, the set
B1 = {c1i | 1 ≤ i ≤ m + n} ∪ {[c1i , c1j ] | 1 ≤ i ≤ m, m < j ≤ m + n}
is an Op -basis of g such that
M t SB (x)M = SB1 (SxT ).
Moreover, because of the shape of M , the Op -module automorphism ϕ defined by
ci 7→ c1i for i ∈ {1, . . . , m + n} and zij 7→ [c1i , c1j ] for 1 ≤ i ≤ m, m < j ≤ m + n gives
a Lie lattice automorphism of (Op )m+n+mn . Hence ker RB (x) and ker RB1 (SxT )
are isomorphic as Lie lattices and the derived sublattice of the former is isolated if
and only if the derived sublattice of the latter is; which is indeed the case as
!
0
Bm−k
m×n (Op )
SB1 (SxT ) =
−Bm−k
0
n×m (Op )
by definition of B1 . This also shows that ker RB1 (SxT ) is isomorphic to
(Gk×(n−m+k) ⊗Z Op ) ⊕ Op mn−k(n−m+k)
and we conclude.
4.1. Proof of Theorem B. In order to apply Theorem 2.10 we need to have a
classification by RB -kernels. This is easily obtained from the previous argument
for geometrical smoothness. Indeed Lemma 4.2 shows that for x ∈ Matm×n (Fq )
with rkFq x = k ≤ m the derived subalgebra of ker RB (x) is always the same. This
implies that the set of loci of constant rank in Matm×n (Fq ) is a classification by
RB -kernels. For convenience we denote the members of this classification simply
by elements of [m − 1]0 . Namely the kernel class corresponding to rank m − i in
Matm×n (Fq ) will be denoted by i ∈ [m − 1]0 . Notice that with this notation we
exclude the locus of rank 0 (which is just the zero matrix), this is not a limitation
10
MICHELE ZORDAN
as we are only required to plug primitive coordinate vectors into RB in order to
compute its Poincaré series (cf. Definition 2.4).
We shall now rephrase Theorem 2.10 with the current notation and the following
convention. As in Section 2, Fm+n+mn
is viewed as an Fq -Lie algebra, so its derived
q
subalgebra is abelian and we identify it with Matm×n (Fq ) endowed with the trivial
m+n+mn
Lie bracket. For x ∈ Fmn
and so
q , ker RB (x) is a Lie subalgebra of Fq
′
(ker RB (x)) may be viewed as a subalgebra of Matm×n (Fq ). First we rephrase
Definition 2.8 in this notation.
Definition 4.3. Let I be an ordered subset of [m − 1]0 . If
i. I = ∅, we define FI (RB ) = {∅},
ii. I = {i1 , . . . , iℓ }, we define FI (RB ) to be the set of ℓ-tuples (x1 , . . . , xℓ ) of
Pℓ
matrices in Matm×n (Fq ) such that for j = 1, . . . , ℓ and yj = k=j xj :
(a) rkFq yj = m − ij ,
(b) xj−1 is in the derived Lie subalgebra of ker RB (yj ) when this is viewed
inside Matm×n (Fq ) with the convention above.
Secondly, we compute the exponents of q in the formula of Theorem 2.10. Namely
let x ∈ Matm×n (Fq ) of rank m − k (for some k ∈ [m − 1]0 ). Then, by Lemma 4.2,
ker RB (x) ∼
= (Gk×(n−m+k) ⊗Z Fq ) ⊕ Fqmn−k(n−m+k)
(ker RB (x))′ ∼
= Fk(n−m+k)
= (Gk×(n−m+k) ⊗Z Fq )′ ∼
q
as Fq -Lie algebras. It follows that
d − dk = 2m − 2k
d′ − d′k = (m − k)(n + k),
where k denotes a kernel class, as explained at the beginning of this section, and
dk , d′k are as in Definition 2.5.
In conclusion if the formula in Theorem 2.10 is rewritten with the notation above,
one finds that the representation zeta function of G is equal to
X
X q (m−i)(n+i)−(m−i)s
.
|FI (RB )|q −(m−i1 )(n+i1 )
1 − q (m−i)(n+i)−(m−i)s
I⊆[m−1]
i∈I
0
I={i1 ,...,iℓ }<
Notice that here we are summing over sequences of kernel classes i1 , . . . , iℓ with
di1 < · · · < diℓ and not the other way around as in Theorem 2.10. This explains
the multiplication by q −(m−i1 )(n+i1 ) and not by q −(m−iℓ )(n+iℓ ) . Theorem B now
follows by repeatedly applying the following observation.
Lemma 4.4. Let x ∈ Matm×n (Fq ) with rank m−k1 for some k1 ∈ [m−1]0 . Then for
all k2 ∈ [k1 −1]0 the number of matrices y in the derived Lie subalgebra of ker RB (x)
such that x + y has rank m − k2 is equal to the number of k1 × (n − m + k1 ) matrices
of rank k1 − k2 with entries in Fq .
Proof. Let S ∈ GLm (Fq ) and T ∈ GLn (Fq ) such that SxT = Bm−k
m×n (Fq ). An
argument similar to the proof of Lemma 4.2 shows that y ∈ (ker SB (x))′ if and only
if SyT is supported only in the entries yij with m− k1 < i ≤ m and m− k1 < j ≤ n.
The conclusion now follows immediately.
Theorem B in the stated form is recovered by applying the following result.
Lemma 4.5 ([13, Proposition 3.1]). Let i, j ∈ N with i ≤ j and let k ∈ [i]0 . Then
j
|Mati−k
(F
)|
=
(q −k−1 ; q −1 )i−k · q (i−k)(j+k) ,
q
i×j
j − i + k q−1
REPRESENTATION GROWTH OF SOME TORSION-FREE NILPOTENT GROUPS
11
where Mati−k
i×j (Fq ) is the set of m × n matrices with entries in Fq and rank i − k.
4.2. Proof of Theorem A and Corollary C. We shall now rewrite the representation zeta function of Gm×n (Op ) thus proving Theorem A and Corollary C.
4.2.1. Proof of Theorem A. The key will be, as in [20], a multiplicative formula for
the local factors, i.e. for the representation zeta function of Gm×n (Op ). We shall
need the following result.
Lemma 4.6 ([20, Proposition 1.5]). For j ∈ N we have that
X
Y (X i Z)j−i
j
(X −j Y Z; X)j
=
.
(Y X −i1 −1 ; X −1 )j−i1
i
j−i
1 − (X Z)
(Z; X)j
I X −1
i∈I
I⊆[j−1]0
I={i1 ,...,iℓ }<
We need now to write the zeta function (1.2) in a form allowing us to use Lemma 4.6.
Let a, j ∈ N and let I = {i1 , . . . , iℓ }< ⊆ [j−1]0 , we write I +a = {i1 +a, . . . , iℓ +a}< .
We shall need the following translation property for X-multinomial coefficients. Its
proof is a standard computation which we do not report here.
Lemma 4.7. Let a, j ∈ N0 and I ⊆ [j − 1]0 , we have that
i1 +1+a
j+a
j (X
; X)j−i1
=
.
I +a X
I (X i1 +1 ; X)j−i1
Theorem 4.8. The representation zeta function of Gm×n (Op ) may be rewritten as
ζGm×n (Op ) (s) =
(q −s ; q)m
.
(q n−s ; q)m
Proof. First, by Lemma 4.7,
n
(4.2)
(X −i1 −1 ; X −1 )m−i1
n − m + I X −1
m
=
(X −(n−m) X −(i1 +1) ; X −1 )m−i1 .
I X −1
Second, rewriting the geometric series in Theorem B, we find that
Y q (m−i)(n+i)−(m−i)s
Y (q i q n−s )m−i
(4.3)
.
=
1 − (q i q n−s )m−i
1 − q (m−i)(n+i)−(m−i)s
i∈I
i∈I
Using (4.2) and (4.3), (1.2) may be rewritten as
Y (q i q n−s )m−i
X m
.
(q −(n−m) q −(i1 +1) ; q −1 )m−i1
ζGm×n (Op ) =
1 − (q i q n−s )m−i
I q−1
i∈I
I⊆[m−1]0
By Lemma 4.6 we conclude.
Theorem A now follows from the Euler product factorization enjoyed by the Dedekind zeta function.
4.2.2. Proof of Corollary C. The formula for the topological representation zeta
function is a direct consequence of Theorem 4.8. Indeed the topological zeta function of Gm×n (Op ) is the constant term of the power series expansion in q − 1 of
its representation zeta function. Thus, since the constant term of a convolution
of power series is the product of the constant terms of the convolution factors, it
suffices to compute the expansion of
fi (s) =
1 − q s−i
1 − q s−n−i
12
MICHELE ZORDAN
in q + 1 for i ∈ {0, . . . , m − 1}. By formally expanding q z = (1 + (q − 1))z in q − 1
using the binomial series we find that this is
s−i
s−n−i
and we conclude.
References
[1] N. Avni, B. Klopsch, U. Onn, and C. Voll, Representation zeta functions of some compact padic analytic groups, Zeta functions in algebra and geometry, 2012, pp. 295–330. MR2858928
[2]
, Representation zeta functions of compact p-adic analytic groups and arithmetic
groups, Duke Math. J. 162 (2013), no. 1, 111–197. MR3011874
[3]
, Similarity classes of integral p-adic matrices and representation zeta functions of
groups of type A2 , Proc. Lond. Math. Soc. (3) 112 (2016), no. 2, 267–350. MR3471251
[4] A. Carnevale, S. Shechter, and C. Voll, Enumerating traceless matrices over compact discrete
valuation rings, arXiv:1709.02717 (September 2017).
[5] D. H. Dung and C. Voll, Uniform analytic properties of representation zeta functions of
finitely generated nilpotent groups, Trans. Amer. Math. Soc. 369 (2017), no. 9, 6327–6349.
MR3660223
[6] F. J. Grunewald, D. Segal, and G. C. Smith, Subgroups of finite index in nilpotent groups,
Invent. Math. 93 (1988), no. 1, 185–223. MR943928
[7] F. Grunewald and D. Segal, Reflections on the classification of torsion-free nilpotent groups,
Group theory, 1984, pp. 121–158. MR780569
[8] J. Häsä and A. Stasinski, Representation growth of compact linear groups, arXiv:1710.09112
(October 2017).
[9] R. E. Howe, On representations of discrete, finitely generated, torsion-free, nilpotent groups,
Pacific J. Math. 73 (1977), no. 2, 281–305. MR0499004 (58 #16984)
[10] E. Hrushovski, B. Martin, S. Rideau, and R. Cluckers, Definable equivalence relations and
zeta functions of groups, arXiv:0701011 to appear on J. Eur. Math. Soc. (JEMS) (August
2017).
[11] A. Jaikin-Zapirain, Zeta function of representations of compact p-adic analytic groups, J.
Amer. Math. Soc. 19 (2006), no. 1, 91–118 (electronic). MR2169043 (2006f:20029)
[12] B. Klopsch, N. Nikolov, and C. Voll, Lectures on profinite topics in group theory, London
Mathematical Society Student Texts, vol. 77, Cambridge University Press, Cambridge, 2011.
Lectures from the course on Asymptotic Methods in Infinite Group Theory held in Oxford,
September 2007, Edited by Dan Segal. MR2797110
[13] D. Laksov and A. Thorup, Counting matrices with coordinates in finite fields and of fixed
rank, Math. Scand. 74 (1994), no. 1, 19–33. MR1277786
[14] M. Larsen and A. Lubotzky, Representation growth of linear groups, J. Eur. Math. Soc.
(JEMS) 10 (2008), no. 2, 351–390. MR2390327 (2009b:20080)
[15] A. Lubotzky and B. Martin, Polynomial representation growth and the congruence subgroup
problem, Israel J. Math. 144 (2004), 293–316. MR2121543 (2006b:20065)
[16] P. M. Paajanen, On the degree of polynomial subgroup growth in class 2 nilpotent groups,
Israel J. Math. 157 (2007), 323–332. MR2342452
[17] T. Rossmann, Topological representation zeta functions of unipotent groups, J. Algebra 448
(2016), 210–237. MR3438311
[18] D. Segal, Polycyclic groups, Cambridge Tracts in Mathematics, vol. 82, Cambridge University
Press, Cambridge, 1983. MR713786
[19] R.
Snocken,
Zeta
functions
of
groups
and
rings,
Ph.D.
Thesis,
https://eprints.soton.ac.uk/372833/, 2014.
[20] A. Stasinski and C. Voll, Representation zeta functions of nilpotent groups and generating
functions for Weyl groups of type B, Amer. J. Math. 136 (2014), no. 2, 501–550. MR3188068
[21] C. Voll, Functional equations for zeta functions of groups and rings, Ann. of Math. (2) 172
(2010), no. 2, 1181–1218. MR2680489 (2011f:20057)
[22] M. Zordan, Poincaré series of Lie lattices and representation zeta functions of arithmetic
groups, arXiv:1704.04165 (April 2017).
KU Leuven, Departement Wiskunde, Celestijnenlaan 200B, B-3001 Leuven, Belgium.
E-mail address: [email protected]
| 4math.GR
|
An Anytime Algorithm for Task and Motion MDPs
arXiv:1802.05835v1 [cs.AI] 16 Feb 2018
‡
Siddharth Srivastava∗‡ , Nishant Desai† , Richard Freedman§ , Shlomo Zilberstein§
Arizona State University, † United Technologies Research Center, § University of Massachusetts
Abstract
Integrated task and motion planning has emerged
as a challenging problem in sequential decision
making, where a robot needs to compute highlevel strategy and low-level motion plans for solving complex tasks. While high-level strategies require decision making over longer time-horizons
and scales, their feasibility depends on low-level
constraints based upon the geometries and continuous dynamics of the environment. The hybrid nature of this problem makes it difficult to scale; most
existing approaches focus on deterministic, fully
observable scenarios. We present a new approach
where the high-level decision problem occurs in a
stochastic setting and can be modeled as a Markov
decision process. In contrast to prior efforts, we
show that complete MDP policies, or contingent
behaviors, can be computed effectively in an anytime fashion. Our algorithm continuously improves
the quality of the solution and is guaranteed to be
probabilistically complete. We evaluate the performance of our approach on a challenging, realistic
test problem: autonomous aircraft inspection. Our
results show that we can effectively compute consistent task and motion policies for the most likely
execution-time outcomes using only a fraction of
the computation required to develop the complete
task and motion policy.
1
Introduction
In order to be truly helpful, robots will need to be able to
accept commands from humans at high-levels of abstraction,
and autonomously execute them. Consider the problem of inspecting an aircraft (Fig. 1). In order to autonomously plan
and execute such a task, the robots (UAVs in this case) will
need to be able to make high-level inspection decisions on
their own, while satisfying low-level constraints that arise
from environment geometries and the limited capabilities of
the UAVs. High-level decisions can include selecting where
to go next, with whom to communicate, and what to inspect.
∗
Some of the work was done while this author as at United Technologies Research Center
Figure 1: The aircraft inspection scenario
These decisions need to take into account the uncertainty in
the UAV’s actions.
For instance, at the start of an aircraft’s inspection, one may
know that the left wing has a structural problem, but the location of the fault may not be known precisely. When a UAV
inspects the left wing, its sensors may succeed with probability 0.9, and so on. In order to solve this task autonomously,
the UAV needs to select which pose to fly to next, which trajectory to use in order to do so, and the order in which to carry
out inspections while making sure that it always has sufficient
battery to return to the docking station and that it does not
collide with any object in the environment. The feasibility
of a high-level strategy for inspection therefore depends on
the battery power required for each high-level operation such
as “move to left wing”; “inspect left wing”, etc., which in
turn depends on the low-level motion plan selected, which in
turn depends on the hangar’s geometric layout and the physical geometry of the UAV. Throughout this paper, we will use
the term “high-level” to represent a discrete MDP and “lowlevel” to refer to a motion planning problem.
The framework of Markov decision processes (MDPs) can
express discrete sequential decision making (SDM) problems. Numerous advances have been made in solving
MDPs [Russell et al., 2015]. However, the scalability of
these approaches relies upon a few key properties, including a
bounded branching factor (or the set of possible actions) and
the ability to express a problem accurately using discrete state
variables. Both of these properties fail to hold in problems
such as those described above. Recent work on deterministic,
integrated task and motion planning [Kaelbling and LozanoPérez, 2011; Erdem et al., 2011; Srivastava et al., 2014; Dantam et al., 2016] shows that hierarchical approaches are useful for such problems.
Computing task and motion policies for MDPs presents a
new set of challenges not encountered in computing task and
motion plans for deterministic scenarios. In particular, selecting an action for a state while ensuring a feasible refinement
requires knowing the history of actions used to reach that
state, since effects on properties that were abstracted away
(such as battery usage) cannot be modeled accurately at the
high level. A direct application of classical task and motion
planning techniques is further limited by the number of possible high-level action paths that can be taken during an execution. Indeed, the task and motion planning literature makes
it clear that computing a single high-level sequence of actions that is feasible with low-level constraints is a challenge;
the extension to MDPs expands the problem to computing
a feasible high-level sequence of actions for every possible
stochastic outcome of a high-level action.
In this paper, we investigate the problem of computing task
and motion policies and show that principles of abstraction
can be used to effectively model the problem, as well as to
solve it by dynamically refining the abstraction used. We address the problem of computational complexity by developing
an anytime algorithm that rapidly produces feasible policies
for a high likelihood of scenarios that may be encountered
during execution. Our methods can therefore be used to start
the execution before the complete problem is solved; computation could continue during execution. The continual policy computation reports the probability of encountering situations which have not been resolved yet. This can be used to
select the point at which execution is started in a manner appropriate to the application. In the worst case, if an unlikely
event is encountered before the ongoing policy computation
resolves it, execution could be brought to a safe state; in situations where this is not possible, one could wait for the entire
policy to be computed with motion plans. In this way our
approach offers a trade-off between pre-execution guarantees
and pre-computation time requirements.
The rest of this paper is organized as follows. Sec. 2 introduces the main concepts that we draw upon from prior
work. Sec. 3 presents our formalization of abstractions and
representations. This is followed by a description of our algorithms (Sec. 4). Sec. 5 presents an empirical evaluation of
our approach in a test scenario that we created using opensource 3D models of aircraft and various hangar components.
Sec. 6 discusses the relationship of the presented work and
contributions with prior work.
2
Background
A Markov decision process (MDP) hS, A, T, R, γi is defined
by a set of states S, a set of actions A, a transition function
T : S × A → µS that gives the probability distribution over
result states upon the application of an action on a state; a reward function R : S → R; and a discounting factor γ ≤ 1.
We will use T (s, a) as a function that maps a state to its probability. We will be particularly interested in MDPs with absorbing states and γ = 1, or, stochastic shortest path problems [Bertsekas and Tsitsiklis, 1991]. In this class of MDPs,
the reward function yields negative values (action costs) for
states except the absorbing states G. Absorbing states give
zero rewards; once the agent reaches an absorbing state, it
stays in it: ∀a ∈ A, g ∈ G, R(g) = 0; T (g, a)(g) = 1. We
will consider SSPs that have a known initial state s0 and a
finite time horizon, h, which represents an uppper bound on
the number of discrete decision making steps available to the
agent.
Solutions to MDPs are represented as policies. A policy
π : S × {1, . . . , h} → A maps a state and the timestep at
which it is encountered, to the action that the agent should execute while following π. Given an MDP, the optimal “policy”
of the agent is defined as one that maximizes the expected
Ph
long-term reward i=1 ri , where ri is the reward obtained
at timestep i following the function R. Our notion of policies includes non-stationary policies since the optimal policy
in a finite horizon MDP need not be stationary. In principle,
dynamic programming can be used to compute the optimal
policy in this setting just as in the infinite horizon setting:
V 0 (s)
i
V (s)
= R(s)
= R(s) + maxa
(1)
X
0
T (s, a)(s )V
i−1
0
(s ) (2)
s0
Here V i is the i-step-to-go value function. Since we are
given the initial state s0 , non-stationary policies can be expressed as finite state machines (FSMs). We will consider
policies that are represented as tree-structured FSMs, also
known as contingent plans. Several algorithms have been developed to solve SSPs. The LAO* algorithm [Hansen and Zilberstein, 2001] was developed to incorporate heuristics while
computing solution policies for SSPs. Kolobov et al. [2011]
developed general methods for solving SSPs in the presence
of dead-ends.
Specifying real-world sequential decision making problems as MDPs using explicitly enumerated state lists usually results in large, unweildy formulations that are difficult
to modify, maintain, or understand. Lifted, or parameterized representations for MDPs such as FOMDPs [Sanner and
Boutilier, 2009], RDDL [Sanner, 2010] and PPDDL [Younes
and Littman, 2004] have been developed for overcoming
these limitations. Such languages separate an MDP domain,
constituting parameterized actions, functions and predicate
vocabularies, from an MDP problem, which expresses the
specific objects of each type and a reward function. We refer to Helmert [2009] for a general introduction to these concepts. W.l.o.g, we consider the vocabulary to consist of predicates alone, since functions can be represented as special
predicates. A grounded predicate is a predicate whose parameters have been substituted by the objects in an MDP problem. For instance, Boolean valuations of the grounded predicate faultLocated(LeftWing) express whether the LeftWing’s
fault’s precise location was identified. In our framework,
states are defined as valuations of grounded predicates in a
given problem. Although this framework usually expresses
discrete properties, it can be extended naturally to model actions that have continuous action arguments and depend on
and affect geometric properties of the environment.
Example 1. Fig. 2 shows the specification for an inspect action in the aircraft inspection domain in a language similar to PPDDL (some syntactic elements have
Action: inspect(Structure s, Trajectory tr)
precond batterySufficient(tr)∧ inspects(tr, s)∧ collisionFree(tr)
effect faultLocated(s)
0.8
¬faultLocated(s)
0.2
decrease(batteryLevel(c(tr)))
Figure 2: Specification of a stochastic action model
been simplified for readability).
This action models
the act of inspecting a structure s while following the
path tr. We use batterySufficient(tr) as an abbreviation
for batteryRemaining−batteryRequired(tr). Intuitively, the
specification states that if this action is executed in a state
where the battery is sufficient and the selected trajectory satisfies constraints for being an inspection trajectory (the precondition is satisfied), it will result in locating the fault with
the probability 0.8. In any case, the battery’s charge will be
depleted by an amount depending on the trajectory used for
inspection c(tr). The inspects(c, tr) predicate is true if the
trajectory tr “covers” the given structure. Different interpretations for such predicates would result in different classes of
coverage patterns.
3
Formal Framework
Let X be a set of states and S a set of abstract states. We define a state abstraction as a surjective function α : X → S.
We focus on predicate abstractions, where the abstraction
function effectively projects the state space into a space without a specified set of predicates. Given a set of predicates
P that are retained by a predicate abstraction, the states of
the abstract state space are equivalence classes defined by the
equivalence relation s1 ∼ s2 iff s1 and s2 agree on the valuations of every predicate in P, grounded using the objects in
the problem.
For any s ∈ S, the concretization function γα (s) = {x ∈
X : α(x) = s} denotes the set of concrete states represented
by the abstract state s. For a set C ⊆ X, [C]α denotes the
smallest set of abstract states representing C. Generating the
complete concretization of an abstract state can be computationally intractable, especially in cases where the concrete
state space is continuous and the abstract state space is discrete. In such situations, the concretization operation can be
implemented as a generator that incrementally computes or
samples elements from an abstract state’s concretization.
Action abstraction functions can be defined similarly.
The main form of an action abstraction function is to drop action arguments, which leads to predicate abstractions to eliminate all predicates that used the dropped arguments in the action’s description. This process can also model non-recursive
temporal abstractions since a macro or a high-level action
with multiple implementations [Marthi et al., 2007] can be
modeled as an action whose arguments include the arguments
of its possible implementations as well as an auxiliary argument for selecting the implementation. The concretization
of an action abstraction function is the set of actions corresponding to different instantiations of the dropped action
arguments. Concretization functions for action abstraction
functions can also be implemented as generators.
Formally, the concretization of each high-level action
corresponds to a set of motion planning problems. We
will use the notation a(x1 7→ o1 ) to denote a grounded
action, whose x1 argument has been instantiated with
the element o1 defined by the underlying MDP problem (Sec. 2). Let a(x̄, ȳ) be a concrete action where x̄
(ȳ) are ordered, typed discrete (continuous) arguments.
The concretization of the instantiated abstract action
γ([a](x̄ 7→ ō)) is the set of actions {a(x̄ 7→ ō, ȳ 7→ ō0 ) :
ō0 is a tuple of elements with types and arity specified by y}.
Predicates in action preconditions specify the constraints
that these arguments need to satisfy. Common examples for
continuous arguments include robot poses and motion plans;
predicates about them may include collisionFree(tr), which
is true exactly when the trajectory tr has no collisions as well
as inspects (Eg. 1).
Both state and action abstractions affect the transition function of the MDP. The actual transition probabilities of an abstract MDP depend on the policy being used and are therefore difficult to estimate accurately [Bai et al., 2016; Li et
al., 2006; Singh et al., 1995]. In this paper, we will use an
optimistic estimate of the true transition probabilities when
expressing the abstract MDP. Such estimates are related to
upper bounds for reachability used in prior approaches for
reasoning in the presence of hierarchical abstractions (e.g.,
[Marthi et al., 2007; Ha and Haddawy, 1996]).
Example 2. Consider the action presented in Eg. 1 Such actions are difficult to plan with however, since the tr argument
is a high-dimensional real-valued vector. We can abstract
away this argument to construct the following abstraction:
Action: [inspect](Structure s)
precond batterySufficient
effect faultLocated(s)
0.8
¬faultLocated(s)
0.2
? {batteryLevel, batterySufficient}
Dropping the tr argument from each predicate that results
in abstract predicates of lower arities. The zero-arity batterySufficient becomes a Boolean state variable and batteryLevel
becomes a numeric variable. The symbol ? indicates that
this action affects the predicates batteryLevel and batterySufficient, but its effects on these predicates cannot be determined due to abstraction.
An optimistic representation of this abstract action would
state that it does not reduce batteryLevel and consequently,
does not make batterySufficient false.
This approach for abstraction is computationally better
than a high-level representation that discretizes the continuous variables, as it does not require the addition of constants
representing discrete pose or trajectory names to the vocabulary. This is desirable because the size of the state space
would be exponential in the number of such discretized values that are included.
4
Overall Algorithmic Framework
The ATM-MDP algorithm (Alg. 1) presents the main outer
loop of our approach for computing a task and motion policy.
It assumes the availability of an SSP solver that can generate
tree-structured policies (starting at a given initial state) for
Algorithm 1: Anytime Task and Motion MDP (ATM-MDP)
1
2
3
4
5
6
7
8
9
10
11
Data: domain D, problem P, motionPlanner M P , SSP Solver
SSP
Input: threshold t
Result: Task and motion policy for hD, Pi
policyTree ← SSP .getContingentPlan(P.f~0 , D, P);
currentState ← P.f~0 ; proportionRefined ← 0.0; replanBias ←
0.5;
partialTraj ← None;
leafQueue ← estimatePathCosts(policyTree,
partialTraj);
while resource limit not reached and leafQueue.size() 6= 0 and
proportionRefined < t do
pathToRefine ← ancestors(leafQueue.pop());
while resource limit not reached and pathToRefine.length()
6= 0 do
(success, partialPathTraj, failureNode, failureReason)
← refinePath(pathToRefine, partialTraj,
policyTree, M P );
if not success and failureReason 6= None then
policyTree ← SSP .replan(failureNode,
failureReason);
break;
Algorithm 2: Subroutine refinePath
1
2
3
4
5
6
7
8
9
10
11
12
13
14
else
12
13
14
15
16
if not success then
for node ∈ partialPathTraj do
partialTraj[node] ←
partialPathTraj[node]
leafQueue ← estimatePathCosts(policyTree,
partialTraj);
proportionRefined ←
computeProportionRefined(policyTree,
partialTraj)
solving an SSP, a motion planner for refinement of actions
within the policy, and a module that determines the reason for
infeasibility of a given motion planning problem. The overall
algorithm operates on root-to-leaf paths in the SSP solution.
The main computational problem is that the number of possible paths to refine grows exponentially with the time horizon. Waiting for a complete refinement would result in a lot
of wasted time as most paths may correspond to outcomes
that are unlikely to be encountered. Every path is associated with the probability p that an execution would follow
that path; and a cost c of refining that pat. Ideally, we would
like to compute an ordering of these paths so that at every
time instant, we compute as many of the most likely paths
as can be computed up to that time instant. Unfortunately,
achieving this would be infeasible as it would require solving
multiple knapsack problems. Instead, we order the paths by
the ratio p/c for refinement (lines 4-15).
Theorem 1. Let t be the time since the start of the algorithm
at which the refinement of any root-to-leaf path is completed.
If path costs are accurate and constant then the total probability of unrefined paths at time t is at most 1 − opt(t)/2,
where opt(t) is the best possible refinement (in terms of
the probability of outcomes covered) that could have been
15
16
Input: pathToRefine, partialTraj, policyTree, motionPlanner
Output: success: indicator of successful refinement;
partialPathTraj: refined path up to the first failure;
failureNode, failureReason: failure information
node ← head(pathToRefine); partialPathTraj ← None;
for node ∈ pathToRefine do
a ← policyTree[node];
if partialTraj = None then
pose1 ← InitialPose;
else
pose1 ←
extractPose(partialTraj[parent(node)]);
while resource limit not reached and
partialPathTraj[node] = None do
pose2 = targetPoseGen(a);
if GetMotionPlan(pose1 , pose2 ) succeeds
then
partialPathTraj[node] ← ComputePath;
break;
if partialPathTraj[node] = None then
if Bernoulli(replanBias).sample() then
return (False, partialPathTraj, node,
FailureReason);
else
partialPathTraj.remove(node.parent());
return (False, partialPathTraj, node.parent(),
None )
return (True, partialPathTraj, None, None);
achieved in time t.
The proof follows from the fact that the greedy algorithm
achieves a 2-approximation for the knapsack problem. In
practice, the true cost of refining a path cannot be determined
prior to refinement. We therefore estimate the cost as the
product of the parameter ranges covered by the generator of
each action in the path. This results in lower bounds on the
ratios p/c modulo constant factors, since a path could be refined before all the generator ranges are exhausted. In this
way it doesn’t over-estimate the relative value of refining a
path. As we show in the empirical section, the resulting algorithm yields the concave performance profiles desired of
anytime algorithms.
The while loop iterates over these paths while recomputing
the priority queue keys after each iteration. Within each iteration, the algorithm tries to compute a full motion planning
refinement of the path. First, the entire path (pathToRefine) is
extracted from the leaf (line 6). The refinePath subroutine attempts to find a motion planning refinement (concretization)
for pathToRefine. If it is unable to find a complete refinement for this path, it either (a) returns with a reason for failure
along with a partial trajectory going up to the deepest node in
the path for which it was able to compute a feasible motion
plan, or (b) backtracks to return a partial trajectory that will
result in a future refinePath call for a parent node of a node
for which a motion planning refinement couldn’t be found.
For partial trajectories under (a) (line 9), Alg. 1 calls an
Figure 3: Left: Backtracking from node B invalidates the subtree
rooted at A. In doing so, the work done in refining the node A’s left
child, in gray, is lost. Right: In some cases, replanning from node B
requires less work than re-refining the invalidated subtree.
SSP solver after adjusting its initial state and domain definitions to include the FailureReason. The policy computed by
the SSP solver is then merged with the existing policy and
the while loop continues. For partial trajectories along case
(b) (line 12), the path is added back to the queue with a partial, successful trajectory that results in backtracking.
If refinePath is successful in computing a full refinement,
the while loop continues with an updated priority queue. In
each iteration of the while loop, we compute the total probability of refined paths – this probability gives us the likelihood
of being able to successfully execute the policy in its current
state of refinement.
The refinePath subroutine (Alg. 2) attempts to compute a
motion plan for each action in a given path. More precisely,
it uses a generator to sample the possible concretizations for
each action and test their feasibility. A feasible solution to
any one of these motion planning problems is considered a
feasible refinement of that abstract action. refinePath starts
by selecting the first node in the path that needs to be refined
in line 1 (Alg. 1 may result in situations where a prefix of a
path has already been refined by a prior call to refinePath, due
to line 14 in that algorithm).
It then iterates over possible target poses for the selected
action (lines 8 through 11). If a feasible motion plan is found,
then the algorithm refines the next action in the path. If not,
it stochastically chooses to either re-invoke the SSP by returning a FailureReason, or to backtrack by invalidating the
current node’s path (line 15) by removing it from partialPathTraj and returning to follow lines 12-13 in Alg. 1.
Though a backtracking search through all possible motion
plans is required to guarantee the completeness of the algorithm, we find in practice that replanning with a new initial
state and replacing the subtree rooted at a failed node with a
new SSP solution is often more time efficient. This is because
backtracking to an ancestor of the failed node invalidates the
motion plans associated with all paths passing through that
ancestor, often causing a large amount of previously completed work to be thrown out. This situation is illustrated in
Figure 3. For this reason, we stochastically choose between
backtracking and replanning and settle for probabilistic completeness of the search algorithm.
Properties of the Algorithm Our algorithm solves the dual
problems of synthesizing a strategy as well as computing motion plans while ensuring that the computed strategy has a
feasible motion plan. It factors a hybrid planning problem
into a succession of discrete SSPs and motion planning problems. The algorithm can compute solutions even when most
discrete strategies have no feasible refinements. A few additional salient features of the algorithm are:
• The representational mechanisms for encoding SSPs do
not require discretization, thus providing scalability.
• The SSP model dynamically improves as the motion
planning problems reveal errors in the high-level model
in terms of FailureReasons.
• Prioritizing paths of relative value gives the algorithm a
desirable anytime performance profile. This is further
evaluated in the empirical section.
5
Empirical Evaluation
We implemented the algorithms presented in Sec. 4 using an
implementation of LAO* [Hansen and Zilberstein, 2001] as
the SSP solver. We used the OpenRAVE [Diankov, 2010] system for modeling and visualizing test environments and its
collision checkers and RRT [LaValle and Kuffner Jr, 2000]
implementation for motion planning. Since there has been
very little research on the task and motion planning problem
in stochastic settings, there are no standardized benchmarks.
We evaluated our algorithms by creating a hangar model in
OpenRAVE for the aircraft inspection problem (Fig. 1). UAV
actions in this domain include actions for moving to various
components of the aircraft, such as the left and right wings,
nacelles, fuselage, etc. Each such action could result in the
UAV reaching the specified component or a region around
the component. The inspection action for a component had
the stochastic effect of localizing a fault’s location. The environment included docking stations that the UAV could reach
and recharge on reserve battery power. Generators for concretizing all actions except the inspect action uniformly sampled poses in the target regions. Some of these poses naturally lead to shorter trajectories and therefore lower battery
usage, depending on the UAV’s current pose. However, we
used uniform-random samples to evaluate the performance of
the algorithm while avoiding domain-specific enhancements.
The generator for inspect(s) simulated an inspection pattern
by randomly sampling five waypoint poses in an envelope
around s and ordering them along the medial axis of the component. We used a linear function of the trajectories to keep
track of battery usage at the low level and to report insufficient
battery as the failureReason when infeasibility was detected.
This function was used to provide failure reasons to the highlevel when the battery level was found to be insufficient.
Fig. 4 shows the performance of our approach for producing execution strategies with motion planning refinements as
a function of the time for which the algorithm is allowed to
run. The red lines show the number of nodes in the high-level
policy that have been evaluated, refined, and potentially replaced with updated policies that permit low-level plans. The
blue lines show the probability with which the policy available at any time during the algorithm’s computation will be
able to handle all possible execution-time outcomes. The different plots show how these relations change as we increase
the level of uncertainty in the domain. The horizon is fixed at
ten high-level decision epochs (each of which can involve ar-
Figure 4: Performance of our anytime algorithm for solving MDPs using dynamic abstractions. The plots from left to right corresponds to
formulation of the problem with 5%, 10%, and 20% rates of failure of the abstract actions described in the text. The blue lines (red lines)
plot the probability mass of possible outcomes (proportion of nodes in the policy graph) that is covered by the partially computed policy as
computation time (x axis, in seconds) evolves.
bitrarily long movements) and the number of parts with faults
is fixed at two. The policy generated by LAO* is unrolled
into a tree prior to the start of refinement. The reported times
include the time taken for unrolling.
Our main result is that that our anytime algorithm balances complexity of computing task and motion policies with
time very well and produces desirable concave anytime peformance profiles. Fig. 4 shows that when noise in the agent’s
actuators and sensors is set at 5%, with 10% of computation
our algorithm computes an executable policy that misses only
the least likely 10% of the possible execution outcomes. This
policy is computed in less than 10 seconds. In the worst case,
with a 20% error rate in actuators and sensors (sensors used
in practice are much more reliable), we miss only about 20%
of the execution trajectories with 40% of the computation.
6
Other Related Work
There has been a renewed interest in integrated task and motion planning algorithms. Most research in this direction has
been focused on deterministic environments [Cambon et al.,
2009; Plaku and Hager, 2010; Hertle et al., 2012; Kaelbling
and Lozano-Pérez, 2011; Garrett et al., 2015; Dantam et al.,
2016]. Kaelbling and Lozano-Pérez [2013] consider a partially observable formulation of the problem. Their approach
utilizes regression modules on belief fluents to develop a
regression-based solution algorithm. Şucan and Kavraki
[2012] use an explicit multigraph to represent the plan or
policy for which motion planning refinements are desired.
Hadfield-Menell et al. [2015] address problems where the
high-level formulation is deterministic and the low-level is
determinized using most likely observations. In contrast, our
approach employs abstraction to bridge MDP solvers and motion planners to solve problems where the high-level model is
stochastic. In addition, the transitions in our MDP formulation depend on properties of the refined motion planning
trajectories (e.g., battery usage).
Principles of abstraction in MDPs have been well studied [Hostetler et al., 2014; Bai et al., 2016; Li et al., 2006;
Singh et al., 1995]. However, these directions of work assume that the full, unabstracted MDP can be efficiently expressed as a discrete MDP. Marecki et al. [2006] consider
continuous time MDPs with finite sets of states and actions.
In contrast, our focus is on MDPs with high-dimensional,
uncountable state and action spaces. Recent work on deep
reinforcement learning (e.g., [Hausknecht and Stone, 2016;
Mnih et al., 2015]) presents approaches for using deep neural
networks in conjunction with reinforcement learning to solve
MDPs with continuous state spaces. We believe that these
approaches can be used in a complementary fashion with our
proposed approach. They could be used to learn maneuvers
spanning shorter-time horizons, while our approach could be
used to efficiently abstract their representations and to use
them as actions or macros in longer-horizon tasks.
Efforts towards improved representation languages are orthogonal to our contributions [Fox and Long, 2002]. The fundamental computational complexity results indicating growth
in complexity with increasing sizes of state spaces, branching factors, and time horizons remain true regardless of the
solution approach taken. It is unlikely that a uniformly precise model, a simulator at the level of precision of individual atoms, or even circuit diagrams of every component used
by the agent will help it solve the kind of complex tasks on
which humans would appreciate assistance. On the other
hand, not using any model at all would result in dangerous
agents that would not be able to safely evaluate the possible
outcomes of their actions. Our results show that these divides
can be bridged using hierarchical modeling and solution approaches that simplify the representational requirements and
offer computational advantages that could make autonomous
robots feasible in the real world.
7
Conclusions
Our experiments showed that starting with an imprecise
model, refining it based on the information required to evaluate different courses of action is an efficient approach for
the synthesis of high-level policies that are consistent with
constraints that may be imposed by aspects of the model that
are more abstract or imprecise. While full models of realistic
problems can overwhelm SDM solvers due to the uncountable branching factor and long time horizons, our hierarchical approach allows us to use SDM solvers while addressing
more realistic problems involving physical agents.
Acknowledgements
This material is based upon work supported by the Defense
Advanced Research Projects Agency (DARPA) and Space
and Naval Warfare Systems Center Pacific (SSC Pacific) under Contract No. N66001-16-C-4050. Any opinions, findings
and conclusions or recommendations expressed in this material are those of the authors and do not necessarily reflect the
views of the DARPA or SSC Pacific.
References
Aijun Bai, Siddharth Srivastava, and Stuart J Russell. Markovian state and action abstractions for MDPs via hierarchical MCTS. In Proc. IJCAI, 2016.
Dimitri P Bertsekas and John N Tsitsiklis. An analysis of
stochastic shortest path problems. Mathematics of Operations Research, 16(3):580–595, 1991.
Stephane Cambon, Rachid Alami, and Fabien Gravot. A hybrid approach to intricate motion, manipulation and task
planning. IJRR, 28:104–126, 2009.
Neil T Dantam, Zachary K Kingston, Swarat Chaudhuri, and
Lydia E Kavraki. Incremental task and motion planning: A
constraint-based approach. In Proc. RSS, 2016.
Rosen Diankov. Automated Construction of Robotic Manipulation Programs. PhD thesis, Carnegie Mellon University,
2010.
Esra Erdem, Kadir Haspalamutgil, Can Palaz, Volkan
Patoglu, and Tansel Uras. Combining high-level causal
reasoning with low-level geometric reasoning and motion
planning for robotic manipulation. In Proc. ICRA, 2011.
Maria Fox and Derek Long. PDDL+: Modeling continuous
time dependent effects. In Proceedings of the 3rd International NASA Workshop on Planning and Scheduling for
Space, 2002.
Caelan Reed Garrett, Tomás Lozano-Pérez, and Leslie Pack
Kaelbling. FFrob: An efficient heuristic for task and motion planning. In Proc. WAFR. 2015.
Vu Ha and Peter Haddawy. Theoretical foundations for
abstraction-based probabilistic planning. In Proc. UAI,
1996.
Dylan Hadfield-Menell, Edward Groshev, Rohan Chitnis, and
Pieter Abbeel. Modular task and motion planning in belief
space. In Proc. IROS, 2015.
Eric A Hansen and Shlomo Zilberstein. LAO∗ : A heuristic
search algorithm that finds solutions with loops. Artificial
Intelligence, 129(1-2):35–62, 2001.
Matthew Hausknecht and Peter Stone. Deep reinforcement
learning in parameterized action space. In Proc. ICLR,
2016.
Malte Helmert. Concise finite-domain representations for
pddl planning tasks. Artificial Intelligence, 173(5):503 –
535, 2009.
Andreas Hertle, Christian Dornhege, Thomas Keller, and
Bernhard Nebel. Planning with semantic attachments: An
object-oriented view. In Proc. ECAI, 2012.
Jesse Hostetler, Alan Fern, and Tom Dietterich. State aggregation in monte carlo tree search. In Proc. AAAI, 2014.
Leslie Pack Kaelbling and Tomás Lozano-Pérez. Hierarchical
task and motion planning in the now. In Proc. ICRA, 2011.
Leslie Pack Kaelbling and Tomás Lozano-Pérez. Integrated
task and motion planning in belief space. The International
Journal of Robotics Research, 32(9-10):1194–1227, 2013.
A Kolobov, Mausam, DS Weld, and H Geffner. Heuristic
search for generalized stochastic shortest path mdps. In
Proceedings of the International Conference on Automated
Planning and Scheduling (ICAPS), 2011.
S.M. LaValle and J.J. Kuffner Jr. Rapidly-exploring random
trees: Progress and prospects. 2000.
Lihong Li, Thomas J Walsh, and Michael L Littman. Towards
a unified theory of state abstraction for mdps. In ISAIM,
2006.
Janusz Marecki, Zvi Topol, Milind Tambe, et al. A fast analytical algorithm for mdps with continuous state spaces. In
AAMAS-06 Proceedings of 8th Workshop on Game Theoretic and Decision Theoretic Agents, 2006.
Bhaskara Marthi, Stuart J Russell, and Jason Wolfe. Angelic
semantics for high-level actions. In Proc. ICAPS, 2007.
Volodymyr Mnih, Koray Kavukcuoglu, David Silver, Andrei A Rusu, Joel Veness, Marc G Bellemare, Alex Graves,
Martin Riedmiller, Andreas K Fidjeland, Georg Ostrovski,
et al. Human-level control through deep reinforcement
learning. Nature, 518(7540):529–533, 2015.
E. Plaku and G. D. Hager. Sampling-based motion and symbolic action planning with geometric and differential constraints. In Proc. ICRA, 2010.
Stuart Russell, Daniel Dewey, and Max Tegmark. Research
priorities for robust and beneficial artificial intelligence. AI
Magazine, 36(4):105–114, 2015.
Scott Sanner and Craig Boutilier. Practical solution techniques for first-order MDPs. Artificial Intelligence, 173(56):748–788, 2009.
Scott Sanner. Relational dynamic influence diagram language
(rddl): Language description. http://users.cecs.
anu.edu.au/˜ssanner/IPPC_2011/RDDL.pdf,
2010.
Satinder P Singh, Tommi Jaakkola, and Michael I Jordan. Reinforcement learning with soft state aggregation. In Proc.
NIPS, pages 361–368, 1995.
Siddharth Srivastava, Eugene Fang, Lorenzo Riano, Rohan
Chitnis, Stuart Russell, and Pieter Abbeel. A modular
approach to task and motion planning with an extensible
planner-independent interface layer. In Proc. ICRA, 2014.
Ioan A Şucan and Lydia E Kavraki. Accounting for uncertainty in simultaneous task and motion planning using task
motion multigraphs. In Proc. ICRA, 2012.
Håkan LS Younes and Michael L Littman. PPDDL 1.0: An
extension to pddl for expressing planning domains with
probabilistic effects. Technical Report CMU-CS-04-162,
2004.
| 2cs.AI
|
On the Boundary between Decidability and
Undecidability of Asynchronous Session Subtyping
arXiv:1703.00659v3 [cs.PL] 12 Feb 2018
Mario Bravetti
University of Bologna, Department of Computer Science and Engineering / INRIA FOCUS
Marco Carbone
Department of Computer Science, IT University of Copenhagen
Gianluigi Zavattaro
University of Bologna, Department of Computer Science and Engineering / INRIA FOCUS
Abstract
Session types are behavioural types for guaranteeing that concurrent programs
are free from basic communication errors. Recent work has shown that asynchronous session subtyping is undecidable. However, since session types have
become popular in mainstream programming languages in which asynchronous
communication is the norm rather than the exception, it is crucial to detect
significant decidable subtyping relations. Previous work considered extremely
restrictive fragments in which limitations were imposed to the size of communication buffer (at most 1) or to the possibility to express multiple choices
(disallowing them completely in one of the compared types). In this work, for
the first time, we show decidability of a fragment that does not impose any
limitation on communication buffers and allows both the compared types to
include multiple choices for either input or output, thus yielding a fragment
which is more significant from an applicability viewpoint. In general, we study
the boundary between decidability and undecidability by considering several
fragments of subtyping. Notably, we show that subtyping remains undecidable
even if restricted to not using output covariance and input contravariance.
Keywords: Session Types, Asynchronous Subtyping, Undecidability
Preprint submitted to Journal of LATEX Templates
February 13, 2018
1. Introduction
Session types [1, 2] are types for controlling the communication behaviour
of processes over channels. In a very simple but effective way, they express the
pattern of sends and receives that a process must perform. Since they can guarantee freedom from some basic programming errors, session types are becoming
popular with many main stream language implementations, e.g., Haskell [3],
Go [4] or Rust [5].
As an example, consider a client that invokes service operations by following
the protocol expressed by the session type
⊕{op1 : &{resp1 : end}, op2 : &{resp2 : end}}
indicating that the client decides whether to call operation op1 or op2 and then
waits for receiving the corresponding response (resp1 or resp2, respectively).
For the sake of simplicity we consider session types where (the type of) communicated data is abstracted away. The symmetric behaviour of the service is
represented by the complementary (so-called dual) session type
&{op1 : ⊕{resp1 : end}, op2 : ⊕{resp2 : end}}
indicating that the server receives the call to operation op1 or op2 and then
sends the corresponding response (resp1 or resp2, respectively).
We call output selection the construct ⊕{l1 : T1 , . . . , ln : Tn }. It is used to
denote a point of choice in the communication protocol: each choice has a label
li and a continuation Ti . In communication protocols, when there is a point of
choice, there is usually a peer that internally takes the decision and the other
involved peers receive communication of the selected branch. Output selection
is used to describe the behaviour of the peer that takes the decision: indeed, in
our example it is the client that decides which operation to call. Symmetrically,
we call input branching the construct &{l1 : T1 , . . . , ln : Tn }. It is used to
describe the behaviour of a peer that receives communication of the selection
done by some other peers. In the example, indeed, the service receives from the
2
client the decision about the selected operation.1
When composing systems whose interaction protocols have been specified
with session types, it is significant to consider variants of their specifications
that still preserve safety properties. In the above example, the client can be
safely replaced by another one with session type
⊕{op1 : &{resp1 : end}}
indicating that it can call only one specific service operation. But also the
service can be safely replaced by another one accepting an additional operation:
&{op1 : ⊕{resp1 : end}, op2 : ⊕{resp2 : end}, op3 : ⊕{resp3 : end}}
Subtyping relations have been formally defined for session types, e.g., by Gay
and Hole [7] and Chen et al. [8], in order to precisely capture this safe replacement notion. For instance, a subtyping relation like that of Gay and Hole [7]
(denoted by ≤s ), where processes are assumed to simply communicate via synchronous channels2 , would imply, for the client, that:
⊕{op1 : &{resp1 : end}} ≤s ⊕ {op1 : &{resp1 : end}, op2 : &{resp2 : end}}
according to the so-called output covariant property, while, for the server
&{op1 : ⊕{resp1 : end}, op2 : ⊕{resp2 : end}, op3 : ⊕{resp3 : end}}
≤s &{op1 : ⊕{resp1 : end}, op2 : ⊕{resp2 : end}}
according to the so-called input contravariant property.
When processes communicate via asynchronous channels, a more generous
notion of subtyping ≤ like that of Chen et al. [8] can be considered. E.g., a process using an asynchronous channel to call a service operation, that receives the
1 In
session type terminology [1, 6], the output selection/input branching constructs are
usually simply called selection/branching; we call them output selection/input branching
because we consider a simplified syntax for session types in which there is no specific separate
construct for sending one output/receiving one input. Anyway, such output/input types can
be seen as an output selection/input branching with only one choice.
2 Here, we focus on the so-called process-oriented subtyping, as opposed to channel-based
subtyping [9].
3
corresponding response and then sends a huge amount of data (requiring heavy
computation), could be safely replaced by a more efficient one that computes
and sends all the data immediately without waiting for the response:
⊕{op : ⊕{huge data : &{resp : end}}} ≤ ⊕ {op : &{resp : ⊕{huge data : end}}}
Intuitively, this form of asynchronous subtyping reflects the possibility to anticipate the output of the huge data w.r.t. to the input of the response because
such data are stored in a buffer waiting for their reader to consume them.
1.1. Previous Results
Recently, Bravetti et al. [10] and Lange and Yoshida [11] have independently
shown that asynchronous subtyping (the subtyping relation with output anticipation) is undecidable. In particular, in Bravetti et al., this is done by showing
undecidability of the much simpler single-choice relation << that is defined as
a restriction of asynchronous subtyping ≤ where related T << S types are such
that: all output selections in T have a single choice (output selections are covariant, thus S is allowed have output selections with multiple choices) and all
input branchings in S have a single choice (input branchings are contravariant,
thus T is allowed to have input branchings with multiple choices). Moreover,
those papers prove decidability for very small fragments of the asynchronous
subtyping relation: the most significant one basically requires one of the two
compared types to be such that all its input branchings and output selections
have a single choice. In particular, in Bravetti et al. this is done by showing decidability of the two relations <<sin (single-choice input <<) and <<sout
(single-choice output <<), both defined as further restrictions of << where for
related T <<sin S (T <<sout S, resp.) types we additionally require that: all input
branchings (output selections, resp.) in T and S have a single choice. Other
decidable fragments, considered by Lange and Yoshida, pose limitations on the
communication behaviour that causes communication buffers to store at most
one message, or they are used in half duplex modality (messages can be sent in
one direction, only if the buffer for the opposite direction has been emptied). Al4
branching/selection structure
buffer
2≤=
< ≤[ DC
≤tin
O h
6 ≤tout
O
≤tin,tout
J T
≤bound
O
undecidable
decidable
≤sinV
≤H sout
..
.
O
≤O 2
≤sin,tout
b
≤tin,sout
<
≤sin,sout
≤O 1
≤0 = ≤s
≤s synchronous subtyping [7]
≤ asynchronous subtyping [8]
/ set inclusion
Figure 1: Lattice of the asynchronous subtyping relations considered in this paper.
though asynchronous subtyping is undecidable, it is important to reason about
more significant cases for which such a relation is decidable. This because session
types have become popular in mainstream programming languages, and, in such
cases, asynchronous communications are the norm rather than the exception.
1.2. Contributed Results
The aim of this paper is to detect significant decidable fragments of asynchronous session subtyping and to establish a more precise boundary between
decidability and undecidability. In particular, concerning decidability, as discussed above, the few decidable fragments of asynchronous subtyping known so
far are extremely restrictive: our relations <<sin , <<sout [10] and the decidable
5
relations considered by Lange and Yoshida [11]. Here, for the first time, we show
decidability of a fragment that does not impose any limitation on communication buffers and allows both the subtype and the supertype to include multiple
choices (either for input branchings or for output selections), thus opening the
possibility for some practical applicability in restricted specific scenarios (e.g.
session types for clients/services in web-service systems, see below). More precisely, while <<sin (<<sout , resp.), being it defined as a restriction of <<, admits
multiple choices only for output selections in the supertype (input branchings
in the subtype, resp.), here we consider and show decidability for a much larger
relation, we denote by ≤sin (≤sout , resp.). Such a relation is defined as the restriction of the whole ≤ relation (instead of the << relation), where, in related
types, all input branchings (output selections, resp.) must have a single choice.
Therefore, differently from <<sin (<<sout , resp.), in ≤sin (≤sout , resp.) both the
subtype and the supertype can include multiple choices for output selections (for
input branchings, resp.). The combination of non restricted buffers and presence of multiple choices on both related types requires a totally new approach
for guaranteeing the termination of the subtyping algorithm (both for the termination condition itself and for the related decidability proof). For instance,
if multiple choices are admitted for input branchings, the termination condition
has to deal with complex recurrent patterns to be checked on the leafs of trees
representing input branchings (with multiple choices), instead of detecting simple repetitions on strings representing sequences of single-choice inputs (as in
previous work [10, 11]).
Concerning undecidability, all previous results [10, 11] exploit the capability
of asynchronous subtyping of matching input branchings/output selections by
means of covariance/contravariance. We here show that asynchronous subtyping
remains undecidable even if we restrict it by disallowing this feature. As asynchronous subtyping is based on the combination of output covariance/input contravariance and output anticipation deriving from asynchronous communication,
our result means that (provided that the syntax of types is not constrained) the
source of undecidability is to be precisely localized into the output anticipation
6
capability. The undecidability proof has a structure similar to that of Bravetti et
al. [10], where the termination problem for queue machines, a well-known Turingequivalent formalism, is encoded into asynchronous session subtyping. However,
differently from Bravetti et al. [10], not having covariance/contravariance makes
it impossible to encode queue machines deterministically. As we will see, the
need to cope with nondeterminism makes it necessary to restrict the class of
encodable queue machines to a new ad-hoc fragment that we introduce in this
paper and we prove being Turing-equivalent: single consuming queue machines.
Moreover a much more complex encoding, that uses nondeterminism, must be
adopted.
In general, the contribution of this work is to analyze restrictions of asynchronous subtyping and classifying them into decidable and undecidable fragments. More precisely, as detailed in the following, we focus on two kinds of
restrictions of asynchronous subtyping: limitations to the branching/selection
structure and to the communication buffer, giving rise to the numerous relations
shown in the lefthand part and righthand part of Figure 1, respectively (see Section 2 for formal definitions of all the relations). The relations are depicted
as a lattice according to their inclusion as sets of pairs. Notice that decidability/undecidability is not logically related to set inclusion (e.g. the emptyset and
the set of all pairs are both decidable and are the bottom and the top of the
lattice).
Concerning asynchronous subtyping ≤ itself, we consider the orphan message
free notion of subtyping introduced in Chen et al. [8]: it is commonly recognized
that a convenient notion of asynchronous subtyping should prevent existence of
messages that remain “orphan”, i.e. that are never consumed from the communication buffer. Operationally, this implies that inputs in the supertype cannot
be indefinitely delayed by output anticipation: eventually such inputs must be
performed by the subtype so to correspondingly consume messages from the
buffer. As a side result, in this paper we introduce a new, elegant, way of defining orphan message free asynchronous subtyping. The new definition is based
on just adding a constraint about closure under duality to the standard (non
7
orphan message free) coinductive definition of asynchronous subtyping [12]. We
thus use such a novel approach based on dual closeness to give a concise definition of asynchronous subtyping: we call ≤DC the obtained relation. We then
show ≤DC (also included in Figure 1) to be equal to orphan message free subtyping ≤ of Chen et al. [8]. We also show that the most significant decidable and
undecidable fragments of ≤, i.e. the ≤sin ∪ ≤sout relation and subtyping without
covariance/contravariance, are dual closed subtyping relations according to our
new definition.
Limitations to the branching/selection structure. We limit the asynchronous subtyping ≤ capability of managing input branchings/output selections, giving rise to the subtyping relations shown in the lefthand part of Figure
1, as follows:
• requiring that in both the subtype and the supertype output selections
(input branchings, resp.) have a single choice: in this case the sout (sin,
resp.) subscript is added to ≤;
• requiring that each output selection (input branching, resp.) performed
by the subtype is matched by an output selection (input branching, resp.)
performed by the supertype with the exactly the same total set of labels,
i.e. output covariance (input contravariance, resp.) is not admitted: in
this case the tout (tin, resp.) subscript is added to ≤.
We now summarize our decidability/undecidability results for these relations.
• Decidability of asynchronous subtyping for single-in types (≤sin ) or singleout types (≤sout ). We consider the class of single-in (single-out, resp.) session types, i.e. types where all output selections (input branchings, resp.)
have a single choice. We present and prove correct an algorithm for deciding whether two single-out (resp. single-in, by exploiting the closure
under duality property) types are in the subtyping relation. From a modeling viewpoint, assuming binary sessions to happen between a single-in
8
and a single-out party, this entails that internal decisions are taken by
the single-in party, while the single-out one passively accepts them. This
kind of behaviour can occur in, e.g., web-service systems where a client
internally chooses a request-response operation [13] and then waits for a
corresponding (non branching) input, while the server accepts invocations
on several operations but then it reacts by answering on a related response channel (independently of the actual returned data/error); see the
examples at the beginning of this Introduction section. Our algorithms
for subtyping of single-out/single-in types could thus be used in typing
systems for server/client code. With minor variants to the machinery introduced to show decidability of ≤sin and ≤sout , we also show that ≤sin,tout ,
≤tin,sout and ≤sin,sout are decidable.
• Undecidability of Asynchronous Subtyping without Output Covariance and
Input Contravariance (≤tin,tout ).
As discussed above, subtyping for
session types makes use of output covariance and input contravariance:
an output ⊕{li : Ti }i∈I is a subtype of an output with more labels
⊕{lj : Tj }j∈J , for I ⊂ J; and an input &{li : Ti }i∈I is a subtype of
an input with less labels &{lj : Tj }j∈J , for J ⊂ I. Existing results on the
undecidability of asynchronous subtyping exploit its capability of relating
types with a different number of branches. We consider a restricted form
of subtyping, the ≤tin,tout relation, which disallows this feature, i.e. which
does not use output covariance and input contravariance. We show that,
also with such a restriction, subtyping remains undecidable by encoding
the termination problem for single consuming queue machines, a Turingequivalent formalism that (as already explained) we introduce on purpose,
into ≤tin,tout . The same encoding we use for ≤tin,tout shows also undecidability of ≤tin , ≤tout and ≤ (thus also providing an alternative proof for
the undecidability of ≤ with respect to those by Bravetti et al. [10] and
Lange and Yoshida [11]).
Limitations to the communication buffer. We limit the communication
9
buffer capability, giving rise to the subtyping relations shown in the righthand
part of Figure 1, by restricting the capability of ≤ to anticipate outputs: this is
equivalent to putting an upper limit to communication buffers between two parties, a common fact in practice. In this context our decidability/undecidability
results are the following ones.
• Decidability of k-bounded Subtyping (≤k ), with k ≥ 0. In k-bounded asynchronous subtyping we restrict the capability of ≤ to anticipate outputs:
they can only be anticipated w.r.t. a number of inputs that is less or equal
to k. We give and prove correct an algorithm for deciding whether any
two session types are in a k-bounded subtyping relation. Notice that, in
the case k = 0 we obtain synchronous subtyping ≤s [7]. Moreover, if we
consider k = 1 we have a notion of subtyping along the lines of that, we
already mentioned, obtained by Lange and Yoshida [11] imposing restrictions on the communication behaviour.
• Undecidability of Bounded Asynchronous Subtyping (≤bound ). We say that
a pair of session types is in bounded asynchronous subtyping relation if
there exists a k such that such pair is in k-bounded subtyping relation.
Bounded asynchronous subtyping relates types that do not unboundedly
put messages in a buffer. For instance, the types
µt. ⊕ {huge data : ⊕{huge data : &{ack : t}}}
µt.&{ack : ⊕{huge data : t}}
are related by asynchronous subtyping but not by bounded asynchronous
subtyping: the augmented data production frequency of the subtype requires to store an unbounded amount of huge data. Since in practice
buffers are bounded, this could have been an acceptable candidate notion
for replacing standard asynchronous subtyping, however we prove that it
is undecidable as well. We do this by showing undecidability of a property
for queue machines: bounded non termination.
10
Outline. In Section 2 we present session types and definition of asynchronous
subtyping ≤, the novel dual closed reformulation ≤DC , the fragments of ≤ shown
in Figure 1 and a discussion about their properties. Section 3 presents all decidability results, notably for k-bounded subtyping (≤k ) and for subtyping over
single-in types (≤sin ) and over single-out types (≤sout ). Section 4 presents all
undecidability results, notably for bounded subtyping (≤bound ) and for subtyping without output covariance and input contravariance (≤tin,tout ). Section 5
discusses related work and Section 6 presents concluding remarks. Detailed
proofs of theorems, lemmas and propositions can be found in the Appendix.
We chose to put proof technicalities, that often include additional definitions
and intermediate results, in the Appendix so not to disrupt the paper prose.
2. Session Types and Asynchronous Subtyping
We begin by formally introducing the various ingredients needed for our
technical development.
We start with the formal syntax of binary session types. Similarly to Chen
et al. [8] we do not use a dedicated construct for sending an output/receiving an
input, we instead represent outputs and inputs directly inside choices. More
precisely, we consider output selection ⊕{li : Ti }i∈I , expressing an internal
choice among outputs, and input branching &{li : Ti }i∈I , expressing an external choice among inputs. Each possible choice is labeled by a label li , taken
from a global set of labels L, followed by a session continuation Ti . Labels in a
branching/selection are assumed to be pairwise distinct.
Definition 2.1 (Session Types). Given a set of labels L, ranged over by l,
the syntax of binary session types is given by the following grammar:
T
::=
⊕{li : Ti }i∈I
|
&{li : Ti }i∈I
|
µt.T
|
t
|
end
A session type is single-out if, for all of its subterms ⊕{li : Ti }i∈I , |I| = 1.
Similarly, a session type is single-in if, for all of its subterms &{li : Ti }i∈I ,
|I| = 1.
11
In the sequel, we leave implicit the index set i ∈ I in input branchings and
output selections when it is already clear from the denotation of the types.
Note also that we abstract from the type of the message that could be sent over
the channel, since this is orthogonal to our theory. Types µt.T and t denote
standard tail recursion for recursive types. We assume recursion to be guarded:
in µt.T , the recursion variable t occurs within the scope of an output or an
input type. In the following, we will consider closed terms only, i.e., types with
all recursion variables t occurring under the scope of a corresponding definition
µt.T . Type end denotes the type of a channel that can no longer be used.
In our development, it is crucial to count the number of times we need to
unfold a recursion µt.T . This is formalised by the following function:
Definition 2.2 (n-unfolding).
unfold0 (T ) = T
unfold1 (⊕{li : Ti }i∈I ) = ⊕{li : unfold1 (Ti )}i∈I
unfold1 (µt.T ) = T {µt.T /t}
unfold1 (&{li : Ti }i∈I ) = &{li : unfold1 (Ti )}i∈I
unfold1 (end) = end
unfoldn (T ) = unfold1 (unfoldn−1 (T ))
The definition of asynchronous subtyping uses the notion of input context,
a type context consisting of a sequence of inputs preceding holes where types
can be placed:
Definition 2.3 (Input Context). An input context A is a session type with
multiple holes defined by the syntax:
A ::=
[ ]n
|
&{li : Ai }i∈I .
An input context A is well-formed whenever all its holes [ ] , with n ∈ N+ , are
n
consistently enumerated, i.e. there exists m ≥ 1 such that A includes one and
only one [ ]n for each n ≤ m. Given a well-formed input context A with holes
indexed over {1, . . . , m} and types T1 ,. . . , Tm , we use A[Tk ]k∈{1,...,m} to denote
the type obtained by filling each hole k in A with the corresponding term Tk .
From now on, whenever using input contexts we will assume them to be
well-formed, unless otherwise specified.
For example, consider the input context
A = &{l1 : []1 , l2 : []2 }
12
we have:
A[⊕{l : Ti }]i∈{1,2} = & l1 : ⊕{l : T1 }, l2 : ⊕{l : T2 }
We start by considering the standard notion of asynchronous subtyping ≤
given by Chen et al. [8]. We choose it because of its orphan message free property that is commonly recognized to be convenient: only subtypes are allowed
that do not cause incoming messages to remain “orphan” (because they are
never consumed from the communication buffer). In the definition of asynchronous subtyping given by Chen et al., orphan message freedom causes a
specific dedicated constraint to be included (which is, e.g., instead not present
in the asynchronous subtyping definition by Mostrous and Yoshida [12]). We
now formally present the asynchronous subtyping relation ≤, rephrased w.r.t.
that of Chen et al. [8] in a technical format that is convenient for showing our
results, which follows a coinductive simulation-like definition.
Definition 2.4 (Asynchronous Subtyping, ≤). R is an asynchronous subtyping relation if (T, S) ∈ R implies that:
1. if T = end then ∃n ≥ 0 such that unfoldn (S) = end;
2. if T = ⊕{li : Ti }i∈I then ∃n ≥ 0, A such that
• unfoldn (S) = A[⊕{lj : Sk j }j∈Jk ]k∈{1,...,m} ,
• ∀k ∈ {1, . . . , m}. I ⊆ Jk ,
• ∀i ∈ I, (Ti , A[Ski ]k∈{1,...,m} ) ∈ R and
• if A 6= [ ]1 then ∀i ∈ I. & ∈ Ti (no orphan message constraint);
3. if T = &{li : Ti }i∈I then ∃n ≥ 0 such that unfoldn (S) = &{lj : Sj }j∈J ,
J ⊆ I and ∀j ∈ J.(Tj , Sj ) ∈ R;
4. if T = µt.T ′ then (T ′ {T /t}, S) ∈ R.
where with “& ∈ Ti ” we mean that Ti contains at least an input branching.
T is an asynchronous subtype of S, written T ≤S, if there is an asynchronous
subtyping relation R such that (T, S) ∈ R.
13
Intuitively, two types T and S are related by ≤, whenever S is able to
simulate T , but with a few twists: type S is allowed to anticipate outputs nested
in its syntax tree (asynchrony); and, output and input types enjoy covariance
and contravariance, respectively. Moreover, the above definition includes the
no orphan message constraint [8], namely: we allow the supertype inputs to be
delayed only if also the subtype contains some input.
A synchronous subtyping relation ≤s like that of Gay and Hole [7] is obtained
by requiring that, in item 2. of the above Definition 2.4, it always holds A = [ ]1 .
Example 2.5. Consider T = µt. ⊕ {l : &{l1 : t, l2 : t}} and S = µt.&{l1 :
&{l2 : ⊕{l : t}}}. We have T ≤ S because the following is an asynchronous
subtyping relation:
{
(T, S) , (⊕{l : &{l1 : T, l2 : T }}, S) , (&{l1 : T, l2 : T }, &{l1 : &{l2 : S}}) ,
(T, &{l2 : S}) , (⊕{l : &{l1 : T, l2 : T }}, &{l2 : S}) ,
(&{l1 : T, l2 : T }, &{l2 : &{l1 : &{l2 : S}}}),
(T, &{l1 : &{l2 : S}}) , (⊕{l : &{l1 : T, l2 : T }}, &{l1 : &{l2 : S}}) ,
(&{l1 : T, l2 : T }, &{l1 : &{l2 : &{l1 : &{l2 : S}}}}) ,
(T, &{l2 : &{l1 : &{l2 : S}}}) , . . .
}
Note that the relation contains infinitely many pairs that differ in the sequence of
inputs, alternatively labeled with l1 and l2 , that are accumulated at the beginning
of the r.h.s. type.
We now introduce an alternative way of defining orphan message free asynchronous subtyping, which is more elegant/concise: it obtains the orphan message freedom property by requiring closure under duality of the type relation
being defined instead of making use of an explicit orphan message free constraint
as in Definition 2.4.
For session types, we define the usual notion of duality: given a session type
T , its dual T is defined as: ⊕{li : Ti }i∈I = &{li : T i }i∈I , &{li : Ti }i∈I = ⊕{li :
T i }i∈I , end = end, t = t, and µt.T = µt.T . In the sequel, we say that a
relation R on session types is dual closed if (S, T ) ∈ R implies (T , S) ∈ R.
14
Definition 2.6 (Asynchronous Dual Closed Subtyping, ≤DC ). R is an
asynchronous dual closed subtyping relation whenever it is dual closed and
(T, S) ∈ R implies 1., 3., and 4. of Definition 2.4, plus a modified version
of 2. where the last constraint (the no orphan message constraint) is removed.
T is an asynchronous dual closed subtype of S, written T ≤DC S, if there is
an asynchronous dual closed subtyping relation R such that (T, S) ∈ R.
We observe that our definition is formally different from the ones found in
literature. In particular, with respect to that by Mostrous and Yoshida [12],
it additionally requires the subtyping relation to be dual closed. Below, we
state that the dual closeness requirement is equivalent to imposing the orphan
message free constraint, i.e. the last item of condition 2 in Definition 2.4 (both
guarantee orphan-message freedom):
Theorem 2.7. Given two session types T and S, we have T ≤ S if and only if
T ≤DC S.
2.1. Subtyping Relation Restrictions
As already discussed in the Introduction, we focus on two kinds of restrictions
of asynchronous subtyping: limitations to the branching/selection structure and
to the communication buffer, giving rise to the numerous relations shown in the
lefthand part and righthand part of Figure 1, respectively.
We now define fragments of ≤ obtained by posing limitations to the branching/selection structure.
Definition 2.8. Restrictions of the asynchronous subtyping relation are denoted by adding subscripts to the ≤ notation, with the following meaning:
• whenever we add subscript sout (sin, resp.) we additionally require in
Definition 2.4 both T and S to be single-out (single-in, resp.),
• whenever we add subscript tout (tin, resp.) we additionally require in
Definition 2.4 I = Jk in point 2. (I = J in point 3., resp.).
15
The latter means that each output selection (input branching, resp.) performed
by the subtype is matched by an output selection (input branching, resp.) performed by the supertype with the exactly the same total set of labels, i.e. output
covariance (input contravariance, resp.) is not admitted.
Notice that, while it holds that ≤ = ≤DC , not all fragments of ≤ are asynchronous dual closed subtyping relations. For instance this does not hold for
≤sin , ≤sout , ≤tin and ≤tout , which perform a limitation, but not its “dual” one.
It holds, instead, for the following two relations that we will show to be in the
boundary between decidability and undecidability.
Proposition 2.9. The ≤tin,tout relation and the ≤sin ∪≤sout relation are asynchronous dual closed subtyping relations.
Dual closeness of the ≤sin ∪ ≤sout relation is a direct consequence of the
fact that T ≤sin S if and only if S ≤sout T , which obviously derives from dual
closeness of ≤ and from the dual of a single-in type being a single-out type and
vice-versa. This fact, together with the following proposition, will be used to
infer decidability of ≤sin and ≤sin,tout relations from that of ≤sout and ≤tin,sout ,
respectively.
Proposition 2.10. The ≤sin,tout and ≤tin,sout relations are such that: T ≤sin,tout S
if and only if S ≤tin,sout T .
We now consider variants of ≤ obtained by posing limitations to the communication buffer.
We can define a variant decidable relation by putting an upper-bound to the
messages that can be buffered. Technically speaking, when an output in the
r.h.s. is anticipated during the subtyping simulation, we impose a bound to the
number of inputs that are in front of such output.
We say that an input context A is k-bounded if the maximal number of nested
inputs in A is less or equal to k.
16
Definition 2.11 (k-bounded Asynchronous Subtyping, with k ≥ 0). The
k-bounded asynchronous subtyping ≤k is defined as in Definition 2.4, with the
only difference that the input context A in item 2. is required to be k-bounded.
Notice that the case k = 0 yields synchronous subtyping: since A = [ ]1 is
the only 0-bounded input context, we obviously have ≤0 = ≤s .
Lange and Yoshida [11] show the decidability of asynchronous subtyping for
a subclass of session types, called alternating, that in our setting corresponds
to impose that every output in a subtype is immediately followed by an input,
while every input in a supertype is followed by an output. For instance, this
property is satisfied by the following pair of types:
T = µt. ⊕ {l2 : &{l1 : t}}
S = µt.&{l1 : ⊕{l2 : t}}
It is not difficult to see that T ≤1 S. The key point of the proof of decidability of
asynchronous subtyping for alternating session types by Lange and Yoshida is
the observation that if T and S are alternating, then T ≤S if and only if T ≤1 S.
As we explained in the Introduction, we also consider the more generic notion
of bounded asynchronous subtyping. This relation is in our opinion of interest
because it reflects real cases in which it is possible to assume bounded buffers,
without an a priory knowledge of the actual bound.
Definition 2.12 (Bounded Asynchronous Subtyping, ≤bound ). We say that
T is a bounded asynchronous subtype of S, written T ≤boundS, if there exists k
such that T ≤k S.
3. Decidability Results
We now present decidability results for k-bounded asynchronous subtyping
and asynchronous subtyping for single-out/single-in session types.
3.1. A Subtyping Procedure
We start by giving a procedure (an algorithm that does not necessarily
terminate) for the general subtyping relation, which is known to be undecidable [10, 11]. Such a procedure is inspired by the one proposed by Mostrous et
17
al. [14] for asynchronous subtyping in multiparty session types. In order to do
so, we introduce two functions on the syntax of types. The function outDepth
calculates how many unfolding are necessary for bringing an output outside a
recursion. If that is not possible, the function is undefined (denoted by ⊥).
Definition 3.1 (outDepth). The partial function outDepth(T ) is inductively
defined as follows:
outDepth(⊕{li : Ti }i∈I ) = 0
outDepth(&{li : Ti }i∈I ) = max{outDepth(Ti ) | i ∈ I}
outDepth(µt.T ) = 1+outDepth(T {end/t})
outDepth(end) =⊥
where max{outDepth(Ti ) | i ∈ I} =⊥, if outDepth(Ti ) =⊥ for some i ∈ I;
similarly, 1+ ⊥=⊥.
As an example of application of outDepth consider, for any T1 and T2 ,
outDepth(⊕{l1 : T1 , l2 : T2 }) = 0. On the other hand, consider the type
o
n
Tex = & l1 : µt. ⊕ l2 : T1 , l3 : µt.& l4 : µt′ . ⊕ {l5 : T2 } : clearly,
outDepth Tex = 2. We then define outUnf(), a variant of the unfolding function given in Definition 2.2, which unfolds only where it is necessary, in order
to reach an output:
Definition 3.2 (outUnf ). The output unfolding outUnf(T ) is a partial function defined whenever outDepth(T ) is defined. Given outDepth(T ) = n, outUnf(T )
is computed using the same inductive rules of unfoldn (T ), excluding the rule for
⊕{li : Ti }i∈I that, instead of recursively unfolding Ti , returns the same term
⊕{li : Ti }i∈I .
The function above differs from unfoldn : for example, unfold2 Tex would unfold
twice both subterms µt. ⊕{l2 : T1 } and µt.& l4 : µt′ . ⊕{l5 : T2 } . On the other
hand, applying outUnf to the same term would unfold once the term reached
with l1 and twice the one reached with l3 .
In the subtyping procedure defined below we make use of outUnf in order to
have that recursive definitions under the scope of an output are never unfolded.
This guarantees that during the execution of the procedure, even if the set of
18
reached terms could be unbounded, all the subterms starting with an output
are taken from a bounded set of terms. This is important to guarantee termination3 of the algorithm that we will define in Section 3.3 as an extension of the
procedure described below.
Subtyping Procedure. An environment Σ is a set containing pairs (T, S),
where T and S are types. Judgements are triples of the form Σ ⊢ T ≤a S which
intuitively read as “in order to succeed, the procedure must check whether T
is a subtype of S, provided that pairs in Σ have already been visited”. Our
subtyping procedure, applied to the types T and S, consists of deriving the state
space of our judgments using the rules in Figure 2 bottom-up starting from the
initial judgement ∅ ⊢ T ≤a S. More precisely, we use the transition relation
Σ ⊢ T ≤a S → Σ′ ⊢ T ′ ≤a S ′ to indicate that if Σ ⊢ T ≤a S matches the
conclusions of one of the rules in Figure 2, then Σ′ ⊢ T ′ ≤a S ′ is produced by
the corresponding premises. The procedure explores the reachable judgements
according to this transition relation. We give highest priority to rule Asmp, thus
ensuring that at most one rule is applicable.4 The idea behind Σ is to avoid
cycles when dealing with recursive types. Rules RecR1 and RecR2 deal with the
case in which we need to unfold recursion in the type on the right-hand side.
If the type on the left-hand side is not an output then the procedure simply
adds the current pair to Σ and continues. On the other hand, if an output must
be found, we apply RecR2 which checks whether such output is available. Rule
Out allows nested outputs to be anticipated (when not under recursion) and
condition A 6= [ ]1 ⇒ ∀i ∈ I.& ∈ Ti makes sure there are no orphan messages.
The remaining rules are self-explanatory. Σ ⊢ T ≤a S →∗ Σ′ ⊢ T ′ ≤a S ′ is the
reflexive and transitive closure of the transition relation among judgements. We
write Σ ⊢ T ≤a S →ok if the judgement Σ ⊢ T ≤a S matches the conclusion of
one of the axioms Asmp or End, and Σ ⊢ T ≤a S →err to mean that no rule can
3 Technically
4 The
speaking, this property of the unfolding is used in the proof of Theorem 3.11.
priority of Asmp is sufficient because all the other rules are alternative, i.e., given a
judgement Σ ⊢ T ≤a S there are no two rules different from Asmp that can be both applied.
19
(A 6= [ ]1 ) ⇒ ∀i ∈ I.& ∈ Ti
∀n.I ⊆ Jn
∀i ∈ I . Σ ⊢ Ti ≤a A[Sni ]n
Σ ⊢ ⊕{li : Ti }i∈I ≤a A[⊕{lj : Snj }j∈Jn ]n
J ⊆ I ∀j ∈ J . Σ ⊢ Tj ≤a Sj
In
Σ ⊢ &{li : Ti }i∈I ≤a &{lj : Sj }j∈J
Σ, (T, S) ⊢ T ≤a S
Asmp
Out
Σ ⊢ end ≤a end
End
Σ, (µt.T, S) ⊢ T {µt.T /t} ≤a S
RecL
Σ ⊢ µt.T ≤a S
T = end ∨ T = &{li : Ti }i∈I Σ, (T, µt.S) ⊢ T ≤a S{µt.S/t}
RecR1
Σ ⊢ T ≤a µt.S
outDepth(S) ≥ 1
Σ, (⊕{li : Ti }i∈I , S) ⊢ ⊕{li : Ti }i∈I ≤a outUnf(S)
RecR2
Σ ⊢ ⊕{li : Ti }i∈I ≤a S
Figure 2: A Procedure for Checking Subtyping
be applied to Σ ⊢ T ≤a S. Due to input branching and output selection, the
rules In and Out could generate branching also in the state space to be explored
by the procedure. Namely, given a judgement Σ ⊢ T ≤a S, there are several
subsequent judgements Σ′ ⊢ T ′ ≤a S ′ sucht that Σ ⊢ T ≤a S → Σ′ ⊢ T ′ ≤a S ′ .
The procedure could (i) successfully terminate because all the explored branches
reach a successful judgement Σ′ ⊢ T ′ ≤a S ′ →ok , (ii) terminate with an error
in case at least one judgement Σ′ ⊢ T ′ ≤a S ′ →err is reached, or (iii) diverge
because no branch terminates with an error and at least one branch never reaches
a succesful judgement.
n
Example 3.3. Consider T = µt. ⊕ l1 : &{l2 : t} and S = µt. ⊕ l1 :
o
& l2 : &{l2 : t} . Clearly, the two types T and S are related by asynchronous
subtyping, i.e. T ≤S. However, the subtyping procedure on ∅ ⊢ T ≤a S does not
terminate:
20
∅ ⊢ T ≤a S →
(T, S) ⊢ ⊕ l1 : &{l2 : T } ≤a S →
(T, S), (⊕ l1 : &{l2 : T } , S)
o
n
→
⊢ ⊕ l1 : &{l2 : T } ≤a ⊕ l1 : & l2 : &{l2 : S}
(T, S), (⊕ l1 : &{l2 : T } , S) ⊢ &{l2 : T } ≤a & l2 : &{l2 : S} →
(T, S), (⊕ l1 : &{l2 : T } , S) ⊢ T ≤a &{l2 : S} →
(T, S), (⊕ l1 : &{l2 : T } , S), (T, &{l2 : S}) ⊢ ⊕ l1 : &{l2 : T } ≤a &{l2 : S} →
(T, S), (⊕ l1 : &{l2 : T } , S), (T, &{l2 : S}), (⊕ l1 : &{l2 : T } , &{l2 : S})
n
o
⊢ ⊕ l1 : &{l2 : T } ≤a & l2 : ⊕ l1 : & l2 : &{l2 : S}
→
(T, S), (⊕ l1 : &{l2 : T } , S), (T, &{l2 : S}), (⊕ l1 : &{l2 : T } , &{l2 : S})
n
o
⊢ &{l2 : T } ≤a & l2 : & l2 : &{l2 : S}
→
...
Notice that the last step above is obtained by application of the rule Out by
considering the input context A = &{l2 : [ ]}.
The example above shows that the procedure could diverge; the next result
proves that this can happen only if the checked types are in subtyping relation.
More precisely, types T and S are not in subtyping relation if and only if the
procedure on ∅ ⊢ T ≤a S terminates with an error; formally
Proposition 3.4. Given the types T and S, we have that there exist Σ′ , T ′ , S ′
such that T ≤
6 S if and only if ∅ ⊢ T ≤a S →∗ Σ′ ⊢ T ′ ≤a S ′ →err .
This means that: if T ≤
6 S then the procedure on ∅ ⊢ T ≤a S surely terminates
with an error; if, instead, T ≤S then the procedure terminates successfully or
diverges.
3.2. k-bounded Asynchronous Subtyping
In the previous subsection we have shown that the standard subtyping procedure does not terminate in general. In order to guarantee termination, Lange
and Yoshida [11] have considered limitations to the communication buffer, like
21
half-duplex (in this case asynchronous and synchronous subtyping coincides) or
alternating protocols (in this case the buffer will store at most one message).
We now prove a more general decidability result. We show that, for every k, we
can define an algorithm for the notion of k-bounded asynchronous subtyping introduced in Section 2.1, building on the subtyping procedure defined previously.
We consider an algorithm, that we denote with ≤ka , obtained from the
above procedure for ≤a simply by imposing that the input context A, used in
rule Out in Figure 2, is always k-bounded. Then, the following result holds:
Theorem 3.5. The algorithm for ≤ka always terminates and, given the types
T and S, there exist Σ′ , T ′ , S ′ such that ∅ ⊢ T ≤ka S →∗ Σ′ ⊢ T ′ ≤a S ′ →err if
and only if T ≤
6 k S.
3.3. Asynchronous Subtyping for Single-Out or Single-In Types
In Example 3.3 we have seen that, if we consider the terms T = µt. ⊕ l1 :
o
n
&{l2 : t} and S = µt. ⊕ l1 : & l2 : &{l2 : t} , the subtyping procedure in
Figure 2 applied to ∅ ⊢ T ≤a S does not terminate. The problem is that the
termination rule Asmp cannot be applied because the term on the r.h.s. (i.e.
n
the supertype) generates always new terms in the form & l2 : & l2 : . . . &{l2 :
o
S} . . . .
Notice that, in this particular example, these infinitely many distinct terms
are obtained by adding single inputs (i.e. single-choice input branchings) in
front of the term in the r.h.s.: we call this linear input accumulation.
For simple cases like this one, solutions have been proposed by Lange and
Yoshida [11] and Bravetti et al. [10]. The idea is to extend the subtyping procedure in Figure 2 with additional termination rules able to detect when it is no
longer necessary to continue because it entered a deterministic loop (where the
only possible future behavior of the procedure is to repeat indefinitely the same
linear input accumulation). This approach holds only under two assumptions,
both satisfied by the subtyping relations considered in Lange and Yoshida [11]
and Bravetti et al. [10]: while checking subtyping output selections in the l.h.s.
22
(i.e. the subtype) are always single-choice and the same holds for input branchings in the r.h.s. (i.e. the supertype). This implies that there is a linear input
accumulation, which is the repetition of a specific sequence of input labels. The
combination of these two assumptions guarantees that the subtyping procedure
proceeds deterministically: this makes it possible to detect whether it enters a
loop because the unique kind of loops are the deterministic ones.
In this section we show that it is possible to relax at least one of these two
assumptions: either deal with the case in which the input accumulation is not
linear, or deal with the case in which output selections in the l.h.s. are not
single-choice. More precisely, the two cases that we consider are the following
ones: subtyping between single-out session types (where input branchings in
the r.h.s. are not constrained to be single-choice as in previous approaches) and
subtyping between single-in session types (where output selections in the l.h.s.
are not constrained to be single-choice as in previous approaches), i.e. the two
relations ≤sout and ≤sin , respectively, that we introduced in Section 2.1. The
idea is to find an algorithm for one of the two cases and apply it also to the
other one by exploiting type duality.
In the single-in case we surely have linear input accumulation but the subtyping procedure is no longer deterministic due to non-single output selections
in the l.h.s. that have multiple possible continuations. This causes the approach
proposed in Lange and Yoshida [11] and Bravetti et al. [10] to fail because now
the procedure can incur in nondeterministic loops (so it is not guaranteed to
repeat indefinitely the accumulation behavior detected by the additional termination rule they consider). On the other hand, in the single-out case we loose
the linear input accumulation but we do not have output selections to cause the
problematic nondeterminism discussed above.
The latter advantage led us to opt for the single-out case, which we were able
to manage by adopting a totally new approach where the input accumulation
is represented in the form of a tree (thus accounting for all possible alternative
accumulated input behaviors at the same time).
We start with an example of subtyping between single-out types that cannot
23
be managed with the approach in Lange and Yoshida [11] and Bravetti et al. [10]
because there is non-linear input accumulation.
n
Example 3.6. Consider T = µt. ⊕ l1 : &{l2 : t, l3 : t} and S = µt. ⊕ l1 :
o
& l2 : &{l2 : t}, l3 : t . We now comment the application of the subtyping
procedure on ∅ ⊢ T ≤a S.
∅ ⊢ T ≤a S →
(T, S) ⊢ ⊕ l1 : &{l2 : T, l3 : T } ≤a S →
(T, S), (⊕ l1 : &{l2 : T, l3 : T } , S) ⊢
o
n
→
⊕ l1 : &{l2 : T, l3 : T } ≤a ⊕ l1 : & l2 : &{l2 : S}, l3 : S
(T, S), (⊕ l1 : &{l2 : T, l3 : T } , S) ⊢
&{l2 : T, l3 : T } ≤a & l2 : &{l2 : S}, l3 : S
At this point, the subtyping procedure has two continuations, one for the label l2
and one for the label l3 . In case of label l3 we reach the judgement:
(T, S), (⊕ l1 : &{l2 : T, l3 : T } , S) ⊢ T ≤a S
on which the termination rule Asmp can be applied. In case of label l2 we have:
(T, S), (⊕ l1 : &{l2 : T, l3 : T } , S) ⊢ T ≤a &{l2 : S} →
(T, S), (⊕ l1 : &{l2 : T, l3 : T } , S), (T, &{l2 : S})
⊢ ⊕ l1 : &{l2 : T, l3 : T } ≤a &{l2 : S} →
(T, S), (⊕ l1 : &{l2 : T, l3 : T } , S), (T, &{l2 : S}),
(⊕ l1 : &{l2 : T, l3 : T } , &{l2 : S})
n
⊢ ⊕ l1 : &{l2 : T, l3 : T } ≤a & l2 : ⊕ l1 : & l2 : &{l2 : S}, l3 : S
(T, S), (⊕ l1 : &{l2 : T, l3 : T } , S), (T, &{l2 : S}),
(⊕ l1 : &{l2 : T, l3 : T } , &{l2 : S})
o
n
→
⊢ &{l2 : T, l3 : T } ≤a & l2 : & l2 : &{l2 : S}, l3 : S
(T, S), (⊕ l1 : &{l2 : T, l3 : T } , S), (T, &{l2 : S}),
(⊕ l1 : &{l2 : T, l3 : T } , &{l2 : S})
⊢ T ≤a & l2 : &{l2 : S}, l3 : S →
...
24
o
→
Notice that in the last judgement, the r.h.s. has a non-linear input accumulation
starting with an input choice on two labels l2 and l3 .
3.3.1. Asynchronous Subtyping for Single-Out Types
We now present our novel approach to asynchronous subtyping that can be
applied to single-out types, hence also to the types in the above Example 3.6,
that will be used as a running example in this section. As anticipated, the main
novelty is the ability to deal with non-linear input accumulation by representing
it as a tree. We need to be able to extract the leafs from these trees: this is
done by the leaf set function defined as follows.
Definition 3.7 (Leaf Set). Given a session type S, we write noIn(S) if S is
not of the form &{li : Si }i∈I . Given a session type T , we define
leafSet(T ) = {T1 , . . . , Tn | noIn(Ti ) and ∃ input context A s.t. T = A[Tk ]k∈{1...n} }
The leaf set of a session type T is the set of subterms different from inputs
that are reachable from its root through a path of inputs. For example, the
leaf set of the term &{l1 : µt. ⊕ {l2 : t}, l3 : &{l4 : ⊕{l2 : µt. ⊕ {l2 : t}}}} is
{µt. ⊕ {l2 : t}, ⊕{l2 : µt. ⊕ {l2 : t}}. If we consider the r.h.s. term in the last
judgement in Example 3.6, we have that leafSet(& l2 : &{l2 : S}, l3 : S ) = {S}.
During the check of subtyping, according to Figure 2 (rule Out), when a
term in the r.h.s. having input accumulation has to mimic an output in front
of the l.h.s., such output must be present in front of all the leafs of the tree.
In this case, the checking continues by anticipating the output from all the
leafs. The following auxiliary function output anticipation indicates the way a
term changes after having anticipated a sequence of outputs. Notice that in
the definition we make use of the assumption on single-out session types, by
considering single-choice output selections.
Definition 3.8 (Output Anticipation). Partial function antOut(T, li1 · · · lin ),
with T single-out session type and li1 · · · lin sequence of labels, is inductively de-
25
fined as follows:
antOut(T, li1 · · · lin ) =
T
if n = 0
k
A[T ]k if outUnf(antOut(T, l · · · l
in−1 )) = A[⊕{lin : Tk }]
k
i1
We say that T can infinitely anticipate outputs, written antOutInf(T ), if there
exists an infinite sequence of labels li1 · · · lij · · · such that antOut(T, li1 · · · lin ) is
defined for every n.
The function antOut(T, l̃) anticipates all outputs in the sequence ˜l. For example,
the function applied to &{l : µt. ⊕ {l1 : ⊕{l2 : t}}, l′ : ⊕{l1 : µt. ⊕ {l2 : ⊕{l1 :
t}}}} and the sequence (l1 , l2 ) would return the same term, while it would be
undefined with the sequence (l1 , l1 ). If we go back to our running Example 3.6,
we have that antOut(S, l1 ) = & l2 : &{l2 : S}, l3 : S . Moreover, we have that
antOutInf(S) holds because the label l1 can be infinitely anticipated.
The definition of antOutInf(T ) is not algorithmic in that it quantifies on
every possible natural number n. Nevertheless, as we show below, it can be
decided by checking whether for every session type obtained from T by means
of output anticipations, all the terms populating its leaf set can anticipate the
same output label. Although such process may generate infinitely many session
types, the terms populating the leaf sets are finite and are over-approximated
by the function reach(T ), which always returns a finite set and is defined as:
Definition 3.9 (Reachable Types). Given a single-out session type T, reach(T )
is the minimal set of session types such that:
1. T ∈ reach(T );
2. &{li : Ti }i∈I ∈ reach(T ) implies Ti ∈ reach(T ) for every i ∈ I;
3. µt.T ′ ∈ reach(T ) implies T ′ {µt.T ′ /t} ∈ reach(T );
4. ⊕{l : T ′ } ∈ reach(T ) implies T ′ ∈ reach(T ).
Notice that reach(T ) is populated by those session types obtained by consuming in sequence the initial inputs and outputs, and by unfolding recursion
only when it is at the top level. As an example, consider the session type S of
26
the Example 3.6. We have
n
reach(S) = S, ⊕ l1 : & l2 : &{l2 : S}, l3 : S
o
, & l2 : &{l2 : S}, l3 : S , &{l2 : S}
For every type T , we have that the terms in reach(T ) are finite; in fact,
during the generation of such terms, eventually the term end or a term already
considered is reached. The latter occurs after consumption of all the inputs and
outputs in front of a recursion variable already unfolded.
Proposition 3.10. Given a single-out session type T , reach(T ) is finite and it
is decidable whether antOutInf(T ).
Subtyping algorithm for single-out types. We are now ready to present
the new termination condition that once added to the subtyping procedure in
Figure 2 makes it a valid algorithm for checking subtyping for single-out types.
The termination condition is defined as an additional rule, named Asmp2, that
complements the already defined Asmp rule by detecting those cases in which
the subtyping procedure in Figure 2 does not terminate.
The new rule is defined parametrically on the session type Z, which is the
type on the right-hand side of the initial pair of types to be checked (i.e. the
algorithm is intended to check V ≤Z, for some type Z). We start from the initial
judgement ∅ ⊢ V ≤t Z and then apply from bottom to top the rules in Figure
2, where ≤a is replaced by ≤t , plus the following additional rule:
S ∈ reach(Z)
antOutInf(S)
|γ| < |β|
leafSet(antOut(S, γ)) = leafSet(antOut(S, β))
Σ, (T, antOut(S, γ)) ⊢ T ≤t antOut(S, β)
Asmp2
We first observe that this termination rule can be applied to the last judgement of our running Example 3.6. We have already seen that S ∈ reach(S),
antOutInf(S) holds, antOut(S, l1 ) = & l2 : &{l2 : S}, l3 : S and that leafSet(& l2 :
&{l2 : S}, l3 : S ) = {S}.
We now observe that antOut(S, ε) = S and
leafSet(S) = {S}, hence we can conclude that we can apply the above termination rule Asmp2 to the last judgement in Example 3.6 by instantiating γ = ε
and β = l1 .
27
The first property of the new algorithm that we prove is termination. Intuitively, we have that this new termination rule guarantees to catch all those
cases where the term on the right grows indefinitely, by anticipating outputs
and accumulating inputs. These infinitely many distinct types are anyway obtainable starting from the finite set reach(Z), by means of output anticipations.
Hence there exists S ∈ reach(Z) that can generate infinitely many of these types:
this guarantees antOutInf(S) to be true. As observed above, the leaves of such
infinitely many terms are themselves taken from the finite set reach(Z). This
guarantees that the algorithm, among the types that can be obtained from S,
visits two terms having the same leaf set. These, even if syntactically different,
are equivalent as far as the subtyping game is regarded.
Concerning the precise definition of the algorithm, in order to avoid the
possibility of applying two distinct rules to the same judgement, we give rule
Asmp2 the same priority as rule Asmp (both rules have highest priority). Also
in this case, we use Σ ⊢ T ≤t S → Σ′ ⊢ T ′ ≤t S ′ to denote that the latter can
be obtained from the former by one rule application, and Σ ⊢ T ≤t S →err , to
denote that there is no rule that can be applied to the judgement Σ ⊢ T ≤t S.
We can now state the termination and soundness of the algorithm:
Theorem 3.11. Given two single-out session types T and S, the algorithm
applied to the initial judgement ∅ ⊢ T ≤t S terminates.
Theorem 3.12. Given two single-out session types T and S, we have that there
exist Σ′ , T ′ , S ′ such that ∅ ⊢ T ≤a S →∗ Σ′ ⊢ T ′ ≤a S ′ →err if and only if there
exist Σ′′ , T ′′ , S ′′ such that ∅ ⊢ T ≤t S →∗ Σ′′ ⊢ T ′′ ≤t S ′′ →err .
Finally, we can conclude the decidability of asynchronous subtyping for
single-out session types.
Corollary 3.13 (Decidability for Single-out Types). Asynchronous subtyping for single-out session types ≤sout is decidable.
We now show that the above decidability results hold also for the ≤tin,sout
relation (where we further restrict the asynchronous subtyping relation not to
28
admit contravariance on input branchings). In the algorithm we just modify
the rule In of Figure 2 by changing the constraint J ⊆ I in the premise into
J = I, thus obtaining modified versions of Σ ⊢ T ≤a S → Σ′ ⊢ T ′ ≤a S ′ (and
Σ ⊢ T ≤a S →err ) and Σ ⊢ T ≤t S → Σ′ ⊢ T ′ ≤t S ′ (and Σ ⊢ T ≤t S →err ).
We have that Proposition 3.4, where relation ≤tin is considered instead of ≤,
termination Theorem 3.11 and soundness Theorem 3.12, where the modified
judgments ≤a and ≤t are considered, still hold (they are proved with exactly
the same proofs as those reported in Appendix B for the original statements).
Corollary 3.14. Asynchronous subtyping for single-out session types without
input contravariance ≤tin,sout is decidable.
3.3.2. Asynchronous Subtyping for Single-in Types
First of all we notice that an obvious consequence of Corollary 3.13 is that
also ≤sin,sout is decidable (we just have to add a preliminary check verifying
that both types are single-in). Moreover, exploiting dual closeness, i.e. the fact
that T ≤sin S if and only if S ≤sout T (see Section 2.1), we can use the algorithm
presented for single-out types also for the case of single-in types.
Corollary 3.15 (Decidability for Single-in Types). Asynchronous subtyping for single-in session types ≤sin is decidable.
We can therefore identify an asynchronous dual closed subtyping relation
that stands in the boundary of decidability.
Corollary 3.16 (Decidability for Single-in or Single-out Types).
The
asynchronous dual closed subtyping relation ≤sin ∪≤sout is decidable.
Finally, similarly as we did for T ≤sin S, by exploiting Proposition 2.10 we
can use the modified algorithm employed for ≤tin,sout subtyping for deciding the
remaining relation ≤sin,tout .
Corollary 3.17. Asynchronous subtyping for single-in session types without
output covariance ≤sin,tout is decidable.
29
4. Undecidability Results
We now move to undecidability results. We first consider bounded asynchronous subtyping ≤bound . The proof in this case is a variant of the proof we
already presented in our previous work [10], where we encoded the problem of
checking (non) termination in queue machines (a well-known Turing powerful
formalism) into checking session subtyping. Technically speaking, we resort to
a different property, namely bounded non termination, that we here show to be
undecidable for queue machines.
The second, and main, undecidability result concerns subtyping without output covariance and input contravariance ≤tin,tout . The proof in this case requires
deep modifications to our proof technique, due to the impossibility to exploit
covariance/contravariance in the queue machine encoding. We deal with the
absence of covariance/contravariance by saturating each point of choice on the
entire considered alphabet. This has a strong impact on the encoding because
it introduces additional choices, in the session types, whose continuations do
not correspond to the behaviour of the considered queue machine. This problem is solved by ensuring that these additional choices and the corresponding
continuations are irrelevant as far as subtyping checking is concerned. Such
solution, however, works only for a fragment of queue machines (that we call
single-consuming queue machines) that we prove to be Turing complete as well.
We consider this second result interesting for the following reason: the previous undecidability proofs [10, 11] made use of both output covariance/input
contravariance (already present in synchronous session subtyping) and output
anticipation (specific for asynchronous subtyping), hence our new proof shows
that (provided that the syntax of types is not constrained, e.g, imposing single
choices for output selections or input branchings) the source of undecidability is
to be precisely localized into the latter, as the former is not necessary to prove
undecidability.
We first report the definition of queue machines.
30
4.1. Queue Machines
Queue machines are a formalism similar to pushdown automata, but with a
queue instead of a stack. Queue machines are Turing-equivalent [15].
Definition 4.1 (Queue Machine). A queue machine M is defined by a sixtuple (Q, Σ, Γ, $, s, δ) where:
• Q is a finite set of states;
• Σ ⊂ Γ is a finite set denoting the input alphabet;
• Γ is a finite set denoting the queue alphabet;
• $ ∈ Γ − Σ is the initial queue symbol;
• s ∈ Q is the start state;
• δ : Q × Γ → Q × Γ∗ is the transition function.
A configuration of a queue machine is an ordered pair (q, γ) where q ∈ Q
is its current state and γ ∈ Γ∗ is the content of the queue (Γ∗ is the Kleene
closure of Γ). The starting configuration on an input string x ∈ Σ∗ is (s, x$).
The transition relation →M from one configuration to the next one is defined as
(p, Aα) →M (q, αγ), when δ(p, A) = (q, γ). A machine M accepts an input x if it
blocks by emptying the queue. Formally, x is accepted by M if (s, x$) →∗M (q, ǫ)
where ǫ is the empty string and →∗M is the reflexive and transitive closure of
→M . Intuitively, a queue machines is a Turing machine with a special tape that
works as a FIFO queue.
The Turing completeness of queue machines is discussed by Kozen [15] (page
354, solution to exercise 99). A configuration of a Turing machine (tape, current head position and internal state) can be encoded in a queue, and a queue
machine can simulate each move of the Turing machine by repeatedly consuming and reproducing the queue contents, only changing the part affected by
the move itself. The undecidability of termination for queue machines follows
directly from such encoding.
31
a) Finite Control
µq.&{A : ⊕{B1A : · · · ⊕ {BnAA : [[q ′ ]]S∪q }}}A∈Γ
if q 6∈ S and δ(q, A) = (q ′ , B1A · · · BnAA )
S
[[q]] =
q
if q ∈ S
b) Queue
[[C1 · · · Cm ]] = &{C1 : . . . &{Cm : µt. ⊕ {A : &{A : t}}A∈Γ }}
Figure 3: Encoding of the Finite Control and the Queue of a Queue Machine
4.2. Bounded Asynchronous Subtyping
We now consider the notion of bounded asynchronous subtyping ≤bound we
introduced in Section 2.1,
The proof of undecidability of ≤bound follows the approach we already used to
prove the undecidability of single-choice asynchronous subtyping << [10] (that
we have commented in the Introduction). The idea is to define, given a queue
machine M and its input x, two session types S and T , such that S is a subtype
of T if and only if M does not accept x. More precisely, the type S models the
finite control of the queue machine M while the type T models the queue that
initially contains the sequence x$.
More precisely, the encoding of queue machines is as follows [10].
Definition 4.2 (Queue Machine Encoding). Let M = (Q, Σ, Γ, $, s, δ) be a
queue machine, and let C1 , · · · , Cm ∈ Γ, with m ≥ 0, q ∈ Q and S ⊆ Q. The finite control encoding function [[q]]S and the queue encoding function [[C1 · · · Cm ]]
are defined as in Figure 3(a) and Figure 3(b) respectively. The initial encoding
of M with input x is given by the pair of types [[s]]∅ and [[x$]].
The basic idea behind the encoding of the finite control is to use a recursively defined type with a recursion variable q for each state q of the encoded
queue machine M . The type corresponding to the recursion variable q starts
with an input with multiple choices, one for each possible symbol that can be
32
consumed from the queue. The continuation is composed of a sequence of singlechoice inputs labeled with the symbols B1A . . . BnAA , where B1A . . . BnAA are the
symbols enqueued by the queue machine when, in state q, consumes A from
the queue. Assuming that q ′ is the new state of M after execution of this step
(i.e. δ(q, A) = (q ′ , B1A . . . BnAA )), the type becomes the one corresponding to the
recursion variable q′ .
On the other hand, the type modeling the queue with contents C1 . . . Cm
is denoted with [[C1 . . . Cm ]]: this type starts with a sequence of single-choice
inputs labeled with the symbols C1 . . . Cm , followed by a recursive type. Such
type starts with an output with multiple-choices, one for each symbols that
can be enqueued, followed by a single-choice input having the same label. This
particular type has the following property: if one label A of the multiple-choice
output is selected for anticipation during the subtyping simulation game, the
corresponding single-choice input labeled with A is enqueued at the end of the
sequence of inputs preceding the recursive definition. This perfectly corresponds
to the behaviour of the queue in the modeled queue machine.
As mentioned above, this encoding has been already used to prove the undecidability of << [10]. More precisely, we proved that given a queue machine
M = (Q, Σ, Γ, $, s, δ) and an initial input x, we have that [[s]]∅ <<[[x$]] if and only
if x is not accepted by M (i.e. M does not terminate on input x). The same
result does not hold for the bounded asynchronous subtyping because there are
cases in which M does not accept x but [[s]]∅ 6 ≤bound[[x$]], in particular, those
cases in which the subtyping simulation game generates unbounded accumulation of inputs. For this reason we have to consider a more complex undecidable
property for queue machines: bounded non termination, i.e., the ability of a
queue machine to have an infinite computation while keeping the length of the
queue bounded. We now define the notion of boundedness for queue machines
and then prove that bounded non termination is undecidable.
Definition 4.3 (Queue Machine Boundedness). Let M be a queue machine
and x a possible input. We say that M is bound on input x if there exists k such
33
that, for every q and γ such that (s, x$) →∗M (q, γ), we have that |γ| ≤ k.
Lemma 4.4. Given a queue machine M and an input x, it is undecidable
whether M does not terminate and is bound on x.
Following the proof technique we already used to prove undecidability of <<,
i.e. by reducing the termination problem for queue machines into subtyping
checking [10], we can prove also the undecidability of ≤bound by reduction from
the bounded non termination problem.
Theorem 4.5. Given a queue machine M = (Q, Σ, Γ, $, s, δ) and an input
string x, we have that [[s]]∅ ≤bound [[x$]] if and only if M does not terminate and
is bound on x.
Corollary 4.6. Bounded asynchronous subtyping ≤bound is undecidable.
4.3. Undecidability of Asynchronous Subtyping without Output Covariance and
Input Contravariance
We now move to the proof of undecidability of ≤tin,tout , the asynchronous
subyping relation, we introduced in Section 2.1, that does not admit output
covariance and input contravariance by imposing matching choices to have the
same set of labels.
The proof technique is still based on an encoding of queue machines, but we
have to significantly improve the encoding discussed in the previous subsection.
In fact, the encoding of Figure 3 exploits both input contravariance (in the
matching between the multiple-choice input at the beginning of the encoding
of the finite control and the initial single-choice inputs of the queue encoding)
and output covariance (in the matching between the multiple-choice output at
the beginning of the recursive part of the queue encoding and the single-choice
outputs in the encoding of the finite control).
The new encoding that we propose saturates all choices, both inputs and
outputs, with labels corresponding to the entire queue alphabet. The addition
of these labels and of the corresponding continuations, introduces new possible
34
paths in the subtyping simulation game. We are able to make these additional
behaviour irrelevant, but at the price of restricting the class of encoded queue
machines. These queue machines are named single consuming queue machines;
their characteristic is to guarantee that in two subsequence actions, at least one
of the two will enqueue symbols.
Definition 4.7 (Single Consuming Queue machine). We say that a queue
machine M = (Q, Σ, Γ, $, s, δ) is single consuming if δ(q, a) = (q ′ , ǫ), for some
q, a and q ′ , implies that there exist no b and q ′′ such that δ(q ′ , b) = (q ′′ , ǫ).
We have that single consuming queue machines are still Turing-complete (see
Appendix C.2 for the detailed proof based on an encoding of queue machines
into single consuming queue machines):
Theorem 4.8. Given a single consuming queue machine M and an input x,
the termination of M on x is undecidable.
We prove the undecidability of ≤tin,tout by encoding single consuming queue
machines into the subtyping simulation game. Following the approach already
discussed in the previous subsection, given a queue machine, our encoding generates a pair of types, say T and S, such that T encodes the finite control and
S encodes the queue. Then, the subtyping T ≤tin,tout S simulates the execution
of the machine.
We are now ready to present the definition of the new encoding where we
make use of the following new notation: {li : Ti }i∈I ⊎ {lj : Tj }j∈J = {lk :
Tk }k∈I∪J .
Definition 4.9 (Encoding Single Consuming Queue machines). Let M =
(Q, Σ, Γ, $, s, δ) be a queue machine such that q ∈ Q, S ⊆ Q and C1 , · · · , Cm ∈ Γ,
S
with m ≥ 0. The finite control encoding function [[[q]]] and the queue encoding
function [[[C1 · · · Cm ]]] are defined as in Figure 4(a) and Figure 4(b) respectively.
As discussed in the previous subsection, the idea is that the type encoding
the finite control is able to perform an input on each of the symbols in Γ, and
35
a) Finite Control
µq.&{A : {{B1A · · · BnA }}S∪{q}
}A∈Γ
A
q′
S
[[[q]]] =
q
if q 6∈ S and
δ(q, A) = (q ′ , B1A · · · BnAA )
if q ∈ S
b) Queue
µt ⊕ A : & A : t ⊎ A′ : T ′′ ′
A ∈Γ\{A}
A∈Γ
[[[C1 . . . Cm ]]] =
& C : [[[C . . . C ]]] ⊎ A′ : T ′′
1
2
m
A′ ∈Γ\{C1 }
if m = 0
otherwise
where:
T
{{B1 · · · Bm }}r
[[[r]]]T
=
⊕ B1 : {{B2 . . . Bm }}T
r
T ′ = µt.& A1 : ⊕{A2 : t}A2 ∈Γ
if m = 0
⊎ A′ : T ′
A′ ∈Γ\{B1 }
otherwise
A1 ∈Γ
T ′′ = µt.& A1 : & A2 : ⊕{A3 : t}A3 ∈Γ
A2 ∈Γ A1 ∈Γ
Figure 4: Encoding of the Finite Control and the Queue of a Single Consuming Queue Machine
continue according to the definition of the transition function δ. The type representing the queue then matches such input with the correct symbol depending
on the state of the queue. For instance, in the encoding described in the previous subsection, if we denote with T and S the types representing the finite
control and the queue respectively, and if Γ = {A, B} and symbol A is on the
head of the queue, we have T = &{A : . . . , B : . . .} and S = &{A : . . .}: type
T is able to react to any symbol that may be present on the queue (like the
transition function δ), while type S reacts with the actual value on the queue,
symbol A. Unfortunately, such idea exploits contravariance for inputs. Therefore, it must be the case, in the new encoding, that the input in S is of the form
&{A : . . . , B : . . .}. We make sure that if label A is selected then the simulation
of the queue machine continues. Otherwise, an infinite subtyping simulation
36
game is started (starting from B in the example).
Also the insertion of symbols in the queue was simulated in the encoding of
the previous subsection by exploiting output covariance. The type representing
the finite control performs a single-choice output that is matched by a multiplechoice output having the effect of adding a corresponding symbol at the end of
the input accumulated in the type modeling the queue. Also in this case, we
have to add choices to the type modeling the finite control: also in this case we
ensure that these extra paths start an infinite subtyping simulation game.
These additional paths make the subtyping simulation game highly nondeterministic and such that several paths that the game can take differ from
what the encoded machine does. We discuss in detail the various cases which
our encoding in Figure 4 can be in:
1. The encoding of the finite control reads the correct symbol. We represent
the machine reading a symbol A from the queue while being in state q,
S∪{q}
with an input type of the form &{A : {{B1A · · · BnAA }}q′
}A∈Γ , where
each branch corresponds to a possible symbol that can be read. On the
other hand, a queue C1 · · · Cm is encoded as an input type of the form
& C1 : [[[C2 . . . Cm ]]] ⊎ A′ : T ′′ A′ ∈Γ\{C } where the branch with label
1
C1 represents the actual content of the queue. Hence, in the simulation
game, if the finite control reads symbol A and this is matched by the
S∪{q}
correct symbol in the queue, then the type {{B1A · · · BnAA }}q′
deals with
inserting symbols B1A · · · BnAA into the queue.
2. The encoding of the finite control reads the wrong symbol. In this case,
the encoding of the finite control picks a symbol that is not that in the
queue head. In order to match it, the encoding of the queue will take
′
A : T ′′ A′ ∈Γ\{C1 } . Type T ′′ is designed in a way that it can match
every move of the finite control, by repeatedly alternating two inputs with
a subsequent output on every queue symbol. Note that, since inputs
cannot be anticipated, matching every move is feasible only if the encoded
machine is single consuming.
37
3. The encoding of the finite control writes the correct symbol. Once the finite
control has read a symbol, it performs {{B1 · · · Bm }}Tr , which simulates
the writing of B1 · · · Bm into the queue. If m = 0 then it moves to the
encoding of the next state according to function δ. Otherwise, it translates
T
to the type ⊕ B1 : {{B2 . . . Bm }}r ⊎ A′ : T ′ A′ ∈Γ\{B1 } . The queue,
in order to match B1 (and B2 , . . . , Bm ) can always anticipate outputs
with the term µt ⊕ A : & A : t ⊎ A′ : T ′′ A′ ∈Γ\{A} A∈Γ which, after
consuming a label A will add an input with label A, simulating the adding
of A to the queue.
4. The encoding of the finite control writes the wrong symbol. In this case, the
finite control writes a symbol to the queue with ⊕ B1 : {{B2 . . . Bm }}Tr ⊎
′
A : T ′ A′ ∈Γ\{B1 } . However, the simulation executes the wrong output
(with any A′ 6= B1 ) and continues as T ′ . In this case, T ′ continues removing and adding any value from the queue, indefinitely. Note that it may
remove the wrong value from the queue overlapping with case 2. In this
case, the requirement that the queue machine is single consuming is not
necessary.
Example 4.10. In order to further clarify our encoding, consider a queue machine with states {s, q} (where s is the starting state), queue alphabet Γ = {X, Y }
and transition relation δ such that δ(s, A) = (q, A) and δ(q, A) = (s, ǫ), for every A ∈ Σ. Clearly, the machine terminates on any input. The encoding of the
finite control is the following session type:
X : ⊕{X : [[[q]]]s , Y : T ′ }
∅
[[[s]]] = µs.&
Y : ⊕{Y : [[[q]]]s , X : T ′ }
[[[q]]]
{s}
X : ⊕{X : s, Y : T ′ }
= µq.&
Y : ⊕{Y : s, X : T ′ }
Assume, e.g., that the queue initially contains the string XY . The machine will
empty the queue by visiting state q twice and terminate in state s with the empty
queue. If we now run the subtyping simulation game between the encoding of
38
finite control above and the encoding of the queue we will end up with two types
that are not in subtyping: the encoding of the state s starting with an input and
the encoding of the empty queue that does not match it.
The encoding of the finite-control and of the queue are such that the following properties hold: given a queue machine M with initial state s and initial
queue symbol $, if M does not accept x then it is possible to define an asynchronous subtyping relation that includes the pair ([[[s]]]∅ , [[[x$]]]); moreover, if
∅
[[[s]]] ≤tin,tout [[[x$]]] then it is possible to conclude that M does not terminate (i.e.
does not accept) on input x. We thus have the following:
Theorem 4.11. Given a single consuming queue machine M = (Q, Σ, Γ, $, s, δ)
∅
and an input string x ∈ Σ∗ , we have [[[s]]] ≤tin,tout [[[x$]]] if and only if M does not
terminate on x.
We can therefore conclude that subtyping without output covariance and
input contravariance is undecidable.
Corollary 4.12 (Undecidability of Subtying without Co/contravariance).
The asynchronous dual closed subtyping relation ≤tin,tout is undecidable.
In the same way we can also show that ≤tin and ≤tout are undecidable
and provide an alternative proof of undecidability of ≤. This because, since
for the types obtained with the encoding (for which the ability to match via
covariance/contravariance is irrelevant) obviously such relations coincide, i.e.
∅
∅
∅
[[[s]]] ≤tin,tout [[[x$]]] if and only if [[[s]]] ≤tin [[[x$]]] if and only if [[[s]]] ≤tout [[[x$]]] if and
∅
only if [[[s]]] ≤[[[x$]]], Theorem 4.11 holds also if we replace the ≤tin,tout relation
with one of such relations.
Corollary 4.13. Asynchronous subtyping relations ≤tin , ≤tout and ≤ are undecidable.
39
5. Related Work
Subtyping for Session Types. Subtyping for session types was first introduced by Gay and Hole [7]5 for a session-based π-calculus where communication
is synchronous, i.e., an output directly synchronises with an input. In such case,
the relation allows no output anticipation. However, as in our case, outputs are
covariant and inputs are contravariant.
To the best of our knowledge, Mostrous et al. [14] were the first to adapt
the notion of session subtyping to an asynchronous setting. Their computation
model is a session π-calculus with asynchronous communication that makes
use of session queues for maintaining the order in which messages are sent.
They introduce the idea of output anticipation, which is also a main feature
of our theory. Mostrous and Yoshida [12] extended the notion of asynchronous
subtyping to session types for the higher-order π-calculus. In the same article,
Mostrous and Yoshida observe that their definition of asynchronous subtyping
allows orphan messages. Orphan message are prohibited with the definition
of subtyping given by Chen et al. [8]. In their article, they show that such a
definition is both sound and complete w.r.t. type safety and orphan message
freedom.
Undecidability Results. Mostrous et al. [14] proposed a procedure to check
asynchronous subtyping for multiparty session types. Differently from what
stated therein, the procedure does not terminate due to unbounded message accumulation in the queues, e.g. for the terms in Example 3.3. Such a procedure
inspired the one we presented in Section 3.1. The problem of unbounded accumulation was observed by Mostrous and Yoshida [12]. The impossibility to define a
correct algorithm has been independently proved by Lange and Yoshida [11] and
Bravetti et. al [10]. Lange and Yoshida [11] reduce Turing machine termination
into a notion of compatibility for communicating automata and, then, transfer
5 The
Gay and Hole subtyping is contravariant on outputs and covariant on inputs. This is
because a channel-based subtyping [9] is considered instead of our process-oriented subtyping.
40
such a result to session types. This proof technique applies only to dual closed
subtyping relations, like the one by Chen et al. [8]. The proof by Bravetti et.
al [10], on the other hand, exploits a direct encoding of queue machines into
session subtyping. This made it possible to prove undecidability of all the other
notions of asynchronous subtyping in the literature. Unlike the encoding in this
paper (Figure 4), both encodings take advantage of the use of output covariance
and input contravariance. For example, by exploiting this feature, the queue
machine encoding by Bravetti et al. [10] (Figure 3) is much simpler than the
encoding we need to use here. We notice that our results on undecidability focus on binary session types. However, it is immediate to generalise this kind of
undecidability results from binary to multiparty sessions (binary session types
are just multiparty session types with only two roles [10]).
Decidability Results. Synchronous subtyping for binary session types is decidable [7]. Both Bravetti et al. [10] and Lange and Yoshida [11] investigate
fragments of session types for which asynchronous subtyping becomes decidable.
However, such fragments are much more limited, and far from having practical
applications, with respect to those considered here. Both address cases where
one of the compared types is a single-choice session type, i.e. all its branchings
and selections are single-choice. Thus they are both, basically, special cases of
our subtyping for single-in or single-out types (≤sin ∪≤sout ). In particular, Lange
and Yoshida give an algorithm for deciding subtyping between a general session
type and a single-choice session type. Although it may seem that such case is
not properly included in our decidable subtyping relation for single-out/singlein types, covariance and contravariance ensure that all types containing at least
one multiple input branch and one multiple output selection (both reachable
in the subtyping simulation game) cannot be related with a single-choice type.
Bravetti et al. [10] prove decidability for relations <<sin and <<sout that pose
an analogous restriction to the branching/selection structure, but that allow for
orphan messages. <<sin and <<sout are fragments where related types (T, S) are
such that, either T is single-choice and S is single-in (<<sin ), or T is single-out
and S is single-choice (<<sout ). For types that do not produce orphan messages,
41
the sutyping of Bravetti et al. [10] is just a special case of our single-in (≤sin )
and single-out (≤sout ) session subtyping.
Additionally, Lange and Yoshida state the decidability of subtyping for halfduplex communication [16] and alternating machines: the former coincides with
synchronous subtyping while the latter can be reduced to 1-bounded asynchronous subtyping as discussed in Section 3.2.
Comparison with our Previous Work [10]. Concerning decidability results, the approaches taken in this and our previous paper are both based on
initially considering a (non always terminating) subtyping procedure ≤a , inspired from Mostrous et al. [14], and then presenting a subtyping algorithm
≤t obtained by adding a new termination Asmp2 rule to the definition of the
≤a procedure (the further additional Asmp3 rule in our previous work is not
needed when dealing with orphan message-free subtyping, as we do here). Such
an Asmp2 rule is crucial to guarantee the termination of the ≤t algorithm: the
structure of the Asmp2 rule and the related proof of algorithm correctness and
termination constitute the main contribution of this paper (as far as decidability
results are concerned). In particular, since here multiple choices are admitted
for input branchings, the termination condition Asmp2 needs to deal with input
accumulation in the form of a tree, instead of simple linear accumulation as in
our previous work (see the discussion at the beginning Section 3.3 for details
about this comparison). As a consequence Asmp2 is completely modified: it has
to deal with complex recurrent patterns to be checked on the leaves of trees representing input branchings (with multiple choices), instead of detecting simple
repetitions on strings representing sequences of single-choice inputs. A detailed
comparison follows. Concerning the subtyping procedure ≤a , the one considered in this paper (Figure 2) is novel in just two details: an additional orphan
message-free condition is considered in the Out rule (because here, we consider
orphan message-free subtyping) and the usage of the outUnf unfolding instead
of unfoldn (so to perform the minimal needed per-branch unfolding). The latter
is needed in this paper because, when the procedure is turned into a subtyping
algorithm ≤t , we have to deal, in the new Asmp2 termination condition, with
42
input trees instead of strings. Correctness of ≤a w.r.t. asynchronous subtyping is stated by Proposition 3.4 in this paper, which corresponds to Lemma
5.1 of our previous work [10]. Differently from the proof of that Lemma, here
we need to cope with the different way of performing unfolding of right-hand
terms in the asynchronous subtyping definition (via unfoldn ) and in ≤a (via
outUnf). Termination of the ≤t algorithm is stated by Theorem 3.11 in this
paper, which corresponds to Lemma 5.2 of our previous work [10]. Here due to
the new structure of the Asmp2 rule, the proof is based on a much more complex
characterization of right-hand types produced, starting from the initial Z one,
by the subtyping algorithm. Such types are characterized as antOut(S, γ) with
γ being a sequence of output labels and S being a type such that S ∈ reach(Z)
and antOutInf(S). For the proof in this paper it is thus crucial to show, that
reach(Z) is finite and antOutInf(S) is decidable, as stated by Proposition 3.10,
which has no counterpart in our previous work. Correctness of the ≤t algorithm
w.r.t. ≤a procedure is stated by Theorem 3.12 in this paper, which corresponds
to Lemma 5.3 of our previous work. In both papers correctness follows from the
following property: whenever the termination rule Asmp2 is applied by the ≤t
algorithm, the subtyping procedure ≤a is guaranteed to proceed forever. However, being Asmp2 completely different, here we need to perform a totally new
proof. In particular, in our previous work we simply had to prove the existence
of a cycle causing the initial input string in the right-hand type to get longer
in a repetitive way, here, instead, we have to manage the more complex case
of input trees: this entails considering all possible branchings in order to prove
that the subtyping procedure proceeds forever.
Concerning undecidability results, both the approaches taken in the two
papers are based on resorting to undecidability of queue machines via a suitable
encoding of them into subtyping. In particular, for showing undecidability of
bounded asynchronous subtyping (≤bound ) we resort to the same encoding as in
our previous work, but we consider a more complex property of queue machines
that we show to be undecidable: bounded non termination. On the other hand,
for showing undecidability of asynchronous subtyping without output covariance
43
and input contravariance (≤tin,tout ), we need to introduce the novel class of single
consuming queue machines and to perform a much more complex encoding
which, differently from the previous one, uses nondeterminism (in order to avoid
usage of covariance and contravariance, spurious nondeterministic paths that
proceed forever have to be added by the encoding). The undecidability proof
thus becomes significantly more complex in that the above encoding makes it
much more intricated to relate the queue machine behaviour with the subtyping
game and the novel class of queue machines. Moreover, introducing a new
restricted class of queue machines requires showing them to be Turing-equivalent
(by providing an encoding of standard queue machines into them).
6. Conclusion
In this article, we have shed light on the boundaries between decidability
and undecidability of asynchronous session subtyping by analyzing two kinds
of restrictions: to the branching/selection structure of inputs/outputs and to
the capabilities of the communication buffer. In particular, considering all the
relations in Figure 1, we have shown: decidability for those in the lower part,
notably of k-bounded subtyping and of subtyping over single-out or single-in
session types; and the undecidability for those in the upper part, notably of
bounded subtyping and of subtyping without output covariance and input contravariance.
As future work, we plan to develop typing systems for server/client code
in the context of web services, exploiting our subtyping algorithms for singleout/single-in session types. Note that, in practice, server code typically connects,
as a client, to other services (e.g. a database server) using another binary session,
according to the commonly used multitier architecture. Thus, in general, when
typing code, we would use for a specific session one of the two algorithms above
depending if the code is playing the role of the client or of the server in that
session.
Moreover, we plan to investigate whether other kinds of restriction w.r.t.
44
the two above allow us to obtain a decidable relation (thus retaining general
branching/selection structure for both inputs and outputs and not limiting communication buffers).
References
References
[1] K. Honda, V. T. Vasconcelos, M. Kubo, Language primitives and type discipline for structured communication-based programming, in: 7th European
Symposium on Programming (ESOP’98), Vol. 1381 of LNCS, Springer,
1998, pp. 122–138. doi:10.1007/BFb0053567.
[2] K. Honda, N. Yoshida, M. Carbone, Multiparty asynchronous session types,
in: Proceedings of the 35th ACM SIGPLAN-SIGACT Symposium on
Principles of Programming Languages (POPL 2008), 2008, pp. 273–284.
doi:10.1145/1328438.1328472.
[3] S. Lindley, J. G. Morris, Embedding session types in Haskell, in: Proceedings of the 9th International Symposium on Haskell (Haskell 2016), 2016,
pp. 133–145. doi:10.1145/2976002.2976018.
[4] N. Ng, N. Yoshida, Static deadlock detection for concurrent Go by
global session graph synthesis, in: Proceedings of the 25th International
Conference on Compiler Construction (CC 2016), 2016, pp. 174–184.
doi:10.1145/2892208.2892232.
[5] T. B. L. Jespersen, P. Munksgaard, K. F. Larsen, Session types
for Rust,
in:
Proceedings of the 11th ACM SIGPLAN Work-
shop on Generic Programming, WGP@ICFP 2015, 2015, pp. 13–22.
doi:10.1145/2808098.2808100.
[6] K. Honda, N. Yoshida, M. Carbone, Multiparty asynchronous session types,
J. ACM 63 (1) (2016) 9. doi:10.1145/2827695.
45
[7] S. J. Gay, M. Hole, Subtyping for session types in the pi calculus, Acta Inf.
42 (2-3) (2005) 191–225. doi:10.1007/s00236-005-0177-z.
[8] T. Chen, M. Dezani-Ciancaglini, N. Yoshida, On the preciseness of subtyping in session types, in: 16th International Symposium on Principles and
Practice of Declarative Programming (PPDP’14), ACM, 2014, pp. 135–146.
doi:10.1145/2643135.2643138.
[9] S. J. Gay, Subtyping supports safe session substitution, in: A List of
Successes That Can Change the World - Essays Dedicated to Philip
Wadler on the Occasion of His 60th Birthday, 2016, pp. 95–108.
doi:10.1007/978-3-319-30936-1_5.
[10] M.
Bravetti,
M.
Carbone,
Undecidability of asynchronous session subtyping,
G.
Inf.
Zavattaro,
Comput.
To
appear.
URL http://arxiv.org/abs/1611.05026
[11] J. Lange, N. Yoshida, On the undecidability of asynchronous session
subtyping, in: Foundations of Software Science and Computation Structures - 20th International Conference, FOSSACS 2017, 2017, pp. 441–457.
doi:10.1007/978-3-662-54458-7_26.
[12] D. Mostrous, N. Yoshida, Session typing and asynchronous subtyping for the higher-order π-calculus, Inf. Comput. 241 (2015) 227–263.
doi:10.1016/j.ic.2015.02.002.
[13] E.
Christensen,
F.
Curbera,
G.
Meredith,
S.
Weerawarana,
Web Services Description Language (WSDL), Tech. rep., W3C (2001).
URL https://www.w3.org/TR/wsdl
[14] D. Mostrous, N. Yoshida, K. Honda, Global principal typing in partially
commutative asynchronous sessions, in: 18th European Symposium on Programming (ESOP’09), Vol. 5502 of LNCS, Springer, 2009, pp. 316–332.
doi:10.1007/978-3-642-00590-9_23.
46
[15] D. Kozen, Automata and computability, Springer, New York, 1997.
[16] G.
Cécé,
duplex
A.
Finkel,
communication,
Verification
Inf.
Comput.
doi:10.1016/j.ic.2005.05.006.
47
of
202
programs
(2)
with
(2005)
half-
166–190.
Appendix A. Proofs of Section 2
Appendix A.1. Proof of Theorem 2.7 and Propositions 2.9 and 2.10
We start by proving Theorem 2.7. This is done by separately showing, as
preliminary lemmas, both implications (one in each direction) to hold. The
proof of such lemmas will be then also exploited to prove Propositions 2.9 and
2.10.
Lemma Appendix A.1. Given two session types T and S, we have that
T ≤DC S implies T ≤S.
Proof. Given an asynchronous dual closed subtyping relation R we show that
R is also a (orphan-message-free) subtyping relation. To this aim we need to
prove that if (T, S) ∈ R and T = ⊕{li : Ti }i∈I then the additional item in 2. of
Definition 2.4 holds, i.e.
• if A 6= [ ]1 then ∀i ∈ I.& ∈ Ti
From A 6= [ ]1 it follows that S, after some possible unfoldings, starts with an
input (it must be in the form A[Sk ]k∈{1,...,m} ). As R is an asynchronous dual
closed subtyping relation we have (S, T ) ∈ R. We observe that S, after some
possible unfoldings, starts with an output and T = &{li : T i }i∈I . For item 2. of
Definition 2.6, we have that T = A′ [⊕{lj : Vk j }j∈Jk ]k∈{1,...,m} , for some input
context A′ . This means that all T i contain at least an output selection, which
implies that all Ti contain at least one input branching.
The following lemma states that T ≤S implies T ≤DC S. This is proved by
showing that given T ≤S we have also T ≤DC S because there exists an asynchronous dual closed subtyping relation R s.t. (T, S) ∈ R. Such relation is defined as follows: R = {(T, S), (S, T ) | T ≤S}. The proof that each pair (T, S) ∈
R satisfies the items in Definition 2.6 has only one complex case, namely, the one
in which we assume S≤T , T = ⊕{li : Ti }i∈I and S = µt1 . . . . µtn .&{lj : Sj }j∈J
(case numbered 2b in the proof). In this case we have to reason on all initial
output paths of S and use the no orphan message constraint of Definition 2.4
48
to be sure that such paths cannot be infinite, i.e. they eventually end in a state
performing an input choice, and moreover each of these reachable input choices
must match with the initial input choice of T .
Lemma Appendix A.2. Given two session types T and S, we have that T ≤S
implies T ≤DC S.
Proof. We show that, given T ≤S, it is possible to define an asynchronous dual
closed subtyping relation R s.t. (T, S) ∈ R. Consider
R = {(T, S), (S, T ) | T ≤S}
The relation R is dual closed by definition. It remains to show that it satisfies
the four items in Definition 2.6. Let (T, S) ∈ R. There are two cases: T ≤S or
S≤T . In the first case all the item holds by definition of orphan-message-free
subtyping relation. We consider now the second case, i.e. S≤T , and proceeds
with a case analysis.
1. T = end.
We have T = end. Having S≤end, by definition of ≤, in particular by
n applications of item 4. (with n ≥ 0) and one application of item 1., it
follows that S = µt1 . . . . µtn .end. Hence S = µt1 . . . . µtn .end, then we
can conclude what requested, i.e., ∃n ≥ 0 such that unfoldn (S) = end.
2. T = ⊕{li : Ti }i∈I .
We have T = &{li : T i }i∈I . Having S≤&{li : T i }i∈I , by definition of ≤,
we have two possible cases.
(a) By n applications of item 4. (with n ≥ 0) and one application of
item 3., it follows that S = µt1 . . . . µtn .&{lj : S j }j∈J , with I ⊆ J
and unfoldn (S) = &{lj : S ′ j }j∈J with S ′ i ≤T i for every i ∈ I. Hence
S = µt1 . . . . µtn . ⊕ {lj : Sj }j∈J , then we can conclude what requested,
i.e., unfoldn (S) = [⊕{lj : S ′ j }j∈J ]1 , I ⊆ J and ∀i ∈ I.(Ti , Si′ ) ∈ R.
Notice that we have used the fact that unfoldn (S) = unfoldn (S) and
we have considered an input context A = []1 .
49
(b) By n applications of item 4. (with n ≥ 0) and one application of item
2., it follows that T = &{li : T i }i∈I = A[⊕{lp : Tk p }p∈Jk ]k∈{1,...,m}
(hence with A 6= []1 ), and S = µt1 . . . . µtn . ⊕ {lj : S j }j∈J , with
∀k ∈ {1, . . . , m}.J ⊆ Jk and unfoldn (S) = ⊕{lj : S ′ j }j∈J with
∀j ∈ J.S ′ j ≤A[Tkj ]k∈{1,...,m} . Hence S = µt1 . . . . µtn .&{lj : Sj }j∈J .
We now observe that there exists an input context A′ and n′ , m′
′
′
such that unfoldn (S) = A′ [⊕{lh : Skh }h∈Lk ]k∈{1,...,m } with ∀k ∈
{1, . . . , m′ }.I ⊆ Lk . This follows from the fact that S≤&{li : T i }i∈I :
by repeated application of the rule 2. of Definition 2.4 (that includes the no orphan message constraint), we have the guarantee
that along all branches of S (and its unfoldings) it is guaranteed
to reach an input branching, and by application of rule 3. (in particular the contra-variance on input branchings), the labels of such
choices include the set of labels of the initial input branching of
&{li : T i }i∈I .
We conclude by showing that what is requested,
′
i.e., ∀i ∈ I.(Ti , A′ [Ski ]k∈{1,...,m } ) ∈ R, actually holds. This follows from the fact that A′ [Ski ]k∈{1,...,m′ } ≤Ti , which is a consequence
′
of S≤T . In fact, this implies that also unfoldn (S)≤T because an
orphan-message-free subtyping relation is still such even if we add
pairs (unfoldr (V ), Z) assuming (V, Z) already in the relation. Hav′
ing unfoldn (S) = unfoldn (S) = A′ [&{lh : Skh }h∈Lk ]k∈{1,...,m } and
′
′
T = &{li : T i }i∈I , it is easy to see that, given an orphan-message-free
′
subtyping relation R′ such that (A′ [&{lh : Skh }h∈Lk ]k∈{1,...,m } , &{li :
T i }i∈I ) ∈ R′ , the relation obtained by enriching R′ with the pairs
′
(A′′ [Ski ]k∈K⊆{1,...,m }, T ′ i ), where A′′ and types Ti′ , with i ∈ I, are
′
such that (A′′ [&{lh : Sk h }h∈Lk ]k∈K⊆{1,...,m }, &{li : T ′ i }i∈I ) ∈ R′ ,
is still an orphan-message-free subtyping relation. Above we adopt
an abuse of notation for input contexts: B[Wk ]k∈K⊆{1,...,t} does not
have holes numbered consistently from 1 to t, but some numbers in
{1, . . . , t} could be missing.
3. T = &{li : Ti }i∈I .
50
We have T = ⊕{li : T i }i∈I . Having S≤ ⊕ {li : T i }i∈I , by definition of ≤,
in particular by n applications of item 4. (with n ≥ 0) and one application
of item 2., it follows that S = µt1 . . . . µtn . ⊕ {lj : S j }j∈J , with J ⊆ I,
and unfoldn (S) = ⊕{lj : S ′ j }j∈J with S ′ j ≤T j for every j ∈ J. Hence
S = µt1 . . . . µtn .&{lj : Sj }j∈J , then we can conclude what requested, i.e.,
unfoldn (S) = &{lj : S ′ j }j∈J , J ⊆ I and ∀j ∈ J.(Tj , Sj′ ) ∈ R. Notice that
we have used the fact that unfoldn (S) = unfoldn (S).
4. T = µt.T ′ .
We first observe that V ≤µt.Z implies V ≤Z{µt.Z/t}. This directly follows
from the fact that if (V, µt.Z) belongs to an orphan-message-free subtyping relation, then the same relation enriched with the pair (V, Z{µt.Z/t})
is still an orphan-message-free subtyping relation. We now proceed by
considering T = µt.T ′ .
As S≤T , we have S≤µt.T ′ .
By the above
observation we have S≤T ′ {µt.T ′ /t} that implies what requested, i.e.,
(T ′ {µt.T ′ /t}, S) ∈ R.
Theorem 2.7. Given two session types T and S, we have T ≤S if and only if
T ≤DC S.
Proof. Direct consequence of Lemmas Appendix A.1 and Appendix A.2.
Proposition 2.9.
The ≤tin,tout relation and the ≤sin ∪≤sout relation are asyn-
chronous dual closed subtyping relations.
Proof. We first show that ≤tin,tout is an asynchronous dual closed subtyping relation. We consider R = {(S, T ) | T ≤tin,tout S} and show that it is an asynchronous
subtyping relation when in Definition 2.4 we require I = Jk in item 2. and I = J
in item 3. to hold. This implies {(S, T ) | T ≤tin,tout S} ⊆ ≤tin,tout , thus showing
that ≤tin,tout is dual closed. Given (T, S) ∈ R, we show that S≤tin,tout T implies
items 1.-4. of Definition 2.4 (where we require I = Jk in item 2. and I = J in
item 3.), apart from the no orphan message constraint of item 2., by case analysis on the structure of type T exactly as in the proof of Lemma Appendix A.2
(where ≤tin,tout is considered instead of ≤ and all subset inclusions related to
covariance/contravariance are replaced by subset equalities). Concerning the
51
no orphan message constraint of item 2., in the case 2.a of the proof of Lemma
Appendix A.2 just an [ ]1 input context arises (so it obviously holds); in the case
2.b, instead, a generic input context A′ arises: if A′ 6= [ ]1 then this means that
S, after some possible unfoldings, starts with an output and the constraint is an
immediate consequence of the fact that S≤tin,tout T (as in the proof of Lemma
Appendix A.1).
We now show that ≤sin ∪ ≤sout is an asynchronous dual closed subtyping
relation. We use T in and T out to denote the set of single-in and single-out session
types, respectively. We have ≤sin ∪ ≤sout = (≤ ∩ T in × T in ) ∪ (≤ ∩ T out × T out ) =
(≤DC ∩ T in × T in ) ∪ (≤DC ∩ T out × T out ), due to Theorem 2.7. We now show that,
for any (T, S) ∈ ≤sin ∪ ≤sout , all constraints considered by Definition 2.6 hold.
We take (T, S) ∈ ≤DC ∩ T in × T in , the other case (T, S) ∈ ≤DC ∩ T out × T out
is dealt with symmetrically. Since (T, S) ∈ ≤DC we have that (T, S) satisfies
all constraints in items 1.-4. of Definition 2.6: we just have to additionally
observe that, since all reached pairs belong to ≤DC , they also obviously belong
to ≤DC ∩ T in × T in . Concerning the duality constraint, from (T, S) ∈ ≤DC , we
have (S, T ) ∈ ≤DC , hence (S, T ) ∈ (≤DC ∩ T out × T out ).
Proposition 2.10.
The ≤sin,tout and ≤tin,sout relations are such that: T ≤sin,tout S
if and only if S ≤tin,sout T .
Proof. Concerning the only if part, we show {(S, T ) | T ≤sin,tout S} ⊆ ≤tin,sout as
follows. We consider R = {(S, T ) | T ≤sin,tout S} and show that it is an asynchronous subtyping relation when in Definition 2.4 we require I = J in item 3.
to hold and related types to be both single-out. Given (S, T ) ∈ R, we obviously
have that S and T are both single-out and we show that T ≤sin,tout S implies
items 1.-4. of Definition 2.4 (where in item 3. we require I = J) as in the proof
of Proposition 2.9. The only difference is that, when resorting to the case analysis in the proof of Lemma Appendix A.2 we consider ≤sin,tout instead of ≤ and
we replace all subset inclusions related to covariance/contravariance in item 3.
and the subset inclusion J ⊆ Jk in item 2. by equalities.
Concerning the if part, we show {(S, T ) | T ≤tin,sout S} ⊆ ≤sin,tout in a completely symmetric way by observing that R = {(S, T ) | T ≤tin,sout S} is an asyn52
chronous subtyping relation when in Definition 2.4 we require I = Jk in item 2.
to hold and related types to be both single-in. In this case, when resorting to the
case analysis in the proof of Lemma Appendix A.2 we consider ≤tin,sout instead
of ≤ and we replace all subset inclusions related to covariance/contravariance
in item 2., apart from J ⊆ Jk , by equalities.
Appendix B. Proofs of Section 3
Appendix B.1. Proof of Proposition 3.4
Proposition 3.4 states that the procedure defined in Figure 2 is a semialgorithm for checking whether T ≤
6 S. The proof of the proposition is divided
in two parts. The first one shows that if it is not possible to reach a judgement in which no rule can be applied, i.e. there exist no Σ′ , T ′ , S ′ such that
∅ ⊢ T ≤a S →∗ Σ′ ⊢ T ′ ≤a S ′ →err , then it is possible to define an asynchronous
subtyping relation R such that (T, S) ∈ R. The second part shows that if T ≤S
then the procedure either continues indefinitely or terminates successfully by
application of the Asmp or End rules.
Proposition 3.4.
Given the types T and S, we have that there exist Σ′ , T ′ , S ′
such that T ≤
6 S if and only if ∅ ⊢ T ≤a S →∗ Σ′ ⊢ T ′ ≤a S ′ →err .
Proof. We prove the two implications separately.
We start with the only if part and proceed by contraposition, i.e, we assume
that it is not true that ∃Σ′ , T ′ , S ′ . ∅ ⊢ T ≤a S →∗ Σ′ ⊢ T ′ ≤a S ′ →err and show
that T ≤S.
In order to do this we need to perform a preliminary observation: under
the assumption that it is not true that ∃Σ′ , T ′ , S ′ . ∅ ⊢ T ≤a S →∗ Σ′ ⊢
T ′ ≤a S ′ →err , even if we remove rule Asmp from the procedure it is still impossible to reach a judgement Σ′ ⊢ T ′ ≤a S ′ →err , i.e. a judgement on which no rule
can be applied. This can be showed as follows. Let →noAsmp be our decision
procedure under the assumption that Asmp is not used. Observe that the set of
pairs Σ in the judgements is irrelevant for the decision procedure →noAsmp because Asmp is the unique rule influenced by it. Now, by contraposition, assume
53
∅ ⊢ T ≤a S →∗noAsmp Σ′ ⊢ T ′ ≤a S ′ →err . Since the standard procedure →∗
cannot reach →err , we must have that there exists an intermediary judgement
Σ′′ ⊢ T ′′ ≤a S ′′ where Asmp is applied. That is, there exists Σ′′ ⊢ T ′′ ≤a S ′′ such
that ∅ ⊢ T ≤a S →∗ Σ′′ ⊢ T ′′ ≤a S ′′ (notice the use of the standard procedure),
(T ′′ , S ′′ ) ∈ Σ′′ and Σ′′ ⊢ T ′′ ≤a S ′′ →∗noAsmp Σ′ ⊢ T ′ ≤a S ′ . Within the sequence of rule applications Σ′′ ⊢ T ′′ ≤a S ′′ →∗noAsmp Σ′ ⊢ T ′ ≤a S ′ we consider
the last judgement Σ′′′ ⊢ T ′′′ ≤a S ′′′ such that (T ′′′ , S ′′′ ) ∈ Σ′′ (such judgement
exists as the first one Σ′′ ⊢ T ′′ ≤a S ′′ already has this property). It is not restrictive to assume that in the sequence Σ′′′ ⊢ T ′′′ ≤a S ′′′ →∗noAsmp Σ′ ⊢ T ′ ≤a S ′
there are no two judgements Σ1 ⊢ T1 ≤a S1 and Σ2 ⊢ T2 ≤a S2 with T1 = T2
and S1 = S2 (otherwise we can shorten the sequence Σ′′′ ⊢ T ′′′ ≤a S ′′′ →∗noAsmp
Σ′ ⊢ T ′ ≤a S ′ by removing the steps between the judgements Σ1 ⊢ T1 ≤a S1
and Σ2 ⊢ T1 ≤a S1 ,6 obtaining a new one having the same properties but without the second judgement Σ2 ⊢ T1 ≤a S1 ). Consider now, in the standard
application of the procedure ∅ ⊢ T ≤a S →∗ Σ′′ ⊢ T ′′ ≤a S ′′ , the intermediary judgement Σi ⊢ T ′′′ ≤a S ′′′ that added (T ′′′ , S ′′′ ) to the environment. We
have that (T ′′′ , S ′′′ ) 6∈ Σi (otherwise rule Asmp was applied that does not introduce any new pair in Σi ) and moreover Σi ⊂ Σ′′ as the sets of type pairs
grow monotonically during the execution of the decision procedure →∗ . These
last observations guarantee that there exists a standard application of the procedure ∅ ⊢ T ≤a S →∗ Σi ⊢ T ′′′ ≤a S ′′′ →∗ Σ′i ⊢ T ′′′′ ≤a S ′′′′ →err simply
by considering from Σi ⊢ T ′′′ ≤a S ′′′ the same rules used in the sequence
Σ′′′ ⊢ T ′′′ ≤a S ′′′ →∗noAsmp Σ′ ⊢ T ′ ≤a S ′ . This holds because the pairs of types
in the judgements traversed by this sequence are not in Σ′′ (due to the assumption on the judgement Σ′′′ ⊢ T ′′′ ≤a S ′′′ ) hence also not in Σi , and moreover
such pairs are all distinct (guaranteed by the assumption on the absence of two
judgements Σ1 ⊢ T1 ≤a S1 and Σ2 ⊢ T2 ≤a S2 with T1 = T2 and S1 = S2 ).
Consider now the relation R = {(T ′ , S ′ ) | ∃Σ′ . Σ′ ⊢ T ′ ≤a S ′ ∈ S} where S
6 The
sequence of rules applied to Σ2 ⊢ T1 ≤a S1 can be applied also to Σ1 ⊢ T1 ≤a S1
because, as already observed, →noAsmp is not sensitive to the differences between Σ1 and Σ2 .
54
is the minimal set of judgements satisfying the following:
• ∅ ⊢ T ≤a S ∈ S;
• if Σ′ ⊢ T ′ ≤a S ′ ∈ S and Σ′ ⊢ T ′ ≤a S ′ → Σ′′ ⊢ T ′′ ≤a S ′′ , without
applying rule Asmp or RecR2 , then Σ′′ ⊢ T ′′ ≤a S ′′ ∈ S;
• if Σ′ ⊢ T ′ ≤a S ′ ∈ S and Σ′ ⊢ T ′ ≤a S ′ → Σ′′ ⊢ T ′′ ≤a S ′′ by applying
′
RecR2 , then Σ′′ ⊢ T ′′ ≤a unfoldoutDepth(S ) (S ′ ) ∈ S.
We observe that to each judgement Σ′ ⊢ T ′ ≤a S ′ ∈ S it is always possible
to apply at least one rule. In fact, if this is not possible, we would have also
∅ ⊢ T ≤a S →∗noAsmp Σ′′ ⊢ T ′′ ≤a S ′′ →err for a judgement Σ′′ ⊢ T ′′ ≤a S ′′ with
T ′′ = T ′ and S ′′ less unfolded than S ′ . In fact, the unique difference between
the judgements in S and those reachable without adopting Asmp is that those
in S are more unfolded (see the difference between outUnf(S) used in rule RecR2
′
and unfoldoutDepth(S ) (S ′ ) used in the definition of S).
We finally show that R is an (orphan-message-free) subtyping relation according to Definition 2.4. Let (T ′ , S ′ ) ∈ R. Then Σ′ ⊢ T ′ ≤a S ′ ∈ S and it is
possible to apply at least one rule to Σ′ ⊢ T ′ ≤a S ′ . We proceed by cases on T ′ .
• If T ′ = end then item 1. of Definition 2.4 for pair (T ′ , S ′ ) is shown by
induction on k = nrec(S ′ ), i.e. the number of unguarded (not prefixed by
some input or output) occurrences of recursions µt.S ′′ in S ′ for any S ′′ , t.
– Base case k = 0. The only rule applicable to Σ′ ⊢ T ′ ≤a S ′ is End,
that immediately yields the desired pair of R.
– Induction case k > 0. The only rules applicable to Σ′ ⊢ T ′ ≤a S ′
are Asmp and RecR1 . In the case of Asmp we have that (T ′ , S ′ ) ∈ Σ′ ,
hence there exists Σ′′ with (T ′ , S ′ ) ∈
/ Σ′′ such that Σ′′ ⊢ T ′ ≤a S ′ ∈ S.
RecR1 can be applied to Σ′′ ⊢ T ′ ≤a S ′ . So for some Σ′′′ (= Σ′
or = Σ′′ ) we have that the procedure applies rule RecR1 to Σ′′′ ⊢
T ′ ≤a S ′ . Hence Σ′′′ ⊢ T ′ ≤a S ′ → Σ′′′′ ⊢ T ′ ≤a unfold1 (S ′ ). Since
55
nrec(unfold1 (S ′ )) = k − 1, by induction hypothesis item 1. of Definition 2.4 holds for pair (T ′ , unfold1 (S ′ )), hence it holds for pair
(T ′ , S ′ ).
• If T ′ = ⊕{li : Ti }i∈I then item 2. of Definition 2.4 for pair (T ′ , S ′ ) is shown
as follows.
– If outDepth(S ′ ) = 0 then the only rule applicable to Σ′ ⊢ T ′ ≤a S ′ is
Out, that immediately yields the desired pairs of R.
– If outDepth(S ′ ) ≥ 1 then the only rules applicable to Σ′ ⊢ T ′ ≤a S ′
are Asmp and RecR2 . In the case of Asmp we have that (T ′ , S ′ ) ∈ Σ′ ,
hence there exists Σ′′ with (T ′ , S ′ ) ∈
/ Σ′′ such that Σ′′ ⊢ T ′ ≤a S ′ ∈
S. RecR2 can be applied to Σ′′ ⊢ T ′ ≤a S ′ . So for some Σ′′′
(= Σ′ or = Σ′′ ) we have that the procedure applies rule RecR2
′
to Σ′′′ ⊢ T ′ ≤a S ′ . Hence (T ′ , unfoldoutDepth(S ) (S ′ )) ∈ R. Since
′
outDepth(unfoldoutDepth(S ) (S ′ )) = 0, we end up in the previous case.
′
Therefore item 2. of Definition 2.4 holds for pair (T ′, unfoldoutDepth(S )(S ′ )),
hence it holds for pair (T ′ , S ′ ).
• If T ′ = &{li : Ti }i∈I then item 3. of Definition 2.4 for pair (T ′ , S ′ ) is shown
by induction on k = nrec(S ′ ).
– Base case k = 0. The only rule applicable to Σ′ ⊢ T ′ ≤a S ′ is In, that
immediately yields the desired pairs of R.
– Induction case k > 0. The only rules applicable to Σ′ ⊢ T ′ ≤a S ′
are Asmp and RecR1 . In the case of Asmp we have that (T ′ , S ′ ) ∈ Σ′ ,
hence there exists Σ′′ with (T ′ , S ′ ) ∈
/ Σ′′ such that Σ′′ ⊢ T ′ ≤a S ′ ∈ S.
RecR1 can be applied to Σ′′ ⊢ T ′ ≤a S ′ . So for some Σ′′′ (= Σ′
or = Σ′′ ) we have that the procedure applies rule RecR1 to Σ′′′ ⊢
T ′ ≤a S ′ . Hence Σ′′′ ⊢ T ′ ≤a S ′ → Σ′′′′ ⊢ T ′ ≤a unfold1 (S ′ ). Since
nrec(unfold1 (S ′ )) = k − 1, by induction hypothesis item 3. of Definition 2.4 holds for pair (T ′ , unfold1 (S ′ )), hence it holds for pair
(T ′ , S ′ ).
56
• If T ′ = µt.T ′ then item 4. of Definition 2.4 for pair (T ′ , S ′ ) holds because
the only rule applicable to Σ′ ⊢ T ′ ≤a S ′ is RecL that immediately yields
the desired pair of R.
We now prove the if part and proceed by contraposition. We assume that
T ≤S and show that there exist no Σ′ , T ′ , S ′ , such that ∅ ⊢ T ≤a S →∗ Σ′ ⊢
T ′ ≤a S ′ →err . So we can assume the existence of a relation R that is an
(orphan-message-free) subtyping relation, according to Definition 2.4, such that
(T, S) ∈ R.
We say that Σ ⊢ T ≤a S →w Σ′ ⊢ T ′ ≤a S ′ if Σ ⊢ T ≤a S →∗ Σ′ ⊢ T ′ ≤a S ′
and: the last rule applied is one of Out, In or RecL rules; while all previous ones
are RecR1 or RecR2 rules. As another notation we use input-output-end contexts
B defined as the input contexts in Definition 2.3 with the difference that also
the output construct and end are part of the grammar in the definition.
We start by showing that ∃Σ. ∅ ⊢ T ≤a S →∗w Σ ⊢ T ′ ≤a S ′ implies
S ′ = B[Sk ]k∈{1...m} , Sk = µtk .Sk′ , for some tk and Sk′ , and ∃n1 , . . . , nm . (T ′ ,
B[unfoldnk (Sk )]k∈{1...m} ) ∈ R. The proof is by induction on the length of
such computation →∗w . The base case is for a 0 length computation: it yields
(T, S) ∈ R which holds.
For the inductive case we assume it to hold for
all computations of a length k and we show it to holds for all computations
of length k + 1, by considering all judgements Σ′ ⊢ T ′′ ≤a S ′′ such that
Σ ⊢ T ′ ≤a S ′ →w Σ′ ⊢ T ′′ ≤a S ′′ . This is shown by first considering the
case in which rule Asmp applies to Σ ⊢ T ′ ≤a S ′ : in this case there is no such a
judgement and there is nothing to prove. Then we consider the case in which
T ′ = end and Σ ⊢ end ≤a S ′ →∗ Σ′′′ ⊢ end ≤a end (by applying RecR1 rules)
and rule End applies to Σ′′′ ⊢ end ≤a end. Also in this case there is no such
a judgement Σ′ ⊢ T ′′ ≤a S ′′ and there is nothing to prove. Finally, we proceed
by an immediate verification that judgements Σ′ ⊢ T ′′ ≤a S ′′ produced in remaining cases are required to be in R by items 2., 3. and 4. of Definition 2.4:
T ′ = ⊕{li : Ti }i∈I (→w is a possibly empty sequence of RecR2 applications followed by Out application), T ′ = &{li : Ti }i∈I (→w is a possibly empty sequence
57
of RecR1 applications followed by In application) or T ′ = µt.T ′ (→w is simply
RecL application).
We finally observe that, given a judgement Σ ⊢ T ′ ≤a S ′ such that S ′ =
B[Sk ]k∈{1...m} , Sk = µtk .Sk′ , for some tk and Sk′ , and ∃ n1 , . . . , nm . (T ′ ,
B[unfoldnk (Sk )]k∈{1...m} ) ∈ R we have:
• either rule Asmp applies to Σ ⊢ T ′ ≤a S ′ , or
• T ′ = end and, by item 1. of Definition 2.4, there exists Σ′ such that
Σ ⊢ end ≤a S ′ →∗ Σ′ ⊢ end ≤a end (by applying RecR1 rules) and rule
End is the unique rule applicable to Σ′ ⊢ end ≤a end, with RecR1 being
the unique rule applicable to intermediate judgements, or
• by items 2., 3. and 4. of Definition 2.4, there exist Σ′ , T ′′ , S ′′ such that
Σ ⊢ T ′ ≤a S ′ →∗w Σ′ ⊢ T ′′ ≤a S ′′ , with each intermediate judgement
having a unique applicable rule. In particular this holds for T ′ = ⊕{li :
Ti }i∈I (→w is a possibly empty sequence of RecR2 applications followed
by Out application), T ′ = &{li : Ti }i∈I (→w is a possibly empty sequence
of RecR1 applications followed by In application) or T ′ = µt.T ′ (→w is
simply RecL application).
Appendix B.2. Proof of Theorem 3.5
Theorem 3.5 states that ≤ka indeed provides an algorithm for checking
whether T ≤
6 k S, this is proved in the following by also resorting to the proof of
Proposition 3.4.
Theorem 3.5.
The algorithm for ≤ka always terminates and, given the types
T and S, there exist Σ′ , T ′ , S ′ such that ∅ ⊢ T ≤ka S →∗ Σ′ ⊢ T ′ ≤a S ′ →err if
and only if T ≤
6 k S.
Proof. We first observe that the decision algorithm for k-bounded asynchronous
subtyping terminates. By contraposition, if the algorithm does not terminate,
there exists an infinite sequence Σ ⊢ T ≤a S → Σ1 ⊢ T1 ≤a S1 →∗ Σi ⊢
Ti ≤a Si →∗ . Along this infinite sequence infinitely many distinct pairs (T, S)
will be added to Σ. As only finitely many distinct terms can be reached as
58
first element of the pairs, there will be infinitely many distinct terms as second
element. Such terms will have unbounded depth, but this is not possible due
to the constraint added to rule Out that impose the use of k-bounded input
contexts.
We now prove that, given the types T and S, there exist Σ′ , T ′ , S ′ such that
∅ ⊢ T ≤ka S →∗ Σ′ ⊢ T ′ ≤a S ′ →err if and only if T ≤
6 k S.
We start with the if part and proceed by contraposition. We assume that
it is not true that ∃Σ′ , T ′ , S ′ . ∅ ⊢ T ≤ka S →∗ Σ′ ⊢ T ′ ≤ka S ′ →err and we
build a relation R that we show to be a k-bounded Asynchronous Subtyping
relation. The relation R is built from the judgments Σ′′ ⊢ T ′′ ≤ka S ′′ exactly
as we did for the ≤a subtyping procedure in the first part (the if part) of
the proof of Proposition 3.4. In such a proof we show R to be an orphanmessage-free subtyping relation, hence we just have to show it to be k-bounded.
It is immediate to observe that, since when applying rule Out to a judgment
Σ′′ ⊢ T ′′ ≤ka S ′′ we require the input context A to be k-bounded, we may
include in R only pairs (T ′′ , S ′′ ) that satisfy the same constraint in item 2 of kbounded Asynchronous Subtyping relation definition (Definition 2.11), because
otherwise we would have Σ′′ ⊢ T ′′ ≤ka S ′′ →∗ Σ′′′ ⊢ T ′′′ ≤ka S ′′′ →err by possibly
applying RecR1 /RecR2 rules. Hence, as justified in Proposition 3.4 this would
lead to violating the assumption that the algorithm does not reach an error.
The justification provided there still holds because judgments Σ′′ ⊢ T ′′ ≤ka S1′′
and Σ′′ ⊢ T ′′ ≤ka S2′′ , with S1′′ and S2′′ that just differ for the level of internal
unfoldings, behave equivalently with respect to errors due to k-boundedness
violations. This because the k-boundedness of context A is established by the
Out rule after unfolding in S1′′ /S2′′ all recursions occurring before the first output
of every possible branch by means of the RecR1 /RecR2 rules.
We now prove the only if part and proceed by contraposition. We assume
that T ≤k S and show that there exist no Σ′ , T ′ , S ′ , such that ∅ ⊢ T ≤ka S →∗
Σ′ ⊢ T ′ ≤ka S ′ →err . If T ≤k S then also T ≤S. So we can assume the existence
of a relation R that is an orphan-message-free subtyping relation such that
(T, S) ∈ R. We then use exactly the same proof as that of the second part
59
(the only if part) of the proof of Proposition 3.4 to establish a correspondance
between judgements Σ′′ ⊢ T ′′ ≤ka S ′′ , such that ∅ ⊢ T ≤ka S →∗w Σ′′ ⊢ T ′′ ≤ka S ′′ ,
and pairs in R (see the construction of the corresponding pair in the proof of
Proposition 3.4). Since R includes only pairs that satisfy the constraint in item
2 of k-bounded Asynchronous Subtyping relation definition (Definition 2.11)
requiring context A to be k-bounded; and since any judgment Σ′′ ⊢ T ′′ ≤ka S ′′
such that ∅ ⊢ T ≤ka S →∗w Σ′′ ⊢ T ′′ ≤ka S ′′ implies there is in R a corresponding
pair (T ′′ , S1′′ ), with S1′′ differing from S ′′ just for the level of internal unfoldings,
we have that reachable judgments Σ′′ ⊢ T ′′ ≤ka S ′′ cannot be such that: Σ′′ ⊢
T ′′ ≤ka S ′′ →∗ Σ′′′ ⊢ T ′′′ ≤ka S ′′′ , by possibly applying RecR1 /RecR2 rules,
and Σ′′′ ⊢ T ′′′ ≤ka S ′′′ →err due to not satisfying the requirement about the
input context A to be k-bounded in the rule Out. This because the difference
in unfolding levels between S ′′ and S1′′ (inside judgment Σ′′ ⊢ T ′′ ≤ka S ′′ and
the corresponding pair (T ′′ , S1′′ ) in R) is not significant: the k-boundedness of
context A is established both in the rule Out and in item 2 of ≤k definition
after unfolding all recursions occurring before the first output of every possible
branch.
This observation makes it possible to carry out the proof as in Proposition 3.4, hence to show that there exist no Σ′ , T ′ , S ′ , such that ∅ ⊢ T ≤ka S →∗
Σ′ ⊢ T ′ ≤ka S ′ →err .
Appendix B.3. Proof of Proposition 3.10
We start by providing the proof of the first part of Proposition 3.10, i.e. finiteness of reach(T ) for single-out session types T , as a separate preliminary lemma,
then we move to proving the second part, i.e. decidability of antOutInf(T ).
Lemma Appendix B.1. Given a single-out session type T , reach(T ) is finite.
Proof. We now define a finite set of session types fin(T ), and then we prove that
it satisfies all the constraints 1., . . . , 4. in Definition 3.9. Hence reach(T ) ⊆ fin(T )
by definition, from which finiteness of reach(T ) follows.
60
It is not restrictive to assume that all the recursion variables of T are distinct: let x1 , . . . , xn be such variables. We consider the rewriting variables
X1 , . . . , Xn . Let Ti be such that µxi .Ti occurs in T ; let T ′ be T with Xi
that replaces µxi .Ti ; and similarly let Ti′ be Ti with Xj that replaces each occurrence of µxj .Tj and xj . We now consider the rewriting rules Xi →1i Ti′
and Xi →2i xi . Given one of the above term S containing rewriting variables,
we denote with close(S) the session type obtained by repeated application of
the rewriting rules in the following way: if Xi occurs inside a subterm µxi .S ′
apply →2i , otherwise apply →1i . We now define another closure function on
sets of terms S: subterms(S) = {S ′ |S ′ is a subterm of S ∈ S}. Consider finally
fin(T ) = {close(S)|S ∈ subterms({T ′ , T1′ , . . . , Tn′ })}. We have that fin(T ) is finite
and it satisfies all the constraints 1., . . . , 4. in Definition 3.9.
We now report some definitions and preliminary results used in the proof of
Proposition 3.10 concerning the second part about decidability of antOutInf(T )
for single-out session types T . We introduce the relation antEqT : intuitively,
(T ′ , T ′′ ) ∈ antEqT if T ′ and T ′′ are terms in reach(T ) capable of anticipating
the same infinite sequence of outputs. For instance, assuming that the following
T ′ = µt. ⊕ {l : &{l1 : ⊕{l′ : t}, l2 : ⊕{l′ : t}}} and T ′′ = µt.&{l1 : ⊕{l :
⊕{l′ : t}}} belong to reach(T ), for some session type T , we have that (T ′ , T ′′ ) ∈
antEqT because they can anticipate the sequence of outputs l and l′ , indefinitely.
Definition Appendix B.2. Let T be a single-out session type. A relation
R over reach(T ) is an antEqT relation if (T ′ , T ′′ ) ∈ R implies: there exist l, A′ , A′′ such that outUnf(T ′ ) = A′ [⊕{l : Ti′ }]i∈{1,...,n} and outUnf(T ′′ ) =
A′′ [⊕{l : Tj′′ }]j∈{1,...,m} , with (Ti′ , Tj′′ ) ∈ R for all i ∈ {1, . . . , n} and j ∈ {1, . . . , m}.
We say that T ′ antEqT T ′′ if there is an antEqT relation R such that (T ′ , T ′′ ) ∈
R.
Notice that antEqT itself is an antEqT relation because, obviously, the union
of two antEqT relations is an antEqT relation and reach(T ) is finite. Moreover
notice that, given a term T ′ ∈ reach(T ), all terms Ti′ (with i ∈ {1, . . . , n}) for
which outUnf(T ′ ) = A′ [⊕{l : Ti′ }]i∈{1,...,n} are always such that Ti′ ∈ reach(T )
61
as well (because outUnf(T ′ ) never unfolds recursions occurring inside terms Ti′ ).
Finally, notice that antEqT is decidable in that it is a relation over reach(T ),
which is a finite set.
Definition Appendix B.3. antSetT is the field of antEqT , that is the set
of session types T ′ ∈ reach(T ) such that there exists T ′′ with (T ′ , T ′′ ) ∈ antEqT
or (T ′′ , T ′ ) ∈ antEqT .
Lemma Appendix B.4. antEqT is an equivalence relation on antSetT .
Proof. The reflexive, symmetric and transitive closure of an antEqT relation is
an antEqT relation, hence this holds true for antEqT as well.
Lemma Appendix B.5. Let T ′ ∈ reach(T ). We have that antOutInf(T ′ ) if
and only if T ′ ∈ antSetT .
Proof. We prove the two implications separately, starting from the if part, e.g.
by assuming T ′ ∈ antSetT . By Lemma Appendix B.4 we have T ′ antEqT T ′ .
We now prove by induction on m that for every m there exists li1 · · · lim such
that antOut(T ′ , li1 · · · lim ) is defined. If m = 1 it is sufficient to consider li1 = l
where outUnf(T ′ ) = A′ [⊕{l : Ti′ }]i∈{1,...,n} (with A′ and Ti′ that exist by Definition Appendix B.2). Consider now that T ′′ = antOut(T ′ , li1 · · · lim−1 ) is defined.
By Definition 3.8, we have T ′′ = A[Tk ]k with outUnf(antOut(T, li1 · · · lim−2 )) =
A[⊕{lim−1 : Tk }]k . As T ′ antEqT T ′ , we can apply m−1 times Definition Appendix B.2
to conclude that Ti antEqT Tj , for every i, j ∈ 1 . . . k. This guarantees the existence of the input contexts Ak , session types Trk , and label l such that such
that outUnf(Tk ) = Ak [⊕{l : Trk }]r . This implies that it is possible to define
antOut(T ′′ , l) hence also antOut(T ′ , li1 · · · lim ) by taking lim = l.
We now move to the only if part assuming that there exists an infinite label
sequence li1 · · · lin · · · such that, for every n, antOut(T ′ , li1 · · · lin ) is defined. Let
R be the minimal relation such that (T ′ , T ′ ) ∈ R and: outUnf(antOut(T ′ , li1 · · · lin−1 ))
= A[⊕{lin : Tk }]k∈{1...mn } , for any n ≥ 1, implies ∀i, j ∈ {1 . . . mn }. (Ti , Tj ) ∈ R.
We now show that R above is an antEqT relation. Considered any (T ′′ , T ′′′ ) in
62
R, we have that there exists h, with h ≥ 1, such that, for some A′ , A′′ , we have:
′
′′
outUnf(T ′′ ) = A′ [⊕{lih : Ti′ }]i∈{1,...,m } and outUnf(T ′′′ ) = A′′ [⊕{lih : Tj′′ }]j∈{1,...,m } ,
with (Ti′ , Tj′′ ) ∈ R for all i ∈ {1, . . . , m′ } and j ∈ {1, . . . , m′′ }. This holds, according to the definition of R: for (T ′′ , T ′′′ ) = (T ′ , T ′ ) by taking h = 1 and
by observing that pairs (Ti′ , Tj′′ ) ∈ R because they are added to R in the case
n = 1; for any (T ′′ , T ′′′ ) added to R in the case n, by taking h = n + 1 and by
observing that pairs (Ti′ , Tj′′ ) ∈ R because they are among the pairs that are
added to R in the case n + 1.
Proposition 3.10.
Given a single-out session type T , reach(T ) is finite and
it is decidable whether antOutInf(T ).
Proof. Direct consequence of Lemmas Appendix B.1, Lemma Appendix B.5
and the finiteness of antSetT .
Appendix B.4. Proof of Theorem 3.11
Theorem 3.11 states that, for single-out session types T and S, the algorithm
provided by T ≤t S indeed terminates. The proof is based on characterizing
terms S ′′ such that (T ′′ , S ′′ ) ∈ Σ′ for any judgment Σ′ ⊢ T ′ ≤t S ′ reached by
the algorithm. In particular, we show that any such term S ′′ can be obtained
by anticipating a sequence of output labels γ in a term R belonging to the finite
set reach(S). This preliminary result is proved by means of the following lemma
and corollary.
Lemma Appendix B.6 states that given a type S ′ at the right hand side of a
judgement reachable during the execution of our algorithm, i.e. ∅ ⊢ T ≤t S →∗
Σ′ ⊢ T ′ ≤t S ′ for some initial types T and S, the type S ′ (and the types
in reach(S ′ )) can be obtained from the initial type S (or one of the types in
reach(S)) by means of anticipation of a sequence of output labels, i.e. for all
Q ∈ reach(S ′ ) there exist R ∈ reach(S) and a sequence of labels γ such that Q =
antOut(R, γ). There is only one case in which this property is not guaranteed
to hold, namely, when the last applied rule in ∅ ⊢ T ≤t S →∗ Σ′ ⊢ T ′ ≤t S ′ is
63
RecR2 . As a counter-example, consider for instance
∅
⊢
µt. ⊕ {l : ⊕{l : &{l′ : t}}} ≤t µt.&{l′ : ⊕{l : t}} →∗
Σ′
⊢
⊕{l : &{l′ : µt. ⊕ {l : ⊕{l : &{l′ : t}}}}} ≤t
&{l′ : &{l′ : ⊕{l : µt.&{l′ : ⊕{l : t}}}}}
where →∗ denotes application of the sequence of rules RecL, RecR2 , Out and
RecR2 . We have that the last type &{l′ : &{l′ : ⊕{l : µt.&{l′ : ⊕{l : t}}}}} is
different from antOut(R, γ), for every R ∈ reach(µt.&{l′ : ⊕{l : t}}) and every
sequence of labels γ.
Lemma Appendix B.6. Consider two single-out session types T and S. Given
a judgement Σ′ ⊢ T ′ ≤t S ′ such that ∅ ⊢ T ≤t S →∗ Σ′ ⊢ T ′ ≤t S ′ , in such a
way that the final rule applied is not RecR2 , we have that for all Q ∈ reach(S ′ )
there exist R ∈ reach(S) and a sequence of labels γ such that Q = antOut(R, γ).
Proof. By induction on the length of the sequence of rule applications ∅ ⊢
T ≤t S →∗ Σ′ ⊢ T ′ ≤t S ′ . In the base case we have S ′ = S. Consider
now Q ∈ reach(S ′ ). Obviously Q = antOut(Q, ǫ) with Q ∈ reach(S) because
reach(S) = reach(S ′ ).
In the inductive case we proceed by case analysis on the last rule application
Σ′′ ⊢ T ′′ ≤t S ′′ → Σ′ ⊢ T ′ ≤t S ′ . We have two possible cases:
• We can apply the induction hypotheses on the judgement Σ′′ ⊢ T ′′ ≤t S ′′ .
Hence for all Q′′ ∈ reach(S ′′ ) there exist R ∈ reach(S) and a sequence of
labels γ such that Q′′ = antOut(R, γ). Consider now Q ∈ reach(S ′ ). We
proceed by cases on the applied rule.
For the rules In, RecR1 and Out with A = [ ]1 we have that S ′ ∈ reach(S ′′ )
hence also Q ∈ reach(S ′′ ) because if S ′ ∈ reach(S ′′ ) then reach(S ′ ) ⊆
reach(S ′′ ) by definition of reach( ).
If the rule is Out with A 6= [ ]1 we have that S ′ = antOut(R, γ · l) with
R ∈ reach(S) and γ such that S ′′ = antOut(R, γ) and l is the label of
the anticipated output. We limit our analysis to the case in which Q 6∈
64
reach(S ′′ ) (in the other cases we can proceed as above). This happens if Q
is obtained by applying rule 2. of Definition 3.9 to remove some but not all
the inputs in front of one of the output anticipated in S ′′ . Consider now
the term V being like Q but without the l output anticipation. Formally,
V is defined as follows. Denoted S ′′ with A[⊕{lj : Skj }j∈Jk ]k we know
that for all k there exists a jk such that ljk = l. This means that S ′ is
A[Skjk ]k . Hence, being Q reachable from S ′ by consuming some inputs
of the input context A only, we have that there exists A′ such that Q is
′
h
A′ [Shj
′ ] , where, considered the hole k corresponding to the hole h, we
h
′
have that jh′ = jk and Shj
′ = Skjk . Therefore, the previously mentioned
h
′
}j∈Jh ]h , where, considered k corresponding to h,
term V is A′ [⊕{lj : Shj
we have that Jh = Jk . We conclude by observing that V ∈ reach(S ′′ ),
hence there exist R′ ∈ reach(S) and γ ′ such that V = antOut(R′ , γ ′ ). But
Q = antOut(R′ , γ ′ · l), hence proving the thesis.
• We cannot apply the induction hypotheses on the judgement Σ′′ ⊢ T ′′ ≤t S ′′
because the rule used to obtain Σ′′ ⊢ T ′′ ≤t S ′′ is RecR2 . As RecR2 cannot be applied in sequence, it is surely possible to apply the induction
hypothesis on the previous judgement Σ′′′ ⊢ T ′′′ ≤t S ′′′ such that Σ′′′ ⊢
T ′′′ ≤t S ′′′ → Σ′′ ⊢ T ′′ ≤t S ′′ . Then we have that for all Q′′′ ∈ reach(S ′′′ )
we have Q′′′ = antOut(R, γ) with R ∈ reach(S) and a sequence of labels
γ. We also have that the rule applied in Σ′′ ⊢ T ′′ ≤t S ′′ → Σ′ ⊢ T ′ ≤t S ′
is Out, which is the only rule that can applied after RecR2 . Let l be the
label of the output involved in the application of the Out rule. Consider
now Q ∈ reach(S ′ ). We consider two possible cases:
– Q is obtained from S ′ by consuming inputs present in the input context A used in the last application of the rule Out. Consider now
Q′′′ obtained from S ′′′ by consuming the same inputs and performing the needed unfoldings. Obviously Q′′′ ∈ reach(S ′′′ ): hence, by
induction hypothesis, Q′′′ = antOut(R, γ) with R ∈ reach(S). We
have Q = antOut(R, γ · l) hence proving the thesis.
65
– Q is obtained from S ′ by consuming strictly more than a sequence
of inputs present in the input context A used in the last application
of the rule Out. This means that Q ∈ reach(W ) where W is a term
starting with an output that populates one of the holes of A in S ′′ .
But the terms starting with an output that can occur in S ′′ , assuming
∅ ⊢ T ≤t S →∗ Σ′′ ⊢ T ′′ ≤t S ′′ , are already in reach(S). In fact the
rules do not perform transformations under outputs, excluding those
strictly performed by top level unfoldings. Hence W ∈ reach(S),
which implies Q ∈ reach(S) from which the thesis trivially follows
(because Q = antOut(Q, ǫ)).
Corollary Appendix B.7. Consider two single-out session types T and S.
Given a judgement Σ′ ⊢ T ′ ≤t S ′ such that ∅ ⊢ T ≤t S →∗ Σ′ ⊢ T ′ ≤t S ′ and a
pair (T ′′ , S ′′ ) ∈ Σ′ , we have that S ′′ = antOut(R, γ) for some R ∈ reach(S) and
a sequence of labels γ.
Proof. Let (T ′′ , S ′′ ) ∈ Σ′ and consider the sequence of rule applications ∅ ⊢
T ≤t S →∗ Σ′′ ⊢ T ′′ ≤t S ′′ that preceeds the application of the rule that
introduces (T ′′ , S ′′ ) in Σ′ . Such rule must be one of RecL, RecR1 or RecR2 :
hence on the judgement Σ′′ ⊢ T ′′ ≤t S ′′ it is possible to apply one of these three
rules. Since after the application of a rule RecR2 the uniqe applicable rule is
Out, we have the guarantee that the last rule in ∅ ⊢ T ≤t S →∗ Σ′′ ⊢ T ′′ ≤t S ′′
is not RecR2 . Hence it is possible to apply Lemma Appendix B.6, from which
the thesis directly follows.
Theorem 3.11.
Given two single-out session types T and S, the algorithm
applied to the initial judgement ∅ ⊢ T ≤t S terminates.
Proof. Assume by contraposition that there exists single-out session types T
and S such that the algorithm applied to the initial judgement ∅ ⊢ T ≤t S
does not terminate. This means that there exists an infinite sequence of rule
applications ∅ ⊢ T ≤t S → Σ1 ⊢ T1 ≤t S1 →∗ Σi ⊢ Ti ≤t Si →∗ . Within this
infinite sequence, there are infinitely many applications of the unfolding rules
RecL, RecR1 or RecR2 , that implies the existence of infinitely many distinct pairs
66
(Tj , Sj ) that are introduced in the environment (assuming that j ranges over the
instances of application of such rules). All these pairs are distinct, otherwise
the precedence of the Asmp rule would have blocked the algorithm. It is obvious
that the distinct r.h.s. Tj are finitely many, because every Tj ∈ reach(T ), which
is a finite set. On the contrary, the distinct Sj are infinitely many, but Corollary
Appendix B.7 guarantees that for each of them, there exists Sj′ ∈ reach(S) and
a sequence of labels γj such that Sj = antOut(Sj′ , γj ).
Due to the finiteness of the possible Tj and Sj′ , there exists T ′′ and S ′′ such
that there exists an infinite subsequence of (Tj1 , Sj1 ), (Tj2 , Sj2 ), . . . , (Tjk , Sjk ), . . .
such that Tji = T ′′ and Sji = antOut(S ′′ , γji ). It is not restrictive to consider
jh < jh+1 for every h. The presence of infinitely many distinct γji for which
antOut(S ′′ , γji ) is defined, guarantees antOutInf(S ′′ ). Moreover, this guarantees
also the possibility to define an infinite subsequence (Tjl1 , Sjl1 ), (Tjl2 , Sjl2 ), . . . ,
(Tjlk , Sjlk ), . . . such that |γjli| < |γjli+1|. We now consider the leaf sets leafSet(Sjli).
These sets are defined on a finite domain because the subterms of such types
starting with a recursive definition or an output, and preceded by inputs only,
are taken from reach(S). This because the algorithm does not apply transformations under recursive definitions or outputs, excluding the effect of the
standard top level unfolding of previous recursive definitions, which is considered in the definition of reach(S). Hence there are only finitely many distinct
leafSet(Sjli ), that guarantees the existence of v < w such that leafSet(Sjlv ) =
leafSet(Sjlw ). Consider now the judgement Σjlw ⊢ Tjlw ≤t Sjlw . We know that
(Tjlv , Sjlv ) ∈ Σjw , Tjlv = Tjlw , Sjlv = antOut(S ′′ , γjlv ), Sjlw = antOut(S ′′ , γjlw ),
S ′′ ∈ reach(S), and |γjlv | < |γjlv |. Hence it is possible to apply to such judgement the rule Asmp2 . As Asmp2 has priority, it should be applied on this
judgement thus blocking the sequence of rule applications. But this contradicts
the initial assumption of non termination of the algorithm.
Appendix B.5. Proof of Theorem 3.12
The soundness Theorem 3.12 states that the ≤t algorithm reaches →err
if and only if the ≤a procedure does so. This is proved in the following by
67
resorting to some preliminary definitions and results.
In Definition Appendix B.2 we have defined the relation antEqT among
types that have the same infinite sequence of outputs that can be anticipated.
But, in order to have a decidable relation, we had to limit to types belonging to
the set reach(T ). Now, we define a more general relation extAntEqT applicable
to types having (once unfolded) the following shape: any possible input context
with holes filled with single outputs having a continuation belonging to reach(T ).
This extension of the antEqT is necessary because the execution of the subtyping
algorithm can generate new terms (as a consequence of output anticipations)
having this specific shape.
Definition Appendix B.8. Let T ′ , T ′′ be single-out session types. We say
that T ′ extAntEqT T ′′ if there exist l, A′, A′′ such that outUnf(T ′ ) = A′ [⊕{l : Ti′ }]i∈{1,...,n}
and outUnf(T ′′ ) = A′′ [⊕{l : Tj′′ }]j∈{1,...,m} , with Ti′ antEqT Tj′′ for all i ∈ {1, . . . , n}
and j ∈ {1, . . . , m}.
Moreover, extAntSetT is the field of extAntEqT .
Notice that, all terms Ti′ , with i ∈ {1, . . . , n} and Tj′′ , with j ∈ {1, . . . , m},
are in antSetT ⊆ reach(T ). Moreover, notice that extAntEqT is obviously an
equivalence relation on extAntSetT .
Lemma Appendix B.9. Let T ′ ∈ antSetT and T ′′ = antOut(T ′ , γ) for some
γ. We have that T ′′ ∈ extAntSetT .
Proof. We have to show that there exist l, A for which we have outUnf(antOut(T ′, γ))
= A[⊕{l : Ti }]i∈{1,...,m} , with Ti antEqT Tj for all i, j ∈ {1, . . . , m}. We denote γl = li1 · · · lih , with h ≥ 1. For any n, with 1 ≤ n ≤ h, considered A′
and terms Tk with k ∈ {1 . . . mn } such that outUnf(antOut(T ′ , li1 · · · lin−1 )) =
A′ [⊕{lin : Tk }]k∈{1...mn } , we have that ∀i, j ∈ {1 . . . mn }. Ti antEqT Tj . This is
easily shown by induction on n, applying the definition of antEqT (the base
case is directly derived from T ′ antEqT T ′ ). The case n = h yields the desired
result.
68
Lemma Appendix B.10. Let T ′, T ′′ ∈ extAntSetT and leafSet(T ′ ) = leafSet(T ′′ ).
We have that T ′ extAntEqT T ′′ .
Proof. It is easy to see that leafSet(T ′ ) = leafSet(T ′′ ) implies leafSet(outUnf(T ′ )) =
leafSet(outUnf(T ′′ )). This because outUnf() causes a leaf T ′′′ belonging to both
leafSet(T ′ ) and leafSet(T ′′ ) to yield the same new set of leaves leafSet(T ′′′ ) in
both T ′ and T ′′ . By definition of extAntSetT we have that exist l′ , A′ such that
outUnf(T ′ ) = A′ [⊕{l′ : Ti′ }]i∈{1,...,n} , with Ti′ antEqT Tj′ for all i, j ∈ {1, . . . , n}.
Similarly, there exist l′′ , A′′ such that outUnf(T ′′ ) = A′′ [⊕{l′′ : Tj′′ }]j∈{1,...,m} ,
with Ti′′ antEqT Tj′′ for all i, j ∈ {1, . . . , m}. From the fact that leafSet(outUnf(T ′ ))
= leafSet(outUnf(T ′′ )) we have that l′ = l′′ and that: for all Ti′ , with i ∈
{1, . . . , n}, there exists Tj′′ , with j ∈ {1, . . . , m}, such that Ti′ = Tj′′ ; and, vice
versa, for all Tj′′ , with j ∈ {1, . . . , m}, there exists Ti′ , with i ∈ {1, . . . , n}, such
that Tj′′ = Ti′ . Therefore we conclude that T ′ extAntEqT T ′′ .
In order to prove Theorem 3.12 we also need to consider a simplified subtyping procedure, whose judgments are denoted by ≤sa , and the notions of IO
steps, blocking judgments and blocking paths.
Simplified Subtyping Procedure. We here denote by ≤sa the judgements
of the subtyping procedure that is defined exactly as our procedure (defined
in Section 3.1 and based on applications of the rules therein over judgments of
the form Σ ⊢ T ≤a S) with the only difference that the Asmp rule is removed
(i.e. the subtyping procedure whose transitions were denoted by →noAsmp in
the proof of Proposition 3.4). Since, in the absence of the Asmp rule the content
of environment Σ is never accessed for reading, it has no actual effect on the
procedure (on rule applications) and can be removed as well, together with
updates on such environment made by the rules. As a consequence we will
denote ≤sa judgments just by ⊢ T ≤sa S for some T and S. Here, differently
from the →noAsmp notation used in the proof of Proposition 3.4, since we adopt
a new notation for judgements, we will simply use: ⊢ T ≤sa S → ⊢ T ′ ≤sa S ′ to
denote that the latter can be obtained from the former by one rule application.
Finally, as usual, ⊢ T ≤sa S →err denotes that there is no rule that can be
69
applied to the judgement ⊢ T ≤sa S.
Definition Appendix B.11. A blocking judgment ⊢ T ≤sa S, denoted by
⊢ T ≤sa S →blk , is a judgment such that, for some T ′ , S ′ we have: ⊢ T ≤sa S →∗
⊢ T ′ ≤sa S ′ →err by applying rules RecL, RecR1 and RecR2 only.
a
Definition Appendix B.12. An IO step a, denoted by −→io , with a ∈ {&l , ⊕l |
l ∈ L} is a sequence of ≤sa rule applications →∗ such that the last applied rule
is an In (in the case a = &l , where l is the input label singling out which of the
rule premises we consider), or an Out rule (in the case a = ⊕l , where l is the
output label singling out which of the rule premises we consider) and all other
rule applications concern RecL, RecR1 and RecR2 rules only.
Definition Appendix B.13. a1 . . . an , with n ≥ 0, is a blocking path for
a
a
n
1
judgment ⊢ T ≤sa S if there exist T ′ , S ′ such that ⊢ T ≤sa S −→
io . . . −→io ⊢
T ′ ≤sa S ′ →blk (where T ′ = T and S ′ = S in the case n = 0).
Lemma Appendix B.14. Let S ∈ reach(Z) and ⊢ T ≤sa antOut(S, γ), ⊢
T ≤sa antOut(S, β) be such that: |γ| < |β| and antOut(S, β) extAntEqZ antOut(S, γ).
If a1 . . . an , with n ≥ 0, is a blocking path for ⊢ T ≤sa antOut(S, β) then there
exists an m long prefix of a1 . . . an , with 0 ≤ m ≤ n, that is a blocking path for
⊢ T ≤sa antOut(S, γ).
Proof.
The proof is by induction on n ≥ 0.
We start by proving the base case n = 0. That is ⊢ T ≤sa antOut(S, γ) →blk ,
i.e. for some T ′ , S ′ we have: ⊢ T ≤sa antOut(S, γ) →∗ ⊢ T ′ ≤sa S ′ →err by
applying rules RecL, RecR1 and RecR2 only.
a
We first observe that ⊢ T ≤sa antOut(S, γ) −→io is not possible for any
&
l
a ∈ {&l , ⊕l | l ∈ L}. This because: if we had ⊢ T ≤sa antOut(S, γ) −→
io
for some l ∈ L, then antOut(S, β) = &{li : Ti }i∈I with l = li for some i ∈
&
l
I, hence we would have that also ⊢ T ≤sa antOut(S, β) −→
io ; and if we had
⊕
l
⊢ T ≤sa antOut(S, γ) −→
io for some l ∈ L, then, since antOut(S, β) extAntEqZ
⊕
l
antOut(S, γ), we would have that also ⊢ T ≤sa antOut(S, β) −→
io .
70
Therefore, given that it is not possible that ⊢ T ≤sa antOut(S, γ) →∗ ⊢
end ≤sa end by applying rules RecL, RecR1 and RecR2 only (because otherwise
antOut(S, β) would not be defined), we conclude ⊢ T ≤sa antOut(S, γ) →blk
(notice that the number of times a RecL, RecR1 or RecR2 is applicable to a
judgment is finite because we do not have unguarded recursion and RecR2 cannot
be consecutively applied for more than one time).
We now consider the induction case for blocking path a1 . . . an of length
n ≥ 1.
We first consider the case a1 = &l for some l ∈ L. Given that antOut(S, β)
&
l
is defined and that ⊢ T ≤sa antOut(S, β) −→
io , we deduce that antOut(S, γ) is:
either ⊕{l′ : T ′ } (possibly preceded by some recursion operators), for some l′ , T ′ ;
or &{li : Ti }i∈I (possibly preceded by some recursion operators), for some terms
Ti and labels li such that l = li for some i ∈ I. In the first case we have ⊢
T ≤sa antOut(S, γ) →blk , hence the the lemma trivially holds; in the second case
&
l
we have ⊢ T ≤sa antOut(S, γ) −→
io and we proceed with the proof. We have that
&
l
′
′
′
there exist T ′ , S ′ , σ such that ⊢ T ≤sa antOut(S, γ) −→
io T ≤sa antOut(S , γ )
&
l
′
′
′
′
′
and ⊢ T ≤sa antOut(S, β) −→
io T ≤sa antOut(S , β ), with γ = σγ and β = σβ .
In particular S ′ is obtained from S by removing all its initial (single-)outputs
(and intertwined recursions, that are unfolded) until the first input &{li : Ti }i∈I
is reached, which is also removed, thus yielding S ′ = Ti for the i ∈ I such
that l = li . This corresponds, in the definition of reach(Z) (Definition 3.9), to
repeatedly applying, starting from S ∈ reach(Z), rules 3 and 4 and finally rule 2,
thus yielding S ′ ∈ reach(Z). Notice that σ is the sequence of labels of the initial
outputs that were removed during this procedure and that, obviously, |γ| < |β|.
Now, in order to be able to apply the induction hypothesis we have also to
show that antOut(S, β ′ ) extAntEqZ antOut(S, γ ′ ). We observe that antOut(S, γ ′ )
extAntEqZ antOut(S, γ). This holds because antOut(S, γ) is a &{li : Ti }i∈I
term, with l = li for some i ∈ I, possibly preceded by some recursion operators, and from the following observations: obviously, for any t, T ′′ , it holds
µt.T ′′ extAntEqZ T ′′ {µt.T ′′ /t}; and leafSet(Ti ) ⊆ leafSet(&{li : Ti }i∈I ). In the
same way, we have antOut(S, β ′ ) extAntEqZ antOut(S, β).
71
It is therefore possible to apply the induction hypothesis to T ′ ≤sa antOut(S ′, γ ′ )
and T ′ ≤sa antOut(S ′ , β ′ ) that possesses the shorter blocking path a2 . . . an .
Finally, we consider the case a1 = ⊕l for some l ∈ L. Since ⊢ T ≤sa antOut(S, β)
⊕l
−→io and antOut(S, β) extAntEqZ antOut(S, γ), we have that also ⊢ T ≤sa
⊕
l
′
antOut(S, γ) −→
io . In particular, we have that there exists T such that ⊢
⊕
⊕
l
l
′
′
T ≤sa antOut(S, γ) −→
io T ≤sa antOut(S, γl) and ⊢ T ≤sa antOut(S, β) −→io T
≤sa antOut(S, βl), where, obviously, |γl| < |βl|. Moreover, since antOut(S, β)
extAntEqZ antOut(S, γ) it is immediate to show (by applying the definitions of
antOut, extAntEq and antEq ) that also antOut(S, βl) extAntEqZ antOut(S, γl).
It is therefore possible to apply the induction hypothesis to T ′ ≤sa antOut(S, γl)
and T ′ ≤sa antOut(S, βl) that possesses the shorter blocking path a2 . . . an .
Theorem 3.12.
Given two single-out session types T and S, we have that
there exist Σ′ , T ′ , S ′ such that ∅ ⊢ T ≤a S →∗ Σ′ ⊢ T ′ ≤a S ′ →err if and only if
there exist Σ′′ , T ′′ , S ′′ such that ∅ ⊢ T ≤t S →∗ Σ′′ ⊢ T ′′ ≤t S ′′ →err .
Proof. We consider the two implications separately starting from the if part.
Assume that ∅ ⊢ T ≤t S →∗ Σ′′ ⊢ T ′′ ≤t S ′′ →err . In this sequence of rule
applications, the new rule Asmp2 is never used otherwise the sequence would
terminate successfully by applying such a rule. Hence, by applying the same
sequence of rules, we have ∅ ⊢ T ≤a S →∗ Σ′ ⊢ T ′ ≤a S ′ with T ′′ = T ′ , S ′′ = S ′
and Σ′′ = Σ′ . We have that Σ′ ⊢ T ′ ≤a S ′ →err , otherwise if a rule could be
applied to this judgement, the same rule could be applied also to Σ′′ ⊢ T ′′ ≤t S ′′
thus contradicting the assumption Σ′′ ⊢ T ′′ ≤t S ′′ →err .
We now move to the only if part. Assume that ∅ ⊢ T ≤a S →∗ Σ′ ⊢
T ′ ≤a S ′ →err and that, by contradiction, ∅ ⊢ T ≤t S →∗ Σ′′ ⊢ T ′′ ≤t S ′′ →err
does not hold.
From ∅ ⊢ T ≤a S →∗ Σ′ ⊢ T ′ ≤a S ′ →err (since in this sequence of rule
applications the Asmp rule is never used, otherwise the sequence would terminate
successfully by applying such a rule), by applying the same sequence of rules,
we have ⊢ T ≤sa S →∗ ⊢ T ′ ≤sa S ′ .
We now observe that, since we assumed (by contradiction) that we do not
72
get the error when using the ≤t procedure, there must exist at least a triple
Σ′′′ , T ′′′ , S ′′′ such that: ∅ ⊢ T ≤t S →∗ Σ′′′ ⊢ T ′′′ ≤t S ′′′ (and correspondingly
⊢ T ≤sa S →∗ ⊢ T ′′′ ≤sa S ′′′ because the Asmp and Asmp2 rules, that would
have led to successful termination, cannot have been applied), Σ′′′ ⊢ T ′′′ ≤t S ′′′
successfully terminates by applying the Asmp or Asmp2 rule, and ⊢ T ′′′ ≤sa S ′′′
has a blocking path.
Let us now consider one of such triples Σ′′′ , T ′′′ , S ′′′ (possessing the above
stated properties) that has a blocking path of minimal length, i.e. there is no
other Σ′′′ , T ′′′ , S ′′′ triple of the kind above such that ⊢ T ′′′ ≤sa S ′′′ has a shorter
blocking path. Let a1 . . . an be such a path. Since the Asmp or Asmp2 rule is
applied to Σ′′′ ⊢ T ′′′ ≤t S ′′′ , we have S ′′′ = antOut(S, β) (in the case of Asmp
this is obtained by Corollary Appendix B.7).
We now consider γ such that (T ′′′ , antOut(S, γ)) ∈ Σ′′′ was used in the
premise of Asmp or Asmp2 rule: γ = β in the case of the Asmp rule, |γ| < |β|
in the case of the Asmp2 rule. Moreover, let us also consider Σγ to be the
environment such that ∅ ⊢ T ≤t S →∗ Σγ ⊢ T ′′′ ≤t antOut(S, γ), where
Σγ ⊢ T ′′′ ≤t antOut(S, γ) is the judgment to which the rule that caused
(T ′′′ , antOut(S, γ)) to be inserted in the environment was applied.
We now observe that there exists a m long prefix of a1 . . . an , with 0 ≤ m ≤ n,
that is a blocking path for ⊢ T ′′′ ≤sa antOut(S, γ). This is obvious in the case
γ = β; it is due to Lemma Appendix B.14 in the case |γ| < |β|: we obtain
antOut(S, β) extAntEqZ antOut(S, γ) as needed by such a Lemma from the statements in the premise of rule Asmp2 and by applying Lemmas Appendix B.5,
Appendix B.9 and Appendix B.10.
Since we assumed (by contradiction) that ∅ ⊢ T ≤t S →∗ Σ′′ ⊢ T ′′ ≤t S ′′ →err
does not hold, this would be possible only if there existed a triple Σ′′′′ , T ′′′′ , S ′′′′
such that: there is a sequence of rule applications Σγ ⊢ T ′′′ ≤t antOut(S, γ) →∗
Σ′′′′ ⊢ T ′′′′ ≤t S ′′′′ that is a prefix of the sequence of rule applications of the
blocking path for ⊢ T ′′′ ≤sa antOut(S, γ); and Σ′′′′ ⊢ T ′′′′ ≤t S ′′′′ successfully
terminates by applying the Asmp or Asmp2 rule. Notice that such a sequence
Σγ ⊢ T ′′′ ≤t antOut(S, γ) →∗ Σ′′′′ ⊢ T ′′′′ ≤t S ′′′′ should necessarily include the
73
application of, at least, an In rule (causing the algorithm to branch), because
otherwise (given that Σγ ⊢ T ′′′ ≤t antOut(S, γ) →∗ Σ′′′ ⊢ T ′′′ ≤t antOut(S, β))
we could not have that Σ′′′′ ⊢ T ′′′′ ≤t S ′′′′ successfully terminates by applying
the Asmp or Asmp2 rule.
However the existence of such a triple Σ′′′′ , T ′′′′ , S ′′′′ is not possible, because
⊢ T ′′′′ ≤sa S ′′′′ would have a k long blocking path with k < n (being such
a path strictly shorter than that of ⊢ T ′′′ ≤sa antOut(S, γ)), thus violating the
minimality assumption about the blocking path length of the Σ′′′ , T ′′′ , S ′′′ triple.
Appendix C. Proofs of Section 4
Appendix C.1. Proof of Lemma 4.4 and Theorem 4.5
We here prove Lemma 4.4 and Theorem 4.5 that show undecidability of
≤bound by reduction from the bounded non termination problem.
Lemma 4.4.
Given a queue machine M and an input x, it is undecidable
whether M does not terminate and is bound on x.
Proof. We first prove that boundedness is undecidable. If, by contraposition,
boundedness was decidable, termination could be decided by first checking
boundedness, and then perform a finite state analysis of the queue machine
behaviour. More precisely, termination on bounded queue machines can be decided by forward exploration of the reachable configurations until a terminating
configuration is found, or a cycle is detected by reaching an already visited
configuration.
We now conclude by observing that given a queue machine M and the input
x, it is not possible to decide whether M does not terminate and is bound on
x. Assume by contraposition one could decide the above property of queue
machines. Then boundedness could be decided as follows: transform M in
a new machine M ′ that behaves like M plus an additional special symbol #
which is enqueued every time it is dequeued; boundedness of M on input x can
74
be decided by checking the above property on M ′ and input #x (in fact M ′
never terminates and is bound on #x if and only if M is bound on x).
Theorem 4.5.
Given a queue machine M = (Q, Σ, Γ, $, s, δ) and an input
string x, we have that [[s]]∅ ≤bound [[x$]] if and only if M does not terminate and
is bound on x.
Proof. We need a preliminary result: given (q, γ) →M (q ′ , γ ′ ), if [[q]]∅ ≤[[γ]] then
we also have that [[q ′ ]]∅ ≤[[γ ′ ]]. In fact, assuming γ = C1 · · · Cm and δ(q, C1 ) =
(q ′ , B1C1 · · · BnCC11 ), we have γ ′ = C2 · · · Cm B1C1 · · · BnCC11 . Having [[q]]∅ ≤[[γ]], by
one application of item 4. of Definition 2.4, one application of item 3., and nC1
applications of item 2., we can conclude that [[q ′ ]]∅ ≤[[γ ′ ]].
We now observe that if M is not bound on x we have that it is not possible to
have [[s]]∅ ≤bound [[x$]]. Assume by contraposition that [[s]]∅ ≤bound [[x$]]. From the
previous preliminary result, we have that also [[q ′ ]]∅ ≤bound [[γ ′ ]] for each reachable
configuration (q ′ , γ ′ ). But due to unboundedness of M on x we have that,
for every k, there is an enqueue operation that is executed when the queue is
longer than k. Assume this happens when the configuration (q ′ , γ ′ ) performs its
computation action. In order to relate [[q ′ ]]∅ and [[γ ′ ]], we need a relation that
contains pairs with the l.h.s. starting with an output and the r.h.s. with an
input context of depth greater than k. But this cannot hold if we fix a maximal
depth smaller than k to the input context.
Now we observe that [[s]]∅ ≤bound [[x$]] if and only if M does not terminate and
is bound on x. Following the (Only if part) of the proof of Theorem 3.1 [10]
stating the undecidability of <<, we prove that if [[s]]∅ ≤bound [[x$]] then M does
not terminate. Moreover, we also have that M is bound on x in the light of the
previous observation.
Consider now that M does not terminate. As in the (If part) of the same
proof mentioned above, we define C = {(qi , γi ) | (s, x$) = (q0 , γ0 ) →M (q1 , γ1 ) →M
75
· · · →M (qi , γi ), i ≥ 0} and the following relation R on types:
R
{
=
[[q]]∅ , [[C1 · · · Cm ]] ,
&{A : ⊕{B1A : · · · ⊕ {BnAA : [[q ′ ]]∅ }}}A∈Γ , [[C1 · · · Cm ]] ,
⊕ {B1C1 : ⊕{B2C1 : · · · ⊕ {BnCC11 : [[q ′ ]]∅ }}} , &{C2 : · · · &{Cm : Z}} ,
⊕ {B2C1 : · · · ⊕ {BnCC11 : [[q ′ ]]∅ }} , &{C2 : · · · &{Cm : &{B1C1 : Z}}} ,
···
[[q ′ ]]∅ , &{C2 : · · · &{Cm : &{B1C1 : · · · &{BnCC11 : Z}}}}
|
(q, C1 · · · Cm ) ∈ C, δ(q, C1 ) = (q ′ , B1C1 · · · BnCC11 ),
Z = µt. ⊕ {A : &{A : t}}A∈Γ }
Following the proof of Theorem 3.1 [10] we show that this relation is an asynchronous subtyping relation. Moreover boundedness of M on x guarantees
boundedness on the length of the reachable queue contents C1 · · · Cm , that implies boundedness of the depth of the input contexts of the r.h.s. of all the pairs
in R. This proves that [[s]]∅ ≤bound [[x$]].
Appendix C.2. Proof of Theorem 4.8
Theorem 4.8 states that, given a single consuming queue machine M and
an input x, termination of M on x is undecidable. The theorem is proved
by resorting to Turing completeness of queue machines. In order to do this
we preliminarily provide an encoding [[M ]] from a queue machine M into a
single-consuming queue machine and two lemmas that guarantee that, given
a queue machine M and an input x, M terminates on x if and only if the
single-consuming queue machine [[M ]] terminates on x.
Definition Appendix C.1. Let M = ({q1 , .., qn }, Σ, Γ, $, s, δ) be a queue machine and let # be a special character not in Γ. We denote with [[M ]] the following single-consuming queue machine ({q1 , .., qn , q1′ , .., qn′ }, Σ, Γ ∪ {#}, $, s, δ ′)
with δ ′ defined as follows:
• δ ′ (qi , a) = (qj′ , ǫ) if δ(qi , a) = (qj , ǫ)
76
• δ ′ (qi , a) = (qj , γ) if δ(qi , a) = (qj , γ) with γ 6= ǫ
• δ ′ (qi , #) = (qi′ , ǫ)
• δ ′ (qi′ , a) = (qj , #) if δ(qi , a) = (qj , ǫ)
• δ ′ (qi′ , a) = (qj , γ) if δ(qi , a) = (qj , γ) with γ 6= ǫ
• δ ′ (qi′ , #) = (qi , #)
Given a configuration (q, γ) of [[M ]], we denote with {{(q, γ)}} the configuration
(z, β) where z = q, if q ∈ {q1 , .., qn }, or z = qi , if q = qi′ , while β is obtained
from γ by removing each instance of the special symbol #.
Example Appendix C.2. We now comment the construction [[M ]] that, given
a queue machine M , returns a single-consuming queue machine. As an example,
consider M = ({q1 , q2 }, {a}, {a, $}, $, q1, δ) with δ such that δ(q1 , a) = (q2 , ǫ),
δ(q1 , $) = (q1 , $), δ(q2 , a) = (q2 , a), and δ(q2 , $) = (q2 , ǫ). This machine accepts
the input string ”a” by consuming in sequence a and then $. Consider now
[[M ]] = ({q1 , q2 , q1′ , q2′ }, {a}, {a, $, #}, $, q1, δ ′ ) with δ ′ such that δ ′ (q1 , a) = (q2′ , ǫ),
δ ′ (q1 , $) = (q1 , $), δ ′ (q1 , #) = (q1′ , ǫ), δ ′ (q2 , a) = (q2 , a), δ ′ (q2 , $) = (q2′ , ǫ),
δ ′ (q2 , #) = (q2′ , ǫ), δ ′ (q1′ , a) = (q2 , #), δ ′ (q1′ , $) = (q1 , $), δ ′ (q1′ , #) = (q1 , #),
δ ′ (q2′ , a) = (q2 , a), δ ′ (q2′ , $) = (q2′ , #), and δ ′ (q2′ , #) = (q2 , #). This new queue
machine also accepts the input string ”a” but it does simply consume a and $
in sequence, but when $ is dequeued the special symbol # is enqueued (which
is subsequently consumed thus emptying the queue). Notice that the queue machine [[M ]] cannot consume two symbol in sequence because after the first one
is consumed, it enters in one of the primed state qi′ that always enqueue some
symbol.
Lemma Appendix C.3. Let M = (Q, Σ, Γ, $, s, δ) be a queue machine and
let x ∈ Σ∗ . If (s, x$) →∗M (q, γ) then there exists a configuration (q ′ , γ ′ ) such
that (s, x$) →∗[[M]] (q ′ , γ ′ ) with {{(q ′ , γ ′ )}} = (q, γ).
Proof. By induction on the number of steps in the sequence (s, x$) →∗M (q, γ).
The base case is trivial. In the inductive case we perform a case analysis. The
77
unique non trivial case is when the configuration reached by [[M ]] according to
the inductive hypothesis has the queue starting with the special symbol #. In
this case, [[M ]] must perform more transitions, first to consume all the instances
of # in front of the queue and then to mimick the new transition of M .
Lemma Appendix C.4. Let M = (Q, Σ, Γ, $, s, δ) be a queue machine and
let x ∈ Σ∗ . If (s, x$) →∗[[M]] (q, γ) then (s, x$) →∗M {{(q, γ)}}.
Proof. By induction on the number of steps in the sequence (s, x$) →∗[[M]] (q, γ).
The base case is trivial. In the inductive case we perform a case analysis. The
unique non trivial case is when γ starts with the special symbol #. In this case,
M does not perform any new transition as if (q ′ , γ ′ ) is the new configuration we
have that {{(q, γ)}} = {{(q ′ , γ ′ )}}.
Theorem 4.8.
Given a single consuming queue machine M and an input x,
the termination of M on x is undecidable.
Proof. The thesis directly follows from the Turing completeness of queue machines, and the two above Lemmas that guarantee that given a queue machine
M and an input x, M terminates on x if and only if the single-consuming queue
machine [[M ]] terminates on x. This is guaranteed by the fact that if [[M ]] reaches
a configuration with the queue containing only instances of #, it is guaranteed
to eventually terminate by emptying the queue.
Appendix C.3. Proof of Theorem 4.11
Theorem 4.11 states that a single-consuming queue machine does not terminate if and only if the types obtained by the encoding of Figure 4 are in
the ≤tin,tout relation. The proof is done by separately showing, as preliminary
lemmas, both implications (one in each direction) to hold.
Concerning such lemmas and their proof, we need to introduce some preliminary notation. Given a sequence of queue symbols γ, we denote with [[[γ]]]u
the set of session types that can be obtained from [[[γ]]] by independently replacing each occurrence, inside it, of the term T ′′ defined in Figure 4 with
78
antOut(T ′′ , li1 . . . lin ), for some sequence of labels li1 . . . lin with n ≥ 0 (distinguished label sequences can be considered for replacing different occurrences of
T ′′ inside [[[γ]]]). Observe that [[[γ]]]u is well defined because T ′′ can anticipate
every possible sequence of outputs. Moreover, for simplicity, we will consider
the asynchronous subtyping relation ≤ instead of ≤tin,tout . Nevertheless, we will
apply such relation on types that have all their choices labeled on the same set
of labels, hence the two relations obviously coincide on such types.
Lemma Appendix C.5. Given a single-consuming queue machine M = (Q, Σ,
Γ, $, s, δ) and an input string x ∈ Σ∗ , if [[[s]]]∅ ≤[[[x$]]] then M does not terminate
on x.
Proof. We need a preliminary result: given (q, γ) →M (q ′ , γ ′ ) and a term
∅
S ∈ [[[γ]]]u , if [[[q]]]∅ ≤S then there exists S ′ ∈ [[[γ ′ ]]]u such that [[[q ′ ]]] ≤S ′ . In
fact, assuming γ = C1 · · · Cm and δ(q, C1 ) = (q ′ , B1C1 · · · BnCC11 ), we have γ ′ =
∅
C2 · · · Cm B1C1 · · · BnCC11 . Consider now S ∈ [[[γ]]]u . Having [[[q]]] ≤S, by one application of item 4. of Definition 2.4, one application of item 3., and nA applications
∅
of item 2., we can conclude that there exists S ′ ∈ [[[γ ′ ]]]u such that [[[q ′ ]]] ≤S ′ .
We now prove the thesis by showing that assuming that M accepts x we
∅
have [[[s]]] 6≤ [[[x$]]]. By definition of queue machines, we have that: M accepts
x implies (s, x$) →∗M (q, ǫ). Assume now, by contraposition, that [[[s]]]∅ ≤[[[x$]]].
As (s, x$) →∗M (q, ǫ), by repeated application of the above preliminary result
we have that exists S ′ ∈ [[[ǫ]]]u such that [[[q]]]∅ ≤S ′ . But this cannot hold because
∅
[[[q]]] is a recursive definition that upon unfolding begins with an input that
implies (according to items 4. and 3. of Definition 2.4) that also S ′ (once
unfolded) starts with an input. But this is false, in that, by definition of the
n
o
queue encoding [[[ǫ]]] = µt ⊕ A : & A : t ⊎ A′ : T ′′ A′ ∈Γ\{A}
.
A∈Γ
Lemma Appendix C.6. Given a single-consuming queue machine M = (Q, Σ,
Γ, $, s, δ) and an input string x ∈ Σ∗ , if M does not terminate on x then
∅
[[[s]]] ≤[[[x$]]].
∅
Proof. Assuming that M does not accept x we show that [[[s]]] ≤[[[x$]]]. When
79
a queue machine does not accept an input, the corresponding computation
never ends. In our case, this means that there is an infinite sequence (s, x$) =
(q0 , γ0 ) →M (q1 , γ1 ) →M · · · →M (qi , γi ) →M · · · . Let C be the set of reachable configurations, i.e. C = {(qi , γi ) | i ≥ 0}. We now define a relation
R on types, where T ′ and T ′′ are as in Figure 4, T0 = ⊕{A : T ′′ }A∈Γ and
Tn = &{A : Tn−1 }A∈Γ :
R =
{
∅
∅
[[[q]]] , SC1 ···Cm , &{A : {{B1A · · · BnAA }}q′ }A∈Γ , SC1 ···Cm ,
∅
∅
{{B1C1 · · · BnCC11 }} ′ , SC2 ···Cm , {{B2C1 · · · BnCC11 }} ′ , SC2 ···Cm B C1 ,
q
q
1
···
∅
{{BnCC11 }}
q′
|
S
{
, SC2 ···Cm B C1 ···B C1
1
nC −1
1
(q, C1 · · · Cm ) ∈ C, δ(q, C1 ) = (q ′ , B1C1 · · · BnCC11 ), Sγ ∈ [[[γ]]]u }
∅
[[[q]]]∅ , Tn , &{A : {{B1A · · · BnAA }}q′ }A∈Γ , Tn ,
∅
∅
{{B1C1 · · · BnCC11 }} ′ , Tm , {{B2C1 · · · BnCC11 }} ′ , Tm ,
q
q
···
∅
{{BnCC11 }} ′ , Tm
q
|
(q, C1 · · · Cm ) ∈ C, δ(q, C1 ) = (q ′ , B1C1 · · · BnCC11 ),
if ∃q ′′ , C s.t. δ(q, C) = (q ′′ , ǫ) then n ≥ 2 else n ≥ 1, m ≥ 0 }
S
{
|
S
{
|
T ′ , Tn , & A1 : ⊕{A2 : T ′ }A2 ∈Γ
A1 ∈Γ
, Tn , ⊕ {A2 : T ′ }A2 ∈Γ , Tm
A1 ∈Γ
, Sγ , ⊕ {A2 : T ′ }A2 ∈Γ , Sγ
n ≥ 1, m ≥ 0 }
T ′ , Sγ , & A1 : ⊕{A2 : T ′ }A2 ∈Γ
γ ∈ Γ∗ , Sγ ∈ [[[γ]]]u }
We have that the above R is an asynchronous subtyping relation because each of
the pairs satisfies the conditions in Definition 2.4 thanks to the presence of other
∅
pairs in R. We can conclude observing that (s, x$) ∈ C implies that ([[[q]]] , [[[x$]]])
∅
belongs to the above asynchronous subtyping relation R, hence [[[q]]] ≤[[[x$]]].
80
Theorem 4.11.
Given a single consuming queue machine M = (Q, Σ, Γ, $, s, δ)
and an input string x ∈ Σ∗ , we have [[[s]]]∅ ≤tin,tout [[[x$]]] if and only if M does not
terminate on x.
Proof. Direct consequence of Lemmas Appendix C.5 and Appendix C.6.
81
| 6cs.PL
|
Deleting vertices to graphs of bounded genus
Tomasz Kociumaka∗
Marcin Pilipczuk†
arXiv:1706.04065v1 [cs.DS] 13 Jun 2017
Abstract
We show that a problem of deleting a minimum number of vertices from a graph to obtain a graph
2
embeddable on a surface of a given Euler genus is solvable in time 2Cg ·k log k nO(1) , where k is the size of
the deletion set, Cg is a constant depending on the Euler genus g of the target surface, and n is the size
of the input graph. On the way to this result, we develop an algorithm solving the problem in question in
time 2O((t+g) log(t+g)) n, given a tree decomposition of the input graph of width t. The results generalize
previous algorithms for the surface being a sphere by Marx and Schlotter [10], Kawarabayashi [6], and
Jansen, Lokshtanov, and Saurabh [5].
1
Introduction
In recent years, a significant effort has been put into the study of parameterized complexity of recognizing
near-planar graphs [10, 6, 5], that is, graphs that become planar after deleting a small number of vertices.
Since, by the classic result of Lewis and Yannakakis [9], the decision version of the problem is NP-hard, it is
natural to look for fixed-parameter algorithms with various parameters.
The parameter of the size of the deletion set naturally comes from the supposed applications: a number
of efficient algorithms for planar graphs generalize well to near-planar graphs, if one supply them with the
deletion set. Formally, we define the problem Planar Vertex Deletion as follows: given a graph G and
an integer k, decide, if one can delete at most k vertices from G to obtain a planar graph.
Clearly, for a fixed integer k, the yes-instances to Planar Vertex Deletion form a minor-closed graph
class. Consequently, from the Graph Minors theory we obtain a nonuniform fixed-parameter algorithm for
Planar Vertex Deletion (cf. [1, Section 6]).
Marx and Schlotter [10] showed an explicit, uniform fixed-parameter algorithm by a typical irrelevant
vertex approach. First, they observe that the formulation of the problem as hitting all models of the
forbidden minors for planar graphs (i.e., K5 and K3,3 ) leads to a fixed-parameter algorithm on bounded
treewidth graphs by relatively standard techniques. Second, it is quite easy to believe (but quite technical
to formally prove) that a middle part of a large, flat (planar), and grid-like subgraph of the input graph will
never be part of an optimal deletion set, and can be removed without changing the answer to the problem.
The combination of the excluded grid theorem and the technique of iterative compression gives here a winwin approach: if the treewidth of the graph is not sufficiently bounded, an irrelevant part can be uncovered
and removed.
There are two sources of potential inefficiencies in the approach of Marx and Schlotter. First, the routine
for graphs of bounded treewidth that finds a minimum set hitting all forbidden minor models work in
time double-exponential in the treewidth bound. Since the treewidth bound needs to be significantly larger
than the size of the deletion set for the irrelevant vertex argument to work, we obtain at least a doubleexponential dependency on the parameter. Second, the technique of iterative compression, at least applied
in a straightforward manner, gives at least quadratic dependency on the input size.
∗ University
of Warsaw, Poland, [email protected].
of Warsaw, Poland, [email protected]. Research supported by Polish National Science Centre grant DEC2012/05/D/ST6/03214.
† University
1
Later, Kawarabayashi [6] showed a fixed-parameter algorithm with linear dependency on the input size.
Finally, Jansen, Lokshtanov, and Saurabh [5] showed an algorithm with running time 2O(k log k) n, that is,
with nearly single-exponential dependency on the parameter and linear dependency on the input size.
On high level, the work of [5] follows the approach of Marx and Schlotter, but improves upon both
components. First, they show that a routine that explicitly constructs partial embeddings of graphs of
bounded treewidth solves the problem in question in time 2O(t log t) n, given a tree decomposition of width t.
Second, they show arguments in the spirit of the aforementioned irrelevant vertex rule that reduce the graph
to treewidth linearly bounded in the size of the solution (deletion set). Third, they apply a more involved
iterative compression approach that in one step compresses the graph by a multiplicative factor, yielding a
linear dependency on the input size.
A simple reduction from the Vertex Cover problem (replace every edge uv with a K5 with vertices u,
v, and 3 new vertices) shows that, unless the Exponential Time Hypothesis (ETH) [4] fails, the dependency
on the parameter k needs to be 2Ω(k) for any parameterized algorithm for Planar Vertex Deletion.
Although it is open whether Planar Vertex Deletion can be solved in 2o(k log k) nO(1) time, we note that
a lower bound by the second author [12] asserts that, unless the ETH fails, the bounded treewidth subroutine
requires dependency 2Ω(t log t) on the treewidth of the graph. This implies that a hypothetical algorithm that
solves Planar Vertex Deletion in 2o(k log k) nO(1) time needs to follow significantly different approach
than the one we know currently.
In the light of the aforementioned developments, in this paper, we initiate the study of the Genus
Vertex Deletion: given a graph G and integers g and k, decide, if one can delete at most k vertices from
G to obtain a graph embeddable on a surface of Euler genus at most g.
Our main result is the following:
Theorem 1.1. Genus Vertex Deletion can be solved in time 2Cg k
depending on g only.
2
log k O(1)
n
, where Cg is a constant
In the proof of Theorem 1.1, we follow the general approach of Marx and Schlotter [10]. Our main
technical contribution is the generalization of the bounded treewidth subroutine of [5] to the bounded genus
case.
Theorem 1.2. Given a Genus Vertex Deletion instance (G, g, k) with |V (G)| = n, and a tree decomposition of G of width t, one can solve the Genus Vertex Deletion problem on (G, g, k) in time
2O((t+g) log(t+g)) n.
The proof of Theorem 1.2 follows the same principle as the corresponding routine of [5] — building partial
embeddings for graphs with small boundaries — but requires significant technical hurdle to be presented
formally and in full detail.
The task of the algorithm of Theorem 1.2 can be also seen as a generalization of an algorithm that
computes the Euler genus in graphs of bounded treewidth: by substituting k = 0 we obtain an algorithm
with running time 2O((t+g) log(t+g)) n that checks if the input graph is embeddable on a surface of Euler genus
at most g. Such a result is not new: a bounded treewidth routine is also part of the current algorithms that
compute embeddings in linear time [11, 7]. In particular, the work of Kawarabayashi, Mohar, and Reed [7]
claims an algorithm with running time f (t) · n for some function f (i.e., without the exponential dependency
on g). However, the work [7] is an extended abstract from FOCS 2008 that, to the best of our knowledge,
never substantiated in a full version, and we were unable to reproduce the details of this algorithm from the
description in [7].
For the second part, namely the irrelevant vertex argument, we generalize the arguments of Marx and
Schlotter [10]:
Theorem 1.3. For every integer g there exists a constant Cg such that the following holds. Given a graph
G, an integer k, and a set M ⊆ V (G) such that G − M is embeddable into a surface of genus at most g, one
can in time Cg nO(1) find one of the following:
1. a tree decomposition of G of width at most Cg |M |1/2 k 3/2 ;
2
2. a vertex w ∈ V (G) such that every solution to the Genus Vertex Deletion instance (G, g, k)
contains w; or
3. a vertex v ∈ V (G) such that (G, g, k) is a yes-instance to Genus Vertex Deletion if and only if
(G − {v}, g, k) is.
Theorem 1.1 follows now from Theorems 1.2 and 1.3 in a standard manner. By a standard application of
the irrelevant vertex technique (cf. [1, Section 4], we can assume that, apart from the input Genus Vertex
Deletion instance (G, g, k), we are additionally given a set M ⊆ V (G) of size k + 1 such that G − M is
embeddable on a surface of Euler genus at most g (i.e., M is a solution of slightly too large size), at the cost
of an additional O(n) factor in the running time bound. We iteratively apply the algorithm of Theorem 1.3
to (G, g, k) and M : if a vertex w or v is returned, we delete it and restart (decreasing the parameter k by one
in case of a vertex w); if a tree decomposition is returned, we solve the Genus Vertex Deletion problem
with the algorithm of Theorem 1.2.
Since |M | = k + 1, the algorithm of Theorem 1.2 is applied to a tree decomposition of width of the order
of Cg k 2 , yielding the bound promised in Theorem 1.1.
While the lower bound of [12] shows that the bounded-treewidth routine of Theorem 1.2 has optimal
running time (assuming ETH), we conjecture that the running time bound of Theorem 1.1 is not optimal, and
can be improved similarly as it was in the planar case [5]. For this reason, we do not optimize the parameter
dependency in Theorem 1.3, favouring the clarity of the arguments. In other words, we view our contribution
as Theorem 1.2 being the main technical merit, while Theorem 1.3 and the resulting Theorem 1.1 being an
example application.
The paper is organized as follows: after introducing notation for permutations and tree decompositions
in Section 2, we discuss combinatorial embeddings in Section 3. Theorem 1.2 is proved in Section 4, and
Theorem 1.3 is proved in Section 5.
2
2.1
Preliminaries
Permutations, involutions, cycles
For a nonnegative integer t, we denote [t] = {1, 2, . . . , t}. The group of all permutations of a set U is denoted
by Sym(U ). Given a set of permutations S of a set U , by hSi we denote the subgroup of Sym(U ) generated
by S. Given a subgroup Γ of the group of Sym(U ), by orb(Γ) we denote the family of orbits of Γ. For
a permutation σ, orb(σ) is a shorthand for orb(hσi); this also allows us to speak about orbits of a single
permutation. An orbit is trivial if it consists of a single element.
A permutation σ is an involution if σ(σ(i)) = i for every i ∈ domain(σ) and is fixed-point free if σ(i) 6= i
for every i ∈ domain(σ). Note that a permutation is a fixed-point free involution if and only if every its orbit
is of size exactly two.
A permutation σ is a cycle permutation if it has exactly one nontrivial orbit; note that σ needs to act
cyclically on this orbit. A cycle is an unordered pair consisting of a cycle permutation and its inverse.
We will need the operation of restricting a cycle permutation σ to a subset A of the elements of the
nontrival orbit v of σ: the result is a permutation σA , where σA (e) = e for every e ∈
/ A while for every e ∈ A
we have σA (e) = σ k (e) where k is the minimum positive integer with σ k (e) ∈ A. In other words, we shorten
the nontrivial orbit v by crossing out the elements not belonging to A. Note that if |A| ≥ 2, then σA is also
a cycle permutation, while for |A| ≤ 1 we have σA being an identity. The definition of restricting naturally
extends to cycles by restricting both components.
For a sequence P of elements of some set, by P̄ we denote its reverse. A subsequence of a cycle is a
sequence consisting of consecutive elements of the nontrivial orbit of the cycle.
In our work we will often analyze the subgroup of the permutation group spanned by two fixed-point free
involutions. Let α and β be two fixed-point free involutions of a set U and let Γ = hα, βi. Observe that for
any orbit v of Γ and any element e ∈ v, the orbit v consists of elements
e, α(e), β(α(e)), α(β(α(e))), β(α(β(α(e)))), α(β(α(β(α(e))))), . . . .
3
The above order is cyclic: |v| is always even, and if |v| = 2k, then (β ◦ α)k (e) = e. To fix notation, let us
hα,βi
define a cycle permutation oe
of U as follows:
i
α(f ) if f ∈ v and f = (β ◦ α) (e) for some i,
oehα,βi (f ) = β(f ) if f ∈ v and f = α ◦ (β ◦ α)i (e) for some i,
f
iff ∈
/ v.
hα,βi
Note that oe
is a cycle permutation whose nontrivial orbit is v. Furthermore, while its definition formally
depends on the choice of e and the order of α and β, different choices of e ∈ v and a potential swap of the
hα,βi
roles of α and β lead either to oe
or its inverse. The definition of a cycle is suited to accommodate that:
hα,βi
hα,βi
For the cycle permutation oe
, its cycle is denoted as ôe
. By the previous argumentation, the cycle
does not depend on the order of α and β, nor of the choice of the element e within the same orbit.
Let σ̂1 and σ̂2 be two cycles of disjoint sets U1 and U2 , with nontrivial orbits v1 and v2 . A merge of
σ̂1 and σ̂2 is a cycle σ̂ of U := U1 ∪ U2 , whose nontrivial orbit v consists of exactly elements of v1 ∪ v2 .
Furthermore, for some choice of cycle permutations σ1 ∈ σ̂1 , σ2 ∈ σ̂2 , σ ∈ σ̂, for every i = 1, 2 and e ∈ vi ,
if k is the minimum positive integer for which σ k (e) ∈ vi , then σ k (e) = σi (e). In other words, if we restrict
the cycle of the nontrivial orbit of σ̂ to the elements of Ui only, we obtain the cyclic order of the nontrivial
orbit of σ̂i .
2.2
Tree decompositions
Given a graph G, a tree decomposition of G is a pair (T, β) where T is a rooted tree and is a function
β : V (T ) → 2V (G) that assigns to every t ∈ V (T ) a bag β(t) ⊆ V (G) such that the following holds:
• for every edge e ∈ E(G), there is a node t ∈ V (T ) with e ⊆ β(t);
• for every vertex v ∈ V (G), the set {t ∈ V (T ) : v ∈ β(t)} is nonempty and induces a connected subgraph
of T .
The width of a decomposition is the maximum size of a bag, minus one.
For a tree decomposition (T, β) of a graph G, we define two auxiliary functions α and G↓ . For a node
t ∈ V (T ),
• by α(t) ⊆ V (G) we denote the union of all bags of descendants of t (including t itself);
• by G↓ (t) we denote the graph G[α(t)] − E(G[β(t)]), that is, the graph induced by G↓ (t) with the edges
inside the bag β(t) removed.
For dynamic programming algorithms, it is often convenient to work with so-called nice tree decompositions, where the bag of the root is empty, and every node t ∈ V (T ) is of one of the following four
types:
leaf node has no children and its bag is empty;
introduce node has one child t0 such that β(t) = β(t0 ) ∪ {v} for some vertex v ∈
/ β(t0 );
forget node has one child t0 such that β(t) = β(t0 ) \ {v} for some vertex v ∈ β(t0 );
join node has two children t1 and t2 with β(t) = β(t1 ) = β(t2 ).
It is well-known (see, e.g., [1, 8]) that, in polynomial time, one can turn any tree decomposition into an
equivalent nice one without increasing its width.
4
3
3.1
Embeddings and operations on them
Graph and hypergraph embeddings
We start with a clean but abstract notion of a hypergraph embedding, and then we restrict ourselves only
to graph embeddings.
Definition 3.1 (hypergraph embedding). A hypergraph embedding is a tuple (F, θ, σ, φ), where F is a finite
set, whose elements are called flags, and θ, σ, and φ are three fixed-point free involutions of the set F.
Given a hypergraph embedding E = (F, θ, σ, φ), we use the notation F(E) := F etc. Note that in any
hypergraph embedding, since θ, σ, and φ are fixed-point free involutions, the number of flags needs to be
even.
Definition 3.2 (connected components, vertices, edges, faces). Let E = (F, θ, σ, φ) be a hypergraph embedding. Then
a connected component is an orbit of ccE := hθ, σ, φi;
a vertex is an orbit of VertsE := hσ, φi;
an edge is an orbit of EdgesE := hθ, σi;
a face is an orbit of FacesE := hθ, φi.
Note that, as all three permutations of E are fixed-point free involutions, every vertex, edge, or a face of
E can be identified with a cycle, whose nontrivial orbit is the vertex/edge/face in question.
Given a hypergraph embedding E = (F, θ, σ, φ), a size-k edge, vertex, or face is an orbit of EdgesE ,
VertsE , or FacesE that consists of k flags. Note that k is always a positive even integer in this context.
Also, observe that a size-2 edge corresponds to two equal orbits or θ and σ, a size-2 vertex corresponds to
two equal orbits of σ and φ, while a size-2 face corresponds to two equal orbits of θ and φ.
We now define the genus of an embedding.
Definition 3.3 (genus of a hypergraph embedding). Given a hypergraph embedding E = (F, θ, σ, φ), its
genus is defined as
ĝE := 21 |F| − |orb(VertsE )| − |orb(EdgesE )| − |orb(FacesE )| + 2|orb(ccE )|.
(1)
Since in any hypergraph embedding the number of flags is even, the genus of a hypergraph embedding is
always an integer.
In a graph embedding, every edge consists of four flags.
Definition 3.4 (graph embedding). A hypergraph embedding E is a graph embedding, or simply an embedding, if every edge consists of exactly four flags.
In other words, a hypergraph embedding E = (F, θ, σ, φ) is a graph embedding if no two orbits of θ and
σ coincide, but the involutions θ and σ commute.
Observe that the formula for genus simplifies in case of a graph embedding.
Observation 3.5. If E = (F, θ, σ, φ) is an embedding, then its genus equals
ĝE := |orb(EdgesE )| − |orb(VertsE )| − |orb(FacesE )| + 2|orb(ccE )|.
(2)
From this point, we use only graph embeddings in this work, and call them simply embeddings.
If two objects (face, edge, vertex) share a flag, we say that these objects are incident. Note that an object
with k flags can be incident to at most k/2 objects of each of the other types (e.g., a size-k vertex can be
incident to at most k/2 edges and k/2 faces). In particular, in an embedding, edge can be incident to one
or two faces and one or two vertices. An edge incident with only one vertex is a loop.
5
φ
θ
σ
θ
Figure 1: Flags (gray triangles) and involutions in an embedding.
Let us now relate the aforementioned definition of a (combinatorial) embedding with the natural intuition.
Let G be a graph, embedded on a surface. We visualize every edge as a (thin, and possibly bend) rectangle,
with a flag attached at every corner of the rectangle (see Figure 1). The involution θ pairs up flags on an
edge that lie on the same side. The involution σ pairs up flags on an edge that lie at the same endpoint.
Finally, the involution φ pairs up neighboring flags of consecutive edges around a vertex. In this manner, an
orbit of VertsE = hσ, φi yields a cyclic order of flags around a vertex, an orbit of EdgesE = hθ, σi yields a
cyclic order of the four flags of an edge, while an orbit of FacesE = hθ, φi yields a cyclic order of flags around
a face. The formula (2) corresponds to the standard notion of an Euler genus of an embedding.
3.2
Basic operations on embeddings
We will need an operation of deleting an edge, and a reverse operation of drawing a new edge.
3.2.1
Deleting an edge
We start with the former. Let E = (F, θ, σ, φ) be an embedding and let e be an edge of E with flags
x, x0 = σ(x), y = θ(x), y 0 = σ(y) = θ(x0 ).
We define an embedding E − e, the result of deletion of the edge e from E, in the following manner.
First, we disconnect the edge e from the rest of the embedding. For this, we modify φ so that {x, x0 }
and {y, y 0 } become its orbits: If {x, x0 } is not already an orbit of φ, we set a = φ(x) and a0 = φ(x0 ), and
we replace {x, a} and {x0 , a0 } with {x, x0 } and {a, a0 }. Next, if {y, y 0 } is not already an orbit of φ, we set
b = φ(y), b0 = φ(y 0 ), and we replace {y, b} and {y 0 , b0 } with {y, y 0 } and {b.b0 }. As the third and final step,
we delete the flags of e and the corresponding orbits of all three involutions.
Note that the first two steps cannot be conveyed in parallel because φ(y) and φ(y 0 ) might be altered
during the first step. However, a direct check shows that {x, x0 } and {y, y 0 } both become orbits of φ and
that φ keeps being an involution. What is more, the order of the first two steps (i.e., the arbitrary decision
of processing {x, x0 } prior to {y, y 0 }) is irrelevant and edge deletion results in deleting the flags of e from the
cycles corresponding to the vertices incident with e. This interpretation supports the following observation.
Observation 3.6. Let E0 = E − e for an embedding E and an edge e. For every vertex v0 of E0 , there exists
a distinct vertex v of E, such that the cycle of v0 is the cycle of v with the flags of e removed. In the other
direction, for every vertex v of E, either all flags of v are contained in e, or there exists a vertex v0 of E0 with
the cycle equal to the cycle of v with the flags of e removed.
Next, we describe how genus changes subject to edge deletion.
Lemma 3.7. Let e be an edge in the embedding E, and let E0 = E − e. Then the genus of E0 is not larger
than the genus of E. Furthermore, if e is incident to two faces or to a size-2 vertex, then the genera of E0
and E are equal.
6
Proof. Let E = (F, θ, σ, φ) and E0 = (F0 , θ0 , σ0 , φ0 ). Clearly, |orb(EdgesE0 )| = |orb(EdgesE )| − 1. We now
investigate how the number of vertices, faces, and connected components can change while deleting an edge
e. Let x1 , x2 = σ(x1 ), y1 = θ(x1 ), y2 = θ(x2 ) = σ(y1 ) be the four flags of e.
For vertices, recall that the operation of deleting an edge almost preserves the set of vertices: every
vertex v0 of E0 originates from a vertex v of E by deleting the flags of e from the cycle of v. Thus,
|orb(VertsE )| − |orb(VertsE0 )| equals the number of vertices of E that have all flags contained in e. There
may be 0, 1, or 2 of them. Furthermore, if e is incident to two distinct faces, there are no such vertices,
unless e is a loop at a vertex v incident only to e; in this case, e is incident to two faces of size 2, and its set
of flags is a whole connected component of E.
For faces, consider first the case when e is incident to two faces, f1 and f2 , and assume that xi , yi lie on
fi for i = 1, 2. Then, if the cycle of fi is xi , yi , Pi for some sequence of flags Pi , then the deletion of e replaces
the faces f1 and f2 with one face with cycle P1 P̄2 , where P̄2 is the sequence P2 reversed. Consequently,
|orb(FacesE )| − |orb(FacesE0 )| = 1 in this case.
Consider now a case when e is incident to one face f . If the cycle of f is x1 , y1 , P1 , x2 , y2 , P2 for some
sequences of flags P1 and P2 , then the deletion of e replaces f with a face with a cycle P1 P̄2 . If the cycle of
f is x1 , y1 , P1 , y2 , x2 , P2 for some P1 , P2 , then the deletion of e replaces f with two faces with cycles P1 and
P2 , respectively. Consequently, in this case we have |orb(FacesE )| − |orb(FacesE0 )| ∈ {0, −1}.
For connected components, note that |orb(ccE )| − |orb(ccE0 )| ∈ {−1, 0, 1}: either e connects two distinct
vertices that land in different connected components of E0 , or the set of connected components essentially
does not change (one connected component loses the flags of e) or e is a whole connected component and it
disappears in E0 .
Let us now wrap up the argument. If the set of flags of e is a whole connected component, then
|orb(ccE )| − |orb(ccE0 )| = 1. This may happen if e is a loop at a vertex incident only to e, or if e connects
two vertices incident only to e. In the first case, let k ∈ {1, 2} be the number of faces e is incident with. We
have
|orb(FacesE ) − |orb(FacesE0 )| = k,
|orb(VertsE )| − |orb(VertsE0 )| = 1,
|orb(ccE )| − |orb(ccE0 )| = 1.
Consequently, ĝE − ĝE0 = 2 − k ∈ {0, 1}. In particular, the genera of E and E0 are equal if e is incident to
two faces.
In the second case, when e connects two size-2 vertices, e is incident with one face with the same set of
flags. Thus
|orb(FacesE )| − |orb(FacesE0 )| = 1, |orb(VertsE )| − |orb(VertsE0 )| = 2, |orb(ccE )| − |orb(ccE0 )| = 1.
Consequently, ĝE − ĝE0 = 0.
Now consider a case when e is incident to two faces, but is not a whole connected component. Then,
at least one of this faces is of size larger than 2, say f1 with cycle x1 , y1 , P1 . Then, P1 remains in E0 as
a sequence of flags, where every two consecutive flags form an orbit either of φ or of θ. Consequently, the
flags of the vertices incident with e remain in the same vertices and in the same connected component, as
P1 connects the first and the last flags of P1 . We infer that in this case we have:
|orb(FacesE )| − |orb(FacesE0 )| = 1, |orb(VertsE )| − |orb(VertsE0 )| = 0, |orb(ccE )| − |orb(ccE0 )| = 0.
Consequently, ĝE = ĝE0 . Note that at this point we have concluded the proof that ĝE = ĝE0 if e is incident
to two faces.
Consider now a case when e is incident to one face, and there exists its incident vertex v with the set
of flags contained in e. As we have already excluded the case when the flags of e is a whole connected
component, e is incident with two vertices v and v 0 , and the flags of v 0 are not contained in e. Furthermore,
v is of size 2, say v consists of flags y1 and y2 . Then, the cycle of the face incident with e is x1 , y1 , y2 , x2 , P
for some sequence of flags P . Consequently,
|orb(FacesE )| − |orb(FacesE0 )| = 0, |orb(VertsE )| − |orb(VertsE0 )| = 1, |orb(ccE )| − |orb(ccE0 )| = 0.
7
We infer that ĝE = ĝE0 . Note that at this point we have concluded the proof that ĝE = ĝE0 if e is incident to
a size-2 vertex.
Assume now that e is incident to one face, and every vertex incident with v has its set of flags not contained
in e. Let f be the face incident with e. If the cycle of f is x1 , y1 , P1 , x2 , y2 , P2 , then the resulting face with
cycle P1 P̄2 provides connectivity in E0 between the (remainings of) vertices incident with e. Consequently,
In this case
|orb(FacesE )| − |orb(FacesE0 )| = 0, |orb(VertsE )| − |orb(VertsE0 )| = 0, |orb(ccE )| − |orb(ccE0 )| = 0,
and have then ĝE − ĝE0 = 1.
In the last case, if the cycle of f is x1 , y1 , P1 , y2 , x2 , P2 , we have
|orb(FacesE )| − |orb(FacesE0 )| = −1,
|orb(VertsE )| − |orb(VertsE0 )| = 0,
|orb(ccE )| − |orb(ccE0 )| ∈ {0, −1},
and then ĝE − ĝE0 ∈ {0, 2}.
A direct corollary of Lemma 3.7 is the following.
Corollary 3.8. The genus of an embedding is always nonnegative.
Proof. Observe that any embedding can be turned into an empty embedding (i.e., one with no flags) by
successive edge deletions. Since edge deletion cannot increase the genus (Lemma 3.7), and the empty
embedding has genus zero, the lemma follows.
3.2.2
Drawing a new edge
Let us now define a reverse operation to edge deletion. Before, let us introduce a notion of position pE (x)
of a flag x. Let x, y = θ(x), y 0 = σ(y), x0 = σ(x) be the flags contained in the edge e containing x. We set
pE (x) = (φ(x), φ(y)) if φ(y) ∈
/ {x, x0 } and pE (x) = (φ(x), φ(σ(φ(y)))) otherwise. It is easy to observe that
defining the process of deletion of e, we set a and b so that (a, b) = pE (x). This lets us use pE (x) to undo
the deletion. We often do not want to specify the flags of the new edge. In this case, we write a = ⊥ instead
of a = x0 , a = > instead of a = y, a = >0 instead of a = y 0 , and b = ⊥ instead of b = y 0 .
Let E = (F, θ, σ, φ) be an embedding, and let a ∈ F ∪ {⊥, >, >0 } and b ∈ F ∪ {⊥}. By drawing a new edge
(x, y, y 0 , x0 ) at (a, b) we mean the following operation, resulting in an embedding E0 . First, we add new flags
{x, y, y 0 , x0 } to F and we set {x, x0 } and {y, y 0 } to be new orbits of σ and φ, and setting {x, y} and {x0 , y 0 }
to be two new orbits of θ. Next, we replace a = ⊥, a = >, a = >0 , and b = ⊥ with a = x0 , a = y, a = y 0 ,
and b = y 0 , respectively. Finally, we adjust φ in the following two steps: If b 6= y 0 , we define b0 = φ(b) and
replace the orbits {y, y 0 } and {b, b0 } with {y, b} and {y 0 , b0 }. Then, if a 6= x0 , we define a0 = φ(a) and the
orbits {x, x0 } and {a, a0 } with {x, a} and {x0 , a0 }.
We defined this operation so that it is clear that if e = (x, y, y 0 , x0 ) is an edge of E, then one can retrieve
E from E − e by drawing a new edge (x, y, y 0 , x0 ) at pE (x). It is also easy to verify that drawing a new edge
is a well defined and that if E0 is obtained from E by drawing a new edge e, then E = E0 − e.
Note that a = ⊥ and b = ⊥ both result in creating a new vertex (two new vertices if a = ⊥ and b = ⊥
hold simultaneously). Moreover, if a = ⊥, then x belongs to a new size-2 vertex (x, x0 ), while a = >, then x
belongs to a new size-2 face (x, y).
We identify one more special case of drawing a new edge. If a and b are distinct flags that lie on the same
face f , and furthermore, the order of flags on the cycle of f is φ(a), a, P, b, φ(b), P 0 for some (possibly empty)
sequences of flags P and P 0 , then we say that the new edge is drawn along the boundary of f . Observe that
if this is the case, then the new edge e is incident to two faces: in the new embedding E0 , the face f has
been split into a face with cycle P, a, x, y, b and a face with cycle P 0 , φ(b), y 0 , x0 , φ(a). We explicitly allow
here also the case b = φ(a); then the cycle of f is a, P, b for some sequence P , and the new embedding has
new faces with cycles x0 , y 0 and P, a, x, y, b.
Consequently, from Lemma 3.7 we immediately obtain the following.
8
Lemma 3.9. If E0 is created from E by drawing a new edge, then the genus of E0 is not smaller than the
genus of E. Furthermore, the genera of E and E0 are equal if the edge has been drawn along a face boundary
or at (a, b) with a ∈ {⊥, >}.
3.3
t-boundaried embeddings
Definition 3.10 (t-boundaried embedding). A t-boundaried embedding is a tuple (E, t, L) where E =
(F, θ, σ, φ) is an embedding, t is a nonnegative integer, and L is an injective function from a subset of
orb(VertsE ) to [t]. The elements of the domain of L are called labelled vertices, and the elements of [t] are
labels.
A genus of a t-boundaried embedding (E, t, L) is the genus of the underlying embedding E.
The main motivation to introduce t-boundaried embeddings is to then merge them.
Definition 3.11 (merge of two t-boundaried embeddings). Let Eb1 and Eb2 be two t-boundaried embeddings,
where Ebi = (Ei , t, Li ) and Ei = (Fi , θi , σi , φi ) for i = 1, 2, and the sets of flags F1 and F2 are disjoint. A
t-boundaried embedding Eb = (E, t, L) with E = (F, θ, σ, φ) is a merge of Eb1 and Eb2 if it is created by the
following process.
First, we take Eb to be a disjoint union of Eb1 and Eb2 , that is, we take:
F = F1 ∪ F2
θ = θ1 ∪ θ2
σ = σ1 ∪ σ2
φ = φ1 ∪ φ2
L = L1 ∪ L2
Then, for every label ` ∈ [t] that is contained in both the range of L1 and L2 , that is, there exists a cycle
Ci corresponding to the orbit L−1
i (`) for i = 1, 2, we modify φ on the elements of C1 and C2 so that these
elements form a single orbit of hσ, φi whose cycle C is a merge of C1 and C2 . We assign this cycle the label
` in the assignment L.
A genus-minimum merge of Eb1 and Eb2 is a merge that minimizes its genus.
Note that, as we only modify φ in the merge operation, the set of edges of a merge is a union of the edges
of the components. In particular, it follows that every edge of a merge consists of four flags, and thus it is
indeed an embedding.
Definition 3.12 (equivalence of t-boundaried embeddings). Two t-boundaried embeddings Eb1 and Eb2 are
b the genera of genus-minimum merges of Eb and Ebi for
equivalent if for every t-boundaried embedding E,
i = 1, 2 are equal.
By Observation 3.6, there is a natural correspondence between the vertices of E and E − e for every edge
e of E. This correspondence allows us to extend the definition of edge deletion to t-boundaried embeddings:
if Eb = (E, t, L) and e is an edge in E, then a t-boundaried embedding Eb − e is defined as (E − e, t, L0 ), where
L0 is defined so that for every vertex v0 of E − e, we take the corresponding vertex v of E, and copy its label
to v0 if v is labelled. Note that the range of L0 may be a proper subset of the range of L if some labelled
vertex in Eb has all its flags contained in the deleted edge e.
Furthermore, the characterization of vertex cycles in Observation 3.6 lets us relate edge deletion to
merging.
Observation 3.13. Consider a merge Eb of t-boundaried embedding Eb1 and Eb2 . For every edge e of Eb1 , we
have that Eb − e is a merge of Eb1 − e and Eb2 .
Similarly, we can define the operation of drawing a new edge in a t-boundaried embedding; if a new
vertex is created by this operation (due to a or b being equal to ⊥), we need to specify its label (or the fact
b we add a special value ⊥` available for a and b,
that it is unlabelled). Thus, for each label ` unused in E,
denoting the fact that the new vertex remains is labelled `; ordinary ⊥ denotes the fact that it is unlabelled.
A counterpart of Observation 3.13 requires dealing with a special situation.
9
Observation 3.14. Consider a merge Eb of t-boundaried embedding Eb1 and Eb2 , and let Eb10 be an embedding
obtained from Eb1 by drawing an edge at (a, b). Then a merge of Eb10 and Eb2 can be obtained from Eb by drawing
an edge at (a0 , b0 ) , where a0 = a and b0 = b except for the following situation: if a = ⊥` (or b = ⊥` ) for a
label ` used in Eb but not in Eb1 , then a0 (resp. b0 ) is an arbitrary flag of Eb contained in a vertex with label `.
3.4
Nice embeddings
We say that a vertex or a face is isolated if its set of flags is a whole connected component.
Definition 3.15 (nice (t-boundaried) embedding). A t-boundaried embedding Eb = (E, t, L) with E =
(F, θ, σ, φ) is nice if the following two conditions hold:
1. If an unlabelled vertex consists of less than 6 flags, then it is isolated.
2. If an edge e is incident to two faces, then for every face f incident with e, if we denote by x and
y = θ(x) the two flags contained both in f and e, then there is a flag z ∈ f \ {x, y, φ(x), φ(y)} that is
contained in a labelled vertex.
Our goal in this section is to show that any embedding can be turned into an equivalent nice one. The
main motivation for such a cleaning step is that a nice embedding enjoys a good size bound due to the Euler
formula-style estimations, presented in the next section.
3.4.1
Nice embeddings are small
Lemma 3.16. A t-boundaried nice embedding Eb = (E, t, L) of genus ĝ satisfies
|F(E)| ≤ 48t + 24ĝEb.
Proof. Let E = (F, θ, σ, φ). We perform a discharging argument. The setup is as follows:
• every labelled vertex receives a charge of 2;
• every isolated vertex receives a charge of 1;
• every isolated face receives a charge of 1;
• every edge receives a charge of 65 .
The total initial charge is at most
2t + 2|orb(orb(ccE ))| + 56 |orb(EdgesE )|.
Then we move the charge according to the following rules:
1. Every labelled vertex that is incident to only one face, sends a charge of 1 to the face it is incident
with.
2. Every edge that is incident only to labelled vertices, divides a charge of 23 equally among the faces it is
incident with. In other words, every flag in such an edge sends a charge of 16 to the face it is contained
in.
3. Every edge that is incident with two vertices, one labelled and one unlabelled, gives a charge of 31 to
the unlabelled incident vertex, and divides the remaining charge of 12 equally between the faces it is
incident with. In other words, in the first part every flag in such an edge contained in unlabelled vertex
sends a charge of 16 to the vertex it is contained in. In the second part, every flag in such an edge sends
a charge of 18 to the face it is contained in.
10
4. Every edge that is incident only to unlabelled vertices, divides the charge of 23 equally among the
vertices it is incident with. In other words, every flag in such an edge sends a charge of 16 to the vertex
it is contained in.
Clearly, every edge is left with non-negative charge (0 or 61 ). We now show that at the end of the process,
every vertex and every face has charge of at least one.
First, let us consider a vertex v. If v is labelled or isolated, 1 out of its initial charge remained, and the
claim is straightforward. Next, we assume that v is unlabelled and not isolated. By the first property of a
nice embedding, v is of size at least 6. It received exactly 16 from each of its flags, i.e., at least 1 in total.
Consider now a face f . We make case distinction depending on how many flags of f belong to labelled
vertices and how these flags are located on the cycle corresponding to f .
No flags of f belong to a labelled vertex. Observe that every edge e incident with f is incident only
with one face, as otherwise it would contradict the second property of a nice embedding. Consequently, f is
isolated an it received an initial charge of 1.
Exactly 2 flags of f belong to a labelled vertex. Suppose f is incident with only one labelled vertex
v and shares two flags x and y = φ(x) with v. Let e be the edge containing x. By the second property of
a nice embedding, f is the only face incident with e. Consequently, σ(x) belongs to f . Since σ(x) belongs
to v, which is a labelled vertex, we infer that y = σ(x), the vertex v is of size 2, and f is the only face v is
incident with. Thus, f received a charge of 1 from v.
Exactly 4 flags of f belong to labelled vertices and they are consecutive along f . Let x, y = θ(x),
x0 = φ(x), and y 0 = φ(y) be these four flags, and let e be the edge containing x and y. If e is incident to two
faces, then it violates the last property of a nice embedding. Otherwise, the orbit {σ(x), σ(y)} of θ appears
on f . Since only 4 flags of f belong to labelled vertices, we need to have {σ(x), σ(y)} = {x0 , y 0 }, and f is an
isolated face of size 4. Hence, it received an initial charge of 1.
The face f consists of 6 flags and they all belong to labelled vertices. In this case all flags in f are
contained in edges incident only to labelled vertices. Thus, f receives a charge of 61 from each of these flags,
which is 1 in total.
Along f there exist two nonconsecutive orbits of φ that are included in labelled vertices. Let
{x, x0 } and {y, y 0 } be these orbits. Since they are nonconsecutive, x, x0 , θ(x), θ(x0 ), y, y 0 , θ(y), θ(y 0 ) are eight
pairwise distinct flags in f , each sending a charge of at least 81 to the face f .
Note that the last three cases case cover the case of f having at least 4 flags contained in labelled vertices.
We have shown that there was enough charge so that every vertex and every face received a charge of at
least one. Consequently,
|orb(FacesE )| + |orb(VertsE )| ≤ 2t + 2|orb(ccE )| + 65 |orb(EdgesE )|.
Together with (2), it implies that
ĝEb = |orb(EdgesE )| − |orb(FacesE )| − |orb(VertsE )| + 2|orb(ccE )| ≥ 16 |orb(EdgesE )| − 2t,
i.e.,
|F| = 4|orb(EdgesE )| ≤ 48t + 24ĝEb.
Corollary 3.17. There are 2O((t+ĝ) log(t+ĝ)) nice t-boundaried embeddings of genus at most ĝ, and they can
be enumerated in time 2O((t+ĝ) log(t+ĝ)) .
3.4.2
Making an embedding nice
Let Eb = (E, t, L) be a t-boundaried embedding with E = (F, θ, σ, φ). Our goal now is to obtain an equivalent
nice embedding. To this end, we show that the following three operations lead to equivalent embeddings:
1. deleting an edge incident with an unlabelled size-2 vertex;
11
2. deleting an edge violating the last property of the definition of a nice embedding;
3. suppressing a size-4 unlabelled vertex that is not isolated.
We will henceforth call them simplifying operations.
Lemma 3.18 (deleting an edge incident to a size-2 unlabelled vertex). If e is an edge of Eb incident to an
unlabelled vertex of size 2, then Eb − e and Eb are equivalent.
Proof. We consider merges Eb and Eb − e with an arbitrary t-boundaried embedding Eb1 .
b Observation 3.13 yields that EbM − e is a merge of Eb − e and
First, let EbM be a merge of Eb1 and E.
Eb1 and Lemma 3.7 implies that its genus does not exceed the genus of EbM . Consequently, the genus of a
b
genus-minimum merge of Eb1 and Eb − e is not larger than the genus of a genus-minimum merge of Eb1 and E.
0
In the other direction, let EbM
be a merge of Eb − e and Eb1 . Let e = (x, y, y 0 , x0 ) where {x, x0 } is contained
in an unlabelled size-2 vertex so that pEb(x) is of the form (x0 , b). Consequently, Eb can be obtained from Eb − e
by drawing a new edge e at (⊥, b) for some b. By Observation 3.14, a merge EbM of Eb1 and Eb can be obtained
0
0
from EbM
by drawing a new edge at (a0 , b0 ). Since a = ⊥, we have a0 = ⊥, so the genera of EbM and EbM
are
b
b
equal due to Lemma 3.9. Consequently, the genus of a genus-minimum merge of E1 and E is not larger than
the genus of a genus-minimum merge of Eb1 and Eb − e. This completes the proof of the lemma.
Lemma 3.19 (deleting an edge violating the last property of the definition of nice embedding). Let e be
an edge in a t-boundaried embedding Eb that is incident to two faces. Furthermore, assume that one face f
incident with e has the following property: if x and y = θ(x) are the two flags shared between e and f , then
each flag z ∈
/ {x, φ(x), y, φ(y)} on f belongs to an unlabelled vertex. Then Eb is equivalent with Eb − e.
Proof. Again, we consider merges of Eb and Eb − e with an arbitrary t-boundaried embedding Eb1 .
In one direction, the proof is the same as in the proof of the previous lemma: Observation 3.13 and
Lemma 3.7 imply that the genus of a genus-minimum merge of Eb1 and Eb − e is not larger than the genus of
b
a genus-minimum merge of Eb1 and E.
0
b
In the other direction, let EM be a merge of Eb1 and Eb − e. We consider two cases depending on whether
f is of size 2. If so, then pE(x) is of the form (y, b), so Eb can be obtained from Eb − e by drawing an new edge
0
by drawing a
at (>, b) for some b. By Observation 3.14, a merge EbM of Eb1 and Eb can be obtained from EbM
0 0
0
0
b
b
new edge at (a , b ). Since a = >, we have a = >, so the genera of EM and EM are equal due to Lemma 3.9.
Next, suppose that f contains a flag z ∈
/ {x, y}. Since e is incident to 2 faces, we conclude that σ(x) and
σ(y) do not belong to e, and Eb can be obtained from Eb − e by drawing a new edge at (a, b) for a = σ(x) and
0
by drawing a new edge at
b = σ(y). By Observation 3.14, a merge EbM of Eb and Eb1 can be obtained from EbM
(a, b). Let the cycle of f be x, a, P, b, y for some (possibly empty) sequence of flags P . By the assumptions
0
of the lemma, every flag of P belongs to an unlabelled vertex. Consequently, in EbM
there exists a face f 0
whose cycle contains consecutive flags a, P, b on its cycle, as no orbit of φ or θ on a, P, b has been altered
This means that a new edge drawn at (a, b) is actually drawn at a face boundary. Hence, the genera of EbM
0
and EbM
are equal due to Lemma 3.9.
Consequently, the genus of a genus-minimum merge of Eb1 and Eb is not larger than the genus of a genusminimum merge of Eb1 and Eb − e, which concludes the proof of the lemma.
For the last basic operation, we need to formally define it. Let Eb = (E, t, L) be a t-boundaried embedding
with an unlabelled size-4 vertex v that is not isolated. Since v is not isolated, v is incident with two edges
e1 and e2 , and each ei is incident with v and a vertex vi 6= v. Note that it is possible that v1 = v2 .
Let x1 and x2 = φ(x1 ) be two flags in v such that xi belongs to ei . Furthermore, let yi = θ(xi ); note
that yi lies in ei . We delete all four flags of v, and replace the orbits of θ on e1 and e2 with {y1 , y2 } and
{σ(y1 ), σ(y2 )}. Clearly, we have replaced e1 and e2 with a new edge with flags y1 , y2 , σ(y1 ), σ(y2 ). We now
formally verify that the output embedding is an equivalent one.
12
Lemma 3.20 (suppressing a size-4 unlabelled vertex that is not isolated). The operation of suppressing a
size-4 unlabelled vertex leads to an equivalent embedding.
Proof. We interpret the suppressing operation as a sequence of one edge drawing and two edge deletions.
Observe that y1 , x1 , x2 , y2 are four distinct flags that lie on the same face f and, furthermore, they are
consecutive in this order along the cycle of f . Let us draw a new edge ef along the boundary of f , at y1
and y2 , obtaining an embedding Ebf . Note that in Ebf there is a new face f 0 with cycle y1 , x1 , x2 , y2 , z2 , z1 ,
where {z1 , z2 } is a new orbit of θ contained in the edge ef . Furthermore, the edge ef with face f 0 fulfills the
b Consequently, Ebf and Eb are equivalent.
assumptions of Lemma 3.19, and its deletion from Ebf gives E.
b
Then, we delete edges e1 and e2 from Ef , obtaining an embedding Eb0 . Our goal is to show that such an
operation leads to an equivalent embedding; note that if we rename the flags of ef to y1 , y2 , σ(y1 ), σ(y2 ) we
obtain the output embedding of the suppressing operation. The proof is similar to that of Lemma 3.19, but
without most of the special cases due to the existence of the edge ef .
Let Eb1 be an arbitrary t-boundaried embedding. By Observation 3.13 and Lemma 3.7 yields that the
genus of a genus-minimum merge of Eb1 and Eb0 is not larger than the genus of a genus-minimum merge of Eb1
and Ebf .
0
0
For the other direction, let EbM
= (EM
, t, L0M ) be a merge of Eb1 and Eb0 . Note that pEbf (x1 ) = (x2 , z1 ) and
pb
(x2 ) = (x0 , z2 ). Hence, Ebf can be retrieved from Eb0 by drawing edge e2 at (⊥, z2 ) and then edge e1 at
Ef −e1
2
0
(x2 , z1 ). By Observation 3.14, a merge EbM of Ebf and Eb1 can be retrieved from EbM
by drawing an edge e2
at (⊥, z2 ) and then e1 at (x2 , z1 ). Lemma 3.9 yields that the first of these operations asserts that the genus
remains unchanged. As for the second operation, we observe after drawing e2 , we have that x2 , y2 , z2 , z1 is
a fragment of face boundary and that e1 is drawn along it. Hence, Lemma 3.9 also implies that the genera
0
and EbM are equal. We infer that the genus of a genus-minimum merge of Eb1 and Ebf is not larger than
of EbM
the genus of a genus-minimum merge of Eb1 and Eb0 . This concludes the proof of the lemma.
Armed with the three basic operations, we are now ready to prove the following statement.
Lemma 3.21. There exists a polynomial-time algorithm that, given an arbitrary t-boundaried embedding,
computes an equivalent nice one by successive applications of the simplifying operations.
b Otherwise, observe that
Proof. Let Eb be a t-boundaried embedding. If Eb is nice, we can just return E.
any object that violates the niceness of Eb gives rise to an application of one of the simplifying operations.
b the resulting embedding is
An edge violating the last property of a nice embedding can be deleted from E;
equivalent due to Lemma 3.19. Similarly, if there exists an unlabelled size-2 vertex, then Lemma 3.18 allows
us to safely delete the incident edge, and a size-4 unlabelled vertex that is not isolated can be also suppressed
by Lemma 3.20.
Finally, note that each execution of a simplifying operation strictly decreases the number of flags in the
embedding. Consequently, in polynomial time we obtain an equivalent nice embedding.
4
Bounded treewidth graphs
In this section we prove Theorem 1.2. Without loss of generality, we can assume that the input tree
decomposition (T, β) of the input graph G is a nice tree decomposition of width less than k, that is, every
bag of (T, β) is of size at most k.
We first refine (T, β) as follows. First, with every node t we associate a graph G(t), initialized as
G(t) := G↓ (t). In the refinement process, we will maintain the invariant that at every node t, the graph
G(t) is a subgraph of G[α(t)] and a supergraph of G↓ (t); in particular, V (G(t)) = V (G↓ (t)). Note that one
needs only O(t2 ) bits to store G(t) at every node t, as one needs only to store which edges of G[β(t)] belong
to G(t).
For every forget node t we perform the following refinement process. Let t0 be the child of t, and let v be
the forgotten vertex, i.e., {v} = β(t0 ) \ β(t). Let e1 , e2 , . . . , e` be the edges incident with v that have their
13
second endpoint in β(t); note that {e1 , e2 , . . . , e` } are exactly the edges that are present in G↓ (t) but are
not present in G↓ (t0 ). Our goal is to refine the edge tt0 of T by inserting a number of vertices on this edge
so that the transition from G↓ (t0 ) to G↓ (t) is more smooth.
More precisely, we subdivide the edge tt0 ` times, inserting vertices t1 , t2 , . . . , t` in this order, such that t1
is adjacent with t, and t` is adjacent with t0 . Furthermore, to every node ti add a second child si that is a leaf
of the tree T . The new node si is of a new type, namely edge leaf : we set β(si ) = β(t0 ), V (G(si )) = β(t0 ),
and E(G(si )) = {ei }. The node ti is a join node with G(ti ) = G↓ (t0 ) − {e1 , e2 , . . . , ei−1 }.
This completes the description of the tree decomposition refinement. By this step, we have obtained the
following properties: at every introduce or forget node t with child t0 we have G(t) = G(t0 ), while at every
join node t with children t1 and t2 we have E(G(t)) = E(G(t1 )) ] E(G(t2 )). Furthermore, note that we have
introduced two nodes for every edge of G, giving O(nk) new nodes in total.
We compute a labelling function Λ : V (G) → [k] such that Λ is injective on every bag of (T, β); such a
labelling Λ is straightforward to compute in a top-to-bottom fashion, using the fact that every bag of (T, β)
is of size at most k.
With the refined tree decomposition, we perform a bottom-up dynamic programming algorithm. Fix a
node t ∈ V (T ). For a subset A ⊆ α(t) and an embedding E of G(t)[A], we can naturally equip E with a
structure of a k-boundaried embedding by assigning labels Λ|β(t) (which is an injective function). We will
denote such a k-boundaried embedding at node t as bnd(t, E)
At t, we maintain the following DP table. For every set X ⊆ β(t) and for every nice k-boundaried
b ∈
embedding Eb of genus at most g that uses only labels of Λ(β(t) \ X), we remember a value T[t, X, E]
b to indicate the minimum size of a deletion
{0, 1, . . .} ∪ {+∞}. Intuitively, we would like the value T[t, X, E]
b However, for sake of
set Y ⊆ α(t) \ β(t) such that G(t) − (X ∪ Y ) can be embedded equivalently with E.
the formal argument, we strip the above intuition into only its essential parts, and define the following two
properties that we maintain:
b 6= +∞, then for every k boundaried embedding
(a finite value yields small deletion set) If T[t, X, E]
b
E0 that uses only labels of Λ(β(t) \ X) and for every merge Eb0M of Eb and Eb0 , there exists a set
b an embedding E of G(t) − (X ∪ Y ), and a merge EbM of
Y ⊆ α(t) \ β(t) of size at most T[t, X, E],
b
bnd(t, E) and E0 of genus not larger than the genus of Eb0M .
(a good deletion set is represented in the table) for every set Y ⊆ α(t) \ β(t), every embedding E of
G(t) − (X ∪ Y ), every k-boundaried embedding Eb0 that uses only labels of Λ(β(t) \ X), and every merge
Eb0M of bnd(t, E) and Eb0 , there exists a nice k-boundaried embedding Eb and a merge EbM of Eb and Eb0
b ≤ |Y | and the genus of EbM is not larger than the genus of Eb0M .
such that T[t, X, E]
The above invariants include only one inequality from the definition of equivalence: the one that is needed
for the proof of the correctness of the algorithm.
In particular, since the root node r has an empty bag, the minimum value of T[r, ∅, ·] at the root node is
the size of the minimum solution to Genus Vertex Deletion on (G, g). Indeed, by the second invariant,
an optimum solution Y ∗ and the corresponding embedding E ∗ has its corresponding entry Eb∗ when merged
with an empty embedding Eb0 (so that Eb0M = bnd(r, E) = (E, k, ∅)), and we have |Y ∗ | ≥ T[r, ∅, Eb∗ ]. In the
b and again an empty embedding
other direction, the first invariant ensures that for every entry T[r, ∅, E]
b there exists a deletion set Y of size at most T[r, ∅, E]
b and an
Eb0 and the only possible merge Eb0M = E,
b
b
embedding E of G − Y such that bnd(r, E) has genus at most the genus of E0M = E, which is at most g.
b in a bottom-to-up fashion.
We now show how to compute the values T[t, X, E]
The computation is trivial at leaf nodes (empty bag, empty graph G(t)) and at edge leaf nodes (a single
edge with distinct endpoints). Furthermore, observe that at introduce nodes, we just copy the corresponding
entries in T, as the new vertex is isolated in G(t) and the notion of an embedding ignores isolated vertices.
Thus, we are left with tackling forget and join nodes.
b with a value a: we set T[t, X, E]
b :=
To this end, we define an operation of updating a cell T[t, X, E]
b
max(a, T[t, X, E]). We say that the update was successful if the stored value changed.
14
4.1
Forget nodes
Let t be a forget node with child t0 and let v be the forgotten vertex. Note that, due to the refinement step,
we have G(t0 ) = G(t). The only difference between the nodes t and t0 is that if we consider some embedding
E of a subgraph of G(t), then bnd(t0 , E) assigns a label Λ(v) to the vertex v, while in bnd(t, E) the vertex v
remains unlabelled.
b to +∞.
This is straightforward to implement in our DP tables. We start by setting all values T[t, X, E]
0
0
b
b
b
For every cell T[t , X, E] with v ∈ X, we update the cell T[t, X \ {v}, E] with value T[t , X, E] + 1: when
v is deleted, the embedding does not change, but we need to account for the vertex v in the value of the
cell. For every cell T[t0 , X, Eb0 ] with v ∈
/ X, we create an embedding Eb from Eb0 by first forgetting the label of
the vertex v (so it becomes an unlabelled one), and then by applying Lemma 3.21 to obtain an equivalent
b with value T[t0 , X, Eb0 ].
embedding. We update the cell T[t, X, E]
We now formally verify that the entries for node t satisfy the required properties.
b that is less than +∞, a k-boundaried embedding Eb0 and a merge
First property. Take a cell T[t, X, E]
0
0 b0
b
b
b
b
E0M of E and E0 . Let T[t , X , E ] be the last cell that caused an update in the value of T[t, X, E].
0
0
0
b By the first property, applied inductively to the
First, assume v ∈ X ; then X = X ∪ {v} and Eb = E.
b the embedding Eb0 and the merge Eb0M , we obtain a set Y 0 of size at most T[t0 , X 0 , E],
b an
cell T[t0 , X 0 , E],
0
0
0
0
0
0
0
b
b
embedding E of G(t ) − (X ∪ Y ) and a merge EM of bnd(t , E ) and E0 of genus at most the genus of Eb0M .
b Consequently, E 0 is an embedding
Set Y = Y 0 ∪ {v}; then X ∪ Y = X 0 ∪ Y 0 while |Y | = |Y 0 | + 1 ≤ T[t, X, E].
0
of G(t) − (X ∪ Y ), and together with Y and EbM
satisfies the requirements of the first property.
0
In the second case, we have v ∈
/ X , and X = X 0 . Let Eb◦ be the embedding Eb0 with the label of v
b
dropped; recall that E is a result of application of Lemma 3.21 to the embedding Eb◦ . Since Eb◦ is equivalent
b the genus-minimum merge Eb◦ of Eb◦ and Eb0 has genus not larger than the genus of Eb0M . Since neither
to E,
M
◦
◦
0
with additionally label Λ(v) assigned to the vertex
to be EbM
Eb nor Eb0 uses the label Λ(v), we can define EbM
v; note that this operation does not change the genus of the embedding. As Eb0 does not use the label Λ(v),
0
0
, yields
is a merge of Eb0 and Eb0 . The first property, applied to the cell T[t0 , X 0 , Eb0 ], Eb0 , and the merge EbM
EbM
0
0
0
0
0
0
existence of a set Y ⊆ α(t ) \ β(t ), an embedding E of G(t ) − (X ∪ Y ), and a merge EbM of bnd(t0 , E) and
0
Eb0 of genus at most the genus of EbM
. Since G(t0 ) = G(t), X 0 = X, and Eb0 does not use the label Λ(v), EbM
0
b
is a merge of bnd(t, E) and E0 . Furthermore, since the genus of EbM
is at most the genus of Eb0M , the tuple
0
(Y , E, EbM ) fulfills the first property.
Second property. Take a set Y , E, Eb0 , and Eb0M as in the statement of the second property. If v ∈ Y ,
take Y 0 = Y \ {v} and X 0 = X ∪ {v}, and otherwise take Y 0 = Y and X 0 = X. In both cases, we have
X ∪ Y = X 0 ∪ Y 0 , X 0 ⊆ β(t0 ), and Y 0 ⊆ α(t0 ) \ β(t0 ).
Since Eb0 does not use the label Λ(v), Eb0M is also a merge of bnd(t0 , E) and Eb0 . By the second property
0
for t0 , X 0 , Y 0 , Eb0 , and Eb0M , we obtain a k-boundaried embedding Eb0 and a merge EbM
of Eb0 and Eb0 of genus
0
0 b0
0
0
b
at most the genus of E0M such that T[t , X , E ] ≤ |Y |. While processing the cell T[t , X 0 , Eb0 ] the algorithm
b for Eb being a result of an application of Lemma 3.21 to the embedding
attempts an update on a cell T[t, X, E]
◦
0
b
b
E being the embedding E with the label of v dropped. A direct check shows that in regardless of whether
b ≤ |Y |.
v belongs to Y or not, it follows that T[t, X, E]
◦
0
◦
b
b
Let EM be the embedding EM with the label Λ(v) dropped if present; note that the genera of EbM
and
0
◦
◦
EbM are equal. Since the embedding Eb0 does not use the label Λ(v), the merge EbM is a merge of Eb and Eb0 .
b the genus-minimum merge EbM of Eb and Eb0 is of genus at most the genus of
By the equivalence of Eb◦ and E,
◦
b
EM , which in turn is at most the genus of Eb0M . Thus, Eb and EbM fulfill the second property for Y , E, Eb0 , and
Eb0M .
This finishes the description and the proof of correctness of the computations at forget nodes.
15
4.2
Join nodes
Let t be a join node with children t1 and t2 . Note that we have V (G(t1 )) ∩ V (G(t2 )) = β(t), V (G(t1 )) ∪
V (G(t2 )) = V (G(t)) and E(G(t)) = E(G(t1 )) ] E(G(t2 )).
We iterate over every set X ⊆ β(t) and every pair of cells T[t1 , X, Eb1 ] and T[t2 , X, Eb2 ]. Note that both Eb1
and Eb2 are k-boundaried embeddings with labels on vertices of β(t) \ X. For every k-boundaried embedding
Eb0 that is a merge of Eb1 and Eb2 , we use Lemma 3.21 to compute a nice embedding equivalent to Eb0 , and
b with T[t1 , X, Eb1 ] + T[t2 , X, Eb2 ].
update T[t, X, E]
We now formally verify that the entries for node t satisfy the required properties.
b that is less than +∞, a k-boundaried embedding Eb0 and a merge
First property. Take a cell T[t, X, E]
b
b
b
b comes from considering cells
E0M of E and E0 . Let g0 be the genus of Eb0M . Assume the value of T[t, X, E]
0
T[ti , X, Ebi ] and a merge Eb of Eb1 and Eb2 .
b the genus-minimum merge EbA of Eb0 and Eb0 is of genus at most g0 . Since
By the equivalence of Eb0 and E,
M
0
A
b
b
b
E is a merge of E1 and E2 , if we delete the edges of Eb2 from the embedding EbM
, we obtain an embedding Eb1A
A
of not larger genus that is a merge of Eb1 and Eb0 . Furthermore, observe that EbM is a merge of Eb1A and Eb2 .
A
We apply the first property to the cell T[t2 , X, Eb2 ] with Eb1A and EbM
, obtaining a set Y2 , an embedding
A
B
b
b
E2 , and a merge EM of bnd(t2 , E2 ) and E1 with genus at most g0 . Recall that Eb1A is a merge of Eb1 and Eb0 .
B
we obtain an embedding Eb2B that is a merge of bnd(t2 , E2 )
Consequently, if we delete the edges of Eb1 from EbM
B
b
b
and E0 . Furthermore, observe that EM is a merge of Eb2B and Eb1 .
B
, obtaining a set Y1 , an embedding
We now apply the first property to the cell T[t1 , X, Eb1 ] with Eb2B and EbM
B
B
b
b
b
E1 , and a merge EM of bnd(t1 , E1 ) and E2 of genus at most g0 . Since E2 is a merge of bnd(t2 , E2 ) and Eb0 , if
we delete the edges of Eb0 from EbM , we obtain an embedding Eb12 that is a merge of bnd(t1 , E1 ) and bnd(t2 , E2 ).
Let E be the embedding underlying Eb12 and let Y = Y1 ∪ Y2 . Then, E is an embedding of G(t) − (X ∪ Y )
such that EbM is a merge of bnd(t, E) with Eb0 . Since the genus of EbM is at most g0 , the proof of the first
property is finished.
Second property. Take a set Y , E, Eb0 , and Eb0M as in the statement of the second property, and let g0 be
the genus of Eb0M . For i = 1, 2, take Yi = Y ∩ α(ti ) and Ei to be the embedding E restricted to the edges of
G(ti ) (i.e., with the edges of G(t3−i ) deleted). Note that bnd(t, E) is a merge of bnd(t1 , E1 ) and bnd(t2 , E2 ).
First, let Eb1A be the embedding Eb0M with the edges of E2 deleted. Observe that Eb1A is a merge of bnd(t1 , E1 )
and Eb0 , while Eb0M is a merge of Eb1A and bnd(t2 , E2 ). By the second property, applied to t2 , X, Y2 , E2 , Eb1A ,
B
and Eb0M , we obtain a nice k-boundaried embedding Eb2 and a merge EbM
of Eb2 and Eb1A of genus at most g0
b
such that T[t2 , X, E2 ] ≤ |Y2 |.
B
Now, if we delete the edges of E1 from EbM
, we obtain an embedding Eb2B . Observe that Eb2B is a merge of
B
is a merge of bnd(t1 , E1 ) and Eb2B . We apply the second property to t1 , X, Y1 , E1 , Eb2B ,
Eb0 and Eb2 , while EbM
B
b
and EM , obtaining a nice k-boundaried embedding Eb1 and a merge EbM of Eb1 and Eb2B of genus at most g0
such that T[t1 , X, Eb1 ] ≤ |Y1 |.
Let Eb be the embedding EbM with the edges of Eb0 deleted. Since EbM is a merge of Eb1 and Eb2B , which
in turn is a merge of Eb2 and Eb0 , we have that Eb is a merge of Eb1 and Eb2 and EbM is a merge of Eb and Eb0 .
b with the value
Consequently, the algorithm attempts an update on the cell T[t, X, E]
T[t1 , X, Eb1 ] + T[t2 , X, Eb2 ] ≤ |Y1 | + |Y2 | ≤ |Y |.
Since the genus of EbM is at most g0 , the pair Eb and EbM fulfills the second property for Y , E, Eb0 , and Eb0M .
4.3
Summary
We have concluded description and correctness proofs of the operations at the nodes of the tree decomposition. At every node t, a straightforward implementation takes time polynomial in the number of entries
16
b which, by Corollary 3.17, is 2O((k+g) log(k+g)) . This finishes the proof of Theorem 1.2.
T[t, X, E],
5
Irrelevant vertex
In this section, we prove Theorem 1.3:
Theorem 1.3. For every integer g there exists a constant Cg such that the following holds. Given a graph
G, an integer k, and a set M ⊆ V (G) such that G − M is embeddable into a surface of genus at most g, one
can in time Cg nO(1) find one of the following:
1. a tree decomposition of G of width at most Cg |M |1/2 k 3/2 ;
2. a vertex w ∈ V (G) such that every solution to the Genus Vertex Deletion instance (G, g, k)
contains w; or
3. a vertex v ∈ V (G) such that (G, g, k) is a yes-instance to Genus Vertex Deletion if and only if
(G − {v}, g, k) is.
The argumentation is heavily inspired by the corresponding planar case by Marx and Schlotter [10]. As
in most irrelevant vertex arguments, we follow typical outline:
1. Run an approximation algorithm for treewidth and, in case it returns that the treewidth of the graph
is larger than the required threshold, find a large grid minor. Here, the linear dependency on the grid
size and treewidth in bounded genus graphs is known, and one can find the corresponding grid minor
efficiently.
2. Show that a vertex w ∈ M that is connected to many places in the grid that are far apart needs to be
included in every solution, exactly as it is in the planar case. In the absence of such a vertex, a large
part of the grid minor is flat, that is, embeds planarly and does not have any internal connections to
the modulator M .
3. Prove that a middle vertex of such a flat part is irrelevant. Here, the challenge is to argue that in
every solution there exists an embedding of the remaining part that draws the flat part indeed in a flat
manner.
We repeatedly use the following auxiliary result throughout this section:
Lemma 5.1 S
([3, Lemma B.6]). Let Γ be a surface of Euler genus g and let C be a set of g + 1 disjoint circles
in Γ. If Γ \ C has a component D0 whose closure in Γ meets every circle in C, then at least one of the
circles in C bounds a disc in Γ that is disjoint from D0 .
Let G, k, g, and M be such as in the statement of Theorem 1.3. We start by computing an embedding
E0 of G − M into a surface of Euler genus at most g, using either the algorithm of Kawarabayashi, Mohar,
and Reed [7] or the older algorithm of Mohar [11].1 This takes time Cg1 n for some constant Cg1 depending
only on g.
5.1
Finding a large grid
We apply a constant-factor approximation algorithm for treewidth and largest excluded grid minor in graphs
with a fixed excluded minor [2, Theorem 3.11]. The algorithm does not need to compute a near-embedding
of G, as E0 serves this purpose. Thus, the algorithm as described in [2, Theorem 3.11] runs in time Cg2 nO(1)
for some constant Cg2 depending only on g, and is a Cg3 -approximation.
If the resulting tree decomposition is of width at most Cg4 |M |1/2 k 3/2 for some sufficiently large constant
4
Cg , then we directly return it. Otherwise, by choosing the constant Cg4 appropriately, we get a grid minor
1 The
work [7] appeared so far only as an extended abstract in conference proceedings.
17
Figure 2: A wall of sidelength 4 and a subdivision of this wall.
of sidelength Θ(|M |1/2 (k + g)3/2 g 1/2 ). By standard arguments, we henceforth focus on the case when we
get a wall W0 of sidelength `0 as a subdivision in G − M , where `0 = Θ(|M |1/2 (k + g)3/2 g 1/2 ) can be chosen
arbitrarily large; see Figure 2 for the definition of a wall.
√
In the wall W0 , we identify g + 1 pairwise disjoint subwalls of sidelength `1 for every `1 = Θ(`0 / g) =
Θ(|M |1/2 (k + g)3/2 ), leaving enough space between the subwalls so that the part of W0 not contained in any
of the subwalls is connected. By Lemma 5.1, for at least one of the identified subwalls, its surrounding cycle
bounds a disc in the considered embedding of G − M . Consequently, at least one of the identified subwalls
is planarly embedded in E0 . We denote this subwall as W1 and the part of G − M embedded in the disc
separated by the surrounding cycle of W1 by G1 . In what follows, we mostly focus on G1 and W1 .
5.2
Filtering out the modulator neighbors
We now show that if a vertex w ∈ M is adjacent to many scattered vertices of G1 , then it needs to be
included in every solution. We start with constructing a canonical graph of Euler genus larger than g.
Lemma 5.2. For an integer ` > 1, let B` be a graph defined as follows. We take ` disjoint copies of K5−
(i.e., the five-vertex clique K5 with an edge deleted), and denote by vi and wi the endpoints of the removed
edge in the i-th copy. We introduce two further vertices v and w and make v adjacent to all vertices vi and
w adjacent to all vertices wi . Then, the Euler genus of B` is at least `/2.
Proof. Consider an embedding of B` into a surface Γ of Euler genus g. For i = 1, 2, . . . , `, let γi be the
triangle in the i-copy of K5− that does not pass through vi or wi , treated as a closed curve on Γ. Since Γ
has Euler genus g, by Lemma 5.1, at least ` − g of the curves γi must be contractible.
Consider an index i for which γi is contractible, and let V (γi ) be the three vertices of G that lie on γi .
Since G − V (γi ) is connected, γi is a boundary of a face of the embedding of G. Augment the graph G by
adding a new vertex xi inside this face, and connect it to the three vertices of V (γi ).
In this manner, the i-th copy of K5− , together with the new vertex xi , forms a nonplanar graph (a
supergraph of K3,3 ). Consequently, we cannot perform this operation in parallel for more than g indices i.
We infer that ` ≤ 2g, and the lemma is proven.
Recall that the boundary cycle of W1 encloses a disc containing W1 in the considered embedding E0 of
G − M , and G1 is the part of the graph G − M that is enclosed by this boundary cycle. For a vertex
v ∈ V (G1 ) and an integer `, the radius-` ball B(v, `) around v is defined as follows: we take all faces f of
W1 (excluding the infinite face) that contain v either inside or on the boundary, mark all faces that lie in
face-vertex distance (excluding traversal through the infinite face) at most ` from one of these faces f and
put into B(v, `) all vertices of G1 that are contained inside or on the boundary of marked faces.
For every w ∈ M , we create a set I(w) ⊆ V (G1 ) as follows. We start with I(w) = ∅ and every vertex
of V (G1 ) unmarked. As long as there exists an unmarked neighbor v of w in G1 , we insert v into I(w) and
mark all vertices of B(v, `2 ) for `2 = 100(k + g + 1). We claim that if I(w) is too large, the vertex w needs
to be deleted in any solution to Genus Vertex Deletion.
18
Lemma 5.3. Let w ∈ M be a vertex for which |I(w)| > k + 2g + 1. Then w belongs to every solution to
Genus Vertex Deletion instance (G, g, k).
Proof. Suppose the contrary, and let S be a solution that does not contain w. That is, |S| ≤ k, and G − S
admits an embedding E into a surface of Euler genus at most g.
Consider subgraphs B(v, `2 /4) for v ∈ I(w). By the construction of I(w), these subgraphs are vertexdisjoint. Furthermore, as `2 /4 is much larger than k, for every v1 , v2 ∈ I(w), v1 6= v2 , there exists a set
P(v1 , v2 ) of k + 1 vertex-disjoint paths in G1 connecting the outer boundary of B(v1 , `2 /4) with the outer
boundary of B(v2 , `2 /4), without any internal vertex in any of the subgraphs B(v, `2 /4) for v ∈ I(w). Let
I 0 be the set of these vertices v ∈ I(w) such that S is disjoint with V (B(v, `2 /4)); since |S| ≤ k, we have
I 0 > 2g + 1.
Construct a minor H of G − S as follows. Pick some arbitrary v0 ∈ I 0 and contract B(v0 , `2 /4) onto v0 .
For every v ∈ I 0 \ {v0 }, contract onto v0 a path of P(v0 , v) that is disjoint with S. For every v ∈ I 0 \ {v0 },
contract B(v, `2 /4) into a K5− (recall that B(v, `2 /4) contains a large wall, being part of W1 ) such that one
of the two nonadjacent vertices is adjacent to w and the other to v0 . We have obtained that B2g+1 (as
defined in Lemma 5.2) is a minor of G − S, a contradiction with Lemma 5.2.
Consequently, if there exists w ∈ M with |I(w)| > k + 2g + 1, then we can return w as the second result
of the algorithm of Theorem 1.3. Henceforth we will assume that |I(w)| ≤ k + 2g + 1 for every w ∈ M .
Recall that W1 is a wall of sidelength `1 , where `1 = Θ(|M |1/2 (k + g)3/2 ) can be chosen arbitrarily large.
This lets us identify n3 disjoint subwalls
S of sidelength `3 = 100(k + g), where n3 = Θ(|M |(k + g)) can be
chosen
arbitrarily
large.
Note
that
|
w∈M I(w)| = O(|M |(k + g)) and that all neighbors of M are located
S
in w∈M,v∈I(w) B(v, `2 ). Each ball B(v, `2 ) may intersect only a constant number of the identified subwalls
of sidelength `3 . Hence, if n3 = Θ(|M |(k + g)) is sufficiently large, there is a subwall W3 of sidelength
`3 = 100(k + g) such that no vertex of G3 , defined as the part of G1 that is enclosed by the outer boundary
of W3 , is a neighbor of a vertex in M . Note that only the vertices on the outer boundary of G3 may have
neighbors in G − V (G3 ). With G3 and W3 , we proceed to the next section.
5.3
Middle vertex of a flat part is irrelevant
Let v be a vertex on the middle face of W3 . We show that v is irrelevant, that is, it can be returned as
the outcome of Theorem 1.3. Clearly, if (G, g, k) is a yes-instance, then so is (G − {v}, g, k). In the other
direction, suppose that there exists S ⊆ V (G) \ {v} such that G − {v} − S admits an embedding E into a
surface of Euler genus at most g. We would like to enhance this embedding so that it also accommodates v.
The sidelength `3 of W3 is large enough to find in W3 a subdivision of a circular wall W4 of height
h4 = 5(k + g) + 9 and circumference `4 = min(3, k + 1) so that v is located inside the inner cycle of the wall;
see Fig. 3 for the definition of a circular wall. Next, in W4 we identify k + g + 2 rings (which are concentric
circular walls of height 5 and circumference `4 ), separated by layers of height 1; see Fig. 3. Let R0 be the
family of identified rings.
For a ring R ∈ R0 , we say that the central circle of R is the circle between the two middle layers of the
ring, and the boundary circles are the circles separating R from the remainder of W4 . A face of R in the
embedding E0 is small if it is not the outerface nor the face inside the innermost cycle, and central if it is
incident to the central circle. The territory of R, denoted by T (R), is the subgraph of G that consists of R
and everything that is drawn in the embedding E0 in the small faces of R. Note that T (R) is a planar graph,
and the subgraphs {T (R) : R ∈ R0 } are vertex-disjoint.
Let R ⊆ R0 be the family of these rings for which T (R) is disjoint with S. As |R0 | = k + g + 2 and the
subgraphs T (R) are vertex-disjoint, we have |R| ≥ g + 2.
We say that a ring R is embedded plainly in the embedding E if for every central face f of R, the cycle
C f that surrounds f in E0 , is a two-sided cycle in E that bounds a disc on one side, and the graph R − V (C f )
is drawn on the other side. The following lemma is an easy corollary of Lemma 5.1:
Lemma 5.4. There exists R ∈ R that is embedded plainly in E.
19
Figure 3: A circular wall of height 9 and circumference 12 containing two rings. Their central circles are
marked with red squares.
20
e2
P
f
e
e1
Figure 4: Illustration for the proof of Corollary 5.5.
Proof. Suppose the contrary. For every ring R ∈ R, let C(R) be the cycle around a small
face of R that
S
violates the definition of a plainly embedded ring. Consider a graph G0 = G − M − S − R∈R V (C(R)). We
shall prove that it has a large connected component D which contains R0 := R \ V (C(R)) for each R ∈ R.
First, note that R0 is itself connected and the boundary circles of R of are preserved in R0 . Thus, it suffices
to prove that for every two consecutive rings R1 , R2 ∈ R, the inner boundary circle of R2 is connected to
the outer boundary circle of R1 . Observe that the part of W4 between these circles is a circular wall of
circumference `4 and some positive height. Thus, there are
S `4 ≥ k + 1 vertex-disjoint paths between the two
boundary circles. These paths are disjoint with M and R∈R V (C(R)), and at least one of them must be
disjoint with S. Hence, one of these paths is preserved in G0 .
We conclude that the claimed component D indeed exists. It is adjacent to every cycle C(R), so
Lemma 5.1, due to our assumption on how cycles C(R) are embedded in E, yields that the Euler genus
of the embedding E is at least |R|, which is more than g.
Let R ∈ R be a plainly embedded ring, and let CR be the central circle of R. A direct corollary of the
definition of a plainly embedded ring is the following.
Corollary 5.5. In the embedding E, CR is a two-sided cycle, and its incident edges of R are partitioned
between the sides of CR exactly as in the embedding E0 .
Proof. Let e1 and e2 be two edges of R \ E(CR ) that are incident to CR that are two consecutive edges on
the same side of CR in E0 . Furthermore, let P be the path between e1 and e2 in CR that is not incident to
any other edge of R \ E(CR ) on the same side as e1 , and let f be the face of the embedding of R in E0 that
is incident with e1 , e2 , and P ; see also Figure 4. Let e be the other edge of R \ E(CR ) incident to P (the one
that is drawn on the opposite side of CR than e1 in E0 ). Since R is plainly embedded in E, the embedding
E restricted to R has a face surrounded by C f , and thus, as we traverse P in E, the edges e1 and e2 are on
one side, and e is on the other side. Since the choice of e1 , e2 , and f is arbitrary, the claim follows.
Corollary 5.5 allows us to speak about the inner and outer side of CR in E: the sides of CR that contain
incident edges inside CR and outside CR in E0 .
A bridge is a connected component C of G − V (R), together with the edges joining C with V (R).
Furthermore, an edge e ∈
/ E(R) with both endpoints in V (R) is also a bridge on its own. For a bridge B,
the vertices of V (B) ∩ V (R) are attachment points.
A bridge B is central if it has at least one attachment point, but all its attachment point lie in V (CR ).
Note that, since R is contained in W3 , a central bridge is disjoint with M and in the embedding E0 it is
drawn inside one of the small faces of R incident with CR . In particular, a central bridge is a subgraph of
the territory of R.
Let H be the subgraph of G − M that consists of: the part of G − M that is enclosed on the same side of
CR in the embedding E0 as the vertex v (i.e., on the disc, flat side of CR ), together with all central bridges.
In other words, H is a subgraph of G − M induced by the vertices of CR and all connected components
of G − M − V (CR ) that are contained in the same connected component of G − M as CR , except for the
21
connected component of G − M − V (CR ) that contains the outermost concentric cycle of R. The boundary
of H, denoted ∂H, is the set of vertices of H that have incident edges of G that do not belong to H.
Armed with these observations, we now modify the embedding E of G − S − {v} as follows. First, we cut
the surface along the cycle CR , and cap with discs the two resulting holes. This operation can only lead to a
surface of a lower Euler genus. Second, we remove all edges of H from E, and all vertices of H that become
isolated by this operation. Observe that due to Corollary 5.5, now in E there is a face f , containing one of
the discs glued to CR (the one glued on the inner side of CR ).
Note that if we restrict the embedding E0 to H, we obtain a plane embedding EH of H with an additional
property that every vertex of ∂H lies on the outerface.
Consider a bridge B that has an attachment point in V (CR ), but is not central (i.e., is not a subgraph of
H). Note that due to the fact that R is plainly embedded, in E the bridge B needs to be drawn in the same
face of R as in the embedding E0 . Thus, the order of the vertices of ∂H on the outerface of EH is exactly
the same as the order of these vertices in the modified embedding E 0 . Consequently, the embedding EH of
H can be glued into f , identifying correspondingly the vertices of V (H) that remain in E. In this manner,
we obtain an embedding E 0 of G − (S \ V (H)), concluding the proof that v is irrelevant.
This finishes the proof of Theorem 1.3.
6
Conclusions
In this work we have developed fixed-parameter algorithms for the Genus Vertex Deletion problem with
solution size and treewidth parameterizations, putting particular effort into optimizing the dependency on
the treewidth parameter in the running time bound.
We remark that, although our formal statement of Genus Vertex Deletion involves only bounding
the Euler genus of the output graph, only minor changes to our algorithms are required if one demands the
final graph to be embeddable in an orientable surface of some genus. In terms of combinatorial embeddings,
studied in Section 3, an embedding is orientable if the set of flags can be partitioned into two parts (called left
and right) such that every orbit of σ, θ, and φ contains two flags from different sets. The crucial observation
is that deleting an edge, drawing an edge along a face boundary, and suppressing a size-4 vertex that is not
isolated, applied to an orientable embedding results in an embedding that is also orientable. Consequently,
if we allow only orientable embeddings in the dynamic programming algorithm of Section 4, we obtain the
desired variant of Theorem 1.2 for orientable surfaces. Finally, the arguments of Section 5 operate in the
language of modifying an embedding in a fixed surface, and therefore yield also without any changes a variant
of Theorem 1.3 for orientable surfaces.
We would like to conclude with two open questions, stemming from our research.
First: Can we obtain a 2O(Cg k log k) n-time algorithm, following the ideas of [5] for the planar case? Our
bounded treewidth routine suits such an algorithm, but the irrelevant vertex argument does not.
Second, and more challenging: what can we say about possible dependency on the parameter k for the
problem of deleting k vertices to an arbitrary minor-closed graph family? A similar question can be asked
for the parameter treewidth. Here, the main challenge is that it is harder to certify being H-minor-free for
an arbitrary graph H, while one can certify being of bounded genus by giving a corresponding embedding.
References
[1] Marek Cygan, Fedor V. Fomin, Lukasz Kowalik, Daniel Lokshtanov, Dániel Marx, Marcin Pilipczuk, Michal
Pilipczuk, and Saket Saurabh. Parameterized Algorithms. Springer, 2015. doi:10.1007/978-3-319-21275-3.
[2] Erik D. Demaine, Mohammad Taghi Hajiaghayi, and Ken-ichi Kawarabayashi. Algorithmic graph minor theory:
Decomposition, approximation, and coloring. In 46th Annual IEEE Symposium on Foundations of Computer
Science, FOCS 2005, pages 637–646. IEEE Computer Society, 2005. doi:10.1109/SFCS.2005.14.
[3] Reinhard Diestel. Graph Theory, volume 173 of Graduate Texts in Mathematics. Springer-Verlag, Heidelberg,
5th edition, August 2016. URL: http://diestel-graph-theory.com/.
22
[4] Russell Impagliazzo, Ramamohan Paturi, and Francis Zane. Which problems have strongly exponential complexity? Journal of Computer and System Sciences, 63(4):512–530, 2001. doi:10.1006/jcss.2001.1774.
[5] Bart M. P. Jansen, Daniel Lokshtanov, and Saket Saurabh. A near-optimal planarization algorithm. In Chandra
Chekuri, editor, 25th Annual ACM-SIAM Symposium on Discrete Algorithms, SODA 2014, pages 1802–1811.
SIAM, 2014. doi:10.1137/1.9781611973402.130.
[6] Ken-ichi Kawarabayashi. Planarity allowing few error vertices in linear time. In 50th Annual IEEE Symposium
on Foundations of Computer Science, FOCS 2009, pages 639–648. IEEE Computer Society, 2009. doi:10.1109/
FOCS.2009.45.
[7] Ken-ichi Kawarabayashi, Bojan Mohar, and Bruce A. Reed. A simpler linear time algorithm for embedding
graphs into an arbitrary surface and the genus of graphs of bounded tree-width. In 49th Annual IEEE Symposium
on Foundations of Computer Science, FOCS 2008, pages 771–780. IEEE Computer Society, 2008. doi:10.1109/
FOCS.2008.53.
[8] Ton Kloks. Treewidth, Computations and Approximations, volume 842 of LNCS. Springer, 1994. doi:10.1007/
BFb0045375.
[9] John M. Lewis and Mihalis Yannakakis. The node-deletion problem for hereditary properties is np-complete.
Journal of Computer and System Sciences, 20(2):219–230, 1980. doi:10.1016/0022-0000(80)90060-4.
[10] Dániel Marx and Ildikó Schlotter. Obtaining a planar graph by vertex deletion. Algorithmica, 62(3-4):807–822,
2012. doi:10.1007/s00453-010-9484-z.
[11] Bojan Mohar. A linear time algorithm for embedding graphs in an arbitrary surface. SIAM J. Discrete Math.,
12(1):6–26, 1999. doi:10.1137/S089548019529248X.
[12] Marcin Pilipczuk. A tight lower bound for vertex planarization on graphs of bounded treewidth. Discrete Applied
Mathematics, 2016. in press. doi:10.1016/j.dam.2016.05.019.
23
| 8cs.DS
|
1
Direct estimation of density functionals using a polynomial
basis
arXiv:1702.06516v2 [cs.IT] 20 Nov 2017
Alan Wisler, Visar Berisha, Andreas Spanias, Alfred O. Hero
Abstract—A number of fundamental quantities in statistical signal
processing and information theory can be expressed as integral functions
of two probability density functions. Such quantities are called density
functionals as they map density functions onto the real line. For
example, information divergence functions measure the dissimilarity
between two probability density functions and are useful in a number
of applications. Typically, estimating these quantities requires complete
knowledge of the underlying distribution followed by multi-dimensional
integration. Existing methods make parametric assumptions about the
data distribution or use non-parametric density estimation followed by
high-dimensional integration. In this paper, we propose a new alternative.
We introduce the concept of “data-driven basis functions” - functions of
distributions whose value we can estimate given only samples from the
underlying distributions without requiring distribution fitting or direct
integration. We derive a new data-driven complete basis that is similar
to the deterministic Bernstein polynomial basis and develop two methods
for performing basis expansions of functionals of two distributions. We
also show that the new basis set allows us to approximate functions of
distributions as closely as desired. Finally, we evaluate the methodology by
developing data driven estimators for the Kullback-Leibler divergences
and the Hellinger distance and by constructing empirical estimates of
tight bounds on the Bayes error rate.
Index Terms—Divergence estimation, direct estimation, nearest
neighbor graphs, Bernstein polynomial
I. I NTRODUCTION
Information divergence measures play a central role in the fields
of machine learning and information theory. Information divergence
functions, functionals that map density functions to IR, have been
used in many signal processing applications involving classification
[1], segmentation [2], source separation [3], clustering [4], and other
domains. In machine learning, a sub-class of these divergences known
as f -divergences [5], are widely used as surrogate loss functions since
they form convex upper bounds on the non-convex 0-1 loss [6].
Although these measures prove useful in a variety of applications,
the task of estimating them from multivariate probability distributions
using finite sample data can pose a significant challenge. It is
especially challenging for continuous distributions, which is the focus
of this paper. For algorithms and theory for estimating information
measures for discrete distributions the reader is referred to [7]–
[12]. For the continuous case treated here, there are three general
classes of methods for estimating divergence [13]: 1) parametric
methods, 2) non-parametric methods based on density estimation,
and 3) non-parametric methods based on direct (or graph-based)
estimation. Parametric methods are the most common choice for
estimation, and typically offer good convergence rates (1/N ) when
an accurate parametric model is selected. The fundamental limitation
of parametric methods, is that an accurate parametric model is rarely
available in real world problems and using an inaccurate parametric
model can heavily bias the final estimate. As an alternative, when no
parametric form is known, non-parametric density estimates such as
kernel density estimation [14], histogram estimation [15], or k-nearest
neighbor (k-NN) density estimation [16] are used to characterize
the distribution. While these methods are quite powerful in certain
This research was supported in part by Office of Naval Research grants
N000141410722 (Berisha) and N000141712826 (Berisha) and Army Research
Office grant W911NF-15-1- 0479 (Hero).
scenarios, they are generally high variance, sensitive to outliers, and
scale poorly with dimension [13].
An alternative to these two classes of methods is direct (or
graph-based) estimation, which exploits the asymptotic properties of
minimal graphs in order to directly estimate distribution functionals
without ever estimating the underlying distributions themselves.
These methods have been used to estimate density functionals such
as entropy [17]–[19], the information divergence [13], and the Dp divergence [20]. This class of methods can have faster asymptotic
convergence rates [13] and are often simpler to implement than plugin methods which may have many tuning parameters such as kernel
width or histogram bin size. Direct estimators are often inspired by
an asymptotic property of a graph-theoretic quantity that can be
scaled or modified in order to generate an estimator for a given
information-theoretic measure. This approach is customized to a
specific form of the density functional and is sometimes difficult
to generalize. The estimator for the Dp divergence is an example of
this approach [20]. This is in contrast with plug-in estimators, where
the same general approach can be used to estimate any distribution
functional. Estimators based on influence functions attempt to bridge
the gap between the two types of approaches [21]. In that work the
authors present a recipe for estimating any smooth functional by
using a Von Mises expansion - the analog of the Taylor expansion
for distributions; however that approach still requires that part of the
data be used for density estimation.
In this paper, we provide a general approach for estimating a
wide range of distribution functionals. We propose decomposing the
functional onto a complete set of “data-driven” basis functions; where
the term “data-driven” means that the basis is determined directly
from the data and does not involve distribution fitting or direct
integration. We show that a broad class of distribution functionals can
be approximated as closely as desired through linear combinations
of our proposed basis, where the weights of the basis expansion
are determined through convex optimization. Our approach offers a
powerful alternative for estimating information-theoretic distribution
functionals. We demonstrate the flexibility of the approach by
constructing empirical estimators of bounds on the Bayes error rate
using the same basis set.
The remainder of the paper is organized as follows. In the next
section, we review the literature in this area. In Section II we
provide a detailed description of the problem this paper attempts
to solve and establish some of the basic mathematical notation used
throughout this paper. In Section III we introduce a set of graphtheoretic basis functions and prove that a wide range of information
theoretic quantities can be represented by a linear combination of
functions in this set. In sections IV and V, we explore the limitations
of the proposed methodology in the finite sample regime and propose
two alternate fitting routines to identify weights to map these basis
functions to quantities of interest. In section VI we empirically
investigate how the proposed method can be used to estimate popular
divergence measures (the KL-divergence, the Hellinger distance,
the Dp -divergence), and we compare its performance to various
parametric and non-parametric alternatives. In Section VII, we show
how the method can be extended to form tighter bounds on the Bayes
2
error rate for binary classification problems. Section VIII offers some
concluding remarks.
A. Related Work
A natural method for non-parametric estimation of continuous
distribution functionals involves histogram binning followed by plugin estimation [22], [23]. When the bin-size is adjusted as a function
of the number of available samples per bin, this histogram plugin method is known as Grenander’s method of sieves and it enjoys
attractive non-parametric convergence rates [24], [25]. While these
methods may work well for small data dimension (d = 1, 2),
their complexity becomes prohibitive for larger dimensions. Recent
work has focused on non-parametric plug-in estimators that are
more practical in higher dimensions. These approaches generally
only estimate the PDF for the values of samples in the reference
data, then calculate the expected value across the sample data in
place of numerical integration [26]–[28]. To circumvent the slow
convergence associated with these approaches, ensemble methods
[27] and methods based on influence functions [21] have been
proposed which are capable of achieving the parametric rate O(N −1 )
MSE convergence if the underlying densities meet certain smoothness
conditions [27]. Alternatively, estimates of divergence functions that
rely on estimates of the likelihood ratio instead of density estimation
have been proposed for estimating the α-divergence and the L2 divergence [6], [29]–[32]. These methods estimate the likelihood ratio
of the two density functions and plug that value into the divergence
functions. Other approaches that bypass density estimation are the
convex optimization approach of [29] to estimate f -divergences and
the k-NN graph and minimal spanning tree approaches to estimating
Henze-Penrose divergence [20], [33], [34].
Similarly, the approach we propose in this paper bypasses density
estimation through a polynomial basis expansion where the basis
coefficients are determined through a convex optimization criterion.
This provides added flexibility and allows us to easily estimate a large
class of distribution functionals and to establish empirical estimates of
bounds on the Bayes error rate. Conceptually, this approach is similar
to prior work in estimating the entropy of discrete distributions using
polynomial approximations [7]–[9], [11].
Bounds on optimal performance are a key component in the
statistical signal processing literature. For classification problems, it
is often desirable to bound the Bayes error rate (BER) - the minimum
achievable error in classification problems. The well-known Chernoff
upper bound on the probability of error has been used in a number
of statistical signal processing applications [35]. It motivated the
Chernoff α-divergence [13]. The Bhattacharyya distance, a special
case of the Chernoff α-divergence for α = 21 , upper and lower
bounds the BER [36], [37]. Beyond the bounds on the BER based on
divergence measures, a number of other bounds exist based on other
functionals of the distributions [38], [39]. For estimation problems,
the Fisher information matrix (FIM) bounds the variance of the
optimal unbiased estimator (through it’s relationship with the CRLB).
The authors have also previously introduced the Dp divergence, a
non-parametric f -divergence, and showed that it provides provably
tighter bounds on the BER than the BC bound [20]. They extended
this work to estimation of the Fisher information in [40].
Our data-driven basis, consisting of Bernstein polynomials, can be
used to estimate functionals of distributions and to estimate bounds
on Bayes optimal classification performance. Bernstein polynomials
of a different form have been used for density estimation [41]–[46]. In
contrast to this work, our methods do not rely on density estimation.
II. P ROBLEM S ETUP
In this section, we will set up the problem and establish the notation
that will be used throughout the rest of the paper. We are given a
set of data [X, y] containing N instances, where each instance is
represented by a d-dimensional feature vector xi and a binary label
yi . Suppose that this data is sampled from underlying distribution,
fx (x), where
fx (x) = p0 f0 (x) + p1 f1 (x)
(1)
is made up of the two conditional class distributions f0 (x) and f1 (x)
for classes 0 and 1, with prior probabilities p0 and p1 respectively. If
the priors aren’t explicitly known, they can be easily estimated from
the sample data by measuring the ratio of samples drawn from each
class. As a simple application of Bayes theorem, we can define the
posterior likelihood of class 1, η(x), evaluated at a point x = x∗ , as
P [y = 1]fx (x∗ |y = 1)
fx (x∗ )
∗
∗
p1 f1 (x )
p1 f1 (x )
=
=
fx (x∗ )
p0 f0 (x∗ ) + p1 f1 (x∗ )
η(x∗ ) = P [y = 1|x = x∗ ] =
(2)
We can similarly define the posterior probability for class 0 as
P [y = 0|x = x∗ ] =
p0 f0 (x∗ )
,
p0 f0 (x∗ ) + p1 f1 (x∗ )
(3)
and since y is binary,
P [y = 0|x = x∗ ] = 1 − η(x∗ ).
(4)
∗
To simplify the notation, we remove the dependence of η on x from
portions of the analysis that follow.
Suppose that we wish to estimate some functional G(f0 , f1 )
of distributions f0 (x) and f1 (x), which can be expressed in the
following form
Z
G(f0 , f1 ) = g(η(x))fx (x)dx.
(5)
Throughout the rest of this paper, we will refer to the g(η) in (5) as
the posterior mapping function. Many functionals in machine learning
and information theory, such as f -divergences and loss functions, can
be expressed this way. Consider the family of f -divergences as an
example. They are defined as
Z
f0 (x)
f1 (x)dx,
(6)
Dφ (f0 , f1 ) = φ
f1 (x)
where φ(t) is a convex or concave function unique to the given f divergence. By substituting
f0 (x)
p1 (1 − η(x))
=
f1 (x)
p0 η(x)
and
f1 (x) =
η(x)
fx (x)
p1
(7)
(8)
we can redefine (6) as
Z
Dφ (f0 , f1 ) =
φ
p1 (1 − η(x))
p0 η(x)
η(x)
fx (x)dx.
p1
(9)
Thus any f -divergence can be presented in the form outlined in (5)
simply by defining the posterior mapping function g(η) as
p1 (1 − η)
η
g(η) =
φ
.
(10)
p1
p0 η
We propose a procedure for estimating these types of divergence
functionals which bypasses density estimation. We do this by
representing the functional in terms of the asymptotic limit of a linear
combination of graph-theoretic basis functions.
3
In essence Φk (x∗ ) tells us something about the probability that y = 1
for instances on or near x∗ . Since we are more concerned with the
dataset as a whole than the local characteristics in x, we define the
statistic ρr,k,N (X) to be the fraction of instances x ∈ X, for which
Φk (x) = r, r ≤ k. If we define the indicator function Ir,k (x) as
(
1 Φk (x) = r
Ir,k (x) =
(15)
0 otherwise,
then this test statistic ρr,k,N (X) can be represented by
1 X
ρr,k,N (X) =
Ir,k (x).
N x∈X
(a) N4 (x∗ )
(b) N8 (x∗ )
Fig. 1: Illustration of two neighborhoods of x∗ for k = 4 and k = 8,
instances with y = 0 are blue while instances with y = 1 are red.
In the first scenario Φ4 (x∗ ) = 1, since only one instance in N4 (x∗ )
is red. In the second scenario Φ8 (x∗ ) = 3, since three of the eight
instance in N8 (x∗ ) are red.
Suppose that there exists a set of basis functions H0 (η), ..., Hk (η)
that can be similarly expressed as
Z
Hi (f0 , f1 ) = hi (η(x))fx (x)dx.
(11)
If we assume that there exists a set of coefficients such that
g(η) ≈
k
X
The function ρr,k,N (X) is simply the proportion of k-NN
neighborhoods that contain exactly r points from class y = 1. This
statistic has a number of desirable qualities. We show that this statistic
asymptotically converges to a function of the underlying distributions
that can be described in the form outlined in (5). The following is
proven in Appendix A.
Theorem 1. As the number of samples (N ) approaches infinity,
!
Z
k r
L2
lim ρr,k,N (X) →
η (x)(1 − η(x))k−r fx (x)dx
N →∞
r
whenever k/N → 0.
We propose to use the asymptotic form of ρr,k,N (X) defined in
Theorem 1 as a basis function for estimating functionals of the form
(5),
Z
Hr,k (f0 , f1 ) = lim ρr,k,N (X) = hr,k (η(x))fx (x)dx (17)
N →∞
wi hi (η),
(12)
where
i=0
hr,k (η) =
then consequently
G(f0 , f1 ) ≈ Ĝ(f0 , f1 ) =
k
X
wi Hi (η),
(13)
i=0
where the sense of approximation is that the `2 norm of the difference
between the right and left hand sides is small. In the following
section, we will introduce a set of basis functions that have the desired
properties.
III. G RAPH - THEORETIC BASIS F UNCTIONS
Consider the dataset [X, y] previously defined. Suppose we select
an arbitrary instance x∗ from X and examine it along with the set of
its k −1 nearest neighbors x1N N , x2N N , ..., xk−1
N N in X. We can define
the neighborhood set Nk (x∗ ) = [x∗ , x1N N , ..., xk−1
N N ], as the union
of x∗ and its k − 1 nearest neighbors. Using this we define Φk (x∗ )
as the number of instances in the neighborhood set which are drawn
from class 1, or alternatively, the sum of y across all points in the
neighborhood
X
Φk (x∗ ) =
yi .
(14)
1) The base instance x is considered in the neighborhood
indistinguishably from other instances in Nk (x∗ ).
2) Where traditional k-NN classifiers are concerned only with
identifying the majority, we are interested in the exact number
of instances drawn from each class.
(18)
k
r
X
g
hr,k (η).
k
r=0
(19)
Combining this result with Theorem 1, we can show that a linear
combination of this basis can be used to estimate any function of the
form (5) .
Theorem 2. For any G(f0 , f1 ) that can be expressed in the form
Z
G(f0 , f1 ) = g(η(x))fx (x)dx,
where g(η) is continuous on [0, 1], the approximation
Ĝk,N (X) =
i:xi ∈Nk (x∗ )
Figure 1 provides a simple illustration to help explain how Φk (x∗ )
is calculated. Calculating Φk is similar to how nearest neighbor
classifiers make decisions, but with two important differences:
!
k r
η (1 − η)k−r .
r
The function (18) is the rth Bernstein basis polynomial of degree
k [47]. Bernstein’s proof of the Weierstrass Approximation Theorem
[48] asserts that any continuous function g(η) can be uniformly
approximated on η ∈ [0, 1] to any desired accuracy by a linear
combination of functions in (18) of the form
g(η) ≈
∗
(16)
k
r
X
g
ρr,k,N (X)
k
r=0
(20)
satisfies
lim
lim E
k→∞ N →∞
k/N →0
Ĝk,N (X) − G(f0 , f1 )
2
= 0.
(21)
Theorem 2 provides an asymptotically consistent method of
estimating a variety of information-theoretic functions that makes
no assumptions on the underlying distributions and can be calculated
4
without having to perform density estimation. Throughout the rest of
the paper we will refer to the weights g(r/k) in the approximation
specified by (20) in Theorem 2 as the Bernstein weights. We next
turn to the finite sample properties of the estimator ρr,k,N (X).
IV. F INITE S AMPLE C ONSIDERATIONS
The previous Section investigated the asymptotic properties of
linear combinations of the proposed set of empirically estimable
basis functions. The asymptotic consistency of the proposed method
is valuable, however in real world scenarios, data is inherently a
finite resource, and as a result the efficacy of this method is heavily
dependent on its convergence characteristics in the finite sample
regime. In this section, we will take a detailed look into how
restricting both N and k affects our ability to estimate functions
of two distributions. To do this, it is necessary to first break down
the different sources of error in the proposed methodology.
A. Estimation vs. Approximation Error
The goal of this paper is to empirically estimate the functional
G(f0 , f1 ) of the two underlying distributions f0 (x) and f1 (x) using
a linear combination of directly estimable basis functions
Ĝk,N (X) =
k
X
wr Ĥr,k,N (X).
(22)
r=0
We divide the error of this estimate into two types, the approximation
error (eA ) and the estimation error (eest ):
eT = G(f0 , f1 ) − Ĝk,N (X)
= G(f0 , f1 ) − Ĝk (f0 , f1 ) + Ĝk (f0 , f1 ) − Ĝk,N (X) .
|
{z
} |
{z
}
=:eA
(23)
=:eest
This allows us to separate the error in estimating the basis functions
from error in fitting to the posterior mapping function. Understanding
the trade-off between these two error types will be particularly useful
in Section V, where we explore different methods of fitting weights
to the desired density functionals.
B. Considerations for finite k
A finite sample also implies a finite k and impacts the
approximation error. Let us consider the Bernstein weighting scheme
introduced in (20) for the scenario where the size of the basis set (k)
is restricted. Consider the following example problem.
Example: Suppose that we wish to estimate the function
!
3
g(η) =
η(1 − η)2
1
(24)
using the basis set β0,3 (η), β1,3 (η), β2,3 (η), β3,3 (η). Because g(η) =
β1,3 (η), there exists a set of weights such that
3
X
wr βr,3 (η) = g(η),
(25)
r=0
however, using the Bernstein weighting scheme in (20) yields
3
r
X
g
βr,3 (η)
3
r=0
0
1
2
=g
β0,3 (η) + g
β1,3 (η) + g
β2,3 (η)
3
3
3
3
+g
β3,3 (η)
3
4
2
= η(1 − η)2 + η 2 (1 − η)
3
3
6=g(η).
ĝ(η) =
It is clear from this example that the Bernstein weighting procedure
do not always provide ideal weights when k is restricted. Based
on these results, we are motivated to explore alternative weighting
procedures in order to improve the performance of this method for
the finite sample case. In the following Section, we will introduce a
method of finding better weights using convex optimization.
Regardless of how weights are assigned to the approximation,
selection of k remains an important factor affecting performance.
In order to satisfy Theorem 2, k should be functionally dependent
k
on N , such that N approaches infinity, both k → ∞ and N
→ 0,
however this still provides a great degree of freedom in the selection
of k.
In general, there are two major competing factors that must
be considered when selecting k. The first is that the Weierstrass
approximation theorem can exactly represent any posterior mapping
function g(η) as a linear combination of the proposed basis set only
as k → ∞. This provides motivation for choosing a large k-value
to ensure the best possible fit of g(η). The second factor is that the
asymptotic characteristics of ρr,k,N (X) are dependent on all points in
Nk (x∗ ). Moreover the regime for which Theorem 2 holds, requires
k
that N
→ 0, so we are motivated to select k such that N k.
This means that the selection of k must achieve a compromise in
the trade-off between the approximation and estimation errors, since
larger values of k will increase the amount of finite sample error
made in estimating the individual basis functions, while lower values
of k may inhibit our ability to accurately model the desired function
in the asymptotic regime.
To illustrate how our ability to estimate the desired set of basis
functions varies with k, we calculate the estimated and asymptotic
values of ρr,k,N (X) for k = 10 and k = 100 on data drawn
from distributions f0 (x) ∼ N (03 , I3 ) and f1 (x) ∼ N ( √13 13 , I3 ),
where 03 = [0 0 0] and 13 = [1 1 1], and plot the results
in Figure 2. Estimates are calculated at 3 different sample sizes
(N0 = N1 = 100, 1000, 10000) and each estimate shown in Figure
2 has been averaged across 500 Monte Carlo trials. While we can
estimate the basis set for either k-value with a high degree of accuracy
given enough samples, the estimates for k = 10 are noticeably more
accurate. In fact, we are able to do about as well with 1000 samples
for k = 10 as we are with 10000 samples for k = 100.
V. O PTIMIZATION C RITERIA FOR F ITTING D ENSITY
F UNCTIONALS
In this section, we propose a convex optimization criterion to
identify appropriate weights for fitting information-theoretic functions
when k and N are restricted. Inherently, our goal is to minimize the
total error, defined in (23), however minimizing this quantity directly
isn’t feasible since the value of G(f0 , f1 ) is unknown. To circumvent
this challenge we focus on developing a criterion to minimize the
approximation error. We initially develop an optimization criterion
that assumes the posterior is uniformly distributed, then propose
an alternate method which incorporates an estimate of the posterior
density function in order to more accurately model the approximation
error.
A. Uniform Optimization Criteria
(26)
Recall that the approximation error eA can be represented as
Z
eA = (η(x))fx (x)dx,
(27)
where
(η) = g(η) −
k
X
r=0
wr hr,k (η).
(28)
5
0.2
0.04
0.15
0.03
0.1
0.02
0.05
0.01
0
0
0
5
10
0
50
100
Fig. 2: Plot of true and estimated basis values vs. r for data drawn from underlying distributions f0 (x) ∼ N (03 , I3 ) and f1 (x) ∼
N ( √13 13 , I3 ).
Since solving (27) requires high-dimensional integration and
knowledge of the underlying distributions. However, because η is
a function of x, (η) is implicitly a function of x as well, and by
the law of the unconscious statistician [49],
Z
eA = E[(η)] = (η)fη (η)dη,
(29)
where fη (η) is the probability density function of the random variable
η. Rewriting the error in this form simplifies the region of integration
to a well defined space (since η ∈ [0, 1]) and circumvents the high
dimensionality of x. While this eliminates some of the challenges in
calculating the error it also creates new ones stemming from the fact
that fη (η) is unknown and difficult to estimate due to its implicit
dependency on f0 (x) and f1 (x). The task of estimating fη (η) will
be explored in detail in Section V-B, however for the time being we
will bypass this challenge and simply attempt to minimize
Z
∗
eA = |(η)|2 dη.
(30)
It is worth noting that if fη (η) is uniformly distributed
e∗A = E |(η)|2 ≥ e2A .
(31)
While there exists an analytical solution for identifying the weights
which minimize (30) [50], we use a convex optimization procedure
that allows us to also account for the estimation error. If we define
a discretized set of posterior values η̃ = [η̃1 , η̃2 , ..., η̃Ñ ], where 0 ≤
η̃1 < η̃2 < ... < ηÑ ≤ 1, a procedure to identify weights that
minimize (30) can be defined as
Ñ
k
X
2
1 X
g(η̃i ) −
wr hr,k (η̃i ) .
w0 ,...,wk Ñ
r=0
i=1
w0 , ..., wk = argmin
(32)
.
To illustrate the effectiveness of this method, we consider the
example problem of trying to estimate the Hellinger distance (a
problem we will further explore in Section VI). If we assume both
classes have equal prior probability (p0 = p1 = 0.5), then the
posterior mapping function for the squared Hellinger distance is
p
√
g(η) = ( η − 1 − η)2 .
(33)
This function is estimated using this convex weighting procedure as
well as the previously described Bernstein weighting procedure, and
we compare how well each method models the desired function for
values of k varying from 0 to 100. The performances of each method
is evaluated by the following formula
MSE(ĝ, g) =
Ñ
X
2
g(η̃i ) − ĝ(η̃i ) ,
(34)
i=1
and the results are presented in Figure 3 for a range of k
values varying from 1 to 100. This experiment shows that the
proposed convex fitting procedure is able to approximate the
Hellinger posterior mapping function far more accurately than the
Bernstein approximation. This improvement isn’t surprising since the
proposed method directly minimizes the MSE whereas the Bernstein
approximation guarantees consistency only as k → ∞, but it still
helps to illustrate the potential for improvement in the Bernstein
weights that exists for smaller k.
The expression (34) does not take into account finite sample
errors that lead to noisy estimates of the basis functions and thus
does not directly reflect our ability to estimate G(f0 , f1 ) with a
finite sample. The consequences of this could be quite significant.
When using a similar approach for entropy estimation, Paninski
observed that the good approximation properties were a result of
“large oscillating coefficients” which magnify the variance of the
corresponding estimator [11]. Additionally, it does not account for
the possibility that η is distributed non-uniformly. To empirically
examine the finite sample properties of the two approaches, we repeat
the previous experiment, this time estimating the basis functions
empirically on samples of data drawn from distributions f0 (x) ∼
13
, I3 ). We generate N = 1000 samples
N (03 , I3 ) and f1 (x) ∼ N ( √
3
(500 samples per class) in each of the 500 iterations of a Monte Carlo
simulation, and evaluate the MSE as
MSE(G, Ĝ) =
1
NM C
NM C
X
2
G(f0 , f1 ) − Ĝ(f0 , f1 ) ,
(35)
i=1
where NM C represents the number of Monte Carlo iterations. Since
we know that the estimation error is scaled by the magnitude of the
weights, we also evaluate a modified fitting routine which augments
6
Bernstein
Convex
10
10 2
-2
Mean Squared Error
Mean Squared Error
10 0
10 -4
10 -6
10 -8
0
20
40
60
80
Bernstein
Convex
Convex (w/regularization)
10 0
10 -2
0
100
20
40
Fig. 3: Approximation error of each fitting procedure as a function
of the number of basis elements k. This is the idealized case where
the basis estimation error is zero and the total error is solely due to
imperfect approximation of the posterior mapping function g(η).
(32) with a regularization term to penalize solutions with large
weights,
w0 , ..., wk =
Ñ
k
X
1 X
g(η̃i ) −
wr hr,k (η̃i )
w0 ,...,wk Ñ
r=0
i=1
80
100
k
k
argmin
60
2
+
k
λX 2
wr ,
k r=0
(36)
where λ represents a tuning parameter which controls the importance
assigned to minimization of the approximation error relative to the
estimation error. Intuitively, higher λ values will make sense for
smaller data sets to control the variance of the estimator. We set
λ = 0.01 for all experiments conducted in this paper. The results
of this experiment are shown in Figure 4. We immediately see the
necessity of the regularization term, as without it the error becomes
extremely large for a range of k values. More generally, the inclusion
of the regularization term improves the performance at every k value
in this experiment. In comparing the Bernstein weights with the
convex (regularized) weights, we find that 1) the performance of
the convex method is less dependent on the selection of k and
2) the convex weights generally perform better at lower values of
k, while the Bernstein weights outperform at higher k. While the
peak performance of the Bernstein method is higher than the convex
method, there exists no good method of selecting k a priori in order
to reliably achieve this performance. In contrast, the convex method
with regularization is less sensitive to the value of k selected.
B. Density-weighted Optimization Criteria
In the optimization criteria introduced in the previous section we
implicitly make the assumption that the distribution of the random
variable η(x) ∼ fη (η) is uniformly distributed. In this section we
will investigate a data-driven estimator for fη (η). However, before
we proceed it is important to clarify what this distribution actually
is.
We initially introduced η as the posterior likelihood function for
class 1, which we showed in (2) can be represented as a function of
the underlying distributions. When this function’s input is a known
point x∗ , η(x∗ ) represents the probability that y = 1 given that
x = x∗ . However, if the input is a random variable x, then η is also
a random variable, which is distributed according to fη (η).
Fig. 4: The total error of each fitting procedure as a function of k,
when there is both approximation and estimation error.
Figure 5 illustrates fη (η) for two univariate normal distributions
f0 (x) ∼ N (0, 1) and f1 (x) ∼ N (1, 1). Figure 5a displays the two
class distributions across x, Figure 5b displays η(x) as a function
of x, and Figure 5c displays fη (η) as a function of η. We see from
this illustration that while, η(x) is close to 0 or 1 across most of the
region of x that is displayed, η(x) remains close to 0.5 in the regions
where fx (x) is greatest. As a results fη (η) is roughly bell-shaped,
and the likelihood of η existing at the extremities (close to 0 or 1)
is relatively low. Because the probability density in these regions is
low, the accuracy of our fit in these regions is less important, and
can be given less weight in the fitting routine. Figure 6 repeats this
illustration for two well seperated normal distributions. In this case
the distribution of η is such that fη (η) is most dense towards the
extremities, and therefore they should be given more weight in the
fitting routine. Side by side, these two figures present an interesting
contrast. Despite the fact that the set of density functions f0 and
f1 look quite similar in the two scenarios, the minor difference in
separation significantly alters the distribution of η.
In practice the underlying distributions f0 and f1 are unknown,
and as a result, fη (η) is unknown as well. However, if we were able
to sample fη (η) at η = η̃i , a more direct method of minimizing eT
would be to solve
w0 , ..., wk =
argmin
w0 ,...,wk
Ñ
X
i=1
g(η̃i ) −
k
X
k
2
λX 2
wr hr,k (η̃i ) fˆη (η̃i )∆η̃ +
wr ,
k r=0
r=0
(37)
where ∆η̃ = η̃i+1 − η̃i .
Below we show that ρr,k,N (X), the statistic previously defined in
Theorem 1 can be used to sample the PDF of η. This result is stated
in Theorem 3.
Theorem 3. As N → ∞ and k → ∞ in a linked manner such that
k
→ 0 and kr → η ∗
N
kρr,k,N (X) → fη (η ∗ ).
Theorem 3 is useful as it provides a method of sampling fη (η) that
doesn’t depend on estimates of the underlying density functions f0
and f1 . Using this result, we can estimate the density of the posterior
at point η̃i as
fˆη (η̃i ) = k̃i ρr̃i ,k̃i ,N (X)
(38)
7
0.4
1.5
1
f0
f1
0.8
0.3
fη (η)
0.2
η
fx (x)
1
0.6
0.4
0.5
0.1
0
-6
0.2
0
-4
-2
0
2
4
6
0
-4
-2
0
x
2
4
6
0
0.2
0.4
x
(a) Class distribution
0.6
0.8
1
η
(b) Class 1 Posterior
(c) Distribution of class 1 posterior likelihoods
Fig. 5: Illustration of the posterior distribution for two close univariate normal distributions.
0.4
6
1
f0
f1
5
0.8
0.3
fη (η)
0.2
η
fx (x)
4
0.6
0.4
3
2
0.1
0
-6
0.2
1
0
-4
-2
0
2
4
6
0
-4
-2
0
x
(a) Class distribution
2
4
6
0
0.2
0.4
x
0.6
0.8
1
η
(b) Class 1 Posterior
(c) Distribution of class 1 posterior likelihoods
Fig. 6: Illustration of the posterior distribution for two separated univariate normal distributions.
where r̃i and k̃i are integers selected such that η̃i =
r̃i
k̃i
for
i = 1, 2, ..., Ñ . Sampling at exactly η̃i may not always be possible
since the maximum value of k is limited by the sample size, and
k determines the resolution of the sampling scheme. Even if it is
possible, it may not be desirable to recalculate ρr̃i ,k̃i ,N (X) for
different values of ki because of the computational burden. To
overcome these problems we can design our approach such that we
utilize the same set of test statistics ρr,k,N (X) in the estimation of
the posterior distribution as are used in the estimation of the basis
set. One way to do this is to assign the set of discretized posteriors
1 2
, , ..., 1]
(39)
k k
so that it is straightforward to calculate fˆη (η) from the known values
of ρr,k,N (X). An alternate approach is to leave η̃ unchanged and
interpolate fˆη (η) to determine its value at the desired points. The
advantage of this approach is that it doesn’t constrain how η is
sampled. Throughout the rest of this paper, we will employ the latter
method and solve for fˆη (η) by linearly interpolating between its
values on the discretized set (39).
η̃ = [0,
Because this density-weighted fitting routine more directly
minimizes the approximation error of the final estimate, we expect
it to generally outperform the uniform method, particularly in cases
where the density of the posterior is highly non-uniform and where
the desired posterior mapping function g(η) is difficult to model
using the proposed basis set. This hypothesis will be verified in
Sections VI and VII, when we empirically evaluate our methods
with real data. This improvement comes at a computational cost
since the weights are now data-dependent, they must be calculated
online, whereas for the uniform method they can be calculated offline
and stored. Solving for the k weights can be be implemented with
O(k3 ) time complexity [51]. Current state-of-the-art algorithms for
k-NN graphs can be implemented with O(N log N ) time complexity
[52]. As a result the complexity for the Convex (uniform) method is
O(N log N ) assuming that the weights are available. The complexity
for the Convex (density weighted) method is O(k(N )3 + N log N )
since the weights must be determined for each new dataset.
VI. D IVERGENCE E STIMATION
In this section, we show how the proposed method can be applied
to estimating three f -divergences, the Hellinger distance, the KLdivergence, and the Dp -divergence, directly from data.
A. Hellinger Distance
The Hellinger distance squared is an f -divergence used to
quantify the dissimilarity between two probability distributions and
is calculated by
Z p
2
p
1
2
H (f0 , f1 ) =
f0 (x) − f1 (x) dx.
(40)
2
Using the approach proposed in Section V, we can estimate
H 2 (f0 , f1 ) by fitting weights to the posterior mapping function
!2
r
r
1
η
1−η
gH (η) =
−
.
(41)
2
p1
p0
8
To evaluate the efficacy of this method, we conduct four different
experiments in which we attempt to estimate the Hellinger distance
between two distributions from finite sample data. In the first three
experiments, both distributions are normally distributed according to
f (x) ∼ N (µ1d , Σd ), where
σ1,1 σ1,2 . . . σ1,d
σ2,1 σ2,2 . . . σ2,d
(42)
Σd = .
..
.. ,
..
..
.
.
.
σd,1 σd,2 . . . σd,d
for σi,j = β |i−j| . The first experiment evaluates the most basic
case where both Gaussians have spherical covariance. The second
experiment considers the case where there exists a strong fixed
dependency between adjacent dimensions by using an elliptical
covariance structure. The third experiment evaluates the case where
this dependency between adjacent dimensions is now dependent on
which class the data is drawn from. In the fourth experiment, we
return to linearly independent dimensions and consider the case
where one of the distributions isn’t normally distributed, but instead
uniformly distributed within a d-dimensional hypercube
(
1
x ∈ [µ − β, µ + β]d
(2β)d
f (x) =
(43)
0
otherwise.
A detailed description of the distribution types and parameter
setting used in each of the four experiments is presented in Table
I. In addition to using the proposed method, we also estimate
the Hellinger distance using two different plug-in estimators, one
based on a parametric density estimator that assumes the data is
normally distributed and one based on a k-NN density estimate of
the underlying distributions. The proposed method is implemented
using λ = 0.01 and k = 10, while setting η̃ = [0, 0.01, . . . , 1].
To calculate the k-NN estimate, we use the universal divergence
estimation approach described in [53] and implemented in the ITE
toolbox [54]. This method allows us to fix k = 10 and still achieve
an asymptotically consistent estimate of the divergence.
In each of the first three experiments, the parametric model shows
the highest rate of convergence as expected, although in experiment
two it is slightly outperformed at smaller sample sizes by the
proposed method. In the fourth experiment, when the assumption
of Gaussianity in f1 no longer holds, the parametric solution is
significantly biased and as a result, is outperformed by both nonparametric methods at higher sample sizes (N > 2000). Relative
to the k-NN plug-in estimator, the proposed method performs
slightly worse in experiment 1, slightly better in experiment 2
and significantly better in experiments 3 and 4, with the results
being relatively consistent across the various sample sizes. The
improvement in performance shown in experiments 3 and 4 suggests
that the proposed method offers the greatest benefit when there
exists differences in the shapes of the two underlying distributions.
Though the density-weighted procedure consistently outperformed
the uniform method, the observed improvement was relatively minor
in these experiments.
B. Kullback Leibler Divergence
The Kullback Leibler (KL) divergence [55], also sometimes
referred to as the KL risk or relative entropy, is an asymmetric
measure of divergence between two probability density functions.
Using our regular notation the KL-divergence can be calculated by
Z ∞
f0 (x)
dKL (f0 ||f1 ) =
f0 (x) log
dx.
(44)
f1 (x)
−∞
TABLE I: Experiment overview table.
f0 (x)
Family
Experiment 1
Experiment 2
Experiment 3
Experiment 4
Normal
Normal
Normal
Normal
µ
0
0
0
0
β
0
0.8
0.8
0
f1 (x)
Family
Normal
Normal
Normal
Uniform
µ
q
β
1
3
0
q
1
3
0.8
q
1
3
0.9
3
0
While the KL-divergence has the same general purpose as the
Hellinger distance, that is to measure the dissimilarity between two
probability density functions, it also has several key difference. Firstly
since the KL-divergence is asymmetric it doesn’t technically qualify
as a distance function and dKL (f0 ||f1 ) isn’t necessarily equal to
dKL (f1 ||f0 ). This also means that dKL (f0 ||f1 ) and dKL (f1 ||f0 )
will have different posterior mapping functions. We can define the
posterior mapping function for dKL (f0 ||f1 ) as
p1 (1 − η)
1−η
0
gKL
(η) =
log
(45)
p0
p0 η
and the posterior mapping function for dKL (f1 ||f0 ) as
p0 η
η
1
gKL
(η) =
log
p1
p1 (1 − η)
(46)
such that
Z
∞
i
gKL
(η(x))(p0 f0 (x) + p1 f1 (x))dx. (47)
dKL (fi ||f|i−1| ) =
−∞
It is worth noting that when p0 = p1 = 0.5
1
0
gKL
(η) = gKL
(1 − η)
(48)
0
(η).
gKL
1
(η)
gKL
One challenge presented in
is a reflection of
thus
modeling the KL-divergence is that the posterior mapping functions
are difficult to model due to discontinuities at the end points, since
0
0
(1) is undefined though
(0) = ∞ and gKL
gKL
0
lim gKL
(η) = 0.
η→1−
(49)
1
has the same problem at the opposite
Due to their symmetry gKL
endpoints. To handle this we simply select our discretized set of
posteriors η̃1 , η̃2 , ..., η̃Ñ such that 0 < η̃1 < η̃2 < ... < ηÑ < 1. For
the experiments in this Section, we set
η̃1 , η̃2 , η̃3 , ..., η̃100 , η̃101 = , 0.01, 0.02, ..., 0.99, 1 −
−4
(50)
where = 10 . Using this modified set of discretized posteriors,
we repeat the set of four experiments described in Section VI-A to
evaluate the proposed methods ability to estimate the KL-divergence.
The results of this experiment are displayed in Figure 8. Like the
estimates of the Hellinger distance, the parametric method generally
yielded the best performance in the first three experiments, but
suffered from a large asymptotic bias in experiment 4. The proposed
method once again significantly outperformed the k-NN plug-in
estimator in experiments 3 and 4, however the results in experiments
1 and 2 are slightly more nuanced due to the significant difference in
performance between the two optimization criteria in these trials.
In both of these trials the density-weighted criteria significantly
outperforms the uniform method at all sample sizes. In experiment
1 the k-NN plug-in estimator outperforms the regular plug-in
estimator at all sample sizes, but is outperformed by the densityweighted method for N > 300. In experiment 2 the k-NN plug-in
estimator consistently outperforms both proposed methods, however
the improvement over the density-weighted method is marginal.
9
Experiment 1
10 -2
Experiment 2
10 -2
10 -2
Experiment 3
10 -1
Experiment 4
10 -3
10 -3
10 -4
MSE
MSE
MSE
10 -2
MSE
10 -3
10 -4
10 -4
10 -3
10
10 -5
10 2
10 3
-5
10
10 -6
10 2
10 4
Samples per distribution
10 3
10 4
Samples per distribution
Parametric
k-NN Plug-in
-5
10 -6
10 2
10 3
10 -4
10 2
10 4
10 3
10 4
Samples per distribution
Samples per distribution
Convex (uniform)
Convex (density-weighted)
Fig. 7: Plots of MSE vs. Sample size in estimating the Hellinger distance for the four different experiments outlined in Table I.
10 0
Experiment 1
10 -1
10 -1
Experiment 2
10 0
10 -2
Experiment 3
10 1
Experiment 4
10 -1
10 -3
MSE
10 -2
MSE
MSE
MSE
10 0
10 -2
10 -1
10 -3
10 -4
10 -4
10 2
10 3
10 4
Sample Size
10 -3
10 -5
10 2
10 3
10 4
Sample Size
Parametric
10 -4
10 2
10 3
Sample Size
k-NN Plug-in
Convex (uniform)
10 4
10 -2
10 2
10 3
10 4
Sample Size
Convex (density-weighted)
Fig. 8: Plots of MSE vs. Sample size in estimating the KL-divergence for the four different experiments outlined in Table I.
C. Dp -Divergence
sample sizes, though the proposed method seems to be converging
slightly faster than the MST method in experiment 4 and could
The Dp -divergence is an f -divergence defined by
possibly exceed its performance given enough samples. Unlike in
Z
(p0 f0 (x) − p1 f1 (x))2
1
the previous experiments, we found no difference in performance
2
Dp0 (f0 , f1 ) =
dx − (p0 − p1 ) .
between the uniform optimization criteria and the density-weighted
4p0 p1
p0 f0 (x) + p1 f1 (x)
(51) criteria in this experiment. This is due to the fact that gDp (η) is a
The Dp -divergence has the unique property of being directly polynomial and can be perfectly represented by the proposed basis
estimable from data using minimum spanning trees [20]. Because of set, even when k is truncated. Since we are able to achieve a solution
this property, it has been used to form non-parametric estimates of the where (η) = 0 ∀η, the density weighting has no impact on the final
Fisher information [40] as well as upper bounds on the Bayes error results.
rate in a range of classification problems [20], [56]. The posterior
mapping function for the Dp -divergence can be defined as
VII. F ITTING B OUNDS ON P ERFORMANCE
gDp (η) =
(2η − 1)2 − (2p0 − 1)2
4p0 (1 − p0 )
(52)
which simplifies to (2η − 1)2 when p0 = p1 = 0.5. We once
again repeat the experiments described in Section VI-A to evaluate
the proposed methods ability to estimate the Dp -divergence. This
experiment provides the unique opportunity to compare the proposed
method to a more traditional direct estimation procedure. The results
of this experiment are displayed in Figure 9.
As in the previous experiments, the parametric estimate generally
performed the best in the first three experiments, but suffered from a
large asymptotic bias in experiment 4. The proposed methods perform
better than the MST-based estimation in experiments 1 and 2 but
worse in experiments 3 and 4. The relative performance of each
method in this experiment was largely consistent across the various
The optimization criteria in the proposed fitting routines gives us
the ability to not only approximate information-theoretic functions,
but to bound them as well. This is especially useful for forming
bounds on the Bayes Error Rate (BER). The Bayes error rate
represents the optimal classification performance that is achievable
for a given pair of class distributions f0 (x) and f1 (x) with prior
probabilities p0 and p1 respectively and can be calculated by
Z
Z
Bayes =
p0 f0 (x)dx +
p1 f1 (x)dx.
(53)
p0 f0 (x)≤p1 f1 (x)
p1 f1 (x)≤p0 f0 (x)
In essence the BER measures the intrinsic difficulty of a particular
classification problem based on the data. A thorough understanding of
the BER of a particular problem can help design optimal classifiers.
Because of the challenges associated with estimating the BER, much
10
Experiment 2
10 -2
10 -4
10 -5
10 2
10 3
10 4
Samples per distribution
Parametric
10 -2
MSE
10 -4
10 -4
10 -5
10 2
10 3
10 4
Samples per distribution
MST
Experiment 4
10 -1
10 -3
MSE
10 -3
MSE
10 -3
Experiment 3
10 -2
MSE
Experiment 1
10 -2
10 -3
10 -5
10 2
10 3
10 4
Samples per distribution
Convex (uniform)
10 -4
10 2
10 3
10 4
Samples per distribution
Convex (density-weighted)
Fig. 9: Plots of MSE vs. Sample size in estimating the Dp -divergence for the four different experiments outlined in Table I.
of the literature has focused on generating bounds on the BER [20],
[37], which are generally formulated in terms of some measure of
divergence between the two class distributions. One such bound, the
well-known Bhattacharya bound, is given by [37]
1
1p
1
−
1 − BC 2 (f0 , f1 ) ≤ Bayes ≤ BC(f0 , f1 ),
(54)
2
2
2
where
BC(f0 , f1 ) = 1 − H 2 (f0 , f1 ).
(55)
While the Hellinger distance here can be estimated via any of the
methods discussed in Section VI-A, parametric estimates are most
common. Alternatively [20] introduced the bounds
1
1
1q
1
(56)
−
D 1 (f0 , f1 ) ≤ Bayes ≤ − D 1 (f0 , f1 )
2
2
2
2
2 2
where
Z
f0 (x)f1 (x)
D 1 (f0 , f1 ) = 1 − 2
dx.
(57)
2
f0 (x) + f1 (x)
These bounds have the advantage of being provably tighter than
the Bhattacharyya bounds [20]. Furthermore since D 1 represents
2
a particular case of the Dp -divergence, which is estimable directly
from data, these bounds bypass the need for density estimation much
like the approaches proposed in this paper. While these bounds are
significantly tighter than the Bhattacharyya bounds, they still leave
room for improvement. Avi-Itzhak proposed arbitrarily tight bounds
on the BER in [39], however these bounds require density estimation
to be employed in practical problems. In this section, we will use a
modified version of the previously described fitting routine in order
to investigate how tightly we are able to bound the BER using a
linear combination of directly estimable basis functions.
Fig. 10: The Bayes error rate along with the three considered upper
bounds displayed as a function of η.
regularized fitting routine described in (36) yields
w0 , ..., wk =
Ñ
k
X
1 X
g(η̃i ) −
wr hr,k (η̃i )
w0 ,...,wk Ñ
r=0
i=1
argmin
subject to
k
X
wr hr,k (η̃i ) ≥ g(η̃i )
2
+
k
λX 2
wr
k r=0
(60)
∀η̃i .
r=0
Using the fitting routine described in (36) to bound the BER,
requires that we define g(η) appropriately for estimation of the BER,
and constrain our fit such that
k
X
wr hr,k (η̃i ) ≥ g(η̃i ) ∀η̃i .
(58)
r=0
We can express (53) as
Z
Bayes = min p0 f0 (x), p1 f1 (x) dx
Z
= min 1 − η(x), η(x) fx (x)dx
(59)
so g(η) = min 1 − η, η . Incorporating these changes within the
Figure 10 compares the theoretical values of each of these bounds
as a function of η. These results indicate that the proposed method
offers much tighter theoretical bounds than the other two methods,
however this bound is based on the asymptotic properties of the
proposed basis set and doesn’t consider the limitations of a finite
sample estimate.
We evaluate the finite-sample performance of this method by
calculating each of the described bounds across the four experiments
described in Table I. Figure 11 displays the ground truth value
of the BER, along with the theoretical and estimated values for
each of the three bounds (Bhattacharyya, Dp , and convex) across
sample sizes ranging from 100 to 10000. The Bhattacharyya bound
is calculated based on a parametric plug-in estimator which assumes
both class distributions to be normally distributed. The Dp bound is
calculated from the Friedman-Rafsky test statistic using the approach
11
Experiment 1
45
Experiment 3
Experiment 2
50
40
45
Experiment 4
40
40
Error (%)
35
45
Error (%)
Error (%)
Error (%)
35
40
35
30
25
20
30
10 2
10 3
10 4
35
10 2
Samples per distribution
BER (True)
Bhattacharyya (True)
10 3
10 4
Samples per distribution
Bhattacharyya (Estimated)
30
10 2
10 3
10 4
Samples per distribution
Dp (True)
Dp (Estimated)
15
10 2
10 3
10 4
Samples per distribution
Convex (True)
Convex (Estimated)
Fig. 11: Plots of theoretical and estimate upper bounds on the BER as a function of sample size for the four different experiments outlined
in Table I.
TABLE II: Parameters for 2 8-dimensional Gaussian data sets for
which the Bayes error rate is known (from [57]).
D0
D2
µ0
σ0
µ1
σ1
µ0
σ0
µ1
σ1
0
1
2.56
1
0
1
3.86
8.41
0
1
0
1
0
1
3.10
12.06
0
1
0
1
0
1
0.84
0.12
0
1
0
1
0
1
0.84
0.22
0
1
0
1
0
1
1.64
1.49
0
1
0
1
0
1
1.08
1.77
0
1
0
1
0
1
0.26
0.35
0
1
0
1
0
1
0.01
2.73
TABLE III: Comparing upper bounds on the Bayes error rate for the
multivariate Gaussians defined in Table II.
Actual Bayes Error
Mahalanobis Bound
Bhattacharyya Bound
Dp Bound
Convex Bound
Data 1
10%
18.90% ± 0.55%
21.74% ± 0.87 %
16.51 % ± 1.07%
14.17 % ± 0.86%
Data 2
1.90%
14.07 % ± 0.45 %
4.68 % ± 0.27 %
3.99 % ± 0.52 %
3.87% ± 0.43%
described in [20]. Finally the convex bound is calculated as a linear
combination of the proposed directly estimable basis functions using
weights optimized according to (60). The results of this experiment
are largely consistent across the four experiments, the convex method
yields the tightest bound, followed by the Dp bound, and finally
the Bhattacharyya bound. Except for the Bhattacharyya bound in
experiment 4, which is estimated parametrically, all of the bounds
appear to converge to their asymptotic solution. While the convex
bound generally offers a slightly slower convergence rate than the
other two solutions, it remains tighter than the other two bounds
across all sample sizes.
In order to further validate this bound we repeat one of the
experiments conducted in [20] by evaluating the proposed bound
along with the Mahalanobis bound, the Bhattacharyya bound, and
the Dp bound on two 8-dimensional Gaussian data sets described in
[57]. The mean and standard deviations of f0 and f1 for the two data
sets are described in Table II, and all dimensions are independent.
These data sets allow us to analyze the tightness and validity of
the bounds in a higher dimensional setting. For this experiment the
sample size was fixed at N = 1000 and only the empirical value of
each of the bounds was evaluated. Table III displays the mean and
standard deviation of each bound calculated across 500 Monte Carlo
iterations for each of the two data sets. In both data sets the convex
method provides the tightest bounds on the BER.
VIII. C ONCLUSION
This paper introduces a novel method for estimating density
functionals which utilizes a set of directly estimable basis functions.
The most appealing feature of the proposed method is its flexibility.
Where previous methods of direct estimation are generally only
applicable to a specific quantity, we show that the basis set can be
used to generate an asymptotically consistent estimate of a broad class
of density functionals, including all f -divergences and the Bayes
error rate. We validate these findings by experimentally evaluating
the proposed method’s ability to estimate three different divergences
(the KL-divergence, the Hellinger distance, and the Dp -divergence)
for four pairs of multivariate probability density functions. The
results reveal that the proposed method performs competitively with
other non-parametric divergence estimation methods, and seems to
outperform them in cases where the data from the two distributions
have different covariance structures or belong to different families.
Additionally we demonstrate how the method can be modified to
generate empirically-estimable bounds on the Bayes error rate that
are much tighter than existing bounds.
Future work could focus on studying the finite-sample properties
of the basis set proposed in this paper, since this represents a
major source of error for the proposed methodology. An improved
understanding of these properties could enable us to refine the
regularization term in our optimization criteria to more accurately
model each weights contribution to the estimation error or to develop
ensemble methods, like those in [26], for estimating the individual
basis functions. Another worthwhile future direction would be on
determining an orthogonal version of the Bernstein basis expansion
since this would simplify solving for the weights of the expansion.
A PPENDIX A
P ROOF OF T HEOREM 1
Aspects of this proof mirror the methods used to prove the
asymptotic convergence of the k-NN error rate in pages 62-70 of [58],
please consult this text for additional details. Construction of this
proof requires the introduction of an auxilary variable u. We define
the augmented data set [X, y, u], where X is as defined previously,
u is a set of i.i.d. random variables which are uniformly distributed
in [0, 1], and y is defined such that
1 ui ≤ η(xi )
yi =
(61)
0 otherwise.
Now, let us consider an arbitrary point x∗ in the support of fx (x).
For each instance xi , we can define an alternate label
1 ui ≤ η(x∗ )
yi0 =
(62)
0 otherwise.
12
Unlike the real labels yi , these alternate labels yi0 have no dependency
on xi ; they depend only on the fixed point x∗ . From these alternate
labels we construct the alternate statistic
X
Φ0k (x∗ ) =
yi0 .
(63)
i:xi ∈Nk (x∗ )
Φ0k (x∗ )
Since
is simply the sum of k i.i.d. Bernoulli random
variables, we can express the probability that Φ0k (x∗ ) = r as
!
k r ∗
0
∗
P [Φk (x) = r|x = x ] =
η (x )(1 − η(x∗ ))k−r .
(64)
r
We can upper bound the likelihood that Φ0k (x∗ ) 6= Φk (x∗ ) by
X
P Φ0k (x∗ ) 6= Φk (x∗ ) ≤
P [yi 6= yi0 ].
(65)
i:xi ∈Nk (x∗ )
Using the definitions of yi and yi0 above, (65) can be expressed in
terms of the difference between the posterior likelihood at x∗ vs. xi
as
X
P Φ0k (x∗ ) 6= Φk (x∗ ) ≤
E |η(x∗ ) − η(xi )| . (66)
N → ∞, P [B] → 1 and as a result
lim E Ir,k (xi )Ir,k (xj ) = lim E Ir,k (xi ) E Ir,k (xj )
N →∞
N →∞
(73)
= (ρ∗ )2 .
Plugging this into (72) yields
1 ∗ N −1 ∗ 2
ρ +
(ρ ) .
(74)
N
N
Similarly the second term simplifies to
1 X
∗
∗
lim E[ρr,k,N (X)ρ ] = lim ρ E
Ir,k (xi ) = (ρ∗ )2
N →∞
N →∞
N x ∈X
i
(75)
∗ 2
and the third term clearly equals (ρ ) . Substituting these results into
(71) yields
1 ∗ N −1 ∗ 2
lim E (ρr,k,N (X) − ρ∗ )2 = lim
ρ +
(ρ )
N →∞
N →∞ N
N
∗ 2
∗ 2
− 2(ρ ) + (ρ )
(76)
∗
∗ 2
ρ − (ρ )
= lim
= 0.
N →∞
N
lim E[(ρr,k,N (X))2 ] = lim
N →∞
N →∞
xi ∈Nk (x∗ )
Using Lemma 5.4 in [58], we can show that
X
E |η(x∗ ) − η(xi )| → 0
xi ∈Nk (x∗ )
Starting with the expression
k
N
→ 0. Combining (66) and (67)
0 ∗
P Φk (x ) 6= Φk (x∗ ) → 0
as N → ∞ whenever
(68)
k
as N → ∞ whenever N
→ 0. Furthermore, since convergence in
probability implies convergence in distribution [59],
lim P [Φk (x∗ ) = r] = P [Φ0k (x∗ ) = r].
N →∞
(69)
Using (69), we can simplify the expectation of Ir,k (x) to
lim E[Ir,k (x)] = lim E[E[Ir,k (x)|x = x0 ]]
N →∞
N →∞
= E[P [Φ0k (x) = r|x = x0 ]]
!
k
η(x0 )r (1 − η(x0 ))k−r fx (x0 )dx0 = ρ∗
r
Z
=
(70)
where x0 represents a random variable independent of all instances
xi and distributed according to fx (x0 ). Now, let us evaluate the
expression
E (ρr,k,N (X) − ρ∗ )2 = E[(ρr,k,N (X))2 ]
(71)
−2E[ρ∗ ρr,k,N (X)] + E[(ρ∗ )2 ].
Beginning with the first term in (71)
1 X X
2
E[(ρr,k,N (X)) ] = E
Ir,k (xi )Ir,k (xj )
N 2 x ∈X x ∈X
i
j
X
XX
1
1
= 2E
Ir,k (xi ) + 2
Ir,k (xi )Ir,k (xj )
N
N
x ∈X
i
A PPENDIX B
P ROOF OF T HEOREM 2
(67)
xi ,xj ∈X;i6=j
N2 − N
1
E Ir,k (xi )Ir,k (xj )
= E Ir,k (xi ) +
N
N2
(72)
Note that the random variables Ir,k (xi ) and Ir,k (xj ) are
conditionally independent if B is true, where B = {xi ∈ X \
Nk (xj ) ∩ xj ∈ X \ Nk (xi )}. Since every instance in X \ xi is
equally likely to be in Nk (xi ), we can bound the probability of the
complement of B by P [B̄] ≤ 2k/N . From this we know that as
lim
lim Ĝk,N (X) = lim
k→∞ N →∞
k/N →0
lim
k→∞ N →∞
k/N →0
k
r
X
g
ρr,k,N (X) (77)
k
r=0
we first evaluate the limit with respect to N . Since these conditions
mirror those of Theorem 1 and g(r/k) is independent of N , we can
rewrite (77) as
!
k
rZ k
X
lim
η r (1 − η)k−r fx (x)dx
g
k→∞
r
k
r=0
(78)
#
!
Z "
k
X
k r
r
=
lim
η (1 − η)k−r fx (x)dx,
g
k→∞
k
r
r=0
which according to Weierstrass’ Approximation Theorem simplifies
to
Z
= g(η)fx (x)dx
(79)
= G(f0 , f1 ).
Therefore
lim
lim E
k→∞ N →∞
k/N →0
Ĝk,N (X) − G(f0 , f1 )
2
= 0.
(80)
A PPENDIX C
P ROOF OF T HEOREM 3
Starting with the augmented data set defined in Appendix A,
remember that Φ0k (x∗ ) is the sum of k i.i.d. Bernoulli random
variables. Therefore k1 Φ0k (x∗ ) represents the arithmetic mean of these
output values, and
X
1
1
lim Φ0k (x∗ ) = lim
yi0 = η(x∗ ).
(81)
k→∞ k
k→∞ k
∗
i:xi ∈Nk (x )
Φ0k (x)
Now, since
must be an integer, its probability of equaling r
can be expressed as
"
#
P [Φ0k (x) = r] = P r − 1 < Φ0k (x) ≤ r .
(82)
13
Taking the limit of this expression w.r.t. k, allows us to use (81) to
form
"
#
r−1
r
0
lim P r − 1 < Φk (x) ≤ r = lim P
<η≤
k→∞
k→∞
k
k
r
r − 1
= lim Fη
− Fη
.
k→∞
k
k
(83)
Now if we multiply each side by k, the right hand side takes the
form of Newton’s difference quotient and can be simplified to the
probability density function
lim kP [Φ0k (x) = r] = lim
k→∞
Fη ( kr ) − Fη ( kr − k1 )
1
k
k→∞
= lim
k→∞
d
Fη
dη
k
r
= lim fη
k→∞
r
k
(84)
.
Since ρr,k,N → P [Φ0k (x) = r] as N and k approach infinity in a
linked manner such that k/N → 0 and kr → η ∗
kρr,k,N (X) → fη (η ∗ ).
(85)
R EFERENCES
[1] P. J. Moreno, P. P. Ho, and N. Vasconcelos, “A KullbackLeibler divergence based kernel for SVM classification in multimedia
applications,” in Advances in neural information processing systems,
2003, pp. 1385–1392.
[2] A. B. Hamza and H. Krim, “Image registration and segmentation
by maximizing the Jensen-Rényi divergence,” in Energy Minimization
Methods in Computer Vision and Pattern Recognition. Springer, 2003,
pp. 147–163.
[3] K. E. Hild, D. Erdogmus, and J. C. Principe, “Blind source separation
using Rényi’s mutual information,” Signal Processing Letters, IEEE,
vol. 8, no. 6, pp. 174–176, 2001.
[4] A. Banerjee, S. Merugu, I. S. Dhillon, and J. Ghosh, “Clustering with
Bregman divergences,” The Journal of Machine Learning Research,
vol. 6, pp. 1705–1749, 2005.
[5] S. Ali and S. D. Silvey, “A general class of coefficients of divergence of
one distribution from another,” Journal of the Royal Statistical Society.
Series B (Methodological), pp. 131–142, 1966.
[6] X. Nguyen, M. J. Wainwright, and M. I. Jordan, “On surrogate loss
functions and f-divergences,” The Annals of Statistics, pp. 876–904,
2009.
[7] Y. Wu and P. Yang, “Minimax rates of entropy estimation on large
alphabets via best polynomial approximation,” IEEE Transactions on
Information Theory, vol. 62, no. 6, pp. 3702–3720, 2016.
[8] J. Jiao, K. Venkat, Y. Han, and T. Weissman, “Minimax estimation of
functionals of discrete distributions,” IEEE Transactions on Information
Theory, vol. 61, no. 5, pp. 2835–2885, 2015.
[9] Y. Wu and P. Yang, “Optimal entropy estimation on large alphabets via
best polynomial approximation,” in IEEE International Symposium on
Information Theory. IEEE, 2015, pp. 824–828.
[10] G. Valiant and P. Valiant, “Estimating the unseen: an n/log (n)-sample
estimator for entropy and support size, shown optimal via new CLTs,”
in Proceedings of the forty-third annual ACM symposium on Theory of
computing. ACM, 2011, pp. 685–694.
[11] L. Paninski, “Estimation of entropy and mutual information,” Neural
computation, vol. 15, no. 6, pp. 1191–1253, 2003.
[12] ——, “Estimating entropy on m bins given fewer than m samples,”
IEEE Transactions on Information Theory, vol. 50, no. 9, pp. 2200–
2203, 2004.
[13] A. O. Hero, B. Ma, O. Michel, and J. Gorman, “Alpha-divergence
for classification, indexing and retrieval,” Communication and Signal
Processing Laboratory, Technical Report CSPL-328, U. Mich, 2001.
[14] I. Ahmad and P.-E. Lin, “A nonparametric estimation of the entropy
for absolutely continuous distributions (corresp.),” IEEE Transactions
on Information Theory, vol. 22, no. 3, pp. 372–375, 1976.
[15] L. Györfi and E. C. Van der Meulen, “Density-free convergence
properties of various estimators of entropy,” Computational Statistics
& Data Analysis, vol. 5, no. 4, pp. 425–436, 1987.
[16] A. J. Izenman, “Review papers: recent developments in nonparametric
density estimation,” Journal of the American Statistical Association,
vol. 86, no. 413, pp. 205–224, 1991.
[17] L. Kozachenko and N. N. Leonenko, “Sample estimate of the entropy
of a random vector,” Problemy Peredachi Informatsii, vol. 23, no. 2, pp.
9–16, 1987.
[18] A. O. Hero III and O. Michel, “Robust entropy estimation strategies
based on edge weighted random graphs,” in SPIE’s International
Symposium on Optical Science, Engineering, and Instrumentation.
International Society for Optics and Photonics, 1998, pp. 250–261.
[19] D. Pál, B. Póczos, and C. Szepesvári, “Estimation of rényi entropy and
mutual information based on generalized nearest-neighbor graphs,” in
Advances in Neural Information Processing Systems, 2010, pp. 1849–
1857.
[20] V. Berisha, A. Wisler, A. O. Hero III, and A. Spanias, “Empirically
estimable classification bounds based on a nonparametric divergence
measure,” IEEE Transactions on Signal Processing, vol. 64, no. 3, pp.
580–591, 2016.
[21] K. Kandasamy, A. Krishnamurthy, B. Poczos, L. Wasserman, and
J. Robins, “Nonparametric von Mises estimators for entropies,
divergences and mutual informations,” in Advances in Neural
Information Processing Systems, 2015, pp. 397–405.
[22] J. Silva and S. S. Narayanan, “Information divergence estimation
based on data-dependent partitions,” Journal of Statistical Planning and
Inference, vol. 140, no. 11, pp. 3180–3198, 2010.
[23] G. A. Darbellay, I. Vajda et al., “Estimation of the information by an
adaptive partitioning of the observation space,” IEEE Transactions on
Information Theory, vol. 45, no. 4, pp. 1315–1321, 1999.
[24] S. Geman and C.-R. Hwang, “Nonparametric maximum likelihood
estimation by the method of sieves,” Annals of Statistics, pp. 401–414,
1982.
[25] U. Grenander and G. Ulf, “Abstract inference,” Tech. Rep., 1981.
[26] K. Moon and A. Hero, “Multivariate f-divergence estimation with
confidence,” in Advances in Neural Information Processing Systems,
2014, pp. 2420–2428.
[27] K. R. Moon, K. Sricharan, K. Greenewald, and A. O. Hero, “Improving
convergence of divergence functional ensemble estimators,” in IEEE
International Symposium on Information Theory.
IEEE, 2016, pp.
1133–1137.
[28] B. Póczos, L. Xiong, and J. Schneider, “Nonparametric divergence
estimation with applications to machine learning on distributions,” arXiv
preprint arXiv:1202.3758, 2012.
[29] X. Nguyen, M. J. Wainwright, and M. I. Jordan, “Estimating divergence
functionals and the likelihood ratio by convex risk minimization,” IEEE
Transactions on Information Theory, vol. 56, no. 11, pp. 5847–5861,
2010.
[30] Q. Wang, S. R. Kulkarni, and S. Verdú, “Divergence estimation
for multidimensional densities via k-nearest-neighbor distances,” IEEE
Transactions on Information Theory, vol. 55, no. 5, pp. 2392–2405,
2009.
[31] ——, “Universal estimation of information measures for analog
sources,” Foundations and Trends in Communications and Information
Theory, vol. 5, no. 3, pp. 265–353, 2009.
[32] B. Póczos and J. G. Schneider, “On the estimation of alpha-divergences,”
in International Conference on Artificial Intelligence and Statistics,
2011, pp. 609–617.
[33] J. H. Friedman and L. C. Rafsky, “Multivariate generalizations of
the Wald-Wolfowitz and Smirnov two-sample tests,” The Annals of
Statistics, pp. 697–717, 1979.
[34] N. Henze, M. D. Penrose et al., “On the multivariate runs test,” The
Annals of Statistics, vol. 27, no. 1, pp. 290–298, 1999.
[35] H. Chernoff, “A measure of asymptotic efficiency for tests of
a hypothesis based on the sum of observations,” The Annals of
Mathematical Statistics, pp. 493–507, 1952.
[36] A. Bhattacharyya, “On a measure of divergence between two
multinomial populations,” Sankhyā: The Indian Journal of Statistics, pp.
401–406, 1946.
[37] T. Kailath, “The divergence and Bhattacharyya distance measures in
signal selection,” Communication Technology, IEEE Transactions on,
vol. 15, no. 1, pp. 52–60, 1967.
[38] W. A. Hashlamoun, P. K. Varshney, and V. Samarasooriya, “A tight
upper bound on the Bayesian probability of error,” IEEE Transactions
on Pattern Analysis and Machine Intelligence, vol. 16, no. 2, pp. 220–
224, 1994.
[39] H. Avi-Itzhak and T. Diep, “Arbitrarily tight upper and lower bounds
on the Bayesian probability of error,” IEEE Transactions on Pattern
Analysis and Machine Intelligence, vol. 18, no. 1, pp. 89–91, 1996.
[40] V. Berisha and A. O. Hero, “Empirical non-parametric estimation of the
Fisher information,” IEEE Signal Processing Letters, vol. 22, no. 7, pp.
988–992, 2015.
14
[41] A. Leblanc, “On estimating distribution functions using Bernstein
polynomials,” Annals of the Institute of Statistical Mathematics, vol. 64,
no. 5, pp. 919–943, 2012.
[42] B. C. Turnbull and S. K. Ghosh, “Unimodal density estimation using
Bernstein polynomials,” Computational Statistics & Data Analysis,
vol. 72, pp. 13–29, 2014.
[43] S. Ghosal, “Convergence rates for density estimation with Bernstein
polynomials,” Annals of Statistics, pp. 1264–1280, 2001.
[44] G. Igarashi and Y. Kakizawa, “On improving convergence rate of
Bernstein polynomial density estimator,” Journal of Nonparametric
Statistics, vol. 26, no. 1, pp. 61–84, 2014.
[45] A. Tenbusch, “Two-dimensional Bernstein polynomial density estimators,” Metrika, vol. 41, no. 1, pp. 233–253, 1994.
[46] G. J. Babu and Y. P. Chaubey, “Smooth estimation of a distribution
and density function on a hypercube using Bernstein polynomials for
dependent random vectors,” Statistics & probability letters, vol. 76, no. 9,
pp. 959–969, 2006.
[47] G. G. Lorentz, Bernstein polynomials. American Mathematical Soc.,
2012.
[48] S. Bernstein, “Démonstration du théorème de Weierstrass fondée sur le
calcul des probabilités,” Comm. Soc. Math. Kharkow, Ser, vol. 2, no. 13,
pp. 49–194, 1912.
[49] J. A. Gubner, Probability and random processes for electrical and
computer engineers. Cambridge University Press, 2006.
[50] R. T. Farouki, “Legendre–Bernstein basis transformations,” Journal of
Computational and Applied Mathematics, vol. 119, no. 1, pp. 145–160,
2000.
[51] B. Efron, T. Hastie, I. Johnstone, R. Tibshirani et al., “Least angle
regression,” The Annals of statistics, vol. 32, no. 2, pp. 407–499, 2004.
[52] S. Arya, D. M. Mount, N. S. Netanyahu, R. Silverman, and A. Y. Wu,
“An optimal algorithm for approximate nearest neighbor searching fixed
dimensions,” Journal of the ACM (JACM), vol. 45, no. 6, pp. 891–923,
1998.
[53] D. J. Sutherland, L. Xiong, B. Póczos, and J. Schneider, “Kernels on
sample sets via nonparametric divergence estimates,” arXiv preprint
arXiv:1202.0302, 2012.
[54] Z. Szabó, “Information theoretical estimators toolbox,” Journal of
Machine Learning Research, vol. 15, pp. 283–287, 2014.
[55] S. Kullback, Information theory and statistics. Courier Corporation,
1997.
[56] A. Wisler, V. Berisha, K. Ramamurthy, D. Wei, and A. Spanias,
“Emperically-estimable multi-class performance bounds,” in IEEE
International Conference on Acoustics, Speech and Signal Processing
(ICASSP). IEEE, 2016.
[57] K. Fukunaga, Introduction to statistical pattern recognition. Academic
press, 1990.
[58] L. Devroye, L. Györfi, and G. Lugosi, A probabilistic theory of pattern
recognition. Springer Science & Business Media, 2013, vol. 31.
[59] A. W. Van der Vaart, Asymptotic statistics. Cambridge university press,
2000, vol. 3.
| 7cs.IT
|
Kısa Dönem Uzam-Zamansal Trafik Tahmini
Short-Term Spatio-Temporal Traffic Forecasting
Akın Taşcıkaraoğlu1 , Fatma Yıldız Taşcıkaraoğlu2 , İbrahim Beklan Küçükdemiral2
1
Elektrik Mühendisliği Bölümü
Yıldız Teknik Üniversitesi, İstanbul
arXiv:1608.08053v1 [cs.SY] 5 Aug 2016
[email protected]
2
Kontrol ve Otomasyon Mühendisliği Bölümü
Yıldız Teknik Üniversitesi, İstanbul
{fayildiz,beklan}@yildiz.edu.tr
Özetçe
Şehir içi yollarda yaşanan tıkanıklık, gecikme ve çevre
problemlerini en aza indirmek amacıyla gerçekleştirilen çalışmaların önemi özellikle son yıllarda oldukça artmıştır. Bu çalışmalar içerisinde, kısa dönem trafik akımı ve ortalama araç hızı
tahmini yöntemleri; uygulanabilirliklerinin kolay olması, farklı
amaçlar için etkin bir şekilde kulllanılabilmeleri ve maliyetlerinin oldukça düşük olması nedeniyle ön plana çıkmaktadırlar.
Trafik ağındaki istenen noktalara ait bağ akımlarının ve araç
hızlarının gelecekteki olası değerlerinin tahmin edildiği bu yöntemlerin, trafik yönetiminde yaşanabilecek sorunları öngörerek
trafik sıkışıklığını azaltmaya yardımcı olduğu çok sayıda çalışmada belirtilmektedir. Bu yayında, bir bağdaki ortalama araç
hızı tahminleri için çok sayıda noktadan alınan geçmiş verileri
göz önüne alan bir uzam-zamansal yaklaşım sunulmaktadır. Belirtilen yaklaşım, her bir aşamada en faydalı verileri belirleyerek yalnızca bu verilerin giriş veri kümesine dahil edilmesini
sağlayan bir algoritma içermektedir. Algoritma içerisinde seyrek matrislerin elde edilmesi ile tahmin doğruluğunun arttırılabilmesi sağlanırken tahminlerin gerçekleştirilme sürelerinin de
literatürde yer alan yöntemlere oranla kayda değer derecede iyileştirilmesi hedeflenmektedir.
Abstract
The studies carried out with the objective of minimizing
the effects of congestion, delay and environment problems on
the transportation network have gained increasing importance
in the last years. Among these studies, short-term traffic flow
and average vehicle speed forecasting methods have come into
prominence due to their easy implementations, efficient usage
on different areas and cost-effectiveness. A large number of studies have reported that these methods, in which the expected
future values of link flows and average speeds are forecasted
in desired points, can reduce the traffic congestion by anticipating the problems in traffic management. In this paper, a spatio-
temporal approach accounted for historical traffic characteristics data collected from a large number of points is presented
for average speed forecasts in a given link. The proposed approach includes an algorithm that enables to take into account
the most informative data in an input set by determining them
for each stage. It is aimed to increase the forecasting accuracy
by using sparse matrices in the algorithm while decreasing the
calculation times significantly compared to the similar methods
presented in the literature.
1. Giriş
Gelişmiş ülkelerdeki büyük kentlerde yaşanan nüfus ve araç sahipliğindeki hızlı artışın ve şehirleşmenin ulaştırma ağı üzerindeki etkilerine bağlı olarak yaşanan tıkanıklık, gecikme, güvenlik ve çevre problemlerini en aza indirmek amacıyla gerçekleştirilen ve bilişim teknolojileriyle desteklenen faaliyetler, Akıllı
Ulaşım Sistemleri olarak adlandırılmaktadır. Şehirlerde yaşanan trafik tıkanıklığının insanların yaşam kalitesi üzerindeki
doğrudan etkileri nedeniyle akıllı ulaşım sistemlerinin, gelecekteki akıllı şehir planlamalarında çok önemli bir bileşen olarak
yer alması beklenmektedir. Bu nedenle, trafiğin durumundaki
beklenen gelişmeler, trafiğin güvenli ve ekonomik bir şekilde
yönetilmesi açısından oldukça büyük bir öneme sahiptir. Belirli
trafik karakteristiklerinin önceden bilinmesi, trafik yönetim sistemlerine veri desteği sağlamalarının yanı sıra kullanıcılara da
trafik şartları hakkında önemli bilgiler verebilir. Ancak sürekli
değişen yapısı nedeniyle, trafik tıkanıklığı ve tıkanıklığı etkileyen faktörler arasındaki etkileşimleri modellemek için gelişmiş
yöntemlere ihtiyaç duyulmaktadır.
Bu amaçla geliştirilen tahmin yöntemlerinin yapısı öncelikle yolun türüne (otoyol, şehir içi arter, vb.) bağlıdır. Yol türlerinin her birinin kendine özgü bir karakteristiği vardır. Örneğin,
ışıklı kavşaklar ve yaya geçişleri gibi çeşitli kısıtlar, otoyollara
oranla şehir içi arterlerdeki tahmin işlemini daha karmaşık bir
hale getirmektedir [1]. Şehir merkezlerine yakın olan ve çok
sayıda katılıma sahip olan otoyollardaki trafik karakteristiklerinin tahminleri de benzer şekilde şehirlerarası otoyollara kıyasla daha zor bir işlem olarak nitelendirilebilir. Model seçimin-
deki bir diğer önemli faktör ise gerekli tahmin aralığıdır ve bu
süre çoğunlukla tahminlerin kullanılacağı uygulama alanı göz
önüne alınarak belirlenmektedir. Bir genelleme yapacak olursak, yüksek doğrulukta modellemeye olanak sağlamalarından
dolayı, daha küçük tahmin adımlarının gerçek uygulamalar için
daha değerli olduğu söylenebilir.
Akım, yolculuk zamanı, kuyruk uzunluğu, hız ve yoğunluk gibi farklı trafik karakteristikleri arasında trafik akımı ve hız
tahminleri, trafik tıkanıklığının azaltılması ve akıllı ulaşım sistemlerinde ulaşım hareketliliğinin arttırılması açısından anahtar
bir role sahiptir. Özellikle, birkaç dakikadan birkaç saate kadar olan kısa dönem trafik akımı ve ortalama hız tahminleri,
dinamik trafik kontrolünü destekleyerek, gecikmeleri ve trafik
tıkanıklığını azaltmada oldukça etkili olmaktadırlar. Belirtilen
parametrelerin tahmininde kullanılan klasik yöntemler kararlı
trafik şartları için belli bir doğrulukta tahminler sağlamaktadırlar. Ancak, trafiğin en yoğun olduğu saatlerdeki yüksek taşıt sayısı, yetersiz yollar, kazalar, koordinasyonu sağlanamamış trafik ışıkları ve kötü hava şartları gibi çok sayıda etkenin neden
olduğu yoğun trafik durumları için bu yöntemler çoğunlukla hatalı tahminler vermektedirler.
Birbirine fiziksel olarak bağlı yollardan alınan veriler, trafik
şartlarındaki bu kısa ve uzun süreli değişimleri dikkate alarak
tahmin yöntemlerinin doğruluğunu iyileştirebilirler. Özellikle,
trafiğin akış yönüne göre daha başlarda bulunan bağlardan alınan verilerin, sonraki bağların tahmini üzerinde genellikle güçlü
bir etkisi olacaktır. Bu amaçla, bir ön araştırma gerçekleştirilerek belirli bir ağdaki trafik akımına ve ortalama hıza en fazla
etki eden faktörler belirlenebilir ve yalnızca bu faktörler tahmin algoritmalarında kullanılabilir. Alternatif olarak, tıkanık ve
tıkanık olmayan trafik şartları için çoklu rejim (multi-regime)
yöntemleri de tercih edilebilir. Ancak, trafik akımı ile trafik şartlarındaki beklenen ve beklenmeyen olaylar arasındaki karmaşık
ilişkinin tam olarak belirlenebilmesi genellikle mümkün olmamaktadır. Üstelik bu ilişki zamanla sürekli olarak değişmektedir
ve böylece bir model elde etmek daha da zor hale gelmektedir.
Bu sorunların üstesinden gelmek amacıyla, ağ üzerindeki çeşitli noktalardaki belli bir sayıda aday değişken içerisinden her
bir adımda en uygun girişleri belirleyen ve bir sonraki tahminde
yalnızca bu verileri dikkate alan tahmin yaklaşımları literatürde
son yıllarda oldukça ön plana çıkmıştır. Uzam-zamansal modeller olarak adlandırılan bu gelişmiş modeller, geniş bir ağ üzerindeki çeşitli bilgileri dikkate almaları nedeniyle, trafikteki değişen şartlara kolaylıkla uyum sağlayabilirler. Bu nedenle, tekrarlanmayan geçici şartlar dahil farklı trafik şartları için sürdürülebilir bir tahmin kalitesi sağlayabilirler.
Literatürde özellikle son yıllarda, trafik ağlarındaki verilerin toplanması için kullanılmaya başlanan gelişmiş teknolojiler ile birlikte geniş bir aralıkta uzamsal ve zamansal trafik
verilerinin elde edilebilmesi, trafik akımı ve ortalama araç hızı
tahmini alanında bu verileri kullanan çok sayıda yöntemin geliştirilmesine olanak sağlamıştır. Stathopoulos vd. [2] bir şehir
içi ışıklı kavşak için çok değişkenli zaman serileri modellerini
kullanarak kısa dönemli akım tahminleri gerçekleştirmişlerdir.
Benzer yapıdaki bir diğer çok değişkenli uzam-zamansal model ise Dong vd. [3] tarafından geliştirilerek, tıkanık ve tıkanık
olmayan trafik durumlarında hız ve akım tahminleri için modelin etkinliği ayrı ayrı test edilmiştir. Sun vd. [4] ve Tselentis
vd. [5] uzamsal ve zamansal trafik akımı tahminlerinde Baye-
sian ağları yöntemini kullanmışlardır. Bulanık mantık tabanlı
bir şehir içi trafik tahmini yöntemi ise Dimitriou vd. [6] tarafından önerilmiştir. Bu yayında önerilen tahmin yöntemine benzer
olarak, Sun ve Zhang [7] bir ağdaki bağ akımları arasındaki korelasyonları dikkate alarak belirli bir ağ için yalnızca yüksek
korelasyona sahip ağlardan gelen verileri giriş verisi kümesine
dahil etmişlerdir. Sonuç olarak özellikle birbirine yakın olan ağlardan alınan verilerin tahmin doğruluğu üzerinde olumlu etkisi
olduğunu belirtmişlerdir. Ancak yalnızca belirli ağlardan alınan
verilerin tahmin algoritmalarında kullanılması, trafikteki aniden
meydana gelen beklenmeyen durumlarda çoğunlukla başarısız
tahminler üretmektedirler. Bu amaçla bu yayında önerilen yöntem içerisinde, bağlar arasındaki korelasyonlar sürekli olarak
hesaplanarak yalnızca mevcut durum için en yüksek korelasyona sahip bağlardan gelen veriler dikkate alınmaktadır. Trafik
karakteristiklerine ait verilerin model içerisine dahil edilmesi işlemi her bir tahmin ufkunda tekrar değerlendirilmektedir ve bu
sayede yalnızca bazı bağların tahminler üzerindeki etkilerini sürekli olarak kullanan literatürdeki modellerden [8] daha yüksek
tahmin doğruluklarının elde edilmesi hedeflenmektedir. Uzamsal verinin trafik tahminindeki kullanımı ve önemi Vlahogianni
vd. [9] tarafından geniş bir çerçevede incelenmiştir.
Bu çalışmada, uzamsal ve zamansal trafik akım hızı tahmini yöntemlerinin, özellikle ortalama yolculuk zamanlarında
ve tüketilen yakıt miktarlarında sağladığı iyileşmeler göz önüne
alınarak, yüksek tahmin doğruluğuna sahip bir uzam-zamansal
ortalama hız tahmini yöntemi geliştirilmiştir. Tahmin modeli
içerisine farklı bağlara ait verilerin dahil edilebilmesi için, matematiksel olarak basit bir çözüm sunmaları nedeniyle çok değişkenli zaman serileri modelleri kullanılmıştır. Önerilen yöntemin iki ana katkısının olacağı düşünülmektedir. Trafik akım
hızlarının gelecekteki değerlerinin tahmin edilmesi ve bu sayede trafik tıkanıklığı açısından sorun yaşaması muhtemel olan
bölgelerde, trafik ışıklarının yeşil sürelerine müdahale edilmesi
ve/veya taşıt sahiplerinin trafik tabelaları ve akıllı telefon uygulamaları (Google maps, Yandex navigator, vb.) ile bilgilendirerek alternatif yollara yönlendirilmesi, hedeflenen katkılardan
ilkini oluşturmaktadır. Diğer katkı ise özellikle trafikte yaşanacak geçici sorunların çözümünde yöntemin sağlayacağı faydadır. Örneğin, literaturde geleneksel olarak kullanılan modelleme
ve parametre kestirimi yöntemleri ile zamanlama planı belirlenen bir yolda meydana gelebilecek bir trafik kazası, yapılan
planlamaların etkinliğini önemli ölçüde azaltacaktır. Bu ve trafik akımındaki benzeri geçici sorunların etkileri, yapılacak olan
tahminler ile belirli bir süre önce öngörülerek en aza indirilebilir. Modelleme yaklaşımlarının aksine tahmin yöntemleri dinamik olarak güncellenen veri kümelerini kullandıkları için bahsedilen geçici sorunların etkilerinin belirlenmesinde literatürde
sıklıkla kullanılmaktadırlar. Mevcut yayın kapsamında geliştirilen tahmin yönteminin, hem modelleme ve parametre kestirimi
yaklaşımları ile bir arada kullanılabileceği hem de geçici durumlarda bu iki yönteme alternatif olabileceği söylenebilir. Her
iki durumda da tek amaç, trafik kontrol yöntemlerinin etkinliğini en üst seviyeye çıkarmaktır. Bu sayede, mevcut trafik ağları
kullanılarak, insanların trafikte geçirdiği süreleri en aza indirebilmek, trafiğin insanlar üzerindeki sosyal ve ekonomik etkilerini azaltabilmek, taşıtların neden olduğu karbon emisyonlarının miktarını düşürebilmek ve trafiğin neden olduğu diğer tüm
olumsuz etkileri azaltabilmek daha mümkün hale gelecektir.
∗
2. Uzam-Zamansal Tahmin Modeli
Bu yayında bir kısa dönem ortalama hız tahmini yöntemi geliştirilmiştir. Geliştirilen yöntem, önceki bölümde belirtildiği gibi
farklı noktalardan alınan veriler arasındaki korelasyonların dikkate alınması prensibine dayanmaktadır. Şekil 1’de bu yayında
kullanılan veri seti içerisindeki üç farklı bağa ait beş dakika ortalamalı hız verileri gösterilmektedir. Bir güne ait verilerin yer
aldığı bu şekilde birbirine yakın farklı yollardaki ortalama hız
değerlerinin benzer karakteristiğe sahip olduğu açıkça görülmektedir. Üç bağa ait değerler bazen aynı anda azalırken veya
artarken bazı zamanlarda ise bu değişimler belli bir zaman gecikmesi ile diğer bağlarda gözlemlenmektedir. Özellikle zaman
gecikmelerine sahip bilgilerin önerilen tahmin yönteminin etkinliği üzerinde oldukça önemli bir etkisi bulunmaktadır.
Bag 1
Bag 2
Bag 3
Ortalama hiz (km/saat)
60
55
50
45
40
35
30
25
50
100
150
200
Zaman (5 dakikalik ortalama)
250
Şekil 1: Farklı bağlara ait ortalama hız ölçümleri.
Geliştirilen yöntemin temeli, yüksek doğrulukta tahminler
sağlaması ve basit yapısı nedeniyle literatürde kısa süreli tahminler için sıklıkla kullanılan Otoregresif Modele (Autoregressive - AR) dayanmaktadır [10]. Bu modeller ayrıca, geliştirilecek olan modelin ana hedefi olan, çok sayıda bağdan alınan ve
farklı trafik karakteristiklerine ait olan tüm verilerin model içerisine dahil edilmesi fikrine olanak sağlamaktadır. AR modeller, bir sistemin çıkış değişkeninin, sisteme ait geçmiş verilerin
ağırlıklandırılmış doğrusal bir kombinasyonu olarak elde edilebileceği varsayımına dayanırlar. Bu yaklaşımı, yayındaki geniş
bir veri setinin kullanılması hedefine uygun olarak çok değişkenli sistemlere uygulayabiliriz. Burada P (p = 1, 2, . . . , P )
adet çıkış değişkenine (örneğin, P adet bağın ortalama hız değerlerine) ait ölçümlere sahip olduğumuzu varsayalım. Bu çıkış
değişkenleri, P = V × S olmak üzere S (s = 1, 2, . . . , S)
adet düğümden alınan V (v = 1, 2, . . . , V ) adet değişkene ait
ölçümler olabilir. t (t = 1, 2, . . . , M + n) örnek zamanında
∗ ∗
s’inci bağda ölçülen v’ninci değişkenin ölçümünü y v ,s olarak tanımlayalım ve bu durumda hedef (tahmin edilmek istenen)
çıkış değişkeni ytv,s olsun. Sonuç olarak bir Çok Değişkenli AR
model (Multivariate Autoregressive - M-AR), diğer adıyla Vektör Otoregresif model (1)’deki gibi elde edilebilir:
ytv
∗
,s∗
=
V,S n
X
X
v,s v,s
yt−i
xi + evt
∗
,s∗
,
(1)
v=1 i=1
s=1
Burada xv,s
(∀i, v, s) ilgili regresyon katsayılarını, n moi
∗
del derecesini ve evt ,s Gauss gürültüsünü (Gaussien noise)
temsil etmektedir. Eşitlik (1), N := nP olacak şekilde Eşitlik (2)’de gösterildiği gibi genişletilebilir. Model eğitimi aşamasında amaç, e gürültü bileşenini de dikkate alarak b ∈ RM ve
A ∈ RM ×N gözlemlerini en iyi şekilde açıklayan bir x ∈ RN
katsayı vektörünü belirlemektir. Eşitlik (2)’den açıkça görülebileceği üzere x katsayıları bir blok yapı içerisinde toplanmışlardır. Diğer bir ifadeyle, her bir değişkene ait katsayılar n uzunluğundaki bir blok vektöre karşılık gelmektedir.
Farklı trafik karakteristiklerine ait olan ve farklı noktalardan
ölçülen değerler birbirinden farklı büyüklüklere sahip olduklarından, b vektörü ve A matrisindeki tüm değerler bir normalizasyon işlemi ile 0-1 aralığındaki değerlere dönüştürülmektedir.
Bu sayede, bu değerlerden herhangi birinin x vektörü katsayılarının belirlenmesindeki etkisinin diğer değişkenlere oranla çok
daha fazla veya az olması önlenmektedir.
Eşitlik (2)’de tüm n derece değerlerinin eşit olduğu varsayılmıştır. Başka bir ifadeyle, hedef bağ akımı ve diğer bağlardaki değişkenlerin aynı n derecesindeki AR modelleri ile
ilişkilendirildiği kabul edilmiştir. Ancak her bağın ve bağlara
ait her değişkenin (akım, yolculuk süresi, kuyruklanma uzunluğu, araç hızları, yoğunluk, vb.) tahmin değerlerine olan etkisinin farklı olacağı açıktır. Bu nedenle Eşitlik (2)’de verilen
M-AR modelinin farklı n derecelerine sahip değişkenler için
düzenlenmiş hali, ni (i = 1, 2, . . . , P ) i’ninci bağa ait derece,
P
nmax ≥ maxi ni ve N := P
i=1 ni olmak üzere Eşitlik (3)’de
verilmiştir. Bu modelde her biri farklı blok uzunluklarına sahip
olan düzgün dağılmamış blok katsayısı matrisi x elde edilir. Bu
sayede hedef akım çıkışı ile her bir bağa ait değişkenler arasındaki korelasyonlara göre tüm değişkenler arasında bir ayrım
yapılması sağlanmıştır.
Verilen modellerde x vektör katsayıları farklı yollarla elde
edilebilir. Örneğin en küçük kareler yöntemi kullanılarak her
bir tahmindeki hataların karelerinin toplamı en küçük olacak
şekilde gerekli katsayılar elde edilebilir. Ancak hız tahminleri
için geliştirilecek olan model çok sayıda farklı bağın her birinden alınacak olan trafik karakteristiklerine ait verileri içerecektir. Dolayısıyla tahminlerin gerçekleştirilebilmesi için harcanacak süre kısa dönemli bir tahmin algoritması için oldukça fazla
olacaktır. Kısa süreli tahminlerde, tahmin süresinin kayda değer derecede artması, elde edilen yüksek tahmin doğruluklarının önemini azaltmaktadır. Bu nedenle, bu yayında geliştirilen
modelde kullanılacak yöntemlerle seyrek (sparse) x vektörleri
elde edilmeye çalışılmaktadır. Burada bir varsayım yapılarak
çok sayıda bağdan alınan hız verileri arasında yalnızca bazılarının, tahmin edilmeye çalışılan bağa ait ortalama hız ile arasında güçlü bir korelasyon olduğu göz önüne alınmaktadır. Fiziksel olarak da anlamlı olacak olan bu belirli değerlerin tahmin
üzerindeki olumlu etkileri düşünülerek tahmin aşamasında yalnızca bu değerlerin dikkate alınması sağlanmaktadır. Başka bir
ifade ile tahmin doğruluğundan ödün vermeksizin x vektör elemanlarının mümkün olduğunca büyük bir kısmının sıfır olması
sağlanmaktadır. Literatürde bu vektörler blok-seyrek (blocksparse) olarak adlandırılmaktadır. Burada en önemli soru sınırlı
sayıdaki sıfır olmayan x vektörü elemanlarının nasıl belirleneceğidir. Literatürde rüzgar hızının tahmin edildiği çalışmalarda
( [11,12]), sıkıştırmalı algılama (Compressive Sensing) yöntemi
kullanarak hesaplanan x vektörleri sayesinde tahmin doğruluğu
açısından elde edilen kayda değer sonuçlar göz önüne alınarak
?
?
v ,s
yn+1
y v? ,s?
n+2
..
.
? ?
v ,s
yn+M
|
{z
b∈RM
=
}
yn1,1
1,1
yn+1
..
.
1,1
yn+M
−1
...
..
.
..
.
...
y11,1
..
.
..
.
1,1
yM
|
...
ynV,S
...
...
..
.
V,S
yn+1
..
.
...
...
{z
V,S
yn+M
−1
...
...
y1V,S
..
.
..
.
V,S
yM
A∈RM ×N
x1,1
1
..
.
x1,1
n
Blok 1
v? ,s?
en+1
? ?
ev ,s
+ n+2
..
.
? ?
v ,s
e
n+M
{z
|
M
e∈R
Blok P
.
.
.
.
.
.
} xV,S
1
.
..
xV,S
n
| {z }
(2)
}
x∈RN
?
?
,s
ynv max
+1
y v? ,s?
nmax +2
..
.
? ?
,s
ynv max
+M
|
{z
b∈RM
=
}
yn1,1
max
yn1,1
max +1
..
.
yn1,1
max +M −1
...
..
.
..
.
...
...
..
.
..
.
...
...
..
.
..
.
...
|
yn1,1
max −n1 +1
..
.
..
.
yn1,1
max −n1 +M
{z
...
ynV,S
max
ynV,S
max +1
..
.
...
A∈RM ×N
ynV,S
max +M −1
...
..
.
..
.
...
ynV,S
max −nP +1
..
.
..
.
ynV,S
max −nP +M
x1,1
1
..
.
..
.
..
.
1,1
xn1
Blok 1
v? ,s?
enmax +1
ev? ,s?
nmax +2
+
..
.
? ?
v ,s
e
nmax +M
{z
|
e∈RM
Blok P
.
.
.
}
V,S
x1
.
.
.
xV,S
nP
| {z }
x∈RN
(3)
bu yayında blok-seyrek x vektörleri (4)’te verilen optimizasyon
problemi çözülerek hesaplanmıştır.
min kb − Axk2
x
(x blok-seyrek olmak üzere).
(4)
3. Benzetim Çalışmaları
Önerilen yöntemin etkinliğini ölçmek amacıyla literatürde çok
sayıda yayında yer verilen Kaliforniya Otoyol Performans Ölçüm Sistemine (PEMS v14.1) ait bir veri seti kullanılmıştır. Bu
sistem Kaliforniya’da bulunan otoyollar üzerindeki çok sayıdaki noktadan gerçek zamanlı verileri kaydetmektedir. İnternet
üzerinden erişilebilen sayfasında [13] bu veriler beşer dakikalık çözünürlükte sağlanmaktadır. Literatürdeki ilgili yayınların
çok büyük bir kısmında bu veriler yarım saat veya bir saatlik
ortalamalara dönüştürülerek kullanılmaktadırlar. Bu çalışmaların temel hedefi genellikle daha yüksek zamansal çözünürlüğe
sahip veriler kullanarak daha uzun süreli tahminler gerçekleştirmektedir. Ayrıca düzeltme etkisi (smoothing effect) nedeniyle
uzun sürelerde daha kararlı değerler elde edilmektedir ve böylece daha yüksek tahmin doğrulukları sağlanmaktadır. Bu çalışmada ise beş dakikalık veriler doğrudan kullanılarak 30 dakikaya kadar ortalama hız değerleri tahmin edilmiştir. Tahminler
için 9 saatlik (5:00-14:00 arası 108 adet beş dakikalık değer) bir
eğitim seti kullanılarak aynı gün içerisindeki bu zamanı takip
eden 9 saatin (14:00-23:00 arası degerler) tahmini gerçekleştirilmiştir. Özellikle kısa süreli trafik bilgisine ihtiyaç duyulan
uygulamalarda (navigasyon yazılımları, trafiğin durumunu gösteren dinamik trafik tabelaları, vb.) bu yöntemin çok daha etkin
olacağı düşünülmektedir.
Önerilen yöntemin etkinliğini test edebilmek amacıyla öncelikle tahminler farklı yöntemlerle gerçekleştirilmiştir. Bu
amaçla, önerilen modelin temelini oluşturan AR modeli ve yapay sinir ağlarına (YSA) dayanan başka bir modele ait tahminler sırasıyla Şekil 2(a) ve 2(b)’de verilmiştir. Veriler arasında
doğrusal bir ilişkinin tam olarak kurulamadığı trafik akımı ve
ortalama hız gibi değişkenler için literatürde genellikle YSA
tabanlı yöntemler daha iyi sonuçlar vermektedirler. Ancak bu
çalışmada tüm modeller için sınırlı sayıda (108 adet 5 dakikalık
veri) bir eğitim seti kullanıldığından, başarısı büyük oranda eğitim setinin boyutuna bağlı olan YSA modeli için AR modeline
kıyasla gerçek değerlerden daha uzak sonuçlar elde edilmiştir.
Önerilen tahmin yöntemi ile elde edilen sonuçlar ise Şekil 3’de verilmiştir. Şekilden görülebileceği gibi yakın çevredeki noktalardan alınan ölçümlerin tahmin modeli içerisine dahil edilmesi özellikle ani değişimlerin tahmininde büyük bir
fayda sağlamaktadır. Önerilen yaklaşımın bir diğer özelliği ise
zamanla değişen veri ile birlikte en yüksek modelleme doğruluğu için x vektör katsayılarının sürekli olarak tekrar hesaplanmasıdır. Başka bir ifadeyle, tahminlerin gerçekleştirildiği zamana ait olan ölçümlerin giriş verisi olarak kullanılmasıyla, her
}
Olcum degeri
Tahmin degeri
50
45
40
35
Olcum degeri
Tahmin degeri
55
Ortalama hiz (km/saat)
Ortalama hiz (km/saat)
55
30
50
45
40
35
30
20
40
60
80
Zaman (5 dakikalik ortalama)
100
20
40
60
80
Zaman (5 dakikalik ortalama)
100
(a) AR Model
Şekil 3: Önerilen yöntem ile elde edilen tahminler.
Olcum degeri
Tahmin degeri
Ortalama hiz (km/saat)
55
50
45
dikkate aldığı ve NRMSE’nin tahmin edilen değişkenin genliğinden bağımsız olarak yüzde bir değer sağladığı söylenebilir.
Tablo incelendiğinde önerilen yöntemin her üç kriter için de diğer yöntemlerden daha başarılı olduğu ifade edilebilir.
40
Tablo 1: Farklı yöntemlere ait hata ölçütlerinin karşılaştırılması
35
Tahmin yöntemi
MAE
[km/saat]
RMSE
[km/saat]
NRMSE
[%]
YSA
AR
Önerilen yöntem
2,62
2,31
1,74
3,25
2,99
2,12
13,91
12,79
9,05
30
20
40
60
80
Zaman (5 dakikalik ortalama)
100
(b) YSA Modeli
Şekil 2: Farklı tahmin modelleri ile elde edilen tahminler.
0.5
0.4
0.3
0.2
Genlik
bir tahmin ufku için yeni bir x vektörü elde edilmektedir ve bir
sonraki tahmin ufku için gerçekleştirilecek tahminlerde bu vektör kullanılmaktadır.
Yol üzerindeki algılayıcılardan son olarak alınan veriler ile
tahminlerin sürekli olarak güncellenmesi ve bu sayede ağın
akım yönüne göre başındaki ve sonundaki bağlar arasındaki trafik şartlarının değişiminin incelenmesi; sabah ve akşam zirve
saatler, gece yarısı saatleri ve büyük etkinlikler (spor müsabakalar, gösteri, vb.) gibi geçici zamanlar ile kazalar ve olumsuz
hava şartları gibi normal olmayan durumlar için literatürdeki
benzer yöntemlere göre daha iyi sonuçlar vermektedir. Ayrıca
önerilen modelde yinelemeli (recursive) bir yaklaşım benimsenmiştir. Burada model eğitimi aşamasından sonraki ilk değer olan n + M + 1 değeri tahmin edildikten sonra bu değer
n + M + 2 değerinin tahmininde kullanılmaktadır. Bu işlem
tahmin ufku boyunca devam ettikten sonra tüm A matrisi yeni
ölçüm değerleri ile güncellenmektedir.
Önerilen tahmin yönteminin ve kullanılan diğer iki adet
yöntemin etkinliklerinin daha iyi anlaşılabilmesi açısından yöntemler Tablo 1’de çeşitli hata ölçütleri kullanılarak karşılaştırılmıştır. Bu karşılaştırmalarda literatürde en yaygın olarak
kullanılan ortalama mutlak hata (MAE), ortalama karesel hata
(RMSE) ve normalleştirilmiş ortalama karesel hata (NRMSE)
ölçütleri kullanılmıştır. Özetle, MAE’nin tahminlerin istenen
süre boyunca doğruluğu hakkında bilgi verdiği, RMSE’nin yüksek hataların karesini alması nedeniyle özellikle bu hataları
Şekil 4’de bir tahmin periyodunda (6 adet 5 dakikalik değer) önerilen modelin hesapladığı ve tahminlerde kullandığı x
vektörü gösterilmektedir. Kesik çizgilerle ayrılan bölümlerden
görülebileceği gibi her bir bağa ait model derecesi farklıdır. Burada tahmin algoritmasının, daha önceden belirtildiği gibi, o
andaki tahminler için yalnızca bazı noktalardaki değerleri (sıfır olmayan değerler) dikkate aldığı görülebilir. Bu noktalar her
bir yarım saatlik sürede tahmin yöntemi tarafından tekrar belirlenmektedir. Tahmin süresi boyunca değişen x vektörü incelendiğinde bazı noktaların her zaman tahmine bir katkı sağladığı,
bazı noktaların ise tam aksine hiçbir zaman bir değer almadığı
görülmüştür. Verilerin alındığı internet sayfasında tüm noktaları
0.1
0
−0.1
−0.2
−0.3
−0.4
0
20
40
60
80
Endeks
Şekil 4: Seyrek katsayı vektörü.
gösteren bir harita bulunmamaktadır. Ancak yalnızca x vektörünün zamanla değişimine bakarak her bir noktanın, tahminlerin
gerçekleştirildiği noktaya olan uzaklığı ve iki nokta arasındaki
ilişki hakkında bir yorum yapabilmek mümkündür. Verilerin
alındığı noktaların harita üzerindeki konumlarını bilmeksizin,
yalnızca veriler arasındaki korelasyonlara bakarak hangi verilerin giriş setine dahil edilebileceğine karar verebilme özelliği,
önerilen yöntemin en önemli özelliklerinden biri olarak gösterilebilir. Bu özelliği sayesinde önerilen yöntemin, özellikle eksik
ve hatalı verilerin bulunduğu veri setlerinde, literatürdeki benzer yöntemlere kıyasla çok daha başarılı olacağı öngörülmektedir.
4. Sonuçlar
Literatürde özellikle yenilenebilir enerji kaynaklarından elde
edilebilecek güç miktarlarının tahmini için son birkaç yılda oldukça fazla kullanılmaya başlayan uzam-zamansal yöntemlerin
etkinlikleri göz önüne alınarak bu yayında trafiğin durumunu
ifade etmek için en yaygın olarak kullanılan ortalama araç hızlarının gelecek tahminleri için bir uzam-zamansal yöntem sunulmaktadır. Ağ üzerindeki farklı noktalardan alınan ve çok sayıda parametreye ait olan ölçümleri tahmin aşamasında model
içerisine dahil ederek mevcut verilerden en üst seviyede yararlanmayı saglayan bu yöntemin, farklı noktalara ait çeşitli trafik karakteristikleri arasındaki korelasyonları kullanarak tahmin
doğruluğunu önemli derecede arttırdığı gösterilmiştir. Trafik tıkanıklığının etkilerinin ardışık yollar üzerinde belirli bir zaman
gecikmesiyle görülmesi, trafik tahminlerinde bu yöntemlerin etkin olmasını sağlamıştır. Literatürde en iyi girişlerin bir ön analiz ile belirlendiği tahmin yaklaşımlarının aksine bu yayında
önerilen yöntem her bir tahmin ufku için yalnızca bir önceki
tahminde en iyi sonuçları veren değişkenleri ve değerleri model içerisine dahil etmektedir. Bu nedenle, bir bölgedeki trafik
akımındaki değişen trafik şartlarının ve hava koşullarının etkilerinin tahmin aşamasında dikkate alınması sağlanmıştır. Çeşitli
tahmin yöntemleri ile yapılan karşılaştırmalar önerilen yöntemin etkinliğini göstermektedir. Önerilen modelin etkinliğini arttırmak amacıyla giriş veri setinin yolculuk süreleri, meşguliyet
ve kuyruklanma uzunluğu gibi trafik karakteristikleri ile sıcaklık ve yağış miktarı gibi hava değişkenleri kullanarak genişletilmesi yakın gelecekteki bir çalışma olarak planlanmaktadır.
deling and prediction,” Transportation Research Part C:
Emerging Technologies, vol. 11, no. 2, pp. 121–135, 2003.
[3] Chunjiao Dong, Chunfu Shao, Stephen H Richards, and
Lee D Han, “Flow rate and time mean speed predictions
for the urban freeway network using state space models,”
Transportation Research Part C: Emerging Technologies,
vol. 43, pp. 20–32, 2014.
[4] Shiliang Sun, Changshui Zhang, and Guoqiang Yu, “A bayesian network approach to traffic flow forecasting,” Intelligent Transportation Systems, IEEE Transactions on,
vol. 7, no. 1, pp. 124–132, 2006.
[5] Dimitrios I Tselentis, Eleni I Vlahogianni, and Matthew G
Karlaftis, “Improving short-term traffic forecasts: to combine models or not to combine?,” Intelligent Transport
Systems, IET, vol. 9, no. 2, pp. 193–201, 2014.
[6] Loukas Dimitriou, Theodore Tsekeris, and Antony Stathopoulos, “Adaptive hybrid fuzzy rule-based system approach for modeling and predicting urban traffic flow,” Transportation Research Part C: Emerging Technologies, vol.
16, no. 5, pp. 554–573, 2008.
[7] Shiliang Sun and Changshui Zhang, “The selective random subspace predictor for traffic flow forecasting,” Intelligent Transportation Systems, IEEE Transactions on,
vol. 8, no. 2, pp. 367–373, 2007.
[8] Stephen Dunne and Bablu Ghosh, “Weather adaptive traffic prediction using neurowavelet models,” Intelligent
Transportation Systems, IEEE Transactions on, vol. 14,
no. 1, pp. 370–379, 2013.
[9] Eleni I Vlahogianni, Matthew G Karlaftis, and John C Golias, “Short-term traffic forecasting: Where we are and
where we?re going,” Transportation Research Part C:
Emerging Technologies, vol. 43, pp. 3–19, 2014.
[10] A Tascikaraoglu and M Uzunoglu, “A review of combined
approaches for prediction of short-term wind speed and
power,” Renewable and Sustainable Energy Reviews, vol.
34, pp. 243–254, 2014.
5. Teşekkür
[11] Borhan M. Sanandaji, Akin Tascikaraoglu, Kameshwar
Poolla, and Pravin Varaiya, “Low-dimensional models in
spatio-temporal wind speed forecasting,” in Proceedings
of the 2015 American Control Conference – ACC, 2015,
pp. 4485–4490.
Bu çalışma Türkiye Bilimsel ve Teknik Araştırmalar Kurumu
(TÜBİTAK) tarafından 115E984 no’lu "Kent içi yol ağlarının modellenmesi, kalibrasyonu ve yeni bir mikrosimülatörün
yaygın olarak kullanılan simülatörlerle karşılaştırılması" projesi
kapsamında desteklenmektedir.
[12] Akin Tascikaraoglu, Borhan M Sanandaji, Kameshwar
Poolla, and Pravin Varaiya, “Exploiting sparsity of interconnections in spatio-temporal wind speed forecasting
using wavelet transform,” Applied Energy, vol. 165, pp.
735–747, 2016.
6. Kaynakça
[13] “Freeway
Performance
Measurement
System
(PeMS),” http://pems.dot.ca.gov/?dnode=
Clearinghouse, Accessed: 2016-04-26.
[1] Fatma Yildiz Tascikaraoglu, Jennie Lioris, Ajith Muralidharan, Martin Gouy, and Pravin Varaiya, “Pointq model
of an arterial network: calibration and experiments,” arXiv preprint arXiv:1507.08082, 2015.
[2] Anthony Stathopoulos and Matthew G Karlaftis, “A multivariate state space approach for urban traffic flow mo-
| 3cs.SY
|
Polynomial braid combing
Juan González-Meneses and Marithania Silvero∗
arXiv:1712.01552v1 [math.GT] 5 Dec 2017
December, 2017
Abstract
Braid combing is a procedure defined by Emil Artin to solve the word problem in braid groups for
the first time. It is well-known to have exponential complexity. In this paper, we use the theory of
straight line programs to give a polynomial algorithm which performs braid combing. This procedure can
be applied to braids on surfaces, providing the first algorithm (to our knowledge) which solves the word
problem for braid groups on surfaces with boundary in polynomial time and space.
In the case of surfaces without boundary, braid combing needs to use a section from the fundamental
group of the surface to the braid group. Such a section was shown to exist by Gonçalves and Guaschi,
who also gave a geometric description. We propose an algebraically simpler section, which we describe
explicitly in terms of generators of the braid group, and we show why the above procedure to comb braids
in polynomial time does not work in this case.
1
Introduction
Braid groups can be seen as the fundamental group of the configuration space of n distinct points in a closed
disc D. If the points are unordered, the fundamental group is called the full braid group (or just the braid
group) with n strands, and denoted Bn . If the points are ordered, the obtained group is a finite index
subgroup of Bn , called the pure braid group with n strands, Pn .
If one replaces the closed disc D with any connected surface S, one obtains the full braid group Bn (S) and
the pure braid group Pn (S) with n strands on S.
Emil Artin [1] solved the word problem in braid groups (on the disc) for the first time. Actually, he solved
the word problem in Pn , and then used that Bn is a finite extension of Pn by the symmetric group Σn . The
way in which he solved the word problem in Pn is known as braid combing. Artin showed that Pn can be
seen as an iterated semi-direct product of free groups:
Pn = ((· · · ((F2 n F3 ) n F4 ) n · · · Fn−2 ) n Fn−1 .
The braid combing consists on computing the normal form of a pure braid with respect to the above semidirect decomposition. As the word problem in a free group of finite rank is well-known, this solves the word
problem in Pn .
But there is a big issue with braid combing: It is an exponential procedure. If we start with a word of length
m in the standard generators of Pn , the length of the combed braid may be exponential in m, when written
in terms of the generators of the free groups. An explicit example is given in Subsection 3.1. Artin was of
course aware of this; In the very last paragraph of [1], in which he talks about braid combing, he says:
∗ Both authors partially supported by the Spanish research projects MTM2013-44233-P, MTM2016-76453-C2-1-P and
FEDER.
1
“Although it has been proved that every braid can be deformed into a similar normal form the writer is
convinced that any attempt to carry this out on a living person would only lead to violent protests and
discrimination against mathematics. He would therefore discourage such an experiment”.
It this paper we use the theory of straight line programs (a compressed way to store a word as a set of
instructions to create it), to perform braid combing in polynomial time and space. This means that we give
an algorithm which, given a word w of length m in the standard generators of Pn , computes n − 1 compressed
words, each one representing a factor of the combed braid associated to w, in polynomial time and space
with respect to m.
Furthermore, given two pure braids, one can compare the compressed words associated to each of them in
polynomial time. Hence, this procedure gives a polynomial solution of the word problem in pure braid groups,
using braid combing.
This result, in the case of classical braids, does not improve the existing algorithms, as there are quadratic
solutions to the word problem in braid groups of the disc. But it happens that the above procedure is valid
not only for braids on the disc, but also for braids on any compact, connected surface S with boundary.
Hence, this provides the first polynomial algorithm to solve the word problem in Pn (S). The previously
known algorithms [20, 12], based on usual braid combing, are clearly exponential.
In the case of closed surfaces, the braid combing is quite different. There is not such a decomposition of
Pn (S) as a semi-direct product of free groups, and one needs to use instead a decomposition Pn (S) = π1 (S) n
Pn−1 (S\{p1 }), where p1 is a point in S. The existence of such a decomposition was shown by Gonçalves and
Guaschi [9], by giving a suitable section s : π1 (S) → Pn (S) of the natural projection π : Pn (S) → π1 (S) which
they explained geometrically, but not algebraically. In this paper we provide an explicit algebraic section,
for closed orientable surfaces of genus g > 0, which does not coincide with the one defined in [9] (although
we also give an explicit algebraic description of the section in [9]).
Finally, we explain why the procedure used to solve the word problem in surfaces with boundary does not
generalize to closed surfaces in the natural way.
The plan of the paper is the following. In Section 2 we introduce the basic notions of braids on surfaces. Then
in Section 3 we explain braid combing in the case of surfaces with boundary, and give an example showing
that combing is exponential. Straight line programs are treated in Section 4, and in Section 5 we introduce
the notion of compressed braid combing and give the polynomial algorithm to solve the word problem in braid
groups on surfaces with boundary. Section 6 deals with combing on a closed surface: We define the group
section from π1 (S) to Pn (S) when S is a closed surface, and we explain why the compressed braid combing
cannot be generalized to this case in a natural way.
Acnkowledgements: The first author thanks Saul Schleimer, for teaching him about straight line programs
at the Centre de Recerca Matemàtica (Barcelona) in 2012, and for useful conversations.
2
Braids on surfaces
Let S be a compact, connected surface of genus g and p boundary components, and let P = {p1 , . . . , pn } be
a set of n distinct points of S. A geometric braid on S based at P is an n-tuple β = (γ1 , . . . , γn ) of paths
γi : [0, 1] → S, such that
• γi (0) = pi , ∀i ∈ {1, . . . , n},
• γi (1) ∈ P, ∀i ∈ {1, . . . , n},
• {γ1 (t), . . . , γn (t)} are n distinct points of S ∀t ∈ [0, 1].
2
A braid on S based at P is a homotopy class of such geometric braids (notice that homotopies must fix the
endpoints). The usual product of paths endows the set of braids with a group structure, and the resulting
group (which is independent of the choice of P) is called the braid group with n strands on S, and denoted
Bn (S). The path starting at pi will be called the ith strand of the braid.
The above definition can also be explained by saying that the braid group Bn (S) is the fundamental group
of Mn (S)/Σn , where
Mn (S) = {(x1 , . . . , xn ) ∈ S n ; xi 6= xj ∀i 6= j}
is the configuration space of n distinct points in S, and Mn (S)/Σn is the quotient of Mn under the natural
action of the symmetric group Σn which permutes coordinates [4]. In other words, a braid can be seen as
a motion of n distinct points in S, whose initial configuration is P, they move along the surface without
colliding, and their final configuration is again P (though the particular position of each point in P may have
changed). We will sometimes use this dynamic interpretation of a braid throughout this paper.
There are well known presentations for braid groups of surfaces. In the particular cases of the sphere, the torus
and the projective plane, the classical references are [6, 3, 10]. For higher genus, one can find presentations
in [20, 12, 2, 11]. In all these presentations, some generators are related to the generators σ1 , . . . , σn−1 of
the classical braid group (and correspond to strand crossings), while some other generators are related to the
generators of the fundamental group of S (and correspond to motions of the distinguished points along the
surface).
A braid β is said to be pure if γi (1) = pi , for all i ∈ {1, . . . , n}, that is, if after the motion each distinguished
point goes back to its original position. Pure braids form a finite index subgroup of Bn (S), the pure braid group
with n strands on S, denoted Pn (S). Notice that Pn (S) is just the fundamental group of the configuration
space Mn (S). In particular, B1 (S) = P1 (S) = π1 (S).
The main object of study in this paper is braid combing, which is a procedure to produce a particular normal
form for pure braids. For this purpose, we need to make precise a particular presentation of Pn (S). From
now on we consider that S is a compact, connected orientable surface. We will see that the combing in the
non-orientable case is analogous, since one just need to consider the corresponding presentation of Pn (S)
appearing in [11, Theorem 3].
Theorem 2.1. [2] Let S be an orientable surface of genus g ≥ 0 with p > 0 boundary components. The
group Pn (S) admits the following presentation:
• Generators: {Ai,j | 2g + p ≤ j ≤ 2g + p + n − 1, 1 ≤ i < j}.
• Relations:
(PR1) A−1
i,j Ar,s Ai,j = Ar,s
if (i < j < r < s) or (r + 1 < i < j < s)
or (i = r + 1 < j < s where r ≥ 2g or r even).
−1
(PR2) Ai,j
Aj,s Ai,j = Ai,s Aj,s A−1
i,s
(PR3)
−1
Ai,j
Ai,s Ai,j
(PR4)
−1
Ai,j
Ar,s Ai,j
if i < j < s.
=
−1
Ai,s Aj,s Ai,s A−1
j,s Ai,s
=
−1
−1 −1
Ai,s Aj,s A−1
i,s Aj,s Ar,s Aj,s Ai,s Aj,s Ai,s
if i < j < s.
if (i + 1 < r < j < s)
or (i + 1 = r < j < s where r > 2g or r odd).
−1
−1
(ER1) Ar+1,j
Ar,s Ar+1,j = Ar,s Ar+1,s A−1
j,s Ar+1,s
(ER2)
−1
Ar−1,j
Ar,s Ar−1,j
=
if r odd, r < 2g and j < s.
−1 −1
Ar−1,s Aj,s A−1
r−1,s Ar,s Aj,s Ar−1,s Aj,s Ar−1,s
if r even, r ≤ 2g and j < s.
Remark 2.2. We have corrected some missprints of the presentation appearing in [2].
The generator Ai,j can be represented as a motion of a single point of P. Notice that 2g+p ≤ j ≤ 2g+p+n−1,
so we can write j = (2g + p − 1) + k for some k = 1, . . . , n. Then Ai,j represents a motion of the point
3
pk as shown in Figure 1. If 1 ≤ i ≤ 2g the motion of pk corresponds to one of the classical generators of
the fundamental group of a closed surface. If i = 2g + r with r = 1, . . . , p − 1, the point pk moves around
the rth boundary component (notice that there is no generator in which pk moves around the pth boundary
component). If i = (2g + p − 1) + t for some t = 1, . . . , k − 1, the point pk moves around the point pt , as in
the classical generators for the pure braid group of the disc [4].
Figure 1: A geometric representation of each of the four different families of motions of pk represented by generators
Ai,(2g+p−1)+k .
Remark 2.3. The presentation given in [2] is stated for g > 0, but it also holds when g = 0. For instance, if
g = 0 and p = 1, Pn (S) is the classical pure braid group Pn , the only relations that survive are (P R1)−(P R4),
and the presentation of Theorem 2.1 is precisely the presentation of Pn given in [4].
±1
Remark 2.4. By using the relations in Theorem 2.1 we can rewrite each conjugation A−1
i,j Ar,s Ai,j with
j < s as a word in generators whose second subindices equal s. Moreover, we can use these words to derive
−1
analogous relations allowing us to rewrite a word of the form Ai,j A±1
r,s Ai,j with j < s as a word in generators
whose second subindices equal s:
(PR10 ) Ai,j Ar,s A−1
i,j = Ar,s
if (i < j < r < s) or (r + 1 < i < j < s)
or (i = r + 1 < j < s where r ≥ 2g or r even).
if i < j < s.
−1 −1
(PR20 ) Ai,j Aj,s A−1
i,j = Aj,s Ai,s Aj,s Ai,s Aj,s
−1
(PR30 ) Ai,j Ai,s A−1
i,j = Aj,s Ai,s Aj,s
0
(PR4 )
Ai,j Ar,s A−1
i,j
=
if i < j < s.
−1
−1 −1
A−1
j,s Ai,s Aj,s Ai,s Ar,s Ai,s Aj,s Ai,s Aj,s
if (i + 1 < r < j < s)
or (i + 1 = r < j < s where r > 2g or r odd).
if r odd, r < 2g and j < s.
(ER10 ) Ar+1,j Ar,s A−1
r+1,j = Ar,s Aj,s
−1
−1
−1
(ER20 ) Ar−1,j Ar,s A−1
r−1,j = Aj,s Ar,s Ar−1,s Aj,s Ar−1,s Aj,s
if r even, r ≤ 2g and j < s.
These relations together with those appearing in Theorem 2.1 are used frequently throughtout this paper. We
use the expression (PR/ER)-relations to denote the set consisting of these 12 types of relations.
When S is a closed surface, one needs to add an extra relation in the presentation of Pn (S). We write
[a, b] = aba−1 b−1 .
Theorem 2.5. [2] Let S be an orientable closed surface of genus g ≥ 0. The group Pn (S) admits a presentation with generators
{Ai,j | 1 ≤ i ≤ 2g + n − 1, 2g + 1 ≤ j ≤ 2g + n, i < j}
and the same relations as those in Theorem 2.1 together with
2g+k−1
−1
(TR) [A−1
2g,2g+k , A2g−1,2g+k ] · · · [A2,2g+k , A1,2g+k ] =
Y
2g+n
Al,2g+k
l=2g+1
for k = 1, . . . , n.
4
Y
j=2g+k+1
A2g+k,j ,
Notice that this time the generator Ai,j corresponds to a motion of the point pk if j = 2g + k. The picture
corresponds exactly to the case in which p = 1 in Figure 1, provided one removes the boundary component
on the right hand side of the picture. Note that the (PR/ER)-relations also hold in this case.
As above, Theorem 2.5 was stated in [2] for genus g > 0, but the presentation also holds in the case g = 0,
as one gets the known presentation for the pure braid group of the sphere given in [8].
Remark 2.6. Choose some points Q = {q1 , . . . , qm } in a compact, connected surface S, and consider the
non-compact surface S\Q. Then the braid groups Pn (S\Q) and Bn (S\Q) are naturally isomorphic to Pn (S 0 )
and Bn (S 0 ), respectively, where S 0 is the surface obtained from S by removing a small open neighborhood
of Q, that is, by replacing each qi with a boundary component. In other words, removing a point from S is
equivalent to adding a boundary component, as far as braid groups on the surface are concerned.
3
Braid combing on a surface with boundary
From now on we assume that S is an orientable surface with p > 0 boundary components, unless otherwise
stated. The case when S is a non-orientable surface with boundary can be treated analogously, since one
just needs to consider the relations apperaring in the presentation of Pn (S) [11, Theorem 3], instead of the
(P R/ER)-relations. We will discuss the case when S is an orientable closed surface in Section 6.
Braid combing is a process by which a particular normal form of a pure braid is obtained. It was introduced
by Artin [1] for the pure braid groups of the disc, but it can be generalized to pure braid groups of other
surfaces, as we will explain in this section.
Recall that P = {p1 , . . . , pn } are n distinguished points in the surface S. For m = 1, . . . , n, we denote
Pm = {p1 , . . . , pm } and Qn−m = {pm+1 , . . . , pn }. It is well known (see for instance [13]) that the map from
Pn (S) to Pm (S) which ‘forgets’ the last n − m strands determines a short exact sequence
in,m
pn,m
1 → Pn−m (S\Pm ) −→ Pn (S) −→ Pm (S) → 1
(1)
where the base points of the three involved groups are, respectively, Qn−m , Pn and Pm .
In the case of a closed surface, the above sequence is also exact if we assume that m ≥ 3 if S is the sphere
S2 , and that m ≥ 2 if S is the projective plane RP 2 . The bad cases are those which involve finite groups,
as P1 (S2 ) = P2 (S2 ) = 1, P3 (S2 ) = Z/2Z [6], and also P1 (RP 2 ) = Z/2Z and P2 (RP 2 ) = Q8 , the quaternion
group of order 8 [13].
In the sequence (1), an element of Pn−m (S\Pm ) can be seen as a braid in Pn (S) in which the first m strands
are trivial or, in other words, in which the points of Pm do not move. This is equivalent to consider that S
has m extra punctures (or, removing a small neighborhood around each fixed puncture, that S has m extra
boundary components), and n − m points are moving.
To define braid combing, we single out a particular case of the exact sequence (1). When m = n − 1 ≥ 1 one
has
in,n−1
pn,n−1
1 → π1 (S \ Pn−1 ) −→ Pn (S) −→ Pn−1 (S) → 1,
(2)
where we applied that P1 (S \ Pn−1 ) = π1 (S \ Pn−1 ). We can easily describe the injection in,n−1 algebraically:
Let jn = (2g + p − 1) + n. Considering each element of Pn−1 as a boundary component and using the
generators of Theorem 2.1, we see that in,n−1 (Ai,jn ) = Ai,jn for all i = 1, . . . , jn − 1. The projection pn,n−1 is
also very easy to describe: For every i < j, the element pn,n−1 (Ai,j ) is either 1 (if j = jn ), or Ai,j (if j < jn ).
It is known that, since S has nontrivial boundary, the sequence (2) splits [9]. An explicit section s : Pn−1 (S) →
Pn (S) is given by s(Ai,j ) = Ai,j for all i, j.
Now notice that, as n − 1 ≥ 1, the fundamental group of S \ Pn−1 is a free group of rank 2g + p + n − 2 [15],
5
so we have an exact sequence:
in,n−1
pn,n−1
1 → F2g+p+n−2 −→ Pn (S) −→ Pn−1 (S) → 1.
(3)
Therefore, Pn (S) = Pn−1 (S) n F2g+p+n−2 , where Pn−1 (S) can be seen as the subgroup of Pn (S) generated
by {Ai,j }i<j<jn , and F2g+p+n−2 as the subgroup of Pn (S) generated by {Ai,jn }i<jn .
By induction on n, Pn (S) can be written as an iterated semi-direct product of free groups:
Pn (S) = ((· · · ((F2g+p−1 n F2g+p ) n F2g+p+1 ) n · · · F2g+p+n−3 ) n F2g+p+n−2 .
The process of computing the normal form of a braid (on a connected, compact surface with boundary)
with respect to this iterated semi-direct product, is known as braid combing.
Definition 3.1. Let S be an orientable surface of genus g with p > 0 boundary components. The combed
normal form of a braid α ∈ Pn (S) is a decomposition
α = α1 α2 · · · αn
where, for k = 1, . . . , n, αk belongs to the subgroup generated by {Ai,jk }i<jk , with jk = (2g + p − 1) + k.
By the uniqueness of normal forms with respect to semi-direct products, the combed normal form of a braid
is unique. Also, since each of the subgroups described in Definition 3.1 is a free group on the given generators,
one can choose a unique reduced word wi to represent each αi , and this gives a unique word representing α.
We now describe explicitly how to compute the combed normal form of a pure braid α represented by a
given word w in the generators of Pn (S). For that purpose, we just need to find a way to move the letters
A±1
i,j with smaller second index to the left of those with bigger second index. This can be achieved by using
±1
the (PR/ER)-relations to replace two consecutive letters AB where A = A±1
r,s , B = Ai,j and j < s, with
the word BW , where W is a word formed by letters whose second subindex equals s. Iteratively applying
these substitutions (together with free reduction), the word w can be transformed into a reduced word of
the form w1 w2 · · · wn , which also represents α, such that the second subindex of every letter of wk is jk , for
k = 1, . . . , n. The decomposition α = α1 · · · αn , where αk is the braid represented by wk for k = 1, . . . , n, is
thus the combed normal form of α.
We will see in Subsection 3.1 that the length of the word w1 · · · wn is possibly exponential with respect to the
length of w. But we will show that it is possible to represent w1 · · · wn in a compressed way. The key point
consists of moving the letters as explained in the above paragraph but, instead of applying the conjugations
described in (PR/ER)-relations, we will keep track of those conjugations without applying them. We will
now show how to do this. To avoid cumbersome notation we use the expression uv to denote the word v −1 uv,
with u and v two given words.
Suppose that w = u1 u2 · · · um , where ut = A±1
it ,st for t = 1, . . . , m. Thus w is a sequence of m letters. For
every t, let vt be the subsequence of w formed by those letters ur such that r > t and sr < st . In other
words, vt is formed by the letters of w which come after ut and have smaller second index. Notice that vt is
formed precisely by the letters of w that must be swapped with ut when combing the braid.
By construction, we have the following:
Lemma 3.2. Let w = u1 u2 · · · um be as above, representing a braid α. Given k ∈ {1, . . . , n} let ui1 ui2 · · · uit
be the subsequence of w formed by those letters whose second subindex is jk . Define:
v
v
v
wk = ui1i1 ui2i2 · · · uitit .
Then wk represents the same braid as wk , so w = w1 · · · wn represents the combed normal form of α.
Example 3.3. If w = A1,2 A1,4 A1,2 A−1
2,3 A2,4 A1,3 A1,2 ∈ P4 , we have
A1,2
A
A1,2 A−1
2,3 A1,3 A1,2
w = A1,2 A1,2 A1,2 A−1
A 1,2 A
|
{z
} | 2,3 {z 1,3 } | 1,4
w2
w3
6
{z
w4
A
A2,41,3
A1,2
}
.
Note that in P4 we have g = 0 and p = 1, so jk = k. There are no letters of the form Ai,1 , that is why w1 is
the trivial word.
The word w = w1 · · · wn is not too long with respect to w: its length is at most m2 . Indeed, the worst case
occurs if we need to swap every pair of letters of w, so each vi has length m − i, and thus the length of w is
= m2 .
m + 2 m(m−1)
2
Let us see how we can obtain enough information to describe each wk , by a procedure which is linear in m.
Lemma 3.4. Given w = u1 · · · um as above, and given k ∈ {1, . . . , n}, one can compute in time O(m) a
subsequence v of w and a list of pairs of integers (c1 , d1 ), . . . , (ct , dt ), where 1 ≤ |cr | ≤ m and 0 ≤ dr ≤ m
for each r = 1, . . . , t (t < m), which encodes all the information to describe wk .
Proof. Going through the word w = u1 · · · um once, we can determine the position i1 of the first letter
v
v
v
with second index jk , and the word v = vi1 . We know that wk will have the form ui1i1 ui2i2 · · · uitit and, by
construction, that each vir is a suffix of vi1 . Hence, in order to determine vir we just need to provide the
word vi1 and the length dr of vir . Also, in order to determine the letter uir we just need to provide its first
index and the sign of its exponent (±1). Hence we just need to provide an integer cr such that |cr | is the
first index of uir , and whose sign is equal to the exponent of uir .
To obtain these integers, we go through w again, starting at the position i1 and setting d equal to the length
of vi1 . Every time we read a letter u whose second index is smaller than jk , we decrease d by one. Every
time we read a letter u whose second index is jk , we store the pair (c, d) where |c| is the first index of u and
the sign of c is given by the exponent of u (positive if u = A|c|,jk , negative if u = A−1
|c|,jk ), and d is the above
number. Notice that the first step stores (c1 , d1 ) where |c1 | = i1 and d1 is the length of vi1 .
At the end of the whole procedure, we have gone through w twice (so the complexity is O(m)), and we have
obtained a word vi1 and a list of pairs of integers (c1 , d1 ), . . . , (ct , dt ) which encode wk as desired.
Example 3.5. Let w be the word with 7 letters appearing in Example 3.3. To describe w3 we store the
word A1,2 and the pairs (−2, 1), (1, 1). To describe w4 we store the word A1,2 A−1
2,3 A1,3 A1,2 and the pairs
(1, 4), (2, 2).
We have then obtained, for every k = 1, . . . , n, a short word wk representing the factor αk of the braid
combing. But a word representing αk will be useful only if we represent it as a reduced word wk in the
generators {Ai,jk }i<jk of the corresponding free group. We can obtain wk from wk by iteratively applying
the (PR/ER)-relations: Given a word uvi i (suppose that the second subindex of ui is jk ), we can write
vi = a1 · · · at ; then we conjugate ui by a1 using the (PR/ER)-relations, obtaining an element which can be
written as a word whose letters have second index equal jk . Next we conjugate this word by a2 , using the
relations again (taking into account that conjugating a product by a2 is the same as conjugating each factor
by a2 ). And so on. At the end, uvi i is transformed into a word in which the second index of every letter is
jk . After repeating this process with every word in wk and applying free reduction, one obtains wk .
This last step of braid combing is the exponential one! It can produce a word wk which is exponentially long
with respect to m (we will show such an example in Lemma 3.6). For this reason, in Section 5 we introduce
a method based on straight line programs to avoid this last step of braid combing, and to solve the word
problem for braids on surfaces with boundary in polynomial time.
3.1
Braid combing is exponential
In this section we provide an example to show that braid combing is, in general, an exponential procedure.
That is, we present a family of pure braids βm , m ≥ 1, where each βm can be expressed as a word whose
length is linear in m, but whose combed expression is a word of exponential length with respect to m. This
7
implies that there is no hope to produce an efficient algorithm to comb braids, if one wishes to express combed
braids as words in usual braid generators.
The example will be given in P4 , the pure braid group with 4 strands on the disc D2 , with generators
−1
−1
· · · σj−1
, for 1 ≤ i < j ≤ 4. As P4 embeds in Pn (S) for n ≥ 5 (where S is an
Ai,j = σj−1 σj−2 · · · σi+1 σi2 σi+1
orientable surface, or n ≥ 4 if S has genus g > 0) by an embedding which sends generators to generators (of
the presentations defined in Section 2), it follows that, in general, braid combing is exponential in surface
braid groups.
Given m ≥ 1, define:
βm = A−1
1,2 A2,3
−m
A3,4 A−1
1,2 A2,3
m
∈ P4 .
The length of the given word representing βm is 4m + 1, so it is linear in m. We can then show the following:
Lemma 3.6. With the above notation, the combed normal form of βm is a reduced word in A1,4 , A2,4 , A3,4
and their inverses, whose length is exponential in m.
Proof. We consider the inner automorphism φ of P4 which consists of conjugating by A−1
1,2 A2,3 . Then βm =
φm (A3,4 ). Using the short exact sequences explained in the previous section, we see that φ induces an
automorphism on the free group generated by A1,4 , A2,4 , A3,4 , so βm can be written as a word on these
letters and their inverses.
The fact that βm is exponentially long when written as a reduced word in A1,4 , A2,4 , A3,4 and their inverses,
can be deduced from the fact that A−1
1,2 A2,3 is a pseudo-Anosov braid in P3 . However, since we want to make
this example as explicit as possible, we will provide an algebraic proof.
To simplify the notation, let us denote ai = Ai,4 for i = 1, 2, 3. From the presentation of P4 , one can check
the following:
−1 −1
−1 −1
φ(a1 ) = a2 a3 a−1
2 a3 a2 a1 a2 a3 a2 a3 a2
−1 −1 −1 −1
−1
−1 −1
φ(a2 ) = a2 a3 a2 a3 a2 a1 a2 a3 a2 a−1
3 a2 a1 a2 a3 a2 a3 a2
−1
φ(a3 ) = a2 a3 a2
Let us consider the following three words: x = a1 , y = a2 a3 a−1
2 and w = a1 a2 a3 . Then we can write:
φ(a1 ) = yw−1 xwy −1
φ(a2 ) = yw−1 x−1 wy −1 wy −1
φ(a3 ) = y
Suppose that we can decompose a braid in the following way:
α = z1 we1 z2 we2 · · · zt−1 wet−1 zt ,
where zi ∈ {x, y, x−1 , y −1 } and ei ∈ {1, −1} for i = 1, . . . , t − 1. This expression of α is not necessarily
reduced when written in terms of {a±1
i }, but the cancellations are only restricted to subwords of the form
x−1 w or w−1 x, and in these cases only two letters cancel, and no further cancellation is produced. Therefore,
the length of the reduced word α in {a±1
i } associated to α, is at least 2t − 1. If we moreover assume that
z1 = y, then the length of α is greater than 2t. In this case, we say that α admits a xyz-decomposition of
length t.
We point out that φ(w) = w. This follows from the fact that w = a1 a2 a3 = σ3 σ2 σ1 σ1 σ2 σ3 commutes with
A1,2 = σ12 and A2,3 = σ22 .
Finally, we conclude the proof by showing that βm admits a xyz-decomposition of length t ≥ 3m−1 .
We proceed by induction on m. First, we check that β1 = φ(a3 ) = y, so the claim holds in the case m = 1.
Now suppose the claim is true for some m ≥ 1, so βm = z1 we1 z2 we2 · · · zt−1 wet−1 zt for some t ≥ 3m−1 . Then
we have
βm+1 = φ(βm ) = φ(z1 )we1 φ(z2 )we2 · · · φ(zt−1 )wet−1 φ(zt ).
8
Finally, since
φ(x) = yw−1 xwy −1 ,
and φ(y) = yw−1 x−1 wy −1 wyw−1 yw−1 xwy −1 ,
it follows that βm+1 admits a xyz-decomposition, whose length is at least 3t, thus greater than 3m .
4
Straight line programs
We have seen that combing a braid α consists of decomposing it as a product α1 α2 · · · αn in a suitable way.
We have also seen that, in general, if α is given as a word of length m, the length of a factor αi may be
exponential in m. So, how could we make this procedure to have polynomial complexity? The answer is:
We will not describe each αi as a word in the generators of the corresponding free group. Instead, we will
describe each αi as a compressed word (also called straight line program).
The concept of straight line program is well known in Complexity Theory [5], and has been used in Combinatorial Group Theory to reduce the complexity of decision problems [19]. In this section we review the
main aspects related to straight line programs, which we will call compressed words, following [19].
Roughly speaking, a compressed word A consists on two disjoint finite sets of symbols F and A (the latter
is ordered), called terminal and non-terminal alphabets respectively, together with a set of production rules
indicating how to rewrite each non-terminal character in A as a word in F and smaller characters of A. In
this way, the biggest non-terminal character can be rewritten, using the production rules, as a word in F
that we denote ev(A) (the evaluation of A, or the decompressed word). So A can be seen as a small set of
instructions on how to produce a long word ev(A) in F.
Here is the rigorous definition:
Definition 4.1. A compressed word (or straight line program) A consists on a finite alphabet F of terminal
characters together with an ordered finite set of (non-terminal) symbols A = {A1 , . . . , An }, and a set of
production rules P = {Ai → Wi }1≤i≤n allowing to replace each non-terminal Ai ∈ A with its production: a
(possibly empty) word Wi ∈ (F ∪ A)∗ , where every non-terminal Aj appearing in Wi has index j < i. The
greatest non-terminal character in A, An ∈ A, is called the root.
The evaluation of the compressed word A, ev(A), is the (decompressed) word in F obtained by replacing
successively every non-terminal symbol with the right-hand side of its production rule, starting from the root.
P
We define the size of a compressed word A as |A| = i |Wi |. We can assume that |A| ≥ #(F) (if a terminal
character appears in no production rule, we can remove it from F as it will not appear in ev(A)). We can
also assume that |A| ≥ #(A) (if a production rule transforms a non-terminal symbol into the empty word,
we can remove every appearance of the non-terminal symbol from the whole compressed word). Therefore,
the space needed to store A is at most 2|A|, and this is the reason why |A| is called “the size of A”.
Example 4.2. Given n ≥ 1, consider the following compressed word:
* F = {a}
+
A = {A1 , . . . , An }
An :=
.
P = {Ai → Ai−1 Ai−2 }ni=3 ∪ {A2 → a} ∪ {A1 → b}
The sequence of compressed words {An }n≥1 encodes the sequence of Fibonacci words. The first seven decompressed words ev(A1 ), . . . , ev(A7 ) are, respectively:
b,
a,
ab,
aba,
abaab,
abaababa,
abaababaabaab.
We see that every word is the concatenation of the two previous ones. Notice that |An | = 2n − 2, while
|ev(An )| = Fn , the n-th Fibonacci number. So, in this example, the sizes of compressed words grow linearly
in n, while the sizes of decompressed words grow exponentially.
9
Two compressed words A and B are said to be equivalent if ev(A) = ev(B).
A crucial property is that a pair of compressed words can be compared without being decompressed:
Theorem 4.3. [18] Given two compressed words A and B, there is a polynomial time algorithm in |A| and
|B| which decides whether or not ev(A) = ev(B).
Now recall that we want to use compressed words to describe the factors α1 · · · αn of a combed braid. The
leftmost factor α1 does not need to be compressed, as it can be expressed as a subsequence of the original
word. Each of the other factors, α2 , . . . , αn , belongs to a free group. So we would like to compress words
representing elements of a free group.
In order for this procedure to be useful (for instance, to solve the word problem in braid groups), we need
to be able to compare two compressed combed braids. This means that we need to determine whether two
compressed words (A and B) evaluate to words (ev(A) and ev(B)) which represent the same element in a free
group. That is, we want to compare the uncompressed words not just as words, but as elements in a free
group.
Fortunately, this problem has already been satisfactory solved:
Theorem 4.4. [16] Given a compressed word A, there exists a polynomial time algorithm producing a compressed word Ared such that ev(Ared ) is the free reduction of ev(A).
As a consequence, if we have two compressed words A and B, we can compute Ared and Bred in polynomial time, and then we can check in polynomial time whether ev(Ared ) and ev(Bred ) are the same word
(Theorem 4.3). Hence, the compressed word problem in a free group is solvable in polynomial time.
We will therefore be able to use compressed words to perform braid combing, and to compare combed braids,
in polynomial time. We now proceed to describe how to compress the words appearing in the process of
braid combing.
5
Compressed braid combing
Throughout this section, S will be a compact, connected orientable surface with p > 0 boundary components.
The arguments in this section can also be applied in a straightforward way if S is non-orientable, just by
using the appropriate presentations.
Recall that the combing algorithm for a braid α ∈ Pn (S) starts with a word w = u1 · · · um representing α,
where uk = A±1
i,j for k = 1, . . . , m, and produces n words, w1 , . . . , wn , representing the factors α1 , . . . , αn of
the combed normal form of α. Each wk is a reduced word formed by letters whose second subindex is jk .
Moreover, we gave a procedure to compute short words w1 , . . . , wn representing the factors α1 , . . . , αn of the
combed normal form of α in polynomial time, only that the second index of a letter in wk is not necessarily
jk .
In some sense wk is a compressed expression of wk . Notice that the word w1 is actually equal to w1 , and its
length is at most m, so we do not need to use compression for this first factor. Let us study the other cases.
In this section we will see how, starting from wk , one can define a compressed word Ak (a straight line
program) such that wk is the free reduction of ev(Ak ). Moreover, the size of Ak will be of order O((g+p+n)m).
This will allow us to determine and compare the factors of the combed normal forms of braids in Pn (S),
without needing to evaluate them, so it solves the word problem in Pn (S) in polynomial time.
v
v
Let k ∈ {2, . . . , n}, and recall that wk = ui1i1 · · · uitit , where each uir belongs to {A±1
i,jk }i<jk , and each vir is
±1
a word formed by letters from {Ai,js }i<js <jk . Recall also that, by construction, each vir is a suffix of vi1 .
Also, t ≤ m and the length of vi1 is smaller than m (where m is the length of w).
10
By Lemma 3.4, we can obtain in time O(m) a subsequence vi1 of w and a sequence of pairs of integers
(c1 , d1 ), . . . , (ct , dt ) which encode wk .
We want to define a compressed word Ak representing αk , so the terminal alphabet will consist of the
generators of Pn (S) corresponding to the motion of the point pk , that is, {A±1
i,jk }i<jk . To be consistent with
the forthcoming notation, we will denote Xi,0 = Ai,jk and X−i,0 = A−1
,
for
i
= 1, . . . , jk −1. So the terminal
i,jk
alphabet of Ak becomes Fk = {Xi,0 }0<|i|<jk .
On the other hand, the non-terminal symbols of Ak will consist of a single symbol Xk (the root), plus a
symbol corresponding to uv for each u ∈ Fk and each nontrivial suffix v of vi1 . We know that the word uv
can be determined by a pair of integers (c, d), where u = Xc,0 and v is the suffix of vi1 of length d. Hence,
the set of nonterminal symbols is
Ak = {Xc,d } 0<|c|<jk ∪ {Xk }.
0<d≤|vi |
1
We order the elements of Ak (distinct form Xk ) according to their second subindex and, in case of equality,
according to their first subindex. Xk is the biggest element, being the root.
The first production rule of Ak will be
Xk → Xc1 ,d1 · · · Xct ,dt .
Now let Xc,d be a non-terminal symbol corresponding to a word uv , let a1 be the first letter of v, and write
0
v = a1 v 0 , so uv = (ua1 )v . We know, from the relations in Theorem 2.1 and Remark 2.4, that the braid
a1
represented by u can be written as b1 · · · bs , where each bi ∈ Fk and s ≤ 9. Therefore, the braid represented
0
0
by uv can be written as bv1 · · · bvs , which we can encode as Xe1 ,d−1 · · · Xes ,d−1 for some integers e1 , . . . , es .
We thus add the following production rule to Ak :
Xc,d → Xe1 ,d−1 · · · Xes ,d−1 .
Adding these production rules for all non-terminal symbol Xc,d (d > 0) determines the compressed word Ak .
Proposition 5.1. The size of Ak is smaller than 19(2g + p + n)m, and the evaluation ev(Ak ) represents αk .
Proof. The first production rule of Ak has length t ≤ m. The length of the other production rules is at
most 9, and there are as many as elements in Ak \{Xk }, that is, 2(jk − 1)|vi−1 | = 2(2g + p + k − 2)|vi−1 | <
2(2g + p + n)(m − 1). Hence |Ak | ≤ m + 18(2g + p + n)(m − 1) < 19(2g + p + n)m.
By induction on d (starting with d = 0), we see that each Xc,d evaluates to a word which represents the same
braid as uv (where uv is the word corresponding to the symbol Xc,d ). Hence, the evaluation of Xk is a word
v
v
in Fk which represents the same braid as ui1i1 · · · uitit , that is, αk .
Theorem 5.2. Let S be a connected surface of genus g ≥ 0 with p > 0 boundary components. Let α ∈ Pn (S)
be a braid given as a word w of length m in the generators {A±1
i,j }i<j . Then, for every k = 1, . . . , n, there is an
algorithm of complexity O((g +p+n)m) which produces a compressed word Ak of size at most 19(2g +p+n)m
and terminal characters {A±1
i,jk }i<jk , whose evaluation represents αk , the k-th factor of the combed normal
form of α.
Proof. This results follows from Lemma 3.4 and Proposition 5.1, taking into account that computing the production rule corresponding to each uv just requires transcribing the relation in Theorem 2.1 and Remark 2.4
corresponding to u and to the first letter of v, so the complexity of the whole algorithm is proportional to
the size of Ak .
Example 5.3. Consider the word w = A1,4 A1,3 A−1
2,4 A1,2 ∈ P4 . If we set k = 4, we have i1 = 1 and
A
vi1 = A1,3 A1,2 which has length 2. The word w4 , which is equal to A141,3
X1,2 X−2,1 . The production rules for A4 are:
X4 → X1,2 X−2,1
11
A1,2
A1,2
(A−1
, can be codified as
2,4 )
X1,2 → X1,1 X3,1 X1,1 X−3,1 X−1,1
X−2,1 → X1,0 X−2,0 X−1,0
X1,1 → X1,0 X2,0 X1,0 X−2,0 X−1,0
X3,1 → X3,0
X−1,1 → X1,0 X2,0 X−1,0 X−2,0 X−1,0
X−3,1 → X−3,0
The word ev(A4 ) is
X1,0 X2,0 X1,0 X−2,0 X−1,0 X3,0 X1,0 X2,0 X1,0 X−2,0 X−1,0 X−3,0 X1,0 X2,0 X−1,0 X−2,0 X−1,0 ,
corresponding to
−1
−1 −1 −1
−1 −1 −1
A1,4 A2,4 A1,4 A−1
2,4 A1,4 A3,4 A1,4 A2,4 A1,4 A2,4 A1,4 A3,4 A1,4 A2,4 A1,4 A2,4 A1,4 .
Recall that the compressed braid combing explained throughout this section can be applied to the case when
S is a non-orientable surface with p > 0 boundary components (one just needs to modify the production
rules so they encode the appropiate relations given in [11, Theorem 3] instead of the (P R/ER)-relations).
Since the length of each of those relations is at most 9, the bounds given in Proposition 5.1 and Theorem 5.2
also hold for the non-orientable case.
Corollary 5.4 (Word problem). Let S be a compact, connected surface of genus g with p > 0 boundary components. There exists an algorithm which, given two pure braids α1 , α2 ∈ Pn (S) represented by words w1 , w2
of respective lengths m1 , m2 in the generators of Theorem 2.1, determines whether α1 = α2 in polynomial
time and space with respect to g, p, m1 and m2 .
Proof. The algorithm just applies compressed braid combing to w1 and w2 , then transforms each compressed
factor into a reduced compressed factor (Theorem 4.4), and then compares the reduced compressed factors to
see whether they evaluate to the same reduced word (Theorem 4.3). All these steps can be done in polynomial
time, as we have explained.
6
Braid combing on a closed surface
Throughout this section, assume that S is a closed surface (compact, connected without boundary). In this
case we cannot apply braid combing as in the previous section, since the exact sequence (1) splits if and only
if either S is the sphere, the torus or the Klein bottle, or if S is the projective plane with n = 3 and m = 2,
or if S is a surface of bigger genus and m = 1 [13]. Moreover, π1 (S) is no longer a free group, if S is not the
sphere.
One can nevertheless decompose Pn (S) by using another instance of Sequence (1), the one in which m = 1:
in,1
pn,1
1 → Pn−1 (S\{p1 }) −→ Pn (S) −→ π1 (S) → 1.
(4)
This sequence is exact if S is not the sphere or the projective plane, so we will exclude those two cases. In
all other cases, the sequence splits [9, 13]. Since we are assuming that S has no boundary, we will use the
generators of Pn (S) defined in Theorem 2.5.
One can then decompose Pn (S) = π1 (S) n Pn−1 (S \ {p1 }). Then S \ {p1 } can be treated as a surface with
boundary, so Pn−1 (S \ {p1 }) can be decomposed as a semi-direct product of free groups. In the case in which
S is orientable with genus g > 0 we obtain a decomposition:
Pn (S) = π1 (S) n ((· · · ((F2g n F2g+1 ) n F2g+2 ) n · · · F2g+n−3 ) n F2g+n−2 ).
12
Combing a braid in Pn (S), when S is a closed orientable surface distinct from S2 , means to find its normal
form with respect to this group decomposition.
It is clear that, in order give an algorithm for braid combing, one needs to describe an explicit group section
for the projection pn,1 . Notice that the generators of π1 (S), say a1 , . . . , a2g , are naturally associated to
A1,j1 , A2,j1 . . . , A2g,j1 , where j1 = 2g + 1. But (contrary to the case with boundary) the map which sends ai
to Ai,j1 is not a group homomorphism: a section for pn,1 must be defined otherwise.
In [9], such a section is defined topologically, by using a retraction of the surface S, and allowing the distinguished points p2 , . . . , pn , to move along the retraction as the point p1 performs the movement corresponding
to some ai . In [9], this map from π1 (S) to Pn (S) is only described algebraically in the case of an orientable
surface of genus 2.
We will now define a different group section, simpler than the one defined in [9] (although related to it),
which will be explicitly given in terms of the generators of Pn (S).
The generators of Pn (S) are described in Figure 2. These are analogous to the generators described in
Figure 1, the only differences being that S has no boundary (hence the second indices of the generators are
shifted), and that we have placed the base points p1 , . . . , pn in a different place to simplify the forthcoming
figures.
Figure 2: A geometric representation of the generators of Pn (S) when S is closed.
Let us define, for k = 1, . . . , n, the braid Bk = A2g,jk Aj1 ,jk Aj2 ,jk · · · Ajk−1 ,jk , where jt = 2g+t for t = 1, . . . , n.
See Figure 3 for a picture of Bk as a product of generators (a), and also in a simpler geometric way (b).
Figure 3: The braid Bk = A2g,jk Aj1 ,jk Aj2 ,jk · · · Ajk−1 ,jk .
Using the geometric representation of the braid Bk given in Figure 3, it is clear that Bi and Bj commute, for
every 1 ≤ i, j ≤ n. We will be particularly interested in the braid B1 B2 · · · Bn which can be seen in Figure 4.
In Figure 5 we can see the braid (B1 B2 · · · Bn )−1 A2g,j1 (B1 B2 · · · Bn )A−1
2g,j1 in picture (a), which is smoothly
transformed into the braid in picture (d). It is a classical exercise to see how to express the path in picture
(d) as a product of generators of the fundamental group of S. In our case, this allows to express that braid
as a product of the generators of Pn (S), as follows:
−1
−1
(B1 B2 · · · Bn )−1 A2g,j1 (B1 B2 · · · Bn )A−1
2g,j1 = [A1,j1 , A2,j1 ] · · · [A2g−3,j1 , A2g−2 ].
It follows that, in Pn (S), one has
−1
[(B1 B2 · · · Bn )−1 , A2g,j1 ][A−1
2g−2 , A2g−3,j1 ] · · · [A2,j1 , A1,j1 ] = 1.
13
Figure 4: The braid B1 B2 · · · Bn .
Figure 5: The braid (B1 B2 · · · Bn )−1 A2g,j1 (B1 B2 · · · Bn )A−1
2g,j1 .
Now recall that the fundamental group of S has the following presentation:
−1
−1
π1 (S) = a1 , . . . , a2g ; [a−1
2g , a2g−1 ][a2g−1 , a2g−2 ] · · · [a2 , a1 ] = 1 ,
where ai = pn,1 (Ai,j1 ) for i = 1, . . . , 2g. If we notice that pn,1 (B1 B2 · · · Bn ) = a2g , the following result is
immediately obtained:
Theorem 6.1. Let S be an orientable closed surface of genus g ≥ 1. The map s : π1 (S) → Pn (S) which
sends ai to Ai,j1 for i = 1, . . . , 2g − 1, and a2g to B1 · · · Bn , is a group section of the projection pn,1 of the
short exact sequence (4).
Remark 6.2. To our knowledge, the above result gives the first known explicit algebraic section of pn,1 , when
g > 2. Moreover, we can also give an explicit algebraic definition of the group section described geometrically
in [9]: It is the map ρ : π1 (S) → Pn (S) such that ρ(ai ) = Ai,j1 when i is odd, and ρ(ai ) = Ai,j1 B2 · · · Bn
when n is even. The proof that ρ is a group section is similar to the one we did for s. We used s instead of
ρ as it is an algebraically simpler section.
Now we can comb a braid in a closed orientable surface S with genus g > 0, using the above section (which
allows to compute the normal form with respect to the decomposition Pn (S) = π1 (S) n Pn−1 (S\{p1 })), and
then applying the combing procedure of Section 3 to the second factor, obtaining the normal form with
respect to the decomposition:
Pn (S) = π1 (S) n ((· · · ((F2g n F2g+1 ) n F2g+2 ) n · · · F2g+n−3 ) n F2g+n−2 ).
There is one detail to be taken into account. When we decompose Pn (S) = π1 (S) n Pn−1 (S\{p1 }), the group
Pn−1 (S\{p1 }) is considered as a subgroup of Pn (S) (formed by the braids in which the first strand is trivial).
14
In other words, the generators of this group are the braids of the form Ai,jk , where i ≤ jk = 2g + k and
k > 1. But then we consider the group Pn−1 (S\{p1 }) as a braid group of a surface with boundary, S 0 , which
is obtained by removing a small neighborhood of p1 .
The generators of Pn−1 (S 0 ) are shown in Figure 1, where the only boundary component is placed on the
right hand side of the picture. We can express any word in the generators of Pn−1 (S\{p1 }) as a word in the
generators of Pn−1 (S 0 ) thanks to the isomorphism f : Pn−1 (S\{p1 }) → Pn−1 (S 0 ) defined as follows:
Ai,jk−1
h
i h
i
−1
f (Ai,jk ) =
A−1
,
A
·
·
·
A
,
A
2g−1,jk−1
1,jk−1
2g,jk−1
2,jk−1
Ai−1,jk−1
if i ≤ 2g,
k−2
Y
t=1
Ajt ,jk−1
n−1
Y
!−1
Ajk−1 ,jt
if i = 2g + 1,
t=k
if 2g + 1 < i < jk .
These formulae are obtained by interpreting the generators of Pn−1 (S\{p1 }) as points moving in the surface
S 0 , in which the point p1 has been transformed into a boundary component (and moved to the right hand
side, like in Figure 1). All interpretations are straightforward, except the generator A2g+1,jk = Aj1 ,jk . In
Pn−1 (S\{p1 }), this generator corresponds to a movement of the puncture pk around the puncture p1 . In
Pn−1 (S 0 ), however, there is no generator corresponding to a puncture moving around the last (and only)
boundary of S 0 . We must then apply the relation (TR) of Theorem 2.5, to express Aj1 ,jk as a product of
other generators, which are then mapped to Pn−1 (S 0 ) as expressed in the above equation. Once we have
applied the map f , we can comb the resulting braid in Pn−1 (S 0 ) as it was explained in Section 3.
Now we will explain why we cannot apply the techniques in Section 5 to comb a braid in a closed surface.
The idea of combing, as it was done in Section 3, is to move to the left the generators with smaller second
index, which act by conjugation on the generators with bigger second index. If the surface S has boundary,
the action of a generator A±1
i,j on a generator Ar,s , produces a word in which all letters have second index
s. Hence, the k-th final factor of the combed braid only depends on the letters of the original word having
second index jk , and on the letters of bigger second subindex which act on them. This is why we can easily
determine the compressed word associated to the k-th factor.
If the surface S is closed, however, the action of a generator A±1
i,j on a generator Ar,s , does not necessarily
produce a word in which all letters have second index s, due to the necessity of applying the map f . As an
example, consider the relation (ER1) with j = j1 :
−1
−1
A−1
r+1,j1 Ar,s Ar+1,j1 = Ar,s Ar+1,s Aj1 ,s Ar+1,s .
All letters in the resulting word seem to have the same second subindex, but when we apply the map f to
see the braid in Pn−1 (S 0 ), the letter A−1
j1 ,s must be replaced by a word whose letters have second subindex
going from s − 1 to n − 1. This fact does not permit to obtain the factors of a combed braid as compressed
words, as was done in Section 3. A different approach should therefore be used, in order to find a polynomial
solution to the word problem of braid groups on closed surfaces.
Nevertheless, the algebraic description of the section s in Theorem 6.1 allows to perform (non-compressed)
braid combing in the classical way, as explained in this section. This was not possible before, due to the
lack of an algebraically explicit section. This procedure of combing a braid in a closed surface is, however,
exponential.
References
[1] E. Artin. Theory of braids. Ann. of Math. 48 (1947), no. 2, 101-126.
[2] P. Bellingeri. On presentations of surface braid groups. J. Algebra 274 (2004), no. 2, 543-563.
15
[3] J. S. Birman. On braid groups. Comm. on Pure and applied Math. 22 (1969) 41-72.
[4] J. S. Birman. Braids, links, and mapping class groups. Annals of Mathematics Studies, No. 82. Princeton
University Press, 1974.
[5] P. Bürgisser, M. Clausen, M.A. Shokrollahi. Algebraic Complexity Theory, Grundlehren der mathematischen Wissenschaften, vol. 315, Springer-Verlag (1997).
[6] E. Fadell and J. Van Buskirk. The braid groups of E 2 and S 2 ’. Duke Math. J. 29 (1962) 243-258.
[7] L. Gasieniec, M. Karpinski, W. Plandowski, and W. Rytter. Efficient algorithms for Lempel-Ziv encoding.
Proc. 4th Scandinavian Workshop on Algorithm Theory (SWAT 1794), Lecture Notes in Comput. Sci.
1097, Springer-Verlag, Berlin 1996, 392-403.
[8] R. Gillette and J. Van Buskirk. The word problem and consequences for the braid groups and mapping
class groups of the 2-sphere. Trans. Amer. Math. Soc. 131 (1968), 277-296.
[9] D. L. Gonçalves and J. Guaschi, On the structure of surface pure braid groups. J. Pure and Appl. Alg.
186 (2004), 187-218.
[10] D. L. Gonçalves and J. Guaschi, The braid groups of the projective plane and the Fadell-Neuwirth short
exact sequence. Geometriae Dedicata 130 (2007) 93-107.
[11] D. L. Gonçalves and J. Guaschi, Braid groups of non-orientable surfaces and the Fadell-Neuwirth short
exact sequence. J. Pure and Appl. Alg. 214 (2010), 667-677.
[12] J. González-Meneses. New Presentations of Surface Braid Groups. J. Knot Theory Ramifications 10
(2001), no. 3, 431-451.
[13] J. Guaschi and D. Juan-Pineda, A survey of surface braid groups and the lower algebraic K-theory of
their group rings. Handbook of Group Actions - II, (Advanced Lectures in Mathematics, 32), 23-76,
2015.
[14] P. de la Harpe. An invitation to Coxeter groups. Group Theory from a geometrical viewpoint. World
Scientific Publishers, Singapore (1991).
[15] A. Hatcher. Algebraic topology. Cambridge University Press, Cambridge, 2002.
[16] M. Lohrey. Word problems on compressed words. In Automata, languages and programming, Lecture
Notes in Comput. Sci. 3142. Springer-Verlag, Berlin 2004, 906-918.
[17] R. C. Lyndon and P. E. Schupp. Combinatorial Group Theory. Springer-Verlag, 1977.
[18] W. Plandowski, Testing equivalence of morphisms on context-free languages. In Algorithms ESA’94
(Utrecht), Lecture Notes in Comput. Sci. 855, Springer-Verlag, Berlin 1994, 460-470.
[19] S. Schleimer. Polynomial-time word problems. Comment. Math. Helv. 83 (2008), 741-765.
[20] G. P. Scott. Braid groups and the group of homeomorphisms of a surface. Proc. Camb. Phil. Soc. 68
(1970) 605-617.
Juan González-Meneses
Dpto. Álgebra
Facultad de Matemáticas
Instituto de Matemáticas (IMUS)
Universidad de Sevilla
Avda. Reina Mercedes s/n
41012 Sevilla (SPAIN)
[email protected]
Marithania Silvero
Institute of Mathematics
Polish Academy of Sciences
ul. Śniadeckich, 8
00-656 Warsaw (POLAND)
[email protected]
16
| 4math.GR
|
Asphericity of groups defined by graphs
Vadim Bereznyuk
A graph Γ labelled by a set S defines a group G(Γ) whose generators are the set
of labels S and whose relations are all words which can be read on closed paths
of this graph. We introduce the notion of aspherical graph and prove that
such a graph defines an aspherical group presentation. This result generalizes
arXiv:1711.05293v1 [math.GR] 14 Nov 2017
a theorem of Dominik Gruber on graphs satisfying graphical C(6)-condition.
1
Introduction
Every group can be defined by a set of generators and a set of relations among these
generators. The latter set can be excessive: there can be some non-trivial identities among its
elements. For example, in a group ha, b, c | ab−1 , bc−1 , ac−1 i a relation ac−1 can be written
as ab−1 bc−1 . Roughly speaking, a presentation is called aspherical if all identities among its
relations are trivial. It can be formalized in various ways, so there are quite a few different
definitions of asphericity (see, for example, [2]).
It is well known that asphericity follows from the classical small cancellation conditions.
In 2003 Mikhail Gromov briefly introduced a graphical analogue of small cancellation theory
in his paper [5]. After that Yann Ollivier gave a combinatorial proof of a theorem of Gromov,
which in particular states asphericity of groups defined by graphs satisfying graphical C ′ (1/6)condition [6]. In 2015 Dominik Gruber proved asphericity of groups defined by graphs satisfying graphical C(6)-condition [4].
We introduce a notion of aspherical graph and prove that such a graph defines an aspherical
group. It is easy to show that a graph satisfying graphical C(6)-condition is aspherical. Thus
our result generalizes Gruber’s theorem.
Classical small cancellation theory operates with presentations where every two distinct
relations have quite short common parts. In graphical small cancellation theory, a group is
defined by a labelled graph. The set of generators is the set of labels and the set of relations
is the set of all words which can be read on closed paths of the graph. Thus every relation
corresponds to a closed path where this relation can be read. Unlike classical case, two
distinct relations can have a long common part, but only if this common part originates from
the graph. It means that paths of the graph corresponding to these relations have the same
common part as relations themselves.
Recall that a reduction pair in a diagram is a pair of distinct faces of the diagram such
that their boundary cycles share a common edge and such that their boundary cycles, read
starting from that edge, clockwise for one of the faces and counter-clockwise for the other,
are equal as words. A spherical diagram is reduced if there are no reduction pairs. If there
exists no reduced spherical diagram over a presentation, then the presentation is called
diagrammaticaly aspherical. It is well known that presentations satisfying classical small
cancellation conditions are diagrammaticaly aspherical.
In graphical case, we call a pair of faces a graphical reduction pair if these faces share
an edge originating from the graph. A spherical diagram is graphically reduced if there exists
no graphical reduction pair in this diagram. If there exists no graphically reduced diagram
over a presentation whose set of relations is the set of labels of all simple closed paths
of the graph, then we call this graph aspherical. It is easy to show that asphericity of a graph
follows from the graphical small cancellation C(6)-condition.
It turns out that asphericity of a graph implies topological asphericity of the corresponding
group. This is the main result of this paper.
1
Note that, when we define a group by a graph, we can restrict the set of relations to the set
of labels of all simple closed paths. It does not change the group. This set of relations can
be reduced further. We can choose an arbitrary basis of the fundamental group of the graph.
Then the set of relations will be the set of cyclically reduced paths of that basis. Again it
does not change the group. The obtained presentation is topologically aspherical if the graph
is aspherical.
The paper begins with a brief introduction to theory of groups defined by graphs. In Section
1.2 the main result is formulated and in Section 1.3 the proof is outlined. In Section 2 we
give exact definitions of main notions. Section 3 is devoted to the link between identities
among relations of a presentation and spherical diagrams over this presentation. The full
proof of the main theorem can be found in Section 4. In Section 5 we show that our result
generalizes Gruber’s theorem.
The author thanks Anton Klyachko for many useful conversations and remarks.
1.1
Groups defined by graphs and graphical small
cancellation conditions
Let Γ be an oriented graph every edge of which is labelled by an element of a finite set S.
Let p be a path in the graph Γ. If we go along this path then we can read a word ℓ(p)
in the alphabet S ⊔ S −1 . Each edge of the path gives a letter which equals to its label if
we traverse the edge in its direction and equals to its label with exponent −1 if we traverse
the edge in the opposite direction. The word ℓ(p) is called the label of the path p.
Let Rc be a set of labels of all closed paths in Γ, Rs be a set of labels of all simple
closed paths in Γ and Rf be a set of cyclically reduced labels of paths which generate a basis
of a fundamental group of each connected component of the graph Γ (note that Rc and Rs
are determined by the graph Γ itself while Rf depends on the chosen basis of the fundamental
group of Γ).
Then a group G(Γ) are defined by one of the three following presentations: hS | Rc i,
hS | Rs i or hS | Rf i. Clearly, all these presentations define the same group.
a
b
b
c
b
c
Fig. 1: G(Γ) ∼
= ha, b, c | bbc, c−1 bc−1 , b−1 a−1 b−1 i
Definition 1. A lift of a word w in the graph Γ is such a path p̄ in the graph that ℓ(p̄) ≡ w
(i.e., the label of the path p̄ coincides with the word w character by character).
Definition 2. A word w is a piece (with respect to Γ) if it has two (or more) distinct lifts
in the graph Γ.
Definition 3. Let p be a path in a graph labelled by a set S. A lift of the path p in the graph
Γ is such a path p̄ in the graph that ℓ(p̄) ≡ ℓ(p) (i.e., the label of the path p coincides with
the label of the path p̄ character by character).
Definition 4. Let p be a path in a graph labelled by a set S. The path p is a piece (with
respect to Γ) if it has two (or more) distinct lifts in the graph Γ.
2
b
D
Γ
b
a
b
b
a
b
a
c
b
c
b
c
b
c
b
a
Fig. 2: A graph Γ and a diagram D over ha, b, c | Rs i. Dotted edges originate from the graph.
If no edge occurs more than once in p, then the lift p̄ of p induces a unique lift for every
subpath q of p. We call such a lift the lift of q to Γ via p → p̄.
Everywhere further it will be clear about which graph Γ we talk, so we will call words
and paths just “pieces”, not “pieces with respect to Γ”. Similarly it will be clear about which
lift of p we talk, so we will just say “lift via p”, not “lift via p → p̄”.
Consider an example. Let Γ be a graph as in Figure 1. Then the words b, b−1 , c and c−1
are all pieces of lengths 1. The words a and a−1 are not pieces. The words bb and (bb)−1 are
all pieces among reduced words of length 2.
A labelling of a graph Γ is reduced if any two distinct edges starting at the same vertex
have distinct labels and any two distinct edges ending at the same vertex have distinct labels.
Definition 5. Let Γ be a labelled graph and let k ∈ N. We say Γ satisfies graphical C(k)condition (or Γ is a C(k)-graph) if:
• the labelling of Γ is reduced and
• no simple closed path is a concatenation of stricly fewer than k pieces.
A graph as in Figure 1 satisfies graphical C(2)-condition, but does not satisfy graphical
C(3)-condition because the simple closed path with the label bbc is a concatenation of the pieces
bb and c.
Henceforth we assume that a graph Γ satisfies graphical C(2)-condition. It implies that
the graph Γ has a reduced labelling and that any word from Rs has only one lift in this graph.
Let D be a diagram over the presentation hS | Rs i (see the next section for definitions).
Let p be a path consisted of internal edges of D and lying between faces Π1 and Π2 . A word
from Rs is written on the boundary of any face of D. Thus the boundary of any face has
a lift in the graph Γ. This lift is unique due to the above remark.
Lift the boundary of the face Π1 in the graph. After that the path p, as subpath of the
boundary, maps to some path p1 in the graph. Similarly lift the boundary of the face Π2 and
determine a path p2 . We say that the path p originates from the graph Γ if p1 = p2 .
Roughly speaking, a path p originates from the graph Γ if faces Π1 and Π2 share the same
path in the diagram and in the graph Γ itself.
Note that if a path p, lying between faces Π1 and Π2 , does not originate from the graph
Γ then it is a piece. Indeed, if p does not originate from the graph then its lifts via Π1 and
via Π2 are distinct. Therefore this path have two distinct lifts in the graph Γ, i.e., this path
is a piece.
1.2
Main result
Recall that a presentation complex K(S; R) of a presentation hS | Ri is a 2-complex which
has a 1-skeleton which consists of a single vertex and a loop labelled by s for every element s
3
from S, and which have a face with the boundary label r attached to the 1-skeleton for every
element r from R.
Definition 6. A presentation hS | Ri is called aspherical if its presentation complex K(S; R)
is aspherical, i.e., πq (K(S; R)) = 0 for all q ≥ 2.
Definition 7. We say that a diagram D over the presentation hS | Rs i is graphically reduced
if it does not have edges originating from the graph Γ.
Definition 8. We say that a graph Γ is aspherical if it satisfies graphical C(2)-condition and
there exists no graphically reduced spherical diagram over the presentation hS | Rs i.
The following theorem is the main result of this paper.
Theorem. If a graph Γ is aspherical, then the presentation hS | Rf i is aspherical.
1.3
Idea of proof
Propositions 1.3 and 1.5 of the paper [2] implies that a presentation hS | Ri, where all relations
from R are not empty and freely reduced, is aspherical if and only if the presentation is concise,
no relation is a proper power and any identity among relations of this presentation is trivial.
Recall that a presentation hS | Ri is concise if for any two distinct relations r and r′ from
R nor r, neither r−1 is conjugate to r′ .
Also recall a notion of identity among relations of a presentation hS | Ri. Assume we are
given a sequence π = (p1 , . . . , pn ) such that pi = ui riǫi u−1
i , where ri ∈ R, ui ∈ F (S) and
ǫi ∈ {+1, −1}. It is called an identity if a product of its elements is equal to the identity
element of the free group, i.e., p1 · · · pn = 1 in F (s). There are identities which we should
consider as trivial. For this reason Peiffer transformations are introduced:
1. Replace any pair of consecutive elements (pi , pi+1 ), either by the pair (pi pi+1 p−1
i , pi ) or
by the pair (pi+1 , p−1
p
p
).
i
i+1
i+1
2. Delete the pair of consecutive elements (pi , pi+1 ) if pi pi+1 = 1 in F (S).
3. Insert at any place a pair of inverse elements (p, p−1 ).
An identity is called trivial if it can be transformed to an empty identity by the finite
number of Peiffer transformation.
Lemma 2.22 of the paper [4] guarantees the first two conditions: conciseness of a presentation
and absence of proper powers. Thus we only should show that any identity among relations
is trivial.
To prove this fact we use a link between identities among relations of a presentation and
spherical diagrams over this presentation which was obtained in [3]. A plan of the proof is
the following. Assume the contrary, that there exists non-trivial identities over the presentation
hS | Rf i or, equivalently, that there exists non-trivial spherical diagrams over the presentation
hS | Rf i. Consider a part of an 1-skeleton of a spherical diagram which consists of all edges
not originating from a graph Γ. We call such a part a not originating skeleton. Consider
a non-trivial spherical diagram with the smallest not originating skeleton. Delete all edges
originating from the graph from this diagramm. It turns out that the obtained diagram is
a diagram over the presentation hS | Rs i. Moreover, it is graphically reduced because we
deleted all originating edges. A contradiction with asphericity of the graph Γ.
2
Main notions
Definitions of graphs and diagrams are given in this section.
4
2.1
Graphs
We use a definition of graph according to [1], i.e., a graph is a union of two sets V and E
with three maps α : E → V , ω : E → V , ·−1 : E → E. The elements of V is called vertices
and the elements of E is calles edges. If e ∈ E then α(e) is called the initial point and
ω(e) the terminal point. The map ·−1 assigns to every edge its inverse. For convenience we
write ·−1 (e) as e−1 . The map ·−1 should be an involution without fixed elements such as
α(e−1 ) = ω(e) and ω(e−1 ) = α(e). In fact, it is a definition of undirected graph, because for
every edge e graph also contains e−1 .
−1
A labelling of a graph Γ by a set S is such a map ℓ : E → S ⊔ S −1 that ℓ(e−1 ) = ℓ(e) .
A labelled graph is a graph with its labelling. One can think of a labbeled graph as an oriented
graph every edge of which labelled by an element of S. A labelling of a graph Γ is reduced if
any two distinct edges starting at the same vertex have distinct labels and any two distinct
edges ending at the same vertex have distinct labels.
Continue the map ℓ to a set of all paths in the graph Γ. Let p = (e1 , . . . , en ) be a path
in the graph, then put ℓ(p) = ℓ(e1 ) · · · ℓ(en ), where w1 ·w2 is concatenation. Thus ℓ(p) is a word
over an alphabet S ⊔ S −1 (not necessarily reduced). We call ℓ(p) as the label of the path p.
Let p = (e1 , . . . , en ) be a path in a graph. The path p is reduced if it contains no subpaths
(e, e−1 ). The path is trivial if it becomes empty after consecutive deletion of all subpaths
(e, e−1 ). The path is closed if its initial point coincides with its terminal point or if it is
empty. The path is simple if it is not empty and it does not contain non-empty closed
subpaths. The path p is simple closed if it is not trivial, is closed and no proper subpaths
of p is closed. An edge e is a spure if degree of at least one of its vetices is equal to 1.
2.2
Diagrams over graphs
Let R be a set of relations, then we define Rsym as the set obtained from the sets R and R−1
by considering all its elements up to cyclic shifts.
A singular disk diagram in the alphabet S is a finite and simply connected 2-dimensional
CW-complex embedded into R2 such that its 1-skeleton is a graph labelled by the set S.
The closures of its 1-cells and 2-cells are called edges and faces, respectively. The label of a face
is a cyclic word obtained by reading its boundary path in a counterclockwise direction.
A singular disk diagram over a presentation hS | Ri is a singular disk diagram in the alphabet
S every face of which has a label from Rsym . A simple disk diagram is a singular disk diagram
homeomorphic to a disk.
We introduce spherical diagrams following [3]. First we inductively define a sphercial
complex. A 2-sphere is a spherical complex. If D is a spherical complex, then D′ obtained
by attaching a simple curve or a 2-sphere to a point of D is also a spherical complex.
A spherical complex is a complex which can be obtained from a 2-sphere by the finite
number of such attachments. In other words, a spherical complex is a tree embedded into R3
some vertices of which (at least one) are replaced by spheres or by some number of spheres
attached to each other. A spherical diagram over a presentation hS | Ri is a 2-dimensional
CW-complex homeomorphic to some spherical complex, such that its 1-skeleton is a graph
labelled by the set S and such that the labels of all its faces lie in Rsym . A simple spherical
diagram is a spherical diagram homeomorphic to a sphere
We will denote the set of positively oriented (i.e., obtained by reading the boundary
in a counterclockwise direction) boundary paths of a face Π as ∂Π+ , and the set of negatively
oriented as ∂Π− . It is clear that ∂Π− = {γ −1 : γ ∈ ∂Π+ }. If P and Q are two sets of paths
then P ∩ Q = {r : ∃p ∈ P, ∃q ∈ Q such as p = rp′ , q = rq ′ }.
We can consider a lift of ∂Π+ or ∂Π− in the graph Γ as a lift of an arbitrary path from this
set. Then a lift of any else path from this set can be obtained as a cyclic shift of the known
lift.
Definition 9. Let D be a diagram over the presentation hS | Rc i. Let Π1 and Π2 be two
5
−
faces of D (not necessary distinct) and let p ∈ ∂Π+
1 ∩ ∂Π2 . We say that the path p originates
+
+
from the graph Γ if there exists such lifts of ∂Π1 and ∂Π−
2 in Γ, that a lift of p in Γ via ∂Π1
−
and via ∂Π2 are equal.
Note again that if the graph Γ satisfies graphical C(2)-condition then lifts of ∂Π+
1 and
∂Π−
2 are unique.
3
Identities and spherical diagrams
For any identity π = (p1 , . . . , pn ) over a presentation hS | Ri we can construct a spherical
diagram over the same presentation by the so-called van Kampen construction. First, fix
draw a face with a spur such
a point v on the plane. After that for each pi = ui riεi u−1
i
that the spur starts at v, the label of the spur is ui and the label of the face is riεi . Then
the boundary label of the obtained diagram is equal to the word p1 · · · pn . So this boundary
label is trivial because π is an identity. It means that there exists 2 consecutive edges with
opposite labels on the boundary. Glue these edges together. We obtain again a diagram
with a trivial boundary label but with the smaller boundary size. Consecutively gluing pairs
of edges with opposite labels we finally obtain a spherical diagram D which is called a diagram
for π (see [3, Section 1.5] for details).
This procedure can be reversed. By “ungluing” faces of a spherical diagram D over
a presentation hS | Ri along edges we can obtain a bouquet of faces with spurs that gives
us a sequence of elements π over the same presentation. Moreover, π is an identity because
the diagram was spherical. Note that D is a diagram for π because we can reverse the described
procedure.
These procedures are not unique. We may obtain different diagrams and identities changing
the order of gluings and ungluings. But at the same time Proposition 8 from [3] implies that
if D is a diagram for π1 and a diagram for π2 then π1 is trivial if and only if π2 is. It allows
us to define trivial spherical diagrams with correct correspondence to trivial identies. We will
call D a trivial spherical diagram if a trivial identity can be obtained from D. Due to the fact
noted above, every identity obtained from a trivial spherical diagram D will be trivial. Thus
only trivial identities can be obtained from a trivial diagram, and only trivial diagrams can
be obtained from a trivial identity.
Note also that actually a diagram D which is a diagram for some identity π has an marked
point (the initial vertex v from the construction). Replacing of the marked point conjugates
all elements of the identity π by the same element — label of the path connecting old and new
marked points. Clearly, it does not change triviality of the diagram. Therefore we are free
to choose any point of the diagram as the marked point. Thus we wo not explicitly specify
the marked point.
And finally note that a spherical diagram can have spurs. But Proposition 8 from [3]
implies that inserting and deleting of spurs does not change triviality of the diagram.
4
Proof of the theorem
In this section Γ is a labelled by the set S graph and Rc , Rs and Rf are the sets defined at
the beginning of the section 1.1
Definition 10. Let D be a diagram over the presentation hS | Rs i. A not originating skeleton
of D is a graph that consists of all edges of D which do not originate from the graph Γ.
Lemma 1. Suppose Γ satisfies graphical C(2)-condition and let D be a simple spherical
diagram over the presentation hS | Rs i. Then a not originating skeleton of D has no spurs.
Proof. Assume the contrary. Let D be a simple spherical diagram whose not originating
skeleton contains a spur e with endpoint v. It means that all other edges incidental to v
6
Π3
Π2
e2
v
e1
e
Π1
e3
Π4
Fig. 3: Illustration for Lemma 1. Dotted edges originate from the graph.
originate from the graph Γ because otherwise the edge e would not be a spur in the originating
skeleton.
We may assume that a part of the diagram in a neighborhood of the point v is embedded
into the plane. Draw a circle with a center at v with quite a small radius such that the circle
intersects all incidental to v edges and only them. Once if an edge is not a loop and twice if
an edge is a loop. Let us traverse this circle in a counterclockwise direction starting at C ∩ e.
We cross some number of faces Π1 , . . . , Πn . We cross also the edges e1 , . . . , en−1 , en = e,
where ei lie in the intersection of faces Πi and Πi+1 .
The boundary of each face has a unique lift to the graph because Γ satisfies graphical C(2)condition. Let vi be a vertex of the graph obtained by lifting v via the boundary of the face
Πi . The edge ei originates from the graph for i = 1, . . . , n − 1, therefore vi = vi+1 for
+
i = 1, . . . , n − 1. Thus v1 = vn . It means that lifts of the edge e via ∂Π+
1 and via ∂Π1 have
a common vertex in the graph and, moreover, have the same label. It implies that these
lifts coincide because a labelling of the graph is reduced. Thus e originates fron the graph.
A contradiction.
Definition 11. We will call D a minimal non-trivial spherical diagram over the presentation
hS | Rf i if it is a non-trivial spherical diagram such that
1. A not originating skeleton of D has the smallest number of edges among all non-trivial
diagrams and among all such diagrams a not originating skeleton of D has the smallest
number of vertices.
2. D has the smallest number of edges among all diagrams satisfying the first condition.
Note that a minimal non-trivial spherical diagram D is always a simple spherical diagram.
Indeed, assume the contrary, let D not be a simple spherical diagram. We noted earlier that
insertion and deletion of spurs does not affect its triviality. Therefore D does not have spurs.
It means that there exists such a point v in D that the diagram splits into two spherical
diagrams D1 and D2 after cutting D at v. Let π1 be an identity obtained from the diagram
D1 and π2 be an identity obtained from the diagram D2 . These identities are trivial because
D is a minimal non-trivial spherical diagram. Then an identity π = (π1 , π2 ) is trivial as
well. But π can be obtained from the diagram D. Thus D is a trivial spherical diagram that
contradicts its definition.
Lemma 2. Let Γ be a C(2)-graph. Let D be a minimal non-trivial spherical diagram over
the presentation hS | Rf i and let ∆ be a proper subdiagram of D such that it have at least
one face and such that ∂∆ lies in the not originating skeleton of D. Then the boundary label
of ∆ is not trivial as an element of the free group F (S).
Proof. The proof is very similar to the proof of the fact that D is a simple spherical diagram.
Assume the contrary. Denote by ∆′ a diagram obtained from D by cutting off ∆. Since ∆
7
and ∆′ have a trivial boundary label we can obtain some identity π from ∆ and some identity
π ′ from D′ .
Denote by ∆s a spherical diagram obtained from ∆ by gluing its boundary. It is clear that
∆s is a diagram for π. And it is also clear that a not originating skeleton of ∆s is smaller than
a not originating skeleton of D because ∆s was obtained by gluing the boundary of ∆, and this
boundary consists only of not originating edges. Thus ∆s is a trivial spherical diagram and π
is a trivial identity. The same is true for π ′ . Then an identity Σ = (π, π ′ ) is trivial as well. But
D is a diagram for Σ and it means that D is a trivial spherical diagram. A contradiction.
Let D be a simple spherical diagram. Denote by D̄ a simple spherical diagram obtained
from D by erasing all edges originating from the graph Γ (if isolated vertices are left we delete
them). Clearly, a 1-skeleton of D̄ coincides with a not originating skeleton of D. Note that
actually D̄ may have some not simply connected “faces”, but we still consider them as faces
and call them not simply connected faces. Note that the boundary of a not simply connected
face consists of some connected components.
Lemma 3. Let Γ be a C(2)-graph. Let D be a minimal non-trivial spherical diagram over
the presentation hS | Rf i. Then the boundary label of any face of the diagram D̄ is reduced
(the boundary label of any connected component for not simply connected faces).
Proof. Assume the contrary. Let Π̄ be a face with a not reduced boundary label. Let e and
f be the edges where reduction is occured. We will use diamond moves introduced in [3].
For definiteness we assume that ω(e) = ω(f ) and ℓ(e) = ℓ(f ) = a. Consider in detail only
the case when all 3 points ω(e), α(e), α(f ) are distinct. Transform the diagram D as shown
in Fig. 4. Note that we reduce the not originating skeleton at least by one edge because one
of the two new edges appears to be inside the face Π̄. But the not originating skeleton does
not have spurs so this edge originates from the graph.
Π̄
Π̄
a
Π̄
a
a
a
a
a
a
a
Fig. 4: The first case of the Lemma 3.
Diamond moves for other cases can be found in [3, Section 1.4]. We do not consider these
cases in detail because they do not differ much from the first case: after applying a diamond
move two old edges are replaced by two new ones and one of the two new edges appears to be
inside the face Π̄ and therefore originates from the graph. Thus the not originating skeleton
is reduced anyway.
Lemma 4. Let p be a simple closed path in the graph Γ. Then there exists a diagram D
over the presentation hS | Rf i such that ∂D lifts to the path p and all internal edges of D
originates from the graph Γ.
Proof. Forget for a while that Γ is labelled by the set S. Label every edge e of the graph Γ by
the new unique label te . Assume the path p lies in the connected component Γ′ of the graph
Γ and let B ′ be the basis of Π(Γ′ , v) from which some relations of Rf are obtained. Let q be
a path which runs from the vertex α(p) to the vertex v and let B = {qb′ q −1 | b′ ∈ B ′ }. Then
the path p are generated by the paths from B, i.e, p = bǫ11 · · · bǫnn , bi ∈ B. Each bi = ui ri u−1
i
8
where ri is a cyclic reduction of the path bi . Due to the definition of the set Rf , the label
of each path ri lie in Rf .
Fix a point s on the plane. For each bǫi i draw a face with a spur such that the spur starts
at s and has a label ℓ(ui ) and such that the boundary label of the face is ℓ(riǫi ). Thus we
obtain a diagram E such that its boundary label is freely equal to the label of the path
p. If there exists a pair of consecutive edges with the same labels and opposite directions
on the boundary of E then glue them. Doing this several times we obtain a reduced boundary
label. The labelling of Γ is reduced and p is a simple closed path so its label is reduced. Thus
the boundary label of E and the label of p is equal as words.
Note that all internal edges of E originate from the graph because they have unique labels
and therefore have unique lifts to the graph. Now recall the original labelling by the set S.
After replacing unique labels by original ones all internal edges still originate from the graph
(because boundaries of the faces still have unique lifts to the graph and these lifts coincide
with lifts to the graph with unique labels) and the boundary label of the diagram still is
equal to the label of p as words. Thus we obtain the desired diagram for the path p over
the presentation hS | Rf i.
Lemma 5. Let Γ be a C(2)-graph. Let D be a minimal non-trivial spherical diagram over
the presentation hS | Rf i and let Π be a face of D̄. Then if Π is simply connected and its
boundary is a simple closed path then this boundary lifts to a simple closed path in the graph Γ.
Proof. Assume the contrary, that there exists a face Π in D̄ such that it have a simple
boundary which lifts to a not simple closed path in the graph. By Lemma 3, the boundary
label of Π is reduced so there exists a subpath p of the path ∂Π which lifts to a simple
closed path p̃ in the graph. Let s and t be respectively the start and the end of the path
p. Transform the diagram as shown in Fig. 5: cut the diagram along the path p and glue
together s and t. Both ∂Π1 and ∂Π2 lift to the simple closed path p̃ in the graph Γ. By
the previous lemma, there exists a diagram E over the presentation hS | Rf i such that its
boundary coincides with p̃ and all internal edges originate from the graph. Glue E at the place
of Π1 and a diagram symmetric to E at the place of Π2 . Thus we again obtain a spherical
diagram over the presentation hS | Rf i. Denote it by D′ .
Note that Π ∩ Π1 = ∂Π1 originates from the graph because it lifts to the simple closed
path in the graph and due to the graphical C(2)-condition such a lift is unique. Therefore
D′ has a smaller not originating skeleton than D because we glued together the vertices s
and t which lay in a not originating skeleton of D and so we reduced the number of vertices
in the not originating skeleton while the number of not originating edges are still the same,
because all internal edges of E originate from the graph.
Moreover, D′ is not trivial because we glued into D the set of opposite faces and we can
−1
think of it as adding to the not trivial identity π some sequence (p1 , . . . , pn , p−1
n , . . . , p1 )
that corresponds to the n-fold application of the insert Peiffer transformation. Thus we
obtained a not trivial spherical diagram with a smaller not originating skeleton than D
has. A contradiction.
Lemma 6. Let Γ be an aspherical graph and let D be a minimal non-trivial spherical diagram
over the presentation hS | Rf i. Then every face of D̄ is simply connected and the boundary
of every face of D̄ lifts to a simple closed path in the graph Γ.
Proof. Assume the contrary. Then there are two types of bad faces: not simply connected
faces and simply connected faces with a not simple boundary path which does not lift to
a simple closed path in the graph. Note that since a not originating skeleton does not have
spurs every face with a not simple boundary encloses some subdiagram which has at least
one face. Similarly every not simply connected face does.
Consider a face Π which is the innermost bad face. It means that a subdiagram ∆ enclosed
by this face contains only simply connected faces which boundaries lift to simple closed paths.
9
s
Π
s
Π
Π
t
Π1
Π2
t
Fig. 5: Illustration for Lemma 5.
First examine the case when Π is a simple connected face. As before let p be a subpath
of ∂Π which lifts to a simple closed path and let s and t be respectively the start and the end
of the path p. If s 6= t then acting as in the proof of Lemma 4 we obtain a contradiction with
minimality of D. Thus s = t.
Let q be such a subpath of ∂Π that pq ∈ ∂Π+ . Note that either p or q encloses some
subdiagram ∆′ of the diagram ∆. If q does then we can again assume that there is a subpath
p′ in q with the end points s′ and t′ which lifts to a simple closed path. Arguing as before
we obtain s′ = t′ . And now p′ encloses some subdiagram ∆′ of the diagram ∆. Note that ∆′
as a subdiagram of ∆ does not contains bad faces. So in the both cases there is a subpath
which encloses some subdiagram without bad faces and which lifts to a simple closed path.
We denote this path by p and this subdiagram by ∆.
Since ∆ does not contain bad faces boundaries of all faces of ∆ lift to a simple closed
paths. But the boundary of ∆ itself, which equals to a path p, lifts to a simple closed path.
Thus we can think of ∆ as a spherical diagram over the presentation hS | Rs i no edge of which
originates from the graph. But it contradicts asphericity of the graph Γ.
In the case of a not simple connected face we act similarly. Let ∆ be a subdiagram enclosed
by Π. Due to the choice of Π all faces of ∆ lift to simple closed paths. If ∂∆ lifts to a simple
closed path in the graph then we again contradict asphericity of the graph. Otherwise we take
a subpath p of ∂∆ with the end points s and t which lifts to a simple closed path in the graph.
If s 6= t then we contradict minimality of D. If s = t then a subdiagram enclosed by p gives
a spherical diagram over the presentation hS | Rs i where no edge originates from the graph
that contradicts asphericity of the graph.
s
t
∆
Π
Fig. 6: Illustration for Lemma 6.
10
The following lemma can be found in [4].
Lemma 7 (Gruber). Let Γ be a connected C(2)-labelled graph. Let R be the set of cyclic
reductions of words read a set of free generators of π1 (Γ, v) for some v ∈ Γ. Let (D, v) be
a simple disk diagram over R with freely trivial boundary word such that every interior edge
originates from Γ. Then any sequence obtained from (D, v) is a trivial identity sequence.
Now we can prove the theorem.
Proof of the theorem. Assume the contrary, that there exists non-trivial identities. Let D
be a minimal non-trivial spherical diagram over the presentation hS | Rf i. Consider a not
originating skeleton of D. If the skeleton is empty then all edges of the diagram originate
from the graph. In this case we can unglue D along some edge connecting distict vertices and
we obtain a simple disk diagram with a trivial boundary label all edges of which originate
from the graph Γ. Clearly, all these edges lift to the same connected component of the graph.
Then, by the previous lemma, D is a diagram for a trivial identity that contradicts its
definition. Thus the not originating skeleton is not empty. By the all previous lemmas every
face of D̄ is simply connected and lifts to a simple closed path of the graph. Therefore D̄ is
a spherical diagram over the presentation hS | Rs i where no edge originates from the graph
that contradicts asphericity of the graph.
5
Corollary of the main result
Now we show that graphical C(6)-condition implies asphericity of a graph. To do this we will
need a notion of [p, q]-diagrams introduced in [1].
Let p and q be positive integers such that p1 + q1 = 12 . Degree of a face is a number of edges
in its boundary path. A face of a diagram is called interior if it has no common edges with
the boundary of the diagram. If D is a non-empty diagram such that each interior vertex of D
has degree at least p and all faces of D have degree at least q, then D is called [p, q]-diagram.
If D is a non-empty diagram such that each interior vertex of D has degree at least p and
each interior face of D have degree at least q, then D is called (p, q)-diagram.
The following theorem can be found in [1].
Theorem (Curvature formula). Let D be a simply connected [p, q]-diagram which contains
at least two vertices. Then
X p
+ 2 − d(v) ≥ p
q
v∈∂D
This theorem implies that there do not exist any simple spherical (3, 6)-diagrams. Indeed,
note at first that a simple spherical (3, 6)-diagram contains more than one vertex because
otherwise there would exist a face with degree 1 that contradicts the fact that the diagram is
a (3, 6)-diagram. Second, we can unglue the diagram along an edge connecting two distinct
vertices. Then we obtain a [3, 6]-diagram with exactly 2 vertices in the boundary and these
vertices have degree at least 3. It contradicts the curvature formula.
We also need an operation of “forgetting” vertices of degree 2. Let D be a diagram without
spurs. Call an arc such a path in D that its every internal vertex has degree 2 and the end
points have degrees more than 2. Then if we replace every arc in D by an edge, we obtain
a diagram where each vertex has degree at least 3.
Now we are ready to prove that graphical C(6)-condition implies asphericity of the graph.
Proof. Assume the contrary, that there exists a simple spherical diagram D over the presentation hS | Rs i no edge of which originates from the graph. By Lemma 1, the diagram D
has no spurs. Note that for any two adjacent faces Π1 and Π2 their common boundary
is a piece because it does not originate from the graph. Due to graphical C(6)-condition
it means that every face consists of at least 6 arcs. Therefore after “forgetting” vertices
11
of degree 2 the diagram D becomes a simple spherical (3, 6)-diagram which does not exist.
A contradiction.
Corollary. If Γ is a C(6)-graph, then the presentation hS | Rf i is aspherical.
For the first time this result was obtained by Dominic Gruber in [4].
References
[1] R. C. Lyndon, P. E. Schupp, Combinatorial Group Theory, Springer-Verlag, 1977
[2] I. M. Chiswell, D. J. Collins, J. Huebschmann, Aspherical group presentations, Math. Z.
178 (1981), 1–36.
[3] D. J. Collins and J. Huebschmann, Spherical diagrams and identities among relations,
Math. Ann. 261 (1982), 155–183.
[4] D. Gruber, Groups with graphical C(6) and C(7) small cancellation presentations, Trans.
Amer. Math. Soc. 367 (2015), no. 3, 2051-2078, preprint versions: arXiv:1210.0178
[5] M. Gromov, Random Walk in Random Groups, GAFA, Geom. Funct. Anal. 13 (2003),
no. 1, 73–146.
[6] Y. Ollivier, On a small cancellation theorem of Gromov, Bull. Belg. Math. Soc. Simon
Stevin 13 (2006), no. 1, 75–89.
12
| 4math.GR
|
Approximation Complexity of Max-Cut on
Power Law Graphs
arXiv:1602.08369v1 [cs.DS] 26 Feb 2016
Mikael Gast∗
Mathias Hauptmann†
Marek Karpinski‡
Abstract
In this paper we study the MAX-CUT problem on power law graphs (PLGs)
with power law exponent β. We prove some new approximability results on that
problem. In particular we show that there exist polynomial time approximation
schemes (PTAS) for MAX-CUT on PLGs for the power law exponent β in the
interval (0, 2). For β > 2 we show that for some ε > 0, MAX-CUT is NP-hard to
approximate within approximation ratio 1 + ε, ruling out the existence of a PTAS
in this case. Moreover we give an approximation algorithm with improved constant
approximation ratio for the case of β > 2.
1
Introduction
In the study of large-scale complex networks, a large body of theoretical and practical work is devoted to clustering and partitioning problems [NG04; WGH04; New06;
DT13].The aim is to identify and to characterize natural partition structures of existing
real world networks, like protein interaction networks, online social networks and parts
or layers of the World Wide Web.
Given an undirected graph G = (V, E) the MAX-CUT problem asks for a partition
of the vertex set into two parts such as to maximize the number of edges between them.
MAX-CUT is one of the classic 21 NP-complete problems listed in [Kar72] and has
applications ranging from VLSI design and cluster analysis to statistical physics [Bar+88].
In this paper we consider the MAX-CUT problem in the context of large-scale complex networks, more precisely in the context of so called power law graphs (PLG). The
distinguishing feature of power law graphs is that their node degree distribution follows
a power law, that is, the number of nodes of degree i is proportional to i−β , for some
fixed power law exponent β > 0. A power law degree distribution has been observed for
a large number and variety of social, information, technological and biological networks
[CSN09].
∗
Dept. of Computer Science and the Hausdorff Center for Mathematics, University of Bonn. e-mail:
[email protected]
†
Dept. of Computer Science, University of Bonn. e-mail: [email protected]
‡
Dept. of Computer Science and the Hausdorff Center for Mathematics, University of Bonn. Research
supported by DFG grants and the Hausdorff grant EXC59-1/2. e-mail: [email protected]
1
Previous Results In the general setting partition problems like the MAX-CUT (and
MAX-BISECTION) problem were shown to be APX-hard. The greedy algorithm or
random assignment algorithm for MAX-CUT was shown to yield a 1/2-approximation for
the problem [SG76]. Beside some (1/2 +o(1))-approximation algorithms, the currently best
approximation ratio for the problem is due to Goemans and Williamson in their seminal
paper [GW95]. Using semidefinite programming they achieved an αGW -approximation
algorithm for MAX-CUT, where αGW ≈ 0.879 is the trigonometric quantity αGW =
2/π sin Θ for Θ ≈ 2.33.
The APX-hardness of the problem on general instances was shown by Papadimitriou
and Yannakakis [PY91]. Under the assumption P 6= NP, the first explicit inapproximability bound of 84/83 was proven by Bellare, Goldreich, and Sudan [BGS98] and later refined
to the current best bound of 17/16 due to Håstad [Hås01] and Trevisan et al. [Tre+00].
Assuming that the Unique Games Conjecture (UGC) holds true, Khot et al. [Kho+07]
showed that MAX-CUT is UGC-hard to approximate to within αGW + ε. Thus, assuming
the UGC, αGW is the best possible approximation ratio achievable in the general case.
Furthermore, a lot of interest has been devoted to the case of more restricted graph instances of MAX-CUT. One of the first results was a polynomial time algorithm for the case
of planar graphs [Had75]. In the metric variant of MAX-CUT (and MIN-BISECTION),
instances are complete graphs where edge weights are given by the metric distance of
the corresponding endpoints. For general metrics a randomized PTAS for MAX-CUT
is due to Fernandez de la Vega and Kenyon [FK01] and for the case of tree metrics, a
polynomial time algorithm was constructed by Karpinski, Lingas, and Sledneu [KLS13].
A PTAS for metric MIN-BISECTION was shown by Fernandez de la Vega, Karpinski,
and Kenyon [FKK04], whereas for the general case the existence of a PTAS remains an
open question (cf. [Kar02]).
Another important special case is when the corresponding problem instances are
dense, i.e. for some δ > 0, the number of edges is lower bounded by δ · n2 . Arora,
Karger, and Karpinski [AKK95] gave a PTAS for dense MAX-CUT and, more generally,
for dense MAX-k-CSPs. The result was extended also to sub-dense instances in [Fer+05;
Bjö05] and to dense weighted instances in [FK00].
Regarding lower approximation bounds, a lot of interest has been devoted to the case
of MAX-CUT in degree d bounded graphs and d-regular graphs. In a series of papers
[BK99; BK01] Berman and Karpinski showed, among other results, that the MAX-CUT
problem restricted to 3-regular graphs is NP-hard to approximate to within 152/151.
2
Main Results
In this paper we study the MAX-CUT problem on power law graphs. Our main results
are new and improved upper approximation bounds for the problem. In particular we
show that there exists a polynomial time approximation scheme (PTAS) for MAX-CUT
on (α, β)-power law graphs for 0 < β < 2.
For the range β < 1, we observe that (α, β)-power law graphs are dense (in the average
sense, i.e. the number of edges is Ω(n2 )) and the result of Arora, Karger, and Karpinski
[AKK95] can be applied to obtain a PTAS. For β = 1 the graphs are not dense anymore.
For this case we prove that (α, β)-power law graphs are core-dense and use the result of
2
Fernandez de la Vega et al. [Fer+05] yielding a PTAS for the problem.
In the range 1 < β < 2, none of the above results directly apply because instances are
neither dense nor core-dense. In order to construct a PTAS in this case, we partition the
vertex set of power law MAX-CUT instances into two sets of high degree vertices and
low degree vertices. We show that for a suitable choice of the partition parameters the
induced subgraph of high degree vertices is asymptotically dense, and at the same time
the total number of edges induced by the low degree vertices is small. Thus, in order to
1
obtain a 1+ε
-approximate cut, we run the algorithm of [AKK95] on the subgraph of high
degree vertices and afterwards placing the remaining vertices arbitrary.
For β > 2 we show that the Goemans-Williamson algorithm [GW95] can be combined
with a preprocessing to yield an improved constant approximation ratio. Moreover we
show that for β > 2, MAX-CUT cannot be approximated with a constant approximation
ratio arbitrary close to 1. For this purpose we use the lower bound result of [BK99; BK01],
construct an embedding of low degree graphs into power law graphs and obtain in this way
also explicit approximation lower bounds depending on the power law exponent β > 2.
Moreover, a variant of this construction also proves the NP-hardness of MAX-CUT in
power law graphs for the whole range β > 0.
Besides NP-hardness in the exact setting, the status of MAX-CUT in PLGs for β = 2
remains unsettled. However, we consider the case when β is a function of the size of the
PLG that converges to 2 from below. We call this the functional case. In particular we
1
show that for βf = 2 − f (α)
, MAX-CUT in (α, βf )-PLGs admits a PTAS provided the
convergence of βf to 2 is sufficiently slow, namely for all sublinear functions f (α) = o(α).
Organization of the Paper Section 3 provides the definition of the (α, β)-PLG model
due to [ACL01] and related notations. In Section 4 we present our PTAS constructions
1
for MAX-CUT on (α, β)-PLG for 0 < β < 2. The functional case βf = 2 − f (α)
is
considered in Section 4.3. Furthermore we show an improved constant approximation
ratio for the case β > 2 in Section 5. Finally, in Section 6, we prove APX-hardness of
the problem for the case β > 2 and NP-hardness for the whole range β > 0.
3
Preliminaries
In this section we first give the formal definition of (α, β)-power law graphs. Then we
provide notations for sizes and volumes of some subsets of the vertex set of a power law
graph which we call intervals. Later on we will give estimates of these quantities in the
analysis of our upper and lower bound constructions for MAX-CUT.
Definition 3.1. [ACL01] An undirected multigraph G = (V, E) with self loops is called
an (α, β) power law graph if the following conditions hold:
• The maximum degree is ∆ = ⌊eα/β ⌋.
• For i = 1, . . . , ∆, the number yi of nodes of degree i in G satisfies
eα
yi = β
i
3
The following estimates for the number n of vertices of an (α, β)-power law graph are
well known [ACL01]:
n≈
eα/β
1−β
α
α·e
ζ(β) · eα
for 0 < β < 1,
for β = 1,
for β > 1.
m≈
1 e2α/β
2 2−β
1
αeα
4
1
ζ(β −
2
1)eα
for 0 < β < 2,
for β = 2,
for β > 2.
−β
Here ζ(β) = ∞
is the Riemann Zeta Function.
i=1 i
A random model for (α, β)-power law graphs was given in [ACL01] and is constructed
in the following way:
P
1. Generate a set L of degG (v) distinct copies of each vertex v.
2. Generate a random matching on the elements of L.
3. For each pair of vertices u and v, the number of edges joining u and v in G is equal
to the number of edges in the matching of L, which join copies of u to copies of v.
degG (v) = 1
degG (v) = 2
degG (v) = 3
V
...
...
...
...
L
edge
self-loop
multi-edge
Given an (α, β) power law graph G = (V, E) with n vertices and maximum degree ∆
and two integers 1 6 a 6 b 6 ∆, an interval [a, b] is defined as the subset of V
[a, b] = {v ∈ V |a 6 degG (v) 6 b}.
If U ⊆ V is a subset of vertices, the volume vol(U) of U is defined as the sum of node
degrees of nodes in U. We will make use of estimates of sizes and volumes of node intervals
in (α, β)-PLGs.
4
Approximation Schemes for 0 < β < 2
We will now show that for every constant power law exponent β ∈ (0, 2), there is a PTAS
for the MAX-CUT problem in (α, β)-PLGs. It turns out that for β ∈ (0, 1) this follows
directly from the results in [AKK95], since in that case the power law graphs are dense,
(Section 4.1). Recall that a graph G = (V, E) with n vertices is called δ-dense if the
number of edges satisfies |E| > δ · n2 . For β = 1, (α, 1)-PLGs are not dense anymore.
Nevertheless we can establish existence of a PTAS by showing that (α, 1)-PLGs are coredense, a notion which was introduced in [Fer+05].
4
The Case 0 < β 6 1
4.1
First we consider the case when the power law exponent β is strictly less than 1. In this
α/β
case, the number n of nodes is asymptotically equal to e1−β and the number m of edges
satisfies m =
1 e2α/β
.
2 2−β
Thus, in this case the graphs are asymptotically dense.
Theorem 1 ([AKK95]). For every δ > 0, there is a PTAS for MAX-CUT in δ-dense
graphs.
Corollary. There exists a PTAS for MAX-CUT in power law graphs with power law
exponent β < 1.
Now we consider the case when β = 1.
Definition 4.1. [Fer+05] The core-strength of a weighted r-uniform hypergraph H =
(V, E) with |V | = n nodes given by an r-dimensional tensor A : V × . . . × V → R is
n
X
Di
i=1
!r−2
X
i1 ,...,ir ∈V
where
Di =
X
A2i1 ,...,ir
,
Qr
j=1 (Dij + D̄)
Ai,i2 ,...,ir , D̄ =
i2 ,...,ir ∈V
n
1X
Di
n i=1
A class of weighted r-uniform hypergraphs is core-dense if the core-strength is O(1).
In particular, the class C of unweighted graphs is core-dense if
sup
X
G=(V,E)∈C {i,j}∈E
1
= O(1)
(Di + D̄)(Dj + D̄)
where Di is the degree of node i in G and D̄ =
1
n
P
i
Di is the average-degree of G.
Theorem 2. For β = 1, the class of (α, β)-Power Law Graphs is core-dense.
Corollary. For β = 1, there is a PTAS for MAX-CUT in (α, β)-Power Law Graphs.
Proof of the Theorem. Let G be an (α, 1)-PLG. The average-degree of G is asymptotically equal to
α
eα
eα α
X eα
X
1
1 X
eα (eα − 1)
e
1 ∆=e
i = α eα·2 −
·i > α
·i−
D̄ = α
αe i=1 i
αe i=1 i
αe
2
i=1
!
!
= (1+o(1))
eα
2α
Now the core-strength of G is
α
e
X
α3
1
1
eα
6
·
= o(1),
6
eα 2
α
)
i
(i
+
e
(deg(i)
+
D̄)(deg(j)
+
D̄)
i=1
2α
e={i,j}∈E
X
which concludes the proof of the theorem.
5
4.2
The Case 1 < β < 2
We consider now the case when the power law exponent β satisfies 1 < β < 2. Our
approach is as follows. We choose a subset [x∆, ∆] of high-degree vertices and construct
a cut for the subgraph G[x∆,∆] induced by these vertices. Here x ∈ (0, 1) is a parameter of
the construction. We will show that we can choose x in such a way that G[x∆,∆] is dense
and the volume of the residual set of vertices [1, x∆) is small, namely vol([1, x∆)) =
o(|E|). Then we will construct a (1 + ε)-approximate solution for MAX-CUT on G[x∆,∆]
and afterwards place the remaining vertices arbitrarily.
This approach is based on precise estimates for sizes and volumes of node degree intervals of the form [1, x∆] and [x∆, ∆], where x ∈ (0, 1) is the parameter of the construction
and ∆ = ⌊eα/β ⌋ is the maximum degree. These estimates rely on the following lemma,
which is also illustrated in Figure 1.
Lemma 3. Let f : R+ → R+ be a monotone decreasing integrable convex function. Then
we have the following bounds.
(a)
(b)
b
P
a
b
P
a
⌊f (i)⌋ ∈
"
i·⌊f (i)⌋ ∈
b+1
R
a
"
f (t)dt − (b − a + 1),
b+1
R
a
tf (t)dt −
b(b+1)
2
−
b+1
R
a
#
f (t)dt + f (a) − f (b + 1)
(a−1)a
2
,
b+1
R
a
tf (t)dt + |(b + 1)f (b + 1) − af (a)|
f (a)
f (a+1)
f (a+2)
a a+1 a+2
b b+1
Figure 1: Estimating the volumes of node degree intervals.
6
#
Using this lemma, we obtain the following bounds for the size of [x∆, ∆]:
|[x∆, ∆]| =
∆
X
x∆
∈
$
eα
jβ
%
∆+1
∆+1
Z
Z
eα
eα
dt − (∆(1 − x) + 1),
dt +
tβ
x∆
"
"
α
= eα
tβ
x∆
t1−β
1−β
#∆+1
− ∆(1 − x) − 1, eα
x∆
"
eα
eα
−
(x∆)β (∆ + 1)β
t1−β
1−β
#∆+1
x∆
eα
eα
+
−
(x∆)β
(∆ + 1)β
1
1
e
· β−1 · β−1 − 1 − ∆(1 − x) − 1,
β−1 ∆
x
#
eα
eα
1
1
1
1
1
·
· β−1 − β−1 + β · β − β
β − 1 ∆β−1
x
2
∆
x
2
!
#
"
1−β
∆
1
1
1
1
1
x
−1
,
−1+x−
−
+ β − β
= ∆
β−1
∆
β − 1 xβ−1 2β−1
x
2
⊆
We also obtain the following estimate for the volume of the interval [1, x∆].
vol([1, x∆]) =
x∆
X
1
∈
x∆+1
Z
1
eα
j· β
j
$
%
eα
x∆(x∆ + 1)
dt −
,
β−1
t
2
x∆+1
Z
1
eα
tβ−1
dt + eα
!
1
−1
(x∆ − 1)β−1
x∆(x∆ + 1)
eα
(x∆ + 1)2−β − 1 −
⊆
,
2−β
2
!#
eα
1
2−β
α
(x∆ + 1)
−1 +e
−1
2−β
(x∆ − 1)β−1
"
"
2α/β
⊆ (1 − o(1))e
x2−β 2α/β
x2−β
eα/β
x2
, (1 + o(1))
+ (1 + o(1)) β−1
−
e
2−β
2
2−β
x
!
#
In particular, if x is constant within the interval (0, 1), i.e. does not converge to 0 or 1,
then we obtain the following estimates:
α/β
• The size of the interval [x∆, ∆] is |[x∆, ∆]| = (1 + o(1)) eβ−1
1
xβ−1
−
1
2β−1
.
• Thus for x being constant, the kernel function for the total error is O(e2α/β )
2−β
• For x being constant, the volume of [1, x∆] is vol([1, x∆]) = (1 + o(1)) x2−β · e2α/β .
Thus we proceed as follows. Given an (α, β)-PLG G and ε > 0, we first choose x ∈ (0, 1)
such that
1 e2α/β
|E(G[x∆,∆])| >
− vol([1, x∆]) >
22−β
e2α/β
1
1
· |E|,
− x2−β
> 1−
2
2−β
τ (ε)
!
where τ (ε) is a function of ε yet to be defined. Then we choose a second parameter
ε′ > 0 and construct a cut in G[x∆,∆], using the AKK-algorithm with precision parameter
7
ε. This yields a cut in G[x∆,∆] of size at least MAX-CUT(G[x∆,∆]) − ε′ · |[x∆, ∆]|2 . By
the choice of x we have
MAX-CUT(G[x∆,∆] ) > MAX-CUT(G) − vol([1, x∆]) > MAX-CUT(G) −
1
· |E|
τ (ε)
Moreover,
|[x∆, ∆]|2 6
e2α/β
(β − 1)2
1
xβ−1
−1
2
= |E| ·
1
2 · (2 − β)
−1
(β − 1)2 xβ−1
2
Thus the size of the cut constructed in this way is at least
2
2(2 − β)
1
|E|
− ε′ · |E| ·
−
1
τ (ε)
(β − 1)2 xβ−1
2 !
2
4(2
−
β)
1
> MAX-CUT(G) 1 −
− ε′ ·
−1
τ (ε)
(β − 1)2 xβ−1
MAX-CUT(G) −
We want to achieve that this yields a (1 + ε)-approximation, i.e.
2
2
4(2 − β)
ε
1
1
− ε′ ·
−
1
=
1
−
,
>
τ (ε)
(β − 1)2 xβ−1
1+ε
1+ε
which is equivalent to
2
2
4(2 − β)
1
ε
+ ε′ ·
−
1
6
2
β−1
τ (ε)
(β − 1) x
1+ε
1−
We achieve this in a two-step approach: First we define the function τ in such a way that
ε
2
6 21 · 1+ε
, namely τ (ε) = 4(1+ε)
. Then we choose ε′ appropriately, namely such that
τ (ε)
ε
ε′ ·
4(2 − β)
1
−1
2
β−1
(β − 1) x
2
6
ε
1
·
, or equivalently
2 1+ε
1
2
ε′ 6
4(2−β)
(β−1)2
·
ε
1+ε
2τ (ε)
2−β
β−1
−1
2
Altogether we obtain the following result.
Theorem 4. For every fixed 1 < β < 2, there is a PTAS for MAX-CUT in (α, β)-Power
Law Graphs.
Now we consider the case when the power law exponent is β = 2. In this case, the
number of nodes is still linear in eα , but now the number of edges drops down to n·log(n).
More precisely, the number of edges of an (α, 2)-PLG is asymptotically equal to 41 αeα ,
while the number of nodes is ζ(β)eα.
4.3
The Functional Case βf = 2 −
1
f (α)
We have shown in the previous sections that MAX-CUT on PLGs admits a PTAS for
every fixed β < 2. We consider now the functional case when βf = 2 − 1/f (α), where f (α)
is a monotone increasing function with f (α) −→ ∞ as α → ∞. In this section we will
show the following result.
8
Theorem 5. For βf = 2 − 1/f (α) with f (α) −→ ∞ as α → ∞ and f (α) = o(α), there is
a PTAS for MAX-CUT in (α, βf )-PLGs.
The proof of the Theorem is based on the following observation. It is sufficient to
show that we can split a given (α, βf )-PLG G into two parts [1, x∆f ) and [x∆f , ∆f ] such
that the following two conditions are satisfied:
(1) |[x∆f , ∆f ]|2 = O(|E|),
(2) vol[1, x∆f , ] = o(|E|).
Before we give the proof of the Theorem, we have to provide precise estimates for the
sizes and volumes of these node degree intervals. The maximum degree is ∆f = ⌊eα/βf ⌋.
The number of vertices is
∆f
X
j=1
$
eα
j βf
%
∆f
X
∆f
X
eα
eα
∈
−
∆
,
f
βf
βf
j=1 j
j=1 j
This sum can be approximated by the associated integral:
∆f
X
eα
∈
β
f
j=1 j
e
α/βf
Z
+1
1
⊆ eα ·
"
e
eα
dz,
z βf
α/βf
Z
+1
1
−1+1/f (α)
z
−1 + 1/f (α)
eα
eα
eα
dz
+
−
β
β
α/β
β
f
f
f
f
z
1
(⌊e
⌋ + 1)
#eα/βf +1
1
1
eα
1−
⊆
f (α)−1
1
1 − /f (α)
eα· 2f (α)−1
Similarly we obtain:
"
[x∆f , ∆f ] =
∆f
X
x∆f
$
eα
j βf
with
α
e
X∆ f
j=x∆ βf
j
∈
e
α/βf
xe
Z
+1
α/βf
%
∈
∆f
X
e
dz,
z βf
e
α/βf
xe
Z
!
−1+1/f (α)
z
−1 + 1/f (α)
#eα/βf +1
1
+1
α/βf
α
α
!
,
+ (1 − o(1))eα
1
eα
1−
,
f (α)−1
1
1 − /f (α)
eα· 2f (α)−1
1
eα
−
∆
f 1−x+
β
j f
∆f
j=x∆f
α
, eα ·
"
!
∆f
X
j=x∆f
α
+ (1 − o(1))e
#
eα
,
j βf
α
e
e
e
dz
+
−
β
β
α/β
α/β
z f
(xe f ) f
(⌊e f ⌋ + 1)βf
eα
1
1
,
⊆
−
(
1
1
1
1 − f (α) (xeα/βf )1− f (α)
(eα/βf + 1)1− f (α)
eα
1
eα
1
eα
+
−
−
1
1
1
1 − f (α)
(xeα/βf )βf
(eα/βf + 1)βf
(xeα/βf )1− f (α)
(eα/βf + 1)1− f (α)
"
!
f (α)
f (α)
1
α· 2f (α)−1
⊆
·
·e
,
f (α)−1 − 1
f (α) − 1
f
(α)
x
!
f (α)
1
f (α)
1
1
1
· eα· 2f (α)−1 ·
+ 2f (α)−1 − 2f (α)−1
f (α)−1 −
2f (α)−1
f (α) − 1
x f (α)
2 f (α)
x f (α)
2 f (α)
Thus we obtain the following estimates for sizes of node degree intervals.
9
#
Lemma 6. Let G = (V, E) be an (α, βf )-PLG with βf = 2 −
0 < x < 1, the size of the node degree interval [x∆f , ∆f ] satisfies
1
.
f (α)
Then for every
!
"
!
f (α)
1
f (α)
1
− ∆f 1 − x +
|[x∆f , ∆f ]| ∈
,
· eα· 2f (α)−1 ·
f (α)−1 − 1
f (α) − 1
∆f
x f (α)
!
#
f (α)
1
1
1
f (α)
1
α· 2f (α)−1
+ 2f (α)−1 − 2f (α)−1
·
·e
f (α)−1 −
2f (α)−1
f (α) − 1
x f (α)
2 f (α)
x f (α)
2 f (α)
Moreover,
eα
1
|V | ∈
1−
f (α)−1
1
α·
1 − /f (α)
e 2f (α)−1
"
!
eα
1
− ∆f ,
1−
f (α)−1
1
α·
1 − /f (α)
e 2f (α)−1
!
α
+ (1 − o(1))e
Corollary. The number of nodes of an (α, βf )-PLG G = (V, E) with βf = 2 −
satisfies
"
f (α)
2f (α) − 1 α
|V | ∈ (1 − o(1) ·
· eα , (1 − o(1) ·
·e
f (α) − 1
f (α) − 1
#
1
f (α)
#
Now we will estimate volumes of node degree intervals. Given some x ∈ (0, 1), possibly
depending on α, we have
vol[x∆f , ∆f ] ∈
∆Zf +1
x∆f
eα
z
∆Zf +1
x∆f
⊆
α
e
z
1−
1−
eα
z
f (α)−1
f (α)
f (α)−1
f (α)
1−
1−
f (α)−1
f (α)
2f (α)
α· 2f (α)−1
⊆ f (α)e
α
e
f (α)−1
f (α)
z
α
e
dz +
(x∆f )
f (α)−1
f (α)
e
−
(∆f + 1)
f (α)−1
f (α)
2α·f (α)
− (1 − x)e 2f (α)−1 ,
x∆f
∆f +1
f (α)−1
+
e
x
x∆f
!1/f (α)
1
1+
∆f
−
f (α)−1
f (α)
f (α)−1
α· 2f (α)−1
α· 2f (α)−1
1
1+
∆f
2f (α)
α· 2f (α)−1
f (α)e
α
∆f +1
f (α)−1
f (α)
dz − (1 − x)∆2f ,
f (α)−1
f (α)
e
2
f (α)−1
f (α)
2α·f (α)
− x /f (α) − (1 − x)e 2f (α)−1 ,
!1/f (α)
1
f (α)−1
α· 2f (α)−1
− x /f (α) + e
10
1
·
x
1
f (α)−1
f (α)
1
−
2
f (α)−1
f (α)
!
Similarly we obtain
f (α)−1
x∆f +1
z 1− f (α)
vol[1, x∆f ] ∈ eα
f (α)−1
2α·f (α)
2f (α)−1 ,
− (x − ∆−1
f )e
1
eα
1−
f (α)−1 x∆f +1
z f (α)
f (α)−1
1
f (α)−1
α· 2f (α)−1
+1−
e
x
1
α
f (α)−1
f (α)
2α·f (α)
⊆ f (α)eα x f (α) e 2f (α) − 1 − xe 2f (α)−1 ,
f (α)eα 2 /f (α) x
1
1
f (α)
e
f (α)−1
α
2f (α)
−1 +1−
x
Thus the number of edges of an (α, βf )-PLG is
"
|E| ∈ f (α)eα (∆f + 1)
α
f (α)e
1
f (α)
(∆f + 1)
−1
1
f (α)
1
− 1−
∆f
−1
!
f (α)−1
α· 2f (α)−1
+e
2f (α)
2f (α)
2f (α)
· ∆f
2f (α)−2
2f (α)
2f (α)
,
f (α)−1
f (α)
⊆ (f (α) − 1)eα· 2f (α)−1 , f (α)eα· 2f (α)−1 + eα· 2f (α)−1
⊆ (f (α) − 1)eα· 2f (α)−1 , (f (α) + 1)eα· 2f (α)−1
f (α)−1
f (α)
α· 2f (α)−1
e
eα· 2f (α)−1
1
−
2
f (α)−1
f (α)
!#
Concerning (2):
Based on these estimates, we will now show how to choose the parameter x such as to
satisfy both conditions (1) and (2). It turns out that this depends on the order of growth
of the function f (α), We observe that condition (2) is equivalent to
2f (α)
2f (α)
x /f (α) eα· 2f (α)−1 − eα = o eα· 2f (α)−1 − eα
1
2f (α)
α· 2f (α)−1
We have e
1
α· 2f (α)−1
− eα = eα · e
−1
= : Tf,α . Now we may consider three cases:
(a) If f (α) = o(α), then Tf,α = ω(1),
(b) If f (α) = Θ(α), then Tf,α = Θ (eα ),
(b) If f (α) = ω(α), then Tf,α = o (eα ).
We consider the case (a). Then, in order to satisfy the condition (2), we have to choose
x such that x1/f (α) = o(1) and moreover,
1
6 x
f (α)
eα· 2f (α)−1
Now we consider the requirement (1). We observe that, up to constant factors, the
condition in (1) is equivalent to
f (α)
f (α) − 1
!2
2f (α)
α· 2f (α)−1
e
1
·
x
f (α)−1
f (α)
−1
!2
11
2f (α)
α· 2f (α)−1
6 f (α) · e
α
−e
,
which is, by rearranging terms, equivalent to
f (α)
f (α) − 1
Since
f (α)
f (α)−1
α
·e ·e
α
2f (α)−1
1
·
x
f (α)−1
f (α)
−1
!2
α
6 f (α)eα e 2f (α)−1 − 1
= Θ(1), this is, again up to constant factors, equivalent to
1
x
!2
f (α)−1
f (α)
−1
!2
6
(f (α) − 1)2
,
f (α)
which yields the requirement
f (α)
x >
f (α) 2(f (α)−1)
((1 + o(1))f (α))
f (α)
f (α)
f (α)−1
= (1 − o(1)) · f (α) 2(f (α)−1)
f (α)
− f (α)−1
= f (α)
f (α)
− f (α)−1
Thus we obtain that for x = 1/f (α), requirements (1) and (2) are satisfied, which concludes
the proof of the theorem.
5
The Case β > 2
The situation drastically changes when the power law exponent is a constant β > 2. In
this case, the number of nodes is |V | = ζ(β)eα, and the number of edges is 21 ζ(β − 1)eα .
We apply the Goemans-Williamson algorithm to the graph G[2,∆] induced by the vertices
of degree at least 2 in G, and afterwards place all the edges incident to degree-1 nodes in
the cut. This yields a cut which has an expected inverse approximation ratio at least
αGW · ( 21 ζ(β − 1) − 1) + 12
αGW · ( 12 ζ(β − 1) − 1)eα + 12 eα
=
,
1
1
ζ(β − 1)eα
ζ(β − 1)
2
2
where αGW ≈ 0.879 denotes the inverse approximation ratio of the Goemans-Williamson
algorithm. This analysis can be refined as follows. There are eα nodes of degree 1.
Suppose there are µ · eα nodes of degree 1 which are incident to another degree-1 node.
The remaining (1 − µ)eα degree-1 nodes are incident to nodes of higher degree. The
resulting lower bound on the expected inverse approximation ratio is then
αGW · ( 12 ζ(β − 1) − (1 − µ2 )) + (1 − µ2 )
αGW · ( 21 ζ(β − 1) − 21 ) + 21
>
,
1
1
ζ(β − 1)
ζ(β − 1)
2
2
where the lower bound is attained at µ = 1.
6
Approximation Lower Bounds for β > 2
In this section we provide explicit approximation lower bounds for MAX-CUT in (α, β)PLG for β > 2. Explicit lower bounds for the approximability of MAX-CUT in degree
B bounded graphs and B-regular graphs have been obtained in a series of papers [BK99;
12
BK01]. The authors showed, among other results, that the MAX-CUT problem restricted
to 3-regular graphs is NP-hard to approximate to within (152/151) + ε.
Here we will make use of this result and the associated constructions in order to prove
APX-hardness for MAX-CUT in (α, β)-PLG for β > 2. Moreover we will show that even
for β ∈ (0, 2], the problem remains NP-hard in the exact setting. It turns out that for
β > 1 this will be a direct consequence from our reduction for the case β > 2, while in
the case β 6 1 a different construction is needed. This is due to the fact that for β > 1,
the number of constant degree nodes in an (α, β)-PLG is linear in the total number of
vertices, while for β 6 1 this is not true anymore.
The section is organized as follows. In the next paragraph we will describe a generic
construction which reduces the MAX-CUT problem in 3-regular graphs to MAX-CUT in
(α, β)-PLGs, based on an embedding of the former graphs into the later ones. Afterwards
we will use this construction such as to obtain the APX-hardness of MAX-CUT in PLGs
for β > 2, with an explicit approximation lower bound that only depends on the power
law exponent β. This also yields the NP-hardness for β ∈ (1, 2]. Finally we will describe
in subsection 6.1 a different embedding construction which yields the NP-hardness in the
exact setting for β 6 1.
The Construction We will now describe an embedding of 3-regular graphs into power
law graphs. Starting from an instance G of E3-MAX-CUT, we construct an (α, β)-PLG
instance G′ = G ∪ W ∪ M which is the disjoint union of G, a matching M on a subset of
the degree-1 nodes and a subgraph W = W2 ∪ W4 ∪ . . . ∪ W∆ , where ∆ = ⌊eα/β ⌋ is the
maximum degree of G. For each i 6∈ {1, 3}, Wi is a multipath of degree i nodes, namely
a path whose edges have multiplicities alternating between ⌊ 2i ⌋ and ⌈ 2i ⌉, with degree 1
nodes attached to the two endpoints of this path. Wi contains all the degree i nodes of
G and additionally up to i degree-1 nodes.
Given the graph G with N vertices, we choose α smallest possible such that N is not
larger than the number of degree 3 vertices in an (α, β)-PLG, namely such that
eα
N= β
3
∈
eα
eα
−
1,
3β
3β
α
j
α
k
Without loss of generality, we assume that N = 3eβ = 3eβ . In the construction of the
subgraphs Wi , we have to distinguish the following four cases.
α
α
Case 1: i is even and ⌊ eiβ ⌋ is even. Then Wi contains vertices vi,j , j = 1, . . . ni = ⌊ eiβ ⌋ of
degree i and additionally i vertices of degree 1. For 1 < j < ni , vertex vi,j is adjacent to
vi,j−1 and vi,j+1 by multi-edges of multiplicity i/2. vi,1 and vi,ni are adjacent to vi,2 and
vi,ni −1 respectively, each by multi-edges with multiplicity i/2. Additionally, each of these
two nodes is adjacent to i/2 nodes of degree 1.
i
2
i
2
i
2
α
Case 2: i is even and ⌊ eiβ ⌋ is odd. The construction is the same as in case 1.
13
α
α
Case 3: i is odd and ⌊ eiβ ⌋ is even. Again Wi has vertices vi,j , j = 1, . . . ni = ⌊ eiβ ⌋ of degree
i and now additionally i − 1 vertices of degree 1. For j = 3, . . . , ni , vertex vi,j is adjacent
to vi,j−1 with multiplicity ⌈ 2i ⌉, and for j = 1, 3, . . . ni − 2 to vi,j+1 with multiplicity ⌊ 2i ⌋.
Additionally vi,1 has ⌊ 2i ⌋ neighbors of degree 1 and vi,ni has ⌊ 2i ⌋ degree 1 neighbors. Thus
Wi contains i − 1 nodes of degree 1.
⌊ 2i ⌋
⌊ 2i ⌋
⌊ 2i ⌋
⌈ 2i ⌉
α
Case 4: i is odd and ⌊ eiβ ⌋ is odd. In this case, vi,1 is adjacent to ⌈ 2i ⌉ degree 1 nodes and
to vi,2 with multiplicity ⌊ 2i ⌋. The vertex vi,ni has ⌊ 2i ⌋ degree 1 neighbors and is adjacent
to vi,ni −1 with multiplicity ⌈ 2i ⌉. Thus Wi contains ⌊ 2i ⌋ + ⌈ 2i ⌉ = i nodes of degree 1.
⌈ 2i ⌉
⌊ 2i ⌋
⌊ 2i ⌋
⌈ 2i ⌉
α
2
α
Finally M is a perfect matching on ⌊eα ⌋ − ∆
i=4 i + O(∆) = e − Θ(∆ ) = (1 − o(1))e
vertices of degree 1.
Our approximation lower bounds for MAX-CUT in (α, β)-PLG will be based on the
fact that a maximum cut in the subgraph W contains all the edges of W and can be
constructed efficiently.
P
Lemma 7. Given a 3-regular graph G with N vertices, let W ∪M be the auxiliary subgraph
described above. Then W ∪ M has the following properties.
(a) The number of vertices is |V (W ∪ M)| = (1 − o(1)) ζ(β) − 3−β eα .
(b) The number of edges is |E(W ∪ M)| = (1 − o(1))
1
ζ(β
2
− 1) −
1
2·3β−1
eα .
(c) The cut size is MAX-CUT(W ∪M) = |E(W ∪M)|, and such a cut can be constructed
in polynomial time.
Proof. Parts (a) and (b) follow directly from the construction. Concerning (c), we place
for each Wi vertices vi,j on the left hand side of the cut if j is odd and on the right hand
side if j is even. Then we place all the degree 1 nodes optimally. This concludes the
proof of the lemma.
Explicit Lower Bounds for β > 2 We start from the following approximation hardness result for MAX-CUT in 3-regular graphs.
Theorem 8 ([BK01]). For every ε ∈ (0, 1/302), it is NP-hard to decide whether an instance
of E3-MAX-CUT with 156n edges and 104n vertices has a maximum cut of size above
(152 − ε)n or below (151 + ε)n. Thus the MAX-CUT problem in 3-regular graphs is
NP-hard to approximate within any ratio (152/151) − ε.
14
We consider now such a 3-regular graph G with N = 104n vertices and construct
the associated
(α, β)-PLG
G′ = G ∪ W ∪ M. Due to Lemma 7, the graph G′ contains
(1 − o(1)) ζ(β) − 3−β · 3β · N + N vertices, and the auxiliary subgraph W ∪ M contains
(1 − o(1))
1
ζ(β
2
− 1) −
1
2·3β−1
· 3β · N edges. It is now NP-hard to decide if
MAX-CUT(G′ ) > (1 − o(1))
1
1
ζ(β − 1) −
· 3β · 104n + (152 − ε)n
2
2 · 3β−1
or
1
1
MAX-CUT(G ) 6 (1 − o(1)) ζ(β − 1) −
· 3β · 104n + (151 + ε)n
2
2 · 3β−1
′
Simplifying terms, we obtain the following result.
Theorem 9. For every β > 2 and ε > 0, MAX-CUT in (α, β)-Power Law Graphs is
β ζ(β−1)−3)·52+152
− ε.
NP-hard to approximate within a ratio (3
(3β ζ(β−1)−3)·52+151
As a byproduct of the proof of the previous theorem we also obtain the following
result.
Corollary. For 1 < β 6 2, MAX-CUT in (α, β)-Power Law Graphs is NP-hard.
Proof. Theorem 8 yields that the following decision problem is NP-complete: Given a
3-regular graph G with 104n vertices, is MAX-CUT(G) > (152 − ε)n? For β > 1, our
reduction G 7→ G′ = G ∪ W ∪ M is well defined and reduces this to the decision problem
P
eα
α
β
if MAX-CUT(G′ ) > (152 − ε)n + 21 ∆
j=1 j · ⌊ j β ⌋ − 156n, where e = 3 · 104n.
Remark. The reduction is not well-defined anymore for β 6 1, since in that case the
number of degree 1 nodes does not suffice to construct the subgraphs Wi , i = 2, 4, 5, . . . , ∆
in the way as described before. In the next subsection we will provide an alternative
reduction which also yields the NP-hardness for β 6 1.
6.1
NP-Hardness for β 6 1
In order to prove NP-hardness of the MAX-CUT problem in (α, β)-power law graphs for
β 6 1, we construct again a polynomial time reduction from the 3-regular MAX-CUT.
We consider first the case β < 1. Then the number of degree 1 nodes is still equal to ⌊eα ⌋,
α/β
while the total number of nodes is (1 − o(1)) e1−β = ω(eα ). Thus it is even not possible to
spend one degree 1 node per subgraph Wi , since we also have eα = o(∆).
We propose the following alternative construction. Starting from a 3-regular graph G
α
α
with N = 104n vertices, we choose again α such that N = ⌊ 3eβ ⌋ = 3eβ . Now we call an
α
integer i ∈ {4, . . . ∆} critical if both i and ⌊ eiβ ⌋ are odd.
For those node degrees i ∈ {2, 4, 5, . . . , ∆} which are non-critical, we let Wi be a
wheel, i.e. a cycle consisting of multi-edges containing all the degree i nodes from G′ and
having edge multiplicities alternating between ⌊ 2i ⌋ and ⌈ 2i ⌉. In that case we have
MAX-CUT(Wi ) =
1
· i · ni
2
j k
ni ·i
i
−
2
2
if i is even,
if i is odd and ni even.
15
Now we deal with the critical node degrees. Suppose that i1 < i2 < . . . < ic are the critical
degrees in {2, 4, 5, . . . , ∆}. We take a maximum matching Mc on this set of indices such
that without loss of generality, index ic is unmatched iff c is odd. For each pair i, j in
this matching, we construct a subgraph Wi,j containing ni nodes of degree i and nj nodes
of degree j. This subgraph is constructed as follows. We arrange all the degree i nodes
on a cycle consisting of multi-edges with multiplicities alternating between ⌊ 2i ⌋ and ⌈ 2i ⌉.
Let vi be the node which has now degree i − 1 = 2 · ⌊ 2i ⌋. All the other nodes on this
cycle have already degree i. Now we do the same for the nj nodes which are supposed to
have degree j, and we define the vertex vj accordingly. Finally we add a single edge of
multiplicity 1 connecting vi and vj . In Figure 2 we show the construction of Wi,j .
⌈ 2j ⌉
⌊ 2j ⌋
⌊ 2j ⌋
critical node
⌊ 2j ⌋
Wj
edge of multiplicity 1
Wi
⌊ 2i ⌋ critical node
⌊ 2i ⌋
⌈ 2i ⌉
⌊ 2i ⌋
Figure 2: The construction of Wi,j .
Directly from this construction we obtain that the subgraph Wi,j contains ni nodes of
degree i and nj nodes of degree j. Note that from the very definition of critical degrees
it follows that both ni and nj are odd. Every cut in Wi,j misses at least one multi-edge
inside the degree i nodes and one such edge inside the degree j nodes. On the other hand,
we can easily construct such a cut which misses only one edge of multiplicity ⌊ 2i ⌋ and one
edge of multiplicity ⌊ 2j ⌋, by placing the degree i nodes alternatingly on the left and right
hand side of the cut and doing the same for the degree j vertices. Thus we have
i · ni
i
MAX-CUT(Wi,j ) =
−
2
2
j · nj
j
+
−
2
2
+1
Finally if c is odd, then Wic contains all the degree-ic nodes and one node of degree 1,
and the maximum cut size is MAX-CUT(Wic ) = ic ·n2 ic − ⌊ 2i ⌋ + 1.
In order to keep notation simple, we let Wic = ∅ in case when the number of critical
node degrees is even. Moreover we let J ⊂ {2, 4, . . . , ∆} be the set of critical degrees.
S
S
Thus the map G 7→ G′ = G ∪ i6∈J Wi ∪ {i,j}∈Mc Wi,j ∪ Wic ∪ M reduces the question if
a given 3-regular graph G with 104n vertices has a cut of size at least (152 − ε)n to the
question if the (α, β)-power law graph G′ has a cut of size at least
16
(152 − ε)n +
P
i∈([2,∆]\J)∪{ic }
MAX-CUT(Wi ) +
P
{i,j}∈Mc
MAX-CUT(Wi,j ) + MAX-CUT(M)
The same construction also works in the case when β = 1. We obtain the following result.
Theorem 10. For every β ∈ (0, 1], the MAX-CUT problem in (α, β)-Power Law Graphs
is NP-hard.
7
Further Research
We prove some new results on approximability of MAX-CUT in Power Law Graphs.
It remains an open problem to settle the status of MAX-CUT in PLGs for the power
law exponent β = 2, where the phase transition happens from existence of a PTAS
(for constant β < 2 and for functional β slowly converging to 2 from below) to APXhardness (for any constant β > 2). Another problem concerns the design of better
constant factor approximation algorithms for MAX-CUT in the case β > 2, based on SDP
simulations with appropriate classes of inequalities for the low-degree vertices in the power
law graph. Similar methods can be applied to other partition problems on power law
graphs like MAX-BISECTION, Multiway-CUT and k-partition problems. Establishing
good approximability bounds for those problems is another interesting question.
References
[ACL01]
W. Aiello, F. Chung, and L. Lu. “A Random Graph Model for Power Law
Graphs”. Experimental Mathematics 10.1 (2001), pp. 53–66.
[AKK95]
S. Arora, D. Karger, and M. Karpinski. “Polynomial Time Approximation
Schemes for Dense Instances of NP-hard Problems”. Proceedings of the 27th
annual ACM Symposium on Theory of Computing (STOC ’95); also in Journal of Computer and System Sciences 58.1 (02/1999). 05/1995, pp. 284–293.
[Bar+88]
F. Barahona, M. Grötschel, M. Jünger, and G. Reinelt. “An Application of
Combinatorial Optimization to Statistical Physics and Circuit Layout Design”. Operations Research 36.3 (06/1988), pp. 493–513.
[BGS98]
M. Bellare, O. Goldreich, and M. Sudan. “Free bits, PCPs and nonapproximability – towards tight results”. SIAM Journal on Computing 27.3 (06/1998),
pp. 804–915.
[BK99]
P. Berman and M. Karpinski. “On some tighter inapproximability results”.
Proceedings of the 26th International Conference on Automata, Languages
and Programming (ICALP ’99). Vol. 1644. LNCS. 1999, pp. 200–209.
[BK01]
P. Berman and M. Karpinski. “Efficient amplifiers and bounded degree optimization”. ECCC TR01-053 (2001).
[Bjö05]
A. Björklund. “Approximating integer quadratic programs and MAXCUT in
subdense graphs”. Proceedings of the 13th annual European Conference on
Algorithms (ESA ’05). Vol. 3669. LNCS. 10/2005, pp. 839–849.
17
[CSN09]
A. Clauset, C. R. Shalizi, and M. E. J. Newman. “Power-Law Distributions in
Empirical Data”. SIAM Review 51.4 (11/2009), p. 661. arXiv:0706.1062v2.
[DT13]
T. N. Dinh and M. T. Thai. “Community detection in scale-free networks:
Approximation algorithms for maximizing modularity”. IEEE Journal on Selected Areas in Communications 31.6 (06/2013), pp. 997–1006.
[FKK04]
W. Fernandez de la Vega, M. Karpinski, and C. Kenyon. “Approximation
schemes for Metric Bisection and partitioning”. Proceedings of the 15th annual ACM-SIAM Symposium on Discrete Algorithms (SODA ’04). 01/2004,
pp. 506–515.
[Fer+05]
W. Fernandez de la Vega, M. Karpinski, R. Kannan, and S. Vempala. “Tensor
decomposition and approximation schemes for constraint satisfaction problems”. Proceedings of the 37th annual ACM Symposium on Theory of Computing (STOC ’05). 05/2005, p. 747.
[FK00]
W. Fernandez de la Vega and M. Karpinski. “Polynomial time approximation
of dense weighted instances of MAX-CUT”. Random Structures and Algorithms 16.4 (07/2000), pp. 314–332.
[FK01]
W. Fernandez de la Vega and C. Kenyon. “A Randomized Approximation
Scheme for Metric MAX-CUT”. Journal of Computer and System Sciences
63.4 (12/2001), pp. 531–541.
[GW95]
M. X. Goemans and D. P. Williamson. “Improved approximation algorithms
for maximum cut and satisfiability problems using semidefinite programming”. Journal of the ACM 42.6 (11/1995), pp. 1115–1145.
[Had75]
F. Hadlock. “Finding a Maximum Cut of a Planar Graph in a polyminal
time”. SIAM Journal on Computing 4.3 (09/1975), pp. 221–225.
[Hås01]
J. Håstad. “Some optimal inapproximability results”. Journal of the ACM
48.4 (07/2001), pp. 798–859.
[Kar02]
M. Karpinski. “Approximability of the Minimum Bisection Problem: An Algorithmic Challenge”. Proceedings of the 27th International Symposium on
Mathematical Foundations of Computer Science (MFCS ’02). Vol. 2420. LNCS.
08/2002, pp. 59–67.
[KLS13]
M. Karpinski, A. Lingas, and D. Sledneu. “Optimal cuts and partitions in tree
metrics in polynomial time”. Information Processing Letters 113.12 (06/2013),
pp. 447–451. arXiv:arXiv:1212.3471v1.
[Kar72]
R. M. Karp. “Reducibility among Combinatorial Problems”. Complexity of
Computer Computations. 1972, pp. 85–103.
[Kho+07] S. Khot, G. Kindler, E. Mossel, and R. O’Donnell. “Optimal inapproximability results for MAX-CUT and other 2-variable CSPs?” SIAM Journal on
Computing 37.1 (01/2007), pp. 319–357.
[New06]
M. E. J. Newman. “Modularity and community structure in networks”. Proceedings of the National Academy of Sciences 103.23 (02/2006), pp. 8577–82.
arXiv:0602124 [physics].
18
[NG04]
M. E. J. Newman and M. Girvan. “Finding and evaluating community structure in networks”. Physical Review E 69.2 (02/2004), p. 026113.
[PY91]
C. H. Papadimitriou and M. Yannakakis. “Optimization, approximation, and
complexity classes”. Journal of Computer and System Sciences 43.3 (12/1991),
pp. 425–440.
[SG76]
S. Sahni and T. Gonzalez. “P-Complete Approximation Problems”. Journal
of the ACM 23.3 (07/1976), pp. 555–565.
[Tre+00]
L. Trevisan, G. B. Sorkin, M. Sudan, and D. P. Williamson. “Gadgets, Approximation, and Linear Programming”. SIAM Journal on Computing 29.6
(01/2000), pp. 2074–2097.
[WGH04]
A. Y. Wu, M. Garland, and J. Han. “Mining scale-free networks using geodesic
clustering”. Proceedings of the ACM SIGKDD International Conference on
Knowledge Discovery and Data Mining (KDD ’04). 08/2004, p. 719.
19
| 8cs.DS
|
Application Embedding: A Language Approach to Declarative Web
Programming
*
David H. Lorenza,b and Boaz Rosenanc
a Open University of Israel
b Faculty of Computer Science, Technion—Israel Institute of Technology
c
University of Haifa
Abstract
Since the early days of the Web, web application developers have aspired to develop much of
their applications declaratively. However, one aspect of the application, namely its business-logic is constantly
left imperative. In this work we present Application Embedding, a novel approach to application development
which allows all aspects of an application, including its business-logic, to be programmed declaratively.
We develop this approach in a two-step process. First, we draw a mapping between web applications
and Domain-Specific Languages (DSLs). Second, we note that out of the two methods for implementing DSLs,
namely as either internal or external, most traditional web applications correspond to external DSLs, while the
the technique that corresponds to DSL embedding (implementing internal DSLs) is left mostly unexplored.
By projecting the well-known technique of DSL embedding onto web applications, we derive a novel
technique—Application Embedding. Application embedding offers a separation of code assets that encourages
reuse of imperative code, while keeping all application-specific assets, including those specifying its businesslogic, declarative.
As validation, we implemented a simple, though nontrivial web application using the proposed separation
of assets. This implementation includes an application-agnostic imperative host application named FishTank,
intended to be applicable for a wide variety of web applications, and a declarative definition of the different
aspects of the specific application, intended to be loaded on that host.
Our method of separation of code assets facilitates a better separation of work, in comparison to traditional
methods. By this separation, host application developers can focus mostly on the extra-functional aspects of
a web application, namely on improving performance, scalability, and availability, while developers of an
embedded application can focus on the functional aspects of their application, without worrying about extrafunctional concerns. The reusability of the host application makes the effort put into a better implementation
cost-effective, since it can benefit all applications built on top of it.
ACM CCS 2012
Software and its engineering → Software development techniques; Development frameworks and
environments;
Keywords Domain-specific languages (DSLs), Logic programming (LP), Web application
The Art, Science, and Engineering of Programming
Submitted
September 1, 2016
Published
January 27, 2017
doi
10.22152/programming-journal.org/2017/1/2
© David H. Lorenz and Boaz Rosenan
This work is licensed under a “CC BY 4.0” license.
In The Art, Science, and Engineering of Programming, vol. 1, no. 1, 2017, article 2; 38 pages.
Application Embedding: A Language Approach to Declarative Web Programming
1
Introduction
Web application development has gone a long way to make much of the application
code declarative. This includes the introduction of server-side frameworks (e.g., Ruby
on Rails 1 and Django 2 ), client-side frameworks (e.g., AngularJS 3 and React 4 ), and
languages backed by web standards (e.g., HTML and CSS). While these frameworks
and languages make declarative programming applicable to many of these applications’
concerns (most notably, the presentation), one important aspect of web applications
is constantly left imperative: their business logic. 5
In this paper we apply a well-known principle in Computer Science—the notion that
program (code) and state (data) are interchangeable [38, 4]—to another well known
technique—domain-specific language embedding [23, 21]—to form a new approach to
application development, named Application Embedding, allowing application business
logic to be defined declaratively rather than implemented imperatively.
Applying declarative programming, which is stateless by nature, to business logic
implementation, which is imperative by nature, may seem at first sight implausible or
even impossible. Indeed, imperative, general-purpose programming languages are
often used to implement business logic. However, if we take a closer look, imperative
business logic can be attributed to a common but inessential design choice: the ThreeTier Architecture (3TA) [13].
1.1 Traditional Approach
In the 3TA a web application is partitioned into a presentation-tier running the presentation logic on the client-side, a logic-tier running the business logic typically on a
web server on the sever-side, and a data-tier consisting of a database program also
running on the server-side.
The 3TA provides a separation of concerns that promotes scalability. Stateless
components are scaled easily by adding more computer nodes running them. Stateful
components, however, are harder to scale. For this reason, the 3TA segregates the
latter parts in their own tier (namely, the data-tier), requiring them to be application
agnostic (e.g., a general-purpose database), so once implemented, they can be reused
massively across applications, thus amortizing the cost of their development.
*
1
Work done in part while visiting the Faculty of Computer Science, Technion—Israel Institute
of Technology.
http://rubyonrails.org/
https://www.djangoproject.com/
3
https://angularjs.org/
4
https://facebook.github.io/react/
2
5
The term business logic in web applications refers to the part of the application that manipulates the data (state) that the application stores, and it is used in a broader sense than
its original meaning in enterprise applications where it refers strictly to business rules or
business processes. For example, a decision in a social network regarding the placement
of a user status in another user’s timeline is a part of the social network’s business logic,
although this decision has nothing to do with business per se.
2-2
D. H. Lorenz and B. Rosenan
Consequently, however, the 3TA separates computation and state into two separate
entities (the logic-tier and the data-tier, respectively). This means that the logic-tier
needs to perform I/O operations to interact with the application’s state, forcing its
imperative nature. Some web frameworks and multi-tier programming languages
strives to hide away the imperative nature of the logic-tier, thus allowing developers to
define most of the application’s concerns declaratively. However, they do so typically
by relying on the declarative nature of relational databases, thus assuming the application’s data model maps nicely to a normalized relational database. Data models
that are hard to fit into relational data (e.g., due to impedance mismatch [35]), or
large systems that require denormalization to meet performance goals, break these
assumptions and require imperative code to fill the gap.
1.2 A Domain-Specific Language Approach
We present a different separation of concerns. Our method of partitioning keeps all
application-specific parts declarative, while keeping all imperative parts application
agnostic, and thus reusable across different applications. For achieving this we turn to
another domain, namely, Domain-Specific Languages (DSLs).
DSLs can be implemented as either external DSLs, implemented as compilers or
interpreters, or internal [10, 34] or embedded [23, 21] DSLs, implemented as software
libraries in a host language, in a process sometimes referred to as DSL embedding.
For declarative host languages, embedded DSLs are defined declaratively [23]. This
results in a desired separation of code assets, where all DSL-specific code (namely the
DSL definition) is declarative, and the (often imperative) implementation of the host
language remains DSL-agnostic, and therefore reusable across DSLs.
We can apply DSL embedding to web applications by thinking of the state of
applications as programs in a languages. The equivalence between code and data is a
fundamental concept in Computer Science (applied, e.g., in Von Neumann’s storedprogram architecture [38] and in Church’s lambda calculus [4]). This equivalence
is often used to describe code as data (e.g., an encoding of a program), but in some
cases the opposite direction is useful as well. In our case, we note that the logic-tier
of an application following the 3TA acts as an interpreter, acting upon the “program”
stored in the data-tier as the state of the application. The database schema defines the
language this program is written in, while the business logic gives it meaning, acting
as its semantics. Since the concepts stored in the database refer to the application’s
problem domain, its corresponding language can be treated as a DSL.
By drawing this mapping between applications and DSLs we open the door to
comparing their implementation techniques. Indeed, state-of-the-art techniques for
web application development typically compare to external DSLs, while methods that
map to DSL embedding are seldom used for anything other than static websites.
1.3 Contribution
We present Application Embedding—a web application development approach based
on DSL embedding, in which a web application is embedded in a host application,
2-3
Application Embedding: A Language Approach to Declarative Web Programming
Figure 1
A comparison between (a) the logic-tier of an application and (b) an interpreter
just like a DSL is embedded in a host language. Just as in the field of DSLs the host
language is a general-purpose programming language, our host application is a generalpurpose application, fully programmable to host any embedded web application. As in
DSLs, this approach results in a desirable separation of code assets. The imperative
implementation of the host application is reusable, since it is general-purpose, while
all the application-specific code is declarative, consisting of definitions in the language
implemented by the host application.
As validation of this approach, we present the following:
A discussion of the mapping between web applications and DSLs, demonstrated
by providing specifications for a web application using formal semantics of the
corresponding DSL (section 2).
A discussion of the consequences of applying this mapping to embedded DSLs,
deriving requirements for host applications, needed to implement our application
embedding approach (section 3).
A computational model, named the Fish-Tank Model (FTM), that is based on Logic
Programming (LP) and meets these requirements, along with an implementation
of this model, name FishTank (section 4).
A purely declarative definition of a microblogging web application, embedded in
FishTank (section 5).
2
An Application as a Programming Language
Since the early days of Computer Science it is known that programs and their state are
interchangeable. Von Neumann’s stored-program architecture [38] requires programs
to be encoded as data. By this architecture, a program, like its state, is a part of the
state of the machine running it. Curry’s lambda calculus [4] goes beyond that, making
it possible to treat the state of any program as a program by itself.
While it may be confusing to think of state as a program in the general case, it is
much easier to see the correspondence in applications that follow the 3TA. Figure 1
compares a logic-tier of an application (figure 1a), to an interpreter (figure 1b) of a
programming language. Both logic-tiers and interpreters process input from a user to
2-4
D. H. Lorenz and B. Rosenan
(a)
(b)
(c)
Application: App
(Database) Schema
Business Logic
Specialpurpose
Presentation
State (Data)
Web App
Traditional
Logic-tier
Embedded
App
Embedding
Host App
Language: L 〈App〉
Abstract Syntax
Semantics
DSL
Concrete Syntax
Program
External DSL
External
Interpreter
Embedded DSL
Internal
Host Language
Table 1 Comparing Applications with Languages
produce output. Both do so according to their own programming: the logic tier follows
the application’s business logic and the interpreter follows the programming language’s
semantics. The programming is complemented by additional input: the application’s
state (stored in the data-tier) or the program being interpreted, respectively. In the case
of the application, however, state is also updated by the logic-tier, while interpreters
typically do not modify the program they run.
This comparison is summarized in table 1. For an application App we denote the corresponding language by L 〈App〉 (dubbed “App language”). The database schema used
by App provides the abstract syntax for L 〈App〉. App’s presentation (user interface)
gives human-readable and writable form to its state, acting as its concrete syntax. Most
applications are special-purpose, built around one specific problem domain (e.g., social
media, e-commerce). In such applications the state consists of concepts from that
problem domain (e.g., tweets, statuses, products). This makes their corresponding
languages domain-specific, i.e., DSLs (table 1a).
The rest of this section relates the different aspects of an application to the corresponding aspects of a DSL. To demonstrate this correspondence we describe the
behavior of an example web application named Tweetlog via a specification of its
corresponding L 〈Tweetlog〉 DSL. Tweetlog is a microblogging application, where
users can publish short textual tweets and follow other users. The application then
presents users with time-lines, containing tweets made by users they follow. 6
2.1 Schema as Abstract Syntax
For an application App, the schema of the data stored in App’s database corresponds
to the abstract syntax for L 〈App〉, where each program P ∈ L 〈App〉 corresponds to a
single state. It is common practice to treat the state of a database at a certain point in
time as a mathematical set, where each element in that set represents a single piece
of information (record, document, etc). We therefore treat P as a set of statements.
Therefore, the abstract syntax of these statements defines the structure of the state.
6
We provide only anecdotal definitions here. The complete specification can be found in
appendix A.
2-5
Application Embedding: A Language Approach to Declarative Web Programming
To define the schema of Tweetlog’s state we define the abstract syntax of its corresponding L 〈Tweetlog〉 DSL: 7
S
=
|
|
|
User =
Tweet =
|
...
tweeted (User, Time, Tweet)
follows (User, User, Time)
timeline (User, Time, TimelineElem)
...
user (String)
text (Text)
image (Binary, Text)
(1)
Statements in L 〈Tweetlog〉, denoted by the S symbol, are tweets (tweeted statements) and following relations (follows statements), which are both considered raw
statements, since users add these statements explicitly. The abstract syntax also accounts for derived statements, such as time-line entires (timeline statements), which
are calculated by the business logic. These statements may or may not be actually
stored by the application, but they can all be inferred based on the state of the
application, and thus we consider them part of the state.
2.2 Business Logic as Semantics
The business logic of an application App can be thought of as both the static and
dynamic semantics of L 〈App〉. The business logic serves two major roles in an application (which are not mutually exclusive): it mutates the state of the application and acts
upon it. In its former role, the business logic actually refines the set of allowed states,
beyond what the schema allows. This corresponds to the role of the static semantics
in a programming language, which refines the set of allowed programs beyond what
the abstract syntax allows. In its latter role, the business logic provides the meaning
of the application state. This corresponds to the role of dynamic semantics, which
provides meaning to programs.
Out of the two roles, we describe here only the dynamic semantics of L 〈Tweetlog〉.
For simplicity we assume that any set of syntactically-correct statements is a legal
program in L 〈Tweetlog〉, and thus a legal state in Tweetlog. Real-life applications
will guard against certain kinds of syntactically-valid state transitions to maintain
data integrity. A method for providing integrity, as well as confidentiality in similar
settings is provided elsewhere [28].
To define the dynamic semantics for L 〈Tweetlog〉, we use a system of
natural deduction rules as a way to reason about logical judgments relating programs
to their input and output. In Natural Semantics [25] judgments usually correspond to
typical behavior of general-purpose programming languages, such as state transition
of imperative languages or the evaluation of an expression to a value in functional
Notation
7
The complete abstract syntax specification for L 〈Tweetlog〉 can be found in figure 10.
2-6
D. H. Lorenz and B. Rosenan
languages. L 〈Tweetlog〉, however, is neither imperative nor functional. Our way of
reasoning upon it is by stating what statements S are true given program P , denoted
P |= S. First and foremost, a statement S is true given program P if it is included in P :
S∈P
P |= S
The following rules extend this set of statement with derived statements. In addition
to |=, we use ad-hoc judgments, which we denote as a first-order logic atom, using a
name and arguments.
Semantic Rules
We begin by tokenizing each tweet to produce a processed tweet
(procTweet):
P |= tweeted (U,τ, T )
parse tokens X 0 , X , "
replaceText T, plain (X ) , tokenized X 0 , T 0
P |= procTweet U, τ, T 0
proc-tweet
(2)
Here a processed tweet is based on a “raw” tweet, after extracting its text and parsing
(tokenizing) it. The parse judgment asserts that X can be tokenized into a list of
tokens X ’, and replaceText replaces the plaintext X with the list of tokens X ’ in the
body of the tweet. 8
This tokenization is then used to create a search index (searchIndex), indexing
tweets by their tokens (simple words, user IDs, and hashtags):
P |= procTweet (U, τ, T )
replaceText (T, tokenized (X ) , tokenized (X ) , T )
P |= searchIndex (ξ, τ, T )
ξ∈X
search1
(3)
Additionally, tweets are also indexed by the user making them:
P |= procTweet (U, τ, T )
search2
P |= searchIndex (userID (U) , τ, T )
(4)
The search index can be used for regular searches, but in particular it is used to
define the time-line:
P |= follows U1 , user (U2 ) , τ0
P |= searchIndex (userID (U2 ) , τ, T )
time-line1 (5)
P |= timeline (U1 , τ, tweet (U2 , T ))
By this, a user’s time-line for user U1 consists of all tweets related to any user U2
followed by U1 . This includes both tweets made by U2 as well as all replies to these
tweets, i.e., tweets in which user U2 is mentioned.
8
The definition of replaceText is given in equation (7) and equation (8). The definition of
parse is given in figure 12.
2-7
Application Embedding: A Language Approach to Declarative Web Programming
2.3 Presentation as Concrete Syntax
For an application App, the concrete syntax of L 〈App〉 defines how programs (state of
application App) is presented to the user. In programming, we are used to thinking of
concrete syntax in the other direction, defining how textual source code translates to
an abstract syntax tree. In contrast, here we consider concrete syntax as a bidirectional
mapping between the abstract and concrete representations of a program. Projectional
Editing (PE) [18, 39, 14], for example, is a technique by which concrete syntax defines
how an Abstract Syntax Tree (AST) is projected to a view displayed to the user. Apart
for displaying the AST, the view also facilitates editing of the AST. These edits are
made visible to the user by applying the projection again, to update the view.
In DSLs that represent web applications the meaning of concrete syntax is similar
to its meaning in projectional editing. The view in this case consists of HTML on a
web browser, and the projection translates the abstract representation (data in the
database) into its HTML representation. The view is also responsible for mutating the
abstract representation by allowing the creation of new data, and editing or deleting
data.
Recall that the state of the application consists of a set of statements. This means
that mutation of the state includes adding new statements, deleting statements,
and modifying statements. HTML provides forms to perform addition of new data
(statements) and editing existing data. Delete buttons can be associated with data to
be displayed to facilitate deletion.
2.4 State as Program
The mapping of the state of an application to a DSL program is somewhat counterintuitive at first. While programs are typically immutable and relatively small, the state
of an application is constantly mutated by the logic-tier and is typically very large.
In this section we elaborate on these differences and provide intuition as to which
languages map well to applications.
One thing to notice about L 〈App〉 is that
it is purely declarative for almost any App. Recall that program P ∈ L 〈App〉 is an
unordered set of statements. Such a set cannot represent a sequence of commands
to be executed (imperative code), unless individual statements do. Statements can
represent such sequences of commands in applications that allow users to build such
sequences, such as programming environments, where the application stores and
executes user programs. In any other case, we can refer to L 〈App〉 as a declarative
language.
State as a set implies a declarative program
Recall program P ∈ L 〈App〉 represents the
state of application App at a given moment. This makes P transient—applicable to
a single, passing moment. The facts the application’s logic-tier mutates the state
corresponds to the interpreter of L 〈App〉 mutating the program. This mutation is
depicted as a dashed line in figure 1b.
Mutable state implies a mutable program
2-8
D. H. Lorenz and B. Rosenan
Program mutation is considered hard in imperative languages, because mutating
a program as it runs requires a corresponding restructuring of its state. However,
declarative programming languages do not have a notion of state, making them easier
to mutate. In fact, some declarative paradigms have a long tradition of program mutation as alternative to storing state. The Prolog [8, 12] predicates assert/1 and retract/1,
for example, add and remove clauses to and from the program. Many imperative
Prolog programs that interact with users modify a declarative subset of the program
using these predicates in order to store user data. Then, to answer user questions,
they query that declarative subset [5, p163]. 9
Most implementations of traditional programming
languages make assumptions regarding the size of the program. Regardless of whether
the language is compiled or interpreted, a common assumption is that a program
is small enough in size so it can be loaded completely into memory. In contrast,
application developers typically do not make similar assumption regarding the size
of the data (state) stored by the application, and such data is often stored to disk
on the database, and in some cases is even distributed across many database nodes.
Consequently, traditional language implementation techniques are often inadequate
for implementing applications.
However, some existing languages can handle such large programs. One example is
Datalog [6]—a data representation and query LP language for deductive databases [33].
It is not Turing complete, and thus is not a general programming language, but its
“programs” (the contents of a deductive database) are both declarative, large, and
mutable.
Big state implies a large program
3
Application Embedding
Having established the similarity between applications and DSLs, next we examine the
two well-known techniques for implementing DSLs, and correlate them to different
approaches for application development.
3.1 Embedded Applications as Embedded DSLs
In the field of programming languages, there are two well-known techniques for
implementing DSLs: external DSLs are implemented as compilers or interpreters,
while internal DSLs [10, 17, 30, 34] are implemented as software libraries in a host
language in a process sometimes referred to as DSL embedding [23, 21], such that a
program in the DSL is also a program in the host language.
The two approaches have significant trade-offs [19, 27]. On the one hand, external
DSLs give their developers absolute freedom in defining the syntax and semantics of
9
An example Prolog program that uses mutating as an alternative to manipulating state is
presented in appendix D.
2-9
Application Embedding: A Language Approach to Declarative Web Programming
Figure 2
A comparison between (a) a host application and (b) a host language
the DSL, while internal DSLs are bound to restrictions (both syntactic and semantic)
posed by the host language. On the other hand, external DSLs leave the burden of
providing an efficient implementation, as well as editors, debuggers, and other tools
on the DSL developer, while internal DSLs reuse much of the facilities created for the
host language. For example, an optimizing compiler that exists for the host language
will optimize (with varying levels of success) code written in DSLs it hosts.
Recall that we compare the logic-tier
of a web application to an interpreter of a DSL (section 2). Consequently, applications
implemented using the 3TA correspond to external DSLs, implemented as interpreters.
Indeed, applications implemented in languages such as PHP [26], Java [2], Ruby [37],
or JavaScript [16] are strictly external to their respective implementation language.
In contrast, websites built using a Content-Management System (CMS) such as WordPress 10 or Drupal 11 can be seen as embedded in these applications.
We coin the term Application Embedding to describe the implementation of an
embedded application as payload over a host application. Figure 2 augments figure 1
to describe how internal DSLs map to embedded applications such as WordPress
websites. WordPress itself is a standard 3TA web application for which the logic-tier
acts as an interpreter over its state. The state, however, consists of mainly two parts:
definitions of guest applications (e.g., user website content), and application userdata (e.g., comments left on user websites). The two live side-by-side as the state of
WordPress—the host application. This is similar to a Ruby program consisting of a
DSL definition provided as a library [10] and a DSL program, both living side-by-side
as Ruby code, interpreted by the Ruby interpreter which is implemented externally of
Ruby (table 1c).
As expected, traditional and embedded web applications feature similar trade-offs
to the ones existing for external and internal DSLs. Building a website using a CMS
Web applications as external and internal DSLs
10
11
https://wordpress.org/
https://www.drupal.org/
2-10
D. H. Lorenz and B. Rosenan
is usually much easier than building it using, e.g., a web framework. However, a
web framework gives much more freedom in both the application’s visual design (its
“syntax”) and its functionality (“semantics”). In both cases, CMS users are bound
to what the particular CMS provides. Specifically, websites built using CMSs are
typically almost entirely static, where the user building the website can only choose
functionality from a predefined collection (a contact form, comments, “likes”, etc).
3.2 Host Application as a Host Language
Following CMS as a conceptual example, a host application needs to be, first and foremost, a web application. It needs to store state, which consists of both the embedded
application (e.g., the content of a website in a CMS), along with the state of that
application (e.g., comments and likes on pages of such a website). However, unlike
CMSs, we want our host application to support arbitrary functionality. Therefore, the
host application needs to be general-purpose. We define a general-purpose application
as an application App for which L 〈App〉 is a general-purpose programming language.
We can therefore start by discussing requirements for a host language, and then move
on to discussing a corresponding host application.
Datalog is an example of a language that can represent large, mutable,
declarative programs (section 2.4). Unfortunately, two problems hinder the use of
Datalog as a general-purpose host language. First, it is not Turing-complete. As a
A host language
result it is not suitable for expressing arbitrary functionality. Second, the well-known
method for embedding DSLs in LP languages [29] is inapplicable to Datalog, since it
does not support compound terms. An appropriate host language should, like Datalog,
be declarative and support large and mutable programs; but, unlike Datalog, it should
be Turing-complete, and support DSL embedding.
In previous work [28] we introduced the Cloudlog language. Like Datalog, Cloudlog
is a LP data representation and query language. However, unlike Datalog, Cloudlog
is Turing-complete and provides a choice between Prolog-like (top-down) clauses,
evaluated at query time, and Datalog-like (bottom-up) rules, evaluated at update time.
Running logic at update time, a practice known as denormalization, is common in
applications that handle large datasets and particularly in conjunction with NoSQL
databases [15, 31]. In this paper we show that Cloudlog is also capable of hosting
internal DSLs.
A host application hostApp is an implementation of a corresponding host language L 〈hostApp〉. If our host language is a data representation language,
then the corresponding host application is a database. Specifically, if the host language is a LP data representation language such as Cloudlog, the corresponding host
application is a deductive database. In section 4 we present a concrete host application,
FishTank, such that L 〈FishTank〉 = Cloudlog. In addition to being a database, a host
application must also be able to store and serve static content (HTML, CSS, images,
and JavaScript files) which it receives from developers, and provide access to the
database through HTTP calls.
A host application
2-11
Application Embedding: A Language Approach to Declarative Web Programming
4
FishTank: A Concrete Host Application
To validate our approach we introduce FishTank, a host application we have implemented as proof-of-concept, and explain its underlying LP computational model—the
Fish-Tank Model (FTM). The novelty of FTM is not in the result of the computation,
which is a derivation of theorems from axioms through Modus Ponens (MP), but rather
its unique imperative behavior, namely an explicit choice by the users of this model
between update-time and query-time evaluation for each rule.
For the sake of clarity, we describe this model in three steps, starting with a simplified model that only supports update-time evaluation (section 4.1), continuing to the
complete theoretical model also supporting query-time evaluation as well as application of arbitrary logic at update time (section 4.2), and finally, posing constraints on
the theoretical model to make it adequate for an efficient implementation (section 4.3).
After that we discuss the design and implementation of FishTank, our proof-of-concept
implementation of this model (section 4.4).
4.1 A Simplified FTM
FTM is a computational model based on LP. 12 It is based on the notion of a container
(the “tank”) that, at any given point in time, contains a set of logic axioms (the “fish”).
These axioms can be facts (male fish)—first-order-logic atoms, rules (female fish) of
the form F → A where F is an atom (corresponding to a fact) and A is an axiom of
any kind, and clauses (infertile fish), which are discussed in section 4.2. 13
When a fact is added to the fish-tank, all matching rules are located and applied to
that fact through MP. The resulting axiom from each match is recursively added to
the fish-tank. Similarly, when a rule is added, it is applied to all matching facts in the
fish-tank, and all resulting axioms are added. When a fact or a rule are removed from
the tank all descendant axioms are removed as well.
Properties This simple model has a few important properties. First, it is declarative.
The symmetry between facts and rules, in conjunction with the symmetry between
insertion and removal, guarantees that when this system comes into a steady state,
the contents of the container is a closure (over MP) of the original set of axioms that
were inserted but not removed, regardless of the order in which these operations
were applied. Second, under some simplifying assumptions discussed in section 4.3,
this model can be implemented to provide efficient storage and retrieval for large
datasets. Retrieval starts with a query term (an atom), looking for either facts that
are potentially unifiable with it, or rules that can be applied to such facts. Retrieval is
the key for both querying the state of the container and for finding possible matches
12
We call this model the Fish-Tank Model due to the spontaneous nature in which facts and
rules interact to create new axioms, like male and female fish spontaneously finding each
other to produce offsprings in a fish-tank.
13
Here we give an intuitive high level description of the computation model. The precise and
complete definition of the FTM is provided in appendix B.
2-12
D. H. Lorenz and B. Rosenan
for inserted rules or facts. Third, this model is scale-friendly. Since it does not provide
global atomicity, there is no single point of failure or no inherent global bottleneck
limiting scalability.
The question remains, is this model powerful enough to represent the
business logic of any application? The simplified FTM presented so far is somewhat
limited. Consider for example a simple search engine built on top of the FTM. The
input to this search engine is a set of facts representing documents that contain
text, and the output is a result for a query that is given one or more keywords. The
simplified FTM is missing two key capabilities to allow an efficient implementation of
such an engine. The first is the ability to perform arbitrary logic (in this case, extraction
of keywords from the text) when applying rules to facts. While such logic can be
performed in our model, it will require the creation of at least one new axiom per each
step of the parsing process, i.e., for each character in the text. Even if keywords are
successfully extracted, multi-keyword searches (i.e., queries for documents containing
all keywords within a given set) are inefficient under the simplified FTM. As we do
not have access to the query itself when we prepare the results, it is not enough to
index documents by each keyword they contain. We need to index each document
by each possible combination of keywords.
Even if we limit the number of keywords
in a search to k, we still need to have nk index entries per document (where n is the
number of different words in the document). If we could apply logic at query-time,
we could have held an index mapping single keywords to documents (k entries per
document), and when searching for a combination of k keywords we would search
for only one keyword, and then filter out results not containing the remaining k − 1
keywords.
Limitations
4.2 The Complete FTM
The complete FTM addresses the need for query-time logic by introducing dynamic
clauses, predicates, and goals, and addresses the need for arbitrary logic at propagationtime by introducing guards.
Facts and rules exist in the FTM to create
dynamic clauses, which are consulted by queries. Dynamic clauses are axioms of
the form H ` B , where H is an atomic dynamic goal, and B is a dynamic goal of
any kind. Clauses, like rules, convey logical deduction. However, while rules are
applied at update time, clauses are applied at query time. Dynamic clauses define
dynamic predicates—predicates representing relations that change over time. This
is in contrast to static predicates, which once defined, do not change. A dynamic
predicate is expected to consist of a large number of clauses, each providing a separate
set of results. In many cases, many of these clauses will be trivial, i.e., of the form
H ← >, where > is the trivial goal. These are the cases where the propagation of
rules and facts applied all the necessary logic. However, in cases where logic needs to
be applied at query time (e.g., in multiple-keyword searches), a non-trivial body (B )
defines the steps necessary to find the results.
Dynamic clauses, predicates, and goals
2-13
Application Embedding: A Language Approach to Declarative Web Programming
In cases where non-trivial logic needs to be applied during the propagation
of axioms, a guard, can be provided as a part of the rule. A rule is defined as F {G} → A,
where G is a goal representing the guard (a rule of the form F → A is read as F {>} → A,
where the guard is the trivial goal >). During the application of a rule to a fact, the
fact’s guard is evaluated, and a resulting axiom is produced for every result of the
guard. If the guard fails, no axioms will be produced even though the fact matched
the left-hand side of the rule. If the guard succeeded multiple times, multiple axioms
will be produced. In section 5 we use guards to index tweets in Tweetlog. The guard
examines tweets and provides a result for each token in each tweet. As result, the
rule produces a search result for each such token.
Guards are required to be static goals, i.e., goals for which the set of results does
not change over time, to ensure consistent behavior between additions and removals.
When a fact is removed, all corresponding rules are applied to it to find out which other
axioms also need to be removed. For each such application, the guard is consulted. If
the guard behaved differently than the way it originally behaved when the fact (or
the rule) were introduced, the removal would not have been complete.
Guards
4.3 FTM Made Practical
The complete FTM is a theoretical model. To make it practical and efficient, some
simplifying assumptions on usage need to be made.
We distinguish between two kinds of axioms: concrete
and generic. Concrete facts are facts of the form f (t 1 , . . . , t n ), where t 1 is a ground
term (a term that does not contain logic variables). A concrete rule is a rule of the
form F {G} → A where F is a concrete fact. A concrete clause is a clause of the form
p (t 1 . . . t n ) ` B , where t 1 is a ground term. Generic axioms are axioms that are not
concrete.
To promote efficient implementation we assume that there is a small number of
generic axioms at any given time in the fish-tank, and that they are seldom added
or removed. Intuitively, such axioms should be considered a part of the embedded
application’s definition, and not a part of its state. The state should consist exclusively
of concrete axioms. We also assume that the number of facts, rules, or clauses in the
fish-tank, with a certain t 1 value, is bounded at any given time.
Concrete axioms
are indexed by the value of t 1 —their subject. When a concrete
fact f t 10 , . . . , t n0 is added to the fish-tank, matching rules are found by using t 10 as
key. When a rule is added, matching facts are found the same way. Similarly, concrete
queries will only consider concrete clauses they will find using the goal’s first argument.
Generic axioms are harder to manage. Having variables in their first argument
(or having a variable as first argument) there is no easy way to retrieve a bounded
collection of matching axioms. As result, any update (addition or removal) of generic
facts or rules requires processing all rules or facts (respectively) that share the same
fact-name ( f ). While this process may take some time, it does not happen often. It
occurs on software updates and resembles data migration that traditionally occurs in
such events.
Concreteness and indexing
2-14
D. H. Lorenz and B. Rosenan
Figure 3 FishTank Architecture. R, N/E, and M respectively denote RabbitMQ, NodeJS/Ex-
press, and MongoDB.
4.4 Design and Implementation
Our proof-of-concept implementation of the FTM is called FishTank.14 FishTank’s architecture is based on MEAN [11]: MongoDB [3], Express,15 AngularJS, and NodeJS.16 This
is a relatively new architecture for web applications that is designed for scale (the use
of the NoSQL database MongoDB), and a single implementation language: JavaScript.
We use an architecture designed for web applications because FishTank is a web
application.
Figure 3 shows the structure of a typical FishTank cluster. It consists of a MongoDB
cluster (M-nodes), a RabbitMQ cluster (R-nodes) and an array of NodeJS/Express
nodes (N/E nodes) connected to a load-balancer.
The MongoDB cluster stores the axioms along with their multiplicity. 17 All facts
with a certain first argument value, and all rules matching such facts are stored in
the same MongoDB document. MongoDB guarantees atomic updates within a single
document. This is important to avoid the chance that a fact and a rule, added to the
database nearly at the same time, miss each other during propagation.
Axioms resulting from applying rules to facts are not added directly, but instead are
placed in a common work queue, maintained by the RabbitMQ cluster. 18 We do this
to ensure low update latency. By placing these axioms in a highly-available queue we
guarantee that these axioms will eventually be added to the database without having
the user wait until this happens.
Scalability and availability
FishTank’s design follows the 3TA, with a separation, on
the server-side, between the stateless N/E nodes, and the stateful (but application-
14
https://github.com/brosenan/fishtank
http://expressjs.com/
16
https://nodejs.org/
17
Corresponds to T in appendix B.4.
18
Corresponds to Q in appendix B.4.
15
2-15
Application Embedding: A Language Approach to Declarative Web Programming
agnostic) M-nodes and R-nodes. This separation allows us to use a large number of
N/E nodes to address high load. Being stateless, they become redundant, so that
if one of them crashes the load balancer can send subsequent requests to the other
nodes and the service as a whole will remain available. MongoDB and RabbitMQ are
both designed for scalability and availability, so the entire system will not fail due to
a failure of a single M-node or R-node, or even multiple nodes (in most cases).
Being based on the MEAN architecture, JavaScript plays an important
role in the implementation of FishTank. FishTank is implemented as a server (over
NodeJS and Express), and a client library, extending AngularJS. The server-side code is
divided to JavaScript code, providing capabilities such as responding to HTTP requests
and using MongoDB and RabbitMQ, and Cedalion [27] code, implementing the logic
itself. We chose Cedalion due to its expressiveness and due to the facts it is a LP
language, making it a natural choice for implementing a deductive database.
To give a sense of size, the server-side implementation consists of ~1000 Cedalion
statements (a typical Cedalion statement represents a few lines of code, but Cedalion
uses projectional editing, so the classic measure of lines-of-code does not apply), and
~3000 lines of JavaScript code. Both include unit tests, which on the JavaScript side
accounts for about half the code. The client-side library consists of ~650 lines of
code, about half of which are unit tests. However, this library is currently very partial,
implementing just enough functionality to support the implementation of Tweetlog.
Implementation
5
Tweetlog: An Embedded Application
We evaluate our approach by demonstrating that: (1) the FTM is powerful enough to
allow non-trivial applications to be developed over it, and (2) the proposed separation
of code assets successfully segregates all imperative code in the application-agnostic
implementation of FishTank, so that all application-specific assets (on the server, and
with small unavoidable exceptions, on the client) are declarative.
We demonstrate both by presenting an implementation of Tweetlog 19 over FishTank.
We show that this implementation closely relates to our definition of the corresponding
L 〈Tweetlog〉.
Figure 4 shows a screenshot of Tweetlog running on a web browser. The navigation
bar at the top of the page allows users to log in (enter their user name) and to search
for content. The page displayed in figure 4 is the timeline of user Foo, which is the
page a user enters after logging in. This page shows the tweets the user registered
to, and provides boxes for tweeting and for following users. On the left, the sidebar
menu contains links to the timeline (this page), a page showing the tweets of the
logged-in user, his or her followers and followees (followers and followees are not yet
implemented), as well as an about page describing the application.
19
https://github.com/brosenan/tweetlog
2-16
D. H. Lorenz and B. Rosenan
Figure 4 Screenshot of the Tweetlog application
Figure 5
User Abstract Syntax
Tweets are tokenized into words, hashtags, and user IDs. User IDs and hashtags
become links to pages showing tweets related to a user or containing a hashtag,
respectively.
5.1 Abstract Syntax
We use Cedalion concept declarations [27] to define concepts in the abstract syntax of
L 〈Tweetlog〉. For example, figure 5 shows a piece of L 〈Tweetlog〉’s abstract syntax.
Statement 1 in figure 5 declares the type User, and statement 2 in figure 5 is equivalent
to User = user (String) from equation (1). This declaration has two parts: a mandatory
type signature (“where Name::string”) and an optional projection definition (“display
as...”). That second part gives the user concept its concrete syntax within the context of
L 〈Tweetlog〉’s definition, and allows it to display as @Name rather than user (Name).
This projection definition helps making L 〈Tweetlog〉’s definition more concise and
intuitive. However, in our current implementation we chose not to impose this concrete
syntax on L 〈Tweetlog〉 users, and thus this concrete syntax is separated from the
concrete syntax of L 〈Tweetlog〉, which is described next. 20
5.2 Concrete Syntax
The concrete syntax of L 〈Tweetlog〉 is defined using templates embedded in the
application’s main HTML file (index.html). Figure 6 shows the template defining the
concrete syntax of user (String). It comes as a <script> element, of type ng−template,
which is picked up by AngularJS. We add the clg−concept attribute to associate this
template with the user concept from the abstract syntax. The user name is bound to
User, that is used inside the template.
20
The complete abstract syntax can be found in figure 10.
2-17
Application Embedding: A Language Approach to Declarative Web Programming
1 <script type=”ng−template” clg−concept=”t:user(User)” id=”user”>
2 <span><a href=”{{url(’userTweets’,␣[User])}}”>@{{User}}</a></span>
3 </script>
Figure 6
Concrete syntax for user (String)
1 <script type=”ng−template” clg−concept=”t:tweet(User,␣Tweet)” id=”tweet”>
2
<span clg−render−term=”User”></span>: <span clg−render−term=”Tweet”></span>
3 </script>
Figure 7
Concrete syntax for tweet (User, Tweet)
1 <form>
2 <clg−fact pattern=”t:tweeted(t:user(User),␣Time,␣t:text(t:rawText(Tweet)))”
3
name=”tweet”
4
assign=”User:␣$session.user,␣Tweet:␣text,␣Time:␣now()”>
5 </clg−fact>
6 <input type=”text” ng−model=”text”>
7 <input type=”submit” ng−click=”add_tweet();␣text␣=␣’’” value=”Share!”>
8 </form>
Figure 8 Form providing the tweeted (User, Time, Tweet) fact concrete syntax
The template’s body is a <span> element, containing a link (<a>) linking to a page
containing the tweets of the given user. The link is generated within the href attribute,
using the {{...}} syntax. This syntax is replaced by AngularJS with the value to which
the expression inside the braces evaluates to. In this case this is the url() function,
which builds a URL using the given location and parameters. The text displayed in
the link is a “@” followed by the name of the user.
Figure 7 shows the concrete syntax of a timeline entry corresponding to a tweet, is
another example for such template. This time, the arguments (a user ID and a tweet)
are compound terms, that by themselves need to be expanded using other templates.
The clg−render−term directive in the empty <span> element tells the FishTank client
library to render the term held by this attribute (either the user ID or the tweet) in
the enclosing <span> element by applying further templates to either the user ID or
the tweet.
The concrete syntax of an L 〈App〉 must also provide users ways to mutate the state
of an application (section 2.3). Currently, our client-side library only supports addition
of new facts. These are performed using the <clg−fact> directive. Figure 8 shows a
form in which Tweetlog users can add textual tweets. The <clg−fact> directive takes a
pattern (a Cloudlog term containing the pattern of facts to be created, where logic
variables represent the data that changes), a name, that is later used to submit the
new fact, and a variable assignment, mapping values from the AngularJS scope to
the logic variables in the pattern. In this case, the pattern represents a textual tweet.
The first <input> element is a text box in which the tweet text is placed. It uses an
AngularJS ng−model directive to associate the value of the field to the scope variable
2-18
D. H. Lorenz and B. Rosenan
Figure 9 Timeline semantics
“text.” The second <input> element is a submit button, which upon clicking calls the
add_tweet() method, a method added to the scope by the <clg−fact> directive. It also
clears the text.
Event handlers such as this ones are the only pieces of imperative code that exist in
Tweetlog’s definition. Our design leaves client-side event handlers imperative since
imperative programming provides the most expressive way to convey such actions.
However, these pieces of code are very limited in their ability to affect the state of the
application. The function add_tweet(), for example, takes no arguments, so the only
thing it can convey is the desire to create a new tweet based on variables stored on
the client. Other than that, event handlers are free to change the state of the client,
as in clearing the text.
5.3 Semantics
Figure 9 shows the definition at the heart of the Tweetlog application. Statement 1 in
figure 9 represents the proc-tweet rule (equation (2)) from section 2.2. The axiom
keyword tells us the rest of this statement is a Cloudlog axiom (a rule, in this case). A
rule has three parts: (1) a fact to match (in this case, a tweet); (2) a guard, which is a
Cedalion goal to be satisfied for each such fact (in this case, a conjunction of a goal
unpacking and repacking the text from the tweet, and a goal tokenizing the text);
and (3) an axiom to be created for each result emitted by the guard (in this case, a
procTweet fact).
Statement 2 in figure 9 indexes processed tweets by their content (the search1
rule, equation (3)), and statement 3 in figure 9 indexes processed tweets by the user
who made them (the search2 rule, equation (4)). Note that both the tokenization
and indexing of all tweets are done using rules, that operate upon insertion of tweets.
Statement 4 in figure 9 defines the timeline predicate. This is a rule that emits a
rule that emits a clause (with :- replacing ` in the syntax). The clause is trivial, with
> as its body. Being dependent on two kinds of facts, this rule is similar to an SQL
join. As in its specification (the time-line1 rule, equation (5)), it provides an entry in
the timeline for each tweet associated with each followed user. By choosing between
a rule and a (non-trivial) clause we get a choice of temporal order. The order we
choose here is to start from a following relationship (to get the followee), go through
all associated tweets, and then place them in the timeline. Like indexing, this too
happens as tweets and following relationships are introduced, and the only thing
2-19
Application Embedding: A Language Approach to Declarative Web Programming
happening when a user searches for his or her timeline is fetching all timeline clauses
associated with the user. Note that the first argument in the timeline predicate is the
user, so FishTank guarantees fast (O (1) I/O operations) access to it. 21
6
Discussion
We have made certain simplifying assumptions concerning the manner in which
FishTank is to be used, assumptions such as the number of generic axioms and the
rate in which they are being updated (section 4.3). These assumptions have direct
consequences in terms of applicability, correctness, and performance.
6.1 Applicability
Many programming platforms, such as databases and programming languages, make
simplifying assumptions on usage. Prolog, for example, indexes the clauses of predicates based on their first argument [20]. In doing so the language implementation
assumes predicates with a large number of clauses are defined in a way such that the
the first argument differs between different clauses, so that after applying the index,
only a small number of clauses (typically one or zero) is left to be considered. This
assumption is translated to a guideline for developers, to choose the first argument
wisely.
MongoDB assumes documents do not exceed a certain size. This restriction derives
our restriction regarding the number of concrete facts and rules with a certain t 1
value. In MongoDB itself, this assumption becomes a guideline (one of many) for data
modeling. 22
Similar to Prolog and MongoDB, we treat the assumptions made by FishTank to
be guidelines for users to structure their facts and rules in a certain way. The first
argument of facts should convey their subject—an entity they describe. All facts
and rules that are considered user data should relate to a subject. The cap on the
number of facts and rules with a certain t 1 value maps to a cap on the number of facts
describing that subject. For example, in Tweetlog we use the user ID as the subject
for timeline entries. If the number of expected timeline entries per user exceeds what
is allowed in FishTank, a different subject needs to be chosen. One possibility is the
pair (user, day), where day is the event’s timestamp integer-divided by the number of
seconds in a day. This creates subjects that are both small enough to comply with the
restriction, but at the same time relevant in the sense that we will usually query all
entries with a certain subject together.
The only kinds of facts and rules that cannot be associated with a specific entity are
those that speak of data in general—the generic axioms that make up the definition
21
22
To complete the semantics, the tokenization rules (figure 11) are implemented using a
pre-existing DSL (figure 14).
https://docs.mongodb.com/v3.2/core/data-modeling-introduction/
2-20
D. H. Lorenz and B. Rosenan
of the application. The definition of an application is typically significantly smaller
than its data, and changes only on software updates, which are usually significantly
less rapid than user data updates.
6.2 Correctness
The correctness requirement for FishTank seems obvious: queries should be evaluated
by taking into account all facts and rules added to the database (and not removed
from it) until the time of the query, and should be evaluated following Cloudlog’s
semantics. However, since we are talking about a large, distributed system, the notion
of time is not exact. In a distributed system, every computer node has its own time
and as a result a fact that is added to the database on one node is not known the
same exact moment to all nodes. Synchronizing all nodes by means of, e.g., ACID
(Atomic, Consistent, Isolated, and Durable) transactions can help, but will inhibit
the scalablility of the system. We therefore require eventual delivery, that is, if we
stop updating the database at one point, the database will come to rest in a situation
where all facts and rules are taken into account by all queries.
Although this seems simple enough, there is one special case to watch for. If a fact
and a matching rule are added to the database at around the same time, the new rule,
which is supposed to be applied to all existing facts, might not yet be aware of the
new fact being added. Likewise, the new fact may not be aware of the new rule. In
such case, the fact and rule will miss each other and queries will not see their mutual
product.
To overcome this problem our design uses a strongly consistent database (MongoDB).
Strongly consistent databases do not guarantee global sequentiality, but do provide
local sequentiality. In MongoDB’s case, all operations performed on a single document
are guaranteed to be sequential. We leverage this property to achieve correctness by
storing all concrete facts and rules of the same subject (which are thus candidates for
being matches of each other) in the same document. By doing this we make sure that
when adding a fact we are aware of all matching rules and vice versa. 23
6.3 Performance
table 2 summarizes the performance we require of our implementation. All user operations (adding or removing facts and queries for pre-calculated data) are completed
after performing a constant-bound number of I/O operations, taking time linear by
either the number of the results (in a query) or the number of rules matching the fact
being added or removed (in an update). We allow an update operation to continue
propagating asynchronously after notifying the user of its success, as long as we can
guarantee delivery. That is, as long as we verified that the update does not conflict
23
Strong consistency also solves the correctness problem for generic facts and rules, but the
mechanism to support that is more complicated and beyond the scope of this paper.
2-21
Application Embedding: A Language Approach to Declarative Web Programming
Operation
Synchronous
I/O Ops Time
Asynchronous
I/O Ops
Time
Add/Remove concrete fact/rule
Query de-normalized goal
Add/Remove generic fact/rule
O (1)
O (n)
O (n)
O (n)
O (1)
O (n)
N/A
N/A
O (N )
O (N )
O (N n) O (N n)
Table 2 Performance requirements. n is the number of axioms involved in a query or
update. N is the number of axioms in the database.
with any previous updates, and that the necessary actions were logged to a commit
log.
We achieve that by using a reliable work queue (RabbitMQ). Adding or removing a
concrete fact or rule is done by accessing a single MongoDB document (using O (1)
I/O operations), querying all matching rules or facts (n). We then apply them to
one another (O (n) time) to receive a set of facts and rules to be added or removed.
Instead of adding or removing these directly we defer these operations to the work
queue (O (1) I/O operations, O (n) time) and once acknowledged by the queue we
acknowledge making the change.
Adding or removing generic axioms requires walking through all existing concrete
facts and rules, looking for matches. We allow them to take O (N ) time, with N
representing the overall number of axioms in the database. It is acceptable for these
operations to take a long time as they are only done when the L 〈App〉’s definition
changes, i.e., on application software upgrades. This is similar to data migration
operations that are performed when upgrading the software of a traditional web
applications.
MongoDB uses locks to synchronize operations done in the scope of a single document. We rely on this synchronization for correctness. However, under some conditions,
these locks can result in congestion, which will increase query and update time beyond
the desired asymptotic requirements listed in table 2. For example, if m clients are
trying to access (query or update) the same subject (a certain value of t 1 mapped to
a single MongoDB document) at the same time, the locking itself will take O (mn)
time on average (waiting for O (m) operations taking O (n) time each to complete),
exceeding the O (n) requirement. However, we make the assumption that only a
constant-bound number of operations is done at one time on a single subject. As
further discussed in section 6.1, this assumption becomes a guideline for developers.
For example, using a user ID as a part of the subject key (as in our example from
section 6.1 of using (user, day)) can help ensure constant-bound parallel accesses, given
that a single user does not make a large number of updates and queries in parallel.
7
Related Work
Most relevant related work include other approaches that can be considered application
embedding, and multi-tier or tierless approaches for deploying code on multiple tiers
from a single, coherent code artifact.
2-22
D. H. Lorenz and B. Rosenan
7.1 Application Embedding
Content Management Systems (CMSs, discussed in section 3.1) are host applications
that allow users to build websites, which are stored as user data and served from
within the CMS. Websites are typically very easy to build on top of CMSs, as the user
does not need to worry about most technical aspects. However, such websites are
bound to restriction posed by the CMS, which apply to both the website’s graphical
design and its functionality. Functionality is typically very limited in these systems, as
they do not provide any programmability to users.
Apache CouchDB [1] is a NoSQL database that acts as a host application for embedded CouchApps. 24 It is a document store, storing JSON objects (documents). Its
distinctive feature, however, in relation to other NoSQL document stores such as
MongoDB, is a special kind of documents, known as design documents, that can be
stored as a part of the state of a database. These documents contain JavaScript code
that augments the database behavior. Most prominently, a design document can
define views. Each view is defined using a JavaScript map function and optionally a
reduce function. Unlike map/reduce capabilities that exist as a query feature in other
NoSQL databases (such as MongoDB), in CouchDB a view is a part of the state of the
database. This has interesting consequences. It allows map and reduce operations
to be performed incrementally on a materialized view. This means that the view
data is stored as a B+ tree, one tree per view. When a new document is added to
the database the map function in each view is invoked on it to emit zero or more
key/value pairs into the tree. When a document is deleted from the database the map
functions are invoked once more to find view elements to be removed. The reduce
function calculates aggregated values for each affected node in the tree. The result is
automated denormalization of the database. To our knowledge, with the exception of
FishTank, CouchDB is the only other database that has such support.
CouchDB’s map functions are similar to rules in Cloudlog and the FTM. They take a
subset of the facts in the database (JSON documents) and map them into other facts
(key/value pairs in the view tree). However, this cannot be used for denormalized joins
such as the one in statement 4 in figure 9. There are two inhibitors: (1) derived facts
(key/value pairs emitted by a map function) go into a dedicated view tree, and not
the general database, and therefore do not trigger further map function invocations,
and (2) rules cannot emit other rules. As a consequence, each key/value pair in a view
tree is based on exactly one raw document. In this, the FTM and FishTank represent a
more powerful approach.
7.2 Multi-Tier Programming
HOP [36] is a programming language that extends JavaScript with multi-tier features.
A HOP program consists of services, which run on the server. Service code can contain
client-side calls, with special syntax distinguishing between the two. Such code can
24
http://docs.couchdb.org/en/1.6.1/couchapp/
2-23
Application Embedding: A Language Approach to Declarative Web Programming
appear in handlers for HTML events (HTML code can be embedded in HOP service
code). Similarly, server-side code can be embedded in client-side code using different
syntax. In both cases, HOP hides the necessary communication from the developer.
Stips.js [32] is a tier-splitting process that splits a single JavaScript program into
two separate programs, one running on the server and the other running on the client.
The separation between client and server is made explicit by using special comments
(@client and @server), but client code can call functions defined in the server code.
The goal behind HOP and Stips.js is to alleviate the dichotomy between the presentation-tier and the logic-tier. WebDSL [22], Ur/Web [7], and Links [9] go beyond that,
and add the logic-tier to the mix, by adding support for schema definition and queries
as a part of the language. An Ur/Web program, for example, will run on the logic tier,
but will send JavaScript code to the presentation-tier, and SQL code to the data-tier.
Its type system provides strong guarantees for the consistency between these artifact,
making sure that, e.g., type validation in the presentation-tier is consistent with the
type the same variable is given in the database schema.
While tierless programming aims to alleviate the dichotomy between the different
tiers (e.g., between the logic-tier and the data-tier), application embedding aims to
alleviate the dichotomy between the tenants of these tiers, namely the business logic
and the data, by treating the business logic as data.
Multi-tier programming hides the immediate imperative aspects that are a direct
consequence of the 3TA, namely the communication between the tiers. HOP and
Stips.js allow client-side code to call server-side code (and in HOP’s case, vice versa) in
a way that resembles a regular function call. WebDSL, Ur/Web, and Links also provide
a seamless integration to a database, in which the programmer is not concerned about
creating a database connection or making SQL calls, but instead writes the application
in a way that (in the case of Ur/Web and Links) weaves SQL queries in other code
(e.g., in HTML fragments), or (in the case of WebDSL) associates data entities to
output formatting.
However, multi-tier programming platforms do not inherently result in declarative
programming. The manipulation of state remains imperative with these approaches.
WebDSL includes a Java-like imperative programming language defined purely for
the purpose of manipulating state. It treats entities as objects (hiding the fact they
are stored on the logic-tier), but not hiding its imperative nature. Ur/Web and Links
both depend on the declarative nature of SQL and HTML to allow applications to be
defined in a mostly declarative manner, as long as their data model maps cleanly to a
relational data model.
For example, a micro-blogging application similar to Tweetlog could be defined in
Ur/Web and Links in a declarative way, similar to our own implementation, as long
as (1) it supported only one kind of tweets (e.g., textual tweets only), and (2) it did
not need any kind of denormalization. In such settings, the timeline would have been
defined using HTML wrapped in a declarative SQL select query, which conveys the
business logic in this case.
However, if we add a requirement to support, e.g., image tweets (as we do in
Tweetlog), the database schema and SQL query become significantly more complex.
This is due to a problem known as the impedance mismatch [35]—a mismatch between
2-24
D. H. Lorenz and B. Rosenan
what the domain object model requires, and what the relational model provides.
The domain object model in this case includes polymorphism, in which textual and
image tweets share some functionality and differ in some particular aspects. However,
the relational model does not support polymorphism out of the box. By supporting
compound terms, Cloudlog does support polymorphism.
The introduction of denormalization, which is often necessary to keep query time independent of the overall size of the data, is another case where mult-tier programming
does not provide an advantage. Without support for denormalization, applications
written in Ur/Web, Links, and WebDSL will all need to maintain the redundant information explicitly, at update time, using imperative code. FishTank, on the other hand,
performs this kind of work automatically, based on declarative rules.
8
Conclusion
In this paper we describe application embedding, a novel approach to the development
of web applications. With our approach the complete application is divided into two
separate applications: a host application and an embedded application. The former is
implemented using general-purpose, often imperative programming languages, while
the latter is written in a declarative programming language. The host application is
general-purpose, making it suitable for, and thus reusable across multiple applications.
Our approach is rooted in the field of language engineering. It is based on a twostep thought process, in which the first draws a mapping between web applications
and programming languages, and the second uses this mapping to find the novel
equivalent of DSL embedding: application embedding.
As validation we provide an implementation of an application, in two parts. A
reusable host application named FishTank, and an example embedded application
named Tweetlog. Tweetlog’s code is almost entirely declarative, consisting of a definition of the business logic in the form of logic rules, and the presentation given as
HTML with directives defining data bindings. The only imperative parts are expressions
associated with user events, as discussed in section 5.2. FishTank’s code is much more
complicated, but it is general-purpose by design, so that it can be used to implement
potentially many different applications. Establishing this by embedding several other
applications over it, however, is a topic left for future work.
One important aspect not addressed by our proof-of-concept implementation of
FishTank is access control. We expect a real-life host application to address access
control by authenticating users using their identity to derive their rights. In previous
work [28] we describe how Cloudlog can be extended to address access control, an
important part of the business logic of applications.
Modern web applications are complicated. Developing a feature-rich, stable, usable
application requires versatile skills. Today, web application developers are required to
do everything. With lack of separation of concerns, the implementation of the business
logic of a specific application is often tangled with, e.g., scalability, availability, and
consistency considerations, which are often not easy to master. The separation of
code assets presented in this paper decouples these concerns from any particular
2-25
Application Embedding: A Language Approach to Declarative Web Programming
application specific concerns, thus allowing collaboration between teams and even
companies (e.g., through open-source development) to build better host applications
for everyone to benefit.
This research was supported in part by the Israel Science Foundation (ISF) under grant No. 1440/14.
Acknowledgements
References
[1] J. Chris Anderson, Jan Lehnardt, and Noah Slater. CouchDB: The Definitive Guide.
O’Reilly, 1 edition, 2010.
[2] Ken Arnold and James Gosling. The Java Programming Language. The Java
Series. Addison-Wesley, 1996.
[3] Kyle Banker. MongoDB in Action. Manning Publications Co., Greenwich, CT,
USA, 2011.
[4] Henk Barendregt. Self-interpretation in lambda calculus. Journal of Functional
Programming, 1(2):229–234, April 1991.
[5] Ivan Bratko. Prolog Programming for Artificial Intelligence. Addison-Wesley, 2001.
[6] Stefano Ceri, Georg Gottlob, and Letizia Tanca. Logic programming and databases.
Surveys in Computer Science. Springer-Verlag, 1990. doi:10.1007/978-3-642-839528.
[7] Adam Chlipala. Ur/Web: A simple model for programming the web. In Proceedings of the 42nd Annual ACM SIGPLAN-SIGACT Simposium on Principles of
Programming Languages (POPL’15), pages 153–165, Mumbai, India, January 2015.
ACM Press. doi:10.1145/2676726.2677004.
[8] William F. Clocksin and Christopher C. Mellish. Programming in Prolog. Springer
Verlag, Berlin Heidelberg, 1st edition, 1981. doi:10.1007/978-3-642-96661-3.
[9] Ezra Cooper, Sam Lindley, Philip Wadler, and Jeremy Yallop. Links: Web programming without tiers. In Proceedings of the 5th International Symposium on
Formal Methods for Components and Objects (FMCO 2006), number 4709 in Lecture Notes in Computer Science, pages 266–296, Amsterdam, The Netherlands,
November 2007. Springer Verlag. doi:10.1007/978-3-540-74792-5_12.
[10] H. Conrad Cunningham. A little language for surveys: constructing an internal DSL in Ruby. In Proceedings of the 46th ACM Annual Southeast Conference
(ACMSE’08), pages 282–287, Auburn, Alabama, March 2008. ACM Press.
[11] Scott Davis. Architecture of the MEAN Stack: Exploring the New Web Dev Tools:
MongoDB, ExpressJS, AngularJS, PNodeJS. O’Reilly Media, 2015.
[12] Pierre Deransart, Laurent Cervoni, and AbdelAli Ed-Dbali. Prolog: The Standard:
reference manual. Springer-Verlag, London, UK, 1996.
[13] Wayne W. Eckerson. Three tier client/server architectures: Achieving scalability,
performance, and efficiency in client/server applications. Open Information
Systems, 3(20):46–50, 1995.
2-26
D. H. Lorenz and B. Rosenan
[14] Sebastian Erdweg, Tijs Storm, Markus Völter, Meinte Boersma, Remi Bosman,
William R. Cook, Albert Gerritsen, Angelo Hulshout, Steven Kelly, Alex Loh,
Gabriël D. P. Konat, Pedro J. Molina, Martin Palatnik, Risto Pohjonen, Eugen
Schindler, Klemens Schindler, Riccardo Solmi, Vlad A. Vergu, Eelco Visser, Kevin
Vlist, Guido H. Wachsmuth, and Jimi Woning. The state of the art in language
workbenches - conclusions from the language workbench challenge. In Martin
Erwig, Richard F. Paige, and Eric Wyk, editors, Proceedings of the 6th International
Conference on Software Language Engineering (SLE 2013), number 8225 in Lecture
Notes in Computer Science, pages 197–217, Indianapolis, IN, USA, October 2013.
SLE 2013, Springer. doi:10.1007/978-3-319-02654-1_11.
[15] Eric Evans. Cassandra by example, May 2010. http://www.rackspace.com/blog/
cassandra-by-example/.
[16] David Flanagan. JavaScript: the definitive guide. O’Reilly Media, Inc., 6 edition,
2011.
[17] Matthew Flatt. Creating languages in Racket. Commun. ACM, 55(1):48–56,
January 2012. doi:10.1145/2063176.2063195.
[18] Martin Fowler. Projectional editing. Martin Fowler’s Bliki. http://martinfowler.com/bliki/ProjectionalEditing.htmlx.
[19] Martin Fowler. Language workbenches: The killer-app for domain specific languages, 2005.
[20] John Gabriel, Tim Lindholm, EL Lusk, and Ross A Overbeek. A tutorial on the
Warren abstract machine for computational logic. Argonne National Laboratory,
1985.
[21] Jeremy Gibbons and Nicolas Wu. Folding domain-specific languages: Deep and
shallow embeddings (functional pearl). In Proceedings of the 19 th ACM SIGPLAN
International Conference on Functional Programming (ICFP ’14), pages 339–347,
Gothenburg, Sweden, September 2014. ACM. doi:10.1145/2628136.2628138.
[22] Zef Hemel, Danny M. Groenewegen, Lennart C.L. Kats, and Eelco Visser. Static
consistency checking of web applications with WebDSL. Journal of Symbolic
Computation, 46(2):150–182, 2011. doi:http://dx.doi.org/10.1016/j.jsc.2010.08.006.
[23] Paul Hudak. Building domain-specific embedded languages. ACM Computing
Surveys (CSUR), 28(4es), December 1996. doi:10.1145/242224.242477.
[24] Neil D. Jones and Alan Mycroft. Stepwise development of operational and
denotational semantics for Prolog. In 1984 International Symposium on Logic
Programming, Atlantic City, February 1984.
[25] Gilles Kahn. Natural semantics. In Proceedings of the Annual Symposium on
Theoretical Aspects of Computer Science, number 247 in Lecture Notes in Computer
Science, pages 22–39. Springer-Verlag, 1987.
[26] Rasmus Jay Lerdorf, Kevin Tatroe, Bob Kaehms, and Ric McGredy. Programming
PHP. O’Reilly & Associates, Inc., Sebastopol, CA, USA, March 2002.
[27] David H. Lorenz and Boaz Rosenan. Cedalion: A language for language oriented
programming. In Proceedings of the 26th Annual ACM SIGPLAN Conference on
Object-Oriented Programming Systems, Languages, and Applications (OOPSLA’11),
2-27
Application Embedding: A Language Approach to Declarative Web Programming
[28]
[29]
[30]
[31]
[32]
[33]
[34]
[35]
[36]
[37]
[38]
[39]
pages 733–752, Portland, Oregon, USA, October 2011. ACM. doi:http://doi.acm.
org/10.1145/2048066.2048123.
David H. Lorenz and Boaz Rosenan. Separation of powers in the cloud: Where
applications and users become peers. In Proceedings of the 2015 ACM International
Symposium on New Ideas, New Paradigms, and Reflections on Programming &
Software (Onward! 2015), pages 146–159, Pittsburgh, PA, USA, October 2015.
ACM. doi:10.1145/2814228.2814237.
Tim Menzies. DSLs: A logical approach, 2001. Lecture Notes, EECE 571F, http:
//courses.ece.ubc.ca/571f/lectures.html.
Tim Menzies, John Black, , Joel Fleming, , and Murray Dean. An expert system
for raising pigs. In The First Conference on Practical Applications of Prolog (PAP’92),
1992.
Jay Patel. Cassandra data modeling best practices, part 1, July 2012. http://www.
ebaytechblog.com/2012/07/16/cassandra-data-modeling-best-practices-part-1/.
Laure Philips, Wolfgang De Meuter, and Coen De Roover. Poster: Tierless programming in JavaScript. In Proceedings of the 37th International Conference on
Software Engineering (ICSE’15), pages 831–832, Florence, Italy, 2015. IEEE.
Raghu Ramakrishnan and Jeffrey D. Ullman. A survey of deductive database
systems. The Journal of Logic Programming, 23(2):125–149, 1995. doi:10.1016/07431066(94)00039-9.
Lukas Renggli and Tudor Gîrba. Why Smalltalk wins the host languages shootout.
In Proceedings of the International Workshop on Smalltalk Technologies (IWST’09),
pages 107–113, New York, NY, USA, 2009. ACM. doi:10.1145/1735935.1735954.
Pramod J Sadalage and Martin Fowler. NoSQL distilled: a brief guide to the
emerging world of polyglot persistence. Addison-Wesley, 2012.
Manuel Serrano, Erick Gallesio, and Florian Loitsch. Hop, a language for programming the Web 2.0. In Proceedings of the 1st Dynamic Languages Symposium
(DLS’06), Portland, Oregon, USA, October 2006.
David Thomas and Andrew Hunt. Programming Ruby: the pragmatic programmer’s guide. Addison-Wesley, 2000.
John Von Neumann. First draft of a report on the EDVAC. IEEE Annals of the
History of Computing, 15(4), October 1993. Reprint of original typescript draft,
June 30, 1945. doi:10.1109/85.238389.
Eric Walkingshaw and Klaus Ostermann. Projectional editing of variational
software. SIGPLAN Not., 50(3):29–38, September 2014. doi:10.1145/2775053.
2658766.
2-28
D. H. Lorenz and B. Rosenan
S =
|
|
|
|
User =
Tweet =
|
Text =
|
Time =
TimelineElem =
|
Token =
|
|
Figure 10
A
tweeted (User, Time, Tweet)
follows (User, User, Time)
timeline (User, Time, TimelineElem)
procTweet (User, Time, Tweet)
searchIndex (Token, Time, Tweet)
user (String)
text (Text)
image (Binary, Text)
plain (String)
tokenized (List (Token))
Number
tweet (User, Tweet)
follolwingYou (User)
word (String)
userID (String)
hashtag (String)
Abstract syntax grammar for L 〈Tweetlog〉
A Language Specification for L 〈Tweetlog〉
A.1 Abstract Syntax
Tweetlog’s state is represented as a set P of statements S. We use algebraic data types
to define S, providing its abstract grammar (figure 10). A statement S ∈ P can be
either a tweet (tweeted (User, Time, Tweet)), indicating a user tweeted some tweet
at a certain time, a following relationship (follows (User, User, Time)), indicating that
one user started following another at a certain time, a timeline entry of the form
timeline (User, Time, TimelineElem), indicating that a certain timeline element is in
one user’s timeline corresponding to a certain point in time, or either a processed
tweet of the form procTweet (User, Time, Tweet), or a search index item of the form
searchIndex (Token, Time, Tweet), which are discussed later.
A tweet can either be textual (text (Text)) or image (image (Binary, Text)) containing
the image’s binary content and some caption text. Text is either a plain string or a
tokenized text consisting of a list of tokens: simple words, hashtags, and user IDs.
A timeline entry is either a tweet or an indication that a user is following the
timeline’s owner.
A.2 Semantics
We define the semantics for L 〈Tweetlog〉 using Natural Semantics [25] (big-step
structural operational semantics). This representation is considered effective for
specifying purely-declarative languages, where the evaluation order is not a part of
2-29
Application Embedding: A Language Approach to Declarative Web Programming
tokens ([])
tokens (t : ts)
whitespace
whitespace
token (word ([]))
token (word (c : cs))
wordchar (c)
token (hashtag (t))
token (userID (u))
::=
::=
::=
::=
::=
::=
::=
::=
::=
"
token (t) whitespace tokens (ts)
"
00
whitespace
"
wordchar (c) token (word (cs))
c where c ∈
/ 0 0 ,0 @0 ,0 #0
0 0
# token (word (t))
0 0
@ token (word (u))
Figure 11 BNF definition of a grammar for tokens
the semantics. We use the predicate P |= S to denote that statement S is true under
program P . Trivially, a statement is true under a program if it is stated by the program:
S∈P
P |= S
We begin our semantic definition by defining processed tweets:
P |= tweeted (U, τ, T )
replaceText T, plain (X ) , tokenized
X0 , T0
parse tokens X 0 , X , "
proc-tweet
P |= procTweet U, τ, T 0
(6)
A processed tweet exists for each (raw) tweet, such that its underlying text is tokenized.
replaceText T, X , X 0 , T 0 is a judgment that replaces the text in a tweet according to
tweet type, and is defined as follows:
replaceText (text (T ) , T, T 0 , text (T 0 ))
repl-texttext
(7)
repl-textimg
(8)
replaceText (image (B, T ) , T, T 0 , image (B, T 0 ))
parse tokens X 0 , X , " is a judgement that relates string X with a list of its tokens:
words, hashtags and user IDs. Tokenization is done using the BNF-like context-free
rules in figure 11. The semantics of these rules is given in figure 12.
What we did here was define a DSL (BNF) to assist us in defining the semantics
for L 〈Tweetlog〉. We call such DSLs helper DSLs. One more thing to note here is that
although we parse a string, this part of L 〈Tweetlog〉 is not its concrete syntax, but
rather a part of its semantics. This is so because parsing is a part of what Tweetlog
needs to do, i.e., a part of its business logic. This tokenization will later be given
concrete syntax, e.g., by treating user IDs and hashtags as links.
2-30
D. H. Lorenz and B. Rosenan
parse (", S, S)
parseempty
parse (A, S, R1 )
(9)
parse (B, R1 , R)
parse (A B, S, R)
parse (c, c : R, R)
(10)
parsetoken
P
parse (c where P, c : R, R)
T ::= T 0
parseseq
parsewhere
parse T 0 , S, R
parse (T, S, R)
(11)
(12)
parsedef
(13)
Figure 12 The semantics of the BNF-like DSL
Now that we have a tokenized tweet, we can define a search index for tweets.
Generally, a tweet can be searched by any of its tokens:
P |= procTweet (U, τ, T )
replaceText (T, tokenized (X ) , tokenized (X ) , T )
P |= searchIndex (ξ, τ, T )
ξ∈X
(14)
search1
As a special case, tweets can also be found by searching the user who tweeted them:
P |= procTweet (U, τ, T )
search2
P |= searchIndex (userID (U) , τ, T )
(15)
Now we are ready to define the timeline. A timeline of user U1 contains all tweets
related to user U2 if user U1 follows user U2 .
P |= follows U1 , user (U2 ) , τ0
P |= searchIndex (userID (U2 ) , τ, T )
P |= timeline (U1 , τ, tweet (U2 , T ))
time-line1
(16)
Additionally, U1 ’s timeline also contains entries for each user U2 following U1 :
P |= follows (U2 , U1 , τ)
P |= timeline (U1 , τ, followingYou (U2 ))
time-line2
2-31
(17)
Application Embedding: A Language Approach to Declarative Web Programming
B
Semantics for the Fish-Tank Model
B.1 Abstract Syntax
A logic term is a compound term, consisting of a name and zero or more child terms,
or a number, or a string, or a logic variable:
term =
|
|
|
name (term, . . . , term)
number
string
variable
A static goal is an atom, consisting of a predicate name (pred) and zero or more
argument terms, or a conjunction of goals (∧), or a negation of a goal (¬), or the
trivial goal >.
goal =
|
|
|
pred (term, . . . , term)
goal ∧ goal
¬goal
>
A (static) clause consists of an atom (head) and a (static) goal (body):
clause = pred (term, . . . , term) ` goal
A dynamic goal (dgoal) is an atom, consisting of a dynamic predicate name and zero or
more argument terms, or a conjunction of dynamic goals, or a negation of a dynamic
goal, or a reference to a static goal:
dgoal =
|
|
|
dpred (term, . . . , term)
dgoal ∧ dgoal
¬dgoal
s (goal)
A dynamic clause (dclause) consists of a (dynamic) atom as head, and a dynamic goal
as body:
dclause = dpred (term, . . . , term) ` dgoal
Finally, an axiom is a dynamic clause, or a fact, consisting of a name (factname)
and zero or more argument terms, or a rule, which consisting of a fact-name with
arguments as trigger, a goal as guard, and an axiom as consequence.
axiom = dclause
| factname (term, . . . , term)
| factname (term, . . . , term) {goal} → axiom
2-32
D. H. Lorenz and B. Rosenan
B.2 Static Goal Semantics
The semantics of static goal G is given by the judgment S |= G , where S is a set of
static clauses. Their semantics is given by the following rules.
> is a goal that succeeds unconditionally:
S |= >
static>
A conjunction of two goals succeeds if both goals succeed under S :
S |= G1 S |= G2
static∧
S |= G1 ∧ G2
A negation of a goal succeed if and only if the goal fails under S :
¬ (S |= G)
static¬
S |= ¬G
An atom succeeds if there exists a clause having that atom as head, for which the
body succeeds under S :
G ` G 0 ∈ S S |= G 0
staticG
S |= G
These rules provide big-step operational semantics, ignoring the order of execution,
and we bring them here for simplicity. For execution order see Prolog’s operational
semantics [24].
B.3 Axiom Semantics
The foundation of the fish-tank model is the derivation of axioms from pairs of axioms
through derivation rules based on Modus Ponens. We define the relation 〈α, β, S〉 → γ,
denoted γ is derived from α and β under S , as follows:
fact (α)
β = α {G} → γ
S |= G
〈α, β, S〉 → γ
fact (β)
α = β {G} → γ
〈α, β, S〉 → γ
S |= G
deriveαF
deriveαR
These two rules represent the cases where α is a fact and where α is a rule,
respectively. The judgment fact (α) is successful if and only if α is a fact. The judgment
S |= G can have zero or more solutions (with different assignments to the variables in
G , which may be shared with γ), yielding potentially different results.
2-33
Application Embedding: A Language Approach to Declarative Web Programming
B.4 State of the Fish-Tank
The state of the fish-tank is represented by a tuple (S, T, Q), where S ⊆ {clause} is a set
of (static) clauses ({clause} here denotes the set of all clauses), T ⊆ {axiom} × Z is a
set of tuples (α, n) where α is an axiom and n is an integer indicating how many times
α exists in the fish-tank, and Q ∈({axiom} × Z)∗ is a sequence of such (α, n) pairs.
Adding and removing occurrences of axioms to and from sets is done using the add
function, defined as follows:
(α, n) ∈ T
add1
add (T, α, m) = T \ (α, n) ∪ {(α, n + m)}
∀n ∈ Z, (α, n) ∈
/T
add2
add (T, α, m) = T ∪ {(α, m)}
An insertion of axiom α to the tank is done by appending a request to add the axiom
with multiplicity 1 to Q:
ins
ins(α)
(S, T, Q) ⇒ (S, T, 〈Q, (α, 1)〉)
Similarly, removal of axiom α from the tank is appending a request to subtract 1 from
α’s value:
rem
rem(α)
(S, T, Q) ⇒ (S, T, 〈Q, (α, −1)〉)
The fish-tank is updated spontaneously, removing the first element from Q and
applying it to T . During that process, derived axioms are identified and placed back
to Q.
Q = (α, n) , Q0
Q00 = 〈(γ, n × m) | (β, m) ∈ seq (T ) ∧ 〈α, β, S〉 → γ〉
(S, T, Q) ⇒ (S, add (T, α, n) , 〈Q0 , Q00 〉)
tick
The interesting part of the tick rule above is the calculation of the suffix Q00 we add to
Q. We calculate it using sequence comprehension. To calculate Q00 we first look for all
elements of the form (β, m) in T , where β can interact with α to produce γ. The seq
function here turns a set into a sequence, choosing an arbitrary order. Each derived
axiom γ is added with multiplicity n × m, representing the amount of occurrences it
would have if n occurrences of α interacted with m occurrences of β . The tick rule is
applied continuously, as long as Q is not empty, updating T gradually.
B.5 Dynamic Goal Semantics
The state of the fish-tank is visible from the outside by applying queries on dynamic
goals. We use the judgment (S, T, Q) |= G d to denote that dynamic goal G d succeeds
under state (S, T, Q).
A dynamic goal referencing a static goal succeeds if and only if the static goal
succeeds under S :
S |= G
querys
(S, T, Q) |= s (G)
2-34
D. H. Lorenz and B. Rosenan
A conjunction of dynamic goals succeeds if both goals succeed under the tank’s state:
(S, T, Q) |= G1d
(S, T, Q) |= G2d
(S, T, Q) |= G1d ∧ G2d
query∧
Here too we give big-step operational semantics, leaving the order of execution
undefined for simplicity of presentation. Here too the operational semantics of Prolog
provides a good reference for this.
Atomic goals depend on the existence of matching dynamic clauses in T . For such
clauses we check that their multiplicity is positive (i.e., they actually exist in the tank)
and that their body holds under the tank’s state.
G1d ← G2d , n ∈ T n > 0 (S, T, Q) |= G2d
query←
(S, T, Q) |= G1d
2-35
Application Embedding: A Language Approach to Declarative Web Programming
Figure 13 The abstract syntax of L 〈Tweetlog〉 as a collection of Cedalion type declarations
C
Code Snippets from the Implementation of Tweetlog
The abstract syntax of L 〈Tweetlog〉 as a collection of Cedalion type declarations is
shown in figure 13, matching the order of definitions in appendix A.1. Part of the
semantics of L 〈Tweetlog〉 defining the tokenization of tweets is shown in figure 14.
2-36
D. H. Lorenz and B. Rosenan
Figure 14
1
2
3
4
5
6
7
8
9
10
11
12
13
14
15
16
17
18
19
Part of the semantics of L 〈Tweetlog〉 defining the tokenization of tweets. Based
on an existing Cedalion DSL for parsing and tokenization, with semantics similar
to the one depicted in figure 12.
1
2
3
4
5
6
7
8
9
10
11
12
:− dynamic edge/2.
go :−
read(Cmd),
dispatch(Cmd, Out),
write(Out), nl,
go.
dispatch(add_edge(X, Y), ok) :−
assert(edge(X, Y)).
dispatch(from(X), R) :−
findall(Y, reachable(X, Y), R).
dispatch(to(Y), R) :−
findall(X, reachable(X, Y), R).
|: add_edge(a, b).
ok
|: add_edge(b, c).
ok
|: add_edge(a, e).
ok
|: to(e).
[e,a]
|: from(e).
[e]
|: from(a).
[a,b,c,e]
Figure 16 Usage example of the
program in figure 15
reachable(X, X).
reachable(X, Z) :−
edge(X, Y),
reachable(Y, Z).
Figure 15 Prolog program that
modifies itself as means
of maintaining state
D
Program Mutation in Prolog
Prolog support program mutation as a common pattern for storing state. The Prolog
program in figure 15 defines the imperative predicate go/0, which implements a ReadEvaluate-Print-Repeat (REPL) cycle, allowing users to build and query directed graphs.
The command add_edge(X, Y) adds an edge to a graph from X to Y, implicitly adding
vertexes X and Y if they do not exist. The commands from(X) and to(X) print vertexes
with paths from or to X, respectively. In doing so their respective implementations
consult the predicate reachable/2, which succeeds for all reachable node pairs in the
graph. See figure 16 for example of how this program is used.
When handling add_edge(X, Y), the program uses assert/1 to add a fact of the form
edge(X, Y) to the program, effectively modifying the program. reachable/2 is a declarative
predicate that consults the mutable (dynamic) edge/2 predicate as part of its logic.
2-37
Application Embedding: A Language Approach to Declarative Web Programming
About the authors
David H. Lorenz is an Associate Professor in the Department
of Mathematics and Computer Science at the Open University
of Israel. He is currently a Visiting Professor at the Faculty of
Computer Science, Technion—Israel Institute of Technology. His
research interests include aspect-oriented software engineering,
modularity, and programming, particularly involving domain specific languages. Prof. Lorenz received his PhD in computer science
from the Technion—Israel Institute of Technology. He is a member
of the ACM and a member of the IEEE. Contact him at the Dept.
of Mathematics and Computer Science, Open University of Israel,
Raanana 4310701, Israel; [email protected].
Boaz Rosenan is a PhD Candidate at the University of Haifa, under
the supervision of Prof. David H. Lorenz. His research interests
include language workbenches and domain specific languages. He
received his BA in computer science from the Technion—Israel
Institute of Technologies and his MSc in computer science from
the Open University of Israel. He is a member of the ACM. Contact
him at the Dept. Computer Science, University of Haifa, Haifa
3498838, Israel; [email protected].
2-38
| 6cs.PL
|
A Nearly Instance Optimal Algorithm for Top-k
Ranking under the Multinomial Logit Model
arXiv:1707.08238v2 [cs.DS] 30 Jul 2017
Xi Chen
∗
Yuanzhi Li
†
Jieming Mao
‡
August 1, 2017
Abstract
We study the active learning problem of top-k ranking from multi-wise comparisons
under the popular multinomial logit model. Our goal is to identify the top-k items
with high probability by adaptively querying sets for comparisons and observing the
noisy output of the most preferred item from each comparison. To achieve this goal,
we design a new active ranking algorithm without using any information about the
underlying items’ preference scores. We also establish a matching lower bound on the
sample complexity even when the set of preference scores is given to the algorithm.
These two results together show that the proposed algorithm is nearly instance optimal
(similar to instance optimal [FLN03], but up to polylog factors). Our work extends the
existing literature on rank aggregation in three directions. First, instead of studying a
static problem with fixed data, we investigate the top-k ranking problem in an active
learning setting. Second, we show our algorithm is nearly instance optimal, which is
a much stronger theoretical guarantee. Finally, we extend the pairwise comparison to
the multi-wise comparison, which has not been fully explored in ranking literature.
1
Introduction
The problem of inferring a ranking over a set of n items (e.g., products, movies, URLs) is
an important problem in machine learning and finds numerous applications in recommender
systems, web search, social choice, and many other areas. To learn the global ranking, an
effective way is to present at most l (l ≥ 2) items at each time and ask about the most
favorable item among the given items. Then, the answers from these multi-wise comparisons
will be aggregated to infer the global ranking. When the number of items n becomes large,
instead of inferring the global ranking over all the n items, it is of more interest to identify
∗
Stern School of Business, New York University, email: [email protected]
Department of Computer Science, Princeton University, email: [email protected]
‡
Department of Computer Science, Princeton University, email: [email protected]
†
1
the top-k items with a pre-specified k. In this paper, we study the problem of active topk ranking from multi-wise comparisons, where the goal is to adaptively choose at most l
items for each comparison and accurately infer the top-k items with the minimum number
of comparisons (i.e., the minimum sample complexity). As an illustration, let us consider
a practical scenario: an online retailer is facing the problem of choosing k best designs of
handbags among n candidate designs. One popular way is to display several designs to each
arriving customer and observe which handbag is chosen. Since a shopping website has a
capacity on the maximum number of display spots, each comparison will involve at most l
possible designs.
Given the wide application of top-k ranking, this problem has received a lot of attention
in recent years, e.g., [SW15, STZ17] (please see Section 1.4 for more details). Our work
greatly extends the existing literature on top-k ranking in the following three directions:
1. Most existing work studies a non-active ranking aggregation problem, where the answers of comparisons are provided statically or the items for each comparison are chosen
completely at random. Instead of considering a passive ranking setup, we propose an
active ranking algorithm, which adaptively chooses the items for comparisons based
on previously collected information.
2. Most existing work chooses some specific function (call this function f ) of problem
parameters (e.g., n, k, l and preference scores) and shows that the algorithm’s sample
complexity is at most f . For the optimality, they also show that for any value of f , there
exists an instance whose sample complexity equals to that value and any algorithm
needs at least Ω(f ) comparisons on this instance. However, this type of algorithms
could perform poorly on some instances other than those instances for establishing
lower bounds (see examples from [CGMS17]); and the form of function f can vary the
designed algorithm a lot.
To address this issue, we establish a much more refined upper bound on the sample
complexity. The derived sample complexity matches the lower bound when all the
parameters (including the set of underlying preference scores for items) are given to the
algorithm. They together show that our lower bound is tight and also our algorithm is
nearly instance optimal (see Definition 1.1 for the definition of nearly instance optimal).
3. Existing work mainly focuses on pairwise comparisons. We extend the pairwise comparison to the multi-wise comparison (at most l items) and further quantify the role
of l in the sample complexity. From our sample complexity result (see Section 1.2), we
show that the pairwise comparison could be as helpful as multi-wise comparison unless
the underlying instance is very easy.
1.1
Model
In this paper, we adopt the widely used multinomial logit (MNL) model [Luc59, McF73,
Tra03] for modeling multi-wise comparisons. In particular, we assume that each item i
2
has an underlying preference score (a.k.a. utility in economics) µi for i = 1, . . . , n. These
scores, which are unknown to the algorithm, determine the underlying ranking of the items.
Specifically, µi > µj means that item i is preferred to item j and item i should have a a higher
rank. Without loss of generality, we assume that µ1 ≥ µ2 ≥ · · · µk > µk+1 ≥ · · · ≥ µn , and
thus the true top-k items are {1, . . . , k}. At each time t from 1 to T , the algorithm chooses
a subset of items with at least two items, denoted by St ⊆ {1, ..., n}, for query/comparison.
The size of the set St is upper bounded by a pre-fixed parameter l, i.e., 2 ≤ |St | ≤ l.
Given the set St , the agent will report her most preferred item a ∈ St following the
multinomial logit (MNL) model:
exp(µa )
.
j∈St exp(µj )
Pr[a|St ] = P
(1)
When the size of St is 2 (i.e., l = 2), the MNL model reduces to Bradley-Terry model [BT52],
which has been widely studied in rank aggregation literature in machine learning (see, e.g.,
[NOS17, JKSO13, RA14, CS15]).
In fact, the MNL model has a simple probabilistic interpretation as follows [Tra03].
Given the set St , the agent draws her valuation νj = µj + ǫj for each item j ∈ St , where µj
is the mean utility for item j and each ǫj is independently, identically distributed random
variable following the Gumbel distribution. Then, the probability that a ∈ St is chosen as
the most favorable item is Pr (νa ≥ νj , ∀j ∈ St \{a}). With some simple algebraic derivation
using the density of Gumbel distribution (see Chapter 3.1 in [Tra03]), the choice probability
Pr (νa ≥ νj , ∀j ∈ St \{a}) has an explicit expression in (1). For notational convenience, we
define θj = exp(µj ) for i = 1, ..., n, and the choice probability in (1) can be equivalently
written as Pr[a|St ] = P θa θj . By adaptively querying the set St for 1 ≤ t ≤ T and observing
j∈St
the reported most favorable item in St , the goal is to identify the set of top-k items with
high probability using the minimum number of queries.
For notation convenience, we assume the i-th item (with the preference score θi ) is labeled
as πi ∈ {1, . . . , n} by the algorithm at the beginning. Since the algorithm has no prior
knowledge on the ranking of items before it makes any comparison, the ranking of the
items should have no correlation with the labels of the items. Therefore, π = (π1 , ..., πn ) is
distributed as a uniform permutation of {1, ..., n}.
The notion of instance optimal was originally defined and emphasized as an important
concept in [FLN03]. With the MNL model in place, we provide a formal definition of nearly
instance optimal in our problem. To get a definition of instance optimal in our problem, we
can just replace Õ with O in Definition 1.1. The “nearly” here just means we allow polylog
factors.
Definition 1.1 (Nearly Instance Optimal). Given instance (n, k, l, θ1 , ..., θn ), define
c(n, k, l, θ1 , ..., θn ) to be the sample complexity of an optimal adaptive algorithm on the instance. We say that an algorithm A is nearly instance optimal, if for any instance (n, k, l, θ1 , ..., θn ),
the algorithm A outputs the top-k items with high probability and only uses at most
Õ(c(n, k, l, θ1 , ..., θn )) number of comparisons. (Note that Õ(·) hides polylog factors of n and
1
.)
θk −θk+1
3
1.2
Main results
Under the MNL model described in Section 1.1, the main results of this paper include the
following upper and lower bounds on the sample complexity.
Theorem 1.2. We design an active ranking algorithm which uses
P
X
X
n
θk2
i≥k+1 θi
+k+
+
+
Õ
l
θk
(θk − θi )2 i:i≤k,θ ≤2θ
θ
i≥k+1,θi ≥
k
2
i
k+1
2
θk+1
(θk+1 − θi )2
!
comparisons with the set size at most l (can be 2-wise, 3-wise,...,l-wise comparisons) to
identify the top-k items with high probability.
We note that in Theorem 1.2, the notation Õ(·) hides polylog factors of n and θk −θ1 k+1 .
Next, we present a matching lower bound result, which shows that our sample complexity
in Theorem 1.2 is nearly instance optimal.
Theorem 1.3. For any (possibly active) ranking algorithm A, suppose that A uses comparisons of set size at most l. Even when the algorithm A is given the values of {θ1 , ..., θn } (note
that A does not know which item takes the preference score θi for each i), A still needs
!
P
2
2
X
X
θ
θ
θ
n
i≥k+1 i
k+1
k
+k+
+
+
Ω
2
l
θk
(θk − θi )
(θk+1 − θi )2
θ
i:i≤k,θ ≤2θ
i≥k+1,θi ≥
k
2
i
k+1
comparisons to identify the top-k items with probability at least 7/8.
Here we give some intuitive explanations of the terms in the above bounds before introducing the proof overview:
1. Term nl : Since each comparison has size at most l, we need at least
query each item at least once.
n
l
comparisons to
2. Term k: As the proof will suggest, in order to find the top-k items, we need to observe
most items in the top-k set as chosen items from comparisons. However, we do not
have to observe most items in the bottom-(n − k) set. Therefore, there is no term n − k
in the bound.
P
2
P
P
θi
θk+1
θk2
3. Term i≥k+1
+
+
θk
2
i:i≤k,θi ≤2θk+1 (θk+1 −θi )2 : Roughly speaking,
θk
i≥k+1,θi ≥ 2 (θk −θi )
2
i)
is the amount of information that the compariwhen i > k and θi ≥ θk /2, Θ (θk −θ
θk2
2
θk
son between item i and item k reveals. So intuitively, we need Ω (θk −θ
to tell that
2
i)
item i ranks after item k. Other quantity can also be understood from an information
theoretic perspective.
4
It is also worthwhile to note that when l is a constant, it’s easy to check that
P
2
X
X
θk+1
n
θk2
i≥k+1 θi
+k+
+
+
l
θk
(θk − θi )2 i:i≤k,θ ≤2θ (θk+1 − θi )2
θ
i≥k+1,θi ≥
= O
n
X
i=k+1
k
2
i
k+1
!
k
X
θi2
θk2
+
.
(θk − θi )2 i=1 (θk+1 − θi )2
This is a simpler expression of the instance optimal sample complexity when l is a constant.
Based on the sample complexity results in Theorem 1.2 and 1.3, we summarize the main
theoretical contribution of this paper:
1. We design an active ranking algorithm for identifying top-k items under the popular
MNL model. We further prove a matching lower bound, which establishes that the
proposed algorithm is nearly instance optimal.
2. Our result shows that the improvement of the multi-wise comparison over the pairwise
comparison depends on the difficulty of the underlying instance. Note that the only
term in the sample complexity involving l is nl . Therefore, the multi-wise comparison
makes a significant difference from the pairwise comparison only when nl is the leading
term in the sample complexity.
Therefore,
unless the underlying instance is really easy (e.g., the instance-adaptive term
P
2
P
P
θk+1
θk2
i≥k+1 θi
k+
+
+
θk
2
2 is o(n). One implication
i:i≤k,θ
≤2θ
θk
i≥k+1,θi ≥ 2 (θk −θi )
i
k+1 (θk+1 −θi )
is that most of the θi ’s among θk+1 , ..., θn are much smaller than θk ), the pairwise
comparison is as helpful as the multi-wise comparison.
1.3
Proof overview
In this section, we give some very high level overviews of how we prove Theorem 1.2 and
Theorem 1.3.
1.3.1
Algorithms
To prove Theorem 1.2, we consider two separate cases: l = O(log n) or l = Ω(log n).
1. In the first case, by losing a log-factor, we can just focus on only using pairwise comparisons. Our algorithm first randomly select Õ(n) pairs and proceed by querying all
of them once per iteration. After getting the query results, by a standard binomial
concentration bound, we are able to construct a confident interval of θθji for each pair
(i, j) selected by the algorithm in the beginning. In a high level, our algorithm goes by
declaring θi ≥ θj for pair i, j, if the lower bound of the corresponding confident interval
is bigger or equal to 1, or if there already exists d items (i = i1 ), i2 , . . . , (id = j) such
that we have already declared θir ≥ θir+1 for all r ∈ [d − 1]. We are able to show that, if
5
θ1 ≥ θ2 ≥ · · · ≥ θn , thenfor all i, j ∈ [n] with j ≥ i + n4 , the
algorithm will successfully
2
2
Pn
P
θk
θi
k
declare θi ≥ θj after O
many total queries. Thus,
i=k+1 (θk −θi )2 +
i=1 (θk+1 −θi )2
n
we can remove at least 4 items and recurse on a smaller set.
2. The more interesting case is when l = Ω(log n). As P
we have argued before, it is only
P
θi
θk2
+
+
beneficial to use multi-wise comparisons when k + i≥k+1
θk
θk
i≥k+1,θi ≥ 2 (θk −θi )2
P
2
P
θk+1
i≥k+1 θi
= o(n − k) and therefore
i:i≤k,θi ≤2θk+1 (θk+1 −θi )2 = o(n). This implies that
θk
among θk+1 , ..., θn , there are more than half of θi ’s whose value is smaller than some
constant fraction of θk . Thus, intuitively, if we select a random subset of items that
contains θk and keep querying this set, then, instead of seeing all items in this set with
roughly equal probability, we will be seeing item k much more often than the median
of frequencies of items in the set. Thus, our algorithm can select an item if it “appears
very often when querying a set containing it”. We will show that, if the number of
total queries is
!
P
2
X
X
θk+1
θk2
n
i≥k+1 θi
Ω
,
+k+
+
+
l
θk
(θk − θi )2 i:i≤k,θ ≤2θ (θk+1 − θi )2
θ
i≥k+1,θi ≥
k
2
i
k+1
then we will be able to select all the top k-items while not selecting any of the bottom
n/2 items. Thus, we can remove at least n2 items and recurse on a smaller set.
1.3.2
Lower bounds
To prove Theorem 1.3, we establish several lower bounds and combine them using a simple
averaging argument. Most of these lower bounds follow the following general proof strategy:
1. For a given instance (n, k, l, θ1 , ..., θn ), consider other instances on which no algorithm
can output {π1 , ..., πk } with high probability 1 . For example if we just change θk+1 to
θk , then no algorithm can output {π1 , ..., πk } with probability more than 1/2. This
is because item k and item k + 1 look the same now and thus all the algorithms will
output {π1 , ..., πk } and {π1 , ..., πk−1 , πk+1 } with the same probability in the modified
instance.
2. We then consider a well-designed distribution over these modified instances. We show
that for any algorithm A with not enough comparisons, the transcript of running A
on the original instance distributes very closely to the transcript of running A on the
well-design distribution over modified instances.
3. Finally, since the transcript also includes the output, step 2 will tell us that if A does
not use enough comparisons, then A must fail to output {π1 , ..., πk } with some constant
probability.
1
Recall that πi denotes the initial label of i-th item given as the input to the algorithm, and thus the
true top-k items are labeled by {π1 , ..., πk }.
6
1.4
Related Works
Rank aggregation from pairwise comparisons is an important problem in computer science,
which has been widely studied under different comparison models. Most existing works
focus on the non-active setting: the pairs of items for comparisons are fixed (or chosen completely at random) and the algorithm cannot adaptively choose the next pair for querying.
In this non-active ranking setup, when the goal is to obtain a global ranking over all the
items, Negahban et al. [NOS17] proposed the RankCentrality algorithm under the popular
Bradley-Terry model, which is a special case of the MNL model for pairwise comparisons.
Lu and Boutilier [LB11] proposed a ranking algorithm under the Mallows model. Rajkumar
and Agarwal [RA14] investigated different statistical assumptions (e.g., generalized low-noise
condition) for guaranteeing to recover the true ranking. Shah et al. [SBGW17] studied the
ranking aggregation under a non-parametric comparison model—strong stochastic transitivity (SST) model, and converted the ranking problem into a matrix estimation problem
under shape-constraints. Most machine learning literature assumes that there is a true global
ranking of items and the output of each pairwise comparison follows a probabilistic model.
Another way of formulating the ranking problem is via the minimum feedback arc set problem on tournaments, which does not assume a true global ranking and aims to find a ranking
that minimizes the number of inconsistent pairs. There is a vast literature on the minimum
feedback arc set problem and here we omit the survey of this direction (please see [KMS07]
and references therein). Due to the increasing number of items, it is practically more useful
to identify the top-k items in many internet applications. Chen and Suh [CS15], Jang et al.
[JKSO13], and Suh et al. [STZ17] proposed various spectral methods for top-k item identification under the BTL model or mixture of BTL models. Shah and Wainwright [SW15]
proposed a counting-based algorithm under the SST model and Chen et al. The notion of
instance optimal was originally defined and emphasized as an important concept in [FLN03]
for identifying the top-k objects from sorted lists. [CGMS17] suggested that notion “instance
optimal” is necessary for rank aggregation from noisy pairwise comparisons in complicated
noise models
√ and further improved [SW15] by proposing an algorithm that has
√ competitive
ratio Θ̃( n) compared to the best algorithm of each instance and proving Θ̃( n) is tight.
In addition to static rank aggregation, active noisy sorting and ranking problems have
received a lot of attentions in recent years. For example, several works [BM08, Ail11, JN11,
WMJ13] studied the active sorting problem from noisy pairwise comparisons and explored
the sample complexity to approximately recover the true ranking in terms of some distance
function (e.g., Kendall’s tau). Chen et al. [CBCTH13] proposed a Bayesian online ranking
algorithm under the mixture of BTL models. Dwork et al. [DKNS01] and Ailon et al.
[ACN08] considered a related Kemeny optimization problem, where the goal is to determine
the total ordering that minimizes the sum of the distances to different permutations. For
top-k identification, Braverman et al. [BMW16] initiated the study of how round complexity
of active algorithms can affect the sample complexity. Szörényi et al. [SBPH15] studied the
case of k = 1 under the BTL model. Heckel et. al. [HSRW16] investigated the active
ranking under a general class of nonparametric models and also established a lower bound
on the number of comparisons for parametric models. A very recent work by Mohajer and
7
Suh [MS16] proposed an active algorithm for top-k identification under a general class of
pairwise comparison models, where the instance difficulty is characterized by the key quantity
mini∈{1,...,k} minj:j>i (pij − 0.5)2 . Here, pij is the probability of item i is preferred over item
j. However, according to our result in Theorem 1.3, the obtained sample complexities in
previous works are not instance optimal. We note that the lower bound result in Theorem 1.3
holds for algorithms even when all the values of θi ’s are known (but without the knowledge of
which item corresponds to which value) and thus characterizes the difficulty of each instance.
Moreover, we study the the multi-wise comparisons, which has not been explored in ranking
aggregation literature but has a wide range applications.
Finally, we note that the top-k ranking problem is related to the best k arm identification
in multi-armed bandit literature [BWV13, JMNB14, ZCL14, CCZZ17]. However, in the
latter problem, the samples are i.i.d. random variables rather than comparisons and the
goal is to identify the top-k distributions with largest means.
2
Algorithm
For notational simplicity, throughout the paper we use the words w.h.p. to denote with
probability 1 − 1/nc for sufficiently large constant c.
2.1
Top-k item identification (For logarithmic l)
For l = O(log n), we can always use pairwise comparisons by losing a polylog factor. Therefore, we only focus on the case when l = 2 in this section.
Before presenting the algorithm, let us first consider a graph G = (V = [n], E) where
each edge is labeled with either ≈l , ≥l , ≤l , >l or <l (see Line 9 in Algorithm 1). Based on
the labeling of edges, we give the following definition of label monotone, which will be used
in Algorithm 1.
Definition 2.1 (Monotone). We call a path i1 → i2 → · · · → id strictly label monotone if:
1. For every j ∈ [d − 1], the edge (ij , ij+1 ) is labeled with either ≈l , ≥l or >l .
2. There exists at least one edge (ij , ij+1) with label >l .
Moreover, we call a path “label monotone” if only property 1 holds.
Theorem 2.2. For every m items with θ1 ≥ θ2 ≥ · · · ≥ θn , Algorithm 1, on given a random
permutation of labels Ω = [n] and k, returns top-k items w.h.p. using
!!
n
k
2
2
X
X
θ
θ
k+1
k
O κ7 · k +
+
2
(θ
(θk+1 − θi )2
k − θi )
i=1
i=k+1
total number of pairwise comparisons.
8
For the page limit, we defer the proof of Theorem 2.2 to Appendix A. We only provide
the pseudocode in Algorithm 1. In Algorithm 1, we note that a different letter m (instead
of n) is used for denoting the set size because we will run the algorithm recursively with
smaller sets. And also notice that the parameter κ = Ω(log2 n) regardless of the value of m.
We also defer our result for superlogarithmic l to Appendix A.
Algorithm 1 AlgPairwise
1:
2:
3:
4:
5:
6:
7:
8:
9:
Parameter: κ = Ω(log2 n).
Input: A set of randomly permuted labels Ω with |Ω| = m, k: number of top items.
Uniformly at random sample s = mκ subsets S1 , · · · Ss of Ω, each of size 2. Associate
these subsets with a graph G = (Ω, E), where each edge eu ∈ E consists of all the vertices
in Su for u ∈ [s].
q = 0, Ωg = ∅, Ωb = ∅, S = ∅.
while true do
q ← q + 1.
Query each set 1 time, obtain in total s query results {Ru,q }u∈[s] . (Ru,q indicates the
reported most favorable item)
P
For all u ∈ [s], for {i, j} = Su , let θ̃i,j = 1q p∈[q] 1Ru,p =i .
For each edge (i, j) ∈ E, we label it as:
h
q i
θ̃
1. i ≈l j if θ̃i,j ∈ 1+41√ κ , 1 + 4 κq
j,i
q
q
q
θ̃i,j
κ
2. i ≥l j if θ̃ ∈ 1 + 4 q , 1 + 32κ κq
j,i
h
q
θ̃
3. i >l j if θ̃i,j ∈ 1 + 32κ κq , ∞
j,i
θ̃
1√
1√
,
4. i ≤l j if i,j ∈
κ
κ
θ̃j,i
5. i <l j if
θ̃i,j
θ̃j,i
1+32κ
q
1+4
q
h
i
1√
∈ 0, 1+32κ κ
q
For every i, j ∈ [m], we call j>>l i if there exists a strictly label monotone path of
length at most κ from j to i.
11:
For each i ∈ [m], if there exists at least k many j ∈ [m] such that j>>l i, then add i
to Ωb . (Ωb is the subset of items that we are sure not in top-k.)
12:
For each i ∈ [m], if there exists at least m − k many j ∈ [m] such that i>>l j, then
add i to Ωg . (Ωg is the subset of items that we are sure in top-k.)
13:
Break if |Ωg ∪ Ωb | ≥ m4 .
10:
14:
15:
Ω′ = Ω − Ωg − Ωb , k ′ = k − |Ωg |, S = S ∪ Ωg ∪ AlgPairwise(Ω′ , k ′ ).
Return S.
9
3
Lower bounds
We will prove lower bounds on the number of comparison used by any algorithm which
identifies top-k items even when the values of preference scores {θ1 , ..., θn } are given to the
algorithm. (The algorithm just do not know which item has which θi ). For the page limit,
all the proofs are deferred to Appendix B.
3.1
Lower bounds for close weights
Theorem 3.1. Assume θk > θk+1 and c < 10−4 . For any algorithm A (can be adaptive),
P
θk2
if A uses c i:k+1≤i,θi≥θk /2 (θk −θ
2 comparisons of any size (can be l-wise comparison for
i)
2 ≤ l ≤ n), then A will identify the top-k items with probability at most 7/8.
Theorem 3.2. Assume θk > θk+1 and c < 4 · 10−4 . For any algorithm A (can be adaptive),
P
θ2
if A uses c i:i≤k,θi≤2θk+1 (θk+1k+1
comparisons of any size (can be l-wise comparison for
−θi )2
2 ≤ l ≤ n), then A will identify the top-k items with probability at most 7/8.
3.2
Lower bounds for arbitrary weights
Theorem
3.3. Assume c < 1/18. For any algorithm A (can be adaptive), if A uses
P
θi
c i:i>k θk comparisons of any size (can be l-wise comparison for 2 ≤ l ≤ n), then A will
identify the top-k items with probability at most 7/8.
Theorem 3.4. For any algorithm A (can be adaptive), if A uses k/4 comparisons of any
size (can be l-wise comparison for 2 ≤ l ≤ n), then A will identify the top-k items with
probability at most 2/3.
Theorem 3.5. Assume c < 1/2. For any algorithm A (can be adaptive), if A uses cnl
comparisons of size at most l (can be 2-wise, 3-wise,...,l-wise comparisons), then A will
identify the top-k items with probability at most 7/8.
3.3
Combining lower bounds
Corollary 3.6 (Restatement of Theorem 1.3). For any algorithm A (can be adaptive),
suppose A uses comparisons of size at most l (can be 2-wise, 3-wise,...,l-wise comparisons).
A needs
!
P
2
2
X
X
θ
θ
n
θk
i≥k+1 i
k+1
Ω
+k+
+
+
2
l
θk
(θk − θi )
(θk+1 − θi )2
θ
i:i≤k,θ ≤2θ
i≥k+1,θi ≥
k
2
i
k+1
to identify the top-k items with probability at least 7/8.
Proof. To prove this corollary, we just need to combine all the results in Theorem 3.1,
Theorem 3.2, Theorem 3.3, Theorem 3.4 and Theorem 3.5. And then use the fact that if
b < a1 + · · · + a5 then there exists i ∈ {1, 2, 3, 4, 5} such that b < 5ai .
10
A
Additional Results and Proofs of Section 2
Throughout the proofs we are going to use the following claim which is a simple fact about
the binomial concentration.
∗
Claim A.1 (Binomial concentration).
For every m
√ ∈ N , every p ∈ [0, 1], suppose X ∼
√
B(m, p), then X ∈ [mp − O( mp log n), mp + O( mp log n)] w.h.p (with high probability
respect to n).
A.1
Top-k item identification (For logarithmic l)
In this section, we prove Theorem 2.2 of Section 2.
q
q
h
i
θi,j κ
θi,j κ
Following Claim A.1, we know that for every (i, j) ∈ E, every q, θ̃i,j ∈ θi,j −
, θi,j +
q
q
w.h.p. W.l.o.g, let us just focus on the case that this bound is satisfied for all (i, j) ∈ E and
every q.
We have the following Lemma about the labelling:
Lemma A.1 (Label). For q = Ω(κ3 ), we have:
1. if θi ≥ θj , then i ≈l j, i ≥l j or i >l j.
q
2. if θi ≥ θj 1 + 128κ κq , then i >l j.
3. if i ≥l j or i ≈l j, then
r
κ
θi ≥ θj 1 − 8
q
4. if i >l j, then
Proof of Lemma A.1.
r
κ
θi ≥ θj 1 + 16κ
q
1. We know that for q = Ω(κ3 ) and θi ≥ θj :
q
q
θi
θi
κ
− θi +θj · q
θi 1 − 2κ
θi +θj
q
θ̃i,j
q
q
≥
≥
θj
θj
2θi θj κ
θ̃j,i
+ θi +θ
· κq
θj +
θi +θj
q
j
q
q
1 − 2κ
1 − 2κ
q
q
q
q
≥
≥
θj
θ
1 + 2κ
+ θj · 2κ
q
θ
q
i
≥
i
1
1+4
q
κ
q
11
q
2. Again by θi ≥ θj 1 + 128κ κq and q = Ω(κ3 ), we know that
therefore, we have:
θ̃i,j
θ̃j,i
θj
θi
q
≤ 1 − 64κ κq ,
q
q
q
2κ
2κ
r
1
−
1 − 2κ
1
−
q
q
q
κ
q ≥
q ≥ 1 + 32κ
q
.
≥
≥
q
θj
θj
θj 2κ
2κ
κ
2κ
q
+
1 − 64κ q +
+ θi · q
θi
q
q
θi
3. Let us suppose θi ≤ θj , otherwise we already complete the proof. Now, we have:
q
θi
θi
+
· κq
θi +θj
θi +θj
θ̃i,j
1
q
q
≥
≥
θj
θj
κ
κ
θ̃j,i
1
+
4
−
·
q
θi +θj
θi +θj
q
Which implies that
q
2θi θj κ
θi +
1
q
q
q ≥
κ
2κ
1
+
4
θj 1 −
q
q
Therefore, by θi ≤ θj , we have:
θi
θj
+
1−
q
q
2κ
q
2κ
q
≥
1
1+4
q
κ
q
Which implies that
q
r
r
1 − 2κ
q
θi
2κ
κ
q −
≥
≥1−8
κ
θj
q
q
1+4 q
4. Let us suppose θi ≤ 2θj , otherwise we already complete the proof. Again, we have:
q
θi
r
+ 3κ
θj
q
κ
q
≥ 1 + 32κ
q
1 − 3κ
q
Which implies that
θi
≥
θj
r
r r
r
3κ
κ
2κ
κ
1−
1 + 32κ
−
≥ 1 + 16κ
q
q
q
q
12
Above, the Lemma A.1 implies that w.h.p. the labelling of each edge (i, j) is consistent
with the order of θi , θj . Now, the algorithm will declare i>>l j if there exists strictly label
monotone path from i to j. Using the Lemma above we can show that if such path exists,
then θi > θj . To show the other direction that such paths exists when θj > θi , we first
consider the following graph Lemma that gives the exists of monotone path in random graph
G(m, p).
Lemma A.2 (Graph Path). For every m ≤ n, every random graph G(m, p) on vertices
κ
V = [m], if p ≥ m
, then w.h.p. For every i, j ∈ [m] with j ≥ i + m4 , there exists a path
i = i1 → i2 → · · · id = j such that
1. d ≤ κ.
2. ir ≤ ir+1 for every r ∈ [d − 1].
We call such a path a monotone path from i to j.
√
Proof of Lemma A.2. It is sufficient to consider the case when m = Ω( κ), otherwise w.h.p.
the graph is a complete graph and theorem is automatically true.
We consider a sequential way of generating G(m, p): At each time t = 1, 2, · · · , m, a
vertex t arrives and there exists an edge between t and each
t′ ∈ [t − 1] with probability p.
√
Let us consider a fixed i ≤ 34 m and j ≥ i + m4 . Let τ = 32κ = Ω(log n). We will divide the
set {i, i + 1, · · · , j − 1} into τ subsets H1 , · · · , Hτ such that
j−i
j−i
j−i
, i + (r − 1)
+ 1, · · · , i + r
−1
Hr = i + (r − 1)
τ
τ
τ
√
Since m = Ω( κ) we know that |Hr | ≥ 1.
Let us define the random variable Yr , Xr , Zv as:
Yr = the set of all v ∈ Hr such that there exists a monotone path from i to v of length at most r
and Xr = |Yr |.
For each v ∈ Hr+1 , we define
Zv = 1there is an edge between v and at least one vertex in Yr
Clearly, X1 ≥ 1 and each Zv is i.i.d. random variable in {0, 1} with Pr[Zv = 0 | Xr ] =
(1 − p)Xr . On the other hand, by definition,
X
Zv
Xr+1 ≥
v∈Hr+1
We consider two cases:
1. Xr ≥ p1 , then Pr[Zv = 1] ≥ 14 .
13
2. Xt < 1p , then by (1 − p)x ≤ 1 −
xp
2
for x < 1/p, we have Pr[Zv = 1] ≥
pXr
.
2
Consider a fixed Xr and for each v ∈ Hr+1 , let Zv be the random variable. By standard
Chernoff bound, we have:
1. If Xr ≥ 1p , then w.h.p. Xr+1 ≥
j−i
4τ
2. 1 ≤ Xr < 1p , then w.h.p. Xr+1 ≥
Recall that p ≥
κ
m
and j − i ≥
r
τ
≥
m
.
16τ
(j−i)pXr
2τ
m
,
4
−
q
r
.
τ (j−i)pX
2τ
therefore,
(j − i)pXr
(j − i)pXr
≤
,
2τ
4τ
(j − i)pXr
≥ 2Xr
4τ
Which implies that w.h.p. Xr+1 ≥ 2Xr .
m
. Therefore,
Putting everything together, we know that for τ = Ω(log n), w.h.p. Xτ ≥ 16τ
condition on this event, by
m
m
Pr there is an edge between j and Yτ Xτ ≥
= 1 − (1 − p) 16τ
16τ
m
m
2 16τ
1
κ 16τ
1024τ
≥ 1 − Ω(1)
≥1− 1−
≥1− 1−
m
m
n
We complete the proof.
Having this Lemma, we can present the main Lemma above the algorithm:
Lemma A.3 (Main 3). Suppose q = Ω(κ3 ), then w.h.p. the following holds:
1. Ωg ⊆ [k], Ωb ∩ [k] = ∅.
5 P
2. If k ≤ m2 and q = Ω κm · m
i=k+1
3. If k >
m
2
and q = Ω
κ5
m
θk2
(θk −θi )2
P
· k + ki=1
, then |Ωb | ≥
2
θk+1
(θk+1 −θi )2
m
.
4
, then |Ωg | ≥
m
.
4
Since the algorithm terminates within O(κ) recursions, moreover, in each recursion, the
algorithm makes at most κmq queries. Therefore, Lemma A.3 implies that the algorithm
runs in total queries:
!!
m
k
2
2
X
X
θ
θ
k+1
k
O κ7 · k +
+
2
(θ
(θk+1 − θi )2
k − θi )
i=1
i=k+1
14
Proof of Lemma A.3.
1. It suffices to show that if i>>l j, then θi ≥ θj . To see this, consider
a strictly label monotone path i = i1 → i2 → · · · → id = j with length d ≤ κ. By
q
Lemma A.1, we know that for every r ∈ [d − 1], we have: θir ≥ θir+1 1 − 8 κq .
q
Moreover, there exists an r ′ ∈ [d − 1] such that θir′ ≥ θir′ +1 1 + 16κ κq . Multiply
every thing together, we know that
r κ−1
r
κ
κ
θi ≥ θj 1 − 8
1 + 16κ
≥ θj
q
q
2. Let us denote the set H = { 34 m + 1, 34 m + 2, · · · m}, we will prove that H ⊆ Ωb .
Consider one j ∈ H, by Lemma A.2, w.h.p. for every i ∈ [k], there exists a path
i = i1 → i2 → · · · id = j of length at most κ such that θir ≥ θir+1 for every r ∈ [d − 1].
Now, by Lemma A.1, we know that this path is label monotone. We just need to show
that this path is strictly label monotone. To see this, we know that there exists one
r ′ ∈ [d − 1] such that
1/κ
θi
θir′ ≥ θir′ +1
θj
Let ν =
1
m−k
θk2
i=k+1 (θk −θi )2 .
Pm
Now, since k ≤
conclude that for this j ∈ H and i ∈ [k],
θi2
m
,
2
(θi −θj )2
we can apply Markov inequality and
≤
1
θi
q ≥1+
≥
θj
1
1 − 2ν
θk2
(θk −θj )2
r
≤ 2ν. Which implies that
1
2ν
For q = Ω(κ5 ν), we know that
θi
θj
Therefore, θir′ ≥ θir′ +1
ir′ +1 and thus i>>l j.
1/κ
≥
1/κ
θi
θj
r 1/κ
r
κ
κ
1 + 64κ
≥ 1 + 32κ
q
q
2
q
≥ 1 + 32κ κq . By definition, we shall label ir′ >
3. It can be shown with exactly the same calculation as 2 with H = 1, 2, · · · 14 m and
apply Markov inequality on
!
k
k
2
X
θk+1
θi2
1
1X
= O k+
ν=
k i=1 (θk+1 − θi )2
k
(θk+1 − θi )2
i=1
15
A.2
Top-k item identification (For super logarithmic l)
Before presenting the algorithm, we first argue about which case using bigger l is unnecessary.
We have the following Claim:
Claim A.2 (Bigger l). For every l ≤ m ≤ n, we have:
k
2
X
θk+1
θk2
+
k+
(θk − θi )2 i=1 (θk+1 − θi )2
i=k+1
P
k
2
X
X
θk+1
θk2
m
i≥k+1 θi
≤ +k+
+
+ 4m
+
2
2
l
θk
(θ
−
θ
)
(θ
−
θ
)
k
i
k+1
i
θ
i=1
m
X
i≥k+1,θi ≥
k
2
Therefore, as long as we can show one of the following:
1. k = Ω(m).
2.
P
i≥k+1 θi
θk
= Ω(m).
We can just use the algorithm for l = 2. Otherwise, we shall consider larger l, we will
directly considering the case when l = Ω(log n). Before giving the algorithm, it is convenient
to first consider the following query procedure: For a fixed Q, do:
Algorithm 2 BasicQuery
Parameter: κ = Ω(log2 n)
Input: Ω: set of items with |Ω| = m, k: number of top items to find. l: size of the
subset to query.
3: Uniformly at random sample s = mκ
subsets S1 , · · · Ss of Ω, each of size l. Associate
l
these subsets with a hypergraph G = (Ω, E), where each edge eu ∈ E is consists of all
the vertices in Su for u ∈ [s].
4: Query each set Q time, obtain in total sQ query results {Ru,q }u∈[s],q∈[Q].
1:
2:
16
For a fixed q ≤ Q, let us consider a random variable θ̃i,Su ∈ [0, 1] defined as
θ̃i,Su =
1X
1Ru,r =i
q
r∈[q]
For each i, u such that i ∈ Su , let us define 0-1 valued function 1i,u,α,β,γ such that
1i,u,α,β,γ = 1 if and only if all the following conditions hold:
1. θ̃i,Su ≥ αq .
2. There exists at least γl many of the j ∈ Su such that θ̃j,Su ≤ β θ̃i,Su .
We also consider the random variable Xi,u,α,β,γ associated with this function, where the
randomness is taken over the uniformly at random choice of Su conditional on i ∈ Su , and
the randomness of the outcome of the queries.
We prove the following main Lemma:
1 1
, 2 , β ∈ (0, 32], α = Ω(κ). For every i ∈ [m], the
Lemma A.4 (Indicator). Let γ ∈ 32
following holds:
P
2αl j∈[m] θj
1. If q = Ω α +
and θi ≥ 2βθ(1−2γ)m , then
mθi
Pr[Xi,u,α,β,γ = 1] ≥
15
16
2. For every q, if θi ≤ β2 θ(1−γ)m , then
Pr[Xi,u,α,β,γ = 1] ≤
Proof of Lemma A.4.
9
16
1. We first bound the probability that θ̃i,Su ≥ αq . By
s
"
s
θi,Su ≥
2α
α
=⇒ θ̃i,Su ≥
q
q
θ̃i,Su ∈ θi,Su −
we know that
θi,Su κ
, θi,Su +
q
θi,Su κ
q
#
To lower bound this probability, we just need to consider the probability that θi,Su <
We apply Markov inequality and have that:
i
h
E 2α
qθi,Su
2α
2α
= Pr
>1 <
Pr θi,Su <
q
qθi,Su
1
17
2α
.
q
Notice that
Therefore,
P
2α j∈Su θj
2α
=
qθi,Su
θi q
P
P
2αE[ j∈Su ]θj
2α
2α 2αl j∈[m] θj
1
E
≤
+
≤
=
qθi,Su
θi q
q
mθi q
64
Putting together we obtain
2α
1
α
≤ Pr θi,Su <
<
Pr θ̃i,Su <
q
q
64
Now we can move to the second condition. For now, suppose θ̃i,Su ≥
know that
31
33
θ̃i,Su ∈
θi,S , θi,S
32 u 32 u
α
q
holds, we then
Therefore, θi ≥ 2βθ(1−2γ)m implies that for every j ∈ H = {(1 − 2γ)m, (1 − 2γ)m +
1, · · · m} with j ∈ Su , we have:
s
s
θj,Su κ
θi,Su κ
θ̃j,Su ≤ θj,Su +
≤ θj,Su +
q
q
≤ θj,Su +
3θi,Su
4β
θ̃i,Su
≤
β
≤
1
θi,Su
1
θi,Su ≤
+
θi,S
128
2β
128 u
for β ≤ 32
1
. Therefore,
Since |H| = 2γm, we know that for l = Ω(log n), Pr[|H ∩ Su | < γl] ≤ 64
α
15
Pr[Xi,u,α,β,γ = 1] ≥ 1 − Pr θ̃i,Su <
− Pr[|H ∩ Su | < γl] ≥
q
16
2. The proof follows from the same calculation. Notice that this time we already have
Xi,u,α,β,γ = 1 =⇒ θ̃i,Su ≥ αq .
1 1
, 2 and every τ ∈ 43 , 78 , we consider set
For fixed α = Ω(κ), every β ∈ (0, 32], γ ∈ 32
X
Xi,u,α,β,γ ≥ τ deg(i)
Ωβ,γ,τ = i ∈ [m]
u:u∈[s],i∈Su
We also have the following Corollary of Lemma A.4:
18
33 7
, 8 , w.h.p. j ∈
Corollary A.5.
1. For every i, j ∈ [m] with θi ≥ θj , every τ ∈ 34 × 32
Ωβ,γ,τ =⇒ i ∈ Ωβ,γ, 32 τ .
33
P
2αl j∈[m] θj
2. For every i ∈ [m], if θi ≥ 2βθ(1−2γ)m and q = Ω α +
, then w.h.p. i ∈
mθi
Ωβ,γ,τ .
3. For every q, if i ∈ Ωβ,γ,τ , then w.h.p. θi ≥ β2 θ(1−γ)m .
Having this Corollary, we can do the following algorithm that selects all the θi ≥
32 max{θk , θ 3 m } and removes most of the θj ≤ 14 θk :
4
Algorithm 3 AlgMulti-wise
1:
2:
3:
4:
5:
6:
7:
8:
9:
10:
11:
12:
Parameter: κ = O(log2 n).
Input: Ω: set of items, k: number of top items to find.
Output: S: set of top items. Ω′ : set of remaining items.
Initialization: S = ∅, Ω′ = Ω, m = |Ω|.
Call BasicQuery to obtain {θ̃i,Su }i∈[n],u∈[s].
if k ≤ 21 m then
if 1 ≤ |Ω32, 1 , 13 | and |Ω4, 1 , 13 | < k then
4 16
16 16
S1 = Ω4, 1 , 7 , Ω′′ = Ω − S1 , (S ′ , Ω′ ) = AlgMulti-wise(Ω′′ , k − |S1 |, R), S = S ∪ S1 ∪
16 8
S ′′ .
13
33 2 3
·
≥
· 4.
Notice that we pick those numbers so 78 ≥ 33
32 16
32
else if |Ω4, 1 , 13 | ≥ k then
16 16
Ω′′ = Ω4, 1 , 3 , (S ′ , Ω′ ) = AlgMulti-wise(Ω′′ , k, R), S = S ∪ S ′ .
16 4
′
Return S, Ω .
We have the following lemma.
Lemma A.6. For every m, every θ1 ≥ θ2 ≥ · · · ≥ θm , every k ≤ m, every l ≤ m, Algorithm
3, on given a random permutation of Ω = [m], k satisfies:
1. Output set (S, Ω′ ) of the algorithm satisfies S ⊆ [k].
P
l(k+ j≥k θj )
2. If Q = Ω̃ 1 +
, then the algorithm returns in O(log m) many recursion
mθk
calls, and after the algorithm, let us for simplicity still denote [|Ω′ |] = Ω′ with θ1 ≥
θ2 ≥ · · · θ|Ω′ | and k ′ = k − |S|, then either
(a) For every i ∈ Ω′ , θi ≤ 512θ 7 |Ω′ | .
8
′
(b) Or k ≥
1
|Ω′ |.
2
Proof of the main theorem. After running this algorithm, we can simply apply the algorithm
for l = 2 (By Claim A.2), since one of the following is true:
19
1.
P
i≥k′ +1 θi
θk ′
≥
3
8
×
1
|Ω′ |.
512
2. k ′ > 21 |Ω′ |.
Therefore, putting everything together, we can get the top k items in total number of queries:
P
k
2
X
X
θk+1
θk2
m
i≥k+1 θi
Õ + k +
+
+
2
2
l
θk
(θ
−
θ
)
(θ
−
θ
)
k
i
k+1
i
θ
i=1
i≥k+1,θi ≥
k
2
Now, it just remains to prove this Lemma:
Proof. We first prove the correctness: S ⊆ [k]. We have the following observations:
1. If |Ω4, 1 , 13 | ≥ k, then there must be i ∈ Ω4, 1 , 13 with θi ≤ θk . By Corollary A.5, since
16 16
16 16
13
33 3
1 3.
≥
·
,
we
know
that
[k]
⊆
Ω
4, ,
16
32 4
16 4
2. If there exists j ∈ Ω4, 1 , 7 such that j ∈
/ [k], then by θj ≤ θk , apply Corollary A.5 with
16 8
7
33 13
≥ 32 · 16 , we know that [k] ⊆ Ω4, 1 , 13 , which implies |Ω4, 1 , 13 | ≥ k. Therefore, we
8
16 16
16 16
will not include any item that is not top k to S when recursing from Line 7.
These two observations immediately imply S ⊆ [k].
Now, we will show that for sufficiently large Q, either of the two conditions hold:
1. For every i ∈ Ω′ , θi ≤ 512θ 7 |Ω′ | .
8
2. Or k ′ ≥ 12 |Ω′ |.
Let us for notation simplicity drop the ′ here. Clearly, we just need to consider the case
when k < 12 m, otherwise, the algorithm will just terminate and the second condition is true.
We will first prove that θi ≤ 64θk and then we prove that θk ≤ 8θ 7 m .
8
1. To prove θi ≤ 64θk, we suppose on the contrary that θ1 > 64θk . Apply Corollary A.5
P
2αl( j≥1 θj )
, we have that 1 ∈ Ω32, 1 , 3 , which implies that
with q = Q = Ω α +
mθ1
16 4
|Ω32, 1 , 13 | > 0, so the algorithm won’t terminate, contradict.
4 16
2. Now, we need to show that θk ≤ 8θ 7 n . We also on the contrary suppose that θk >
8
8θ 7 m . Since the algorithm termniates, by the previous claim, we know that in the last
8
recursion, it must be the case that θ1 ≤ 64θk . Which implies that
P
P
2αl
θ
2αl kθk + j≥k θj
j≥1 j
= Ω α +
Q = Ω α +
mθk
mθk
Therefore, if θk > 8θ 7 m , then by Corollary A.5 we know that [k] ⊆ Ω4, 1 , 3 , so the
8
16 4
algorithm won’t terminate.
20
Finally, we consider about the total number of recursions. Clearly, if the algorithm
15
|Ω|. If the algorithm recurses through the
recurses through the second case, then |Ω′ | ≤ 16
first case, then by Corollary A.5, it must be the case that
θ1 ≥ 16θ 7 m
8
Which implies that for all i with θi ≥ θ21 ≥ 8θ 7 m , i ∈ Ω4, 1 , 7 .
8
16 8
Therefore, the total number of recursions of the algorithm is bounded by O(log m). So
the total number of queries of the algorithm is:
P
P
2αl kθk + j≥k θj
θj
κm
m
j≥k
×
× O(log m) = Õ
+k+
O α +
mθk
l
l
θk
Remark: How to obtain the value Q: In the proof above we assumed that we have an
aprior estimation of the value of Q. We can replace this assumption by initially setting Q
to be Q = Q0 = 1, and run algorithm 3 with Q0 queries and then run the algorithm with
pairwise comparision. Once the later algorithm requires more than Q0 × nl queries, then we
stop it, set Q1 = 2Q0 and repeat this procedure with Q = Q1 . We keep on repeating this
for Q2 = 2Q1 , Q3 = 2Q2 , · · · until the later algorithm requires less than Qi × nl queries.
By the Lemma we just proved, the output of the algorithm is correct for every Q. Moreover, if
P
k
2
2
X
X
θk+1
θk
n
n
i≥k+1 θi
Q × = Ω̃ + k +
+
+
2
2
l
l
θk
(θk − θi )
(θk+1 − θi )
θ
i=1
i≥k+1,θi ≥
k
2
Then this process will terminate, and the total query complexity is then bounded by
Õ Q × nl .
B
B.1
Proofs of Section 3
Lower bounds for close weights
Theorem B.1 (Restatement of Theorem 3.1). Assume θk > θk+1 and c < 10−4 . For any
P
θk2
algorithm A (can be adaptive), if A uses c i:k+1≤i,θi≥θk /2 (θk −θ
2 comparisons of any size (can
i)
be l-wise comparison for 2 ≤ l ≤ n), then A will identify the top-k items with probability at
most 7/8.
θ2
k
Proof. For notation convenience, we set wi = (θk −θ
for i such that i ≥ k + 1 and θi ≥ θk /2.
Pn i )2
For other i, we set wi = 0. We also set W = i=1 wi . Then we have T = cW .
21
First of all, we can assume A is deterministic. This is because if A is randomized, we can
fix the randomness string which makes A achieves the highest successful probability.
Let S = (S1 , ..., ST ) be the history of algorithm. Each St is the comparison result of
round t. Notice that since A is deterministic, with S1 , ..., St , we can determine the labels of
items A want to compare in round t + 1 even when A is adaptive. So there is no point to
put the labels of compared items in the history. So we only put the comparison result in the
history, i.e St is a number in [n] and S is a length-T string of numbers in [n].
Again since A is deterministic, the label A outputs is just a deterministic function of S,
we use A(S) to denote it. A outputs correctly if A outputs the label of the top-k items, i.e.
A(S) = {π1 , ..., πk }.
We use p(S, π) to denote the probability that the items are labeled as π and A has history
S. Now consider the case when we set θi equals to θk for i ≥ k+1. In this case the probability
of A(S) = {π1 , ..., πk } should be at most 1/2 as item k and item i have the same weight. We
use pi (S, π) to denote the probability that the items are labeled as π and A has history S
when θi is changed to θk .
Now we prove the following lemma that gives the connection between p(S, π) and pi (S, π).
Lemma B.2. Consider p as a distribution over (π, S). For all c1 > 0, we have
" n
!
#
X wi pi (S, π)
(c1 − 4c)2
Pr
≤ −c1 ≤ exp −
.
ln
(π,S)∼p
W
p(S, π)
72c
i=1
Proof. Define random variable Zt to be the following for t = 1, ..., T when (π, S) is sampled
from distribution p:
n
X
wi pi (S1 ...St , π)
ln
.
Zt =
W
p(S1 ...St , π)
i=1
We have
ZT =
n
X
wi pi (S, π)
ln
.
W
p(S,
π)
i=1
4t
, ..., ZT + 4T
forms a supermartingale.
Now we want to show that sequence 0, Z1 + W4 , ..., Zt + W
W
Suppose
in
round
t,
given
S
,
...,
S
and
π,
Algorithm
A
compares
items in set Ut . Let
1
t−1
P
θ−i = j∈Ut ,j6=i θi . Then we have, with probability θi /(θi + θ−i ),
Zt − Zt−1
X wj
(θk − θi )θ−i
θk − θj
wi
+
ln 1 +
ln 1 −
=
W
(θk + θ−i )θi
W
θk + θ−j
j∈U ,j6=i
t
Here are two simple facts about ln. For 0 ≤ x ≤ 1, ln(1 + x) ≥ x − x2 . For 0 ≤ x ≤ 1/2,
k −θi )θ−i
ln(1 − x) ≥ −x − x2 . It’s easy to check that for i such that wi > 0, we have (θ
≤1
(θk +θ−i )θi
22
and
θk −θi
θk +θ−i
≤ 1/2 . Therefore, by these two facts, for i such that wi > 0, we have
θi
(θk − θi )θ−i
θ−i
θk − θi
wi ln 1 +
+
wi ln 1 −
θi + θ−i
(θk + θ−i )θi
θi + θ−i
θk + θ−i
2 2
(θk − θi )θ−i (θk − θi ) θ−i (θk − θi )θ−i (θk − θi )2 θ−i
wi
−
−
−
≥
θi + θ−i
θk + θ−i
(θk + θ−i )2 θi
θk + θ−i
(θk + θ−i )2
θ−i θk2
(θk − θi )2 θ−i
=
−
= −wi
(θk + θ−i )2 θi
(θk + θ−i )2 θi
2θ−i θk
4θi
4θi
≥ −
≥−
≥−
.
2
(θk + θ−i )
θk + θ−i
θi + θ−i
Therefore we have for all t and S1 , ..., St−1 ,
E[Zt − Zt−1 |S1 , ..., St−1 ] ≥ −
X
i∈Ut
4θi
4
≥− .
W (θi + θ−i )
W
As Z1 , ..., Zt−1 can be determined by S1 , ..., St−1 , we have for all t and Z1 , ..., Zt−1 ,
4(t − 1)
4
4(t − 1)
4t
− Zt−1 +
|Z1 − , ..., Zt−1 −
≥ 0.
E Zt +
W
W
W
W
4t
Therefore sequence 0, Z1 + W4 , ..., Zt + W
, ..., ZT + 4T
forms a supermartingale.
W
Now we want to bound |Zt − Zt−1 |. We know that for 0 ≤ x ≤ 1, | ln(1 + x)| ≤ x and for
0 ≤ x ≤ 1/2, | ln(1 − x)| ≤ 2|x|. Therefore for i such that wi > 0,
wi (θk − θi )θ−i
(θk − θi )θ−i
θk2 θ−i
2θk
wi
|≤
·
≤
≤
| ln 1 +
W
(θk + θ−i )θi
W (θk + θ−i )θi
W θi (θk − θi )(θk + θ−i )
W (θk − θk+1 )
and
wi
wi 2(θk − θi )
θk − θi
2θk2
4θk
θi
| ln 1 −
|≤
·
=
≤
·
.
W
θk + θ−i
W θk + θ−i
W (θk − θi )(θk + θ−i )
W (θk − θk+1 ) θi + θ−i
Therefore, we get
|Zt − Zt−1 | ≤
Also notice that
X
2θk
4θk
θi
6θk
+
·
≤
.
W (θk − θk+1 ) i∈U W (θk − θk+1 ) θi + θ−i
W (θk − θk+1 )
t
θk
(θk − θk+1 )
2
23
≤ wk+1 ≤ W.
Now by Azuma’s inequality, we have
Pr [ZT ≤ −c1 ] ≤
(π,S)∼p
=
!
4T 2
)
W
exp −
6θk
2T ( W (θk −θk+1 ) )2
(c1 − 4c)2 (θk − θk+1 )2 W
exp −
72 · c · θk2
(c1 − 4c)2
(c1 −
≤ exp −
72c
.
Finally we are going to use Lemma B.2 with c1 = 1/3. We define V as indicator function
P
i (S,π)
i
of the event ni=1 w
ln pp(S,π)
≥ −c1 , i.e.
W
P
i (S,π)
i
ln pp(S,π)
≥ −c1 .
1. V = 1 if ni=1 w
W
2. V = 0, otherwise.
The probability that A identify the top item can be written as
Pr [A(S) = {π1 , ..., πk }]
(π,S)∼p
=
≤
≤
≤
≤
≤
≤
≤
≤
≤
Pr [(A(S) = {π1 , ..., πk }) ∧ (V = 0)] + Pr [(A(S) = π1 ) ∧ (V = 1)]
(π,S)∼p
X
Pr [V = 0] +
p(S, π)
(π,S)∼p
(π,S)∼p
(π,S):A(S)={π1 ,...,πk },V =1
!
n
X
Y
wi
(c1 − 4c)2
+
ec1
pi (S, π) W
exp −
72c
i=1
(π,S):A(S)={π1 ,...,πk },V =1
!
n
X
X
wi
(c1 − 4c)2
c1
+
e
· pi (S, π)
exp −
72c
W
i=1
(π,S):A(S)={π1 ,...,πk },V =1
n
X
wi
(c1 − 4c)2
c1
+e
Pr [(A(S) = {π1 , ..., πk }) ∧ (V = 1)]
exp −
72c
W (π,S)∼pi
i=1
n
X
(c1 − 4c)2
wi
c1
exp −
+e
Pr [A(S) = {π1 , ..., πk }]
72c
W (π,S)∼pi
i=1
n
X
(c1 − 4c)2
wi 1
exp −
+ ec1
·
72c
W
2
i=1
(c1 − 4c)2
ec1
exp −
+
72c
2
2
3
(1/3 − 4c)
+
exp −
72c
4
7
1 3
+ = .
8 4
8
24
The last step comes from the fact that c < 10−4 .
The following theorem is very similar to Theorem 3.1. For some technical reason, it’s not
very easy to merge the two proofs. But many parts of proofs of these two theorems are very
similar.
Theorem B.3 (Restatement of Theorem 3.2). Assume θk > θk+1 and c < 4 · 10−4. For any
P
θ2
comparisons of any size
algorithm A (can be adaptive), if A uses c i:i≤k,θi≤2θk+1 (θk+1k+1
−θi )2
(can be l-wise comparison for 2 ≤ l ≤ n), then A will identify the top-k items with probability
at most 7/8.
θ2
Proof. For notation convenience, we set wi = (θk+1k+1
for i such that i ≤ k and θi ≤ 2θk+1 .
Pn −θi )2
For other i, we set wi = 0. We also set W = i=1 wi . Then we have T = cW .
First of all, we can assume A is deterministic. This is because if A is randomized, we can
fix the randomness string which makes A achieves the highest successful probability.
Let S = (S1 , ..., ST ) be the history of algorithm. Each St is the comparison result of
round t. Notice that since A is deterministic, with S1 , ..., St , we can determine the labels of
items A want to compare in round t + 1 even when A is adaptive. So there is no point to
put the labels of compared items in the history. So we only put the comparison result in the
history, i.e St is a number in [n] and S is a length-T string of numbers in [n].
Again since A is deterministic, the label A outputs is just a deterministic function of S,
we use A(S) to denote it. A outputs correctly if A outputs the label of the top-k items, i.e.
A(S) = {π1 , ..., πk }.
We use p(S, π) to denote the probability that the items are labeled as π and A has history
S. Now consider the case when we set θi equals to θk+1 for i ≤ k. In this case the probability
of A(S) = {π1 , ..., πk } should be at most 1/2 as item k + 1 and item i have the same weight.
We use pi (S, π) to denote the probability that the items are labeled as π and A has history
S when θi is changed to θk+1 .
Now we prove the following lemma that gives the connection between p(S, π) and pi (S, π).
Lemma B.4. Consider p as a distribution over (π, S). For all c1 > 0, we have
" n
!
#
X wi pi (S, π)
(c1 − c)2
Pr
≤ −c1 ≤ exp −
.
ln
(π,S)∼p
W
p(S,
π)
18c
i=1
Proof. Define random variable Zt to be the following for t = 1, ..., T when (π, S) is sampled
from distribution p:
n
X
wi pi (S1 ...St , π)
ln
.
Zt =
W
p(S1 ...St , π)
i=1
We have
n
X
wi pi (S, π)
ln
.
ZT =
W
p(S, π)
i=1
25
T
forms a supermartigale.
Now we want to show that sequence 0, Z1 + W1 , ..., Zt + Wt , ..., ZT + W
Suppose
in round t, given S1 , ..., St−1 and π, Algorithm A compares items in set Ut . Let
P
θ−i = j∈Ut ,j6=i θi . Then we have, with probability θi /(θi + θ−i ),
X wj
wi
(θi − θk+1 )θ−i
θj − θk+1
Zt − Zt−1 =
+
ln 1 −
ln 1 +
W
(θk+1 + θ−i )θi
W
θk+1 + θ−j
j∈U ,j6=i
t
Here are two simple facts about ln. For 0 ≤ x ≤ 1, ln(1 + x) ≥ x − x2 . For 0 ≤ x ≤ 1/2,
(θ −θ
)θ−i
≤ 1/2
ln(1 − x) ≥ −x − x2 . It’s easy to check that for i such that wi > 0, we have (θik+1k+1
+θ−i )θi
and
θi −θk+1
θk+1 +θ−i
≤ 1 . Therefore, by these two facts, for i such that wi > 0, we have
(θi − θk+1 )θ−i
θ−i
θi − θk+1
θi
wi ln 1 −
+
wi ln 1 +
θi + θ−i
(θk+1 + θ−i )θi
θi + θ−i
θk+1 + θ−i
2 2
wi
(θi − θk+1 )θ−i (θi − θk+1 ) θ−i (θi − θk+1 )θ−i (θi − θk+1 )2 θ−i
≥
−
−
+
−
θi + θ−i
θk+1 + θ−i
(θk+1 + θ−i )2 θi
θk+1 + θ−i
(θk+1 + θ−i )2
2
θ−i θk+1
(θk+1 − θi )2 θ−i
=
−
= −wi
(θk+1 + θ−i )2 θi
(θk+1 + θ−i )2 θi
2
θk+1
θi
≥ −
≥−
.
(θk+1 + θ−i )θi
θi + θ−i
The last step comes from the fact that
2
θk+1
(θi + θ−i ) ≤ θi2 (θk+1 + θ−i ).
Therefore we have for all t and S1 , ..., St−1 ,
E[Zt − Zt−1 |S1 , ..., St−1 ] ≥ −
X
i∈Ut
1
θi
≥− .
W (θi + θ−i )
W
As Z1 , ..., Zt−1 can be determined by S1 , ..., St−1 , we have for all t and Z1 , ..., Zt−1 ,
t−1
1
t−1
t
− Zt−1 +
|Z1 − , ..., Zt−1 −
≥ 0.
E Zt +
W
W
W
W
T
Therefore sequence 0, Z1 + W1 , ..., Zt + Wt , ..., ZT + W
forms a supermartingale.
Now we want to bound |Zt − Zt−1 |. We know that for 0 ≤ x ≤ 1, | ln(1 + x)| ≤ x and for
0 ≤ x ≤ 1/2, | ln(1 − x)| ≤ 2|x|. Therefore for i such that wi > 0,
2
2θk+1
θ−i
wi
wi 2(θi − θk+1 )θ−i
(θi − θk+1 )θ−i
2θk+1
| ln 1 −
|≤
≤
·
≤
W
(θk+1 + θ−i )θi
W (θk+1 + θ−i )θi
W θi (θi − θk+1 )(θk+1 + θ−i )
W (θk − θk+1 )
and
2
θk+1
wi θi − θk+1
θi − θk+1
θk+1
θi
wi
|≤
·
=
≤
·
.
| ln 1 +
W
θk+1 + θ−i
W θk+1 + θ−i
W (θi − θk+1 )(θk+1 + θ−i )
W (θk − θk+1 ) θi + θ−i
26
Therefore, we get
|Zt − Zt−1 | ≤
X
2θk+1
θi
3θk+1
θk+1
+
·
≤
.
W (θk − θk+1 ) i∈U W (θk − θk+1 ) θi + θ−i
W (θk − θk+1 )
Also notice that
t
θk+1
(θk − θk+1 )
Now by Azuma’s inequality, we have
2
≤ wk ≤ W.
Pr [ZT ≤ −c1 ] ≤ exp −
(π,S)∼p
(c1 −
T 2
)
W
k+1
2T ( W (θ3θk −θ
)2
k+1 )
!
(c1 − c)2 (θk − θk+1 )2 W
= exp −
2
18 · c · θk+1
(c1 − c)2
≤ exp −
.
18c
After we prove Lemma B.4, the rest of the proof is very similar to Theorem 3.1. We omit
the argument.
B.2
Lower bounds for arbitrary weights
Again, the following theorem is very similar to Theorem 3.1.
Theorem B.5 (Restatement
P of Theorem 3.3). Assume c < 1/18. For any algorithm A (can
be adaptive), if A uses c i:i>k θθki comparisons of any size (can be l-wise comparison for
2 ≤ l ≤ n), then A will identify the top-k items with probability at most 7/8.
Proof. For notation convenience, we set wi = θθki for i > k. For i ≤ k, we set wi = 0. We
P
also set W = ni=1 wi . Then we have T = cW .
First of all, we can assume A is deterministic. This is because if A is randomized, we can
fix the randomness string which makes A achieves the highest successful probability.
Let S = (S1 , ..., ST ) be the history of algorithm. Each St is the comparison result of
round t. Notice that since A is deterministic, with S1 , ..., St , we can determine the labels of
items A want to compare in round t + 1 even when A is adaptive. So there is no point to
put the labels of compared items in the history. So we only put the comparison result in the
history, i.e St is a number in [n] and S is a length-T string of numbers in [n].
Again since A is deterministic, the label A outputs is just a deterministic function of S,
we use A(S) to denote it. A outputs correctly if A outputs the label of the top-k items, i.e.
A(S) = {π1 , ..., πk }.
We use p(S, π) to denote the probability that the items are labeled as π and A has history
S. Now consider the case when we set θi equals to θk for i > k. In this case the probability
27
of A(S) = {π1 , ..., πk } should be at most 1/2 as item k and item i have the same weight. We
use pi (S, π) to denote the probability that the items are labeled as π and A has history S
when θi is changed to θk .
Now we prove the following lemma that gives the connection between p(S, π) and pi (S, π).
Lemma B.6. Consider p as a distribution over (π, S). For all c1 > 0, we have
!
#
" n
X wi pi (S, π)
(c1 /c − 1)2 T
≤ −c1 ≤ exp −
.
ln
Pr
(π,S)∼p
W
p(S,
π)
8
i=1
Proof. Define random variable Zt to be the following for t = 1, ..., T when (π, S) is sampled
from distribution p:
n
X
wi pi (S1 ...St , π)
Zt =
ln
.
W
p(S1 ...St , π)
i=1
We have
ZT =
n
X
wi pi (S, π)
ln
.
W
p(S,
π)
i=1
T
forms a supermartingale.
Now we want to show that sequence 0, Z1 + W1 , ..., Zt + Wt , ..., ZT + W
Suppose
in
round
t,
given
S
,
...,
S
and
π,
Algorithm
A
compares
items in set Ut . Let
1
t−1
P
θ−i = j∈Ut ,j6=i θi . Then we have, with probability θi /(θi + θ−i ),
X wj
wi
(θk − θi )θ−i
θk − θj
Zt − Zt−1 =
+
ln 1 +
ln 1 −
W
(θk + θ−i )θi
W
θk + θ−j
j∈Ut ,j6=i
X wj
(θk − θi )θ−i
θk − θj
wi
−
ln 1 +
= − ln 1 −
W
(θi + θ−i )θk
W
θj + θ−j
j∈U ,j6=i
t
We are going to use a simple fact about ln: for all x > −1, ln(1 + x) ≤ x.
θi
(θk − θi )θ−i
θ−i
θk − θi
−
wi ln 1 −
−
wi ln 1 +
θi + θ−i
(θi + θ−i )θk
θi + θ−i
θi + θ−i
wi
(θk − θi )θ−i θi (θk − θi )θ−i
≥ −
−
+
θi + θ−i
(θi + θ−i )θk
θi + θ−i
2
(θk − θi ) θ−i θi
= −
(θi + θ−i )2 θk2
θi
≥ −
.
θi + θ−i
Therefore we have for all t and S1 , ..., St−1 ,
E[Zt − Zt−1 |S1 , ..., St−1 ] ≥ −
28
X
i∈Ut
1
θi
≥− .
W (θi + θ−i )
W
As Z1 , ..., Zt−1 can be determined by S1 , ..., St−1 , we have for all t and Z1 , ..., Zt−1 ,
t−1
1
t−1
t
− Zt−1 +
|Z1 − , ..., Zt−1 −
≥ 0.
E Zt +
W
W
W
W
T
Therefore sequence 0, Z1 + W1 , ..., Zt + Wt , ..., ZT + W
forms a supermartingale.
Now we want to bound |Zt − Zt−1 |. We know that for 0 ≤ x, | ln(1 + x)| ≤ x. Therefore
for i such that wi > 0,
wi
wi (θk − θi )θ−i
(θk − θi )θ−i
(θk − θi )θ−i
1
| ln 1 +
|≤
·
≤
≤
W
(θk + θ−i )θi
W (θk + θ−i )θi
W (θk + θ−i )θk
W
and
wi (θk − θi )
θk − θi
θi (θk − θi )
1
θi
wi
|≤
.
=
≤
·
| ln 1 +
W
θi + θ−i
W (θi + θ−i )
W θk (θi + θ−i )
W θi + θ−i
Therefore, we get
|Zt − Zt−1 | ≤
X 1
1
θi
2
+
·
≤
.
W i∈U W θi + θ−i
W
t
Now by Azuma’s inequality, we have
T 2
)
(c1 − W
Pr [ZT ≤ −c1 ] ≤ exp −
2 2
(π,S)∼p
2T ( W )
!
(c1 /c − 1)2 T
= exp −
.
8
After we prove Lemma B.6, the rest of the proof is very similar to Theorem 3.1 by picking
c1 = 1/3. We omit the argument.
Theorem B.7 (Restatement of Theorem 3.4). For any algorithm A (can be adaptive), if
A uses k/4 comparisons of any size (can be l-wise comparison for 2 ≤ l ≤ n), then A will
identify the top-k items with probability at most 2/3.
Proof. First of all, we can assume A is deterministic. This is because if A is randomized, we
can fix the randomness string which makes A achieves the highest successful probability.
Let S = (S1 , ..., ST ) be the history of algorithm. Each St is the comparison result of
round t. Notice that since A is deterministic, with S1 , ..., St , we can determine the labels of
items A want to compare in round t + 1 even when A is adaptive. So there is no point to
put the labels of compared items in the history. So we only put the comparison result in the
history, i.e St is a number in [n] and S is a length-T string of numbers in [n].
Again since A is deterministic, the label A outputs is just a deterministic function of S,
we use A(S) to denote it. A outputs correctly if A outputs the label of the top-k items, i.e.
A(S) = {π1 , ..., πk }.
We use p(S, π) to denote the probability that the items are labeled as π and A has history
S. Now consider the case when we set θi equals to θk+1 for i ≤ k. In this case the probability
29
of A(S) = {π1 , ..., πk } should be at most 1/2 as item k + 1 and item i have the same weight.
We use pi (S, π) to denote the probability that the items are labeled as π and A has history
S when θi is changed to θk+1 .
We define N(π, S) as the set of items among top-k items such that they are not chosen
as the favorite items by algorithm A in history S with labels π. As there are only k/4
comparisons, N(π, S) ≤ 3/4 for all π, S.
Now we prove the following simple lemma that gives the connection between p(S, π) and
pi (S, π) for all i ∈ N(π, S).
Lemma B.8.
∀π, S, i ∈ N(π, S), pi (S, π) ≥ p(S, π).
Proof. We write p(S, π) as
p(S, π) =
T
Y
p(St , π|S1 ...St−1 ).
t=1
And similarly pi (S, π) as
pi (S, π) =
T
Y
pi (St , π|S1 ...St−1 ).
t=1
Consider the comparison in round t given S1 , S2 , ..., St , π. There are two cases
1. i-th item is not compared in round t: The change of θi does not change p(St , π|S1 ...St−1 ).
So p(St , π|S1 ...St−1 ) = pi (St , π|S1...St−1 ).
2. i-th item is compared in round t: We know the i-th item is not the favorite item of
round t in this history. Therefore decreasing θi to θk+1 will increase p(St , π|S1 ...St−1 ).
So p(St , π|S1 ...St−1 ) ≤ pi (St , π|S1 ...St−1 ).
Thus we always have p(St , π|S1...St−1 ) ≤ pi (St , π|S1 ...St−1 ). By multiplying things together
we get the statement of this lemma.
30
Finally we have
Pr [A(S) = {π1 , ..., πk }]
X
p(π, S)
(π,S)∼p
=
π,S,A(S)={π1 ,...,πk }
≤
≤
X
π,S,A(S)={π1 ,...,πk }
X
π,S,A(S)={π1 ,...,πk }
1
|N(π, S)|
1
|N(π, S)|
X
pi (π, S)
X
pi (π, S)
i∈N (π,S)
i∈{1,...,k}
4 X
pi (π, S)
3k
π,S,A(S)={π1 ,...,πk }
i∈{1,...,k}
X
4 X
pi (π, S)
=
3k
X
≤
≤
4
3k
2
.
=
3
i∈{1,...,k} π,S,A(S)={π1 ,...,πk }
X
i∈{1,...,k}
1
2
Theorem B.9 (Restatement of Theorem 3.5). Assume c < 1/2. For any algorithm A (can
be adaptive), if A uses cnl comparisons of size at most l (can be 2-wise, 3-wise,...,l-wise
comparisons), then A will identify the top-k items with probability at most 7/8.
Proof. We are going to prove by contradiction. Suppose there’s some A uses cnl comparisons
of size at most l and identify the top-k items with probability more than 7/8.
Now consider another task where the goal is just to make sure k-th item appeared in
some comparison or (k + 1)-th item appeared in some comparison. Notice that when some
algorithm fails this new task, then the algorithm cannot output top-k items with probability
better than 1/2 because when both k-th item and (k + 1)-th item are not compared, the algorithm should output them with same probability for identifying top-k items. So algorithm
A should solve the new task with probability more than 3/4.
For the new task, it’s easy to see that the best strategy is to always use l-wise comparison
and compare cnl · l different items. The probability of having either k-th item or (k + 1)-th
item compared is
cn
1
2
≤ 1 − 2c ≤ 3/4.
1− 1−
n
4
Here we need to use the fact that n ≥ 4 (when n < 4, the statement of the theorem is
trivial). Now we get a contradiction.
31
Acknowledgement
We would like to thank Rene Caldentey and Yifan Feng for earlier discussions of this problem.
References
[ACN08]
N Ailon, M. Charikar, and A. Newman. Aggregating inconsistent information:
ranking and clustering. Journal of the ACM, 55(5):23:1–23:27, 2008.
[Ail11]
N. Ailon. Active learning ranking from pairwise preferences with almost optimal
query complexity. In Advances in Neural Information Processing Systems, 2011.
[BM08]
M. Braverman and E. Mossel. Noisy sorting without resampling. In Proceedings
of ACM-SIAM symposium on discrete algorithms, 2008.
[BMW16]
M. Braverman, J. Mao, and M. S. Weinberg. Parallel algorithms for select
and partition with noisy comparisons. In Proceedings of the Symposium on the
Theory of Computing (STOC), 2016.
[BT52]
R. Bradley and M. Terry. Rank analysis of incomplete block designs: I. the
method of paired comparisons. Biometrika, 39(3/4):324–345, 1952.
[BWV13]
S. Bubeck, T. Wang, and N. Viswanathan. Multiple identifications in multiarmed bandits. In Proceedings of the International Conference on Machine
Learning (ICML), 2013.
[CBCTH13] X. Chen, P. N. Bennett, K. Collins-Thompson, and E. Horvitz. Pairwise ranking aggregation in a crowdsourced setting. In Proceedings of the sixth ACM
international conference on Web search and data mining, 2013.
[CCZZ17]
J. Chen, X. Chen, Q. Zhang, and Y. Zhou. Adaptive multiple-arm identification.
In Proceedings of International Conference on Machine Learning (ICML), 2017.
[CGMS17]
X. Chen, S. Gopi, J. Mao, and J. Schneider. Competitive analysis of the top-k
ranking problem. In Proceedings of ACM-SIAM Symposium on Discrete Algorithms (SODA), 2017.
[CS15]
Y. Chen and C. Suh. Spectral MLE: Top-k rank aggregation from pairwise comparisons. In Proceedings of the International Conference on Machine Learning
(ICML), 2015.
[DKNS01]
C. Dwork, R. Kumar, M. Naor, and D. Sivakumar. Rank aggregation methods
for the web. In Proceedings of the International World Wide Web Conference,
2001.
32
[FLN03]
Ronald Fagin, Amnon Lotem, and Moni Naor. Optimal aggregation algorithms
for middleware. J. Comput. Syst. Sci., 66(4):614–656, 2003.
[HSRW16]
Reinhard Heckel, Nihar B. Shah, Kannan Ramchandran, and Martin J. Wainwright. Active ranking from pairwise comparisons and when parametric assumptions dont help. arXiv preprint arXiv:1606.08842v2, 2016.
[JKSO13]
M. Jang, S. Kim, C. Suh, and S. Oh. Top-k ranking from pairwise comparisons:
When spectral ranking is optimal. arXiv preprint arXiv:1603.04153, 2013.
[JMNB14]
K. Jamieson, M. Malloy, R. Nowak, and S. Bubeck. lil’ ucb: An optimal
exploration algorithm for multi-armed bandits. In Proceedings of Conference
on Learning Theory, 2014.
[JN11]
K. Jamieson and R. Nowak. Active ranking using pairwise comparisons. In
Advances in Neural Information Processing Systems, 2011.
[KMS07]
C. Kenyon-Mathieu and W. Schudy. How to rank with few errors. In Proceedings
of the Symposium on Theory of computing (STOC), 2007.
[LB11]
T. Lu and C. Boutilier. Learning mallows models with pairwise preferences.
In Proceedings of the International Conference on Machine Learning (ICML),
2011.
[Luc59]
R. D. Luce. Individual choice behavior: A theoretical analysis. New York:
Wiley, 1959.
[McF73]
D. McFadden. Conditional logit analysis of qualitative choice behaviour. In
P. Zarembka, editor, Frontiers in Econometrics, pages 105–142. Academic Press
New York, New York, NY, USA, 1973.
[MS16]
S. Mohajer and C. Suh. Active top-k ranking from noisy comparisons. In Proceedings of the 54th Annual Allerton Conference on Communication, Control,
and Computing (Allerton), 2016.
[NOS17]
S. Negahban, S. Oh, and D. Sha. Rank centrality: Ranking from pair-wise
comparisons. Operations Research, 65(1):266–287, 2017.
[RA14]
A. Rajkumar and S. Agarwal. A statistical convergence perspective of algorithms for rank aggregation from pairwise data. In Proceedings of the International Conference on Machine Learning (ICML), 2014.
[SBGW17]
N. B. Shah, S. Balakrishnan, A. Guntuboyina, and M. J. Wainright. Stochastically transitive models for pairwise comparisons: Statistical and computational
issues. IEEE Transactions on Information Theory (to appear), 63(2):934–959,
2017.
33
[SBPH15]
B. Szörényi, R. Busa-Fekete, A. Paul, and E. Hüllermeier. Online rank elicitation for plackett-luce: A dueling bandits approach. In Proceedings of Advances
in Neural Information Processing Systems (NIPS), 2015.
[STZ17]
C. Suh, V. Tan, and R. Zhao. Adversarial top-k ranking. IEEE Transactions
on Information Theory, 63(4):2201–2225, 2017.
[SW15]
N. B. Shah and M. Wainwright. Simple, robust and optimal ranking from
pairwise comparisons. arXiv preprint arXiv:1512.08949, 2015.
[Tra03]
Kenneth Train. Discrete Choice Methods with Simulation. Cambridge University Press, 2003.
[WMJ13]
F. Wauthier, M.Jordan, and N. Jojic. Efficient ranking from pairwise comparisons. In Proceedings of the International Conference on Machine Learning
(ICML), 2013.
[ZCL14]
Y. Zhou, X. Chen, and J. Li. Optimal pac multiple arm identification with
applications to crowdsourcing. In Proceedings of the International Conference
on Machine Learning (ICML), 2014.
34
| 8cs.DS
|
SUPER-APPROXIMATION, II:
THE p-ADIC AND BOUNDED POWER OF SQUARE-FREE INTEGERS CASES.
arXiv:1602.00409v2 [math.GR] 10 Feb 2018
ALIREZA SALEHI GOLSEFIDY
Abstract. Let Ω be a finite symmetric subset of GLn (Z[1/q0 ]), Γ := hΩi, and let πm be the group
homomorphism induced by the quotient map Z[1/q0 ] → Z[1/q0 ]/mZ[1/q0 ]. Then the family of Cayley graphs
{Cay(πm (Γ), πm (Ω))}m is a family of expanders as m ranges over fixed powers of square-free integers and
powers of primes that are coprime to q0 if and only if the connected component of the Zariski-closure of Γ
is perfect. Some of the immediate applications, e.g. orbit equivalence rigidity, largeness of certain ℓ-adic
Galois representations, are also discussed.
Contents
1. Introduction
2
1.1. Statement of the main results.
2
1.2. Comparing with the previous related results.
3
1.3. Applications
3
Acknowledgments.
7
2. Preliminary results and notation.
7
2.1. Notation.
7
2.2. Necessity.
8
2.3. A few reductions.
9
2.4. Algebraic homomorphisms and the congruence maps.
10
2.5. Ping-pong players.
12
2.6. Summary of the initial reductions.
14
2.7. Multiplicity bound.
14
3. Expansion, approximate subgroup, and bounded generation.
16
3.1. Largeness of level-Q approximate subgroups implies super-approximation.
16
3.2. Reduction to bounded generation and abelian unipotent radical.
19
4. Super-approximation: bounded power of square-free integers case.
23
4.1. The residue maps do not split.
23
4.2. Statistical non-splitting of residue maps.
25
Date: February 13, 2018.
1991 Mathematics Subject Classification. 22E40.
A. S-G. was partially supported by the NSF grants DMS-1160472, DMS-1303121, DMS-1602137 and A. P. Sloan Research
Fellowship.
1
2
ALIREZA SALEHI GOLSEFIDY
4.3. Bounded generation of perfect groups by commutators.
26
4.4. Proof of Theorem 1 for bounded powers of square-free integers.
28
5. Super-approximation: the p-adic case.
32
5.1. Escape from proper subgroups.
32
5.2. Getting a large ideal by adding/subtracting a congruence subgroup boundedly many times.
42
5.3. Getting a p-adically large vector in a submodule in boundedly many steps.
44
5.4. Proof of super-approximation: the p-adic case.
45
6. Appendix A: quantitative open image for p-adic analytic maps.
46
References
53
1. Introduction
1.1. Statement of the main results. Let Ω be a finite symmetric subset of GLn (Q), and Γ = hΩi. Since
Ω is finite, for some q0 ∈ Z+ we have Γ ⊆
implies that (under certain
Q GLn (Z[1/q0 ]). Strong approximation
Q
algebraic conditions) the closure of Γ in p∤q0 GLn (Zp ) is open in p∤q0 (G(Qp ) ∩ GLn (Zp )), where G is the
Zariski-closure of Γ in GLn . In combinatorial language, this means that the Cayley graphs
Cay (πm (G(Q) ∩ GLn (Z[1/q0 ])), πm (Ω)) ,
where m and q0 are coprime and πm is the group homomorphism induced by the quotient map πm : Z[1/q0 ] →
Z[1/q0 ]/mZ[1/q0 ], have at most C := C(Ω) many connected components. The point being that C does not
depend on m. Super-approximation1 tells us that these sparse graphs are highly connected, i.e. they form a
family of expanders. A family {Xi }i∈I of d0 -regular graphs is called expander if there is c > 0 such that for
any i ∈ I and any subset A of the set of vertices V (Xi ) of Xi we have
c<
|e(A, V (Xi ) \ A)|
,
min(|A|, |V (Xi ) \ A|)
where e(A, V (Xi ) \ A) consists of the edges that connect a vertex in A to a vertex in V (Xi ) \ A.
Expanders are extremely useful in communication and theoretical computer science (e.g. see [HLW06]). In
the past decade they have been found useful in a wide range of pure math problems, e.g. affine sieve [BGS10,
SGS13], sieve in groups [LM13], variation of Galois representations [EHK12], etc. (see [BO14] for a collection
of surveys of related works and applications).
In this article we prove the best possible super-approximation results for two families of residue maps. Let
Vf (Q) be the set of primes in Q, and νp (q) be the p-valuation of q, i.e. the power of p in q.
Theorem 1. Let Ω be a finite symmetric subset of GLN0 (Z[1/q0 ]), Γ := hΩi, and M0 be a fixed positive
integer. Suppose Γ is an infinite group. Then the family of Cayley graphs {Cay(πm (Γ), πm (Ω))}m is a family
of expanders as m runs through either
{pn | n ∈ Z+ , p ∈ Vf (Q), p ∤ q0 } or {q| ∀p ∈ Vf (Q), νp (q) ≤ M0 }
if and only if the connected component G◦ of the Zariski-closure G of Γ in GLN0 is perfect, i.e. [G◦ , G◦ ] = G◦ .
1Following A. Kontorovich’s suggestion, I call this phenomenon super-approximation. It is worth pointing out that this
phenomenon has been called superstrong approximation [BO14] by some authors.
SUPER-APPROXIMATION, II:
THE p-ADIC AND BOUNDED POWER OF SQUARE-FREE INTEGERS CASES.
3
In the appendix, a quantitative open image for p-adic analytic maps is proved, which should be of independent
interest. In a joint work with Zhang [SGZ], this result is used to generalize a theorem of Burger and
Sarnak [BS91] to the setting of super-approximation (see Theorem 9 for the statement of this result.).
1.2. Comparing with the previous related results. The importance of Theorem 1 lies on the fact
that it is an if-and-only-if statement. In fact, by Proposition 10, we get that, if {πmi (Γ), πmi (Ω)} is a
family of expanders for some increasing sequence {mi } of integers that are relatively prime to q0 , then
{πm (Γ), πm (Ω)}m is a family of expanders as m runs through integers specified in Theorem 1.
Super-approximation for square-free numbers is the main theorem of [SGV12]. Of course it is a special case
of Theorem 1 for m ∈ {q| ∀p ∈ Vf (Q), νp (q) ≤ M0 } and M0 = 1. It should, however, be noted that our proof
relies on [SGV12]. Proof of this part of Theorem 1 is heavily influenced by the relevant part of [BV12]. But
here, instead of working with the concrete structure of the congruence quotients of SLn (Z) as it was done in
[BV12], one has to work with arbitrary perfect groups. So it is inevitable to make use of basic theorems from
Bruhat-Tits theory to describe structure of a maximal compact subgroup of H(Qp ) where H is a semisimple
Qp -group, and to use truncated or finite logarithmic maps [Pin98, Section 6] or [SG, Section 2.9]. In addition,
one has to understand how such a maximal compact group acts on an open compact subgroup of U(Qp )
where U is a unipotent Qp -group, which makes the use of the language of schemes necessary.
Proof of Theorem 1 for powers of primes relies on [SG] where the semisimple case is proved.
Prior to [SG], the case of Zariski-dense subgroups of SLn (Z) was studied by Bourgain-Gamburd [BG08-b,
BG09] and Bourgain-Varjú [BV12]. As it is pointed out in [SG, Section 1.3], the proofs by Bourgain and
Gamburd [BG08-b, BG09] and Bourgain and Varjú [BV12] rely heavily on Archimedean dynamics. In
particular, those ideas cannot be implemented for a finitely generated subgroup Γ of GLn (Q) if Γ is a
bounded subgroup of GLn (R). But it is worth pointing out that, if Γ is a Zariski-dense subgroup of SLn (Z),
Bourgain and Varjú [BV12] prove that {Cay(πm (Γ), πm (Ω))}m is a family of expanders with no restriction
on m.
Prior to the current work, no super-approximation result for powers of primes beyond semisimple case was
known. To get such a result, one had to use new ideas and combine techniques from p-adic analytic analysis,
p-adic analytic pro-p groups, and non-commutative algebra.
1.3. Applications.
1.3.1. Random-walk and spectral gap. Let Ω be a finite symmetric subset of a compact group G. Let µ be
the counting probability measure on Ω and Γ := hΩi. Let
Tµ : L2 (Γ) → L2 (Γ), Tµ (f ) := µ ∗ f,
P
where (µ ∗ f )(g) := g′ ∈supp µ µ(g ′ )f (g ′−1 g) and Γ is the closure of Γ in G. It is easy to see that Tµ is a
self-adjoint operator, the constant function 1Γ is an eigenfunction with eigenvalue one, i.e. Tµ (1Γ ) = 1Γ ,
and kTµ k = 1. Let
λ(µ; G) := sup{|λ|| λ ∈ spec(Tµ ), λ < 1}.
We get a fairly good understanding of the random-walk in G with respect to µ if λ(µ; G) < 1, in which case
it is said that either Γ y Γ or the random-walk with respect to µ has spectral gap.
Let us recall a couple of well-known results which give us a connection between having spectral gap and
explicit construction of expanders (see [Lub94, Chapter 4.3], [LZ03, Chapter 1.4], [SG, Remark 15]).
Remark 2.
(1) Let Γ be the group generated by a finite symmetric set Ω. Suppose {Ni }i∈I is a family
of normal, finite-index subgroups of Γ. Suppose, for any i1 , i2 ∈ I, there is i3 ∈ I such that Ni3 ⊆
Ni1 ∩ Ni2 . Then, by Peter-Weyl theorem, one has
λ(PΩ ; lim Γ/Ni ) = sup λ(PιNi (Ω) ; Γ/Ni ),
←−
i
4
ALIREZA SALEHI GOLSEFIDY
where ιNi : Γ → Γ/Ni is the natural quotient map.
(2) Suppose {Ni }i∈I is a family of normal, finite-index subgroups of Γ. Then the family of Cayley graphs
{Cay(Γ/Ni , ιNi (Ω))}i∈I is a family of expanders if and only if
sup λ(PιNi (Ω) ; Γ/Ni ) < 1.
i
Based on these results, we get that Theorem 1 implies the following.
Corollary 3. Let Ω be a finite symmetric subset of GLN0 (Z[1/q0 ]), and PΩ be the counting probability
measure on Ω. Suppose the connected component G◦ of the Zariski-closure G of Γ := hΩi in GLN0 is perfect.
Then for any positive integer M0 we have
Y
GLN0 (Z/pM0 Z)) < 1,
λ(PΩ ;
p∤q0
and
sup λ(PΩ ; GLN0 (Zp )) < 1.
p∤q0
Since a stronger form of the only-if part of Theorem 1 (see Proposition 10) holds, having spectral gap for a
single place implies a uniform spectral gap for all the places.
Corollary 4. Let Ω be a finite symmetric subset of GLN0 (Z[1/q0 ]), and PΩ be the counting probability
measure on Ω. Then for some p ∤ q0 we have λ(PΩ ; GLN0 (Zp )) < 1 if and only if
sup λ(PΩ ; GLN0 (Zp )) < 1.
p∤q0
Spectral gap has a well-known weighted equidistribution consequence.
Corollary 5. In the setting of Corollary 3, there is λ < 1 such that for any prime p which does not divide
q0 and G(Qp ) ∩ GLN0 (Zp )-finite function f0 on G(Qp ) (which means that f0 is invariant under an open
subgroup of G(Qp )) we have
Z
Z
q
X (l)
|
f0 (g)dgk2 |Γp · f0 |λl ,
f0 (g)dg| ≤ kf0 −
PΩ (γ)f0 (γ) −
Γp
Γp
γ∈Γ
where Γp is the closure of Γ in GLN0 (Zp ) and it acts on functions on G(Qp ) via the right translation, i.e.
(l)
(g · f )(g ′ )R:= f (g ′ g), PΩ is the l-th convolution power of PΩ , dg is the probability Haar measure on Γp , and
kf k2 := ( g∈Γp f (g)2 dg)1/2 .
Proof. Since Ω is a symmetric set, we have that
Z
X (l)
f0 (g)dg = Tµl (f0 )(I) − h1Γp , f0 i,
(1)
PΩ (γ)f0 (γ) −
γ∈Γ
Γp
where I is the identity matrix and h, i is the dot product in L2 (Γp ).
On the other hand, by Corollary 3, we have that
(2)
kTµl (f0 ) − h1Γp , f0 ik2 = kTµl (f0 − h1Γp , f0 i1Γp )k2 ≤ λl kf0 − h1Γp , f0 i1Γp )k2 ,
SUPER-APPROXIMATION, II:
THE p-ADIC AND BOUNDED POWER OF SQUARE-FREE INTEGERS CASES.
5
where λ := supp∤q0 λ(PΩ ; GLN0 (Zp )). For any function f , let Γp,f := {g ∈ Γp | g · f = f }. Since for any
g ∈ Γp and f ∈ L2 (Γp ) we have Tµ (g · f ) = g · Tµ (f ), we get that Γp,f ⊆ Γp,Tµ (f ) . So we have
Z
Z
2
l
kTµ (f0 ) − h1Γp , f0 ik2 =
|Tµl (f0 )(gg ′ ) − h1Γp , f0 i|2 dg ′ dg
Γp /Γp,f0
=
Z
Γp /Γp,f0
=
Z
Γp /Γp,f0
Γp,f0
Z
Γp,f0
|Tµl (f0 )(g) − h1Γp , f0 i|2
1
=
[Γp : Γp,f0 ]
(3)
≥
|(g ′ · Tµl (f0 ))(g) − h1Γp , f0 i|2 dg ′ dg
X
Z
dg ′ dg
Γp,f0
|Tµl (f0 )(g)
− h1Γp , f0 i|2
gΓp,f0 ∈Γp /Γp,f0
− h1Γp , f0 i|2
.
|Γp · f0 |
|Tµl (f0 )(I)
Equations (1), (2), and (3) imply the claim.
Corollary 5 is another indication that super-approximation is a suitable name for such a phenomenon as it
implies a quantitative way to approximate points.
1.3.2. Orbit equivalence. Suppose Γ ⊆ G and Λ ⊆ H are dense subgroups of compact groups G and H. We
say the left translation actions Γ y (G, mG ) and Λ y (H, mH ) are orbit equivalent if there exists a measure
class preserving Borel isomorphism θ : G → H such that θ(Γg) = Λθ(g), for mG -almost every g ∈ G.
Surprisingly, if Γ and Λ are amenable, the mentioned actions are orbit equivalent [OW80, CFW81]. In the
past decade there have been a lot of progress on this subject, and as a result now it is known that one gets
orbit equivalence rigidity under spectral gap assumption [Ioa14-a, Ioa14-b].
Corollary 6. Let Ω be a finite symmetric subset of GLN0 (Z[1/q0 ]), and PΩ be the counting probability
measure on Ω. Suppose the connected component G◦ of the Zariski-closure G of Γ := hΩi in GLN0 is perfect.
Let Γp be the closure of Γ in GLN0 (Zp ) where p ∤ q0 .
Let Λ be a countable dense subgroup of a profinite group H. Then Γ y (Γp , mΓp ) and Λ y (H, mH ) are
orbit equivalent if and only if there are open subgroups G0 ⊆ Γp and H0 ⊆ H and a continuous isomorphism
δ : G0 → H0 such that [Γp : G0 ] = [H : H0 ] and δ(G0 ∩ Γ) = Λ ∩ H0 . In particular, H is p-adic analytic.
Proof. This is a direct consequence of Corollary 3 and [Ioa14-a, Theorem A].
1.3.3. Variations of Galois representations in one-parameter families of abelian varieties. Since eight years
ago, because of a surge of works by various people specially Cadoret, Tamagawa [Cad, CT12, CT13], Hui,
and Larsen [Hui12, HL], we have got a much better understanding of the image of ℓ-adic and adelic Galois
representations induced by Tate modules of an abelian scheme. On the other hand, Ellenberg, Hall, and
Kowalski in [EHK12] gave a beautiful connection between variations of Galois representations and certain
spectral gap property2. Here we make an observation that Ellenberg-Hall-Kowalski [EHK12] machinery
combined with Theorem 1 gives an alternative approach towards [CT12, CT13]; in particular, we do not get
any new result on this topic.
In this section, let k be a finitely generated characteristic zero field, k be the algebraic closure of k, and U
be a smooth algebraic curve over k such that U ×k k is connected. Let A → U be an abelian scheme of
dimension g ≥ 1, defined over k. Take an embedding of k into C, and let
ρ0 : π1 (U (C), y0 ) → AutU (A) ⊆ GL2g (Z)
2Instead of being expanders, they only need a weaker assumption on the relevant Schreier graphs, called esperantist graphs.
6
ALIREZA SALEHI GOLSEFIDY
be the monodromy representation. Let Γ := ρ0 (π1 (U (C), y0 )), and G be the Zariski-closure of Γ in (GL2g )Q .
Let Ω be a finite symmetric generating set of Γ.
For any x ∈ U (k), the fiber Ax over x is an abelian variety defined over the residue field k(x) at x. For any
prime l, let Tl,x be the Tate module of Ax , i.e.
Tl,x := lim Ax [lm ](k),
←−
where Ax [lm ](k) is the lm -th torsion elements of Ax (k). So Tl,x ≃ Z2g
l and we get the l-adic Galois representation
ρl,x : Gal(k/k(x)) → AutZl (Tl,x ) ≃ GL2g (Zl ).
Lemma 7. In the above setting, the connected component G◦ of G is perfect.
Proof. By [EHK12, Proposition 16], Γ has a finite index subgroup Λ such that πl (Λ) is a perfect group and
generated by its order l elements for large enough prime l. In particular, the index of any proper subgroup
of πl (Λ) is at least l.
Let Γ◦ := Γ ∩ G◦ . So Γ◦ is Zariski-dense in G◦ . Since [πl (Λ) : πl (Λ ∩ Γ◦ )] ≤ [Γ : Γ◦ ], for large enough prime
l we have that πl (Λ) = πl (Λ ∩ Γ◦ ).
f
On the other hand, the quotient map G◦ −
→ (G◦ )ab := G◦ /[G◦ , G◦ ] is defined over Q. Hence after realizing
◦ ab
′
(G ) as a subgroup of (GLn )Q , for large enough l we have that πl (f (Γ◦ )) is a homomorphic image of
πl (Γ◦ )ab := πl (Γ◦ )/[πl (Γ◦ ), πl (Γ◦ )]. Suppose to the contrary that G◦ is not perfect. Then f (Γ◦ ) is a finitely
generated, infinite, abelian group as it is a Zariski-dense subgroup of (G◦ )ab . Hence |πl (f (Γ◦ ))| → ∞ as l
goes to infinity, which implies that |πl (f (Γ◦ ∩ Λ))| → ∞. Since πl (f (Γ◦ ∩ Λ)) is abelian and a homomorphic
image of πl (Γ◦ ∩ Λ), we get a contradiction.
Now we can give an alternative proof of the main result of [CT12, CT13].
Proposition 8. In the above setting, for any integer d and any prime ℓ there is an integer B := B(d, ℓ, ρ0 )
such that
Ud,ℓ := {x ∈ U (k)| [k(x) : k] ≤ d, [Γℓ : Im(ρℓ,x ) ∩ Γℓ ] > B},
is finite.
Proof. Suppose to the contrary that there is a sequence {xi } of points in U (k) such that
(1) xi 6= xj if i 6= j,
(2) [k(xi ) : k] ≤ d,
(3) [Γℓ : Im(ρℓ,xi ) ∩ Γℓ ] = Ni where Ni ∈ Z+ ∪ {∞}, and Ni → ∞.
By Lemma 38, there is an increasing sequence {mi }i of positive integers and a sequence of open subgroups
{Hi }i of Γℓ such that
lim [Γℓ : Hi ] = ∞, Γℓ ∩ Im(ρℓ,xi ) ⊆ Hi , and Γℓ [ℓmi ] ⊆ Hi .
i→∞
Going to a subsequence, if necessary, we can and will assume that πℓmi (Hi ) = πℓmi (Hj ) for any j ≥ i.
Therefore we can and will assume that H1 ) H2 ) · · · .
φi
For each i, we get an open subgroup of π1et (U ×k k), and so we get an étale cover Ui −→ U of U that is defined
ei ∈ Ui (k(xi )) which is in the fiber over xi ∈ U (k(xi )). Moreover,
over k. Since Im(ρℓ,xi ) ⊆ Hi , there is x
φij
since {Hi }i is a decreasing sequence of open subgroups, we get k-étale covering maps Uj −−→ Ui for any
i ≤ j that satisfy φi ◦ φij = φj . In particular, we have that
[
Ui (k ′ )
(4)
{φij (e
xj )}j ⊆
[k′ :k]≤d
SUPER-APPROXIMATION, II:
THE p-ADIC AND BOUNDED POWER OF SQUARE-FREE INTEGERS CASES.
7
is an infinite subset for any i. On the other hand, since π1et (Ui ×k k) is the profinite closure of π1 (Ui (C), x0 ),
the natural embedding induces an isomorphism between the Schreier graphs
(5)
Sch(π1 (U (C), x0 )/π1 (Ui (C), x
e0,i ); Ω) ≃ Sch(πℓmi (Γℓ )/πℓmi (Hi ); πℓmi (Ω)),
where x
e0,i is a point over x0 . On the other hand, by Lemma 7 and Theorem 1, we have that
(6)
{Cay(πℓmi (Γ); πℓmi (Ω)}i
is a family of expanders. Since the Schreier graphs in (5) are quotients of the Cayley graphs in (6) and
their size goes to infinity, they form a family of expanders. Therefore by the main result of EllenbergHall-Kowalski [EHK12, Theorem 8] we have that the (geometric) gonality γ(Ui ) of Ui goes to infinity.
Hence by a corollary [CT13, Theorem 2.1] of Falting’s theorem [Fal91] on Lang-Mordell conjecture (see
[Fer94, Mc95, Maz00]) we have that
[
Ui (k ′ )
[k′ :k]≤d
is finite for large enough i. This contradicts (4).
1.3.4. Inducing super-approximation. As it was mentioned earlier, in the appendix a quantitative open function theorem (see Lemma 53′ ) is proved, which is of independent interest. This result plays an important
role in my joint work with Zhang [SGZ], where we prove that one can induce super-approximation from a
subgroup under a mild (and needed) algebraic condition.
Theorem 9. Let Ω1 be a finite symmetric subset of GLN0 (Z[1/q0 ]), and Γ1 := hΩ1 i. Let Ω2 be a finite
symmetric subset of Γ1 , and Γ2 := hΩ2 i. Let G◦i be the connected component of the Zariski-closure of Γi in
GLN0 for i = 1, 2. Suppose the smallest normal subgroup of G◦1 which contains G◦2 is G◦1 . Then, if Γ2 has
super-approximation, then Γ1 has, too; that is equivalent to say
Y
Y
Zp ) < 1.
Zp ) < 1 =⇒ λ(PΩ1 ;
λ(PΩ2 ;
p∤q0
p∤q0
Such a result for arithmetic groups was proved by Burger and Sarnak [BS91].
Acknowledgments.
I would like to thank Adrian Ioana for pointing out the application of spectral gap to orbit rigidity. I
am thankful to Cristian Popescu for useful conversations concerning ℓ-adic Galois representations. Special
thanks are due to Efim Zelmanov for his support and encouragement. My sincere appreciation goes to the
referees for all the corrections and suggestions, which made this article much easier to read, and for helping
me to spot some inconsistencies in the first version of this work.
2. Preliminary results and notation.
2.1. Notation. In this note, for two real valued functions f and g, and a set of parameters X, we write
f = OX (g) if there are positive functions C1 (X) and C2 (X) of the set of parameters such that for t ≥ C1 (X)
we have 0 ≤ f ≤ C2 (X)g; notice that this is slightly different from the usual Landau symbol as OX (g)
is assumed to be non-negative. For two real valued functions f and g of a real variable t and a set of
parameters X, we write f = ΘX (g) if there are positive functions C1 (X), C2 (X), and C3 (X) such that
C1 (X)g(t) ≤ f (t) ≤ C2 (X)g(t) for any t ≥ C3 (X). For two real valued functions f and g of a real variable t
and a set of parameters X, we write f ≪X g if there are positive functions C1 (X) and C2 (X) such that for
t ≥ C2 (X) we have f (t) ≤ C1 (X)g(t). So for two non-negative functions f and g we have f = OX (g) if and
only if f ≪X g; and f = ΘX (g) if and only if g ≪X f ≪X g.
8
ALIREZA SALEHI GOLSEFIDY
For a (commutative unital) ring R and a ∈ R, πa : R → R/aR is the quotient map. We denote also the
induced group homomorphism GLn (R) → GLn (R/aR) and all of its restrictions by πa . For a subgroup Γ of
GLn (R), Γ[a] denotes {γ ∈ Γ| πa (γ) = 1}; in particular Γ[1] = Γ.
Q
· · · · · A} for the subset {a1 · · · ac | a1 , . . . , aC ∈ A};
For a subset A of a group G, we write either C A or A
| · A {z
C times
−1
−1
−1
A
P denotes the subset {a | a ∈ A}; and A is called symmetric if A = A . For an additive group G,
C A denotes the subset {a1 + · · · + aC | a1 , . . . , aC ∈ A}.
For a measure µ on a group G, we let supp µ be the support of µ. For a measure µ on a group G and g ∈ G,
we let µ(g) := µ({g}). For a measure µ with finite support on a group G and
Pa complex valued function f
on G, we let µ ∗ f be the convolution of µ and f , that means (µ ∗ f )(g) = s∈supp(µ) µ(s)f (s−1 g). For a
(l)
measure µ with finite support on a group G,
P we let µ be the l-fold convolution of µ, that means for any
(l)
g ∈ G, we have µ (g) = (µ ∗ · · · ∗ µ)(g) = g1 ···gl =g µ(g1 ) · · · µ(gl ). We say a measure µ with finite support
| {z }
l times
on a group G is symmetric if µ(g −1 ) = µ(g) for any g ∈ G. For a finite subset X of a group G, we let PX
be the probability counting measure on X.
The set of (rational) primes is denoted by Vf (Q). For any prime p, fp denotes the finite field with p elements;
Qp is the field of p-adic numbers; Zp is the ring of p-adic integers; and vp : Qp \ {0} → Z (and its restriction
to Q \ {0}) is the p-adic valuation, that means vp (x) = n if and only if x ∈ pn Zp \ pn+1 Zp . For any prime p
and x ∈ Qp , we let |x|p := (1/p)vp (x) if x 6= 0, and |0|p := 0.
2.2. Necessity. In this section, we will prove that getting a family of expanders modulo an infinite sequence
of integers implies that the connected component of the Zariski-closure is perfect (see Proposition 10). In
particular, we get the necessary part of Theorem 1.
Let us remark that proof of [SGV12, Section 5.1] can be adjusted to give the necessary part of Theorem 1 for
any fixed prime p. The main point being that the proof in [SGV12] makes use of the fact that the congruence
kernels modulo square-free numbers define a topology.
The proof here is rather straightforward and has three parts:
(a) Changing Γ to Γ◦ := Γ ∩ G◦ , we can and will assume G is Zariski-connected;
(b) If G has infinite abelianization Gab := G/[G, G], then the order of the abelianization πqi (Γ)ab of
πqi (Γ) gets arbitrarily large as qi goes to infinity.
(c) If S is a finite symmetric generating set of an abelian group A, then the order of A is bounded by a
function of |S| and λ(PS ; A) (see [LW93, Corollary 3.3]).
Proposition 10. Let Ω be a finite symmetric subset of GLn (Z[1/q0 ]) and Γ = hΩi ⊆ GLn (Z[1/q0 ]). Let
G◦ be the Zariski-connected component of the Zariski-closure G of Γ in (GLn )Q . If {Cay(πqi (Γ), πqi (Ω))}i
is a family of expanders for integers q1 < q2 < · · · coprime with q0 , then G◦ is perfect, i.e. [G◦ , G◦ ] = G◦ .
Proof. Let Γ◦ := Γ ∩ G◦ . Then, by [SG, Corollary 17], Γ◦ has a finite symmetric generating set Ω◦ and
{Cay(πqi (Γ◦ ), πqi (Ω◦ ))}i
is a family of expanders. Therefore for some positive integer C we have |πqi (Γ◦ )/[πqi (Γ◦ ), πqi (Γ◦ )]| < C for
any i (see [LW93, Corollary 3.3]).
f
On the other hand, G◦ and [G◦ , G◦ ] are Q-groups, and the quotient map G◦ −
→ G◦ /[G◦ , G◦ ] is a Q◦
◦
◦
homomorphism. Therefore G /[G , G ] can be viewed as a Q-subgroup of (GLn′ )Q such that first the
quotient map f induces a Q-homomorphism from G to (GLn′ )Q and second f (Γ◦ ) ⊆ GLn′ (Z[1/q0 ]). So there
is an integer q ′ such that for any prime p we have kf (γ) − Ikp ≤ |q ′ |−1
p kγ − Ikp . Therefore f induces an
SUPER-APPROXIMATION, II:
THE p-ADIC AND BOUNDED POWER OF SQUARE-FREE INTEGERS CASES.
9
epimorphism from πqi (Γ◦ )/[πqi (Γ◦ ), πqi (Γ◦ )] onto πqi / gcd(qi ,q′ ) (f (Γ◦ )). Hence for any i we have
|πqi / gcd(qi ,q′ ) (f (Γ◦ ))| < C.
Since f (Γ◦ ) is a Zariski-dense subset of G◦ /[G◦ , G◦ ] and modulo arbitrarily large integers it has at most C
elements, we have that G◦ /[G◦ , G◦ ] is finite. So by the connectedness we have that G◦ is perfect.
2.3. A few reductions. In this section, we make a few reductions and describe the group structure of
πQ (Γ) using strong approximation [Nor89].
Lemma 11. It is enough to prove the sufficiency part of Theorem 1 under the following additional assumptions on the Zariski-closure G of Γ in (GLn0 )Q :
There are connected, simply-connected, semisimple Q-group Gs , and unipotent Q-group U such that Gs acts
on U and G ≃ Gs ⋉ U.
Proof. A similar argument as in the proof of [SG, Lemma 18] works here. For the convenience of the reader,
we present an outline of the proof.
By the assumption, G◦ is a perfect group. Therefore the radical U of G◦ is unipotent. Let Gs be a Levi
e s be the simply-connected cover of Gs , and ι : G
e s ⋉ U → Gs ⋉ U = G◦ be the induced
subgroup of G◦ , G
e
covering map. To lift the problem to G, we consider
e
e := ι−1 (Γ ∩ G◦ (Q)) ∩ G(Q),
Λ
e
e Since G◦ (Q)/ι(G(Q))
and Λ := ι(Λ).
is a torsion abelian group and Γ is a finitely generated group, we have
that Λ is a normal finite index subgroup of Γ. Therefore by [SG, Corollary 17] we have that Λ has a finite
symmetric generating set Ω′ and for any family of positive integers C we have that {Cay(πq (Γ), πq (Ω))}q∈C
is a family of expanders if and only if {Cay(πq (Λ), πq (Ω′ ))}q∈C is a family of expanders.
e → (GLn′ )Q . Since Γ ⊆ GLn0 (Z[1/q0 ]), after passing to a (normal) finite index
Let us fix an embedding f : G
0
e ⊆ GLn′ (Z[1/q0 ]). Since ι is an isogeny defined
subgroup of Λ, if needed, we can and will assume that f (Λ)
0
′
e we have
over Q, there is q0 ∈ Z such that for any positive integer q and γ ∈ Λ
πq (f (γ)) = 1 implies πq/ gcd(q0′ ,q) (ι(γ)) = 1.
e maps onto πq/ gcd(q,q′ ) (Λ) via the map induced by ι. Therefore if Theorem 1 is proved for the
Hence πq (Λ)
0
e
e ⊆ GLn′ (Z[1/q0 ]), we get the desired result for the group Γ. Hence we can focus on the group Λ
group Λ
0
e and the claim follows.
which is Zariski-dense in G;
Lemma 12. Let G and G be the Zariski-closure of Γ in (GLn0 )Z[1/q0 ] and (GLn0 )Q , respectively. Let Gs be
a semisimple Q-subgroup of G, and U be a unipotent Q-subgroup of G such that G = Gs ⋉ U.
It is enough to prove the sufficiency part of Theorem 1 under the following additional assumptions on the
b of Γ in Q
closure Γ
p∤q0 GLn0 (Zp ):
Q
Q
b=
(1) Γ
p∤q0 Pp is an open compact subgroup of
p∤q0 G(Zp ).
′
(2) For some non-negative integer q0 and any prime p ∤ q0 , we have Pp = G(Qp ) ∩ GLn0 (Zp )[pkp ], where
kp := vp (q0′ ) and GLn0 (Zp )[pkp ] := {g ∈ GLn0 (Zp )| kg − 1kp ≤ p−kp }. And Pp = Kp ⋉ Up where
Kp := Gs (Qp ) ∩ GLn0 (Zp )[pkp ], Up := U(Qp ) ∩ GLn0 (Zp )[pkp ].
(3) There is a prime p0 such that, for p ≥ p0 , Kp = Gs (Qp ) ∩ GLn0 (Zp ) is a hyperspecial parahoric
subgroup of Gs (Qp ), and Up = U(Qp ) ∩ GLn0 (Zp ).
(4) For p ≥ p0 , the dimension of any non-trivial irreducible representation of πp (Γ) ≃ G(fp ) is at least
|πp (Γ)|Odim G (1) .
Proof. By [SG, Corollary 17], we are allowed to pass to a finite-index subgroup of Γ if needed. By Lemma 11
we can and will assume that G is connected, simply-connected, and perfect. Hence by Strong approximation
10
ALIREZA SALEHI GOLSEFIDY
Q
b is an open subgroup of
we have that Γ
p∤q0 G(Zp ) (see [Nor89, Theorem 5.4]). So passing to a finite-index
subgroup we get part 1. By [Tit79, Section 3.8], for large enough p, Kp is a hyperspecial parahoric. (See
the same reference for the definition of hyperspecial; here we use the property that πp (Kp ) is a product of
quasi-simple groups where the number of factors and their ranks are bounded by dim Gs .) Since the action
of Gs on U is defined over Q, one gets the other claims of part 2 and part 3. Moreover, for large enough
p, πp (Γ) = πp (Kp ) ⋉ πp (Up ) is a perfect group. Hence by [SGV12, Corollary 14] the restriction of any
non-trivial representation ρ of πp (Γ) to πp (Kp ) is non-trivial. Since Kp is hyperspecial, πp (Kp ) is a product
of quasi-simple groups where the number of factors and their ranks are bounded by dim Gs . Therefore by
[LS74] we have dim ρ ≥ |πp (Kp )|Odim Gs (1) ≥ |πp (Γ)|Odim G (1) .
b and so
We notice that the restriction of πq to Γ factors through Γ,
Y
Y
πpni (Kpi ) ⋉ πpni (Upi ).
πpni (Ppi ) ≃
πQ pni (Γ) ≃
i
i
i
i
i
i
i
2.4. Algebraic homomorphisms and the congruence maps. In this section, an auxiliary result on the
relation between the congruence maps and Q-group homomorphisms is proved (see Lemma 13). This relation
is crucial for reducing the proof of the main theorem to the case where the unipotent radical is abelian.
Here is a general formulation of the issue (without a reference to affine group schemes): suppose H1 ⊆
(GLn1 )Q and H2 ⊆ (GLn2 )Q are given, and ρ : H1 → H2 is a Q-groupQhomomorphism. For any positive
integer q and i = 1, 2, let πq be the induced group homomorphism on p GLni (Zp ). For any prime p, let
Hi,p := Hi (Qp ) ∩ GLni (Zp ). One would like to have that ρ and πq commute; that means the following
(wrong) isomorphism
Y
Y
(7)
πq (ρ( H1,p )) ≃ ρ(πq ( H1,p )).
p
p
Q
Q
Equation (7) has two issues: (a) ρ( p H1,p ) is not necessarily a subgroup of p H2,p , and so it does not
make sense to talk about its congruence quotient, (b) ρ is defined by polynomials with rational coefficients
Q
that are not necessarily integer. So it does not make sense to talk about ρ of the finite group πq ( p H1,p ).
In Lemma 13 under the assumption that ρ is surjective with a Q-section, an almost commuting of ρ and πq
is proved; in particular, it is proved that, if q does not have small prime factors, then the isomorphism in
Equation (7) holds.
The main ingredient in the proof of Lemma 13 is the fact that a Q-homomorphism ρ : H1 → H2 between
two algebraic Q-groups H1 and H2 induces a continuous homomorphism between the adelic points H1 (AQ )
and H2 (AQ ).
After proving Lemma 13, it will be used for the quotient maps ρ1 : G → Gs and ρ2 : G → Gs ⋉ U/[U, U]
where Gs is a Levi Q-subgroup of G, and U is the unipotent radical of G.
Lemma 13. Let H ⊆ (GLn1 )Q be a Q-subgroup (with a given embedding). Let ρ : H → (GLn2 )Q be a
Q-homomorphism. Then the following hold.
(1) There is g ∈ GLn2 (Q) such that ρ(H(Q) ∩ GLn1 (Z[1/q0 ])) ⊆ g GLn2 (Z[1/q0 ])g −1 .
(2) There is a positive integer q ′ such that for any prime p and any h ∈ H(Qp ) we have
kρ(h) − 1kp ≤ |q ′ |−1
p kh − 1kp .
ρ
(3) Let 1 → L → H −
→ H̄ → 1 be a short exact sequence of Q-groups. Suppose there is a Q-section
s : H̄ → H such that s(1) = 1. Let q0′ be a positive integer and kp := vp (q0′ ). For any prime p ∤ q0 , let
Qp = H(Qp ) ∩ (1 + q0′ gln1 (Zp )) = {hp ∈ H(Qp )| khp − 1k ≤ p−kp }. Assume that ρ(Qp ) ⊆ GLn2 (Zp ),
SUPER-APPROXIMATION, II:
THE p-ADIC AND BOUNDED POWER OF SQUARE-FREE INTEGERS CASES.
Q
(8)
11
′
kp
and s(ρ(Qp )) ⊆ GLn1 (Zp ) for any prime p ∤ q0 . Then there is a positive integer q ′ := p p such
that for non-negative integers np ≥ kp′ + kp we have
Y
Y
Y
πpnp −kp′ (Qp )/πpnp −kp′ (L(Qp ) ∩ Qp ).
πpnp (ρ(Qp )) ։
πpnp +kp′ (Qp )/πpnp +kp′ (L(Qp ) ∩ Qp ) ։
p∤q0
p∤q0
p∤q0
Proof. For any p there is gp ∈ GLn2 (Qp ) such that gp ρ(H(Qp ) ∩ GLn1 (Zp ))gp−1 ⊆ GLn2 (Zp ). Since ρ is
defined over Q, for large enough p, we can assume that gp = 1. Hence there is gb = (gp ) ∈ GLn2 (AQ ) such
that
Y
Y
GLn2 (Zp ),
GLn2 (Qp )
g −1 ⊆ GLn2 (R)
g∆(ρ(H(Q) ∩ GLn1 (Z[1/q0 ])))b
b
p∤q0
p|q0
where ∆ is the diagonal embedding. On the other hand, we have
Y
Y
GLn2 (Qp )
GLn2 (AQ ) = ∆(GLn2 (Q)) · GLn2 (R)
GLn2 (Zp ) .
p|q0
So there is g ∈ GLn2 (Q) such that
∆(gρ(H(Q) ∩ GLn1 (Z[1/q0 ]))g −1 ) ⊆ GLn2 (R)
Y
GLn2 (Qp )
p|q0
= ∆(GLn2 (Z[1/q0 ])),
p∤q0
Y
p∤q0
GLn2 (Zp ) ∩ ∆(GLn2 (Q))
which gives us the first part.
The second part is an easy consequence of the fact that ρ can be represented by a polynomial with rational
coefficients. Similarly there is a positive integer q ′ such that
|q ′ |p ks(ρ(hp )) − 1kp ≤ kρ(hp ) − 1kp ≤ |q ′ |−1
p khp − 1kp ,
(9)
for any prime p and hp ∈ H(Qp ). On the other hand,
hp ∈ ker(πpn ◦ ρ) if and only if kρ(hp ) − 1kp ≤ p−n .
(10)
These imply that for hp ∈ ker(πpn ◦ ρ) we have πpn−kp′ (s(ρ(hp ))) = 1. And so if hp ∈ Qp and πpn (ρ(hp )) = 1,
then we have
hp = (hp s(ρ(hp ))−1 ) · s(ρ(hp )) ∈ (L(Qp ) ∩ Qp ) ker πpn−kp′ .
(11)
Let φe1 : Qp → πpnp −kp′ (Qp )/πpnp −kp′ (L(Qp ) ∩ Qp ), φe1 (hp ) := πpnp −kp′ (hp )πpnp −kp′ (L(Qp ) ∩ Qp ). Then clearly
φe1 is surjective. By (11), φe1 factors through πpnp (ρ(Qp )); that means
πpnp (ρ(Qp )) ։ πpnp −kp′ (Qp )/πpnp −kp′ (L(Qp ) ∩ Qp ).
Let φe2 : Qp → πpnp (ρ(Qp )), φe2 (hp ) := πpnp (ρ(hp )). So clearly φe2 is surjective. By (9) and (10), we have, if
′
′
−np −kp
kp
= p−np ; and so hp ∈ ker φe2 . Therefore
πpnp +kp′ (hp ) = 1, then kρ(hp ) − 1kp ≤ |q ′ |−1
p khp − 1kp ≤ p p
φe2 factors through π np +k′ (Qp ). By the definition of φe2 we have L(Qp ) ∩ Qp ⊆ ker φe2 . Hence we get a
p
p
surjection πpnp +kp′ (Qp )/πpnp +kp′ (L(Qp ) ∩ Qp ) ։ πpnp (ρ(Qp )). The claim follows.
Let Gs be a simply connected semisimple Q-group which acts on a unipotent Q-group U. Then by the virtue
of the proof of [SGV12, Corollary 14] G = Gs ⋉ U is a perfect group if and only if Gs ⋉ U/[U, U] is perfect.
On the other hand, V := U/[U, U] is a vector Q-group, and the action of Gs on U induces a representation
φ0 : Gs → GL(V) defined over Q. And Gs ⋉ V is perfect if and only if the trivial representation is not a
subrepresentation of φ0 . Consider the following short exact sequences of Q-algebraic groups
(12)
ρ1
1 → U → G −→ Gs → 1
and
ρ2
1 → [U, U] → G −→ Gs ⋉ V → 1.
12
ALIREZA SALEHI GOLSEFIDY
We notice that, since the U is a Q-unipotent group, there is a Q-section from V to U, which can be extended
to a Q-section for the second exact sequence of (12). The first exact sequence clearly splits. We fix certain
Q-embeddings Gs ⊆ G ⊆ (GLn1 )Q and Gs ⋉ V ⊆ (GLn2 )Q . Let Γ ⊆ G ∩ GLn1 (Z[1/q0 ]) be a Zariski-dense
subgroup. Then by Lemma 13 after passing to a finite index subgroup, if necessary, we can assume that
ρ2 (Γ) ⊆ GLn2 (Z[1/q0 ]). By the above comment, we are also allowed to use (8) in Lemma 13 for the p-adic
closure of Γ in G(Qp ) for any p ∤ q0 .
2.5. Ping-pong players. As in [SGV12, SG], we work with a Zariski-dense free subgroup of Γ. Following
[SG, Proposition 22], using [SGV12, Proposition 7, Proposition 17, Proposition 20] we get the following.
Proposition 14. Let Ω be a finite symmetric subset of GLn (Q) which generates a Zariski-dense subgroup
Γ of a Zariski-connected perfect group G. Then there are a finite subset Ω′ of Γ and δ0 > 0 and l0 (which
depend on Ω) such that Ω′ freely generates a Zariski-dense subgroup of G(Q) and
(13)
(l)
PΩ′ (H(F )) ≤ e−δ0 l
−1
for any field extension F/Q, and any proper subgroup H ( G ×Q F and l ≥ l0 , where Ω′ = Ω′ ∪ Ω′ and PΩ′
is the probability counting measure on Ω′ . Moreover, Ad(ρ1 (Ω′ )) freely generates a subgroup of Ad(Gs (Q)),
where ρ1 : G → Gs is the quotient map; in particular, ker(Ad ◦ρ1 ) ∩ hΩ′ i = {1}.
Proof. This is essentially proved in [SGV12, Proposition 7, Proposition 17, Proposition 20]. Here is an
outline of the argument.
By [SGV12, Proposition 17] there are irreducible representations ηi of G :→ GL(Wi ) defined over local fields
Fi and algebraic deformations φi,v : G → Aff(Wi ) of ηi where the parameter v varies in an affine space Vi
with the following properties:
(1)
(2)
(3)
(4)
(5)
for any i, ηi factors through Gs , and at least one of them factors through Ad(Gs );
for any i, ηi (Γ) ⊆ GL(Wi (Fi )) is unbounded;
the linear part of φi,v is ηi for any v ∈ Vi (Fi );
for any non-zero vector v, the affine action of G(Fi ) on Wi (Fi ) via φi,v has no fixed point;
for any proper algebraic subgroup H of G for some i either
(a) there is a non-zero vector w ∈ Wi (Fi ) such that ηi (H(Fi ))[w] = [w] where [w] is the line Fi w,
or
(b) there is a parameter v ∈ Vi (Fi ) of norm 1 and a point w ∈ Wi (Fi ) such that w is fixed by H
under the affine action of G on Wi given by φi,v .
By [SGV12, Proposition 20 and Proposition 21], there is a finite subset Ω′ ⊆ Γ such that
(1) for any i, ηi (Ω′ ) freely generates a subgroup of ηi (Γ); in particular Ad(ρ1 (Ω′ )) freely generates a
subgroup of Ad(Gs (Q)) where ρ1 : G → Gs is the projection to the semisimple part.
(2) there is c0 > 0 such that, for any index i, and non-zero vector w0 ∈ Wi (Fi ), we have
|{γ ∈ Bl (Ω′ )| ηi (γ)[w0 ] = [w0 ]}| ≤ |Bl (Ω′ )|1−c0 ,
and, for any point w1 ∈ Wi (Fi ), and parameter vi ∈ Vi (Fi ) with norm one,
|{γ ∈ Bl (Ω′ )| φi,v (γ)(w1 ) = (w1 )}| ≤ |Bl (Ω′ )|1−c0 ,
′
where Bl (Ω′ ) is the the set of reduced words over Ω of length l.
Hence, by the above geometric description of proper algebraic subgroups, for any proper algebraic subgroup
H of G we have
|Bl (Ω′ ) ∩ H(Q)| < |Bl (Ω′ )|1−c0 .
SUPER-APPROXIMATION, II:
THE p-ADIC AND BOUNDED POWER OF SQUARE-FREE INTEGERS CASES.
−1
(l)
′
13
l
|−1
, where I is
. By a result of Kesten [Kes59, Theorem 3] we have PΩ′ (I) ≤ 2|Ω
|Ω′ |2
P (l) 2
(2l)
′
the identity matrix. As Ω is symmetric, we have PΩ′ (I) = γ PΩ′ (γ) . Hence together with the Kesten
bound we have
l
2|Ω′ | − 1
(2l)
(l)
(14)
PΩ′ (γ) ≤ PΩ′ (I) ≤
.
|Ω′ |2
Let Ω′ := Ω′ ∪ Ω′
We also notice that, since the random-walk in the group generated by Ω′ with respect to the probability
(l)
counting measure PΩ′ on Ω′ can be identified with the random-walk on a regular tree, PΩ′ is constant on
(l)
the set Bk (Ω′ ) of reduced words in Ω′ of length k for any positive integer k. Let pl,k := PΩ′ (γ) for some
γ ∈ Bk (Ω′ ).
For a proper subgroup H of G ×Q F , let H be the Zariski-closure of Γ ∩ H(F ) in G. Then H is a subgroup of
(l)
(l)
G and PΩ′ (H(F )) = PΩ′ (H(Q)). Now we can finish the argument as in the proof of [SGV12, Proposition 7]:
X
(l)
(l)
PΩ′ (H(F )) = PΩ′ (H(Q)) =
|BΩ′ (k) ∩ H(Q)| pl,k
≤
k≥0
X
0≤k<l/10
(2|Ω′ |)l/10
(2|Ω′ |)k
2|Ω′ | − 1
|Ω′ |2
l
+ (2|Ω′ | − 1)−c0 l/10
X
|Bk (Ω′ )|pl,k
k≥l/10
+ (2|Ω′ | − 1)−c0 l/10 ≤ e−δ0 l ,
(|Ω′ |/2)l
for some δ0 and any l ≥ l0 where l0 is a fixed positive integer.
≤
Lemma 15. Let Ω be a finite symmetric subset of GLn0 (Z[1/q0 ]) and Γ = hΩi. Assume the Zariski-closure
G of Γ is a connected, simply connected, perfect Q-group. Suppose Ω′ ⊆ Γ is a finite symmetric set which
generates a Zariski-dense subgroup Λ of G. Then for any set of positive integers C consisting of integers
coprime to q0 if {Cay(πq (Λ), πq (Ω′ ))}q∈C is a family of expanders, then {Cay(πq (Γ), πq (Ω))}q∈C is a family
of expanders.
b of Λ in
Proof. By Nori’s strong approximation (see [Nor89, Theorem 5.4]), we have that the closure Λ
Q
Q
b
p∤q0 GLn0 (Zp ). For any q, any representation ρ
p∤q0 GLn0 (Zp ) is of finite index in the closure Γ of Γ in
b The extension of ρ to Γ
b is denoted by ρ too. Only finitely
of πq (Γ) can be viewed as a representation of Γ.
b
many of them have Λ in their kernel. Hence the restriction to Λ of only finitely many such representation
is trivial. Let ρ be an irreducible representation of πq (Γ) whose restriction to
PΛ is not trivial. Hence the
restriction of ρ to Λ is a subrepresentation of L2 (πq (Λ))◦ := {f ∈ L2 (πq (Λ))|
g∈πq (Λ) f (g) = 0}.
Let u ∈ Vρ be a unit vector which is an ε-almost invariant vector with respect to Ω; that means
kρ(γ)(u) − uk < ε,
Q
for any γ ∈ Ω. Since Ω generates Γ and Ω ⊆ Γ, for some r0 we have Ω′ ⊆ r0 Ω. Thus for any λ ∈ Ω′ we
have
kρ(λ)(u) − (u)k < r0 ε.
′
Assuming {Cay(πq (Λ), πq (Ω ))}q∈C is a family of expanders, there is ε0 > 0 such that
′
max′ kρ(λ)(u) − uk > ε0 .
λ∈Ω
Therefore ε > ε0 /r0 , which implies that no non-trivial irreducible representation of πq (Γ) (for q ∈ C) has an
ε′ -almost invariant vector for some ε′ > 0.
Corollary 16. It is enough to prove the sufficiency part of Theorem 1 under the additional assumptions:
(1) Ω = Ω ∪ Ω
−1
where Ad(ρ1 (Ω)) freely generates a Zariski-dense subgroup of Ad(Gs ).
14
ALIREZA SALEHI GOLSEFIDY
(2) For any field extension F/Q and proper subgroup H of G ×Q F , (13) holds.
Remark 17. Notice that by changing Ω, the closure Γp of Γ in GLn0 (Zp ) is of the form described in
Lemma 12 only for large enough p. For small primes Γp is still an open subgroup of G(Qp ).
2.6. Summary of the initial reductions. In the rest of the article, (unless it is explicitly mentioned
otherwise) we assume that Ω ⊆ GLn0 (Z[1/q0 ]) is a finite symmetric set, and set Γ = hΩi. And let G and G
be the Zariski-closure of Γ in (GLn0 )Z[1/q0 ] and (GLn0 )Q , respectively.
(1) (Zariski closure) By Lemma 11, we can and will assume that G is a connected, simply connected,
perfect group. We let Gs be a Levi Q-subgroup of G, and U be the unipotent radical of G.
b
(2) Q
(Adelic closure) By Lemma 12 and Remark
Q 17, we can and will assume that the closure Γ of Γ in
G(Q
)
which
contains
a
finite
index
subgroup
of the
(Z
)
is
an
open
subgroup
of
GL
p
p
p∤q0
p∤q0 Q n0
form p∤q0 Pp where Pp = Kp ⋉ Up and Kp and Up satisfy the properties mentioned in Lemma 12.
In particular, Pp = G(Qp ) ∩ GLn0 (Zp )[pkp ], Kp = Gs (Qp ) ∩ GLn0 (Zp )[pkp ], and Up = U(Qp ) ∩
GLn0 (Zp )[pkp ] for some non-negative integers kp that are zero for large enough p.
(3) (Generating set) We can and will assume that Ω is as in Corollary 16. In particular, ker(Ad ◦ρ1 )∩Γ =
{1}, and any proper algebraic subgroup of G intersects Γ in an exponentially small subset (see
Inequality (13)).
2.7. Multiplicity bound. In order to execute Sarnak-Xue [SX91] trick, we need to have a control on the
degree of irreducible representations of πq (Γ). Here we prove such a result for an arbitrary modulus q which
is of independent interest.
Proposition 18. Let Γ ⊆ GLn (Z[1/q0 ]) be a finitely generated group whose Zariski-closure G is a Zariskiconnected perfect group. Suppose G is simply-connected; that means G = Gs ⋉ U where Gs is a simplyb
connected semisimple Q
group and U is a unipotent group. Let q be an integer which is co-prime to q0 . Let Γ
be the closure of Γ in p∤q0 GLn (Zp ).
b and δ > 0 which depend on Γ such
Then there are a finite set Σ of complex irreducible representations of Γ
that the following holds.
For any complex irreducible representation ρ of πq (Γ) which does not factor through πq′ (Γ) for any proper
divisor q ′ of q, we have that either
b is in Σ, or
(1) The lift of ρ to Γ
(2) dim ρ ≥ |πq (Γ)|δ .
For a group with entries in a number field and a semisimple Zariski-closure, a similar statement is proved
in [SG, Proposition 33]. An identical argument works here, but for the sake of the convenience a proof is
included. It is worth pointing out that the main (behind the scene) representation theoretic tool is Howe’s
Kirillov theory for compact p-adic analytic groups [How77].
For any odd prime p, let a0 (p) := 1, and let a0 (2) := 2. Then exp : pa0 (p) gln (Zp ) → 1 + pa0 (p) gln (Zp ), and
log going backward, can be defined and satisfy the usual properties. Let b0 (p) := 1 for any prime p ≥ 5, and
let b0 (2) = b0 (3) = 2. Then, by Campbell-Hausdorff formula, we have that for any Lie Zp -subalgebra h of
pa0 (p) gln (Zp ) we have that exp(h) is a subgroup of 1 + pa0 (p) gln (Zp ).
Lemma 19. Let H be an algebraic subgroup of (GLn )Qp . Let Hc := H(Qp ) ∩ (1 + pc gln (Zp )) for any
positive integer c. Let h := Lie(H)(Qp ) ∩ gln (Zp ). Suppose ps0 h ⊆ [h, h]. Let ρ be an irreducible complex
representation of Hc , c ≥ b0 (p) + 1, and
mρ := min{m ∈ Z+ | Hm ⊆ ker(ρ)}.
Then dim ρ ≥ p(mρ −1−c−s0 )/2 .
SUPER-APPROXIMATION, II:
THE p-ADIC AND BOUNDED POWER OF SQUARE-FREE INTEGERS CASES.
15
Proof. Since the analytic Lie algebra of H(Qp ) coincides with Lie(H)(Qp ), we get that
log : Hm → pm h
is a bijection and the exponential map is its inverse (for a detailed proof see [SGZ, Proposition 7]). We
notice also that pm+s0 (pm h) ⊆ [pm h, pm h] ⊆ pm (pm h). Hence by [SG, Lemma 32] we have
dim ρ ≥ p(mρ −1−c−s0 )/2 .
Proof of Proposition 18. Let Gp,m := G(Qp )∩(1 + pm GLn (Zp )) for any prime p and non-negative integer m.
m
b be the closure of Γ in Q
Let Γ
p∤q0 GLn (Zp ), let Γp be the closure of Γ in GLn (Zp ), and Γp [p ] := Γp ∩ Gp,m .
As in Lemma 12, by Nori’s strong approximation [Nor89, Theorem 5.4], there is a finite-index subgroup
Q
b in Q
Λ of Γ such that its closure Λ
p∤q0 Λp , where Λp ’s satisfy the following
p∤q0 GLn (Zp ) is of the form
properties.
(1) For some positive integer c0 we have that, for any prime p ≥ c0 , Λp = Gp,0 and it is a perfect group,
and πp (Λp ) is generated by p-elements;
(2) For any prime p < c0 , we have Λp = Gp,c(p) where c(p) is chosen large enough such that Λp ⊆ Gp,8 .
Since G is perfect, [Lie(G)(Q), Lie(G)(Q)] = Lie(G)(Q). Hence there is a positive integer s0 such that for
any prime p we have
ps0 gp ⊆ [gp , gp ]
where gp := Lie(G)(Qp ) ∩ gln (Zp ).
Let ρ : πq (Γ) → GL(V ) be a complex irreducible representation which does not factor through πq′ (Γ) for any
proper divisor q ′ of q.QThen V = ⊕i ρ(gi )W where W is a simple πq (Λ)-module
Q and gi ’s are some elements of
πq (Γ). Suppose q = i pni i is the prime factorization of q. Since πq (Λ) ≃ i πpni i (Λpi ), there are irreducible
b
In particular,
representations
ηi : Λpi → GL(Wi ) such that W is isomorphic to ⊗i Wi as an Λ-module.
Q
dim W = i dim Wi .
b we have that Λ
b acts non-trivially on V , and so
Except for finitely many irreducible representations ρ of Γ,
b Suppose that
⊗i ηi is a non-trivial irreducible representation of Λ.
b Λ
b ⊆ ker(ρ)},
Σ ⊇ {ρ ∈ Irr(Γ)|
b is the set of all the complex irreducible representations of Γ.
b
where Irr(Γ)
For any i, ηi is an irreducible representation of Λpi = Gpi ,c(pi ) where c(pi ) is 0 for p ≥ c0 and given as above
for c < p0 . Since ρ does not factor through πq′ (Γ) for a proper divisor q ′ of q, for any i we have
mi := min{m ∈ Z+ | Gpi ,m ⊆ ker(ηi )} = max{ni , c(pi )}.
Next we partition the indexes into five types: (1) ni − c(pi ) ≥ 2s0 + 2, and c(pi ) ≥ 8; (2) ni ≥ 2s0 + 10 and
c(pi ) = 0; (3) 1 < ni < 2s0 + 10 and pi ≥ c0 ; (4) ni = 1 and pi ≥ c0 ; and (5) pi < c0 and ni < 2s0 + 10.
Now we find a lower bound for the dimension of ηi for the first three types of indexes.
Case 1. ni − c(pi ) ≥ 2s0 + 2 and c(pi ) ≥ 8.
In this case, by Lemma 19, we have
dim ηi ≥ p(ni −c(pi )−1−s0 )/2 ≥ p(ni −c(pi ))/4 = |πpni (Λpi )|δ1
i
where δ1 is a positive number which depends only on dim G.
Case 2. ni ≥ 2s0 + 10 and c(pi ) = 0.
16
ALIREZA SALEHI GOLSEFIDY
In this case, the restriction of ηi to Gpi ,2 is non-trivial. Hence by Lemma 19 we have
dim ηi ≥ p(ni −2−1−s0 )/2 ≥ pni /4 ≥ |πpni i (Λpi )|δ2
where δ2 is a positive number which depends only on dim G.
Case 3. 1 < ni < 2s0 + 10 and p ≥ c0 .
Notice that πpni (Gpi ,1 ) is a normal pi -subgroup of πpni (Λpi ). Since the dimension of any irreducible repi
i
resentation of a pi -group is a power of pi , we have that either the restriction of ηi to Gpi ,1 factors through
Gpi ,1 /[Gpi ,1 , Gpi ,1 ] or dim ηi ≥ pi ≥ |πpni (Λpi )|δ3 , where δ3 is a positive number which depends on G.
i
Now suppose the restriction of ηi to Gpi ,1Pfactors through Gpi ,1 /[Gpi ,1 , Gpi ,1 ]. Then there is a line li in Wi
which is invariant under Gpi ,1 and Wi = g∈Gp ,0 ηi (g)(li ). Since Gpi ,0 is perfect, the stabilizer of li should
i
be a proper subgroup of Gpi ,0 . Since πpni i (Gpi ,0 ) is generated by pi -elements, we get that the index of the
stabilizer of li in Gpi ,0 is at least pi . And so again dim ηi ≥ pi ≥ |πpni (Λpi )|δ3 .
i
Case 4. ni = 1 and pi ≥ c0 .
By [LS74] (see [SGV12, Section 4.2]), we have dim ηi ≥ |πpni i (Λpi )|δ4 where δ4 is a positive number which
depends only on G.
Q
Since i in Case 5 pni i ≪G 1, by the above cases we get that either q ≪Γ 1 or dim ρ ≥ |πq (Γ)|δ where δ is a
positive number which depends only on Γ. By enlarging Σ, we can assume that it contains all the irreducible
b which factor through πq (Γ) for a positive integer q ≪Γ 1. Such Σ and δ satisfy the
representations of Γ
claimed properties.
3. Expansion, approximate subgroup, and bounded generation.
3.1. Largeness of level-Q approximate subgroups implies super-approximation. In this section,
the Bourgain-Gamburd machine [BG08-a] is used to reduce a proof of spectral gap at level Q to a proof of
the largeness of a generic level-Q approximate subgroup (see Theorem 20).
We work in the setting of Section 2.6. As in [SG], let us introduce the following notation for convenience:
for a finite symmetric subset A of Γ, positive integers l and Q, and a positive number δ, let PQ (δ, A, l) be
the following statement
1
log Q) ∧ (|πQ (A · A · A)| ≤ |πQ (A)|1+δ ).
δ
One can think of A as a generic level-Q approximate subgroup if PQ (δ, A, l) holds for small enough δ.
(15)
(l)
(PΩ (A) > Q−δ ) ∧ (l >
Notice that the random-walk in πQ (Γ) gets interesting only at steps OΩ (log |πQ (Γ)|) = OΩ (log Q) since
before this time the random-walk with respect to the probability measure PπQ (Ω) is similar to a randomwalk in a regular tree. The first two conditions of PQ (δ, A, l) imply that at OΩ (log Q)-range of random-walk
the probability of hitting A is not exponentially small. And so A should be fairly generic. For instance
Proposition 14 implies that A cannot be a subset of a proper algebraic subgroup if δ is small enough. The
genericness of such A was crucial in the proof of [SG, Theorem 35] where the largeness of a generic level-Q
approximate subgroup for a semisimple group is proved.
Let C be either {pn |p ∈ Vf (Q), p ∤ q0 } or {q ∈ Z+ | gcd(q, q0 ) = 1, ∀p ∈ Vf (Q), vp (q0 ) ≤ N }, where vp (q) is
the p-adic valuation of q.
In this section, we deduce Theorem 1 from the following result.
Theorem 20 (Approximate subgroups). In the above setting, for any ε > 0 there is δ > 0 such that
PQ (δ, A, l) implies that |πQ (A)| ≥ |πQ (Γ)|1−ε
SUPER-APPROXIMATION, II:
ΘG (1)
if Q ∈ C and Qε
THE p-ADIC AND BOUNDED POWER OF SQUARE-FREE INTEGERS CASES.
17
≫Ω 1.
To show that Theorem 20 implies Theorem 1, we prove the following proposition based on the same proof as
in [SG, Section 3.2]. In particular, we see that the reduction to understanding generic level-Q approximate
subgroups can be done in much more generality.
−1
Proposition 21. Let Ω ⊆ GLn (Z[1/q0 ]) be a finite subset and Ω := Ω ∪ Ω . Suppose Ω freely generates a
subgroup Γ. Let C ′ be a family of positive integers which consists of numbers that are corpime to q0 . Suppose
C ′ is closed under divisibility; that means if q ∈ C ′ and q ′ |q, then q ′ ∈ C ′ . Suppose the assertion of Theorem 20
holds for Ω and C ′ . Then the family of Cayley graphs {Cay(πq (Γ); πq (Ω))}q∈C ′ is a family of expanders.
Proof. By Remark 2, we have to show supq∈C ′ λ(Pπq (Ω) ; πq (Γ)) < 1. For Q ∈ C ′ , let TQ : L2 (πQ (Γ)) →
L2 (πQ (Γ)), TQ (f ) := πQ [PΩ ] ∗ f and λ(Q) := λ(PπQ (Ω) ; πQ (Γ)). It is well-known that λ(Q) is the second
largest eigenvalue of TQ . Since L2 (πQ (Γ)) is a completely reducible representation of πQ (Γ), there is a unit
function f0 ∈ L2 (πQ (Γ)) such that TQ (f0 ) = λ(Q)f0 and the πQ (Γ)-module Vρ generated by f0 is a simple
πQ (Γ)-module. Since λ(Q) is not one, ρ is not the trivial representation of πQ (Γ). Since there are dim ρ-many
copies of ρ as subrepresentations of L2 (πQ (Γ)), we have that the eigenvalue λ(Q) of TQ has multiplicity at
least dim ρ. Since Q ∈ C ′ and C ′ is closed under divisibility, we can and will assume that ρ does not factor
through πq (Γ) for some proper divisor q of Q. Hence by Proposition 18 we have that either the lift of ρ to
ε0
b the closure of Γ in Q
where ε0
a representation of Γ,
p∤q0 GLn (Zp ), is in a finite set Σ or dim ρ ≥ |πQ (Γ)|
is a positive number depending only on Γ. As we are seeking a uniform upper bound for λ(Q), without loss
of generality we can and will assume that the lift of ρ is not in Σ. And so we have that multiplicity of the
eigenvalue λ(Q) of TQ is at least |πQ (Γ)|ε0 . Therefore for any positive integer l we have
(l)
|πQ (Γ)|ε0 λ(Q)2l ≤ Tr(TQ2l ) = |πQ (Γ)|kPπQ (Ω) k22 .
So it is enough to show
(16)
1
(l)
kPπQ (Ω) k2 ≤ |πQ (Γ)|− 2 +
ε0
4
,
for some l ≪Ω log Q where the implied constant depends only on Ω; it is worth pointing out that we can
and will assume that Q is arbitrarily large depending only on Ω. In particular, we can and will assume the
Largeness of level Q-Approximate Subgroups in the following sense:
The LAS Assumption: there is c2 > 0 such that, if Q > 1/c2 and PQ (c2 , A, l) holds, then |πQ (A)| ≥
|πQ (Γ)|1−ε0 /8 .
On the other hand, the first OΩ (log Q) steps of the random-walk with respect to the probability measure
PπQ (Ω) can be identified with the random-walk with respect to the probability measure PΩ . And the latter
is the random-walk on a regular infinite tree. Hence by the Kesten bound [Kes59, Theorem 3] we have
l
2|Ω| − 1
(l)
(2l)
kPΩ k22 = PΩ (I) ≤
= e−c1 l ,
|Ω|2
where c1 is a positive number which depends only on Ω. Hence for some positive integer l0 = ΘΩ (log Q)
where the implied constants depend only on Ω and for some positive number c3 depending on Ω we have
(17)
(l )
(l )
kPπQ0 (Ω) k2 = kPΩ 0 k2 ≤ |πQ (Γ)|−c3 .
Equation (17) can be considered as an initial L2 -flatness.
An important result of Bourgain and Gamburd roughly indicates that the L2 -norm of a probability measure should substantially drop after a convolution by itself unless an approximate subgroup has a lot of
mass [BG08-a] (see also [Var12, Lemma 15]). This result is proved using Tao’s non-commutative version of
Balog-Szemerédi-Gowers theorem [Tao08]. Here is a precise formulation of their result:
18
ALIREZA SALEHI GOLSEFIDY
Proposition 22 (Bourgain-Gamburd’s L2 -flattening.). Let µ be a symmetric probability measure with finite
1
support on a group G. Let K be a number more than 2. If kµ∗ µk2 ≥ K
kµk2 , then there is a finite symmetric
subset A of G which satisfies the following properties:
−2
−R
R
• (Size of A is comparable with kµk−2
kµk−2
2 ) K
2 ≤ |A| ≤ K kµk2 ;
R
• (Getting an approximate subgroup) |A · A · A| ≤ K |A|;
• (Almost equidistribution on A) mina∈A (µ ∗ µ)(a) ≥ K −R |A|−1 ,
where R is an absolute positive number.
For a small (depending only on Ω and specified later) positive parameter δ, consider the following pseudocode:
Algorithm 1 L2 -flattening process
Give Q ∈ C ′ , δ ∈ (0, 1);
l0 := ⌊log Q/2 log C(Ω)⌋; (as in Equation 17)
(l )
µ0 := PπQ0 (Ω) ;
i := 0;
i
while (kµi k2 > |πQ (Γ)|−1/2+ε0 /4 and kµi ∗ µi k2 ≤ kµi k1+δ
2 ) or (2 l0 < (1/δ) log Q) do
µi+1 := µi ∗ µi ;
i := i + 1;
end while
where C(Ω) := maxs∈Ω ksk, where ksk is the operator norm of s : Rn ×
k(v∞ , vp )p|q0 k := max{kv∞ k, kvp kp }p|q0 and the sup norm on Rn and Qnp .
Q
p|q0
Qnp → Rn ×
Q
p|q0
Claim 1. For any 0 < δ < 1 and Q ∈ C ′ , the number i0 (δ, Q) of steps in Algorithm 1 is at most
log( 2 logδC(Ω) )
Qnp with
log(2/c3 )
log(1+δ)
+
where c3 is given in Equation (17).
Proof of Claim 1. By the Young inequality we have kµi+1 k2 ≤ kµi k2 for any non-negative integer i. Hence
by the initial flatness, Equation (17), we have kµi k2 ≤ kµ0 k2 ≤ |πQ (Γ)|−c3 for any non-negative integer
i. Let j0 be the smallest integer such that 2j0 l0 ≥ (1/δ) log Q. Then for any j0 ≤ i < i0 (δ, Q) we have
(1+δ)
kµi+1 k2 ≤ kµi k2
. And so
(1+δ)i0 −j0
kµi0 (δ,Q) k2 ≤ kµj0 k2
If i0 (δ, Q) >
log(2/c3 )
log(1+δ)
+ log( 2 logδC(Ω) ), then i0 − j0 >
log(2/c3 )
log(1+δ)
.
which implies kµi0 −1 k2 ≤ |πQ (Γ)|−1 . And this
contradicts kµi0 −1 k2 > |πQ (Γ)|−1/2+ε0 /4 .
Claim 2. There is a positive number δ := δ(Ω) depending only on Ω, such that, for any Q ∈ C ′ which is at
least 1/δ, kµi0 (δ,Q) k2 ≤ |πQ (Γ)|−1/2+ε0 /4 , where i0 (δ, Q) is given in Claim 1.
Proof of Claim 2. We proceed by contradiction. So to the contrary assumption for any positive number δ
there is Q(δ) ∈ C ′ such that
(18)
(19)
(20)
Q(δ) > 1/δ; in particular lim Q(δ) = ∞,
kµi0 (δ,Q(δ)) k2 > |πQ(δ) (Γ)|
kµi0 (δ,Q(δ)) ∗ µi0 (δ,Q(δ)) k2 ≥
δ→0
−1/2+ε0 /4
kµi0 (δ,Q(δ)) k1+δ
2 .
, and
SUPER-APPROXIMATION, II:
THE p-ADIC AND BOUNDED POWER OF SQUARE-FREE INTEGERS CASES.
19
Hence by (20) and Proposition 22 (for K = kµi0 (δ,Q(δ)) k−δ
2 ) there is a symmetric subset A(δ) of πQ(δ) (Γ)
such that
(21)
kµi0 (δ,Q(δ)) k−2+Rδ
≤ |A(δ)| ≤ kµi0 (δ,Q(δ)) k−2−Rδ
,
2
2
(22)
|A(δ)|, and
|A(δ) · A(δ) · A(δ)| ≤ kµi0 (δ) k−Rδ
2
(23)
µi0 (δ,Q(δ))+1 (A(δ)) ≥ kµi0 (δ,Q(δ)) kRδ
2 .
Hence there is a function δ ′ (δ) of δ such that
lim δ ′ (δ) = 0 and δ ′ (δ) ≥ δ,
(24)
δ→0
′
(25)
(because of Inequalities (21) and (22),)
|A(δ) · A(δ) · A(δ)| ≤ |A(δ)|1+δ
(δ)
(26)
(because of Inequalities (19) and (23),)
µi0 (δ,Q(δ))+1 (A(δ)) ≥ Q(δ)−δ (δ) .
′
(2j l (δ))
(l(δ))
−1
(A(δ)) ∩ supp(PΩ
We notice that µj = PπQ(δ)0 (Ω) , and let A(δ) := πQ(δ)
then we have
), where l(δ) := 2i0 (δ,Q(δ))+1 l0 (δ);
(27)
A(δ) is a finite symmetric subset of Γ because A(δ) and PΩ are symmetric,
(28)
l(δ) ≥ (1/δ) log Q(δ) ≥ (1/δ ′ (δ)) log Q(δ)
because of the second condition of the loop in Algorithm 1 and the inequality in (24),
′
(l(δ))
PΩ(δ) (A(δ)) = µi0 (δ)+1 (A(δ)) ≥ Q(δ)−δ (δ) by Inequality (26),
(29)
|πQ(δ) (A(δ) · A(δ) · A(δ))| ≤ |πQ(δ) (A(δ))|1+δ
(30)
′
(δ)
by Inequality (25).
′
Statements (27), (28), (29), and (30) imply that PQ(δ) (δ (δ), A(δ), l(δ)) holds for any δ. Now suppose δ is
small enough so that δ ′ (δ) is less than c2 where c2 is given by the LAS Assumption (this can be done thanks
to (24)). Then Q(δ) > 1/δ ≥ 1/δ ′ (δ) ≥ 1/c2 , and so by the LAS Assumption we have
|πQ(δ) (A(δ))| = |A(δ)| ≥ |πQ(δ) (Γ)|1−ε0 /8 .
(31)
On the other hand, by Inequalities (21) and (19), we have
≤ |πQ(δ) (Γ)|1+Rδ/2−ε0 /2−Rδε0 /4 .
|A(δ)| ≤ kµi0 (δ,Q(δ)) k−2−Rδ
2
Hence, if δ is small enough, we get |A(δ)| ≤ |πQ(δ) (Γ)|1+Rδ/2−ε0 /2−Rδε0 /4 ≤ |πQ(δ) (Γ)|1−ε0 /4 , which contradicts (31).
(l)
(Going back to the proof of Proposition 21) By Claim 1 and Claim 2, for Q ≫Ω 1 in C ′ we have kPπQ (Ω) k2 ≤
|πQ (Γ)|−1/2+ε0 /4 where l := 2i0 (δ(Ω),Q) l0 ≪Ω log Q. And this gives us the needed L2 -norm upper bound (see
Inequality (16)) for Q ≫Ω 1. The rest are only finitely many Q’s for which we do have spectral gap.
3.2. Reduction to bounded generation and abelian unipotent radical. The main goal of this section
is to deduce Theorem 20 from the following.
Theorem 23 (Bounded generation). In the setting of Section 2.6, suppose U is abelian. For any positive
integer q, let Γ[q] := Γ ∩ GLn0 (Z[1/q0 ])[q], where
πq
GLn0 (Z[1/q0 ])[q] := ker(GLn0 (Z[1/q0 ]) −→ GLn0 (Z[1/q0 ]/qZ[1/q0 ])).
Then for any 0 < ε ≪Ω 1 there are 0 < δ ≪Ω,ε 1 and positive integer C ≫Ω,ε 1 such that for a finite
symmetric subset A of Γ
Q
(l)
PΩ (A) > Q−δ for some l > (1/δ) log Q implies πQ (Γ[q]) ⊆ C πQ (A) for some q|Q and q ≤ Qε
OG (1)
if Q ∈ C and Qε
≫Ω 1.
20
ALIREZA SALEHI GOLSEFIDY
The idea of the proof of Theorem 20 modulo Theorem 23. It is not clear to the author if a bounded
generation result is true when the unipotent radical U is not abelian (specially since a perfect group might
have a non-trivial center.). So to prove the largeness of a generic level-Q approximate subgroup A of Γ
with Zariki-closure G := Gs ⋉ U, first we go to the quotient Gs ⋉ U/[U, U]. Let ρ2 : G → Gs ⋉ U/[U, U]
be the quotient map. For simplicity let us assume ρ2 commutes with the congruence maps πQ . This
−1
assumption together with the way the generating set Ω is chosen (recall that we have Ω = Ω ∪ Ω and
ρ2 (Ω) freely generates a free group) implies that ρ2 (A) is a generic level-Q approximate subgroup of ρ2 (Γ).
Now by the bounded generation result for ρ2 (Γ), we get a large congruence subgroup of ρ2 (Γ) modulo Q in
a controlled number of steps. Next we use the techniques of studying a nilpotent group in order to translate
our information on U/[U, U] to some data on U. For a unipotent group U (and in general any nilpotent
group), it is a useful method to consider
LU := γ1 (U)/γ2 (U) ⊕ γ2 (U)/γ3 (U) ⊕ · · · ⊕ γc (U)/γc+1 (U),
where γi (U) is the i-th lower central series; that means γ1 (U) := U and γi+1 := [γi (U), U], and γc+1 (U) = {1}.
It is well-known that [gi γi+1 (U), gj γj (U)] := gi−1 gj−1 gi gj γi+j+1 (U) for gi ∈ γi (U) and gj ∈ γj (U) defines a
graded Lie ring structure on LU . And LU is generated by its degree 1 elements, i.e. γ1 (U)/γ2 (U), as a Lie
ring. So we get an onto multi-linear map
φk : γ1 (U)/γ2 (U) × · · · × γ1 (U)/γ2 (U) → γk (U)/γk+1 (U),
{z
}
|
φk (x1 , . . . , xk ) := [x1 , . . . , xk ],
k-times
where [x1 , . . . , xk ] := [[x1 , . . . , xk−1 ], xk ]. Having a set X which maps onto a set of representatives for a large
congruence subgroup modulo Q in γ1 (U)/γ2 (U), using the above multi-linear maps φk , we can get a subset
of U which
Q is large modulo Q. Notice that at this step we can get a control only on the size of the product
set πQ ( OU (1) X). Though we do get a set of representatives of a large congruence subset of the associated
gradedQLie ring, we do not get such a structural understanding of the product set itself. Altogether we get
that | C πQ (A)| is large. On the other hand, since |πQ (A) · πQ (A) · πQ (A)| ≤ |πQ (A)|1+δ , by the Ruzsa
inequality (see [Hel05, Lemma 2.2]), we get
|
Q
C
πQ (A)| ≤ |πQ (A)|
which helps us to finish the proof.
|
Q
3 πQ (A)|
|πQ (A)|
C−2
≤ |πQ (A)|1+(C−2)δ ,
Theorem 23 implies Theorem 20. By the contrary assumption, there is ε0 > 0 such that for any δ > 0 there
are a finite symmetric subset Aδ , a positive integer lδ and Qδ ∈ C such that PQδ (δ, Aδ , lδ ) holds and at
the same time |πQδ (Aδ )| < |πQδ (Γ)|1−ε0 . Since lδ > log Qδ /δ, we have limδ→0 lδ = ∞. Now we claim that
Qδ → ∞ as δ → 0. If not, for infinitely many δ we have Qδ = Q, and πQ (Aδ ) = A is independent of δ. And
so
(l )
lim PπQδ (Ω) (πQ (Aδ )) =
δ→0
1
|A|
.
<
|πQ (Γ)|
|πQ (Γ)|ε0
On the other hand, we have
(l )
(l )
−δ
.
PπQδ (Ω) (πQ (Aδ )) ≥ PΩ δ (Aδ ) > Q−δ
δ = Q
Hence as δ → 0, we get |πQ (Γ)|−ε0 > 1, which is a contradiction.
Let {kp }p∤q0 be such that Pp = G(Qp ) ∩ GLn0 (Zp )[pkp ] where {Pp }p is as in Section 2.6. Let ρ2 : G →
Gs ⋉ U/[U, U] be the quotient map. Let us recall that, by the discussion at the end of Section 2.4, we can
and will assume ρ2 (Pp ) ⊆ GLn2 (Zp ).
SUPER-APPROXIMATION, II:
Q
THE p-ADIC AND BOUNDED POWER OF SQUARE-FREE INTEGERS CASES.
21
L
′
kp
Let q0′ = p∤q0 p be a positive integer given by Lemma 13; that means for any sequence {mp }p∤q0 ∈ p∤q0 Z
such that mp ≥ kp + kp′ we have
(32)
Y
Y
Y
πpmp −kp′ (Pp )/πpmp −kp′ ([U, U](Qp )∩Pp ).
πpmp (ρ2 (Pp )) ։
πpmp +kp′ (Pp )/πpmp +kp′ ([U, U](Qp )∩Pp ) ։
p∤q0
p∤q0
p∤q0
In order to use (32), we modify Qδ so that for any prime p either vp (Qδ ) ≥ kp′ or vp (Qδ ) = 0. To this end,
Q
it is enough to show that we are allowed to change Qδ to Qδ /qδ′ where qδ′ := p∤q0 ,0<vp (Qδ )<2k′ +kp pvp (Qδ ) .
p
Lemma 24. Let Aδ , Qδ , and lδ be as above.
PQδ /qδ′ (ΘC ′ ,Ω (δ), Aδ , lδ ) holds and |πQδ /qδ′ (Aδ )|
Suppose qδ′ |Qδ and
≤ |πQδ /qδ′ (Γ)|1−ε0 /2
qδ′
′
< C (for some constant C ′ ). Then
for 0 < δ ≪C ′ 1.
1/C ′′
Proof. Since limδ→0 Qδ = ∞, for C ′′ > 1 and small enough δ we have Qδ /qδ′ > Qδ
. And so
log(Qδ /qδ′ )
log Qδ
>
.
δ
C ′′ δ
Q
′′
To show that PQδ /qδ′ (C ′′ δ, Aδ , lδ ) holds, it is enough to prove that we have |πQδ /qδ′ ( 3 Aδ )| ≤ |πQδ /qδ′ (Aδ )|1+C δ
for large enough C ′′ .
′′
′ −C δ
P (lδ ) (Aδ ) > Q−δ
, and
δ > (Qδ /qδ )
lδ >
We notice that, if c is small enough (depending on Ω), πQδ /qδ′ induces an injection on the ball of radius
c log(Qδ /qδ′ ) (with respect to the word metric). By [BG09, Remark, page 1060] (see [SG, Lemma 8]), we
′
have P (⌊2c log(Qδ /qδ )⌋) (Aδ · Aδ ) > Q−2δ
. Hence by the Kesten bound (see Inequality (14)) for small enough δ
δ
we have
(33)
|πQδ /qδ′ (Aδ · Aδ )| = |Aδ · Aδ ∩ B⌊2c log(Qδ /qδ′ )⌋ |
′
′
≥ P (⌊2c log(Qδ /qδ )⌋) (Aδ · Aδ )/kP (⌊2c log(Qδ /qδ )⌋) k∞
Θ
′
≥ Q−2δ
· (2c log(Qδ /qδ′ ))ΘΩ (1) ≥ Qδ C ,Ω
δ
Θ
′
(1)
,
(1)
which implies that |πQδ /qδ′ (Aδ )| ≥ Qδ C ,Ω .
Q
Q
Since |πQδ ( 3 Aδ )| ≤ | Qδ (Aδ )|1+δ , by the Ruzsa inequality (see [Hel05, Lemma 2.2]) and [Hel11, Lemma
7.4] we have
Q
Q
|πQδ /qδ′ ( 3 Aδ )|
|πQδ ( 5 Aδ )|
3δ
|πQδ (Aδ )| ≥
≥
,
|πQδ (Aδ )|
|πQδ /qδ′ (Aδ )|
Q
′′
which implies that |πQδ /qδ′ ( 3 Aδ )|/|πQδ /qδ′ (Aδ )| ≤ |πQδ /qδ′ (Aδ )|C δ for large enough C ′′ . Since Qδ goes to
infinity, it is clear that for small enough δ we have |πQδ /qδ′ (Aδ )| ≤ |πQδ /qδ′ (Γ)|1−ε0 /2 .
By Lemma 24 we can and will modify {Qδ }δ and assume that, if kp + kp′ > 0, then either vp (Qδ ) = 0 for
any δ or kp + 2kp′ ≤ minδ vp (Qδ ) < supδ vp (Qδ ) = ∞. In particular, without loss of generality we can and
will assume that q0′2 |Qδ for any δ.
(l )
(l )
On the other hand, by Corollary 16, we have that ker(ρ2 ) ∩ Γ = {1}. So Pρ2δ(Ω) (ρ2 (Aδ )) = PΩ δ (Aδ ) >
′
−ΘΩ (δ)
Q−δ
for suitable constant and small enough δ. Hence by Theorem 23 for any 0 < ε′ ≪Ω 1
δ > (q0 Qδ )
there is a positive integer C = OΩ,ε′ (1) such that for any 0 < δ ≪ε′ ,Ω 1 we have
Q
(34)
πq0′ Qδ (ρ2 (Γ)[qδ ]) ⊆ C πq0′ Qδ (ρ2 (Aδ ))
′
Θ(ε′ )
for some qδ |q0′ Qδ and qδ ≤ (q0′ Qδ )ε ≤ Qδ
.
Q
On the other hand we have πq0′ Qδ ( p∤q0 ρ2 (Up )) ⊆ πq0′ Qδ (ρ2 (Γ)) (recall that Up := U(Qp ) ∩ GLn0 (Zp )[pkp ].
Let U p := GLn2 (Zp ) ∩ U(Qp )); and so we have that U p = Up for large enough p, and Up is an open subgroup
of U p . By Lemma 24, we can and will assume that vp (Qδ ) are either zero or large enough depending on the
22
ALIREZA SALEHI GOLSEFIDY
index of Up in U p for all the primes p. And so by enlarging qδ by a multiplicative constant we can and will
assume that
Q
Q
πq0′ Qδ (ρ2 ( p∤q0 U p )qδ ) ⊆ πq0′ Qδ (ρ2 (Γ)[qδ ]) ⊆ C πq0′ Qδ (ρ2 (Aδ )),
Q
Q
Q
where ρ2 ( p∤q0 U p )qδ := {xqδ | x ∈ ρ2 ( p∤q0 U p )}. So there is Bδ ⊆ C Aδ such that
Y
(35)
πq0′ Qδ (ρ2 (Bδ )) = πq0′ Qδ (ρ2 (
U p )qδ ).
p∤q0
Lemma 25. Let U be a unipotent Q-group with a given Q-embedding in GLn . For any prime p, let U p :=
U(Qp ) ∩ GLn (Zp ). Let φ : U → U/[U, U] be the quotient map. Fix a Q-basis {ei } of (U/[U, U])(Q) such that
Q
P
φ(U p ) ⊆ i Zp ei for any prime p. Let q|Q be two positive integers. Suppose X is a finite subset of p U p
P
such πQ (φ(X)) = πQ (q i Zei ). Then
Q
Q
|πQ ( p U p ))|
,
| C πQ (X)| ≥
CqC
where C depends on the dimension of U and the choice of embeddings.
Proof. Let γi (U) be the i-th lower central series; that means γ1 (U) := U and γi+1 (U) := [γi (U), U]. It is
well-known that the long commutators (g1 , . . . , gk ) := ((g1 , . . . , gk−1 ), gk ), where (g1 , g2 ) := g1−1 g2−1 g1 g2 ,
induce multilinear maps fk from U/[U, U] × · · · × U/[U, U] (k times) to γk (U)/γk+1 (U), and the Q-span of
{fk (ei1 , . . . , eik )} is (γk (U)/γk+1 (U))(Q). For any k, let φk : γk (U) → γk (U)/γk+1 (U). Then for any k we
have
Q
Q
πQ (φk ( 4k X ∩ γk (U)( p Qp )) ⊇ ∪i1 ,...,ik πQ (q k Zfk (ei1 , . . . , eik )),
and so
P
Q
Q
πQ (φk ( 4k dim U X ∩ γk (U)( p Qp ))) ⊇ πQ (q k i1 ,...,ik Zfk (ei1 , . . . , eik ))
In particular, we have
(36)
Q
Q
|πQ (φk ( 4k dim U X ∩ γk (U)( p Qp )))| ≥ (Q/q k )dim γk (U)−dim γk+1 (U) .
Since φk ’s are regular morphisms defined over Q, there is a positive integer q1′ depending on the embedding
of U and the choice of {ei }) such that for any prime power pm we have
!
Y
Y
Y
X
′
φk γk (U)( Qp ) ∩
U p [q1 Q] ⊆
Zp fk (ei1 , . . . , eik )).
(Q
p
p
p
i1 ,...,ik
Q
Q
P
And so φk induces homomorphism φk : πq1′ Q (γk (U)( p Qp ) ∩ p U p ) → πQ ( i1 ,...,ik Zfk (ei1 , . . . , eik )), and
we have φk ◦ πq1′ Q = πQ ◦ φk . Thus by (36) we have
Y
Y
Q
Q
|πQ (φk ( 4k dim U X ∩ γk (U)( Qp )))| ≥ (Q/q dim U )dim U .
(37)
|πq1′ Q ( 4 dim U3 X)| ≥
p
k
Since U is a Q-unipotent, there is a positive integer
Q
−1
q2′ Qdim U ≤ |πQ ( p U p )| ≤ q2′ Qdim U . And so
q2′
Q
|πQ ( 4 dim U3 X)| ≥
and the claim follows.
(depending on the embedding and U) such that
Q
|πQ ( p U p )|
q2′ 2 q1′ dim U qdim U2
;
Q
′
′
By (35) there is Xδ ⊆
p Up such that πq0 Qδ (Xδ ) = πq0 Qδ (Bδ ), which implies that πQδ (ρ2 (Xδ )) =
πQδ (ρ2 (Bδ )). Hence by Lemma 25 we have
Q
Q
Q
Q
Q
|πQδ ( p∤q U p )|
|πQδ ( p∤q U p )|
′
0
0
(38) |πQδ ( OΩ,ε′ (1) Aδ ) ∩ πQδ ( p∤q0 U p )| ≥
≥ |πQδ ( p|Qδ U p )|1−ΘU (ε ) ,
ΘU (1) ≥
Θdim U (ε′ )
Θdim U (1)qδ
′
for small enough δ (depending on ε ) as limδ→0 Qδ = ∞.
Qδ
SUPER-APPROXIMATION, II:
THE p-ADIC AND BOUNDED POWER OF SQUARE-FREE INTEGERS CASES.
On the other hand, by (34), we have
Therefore we have
(39)
23
Q
Q
πQδ ( p∤q0 Kp [qδ ]) ⊆ C πQδ (Aδ ).
Q
Q
Q
′
|πQδ ( C Aδ ) ∩ πQδ ( p|Qδ Kp )| ≥ |πQδ ( p|Qδ Kp )|1−ΘGs (ε ) ,
for small enough δ (depending on ε′ ) as limδ→0 Qδ = ∞.
Q
b we have
By (38), (39), limδ→0 Qδ = ∞, and the fact that p∤q0 Kp ⋉ Up is of finite index in Γ,
Q
′
(40)
|πQδ ( C Aδ )| ≥ |πQδ (Γ)|1−ΘΩ (ε ) ,
for small enough δ (depending on ε′ ).
Q
On the other hand, since |πQδ ( 3 Aδ )| ≤ |πQδ (Aδ )|1+δ by the Ruzsa inequality (see [Hel05]), we have
Q
(41)
|πQδ ( C Aδ )| ≤ |πQδ (Aδ )|1+(C−2)δ ≤ |πQδ (Γ)|(1−ε0 )(1+(C−2)δ) .
By (40) and (41), for any ε′ and small enough δ (in particular it can approach zero), we have
1 − ΘΩ (ε′ ) ≤ (1 − ε0 )(1 + (C(ε′ ) − 2)δ),
which is a contradiction.
4. Super-approximation: bounded power of square-free integers case.
By Section 3.2, to prove Theorem 1 for CN := {q ∈ Z+ | ∀p ∤ q0 , vp (q) ≤ N } it is enough to prove Theorem 23
(Bounded generation) for CN . In particular, we can assume that U is abelian. Since I believe it is interesting
to know if some of the auxiliary results are true for more general perfect groups, I do not assume U is abelian
till towards the end of this section. As a result some extra lemmas are proved which are not necessary for
the proof of Theorem 23 for CN .3
4.1. The residue maps do not split. The main goal of this section is to show Lemma 29 which roughly
asserts that πp does split under mild conditions.
Lemma 26. Let G = H ⋉ U be a Q-group with a given Q-embedding into (GLn )Q , where H is a connected
semisimple Q-group and U is a unipotent Q-group. Let G (resp. H, U) be the Zariski-closure of G (resp.
H, U) in the Z-group scheme (GLn )Z , and g = Lie(G), h = Lie(H), and u = Lie(U). For large enough p, if
V ( g(fp ) is a G(fp )-invariant proper subspace, then V + [u(fp ), u(fp )] is a proper subspace. Moreover if G is
perfect, then V + z(fp ) is also a proper subspace, where z = Lie(Z(G)) is the Lie ring of the schematic center
of G.
Proof. Let AG := fp [Ad(G(fp ))] ⊆ Endfp (g(fp )) be the fp -span of Ad(G(fp )), and AH := fp [Ad(H(fp ))] be
its subalgebra. Since H is connected semisimple, for large enough p, g(fp ) is a faithful completely reducible
AH -module. Thus AH is a semisimple algebra, and so its Jacobson radical J(AH ) is zero. Let a be the ideal
generated by {u − 1| u ∈ U(fp )}. Then we have the following short exact sequence of AG -modules:
0 → a → AG → AH → 0.
Since AH is semisimple, J(AG ) ⊆ a. On the other hand, since U is a unipotent normal subgroup of G, a is
a nilpotent ideal. So overall we have a = J(AG ).
Suppose to the contrary that there is a proper AG -submodule V of g(fp ) such that V + [u(fp ), u(fp )] = g(fp ).
If we show that J(AG )g(fp ) ⊇ [u(fp ), u(fp )], then we get that V + J(AG )g(fp ) = g(fp ). And so by Nakayama’s
lemma we have V = g(fp ), which is a contradiction.
3If the reader is only interested in the proof of Theorem 23 for C , s/he can skip Lemma 26, Lemma 27, and Lemma 31.
N
24
ALIREZA SALEHI GOLSEFIDY
Since U is unipotent, log and exp define Q-morphisms between U and its Lie algebra. And for large enough
p they induce bijections between U(fp ) and u(fp ). Furthermore for any t ∈ fp , x, y ∈ u(fp ) we have
(42)
Ad(exp(tx))(y) = exp(t ad(x))(y).
Thus for t ∈ fp , x, y ∈ u(fp ) we have
t−1 (Ad(exp(tx))(y) − y) = [x, y] +
(43)
dim
XU i
i=1
t
ad(x)i (y) ∈ J(AG )g(fp ).
i!
Therefore for large enough p by the Vandermonde determinant we have that [x, y] ∈ J(AG )g(fp ) which shows
our claim.
To finish the proof, it is enough to notice that G is perfect if and only if H acts without a non-zero fixed
vector on U/[U, U]. And so when G is perfect, for large enough p we have z(fp ) ⊆ [u, u](fp ) = [u(fp ), u(fp )].
Lemma 27. Let U ⊆ (GLn )Qp be a Qp -subgroup, and U := U(Qp ) ∩ GLn (Zp ). Then if p is large enough
depending only on n, then we have
m
m
πp2 (U ) = hg ∈ πp2 (U )| g p 6= 1i, and U [pm ] = U p := {up | u ∈ U },
for any positive integer m.
Proof. Let exp and log define Qp -morphisms between Lie(U) and U. And so for large enough prime p we
have
(44)
if kxkp , kukp ≤ 1 and xn = 0, (u − I)n = 0.
k exp x − Ikp = kxkp , k log ukp = ku − Ikp
Hence exp and log induce homeomorphisms between u := Lie(U)(Qp ) ∩ gln (Zp ) and U , and moreover for any
positive integer m
m
m
U [pm ] = U p := {up | u ∈ U }.
(45)
We have (45) because of the following observation:
u ∈ U [pm ] ⇔ ku − Ikp ≤ p−m ⇔ u = exp x, x ∈ Lie(U)(Qp ) and kxkp ≤ p−m ,
m
⇔ u = exp(pm x′ ) = (exp x′ )p and x′ ∈ u,
m
⇔ u ∈ Up .
If g p = 1 for some g ∈ πp2 (U ), then for some u ∈ U we have g = πp2 (u) and up ∈ ker(πp2 ). So there is a
m
(unique) u′ ∈ U \ U p such that u′p = up for some m ∈ Z≥2 . Another use of the logarithm map, we have
m−1
u = u′p
. Hence
i
πp2 (U ) = {g p | g ∈ πp2 (U ), g p 6= 1, i = 0, 1, 2}.
Lemma 28. Let H be a simply-connected semisimple Q-group with a given embedding into (GLn )Q . Let
πp
P0 = H(Qp ) ∩ GLn (Zp ) and P1 := ker(P0 −→ GLn (Z/pZ)). Then for large enough p we have
Pi = hU(Qp ) ∩ Pi | U ⊆ H unipotent Qp -subgroupi,
for i = 0, 1.
Proof. For large enough p, we have that H is quasi-split over Qp , and splits over an unramified extension of
Qp , and P0 is a hyper-special parahoric subgroup. Let Qi = hU(Qp ) ∩ Pi | U ⊆ H unipotent Qp -subgroupi.
Then clearly Qi is a normal subgroup of P0 . By the Bruhat-Tits theory, there is a smooth Zp -group scheme
H such that
(1) P0 = H(Zp ),
(2) The generic fiber of H is isomorphic to H,
SUPER-APPROXIMATION, II:
THE p-ADIC AND BOUNDED POWER OF SQUARE-FREE INTEGERS CASES.
25
(3) The special fiber Hp := H ⊗Zp fp is a simply-connected semisimple fp -group since H is simply
connected quasi-split over Qp and splits over an unramified extension of Qp ,
(4) πp (P0 ) = Hp (fp ) is a product of quasi-simple groups.
Since H is quasi-split over Qp , πp (Q0 ) intersects any quasi-simple factor subgroup of Hp (fp ) in a non-central
subgroup. As πp (Q0 ) is a normal subgroup of Hp (fp ), Hp (fp )+ is contained in πp (Q0 ). As Hp is simplyπp i
connected, we get that πp (Q0 ) = πp (P0 ). For any positive integer i, let Pi := ker(P −−→ Hp (Zp /pi Zp )). It
is well-known that Pi /Pi+1 as an Hp (fp )-module is isomorphic to Lie(Hp )(fp ) via the adjoint action. Let’s
identify πpi (Q1 ∩Pi+1 ) with hi ⊆ Lie(Hp )(fp ) . Since Q1 ⊳ P , hi is Hp (fp )-invariant. On the other hand, since
H is quasi-split, hi intersects each one of the Lie algebras of the simple factors of Hp . Since for large enough
p the fp -points of the Lie algebra of the simple factors of Hp are simple Hp (fp )-module, hi = Lie(Hp )(fp ).
Hence by induction we have πpi (Qj ) = πpi (Pj ) for j = 0, 1 and any positive integer i, which finishes the
proof.
Lemma 29. Let G ⊆ (GLn )Q be a perfect Q-group. Assume that the semisimple part of G is simplyconnected; that means G ≃ H ⋉ U where H is a simply-connected semisimple Q-group and U is a unipotent
Q-group. Let Pp = G(Qp ) ∩ GLn (Zp ), G be the closure of G in (GLn )Z , and g = Lie(G). Then
(1) For large enough p, the following is a short exact sequence:
1 → g(fp ) → πp2 (Pp ) → πp (Pp ) → 1.
(2) For large enough p and any proper subgroup V ( g(fp ) which is a normal subgroup of πp2 (Pp ), the
following short exact sequence does not split
1 → g(fp )/V → πp2 (Pp )/V → πp (Pp ) → 1
Proof. For large enough p, Qp := H(Qp ) ∩ GLn (Zp ) is a hyperspecial parahoric subgroup which acts on
Up := U(Qp ) ∩ GLn (Zp ), and Pp = Qp ⋉ Up . Furthermore for large enough p, the assumptions of Lemma 28
hold and so πpi (Pp ) = πpi (hPp ∩ U′ (Qp )| U′ unipotent Qp -subgroupi). Hence by Lemma 27 we have
πp2 (Pp ) = hg ∈ πp2 (Pp ∩ U′ (Qp ))| U′ unipotent Qp -subgroup, g p 6= 1i.
On the other hand, if for any Qp -unipotent subgroup U′ of G we have πp2 (Pp [p] ∩ U′ (Qp )) ⊆ V , then by
Lemma 28 we get that πp2 (Pp [p]) ⊆ V which is a contradiction. Hence for some unipotent Qp -subgroup U′
there is x ∈ πp2 (Pp [p] ∩ U′ (Qp )) \ V . By Lemma 27 there is g ∈ GLn (Zp ) ∩ U′ (Qp ) such that πp2 (g p ) = x.
Since U′ is a Qp -subgroup of G and Pp = GLn (Zp ) ∩ G(Qp ), we have that g ∈ Pp , and the order of πp2 (g)
in πp2 (Pp )/V is p2 . Now suppose to the contrary that the given exact sequence splits; then πp2 (Pp )/V ≃
πp (Pp ) ⋉ (g(fp )/V ), and the latter is a linear group over fp . However any p-element of a linear group over fp
is unipotent, and so it is of order p (for large enough p), which is a contradiction.
4.2. Statistical non-splitting of residue maps. In this section, following [BV12], we prove that any
section of πp is statistically far from being a group homomorphism.
Lemma 30. Let G, g, and Pp be as in Lemma 29. Then there is a positive number δ (depending on G) such
that for large enough p (depending on G and its embedding in (GLn )Q ) the following holds: Let V ( g(fp )
be a proper subgroup which is normal in πp2 (Pp ), and ψ : πp (Pp ) → πp2 (Pp )/V be a section of the quotient
map π : πp2 (Pp )/V → πp (Pp ); that means π ◦ ψ = idπp (Pp ) . Let µ = Pπp (Pp ) be the probability Haar measure
(i.e. the normalized counting measure) on πp (Pp ). Then
(µ × µ)({(x, y) ∈ πp (Pp ) × πp (Pp )| ψ(x)ψ(y) = ψ(xy)}) < p−δ .
Proof. We notice that by Lemma 26, for large enough p, V + z(fp ) is a proper subspace of g(fp ) (which is
normal in πp2 (Pp )). And so without loss of generality we can assume that V contains z(fp ).
26
ALIREZA SALEHI GOLSEFIDY
Let A := {(x, y) ∈ πp (Pp ) × πp (Pp )| ψ(x)ψ(y) = ψ(xy)} and suppose to the contrary that (µ × µ)(A) ≥ p−δ .
We will get a contradiction if δ is small enough. Let ν be the push-forward of µ via ψ. For any g ∈ πp (Pp )
we have
X
X
µ(x)µ(y)
(since supp ν = Imψ)
ν(h1 )ν(h2 ) =
(ν ∗ ν)(ψ(g)) =
h1 h2 =ψ(g)
=
X
x, y ∈ πp (Pp )
ψ(x)ψ(y) = ψ(g)
µ(x)µ(y).
(x, y) ∈ A
xy = g
(since g = π(ψ(g)) = π(ψ(x)ψ(y))
= π(ψ(x))π(ψ(y)) = xy.)
In particular, we have (ν ∗ ν)(supp ν) = (µ × µ)(A). And so by the Cauchy-Schwarz inequality, we get
p
| supp ν|kν ∗ νk2 ≥ (ν ∗ ν)(supp ν) = (µ × µ)(A) ≥ p−δ .
Hence kν ∗ νk2 ≥ p−δ kνk2 . Thus by [Var12, Lemma 15] (which is based on [BG08-a] (see Proposition 22))
there is a symmetric subset S ⊆ πp2 (Pp )/V with the following properties:
(P1)
(P2)
(P3)
|πp (Pp )|1−ΘG (δ) ≤ |S| ≤ |πp (Pp )|1+ΘG (δ) ,
Q
| 3 S| ≤ |πp (Pp )|ΘG (δ) |S|,
mins∈S (e
ν ∗ ν)(s) ≥ (|πp (Pp )|ΘG (δ) |S|)−1 .
Since the push-forward π[ν] of ν via π is the probability Haar measure of πp (Pp ), we have µ = π[e
ν ∗ν]. And so
the third part of properties of S implies that µ(π(S)) ≥ |πp (Pp )|−ΘG (δ) . Therefore |π(S)| ≥ |πp (Pp )|1−ΘG (δ) .
On the other hand, by [SGV12, Corollary 14] and the main theorem of [LS74], we have that πp (Pp ) is
a quasi-random group; that means the minimal degree of its non-trivial irreducible representations is at
least |πp (Pp )|ΘG (1) for large enough
p. Thus by Gowers’s result [Gow08] (we
Q use the formulation in [NP11,
Q
Corollary 1]) we have that π( 3 S) = πp (Pp ) if δ is small enough. So S ′ = 3 S has the following properties:
(1) π(S ′ ) = πp (Pp ),
(2) S ′ is a |πp (Pp )|ΘG (δ) -approximate subgroup by [Tao08, Corollary 3.11],
(3) |S ′ | ≤ |πp (Pp )|1+ΘG (δ) by properties (P1) and (P2) of S.
Q
Next we prove that there is 0 6= x ∈ 3 S ′ ∩ (g(fp )/V ). We proceed by contradiction. For any s1 , s2 ∈ S ′ ,
Q
there is s3 ∈ S ′ such that π(s1 s2 ) = π(s3 ). Hence s1 s2 s−1
∈ 3 S ′ ∩ (g(fp )/V ), and so by the contrary
3
assumption we have s1 s2 = s3 . Thus S ′ is a subgroup which contradicts Lemma 29.
Since V contains z(fp ), π(S) = πp (Pp ) acts on (g(fpP
)/V ) via the adjoint action without non-zero fixed point
for large enough p. Hence by [SGV12, Lemma 30], OG (1) π(S ′ ) · x contains a non-zero subspace of g(fp )/V .
Q
Q
Therefore | OG (1) S ′ ∩ (g(fp )/V )| ≥ p, and we get | OG (1) S ′ | ≥ p|πp (Pp )|. Since S ′ is an approximate
subgroup, we get
Q
p|πp (Pp )| ≤ | OG (1) S ′ | ≤ |πp (Pp )|1+ΘG (δ) .
So p ≤ |πp (Pp )|ΘG (δ) which is a contradiction for small enough δ (independent of p).
4.3. Bounded generation of perfect groups by commutators.
Lemma 31. Let P be a finite p-group. Suppose |P | = pn and |P/[P, P ]| = pm . Then
Q
n−m {[g1 , g2 ]| g1 , g2 ∈ P } = [P, P ],
where [g1 , g2 ] := g1 g2 g1−1 g2−1 .
Proof. We proceed by induction on the nilpotency class of P . If P is abelian, there is nothing to prove. Now
suppose the nilpotency class of P is c; that means its c + 1-th lower central series γc+1 (P ) is trivial and γc (P )
SUPER-APPROXIMATION, II:
THE p-ADIC AND BOUNDED POWER OF SQUARE-FREE INTEGERS CASES.
27
is not trivial. It is well-known
that γc (P ) is an fp -vector space which is spanned by the long commutators.
Q
So γc (P ) is contained in dimfp γc (P ) {[g1 , g2 ]| g1 , g2 ∈ P }. By the induction hypothesis, we have
Q
n−m−dimfp γc (P ) {[ḡ1 , ḡ2 ]| ḡ1 , ḡ2 ∈ P/γc (P )} = P/γc (P ).
Q
And so n−m {[g1 , g2 ]| g1 , g2 ∈ P } = P .
Lemma 32. Let G be a Zariski-connected perfect Q-subgroup of (GLn )Q , and let G be its Zariski-closure
in (GLn )Z . Suppose that G is simply-connected; that means its semisimple part is simply-connected. Then
there is a positive integer C depending on G such that for large enough prime p we have
Q
C {[g1 , g2 ]| g1 , g2 ∈ Gp (fp )} = Gp (fp ),
where Gp = G ×Spec Z Spec fp .
Proof. As it was mentioned earlier, for large enough p, Gp ≃ Hp ⋉ Up is a perfect fp -group, where Hp is a
simply connected semisimple fp -group and Up is a unipotent fp -group. Moreover Gp (fp ) = Hp (fp ) ⋉ Up (fp ),
and Vp := Up (fp )/[Up (fp ), Up (fp )] is a completely reducible Hp (fp )-module with no non-zero fixed vector. Let
w(x1 , x2 ) = [x1 , x2 ] and, for any group K, let w(K) := {w(k1 , k2 )| k1 , k2 Q
∈ K}. Then by [Sha09, Theorem
1.1] (see an alternative approach in [NP11, Theorem 3]) we have that 3 w(Hp (fp )) = Hp (fp ) for large
enough p.
Let Hp = Hp (fp ) and
Q Gp = Hp ⋉ Vp . Let ρ : Hp → GL(Vp ) be the homomorphism induced from the action
of Hp on Vp . Then 7 w(Gp ) contains
{ρ(h)((ρ(h′ ) − I)(v))| h, h′ ∈ Hp , v ∈ Vp }.
Q
And so 7 dimf Vp w(Gp ) contains the fp [Hp ]-submodule Mp of Vp generated by {(ρ(h) − I)(v)| h ∈ Hp , v ∈
p
Vp }. Therefore Hp acts trivially on Vp /Mp . Since Vp is a completely reducible Hp -module with no non-zero
fixed vectors, Mp = Vp . Thus
Q
O(dim G) w(Gp ) = Gp .
And so by Lemma 31 applied for P := Up (fp ) the claim follows.
Lemma 33. Let G and G be as in Lemma 32, and N be a positive integer. Suppose the unipotent radical of
G is abelian. Then there is a positive integer C depending on G and N such that for large enough p we have
Q
N
N
C {[g1 , g2 ]| gi ∈ G(Z/p Z)} = G(Z/p Z).
Proof. By Lemma 32 we get the case of N = 1. It is well-known that πp induces the following short exact
sequence
πp
1 → g(fp ) → G(Z/p2 Z) −→ G(fp ) → 1,
where g = Lie(G), and the conjugation action of G(Z/p2 Z) on g(fp ) factors through the adjoint action of
G(fp ). Let u be the Lie algebra of the Zariski-closure of the unipotent radical of G in G. For large enough p,
G(fp ) and g(fp ) are perfect, and g(fp ) is a completely reducible G(fp )-module without any non-zero invariant
vectors (as U is abelian). So by a similar argument as in the proof of Lemma 32, we have
Q
2
OG (1) w(G(Z/p Z)) ⊇ g(fp ),
and so by Lemma 32, we have that
Q
OG (1)
w(G(Z/p2 Z)) = G(Z/p2 Z).
Now using the facts that G(Zp )[pi ]/G(Zp )[pi+1 ] ≃ g(fp ), the group commutators are mapped to the Lie
algebra commutators, and g(fp ) is perfect, we get the desired result. (See [SG, Lemma 34] for the relations
between the group commutators and the Lie algebra commutators.)
28
ALIREZA SALEHI GOLSEFIDY
4.4. Proof of Theorem 1 for bounded powers of square-free integers. Now we go to the proof of
Theorem 1 for CN := {q ∈ Z+ | gcd(q, q0 ) = 1, ∀p ∈ Vf (Q), vp (q0 ) ≤ N }. Having the above lemmas the rest
of the proof is a modification of the argument in [BV12, Section 3].
By Section 3.2, it is enough to prove Theorem 23 for CN . First we point out that it is enough to prove
≥p0
:= {q ∈ Z+ | gcd(q, q0 ) = 1, ∀p ∈ Vf (Q), vp (q0 ) ≤ N, ∀p ≤ p0 , vp (q0 ) = 0}. This
Theorem 23 for CN
Lemma helps us to avoid all the small primes p where πp (Γ) does not behave nicely; in particular we will be
able to assume Lemmas 26-33 to hold for all the prime divisors of Q.
≥p0
Lemma 34. Let p0 be a prime number. Then a bounded generation in the sense of Theorem 23 for CN
implies Theorem 23 for CN where the implied constants depend also on p0 .
Proof. Fix a positive number δ0 := δ0 (ε, Ω, p0 ) and a positive integer C0 := C0 (ε, Ω, p0 ) for which Theorem 23
≥p0
.
holds for CN
Q
p0
, we have [Γ : Γ[q0 ]] ≪G,p0 ,N 1. Suppose Q is
For Q ∈ CN , let q0 := p|Q,p≤p0 pvp (Q) . Since q0 < pN
0
≥p0
can be applied
large enough (depending only on ε, Ω, p0 , N ) such that the bounded generation for CN
≥p0
for Q/q0 ∈ CN . Moreover we assume Q is large enough (again depending only on ε, Ω, p0 , N ) such that
Q−δ0 /2 /[Γ, Γ[q0 ]] > (Q/q0 )−δ0 . Now suppose δ is a positive number which is at most δ0 /4, A is a symmetric
(l)
set, and l is a positive integer which is more than (1/δ) log Q such that PΩ (A) > Q−δ . Then
(2l)
PΩ (A · A ∩ Γ[q0 ]) ≥ Q−2δ /[Γ : Γ[q0 ]] > Q−δ0 /2 /[Γ : Γ[q0 ]] > (Q/q0 )−δ0 .
≥p0
we have that
Hence PQ/q0 (δ0 , A · A ∩ Γ[q0 ], l) holds. Therefore by the bounded generation for CN
Q
(46)
πQ/q0 (Γ[q]) ⊆ πQ/q0 ( C0 (A · A ∩ Γ[q0 ])),
for some q|Q/q0 and q ≤ (Q/q0 )ε . Since Q and Q/q0 are coprime, πQ (Γ) can be diagonally embedded into
πq0 (Γ)⊕ πQ/q0 (Γ), and by enlarging p0 , if needed, we can and will assume that πQ (Γ[qq0 ]) gets identified with
{1}⊕πQ/q0 (Γ[q]) under the diagonal embedding of πQ (Γ) into πq0 (Γ)⊕πQ/q0 (Γ). Notice that πQ (A·A∩Γ[q0 ])
gets mapped to a subset of {1} ⊕ πQ/q0 (Γ[q]) and so by (46) we get that
Q
πQ (Γ[qq0 ]) ⊆ 2C0 πQ (A).
And we have qq0 |Q and qq0 ≤ Qε q01−ε ≤ Q2ε for large enough Q (again depending on ε, Ω, p0 , N ).
Next we show a reduction which works for any family of positive integers C ′ :
Lemma 35. To prove Theorem 23 for Q in a family of positive integers C ′ , it is enough to prove the bounded
generation claim for a subset A for which PQ (δ, A, l) and PQ (δ, ρ1 (A), l) hold (where ρ1 : G → Gs is the
quotient map). That means for any 0 < ε ≪Ω 1 there are 0 < δ ≪Ω,ε 1 and positive integer C ≫Ω,ε 1 for
which the following holds:
Q
(Restricted BG): PQ (δ, A, l) ∧ PQ (δ, ρ1 (A), l) implies πQ (Γ[q]) ⊆ C πQ (A) for some q|Q such that
O (1)
q ≤ Qε if Q ∈ C ′ and Qε G ≫Ω 1.
(l)
Proof. Suppose δ is a positive number and PΩ (A) > Q−δ for a symmetric set A and a positive integer
(l)
l > (1/δ) log Q. Since ker(ρ1 ) ∩ Γ = {1}, we have Pρ1 (Ω) (ρ1 (A)) > Q−δ . Then by a similar argument as in
(33) (based on the Kesten bound (see Inequality (14))) we have
(47)
|πQ (A)| > Qη0 and |πQ (ρ1 (A))| > Qη0 ,
for small enough δ (depending only on Ω) and some positive number η0 := η0 (Ω) which depends only on Ω.
SUPER-APPROXIMATION, II:
THE p-ADIC AND BOUNDED POWER OF SQUARE-FREE INTEGERS CASES.
29
Fix a positive number δ0 := δ0 (ε, Ω) and a positive integer C0 := C0 (ε, Ω) for which the Restricted BG
holds. And let
Q
Q
(48)
IA := {i ∈ Z+ | | 3i+1 πQ (A)| ≤ | 3i πQ (A)|1+δ0 } and
Q
Q
Iρ1 (A) := {i ∈ Z+ | | 3i+1 πQ (ρ1 (A))| ≤ | 3i πQ (ρ1 (A))|1+δ0 }.
(49)
Notice that |πQ (Γ)| ≤ QN0 (G) and |πQ (ρ1 (Γ))| ≤ QN0 (G) for some positive integer N0 which depends only on
the dimension of G. Now let N1 := N1 (δ0 ) be the smallest positive integer such that
(1 + δ0 )N1 η0 > N0 .
(50)
So by Inequalities (47) and (50), the complements of IA and Iρ1 (A) in Z+ have at most N1 elements. Hence
there is a positive integer C1 ≤ 32N1 +1 ≪Ω,ε 1 such that
Q
Q
Q
Q
|πQ ( C1 ρ1 (A)) · πQ ( C1 ρ1 (A)) · πQ ( C1 ρ1 (A))| ≤ |πQ ( C1 ρ1 (A))|1+δ0 and
Q
Q
Q
Q
|πQ ( C1 A) · πQ ( C1 A) · πQ ( C1 A)| ≤ |πQ ( C1 A)|1+δ0 .
Q
Q
Therefore, if δ < δ0 , then PQ (δ0 , C1 A, l) and PQ (δ0 , C1 (ρ1 (A)), l) hold. Thus, by the Restricted BG
and the fact that C1 C0 ≪Ω,ε 1, we are done.
(Going back to the proof of Theorem 23 for CN :) Suppose p0 is a prime number such that p0 ≪Ω 1 and all the
claims of Lemmas 26-33 hold for all the primes p ≥ p0 (later we will assume p0 is large enough depending on
Ω to gain additional properties for πp (Γ) where p is a prime divisor of Q.). So by Lemma 34 we can assume
≥p0
), without loss of generality we
that all the prime factors of Q are at least p0 . By Lemma 35 (for C ′ := CN
can and will assume that PQ (δ, A, l) and PQ (δ, ρ1 (A), l) hold.
Q
N
Let Qs := p|Q p (, and so Q|QN
s ). By [SGV12, Theorem 1], [SG, Lemma 31], and l > δ log Qs we have
that
1
|πQs (A)|
(l)
≤
PπQ (Ω) (πQs (A)) −
s
|πQs (Γ)|
|πQs (Γ)|
for small enough δ. And so we have
(l)
Qs −N δ ≤ Q−δ ≤ PπQ
−Θ
s (Ω)
(πQs (A)) ≤
|πQs (A)| + 1
,
|πQs (Γ)|
(δ)
which implies that |πQs (A)| ≥ |πQs (Γ)|Qs Ω,N . Since we assumed that the prime factors of Q are large
Q
Q
Q
enough, we have that πQs (Γ) ≃ p|Q πp (Γ). Writing Qs = lj=1 pj , to the product set li=1 πpi (Γ) one can
associate a rooted tree with l levels. The root is considered to be the zero level. The vertices of the j-th
Q
level are elements of ji=1 πpi (Γ), and each vertex is connected to its projection in the previous level. By a
regularity argument similar to the proof of [SG, Lemma 12] (see [BGS10, Lemma 5.2] or [Var12, Page 26]),
one can show that there are a set A′ ⊆ A and a sequence of integers {ki }li=1 such that for any g ∈ A′ and
for any 1 ≤ i ≤ l we have
|{x ∈ πpi (Γ)| ∃h ∈ A′ : πp1 ···pi−1 (h) = πp1 ···pi−1 (g) and πpi (h) = x}| = ki ,
and
|A′ | =
l
Y
i=1
ki ≥ (
l
Y
2 log |πpi (Γ)|)−1 |A|.
i=1
By [SGV12, Corollary 14] and the main theorem of [LS74], πp (Γ) is a quasi-random group for large enough
p; that means the dimension of any non-trivial irreducible representation of πp (Γ) is at least |πp (Γ)|ΘG (1)
(for p ≫Ω 1) (see also Proposition 18). So by a result of Gowers [Gow08] (see [NP11, Corollary 1]) there is
a positive number c0 depending on G such that for large enough p the following holds: if A, B, C ⊆ πp (Γ)
30
ALIREZA SALEHI GOLSEFIDY
and |A|, |B|, |C| ≥ |πp (Γ)|1−c0 , then A · B · C = G. We assume that the prime factors of Q are large enough
so that 2 log |πpi (Γ)| < |πpi (Γ)|c0 ε . Hence
l
Y
ki ≥ |πQs (Γ)|1−ΘG (δ)−c0 ε .
i=1
Let I := {i| ki < |πpi (Γ)|
1−c0
}; then we have
|πQs (Γ)|/
Y
|πpi (Γ)|c0 ≥
ki ≥ |πQs (Γ)|1−ΘG (δ)−c0 ε ,
i=1
i∈I
and so
l
Y
Y
G (δ)+2ε
pi ≤ Q Θ
≤ Q4ε
s
s ,
i∈I
for small enough δ (depending on Ω, N and ε). Now by a similar argument as in [Var12, Page 26] (using
the mentioned result of Gowers) one can show that
πQi6∈I pi (A · A · A) = πQi6∈I pi (Γ).
(51)
Let Q′s =
Q
pi and q ′ = Qs /Q′s .
By (51) there is a function f1 : πQ′s (Γ) → πq′N (Γ) such that
Q
1−Θ
(ε)
{(g, f1 (g))| g ∈ πQ′s (Γ)} is contained in πQ′s q′N ( 3 A). Thus there is a fiber of f1 with at least Qs G,N many elements. Hence, as before by a regularization argument and another application of the mentioned Gowers’s result (for ε ≪G,N 1), there is q ′′ |Q′s and a function f2 : πQ′s /q′′ (Γ) → πq′′ (Γ) such that q ′′ ≤ QΘG,N (ε)
and
{(g, f2 (g), 1) ∈ πQ′s /q′′ (Γ) × πq′′ (Γ) × πq′N (Γ)|g ∈ πQ′s /q′′ (Γ)}
Q
is contained in πQ′s q′N ( OG (1) A). Hence, for any g1 , g2 ∈ πQ′s /q′′ (Γ), we have
Q
([g1 , g2 ], 1, 1) = [(g1 , f2 (g1 ), 1), (g2 , 1, f1 (g2 , 1))] ∈ πQ′s q′N ( OG (1) A).
i6∈I
Therefore by Lemma 32
Q
πQ′s q′N (Γ[q ′N ]) ⊆ πQ′s q′N ( QG (1) A).
Q
Let A′ be a subset of QG (1) A such that πQ′s q′N (Γ[q ′N ]) = πQ′s q′N (A′ ). Now we will go to the second level;
Q
that means we will show that there are a positive integer C and q̄|Q′s such that q̄ ≤ QΘG (ε) and πQ′s 2 ( C A′ )
contains πQ′s 2 (Γ[q̄ 2 ]).
Lemma 36. In the above setting, assume U is abelian. Then for any ε > 0 there are δ > 0 and a positive
integer C such that the following holds.
Assume A ⊆ Γ and a square-free integer Qs have the following properties:
(1) Qεs ≫Ω 1.
(2) The prime factors of Qs are sufficiently
large (depending on Ω); in particular the assertions of
Q
Lemma 29 hold, and πQ2s (Γ) ≃ p|Qs πp2 (Γ).
(3) πQs (A) = πQs (Γ).
(4) Q
For any prime factor p of Qs , let Vp ( g(fp ) be a πp (Γ)-invariant under the adjoint action. And
p|Qs Vp ⊆ πQ2s (A).
Then there are Q′s and {Mp }p|Q′s such that
(1) Q′s |Qs and Q1−ε
≤ Q′s .
s
(2) For
), and Mp is πp (Γ)-invariant under the adjoint action.
Q any p, we
Qhave Vp ( Mp ⊆ g(fpQ
(3) p|Q′s Mp ⊕ p|(Qs /Q′s ) {0} ⊆ πQs 2 ( C A).
SUPER-APPROXIMATION, II:
THE p-ADIC AND BOUNDED POWER OF SQUARE-FREE INTEGERS CASES.
31
Proof. Since Vp ⊆ g(fp ), the map πp : πp2 (Γ) → πp (Γ) factors through πp2 (Γ)/Vp . Let us abuse the notation
and still denote the induce homomorphism by πp : πp2 (Γ)/Vp → πp (Γ). Let ψ : πQs (Γ) → A ⊆ Γ be a section
of πQs , and ψp : πQs (Γ) → πp2 (Γ)/Vp be ψp (g) = πp2 (ψ(g))Vp . As in the proof of [BV12, Proposition 3], let
us consider the following expectation with respect to the probability Haar measure:
X
X
log p =
E
(52)
log p · (Pπp (Γ) × Pπp (Γ) )({(x, y)| ψp (xy) = ψp (x)ψp (y)})
p|Qs
p|Qs , ψp (xy)=ψp (x)ψp (y)
≤
X
p−ΘG (1) log p ≤ log(Qεs ),
p|Qs
where the first inequality is given by Lemma 30 and the second inequality holds for Qεs ≫G 1. Inequality (52)
implies that
Y
E log
p|Qs ψp (x)ψp (y)ψp
(xy)−1 6=1
p ≥ log(Q1−ε
).
s
′
1−ε
and for any prime p|Q′s we have zp :=
So there are x, y ∈ πQs (Γ) and Q′s |Qs such that
Q Qs ≥ Qs
−1
ψp (x)ψp (y)ψp (xy)
6= 1. Hence there is a ∈ 3 A such that for any p|Q′s we have that πp2 (a) belongs
to g(fp ) \ Vp where g(fp ) is identified with πp2 (Γ[p]). By our assumptions on G, πp (Γ) is generated by its pelements and g(fp ) is a completely reducible πp (Γ)-module with no non-zero invariant vectorP
(the latter holds
as G is perfect and its unipotent radical is abelian). Therefore by [SGV12, Corollary 31] OG (1) πp (Γ) · zp
is the πp (Γ)-subspace Mp /Vp of g(fp )/Vp that is generated by zp .
Q
In the group πQ2s (Γ)/ p|Qs Vp , for any γ ∈ Γ, we have
Y
Y
g(fp )/Vp .
Vp = (πp (γ) · zp )p|Qs ∈
(53)
πQ2s (γ)πQ2s (a)πQ2s (γ)−1
p|Qs
p|Qs
Q
Q
Q
′2
(as a subset of
V
A)
A,
by
(53)
we
have
that
π
(
Qs
OG (1)
p|Qs p
3
Since πQs (A) = πQs (Γ) and a ∈
Q
πQ2s (Γ)/ p|Qs Vp ) contains
Y
YP
P
Mp /Vp .
OG (1) πp (Γ) · zp =
OG (1) πQ′s (A) · (zp )p|Q′s =
(Here πQs 2 (Γ)/
Q
p|Qs
p|Q′s
p|Q′s
Q
Vp is viewed as (πQ′s 2 (Γ)/ p|Q′ Vp ) ⊕ (π(Qs /Q′s )2 (Γ)/
s
Y
Q
Mp ⊆ πQ2s ( OG (1) A);
Q
p|Qs /Q′s
Vp ).) And so
p|Q′s
and the claim follows.
(Going back to the proof of Theorem 23 for CN :) By Equation (51) and a repeated use of Lemma 36 (for
Θ (ε)
OG (1)-many times), we get that there is q|Qs such that q ≤ Qs G (if Qεs ≫G 1) and
Y
Y
Q
πp2 (Γ) × {1}.
(54)
πQs 2 ( OG (1) A) ⊇ πQs 2 (Γ[q 2 ]) =
p|q
p|(Qs /q)
Let us recall that for large enough p (depending only on Ω) we have that, for any positive integer i, there
i+1
is a πp (Γ)-module isomorphism Ψppi : Γ[pi ]/Γ[pi+1 ] → g(fp ); and for any gi ∈ Γ[pi ] and gj ∈ Γ[pj ] we have
[gi , gj ] := gi gj gi−1 gj−1 ∈ Γ[pi+j ] and
(55)
i+j+1
i+1
j+1
Ψppi+j ([gi , gj ]Γ[pi+j+1 ]) = [Ψppi (gi Γ[pi+1 ]), Ψppj (gj Γ[pj+1 ])],
32
ALIREZA SALEHI GOLSEFIDY
where the right hand side is the Lie bracket in g(fp ). (These maps are called the truncated or finite logarithmic
maps. We refer the interested reader to [SG, Section 2.9] for a general and quick treatment of these maps
and their basic properties.) Equation (55) defines a graded Lie algebra structure on
Γ[p]/Γ[p2 ] ⊕ Γ[p2 ]/Γ[p3 ] ⊕ · · · ⊕ Γ[pN −1 ]/Γ[pN ]
and shows that it is isomorphic to the graded Lie algebra g(fp ) ⊗ (tfp [t]/tN fp [t]) where fp [t] is the ring of
polynomials in a single variable t, [xi ⊗ t̄i , xj ⊗ t̄j ] := [xi , xj ] ⊗ t̄i+j for any xi , xj ∈ g(fp ), and t̄ := t + tN fp [t].
Since, for large enough p (depending only on Ω), g(fp ) is a perfect Lie algebra, g(fp ) ⊗ (tfp [t]/tN fp [t]) is
generated by the degree one elements, that means g(fp ) ⊗ t̄. So by (54) and (55) we get that
Y
Q
πpN (Γ).
(56)
π(Qs /q)N ( OG,N (1) A) = π(Qs /q)N (Γ) =
p|(Qs /q)
Hence there is a function f1 : π(Qs /q)N (Γ) → πqN (Γ[q]) such that G(f1 ) := {(g, f1 (g))|g ∈ π(Qs /q)N (Γ)} is
Q
−1
and for any subset X of a group w(X) :=
contained in πQN
( OG,N (1) A). Let w(x1 , x2 ) := x1 x2 x−1
1 x2
s
N
N
{w(x1 , x2 )| x1 , x2 ∈ X}. Then w(G(f1 )) ⊆ π(Qs /q) (Γ) × πq (Γ[q 2 ]). And so by Lemma 33 we have that
Q
there is a function f2 : π(Qs /q)N (Γ) → πqN (Γ[q 2 ]) such that its graph G(f2 ) is contained in πQN
( OG,N (1) A).
s
Q
Repeating this argument log N times we get that πQN
(Γ[q N ]) is contained in πQN
( OG,N (1) A), which finishes
s
s
the proof of Theorem 23 for CN .
5. Super-approximation: the p-adic case.
To prove Theorem 1 for C := {pm | p ∈ Vf (Q), p ∤ q0 , m ∈ Z+ }, by Section 3.2, it is enough to Prove Theorem
23 for C.
In this section, we work in the setting of Section 2.6. In addition, we assume that U is abelian. So we
assume that V is a vector Q-group, Gs is a connected, simply-connected, semisimple Q-group, there is a
Q-homomorphism G → GL(V) with no non-zero fixed vector, and G = Gs ⋉ V is the Zariski-closure of
Γ = hΩi. We also fix a Q-embedding G ⊆ (GLN0 )Q .
By Lemma 35 (for C ′ := C), we can and will assume that PQ (δ, A, l) and PQ (δ, ρ1 (A), l) hold where ρ1 :
G → Gs is the quotient map.
It is worth pointing out that Theorem 23 for semisimple groups and powers of primes is proved in [SG,
Theorem 36].
5.1. Escape from proper subgroups. In this section, we explore the pro-p structure of an open subgroup
of the p-adic closure Γp of Γ. The main goal is to escape proper subgroups of πQ (Γ) where Q = pn is a power
of a prime p.
Proposition 37. In the above setting, there is a positive number δ (depending on Ω) such that for any
n ≫Ω 1 and any proper subgroup H of πpn (Γ) we have
(l)
Pπpn (Ω) (H) ≤ [πpn (Γ) : H]−δ ,
for l ≥
n
δ
log p.
We start with a (well-known) lemma which gives us the Frattini subgroup of the congruence subgroups of
G(Zp ), where (as before) G is the closure of G in (GLN0 )ZS . This kind of result for semisimple groups and
large p goes back to Weisfeiler [Wei84].
e is a closed
Lemma 38. Let G be as above, and G := G(Zp ). Suppose either p ≫G 1 or m0 ≫G 1. If H
m0
m0 +1
m0
m0
e
e
] = G[p ], then H = G[p ]. And so the Frattini subgroup Φ(G[pm0 ]) =
subgroup of G[p ] and HG[p
m0 +1
G[p
].
SUPER-APPROXIMATION, II:
THE p-ADIC AND BOUNDED POWER OF SQUARE-FREE INTEGERS CASES.
33
m0 +k
e
Proof. By induction on k we prove that HG[p
] = G[pm0 ]. The base of the induction is given by the
e ∩G[pm0 +k ])G[pm0 +k+1 ] = G[pm0 +k ]. Let
assumption. To get the induction step, it is enough to prove that (H
pm0 +1
g := Lie(G)(Zp ). By [SG, Lemma 34], we have Ψpm0 : G[pm0 ]/G[pm0 +1 ] → g/pg is an isomorphism if either
m0 +1
p ≫G 1 or m0 ≫G 1 (see [SG, Section 2.9] for the definition and properties of finite logarithmic map Ψppm0 ).
e such that πpm0 +1 (h) = πpm0 +1 (I + pm0 x). On the
And so by the assumption, for any x ∈ g there is h ∈ H
m0 +k+1
k
other hand, if m0 ≥ 2, then πp (((I +pm0 x′ )p −I)/pm0 +1 ) = πp (x′ ). And so Ψppm0 +k (πpm0 +k+1 (hp )) = πp (x).
m0 +k+1
Since, by [SG, Lemma 34], Ψppm0 +k : G[pm0 +k ]/G[pm0 +k+1 ] → g/pg is a bijection, we get the induction
e is a closed subgroup, we have that H
e = G[pm0 ].
step. Since H
The following lemma which is a module theoretic version of [SG05, Lemma 3.5] is proved next. It will be
needed in the following sections, too.
Lemma 39. Let R ⊆ Mn0 (Zp ) be a Zp -subalgebra (not necessarily with the identity element). Suppose Qnp 0
is a simple Qp [R]-module, where Qp [R] is the Qp -span of R. Then
sup
n
n
v∈Zp 0 \pZp 0
[Znp 0 : Rv] < ∞.
Proof. Suppose to the contrary that there is a sequence of unit vectors vi such that [Znp 0 : Rvi ] goes to
infinity as i goes to infinity. Since the set of unit vectors is a compact set, by passing to a subsequence we
can assume that limi→∞ vi = v where v is a unit vector. Since Qnp 0 is a simple Qp [R]-module, the Qp -span
of Rv is Qnp 0 . And so Rv is of finite-index in Znp 0 . Therefore Rv contains pk0 Znp 0 for some positive integer k0 .
If i is large enough, v − vi ∈ pk0 +1 Zp . So Rvi + pk0 +1 Znp 0 = Rv + pk0 +1 Znp 0 ⊇ pk0 Znp 0 . Since Rvi is complete
and the Frattini subgroup of pk0 Znp 0 is pk0 +1 Znp 0 , Rvi ⊇ pk0 Znp 0 which is a contradiction.
The following lemma roughly shows that an almost invariant subspace of Lie(Gs )(Qp ) is close to an invariant
subspace.
Lemma 40. Let Gs ⊆ (GLN0 )Q be (as before) a semisimple group. Let G := Gs (Qp ) ∩ GLN0 (Zp )[pk0 ] for a
fixed positive integer k0 . Suppose m is large enough depending on the embedding of Gs in (GLN0 )Q and k0 .
Then the following holds:
Suppose W is a subspace of Lie(Gs )(Qp ) such that πpm (W ∩ glN0 (Zp )) is invariant under G. Then there is
a normal closed subgroup GW of Gs and a positive integer C which depends only on the embedding of Gs in
(GLN0 )Q and k0 such that
p2C πpm−C (gW ) ⊆ pC πpm−C (W ∩ glN0 (Zp )) ⊆ πpm−C (gW ),
(57)
where gW := Lie(GW )(Qp ) ∩ glN0 (Zp ).
Proof. Since Gs is semisimple, it is the almost product of Qp -simple factors Gi . And so Lie(G) = ⊕i Lie(Gi )(Qp ).
Let g := Lie(G)(Qp ) ∩ glN0 (Zp ) and gi := Lie(Gi )(Qp ) ∩ glN0 (Zp ). Then there is a positive number c1 which
depends only on the embedding of Gs in (GLN0 )Q such that
M
M
p−c1 gi ;
gi ⊆ g ⊆
(58)
i
i
in particular, for any x ∈ Lie(G)(Qp ), we have
kxkp ≤ max kxi kp ≤ pc1 kxkp ,
(59)
where x =
P
i
i
xi and xi ∈ Lie(Gi )(Qp ).
34
ALIREZA SALEHI GOLSEFIDY
P
Let W := W ∩ glN0 (Zp ). For x ∈ W, let xi ∈ Lie(Gi )(Qp ) be such that x =
i xi . For any gj ∈
Gj (Qp ) ∩ GLN0 (Zp )[pk0 ], Ad(gj ) changes only the j-th component. So, using the assumption that πpm (W)
is invariant under Ad(gj ), we get that
Ad(gj )xj − xj ∈ W + pm glN0 (Zp ).
(60)
Hence aj xj ⊆ W + pm glN0 (Zp ) where aj is the Zp -span of {Ad(gj ) − I| gj ∈ Gj (Qp ) ∩ GLN0 (Zp )[pk0 ]} (here
we restrict to the action on Lie(Gj ), and consider aj as a subset of EndQp (Lie Gj (Qp ))). Notice that aj is
an ideal of the Zp -span Rj of Ad(Gj (Qp ) ∩ GLN0 (Zp )[pk0 ]). And Lie(Gj )(Qp ) is a simple Qp [Rj ]-module,
where Qp [Rj ] is the Qp -span of Rj . Hence Qp [Rj ] is a simple Qp -algebra as it has a simple faithful module.
Therefore the Qp -span Qp [aj ] of aj is Rj , and Lie(Gj )(Qp ) is a simple Qp [aj ]-module. Hence, by Lemma 38,
there is a positive integer c2 which depends only on k0 and the embedding of Gs in (GLN0 )Q such that
M
M
(61)
pc2 kxj k−1 gj ⊆
aj xj ⊆ W + pm glN0 (Zp ).
j
j
P
Let IW := {j| ∃x ∈ W, kxkp = 1 and kxj kp ≥ 1, where x =
i xi , and xi ∈ Lie(Gi )(Qp )}, GW be the
product of the simple factors Gi for i ∈ IW ; and so gW := (⊕i∈IW Lie(Gi )(Qp )) ∩ glN0 (Zp ).
We can and will assume that c2 ≥ c1 , and we claim that (57) holds for C := c1 + c2 and gW . We start with
the first inclusion in (57).
For any j ∈ IW , there is x ∈ W such that its j-th simple component xj has norm at least 1. Therefore by
(61) we have pc2 gj ⊆ W + pm glN0 (Zp ). And so by (58) we have
M
(62)
pc1 +c2 gW ⊆
pc2 gj ⊆ W + pm glN0 (Zp ).
j∈IW
To show the second inclusion in (57), we proceed by contradiction. So assume to the contrary
P that there is
x ∈ W such that pC πpm−C (x) 6∈ πpm−C (gW ). By (58) there are xj ∈ gj such that pc1 x = j xj . By (62),
there are x′ ∈ W and e ∈ g such that
X
xi = x′ + pm e.
(63)
pc 2
i∈IW
Hence
x′′ := pc1 +c2 x − x′ =
(64)
X
pc2 xj + pm e =
(pc2 xj + pm ej ) +
j6∈IW
j6∈IW
P
X
X
pm e j ,
j∈IW
where e = i ei and ei ∈ Lie(Gi )(Qp ), is contained in W. Since e ∈ g, we have kekp ≤ 1. Therefore, by
(59), we have
(65)
kpm ei kp ≤ p−m+c1 .
P
P
On the other hand, since pC πpm−C (x) 6∈ πpm−C (gW ) and pC x = pc2 i∈IW xi + pc2 i6∈IW xi ∈ gW +
P
P
pc2 i6∈IW xi , we have kpc2 j6∈IW xj kp > p−m+C , which implies that there is j0 6∈ IW such that
kpc2 xj0 kp > p−m+C .
(66)
So by the inequalities given in (65) and (66) we have
kpc2 xj0 + pm ej0 k = max{kpc2 xj0 kp , kpm ej0 kp } > p−m+C .
(67)
Therefore
kx′′ kp ≥ p−c1 kpc2 xj0 + pm ej0 k
>p
−c1 −m+C
p
m
≥ kp ei kp .
′′
=p
−m+c2
(by (64) and (59))
≥p
−m+c1
(by (67))
(by (65))
So after
P normalizing x , we get a vector x ∈ W such that kxkp = 1 and, for any i ∈ IW , kxi kp < 1, where
x = i xi and xi ∈ Lie(Gi )(Qp ). This is contrary to the definition of IW .
SUPER-APPROXIMATION, II:
THE p-ADIC AND BOUNDED POWER OF SQUARE-FREE INTEGERS CASES.
35
Lemma 41. As before let G = Gs ⋉ V ⊆ (GLN0 )Q , where Gs is a semisimple connected simply-connected
Q-group and V is a Q-vector group. Let G := G(Qp ) ∩ GLN0 (Zp )[pk0 ] for a fixed positive integer k0 . Suppose
m is large enough depending on the embedding of G in (GLN0 )Q and k0 . Then the following holds:
Suppose W is a subspace of Lie(G)(Qp ) such that the projection of W onto Lie(Gs )(Qp ) is onto and πpm (W ∩
glN0 (Zp )) is invariant under G. Then W = Lie(G)(Qp ).
Proof. Without loss of generality we can and will assume that V is non-zero. Suppose Gs acts on V via
ρ : Gs → GL(V); that means for any g ∈ Gs (R) and v ∈ V(R) we have (g, 0)(1, v)(g, 0)−1 := (1, ρ(g)(v))
in G(R) for any Q-algebra R. To get a clear computation of the adjoint action, we use the dual numbers
Qp [ε] := Qp ⊕ Qp ε where ε2 = 0 (this lemma is the only place where ε is not a real number, and it is an
element of the dual numbers.). Recall that we can identify Lie(G)(Qp ) with
πε
Gs (Qp ) ⋉ Vp (Qp )
ker Gs (Qp [ε]) ⋉ V(Qp [ε]) −→
where πε is the group homomorphism induced by the ring homomorphism πε : Qp [ε] → Qp . Under this
identification the adjoint action is given by conjugation; that means, for x ∈ Lie(G)(Qp ), we have 1 + εx ∈
G(Qp [ε]); and for g ∈ G(Qp ), we have 1 + ε Ad(g)(x) = g(1 + εx)g −1 .
The identity element of G = Gs ⋉V is (1, 0). And so (x, w) ∈ Lie(G)(Qp ) if and only if (1+εx, εw) ∈ G(Qp [ε]);
and to compute Ad(v)(x, w) we have to compute (1, v)(1 + εx, εw)(1, v)−1 in G(Qp [ε]). We have
(1, v)(1 + εx, εw)(1, −v) = (1, v)(1 + εx, 0)(1, εw − v)
= (1 + εx, ρ(1 − εx)(v) − v + εw) = (1 + εx, ε(dρ(x)(v) + w)),
which means Ad(1, v)(x, w) = (x, dρ(x)(v) + w). Notice that we are slightly abusing the notation and use the
addition for the group operation of V though it is realized as a subgroup of (GLN0 )Q . By this computation,
for any (x, w) ∈ W and v ∈ V(Qp ) ∩ GLN0 (Zp ), we have
pk0 dρ(x)(v) := (0, pk0 dρ(x)(v)) = Ad(pk0 v)(x, w) − (x, w) ∈ (W ∩ glN0 (Zp )) + pm glN0 (Zp ).
Thus we have
pk0 dρ Lie(Gs )(Qp ) ∩ glN0 (Zp ) V(Qp ) ∩ glN0 (Zp ) ⊆ (W ∩ glN0 (Zp )) + pm glN0 (Zp ).
Since V is a completely reducible Gs module with no trivial factors, we have
dρ Lie(Gs )(Qp ) ∩ glN0 (Zp ) V(Qp ) ∩ glN0 (Zp ) ⊇ pO(1) (V(Qp ) ∩ glN0 (Zp )),
where the implied constant depends on G and its embedding into (GLN0 )Q . Therefore V(Qp ) ⊆ W ; and the
claim follows.
The next lemma is crucial in proving the bounded generation claim and its proof is fairly involved. By the
strong-approximation, we know that the S-arithmetic group Γ0 := G(Q) ∩ GLN0 (ZS ), when infinite, is dense
in Gp := G(Qp ) ∩ GLN0 (Zp ) for any prime p which is not in S (here G is as before a simply-connected,
e of Gp , Γ0 ∩ H
e is dense in H.
e The next
connected, perfect Q-group.). And so for any open subgroup H
OΓ0 (1)
e
e
are, however, trapped within
lemma shows that elements of Γ0 ∩ H with the S-norm at most [Gp : H]
a proper algebraic subgroup of G.
Let ν∞ be the Archimedean place of Q, and Se := {ν∞
∪ S. Notice that ZS can be diagonally embedded
p}P
Q
x2 for x = (x1 , . . . , xN0 ) ∈ RN0 , and kxkp :=
in ν∈Se Qν as a discrete subgroup. Let kxkν∞ :=
Qi i
N0
0
max{|xi |p }i for x = (x1 , . . . , xN0 ) ∈ Qp . For any (xν ) ∈ ν∈Se QN
let k(xν )kSe := maxν∈Se kxν kν . For any
ν ,Q
Q
N0
0
λ ∈ GLN0 (ZS ), we let kλkS be the operator norm of λ : Se Qν → Se QN
ν .
36
ALIREZA SALEHI GOLSEFIDY
Lemma 42. As before, let G = Gs ⋉ V ⊆ (GLN0 )Q , where Gs is a semisimple, connected, simply-connected
Q-group and V is a Q-vector group. Let G be the Zariski-closure of G in (GLN0 )Z ; in particular G(ZS ) =
G(Q) ∩ GLN0 (ZS ) for any finite set of primes S. Let S be a finite set of primes such that G(ZS ) is infinite.
Suppose either p ≫G 1 or m0 ≫G 1. Let G := G(Zp ). Then there is a positive number δ such that for any
e of G[pm0 ] the set
prime p 6∈ S and an open subgroup H
e := {λ ∈ G(ZS ) ∩ H|
e kλkS ≤ [G(Zp ) : H]
e δ}
Lδ (H)
is in a proper Zariski-closed subgroup of G.
e
Proof. Since GLN0 (ZS ) is a discrete set with respect to the S-norm, we can and will assume that [G(Zp ) : H]
e is a subset of a finite subgroup and
is large enough. Otherwise choosing δ small enough, we have that Lδ (H)
we are done.
Suppose p or m0 are large enough so that Lemma 38 holds. Let G := G(Zp ), g := Lie(G)(Zp ), and
e := min{k ∈ Z+ | H
e ⊇ G[pk ]}.
l(H)
e is an open subgroup, l(H)
e is finite. By Lemma 38, for any m0 ≤ k ≤ l(H)
e − 1, we have
Notice that, since H
k
k+1
k+1
k
k+1
e ∩ G[p ])G[p
that (H
]/G[p
] is a proper fp -subspace of G[p ]/G[p ] ≃ g/pg. Thus we have
(68)
where d = dimfp g/pg.
e
e
0)
e ≤ pd(l(H)−m
,
pl(H)−m0 ≤ [G[pm0 ] : H]
e
e
Our first goal is to find a linear separation of an Ad(H)-orbit
with a good margin (see Claim 1) when l(H)
is large enough depending only on G. To achieve this goal, we start with finding primitive Zp -submodules of
e
g and gs := Lie(Gs )(Qp ) ∩ g that are almost Ad(H)-invariant
(see (71) and (80)).
Notice that, since V is a direct sum of non-trivial simple Gs -modules, for any prime p, G(Qp ) = Gs (Qp ) ⋉
V(Qp ) is compactly generated. Hence, by a result of Kneser [Kne64], G(ZS ) is a finitely generated group.
We fix a finite generating set Ω′ of G(ZS ).
e ≫G,C0 1,
Claim 1. For any constant C0 > 1 (later it will depend on dim G), there is c(G) such that, if l(H)
then for some w ∈ g and L ∈ g∗ := HomZp (g, Zp ) the following holds
e |L(Ad(h)(w))|p ≤ [G(Zp ) : H]
e −c ;
(1) for any h ∈ H,
′
e −c/C0 .
(2) for some γ ∈ Ω , |L(Ad(γ)(w))|p ≥ [G(Zp ) : H]
Proof of Claim 1. Let
l
e ∩ G[pl′ ])G[pl ]/G[pl ]) ⊆ g/pl−l′ g,
V := Ψppl′ ((H
l
′
e and l′ := ⌈l/2⌉ + m0 (Recall that Ψpl′ (1 + pl x) := π l−l′ (x) is a finite logarithmic map.). So
where l := l(H)
p
p
Pd′
there are a Zp -basis {e1 , . . . , ed } of g and positive integers n1 ≤ · · · ≤ nd′ such that V = πpl−l′ ( i=1 pni Zp ei )
(for some d′ ≤ d). Since
0 = n0 ≤ n1 ≤ · · · ≤ nd′ ≤ ⌊l/2⌋ − m0 =: nd′ +1 ,
there is i0 such that
(1/2d)l − m0 − 1 < ni0 − ni0 −1 ≤ (1/2)l − m0 .
(69)
Let W1 :=
(70)
Pi0 −1
i=1
pni Zp ei ⊂ g (if i0 = 1, then W1 = 0). So
ni
e
Ad(H)(W
1 ) ⊆ W1 + p 0 g.
SUPER-APPROXIMATION, II:
THE p-ADIC AND BOUNDED POWER OF SQUARE-FREE INTEGERS CASES.
Pi0 −1
37
Pi0 −1
Let W2 := i=1 Zp ei . Notice that W2 = W1 ∩ g where W1 = i=1 Qp ei is the Qp -span of W1 . Since
e
e we have
for any 1 ≤ i ≤ i0 − 1 and h ∈ H
πpni0 (W1 ) is H-invariant,
ni
Ad(h)(p ei ) =
iX
0 −1
nj
cj p e j +
where
pni0 c′j ej ,
j=1
j=1
cj , c′j
d
X
∈ Zp . Therefore
Ad(h)(ei ) =
iX
0 −1
(cj p
nj −ni
+
pni0 −ni c′j )ej
+
d
X
pni0 −ni c′j ej .
j=i0
j=1
e
Since g is H-invariant,
we have cj pnj −ni +pni0 −ni c′j ∈ Zp . Since nj is increasing, we have pni0 −ni ≥ pni0 −ni0 −1
for any i < i0 . Hence we have
ni −ni0 −1
e
g.
Ad(H)(W
2 ) ⊆ W2 + p 0
(71)
Next we would like to get some invariance in the semisimple part of g. We notice that Lie(G)(Qp ) =
Lie(Gs )(Qp ) ⊕ V(Qp ) (we are identifying V with its Lie algebra.). For any x ∈ Lie(G)(Qp ), we write its
components in the above decomposition by xs and xv ; that means x = xs + xv where xs ∈ Lie(Gs )(Qp ) and
Lie(V)(Qp ). There is a positive integer d1 depending only on G such that
(72)
kxkp ≤ max{kxs kp , kxv kp } ≤ pd1 kxkp
gs ⊕ V ⊆ g ⊆ p−d1 (gs ⊕ V),
and
where gs := Lie(Gs )(Qp )∩g and V := V(Qp )∩g. Let W2,s := {xs | x ∈ W2 } be the projection to Lie(Gs )(Qp )
of W2 . And so, by (71) and the fact that V is a normal subgroup of G, we have
ni −ni0 −1
e
g.
Ad(H)(W
2,s ) ⊆ W2,s + V(Qp ) + p 0
(73)
By (72), pd1 W2,s is a Zp -submodule of gs . Hence gs has a Zp -basis f1 , . . . , fds such that
′
(74)
W2,s =
ds
X
′
pni Zp fi
i=1
for some integers −d1 =:
n′0
≤
n′1
≤ ··· ≤
n′d′s
<
n′d′s +1
n′j0 +1 − n′j0 ≥
(75)
:= ∞. Let j0 be the smallest index such that
ni0 − ni0 −1 − d1
.
2ds
(Notice that, since n′0 := 0 and n′d′s +1 := ∞, there is such j0 .) In particular, we have
n′j0 <
(76)
ni0 − ni0 −1 − d1
.
2
Pj0
Pj0
Let W3,s := i=1
Zp fi . Notice that W3,s = W3,s ∩ gs where W3,s = i=1
Qp fi . So by (73) and (72) for any
e we have
1 ≤ i ≤ j0 and h ∈ H,
′
n′i
Ad(h)(p fi ) =
ds
X
′
pnj cj,s fj + pni0 −ni0 −1
where
∈ Zp and vi ∈ V(Qp ). Therefore
′
(77)
p−d1 c′j,s fj + vi ,
j=1
j=1
cj,s , c′j,s
ds
X
(Ad(h)(fi ))s =
ds
X
j=1
′
′
′
(pnj −ni cj,s + pni0 −ni0 −1 −d1 −ni c′j,s )fj +
ds
X
j=d′s
′
pni0 −ni0 −1 −d1 −ni c′j,s fj .
38
ALIREZA SALEHI GOLSEFIDY
e (Ad(h)(xs ))s ∈ gs , we have pn′j −n′i cj,s + pni0 −ni0 −1 −d1 −n′i c′ ∈ Zp and
Since, for any xs ∈ gs and h ∈ H,
j,s
′
pni0 −ni0 −1 −d1 −ni c′j,s ∈ Zp . By (76) we have
ni0 − ni0 −1 − d1 − n′i >
(78)
ni0 − ni0 −1 − d1
.
2
By (75), for j ≥ j0 + 1 and i ≤ j0 , we have
n′j − n′i ≥
(79)
ni0 − ni0 −1 − d1
.
2ds
By (77), (78), and (79), we get
(80)
⌊(ni0 −ni0 −1 −d1 )/(2ds )⌋
e
gs .
Ad(H)(W
3,s ) ⊆ W3,s + V(Qp ) + p
e
Having (71) and (80), we are ready to find the desired linear separation of an Ad(H)-orbit.
We do this
considering various cases.
Case 1. Suppose
(81)
Ad(G(ZS ))(W3,s ) 6⊆ W3,s + V(Qp ) + p⌊(ni0 −ni0 −1 −d1 )/(4ds C0 )⌋ gs .
By (81), we have that there is λ ∈ Ω′ and 1 ≤ i ≤ j0 such that
(82)
(Ad(λ)(fi ))s 6∈ W3,s + p⌊(ni0 −ni0 −1 −d1 )/(4ds C0 )⌋ gs .
For 1 ≤ i ≤ j0 , let fi∗ : g → Zp be a Zp -linear map such that
fi∗ (fi ) = pd1 , fi∗ (fj ) = 0 for j 6= i, and fi∗ (x) = fi∗ (xs ).
(Here xs is the Lie(Gs )(Qp )-component of x; and notice that because of (72) the image of fi∗ is a subset of
Zp .) By (82), we have that there is j > j0 such that
(83)
|fj∗ (Ad(λ)(fi ))|p > p−⌊(ni0 −ni0 −1 −d1 )/(4ds C0 )⌋ .
e we have
And, by (80) and j > j0 ≥ i, for any h ∈ H,
(84)
|fj∗ (Ad(h)(fi )|p ≤ p−⌊(ni0 −ni0 −1 −d1 )/(2ds )⌋ .
e ≫G,C0 1, we get that (83) and (84) imply w := fi and L := f ∗ satisfy the conditions of
So by (69), for l(H)
j
Claim 1.
Case 2. Suppose Lie(Gs )(Qp ) is the Qp -span W3,s of W3,s .
Let us recall that W1 is the Qp -span of W1 . And W3,s is a subset of the projection of W1 to Lie(Gs )(Qp ).
So by the assumption of Case 2, we have that the projection of W1 onto Lie(Gs )(Qp ) is onto. Since W1 is
a proper subspace of Lie(G)(Qp ), by Lemma 41 and the strong approximation there is c3 ≪G 1 such that
Pi0
Zp ei . So there is λ ∈ Ω′ and 1 ≤ i ≤ i0
πpc3 (W1 ∩g) is not G(ZS )-invariant. Recall that W2 = W1 ∩g = i=1
such that
(85)
Ad(λ)(ei ) 6∈ W2 + pc3 g.
Let {e∗1 , . . . , e∗d } be the dual of {e1 , . . . , ed }. Then, by (85) we have that there is j ≥ i0 such that
(86)
|e∗j (Ad(λ)(ei ))|p > p−c3 .
e we have
And by (71) and j ≥ i0 > i, for any h ∈ H,
(87)
|e∗j (Ad(h)(ei ))|p ≤ p−(ni0 −ni0 −1 ) .
e ≫G,C0 1, we get that (86) and (87) imply w := ei and L := e∗ satisfy the conditions of
So by (69), for l(H)
j
Claim 1.
SUPER-APPROXIMATION, II:
THE p-ADIC AND BOUNDED POWER OF SQUARE-FREE INTEGERS CASES.
39
Case 3. Suppose
Ad(G(ZS ))(W3,s ) ⊆ W3,s + V(Qp ) + p⌊(ni0 −ni0 −1 −d1 )/(4ds C0 )⌋ gs ,
(88)
and Lie(Gs )(Qp ) is not the Qp -span W3,s of W3,s .
By the strong approximation, the projection Kp of the closure of G(ZS ) to Gs (Qp ) contains Gs (Qp ) ∩
GLN0 (Zp )[pk0 ] where k0 depends only on G. So by Lemma 40 and (88), there are a positive integer c4 , which
depends only on G, and a normal subgroup GW3,s of Gs such that
p2c4 πpm−c4 (gW3,s ) ⊆ pc4 πpm−c4 (W3,s ) ⊆ πpm−c4 (gW3,s ),
(89)
where gW3,s := Lie(GW3,s )(Qp ) ∩ g and m := ⌊(ni0 − ni0 −1 − d1 )/(4ds C0 )⌋.
Since W3,s is a proper subspace of Lie(Gs )(Qp ), by the first inclusion in (89) we have that GW3,s is a
proper normal subgroup of Gs . So at least one simple factor Gi′0 is missing in Lie(GW3,s )(Qp ). Let gi′0 :=
Lie(Gi′0 )(Qp ) ∩ g. Then we have that gi′0 ∩ gW3,s = {0} and [gi′0 , gW3,s ] = {0}. Let x0 ∈ gi′0 be a unit vector.
e
Next we show that W3,s + Zp x0 + V(Qp ) is almost Ad(H)-invariant.
e and suppose h ∈ G[pk ] \ G[pk+1 ] for some k < l′ := ⌈l/2⌉ + m0 . Then
Let h ∈ H
log h ∈ (pk glN0 (Zp ) ∩ g) \ pk+1 glN0 (Zp ).
So
l′ −k
log hp
′
′
e
= pl −k log h ∈ pl glN0 (Zp ) ∩ log H.
Hence πpni0 (p−k log h) ∈ πpni0 (W1 ). This implies that there is yh ∈ W2,s such that
p−k log h ∈ yh + pni0 g + V(Qp ).
And so yh =
(90)
Pd′s
′
i=1
ai pni fi , for some ai ∈ Zp , which implies
p−k log h ∈ W3,s + p
min{ni0 ,n′i′ +1 }
0
g + V(Qp ).
By (75), we have
min{ni0 , n′i′0 +1 } ≥
(91)
ni0 − ni0 −1 − d1
− d1 ≥ 2C0 m − d1 .
2ds
Hence by (90) and (91), we have
p−k log h ∈ W3,s + p2C0 m−d1 g + V(Qp ).
(92)
By (89) and (92), we have
(93)
pc4 −k log h ∈ gW3,s + pmin{2C0 m−d1 ,m−c4 } g + V(Qp ).
By (93) and the fact that [gW3,s , x0 ] = {0}, we have
ad(log h)(x0 ) ∈ pmin{2C0 m−d1 ,m−c4 }−c4 g + V(Qp ).
(94)
e ≫G,C0 1 we have
Thus by m = ⌊(ni0 − ni0 −1 − d1 )/(4ds C0 )⌋, (69), and (94), for l(H)
(95)
Ad(h)(x0 ) = exp(ad(log h))(x0 ) ∈ x0 + pmin{2C0 m−d1 ,m−c4 }−c4 g + V(Qp ).
So by (80) and (95), we have
(96)
min{2C0 m−d1 ,m−c4 }−c4
e
g + V(Qp ).
Ad(H)(W
3,s + Zp x0 ) ⊆ W3,s + Zp x0 + p
On the other hand, looking at the action of G(ZS ) on gi′0 , we can choose x0 ∈ gi′0 , L0 ∈ HomZp (gi′0 , Zp ), and
a positive integer c5 which depends only on G such that for some λ ∈ Ω′ we have
(97)
|L0 ((Ad(λ)(x0 ))s )|p ≥ p−c5 , and L0 (x0 ) = 0.
40
ALIREZA SALEHI GOLSEFIDY
Now let Lie(Gs )(Qp ) = ⊕i Lie(Gs,iP
)(Qp ) be the decomposition of Lie(Gs )(Qp ) to its simple factors; and for
any x ∈ Lie(G)(Qp ) let x = xv + i xs,i be the such that xv ∈ Lie(V)(Qp ) and xs,i ∈ Lie(Gs,i )(Qp ). Then
for some positive integer c6 which depends only on G we have
pc6 kxkp ≥ max{kxs,i kp }i .
So there is a well-defined Zp -linear map L : g → Zp , L(x) := L0 (pc6 xs,i′0 ). We notice that
|L(Ad(λ)(x0 ))|p ≥ p−c5 , and L(W3,s + Zp x0 ) = 0.
(98)
e we have
And so by (96), for any h ∈ H,
|L(Ad(h)(x0 ))|p ≤ p− min{2C0 m−d1 ,m−c4 }−c4 .
(99)
e ≫G,C0 1, we get that (98) and (99) imply w := x0 and L satisfy the conditions of Claim
So by (69), for l(H)
1.
e are in a proper quadratic subvariety. Let g := Lie G. So
Next we will show elements with small height in H
there are finitely many ZS -linear maps fi (for 1 ≤ i ≤ s0 ) viewed as regular functions on the affine scheme
(glN0 )ZS of N0 -by-N0 matrices such that for any (unital commutative) ZS -algebra R, we have
g(R) = {x ∈ glN0 (R)| ∀i, fi (x) = 0}.
(100)
N02
Let us view glN0 (R) as R , and write its elements as vectors. This way we write fi (x1 , . . . , xN02 ) :=
P
j aij xj , where aij ∈ ZS . And consider the matrix A := [aij ] ∈ Ms0 ×N02 (ZS ). Since ZS is a PID, there are
γ1 ∈ GLs0 (ZS ), γ2 ∈ GLN02 (ZS ), and positive integers r1 |r2 | · · · |rN02 −d such that
0
0
γ2 ,
(101)
A = γ1
0 diag(r1 , . . . , rN02 −d )
where diag(r1 , . . . , rN02 −d ) is the diagonal matrix with diagonal entries ri ’s (this is known as the Smith
normal form when the diagonal block is at the top left corner.). Let us view the rows of γ2 as a ZS -basis
N2
N2
{e∗1 , . . . , e∗N 2 } of the dual of ZS 0 . And let {e1 , . . . , eN02 } be the dual ZS -basis of ZS 0 . So by (100) and (101),
0
for any ZS -algebra R with no additive torsion element, we have
(102)
x ∈ g(R) ⇐⇒ e∗d+1 (x) = · · · = e∗N 2 (x) = 0 ⇐⇒ x ∈
0
d
M
Rei .
i=1
And so in this case we have
(103)
g∗ (R) := HomR (g(R), R) =
d
M
Re∗i .
i=1
Claim 2. Suppose {e1 , . . . , eN02 } is a ZS -basis of glN0 (ZS ) such that (102) and (103) hold for any ZS -algebra
R with no additive torsion element. For (w1 , . . . , wd ) ∈ Rd , (l1 , . . . , ld ) ∈ Rd , and g ∈ G(R), let
X
e∗i (Ad(g)(ej ))li wj .
ηg (w1 , . . . , wd ; l1 , . . . , ld ) :=
i,j
2d
If δ is small enough, there is (w1 , . . . , wd , l1 , . . . , ld ) ∈ Q
(Q is the algebraic closure of Q) such that
e we have ηh (w1 , . . . , wd ; l1 , . . . , ld ) = 0,
(1) For any h ∈ Lδ (H)
(2) For some γ ∈ Ω′ we have ηγ (w1 , . . . , wd ; l1 , . . . , ld ) 6= 0.
e and some γ ∈ Ω′
Proof of Claim 2. By Claim 1, there are w ∈ g(Zp ) and L ∈ g(Zp )∗ such that for any h ∈ H
we have
e −c ,
e −c/C0 , and |L(Ad(h)(w))|p ≤ [G(Zp ) : H]
(104)
|L(Ad(γ)(w))|p ≥ [G(Zp ) : H]
SUPER-APPROXIMATION, II:
THE p-ADIC AND BOUNDED POWER OF SQUARE-FREE INTEGERS CASES.
41
for some c which depends only on G and some large positive number C0 which will be specified later.
Pd
Pd
Suppose L(Ad(λ)(w))−1 w = i=1 xi ei and L = i=1 yi e∗i ; and let x0 := (x1 , . . . , xd ) and y0 := (y1 , . . . , yd ).
Since g(Zp ) = g = Lie(G)(Qp ) ∩ glN0 (Zp ), it is a primitive Zp -submodule of glN0 (Zp ); and so
(105)
e c/C0 .
kx0 kp = |L(Ad(h)(w))|−1
p kwkp ≤ [G(Zp ) : H]
e and some
And clearly ky0 kp ≤ 1. By (104) and the way x0 and y0 are defined, we get that for any h ∈ H
′
γ ∈ Ω we have
(106)
e −c(1−1/C0 ) .
ηγ (x0 , y0 ) = 1, and |ηh (x0 , y0 )| ≤ [G(Zp ) : H]
To prove the claim, we proceed by contradiction and use an effective version of Nullstellensatz theorem [BY91]
(or [MW, Theorem IV]).
Suppose the following has no solution over Q:
e ηh (x, y) = 0, ηγ (x, y) = 1,
∀ h ∈ Lδ (H),
where γ ∈ Ω′ is the one given in (106). Notice that the total degree of ηg is two for any g ∈ G(Zp ); and the
e is at most [G(Zp ) : H]
e ΘΩ (δ) . So by the effective Nullstellensatz,
S-norm of the coefficients of ηh for h ∈ Lδ (H)
there are polynomials qγ , qh ∈ ZS [T1 , . . . , T2d ] and D0 ∈ ZS such that
X
(107)
qγ (x, y)(ηγ (x, y) − 1) +
qh (x, y)ηh (x, y) = D0 , and
e
h∈Lδ (H)
(108)
e ΘΩ (δ) ,
deg qγ , deg qh ≤ N (d), and kD0 kS ≤ [G(Zp ) : H]
e and N (d) is a positive integer which depends only on d := dim G. So by (108), (105), and
where h ∈ Lδ (H)
the fact that y0 ∈ Zdp , we have
e N (d)c/C0 .
|qh (x0 , y0 )|p ≤ [G(Zp ) : H]
(109)
e we have
Hence by (109) and (106), for any h ∈ Lδ (H)
(110)
e −c
|qh (x0 , y0 )ηh (x0 , y0 )|p ≤ [G(Zp ) : H]
1− N (d)+1
C
0
.
Now suppose C0 = 2N (d) + 2; so by (106), (107), and (110), we have
(111)
|D0 |p = qγ (x0 , y0 )(ηγ (x0 , y0 ) − 1) +
X
qh (x0 , y0 )ηh (x0 , y0 )
e
h∈Lδ (H)
p
Hence by (108) and (111) we get
e c/2 ≤ kD0 kS ≤ [G(Zp ) : H]
e ΘΩ (δ) ,
[G(Zp ) : H]
which gives us a contradiction for δ ≪Ω 1.
d
e −c/2 .
≤ [G(Zp ) : H]
d
(Going back to the proof of Lemma 42.) For a given (w, l) ∈ Q × Q , let
Vw,l (Q) := {g ∈ G(Q)| ηg (w, l) = 0}.
Then Vw,l gives us a closed subvariety of G. And since it is coming from an intersection of a hyperplane in
glN0 (Q) and G(Q), both its number of irreducible components and their degrees are bounded by a positive
integer N1 := N1 (G) which depends only on G.
Hence by [EMO05, Proposition 3.2] and its proof, there is a positive integer N2 := N2 (G) such that for any
subset A of G(Q) which generates a Zariski-dense subgroup of G(Q) we have
Q
(112)
N2 A 6⊆ Vw,l (Q),
42
ALIREZA SALEHI GOLSEFIDY
d
d
d
d
e ⊆ Vw,l (Q).
for any (w, l) ∈ Q × Q . By Claim 2, if δ ≪G 1, there is (w, l) ∈ Q × Q such that LN2 δ (H)
And so
Q
e
e
(113)
N2 Lδ (H) ⊆ LN2 δ (H) ⊆ Vw,l (Q).
e does not generate a Zariski-dense subgroup of G(Q);
Therefore by (112) and (113) we have that Lδ (H)
proving Lemma.
Proof of Proposition 37. By Nori’s strong approximation [Nor89, Theorem 5.4] we know that [G(Zp ) : Γ] ≪Γ
e := {h ∈ G(Zp )| πpn (h) ∈ H}.
1, where Γ is the closure of Γ in G(Zp ). For a proper subgroup H of πpn (Γ), let H
e is in a proper algebraic subgroup H of G for small enough δ1 . So by
So by Lemma 42 we have that Lδ1 (H)
Proposition 14 we have that
P (l) ({γ ∈ Γ| πpn (γ) ∈ H}) ≤ e−δ0 l
(114)
if 1 ≪ l ≪ δ1 log[πpn (Γ) : H] where δ0 is given by Proposition 14 and δ1 is given by Lemma 42.
If for some l ≥
n
δ
(l)
log p we have Pπpn (Ω) (H) ≥ [πpn (Γ) : H]−δ , then by [BG09, Remark, page 1060] (see [SG,
Lemma 8]) for any l′ ≤
enough.
n
δ
(2l′ )
log p we have Pπpn (Ω) (H) ≥ [πpn (Γ) : H]−2δ . This contradicts (114) if δ is small
5.2. Getting a large ideal by adding/subtracting a congruence subgroup boundedly many times.
The main goal of this section is to prove Proposition 43. At the end we also provide a few lemmas that are
needed for using Proposition 43 in the context of modules. The results of this section rely on properties of
p-adic analytic maps that are proved in the appendix.
Proposition 43. Let Gs ⊆ (GLN0 )Q be a semisimple Q-subgroup. Let ρ : Gs → GL(V) be a Q-homomorphism,
where V is a Q-vector group. Suppose no non-zero vector of V(Q) is ρ(Gs (Q))-invariant. For a prime p, let
P [pl ] := {g ∈ Gs (Qp ) ∩ GLN0 (Zp )| πpl (g) = 1}. Then for l ≫Gs ,ρ 1 we have
P
P
l
ΘGs ,ρ (l)
l
Zp [ρ(P [1])].
ΘG ,ρ (1) ρ(P [p ]) ⊇ p
ΘG ,ρ (1) ρ(P [p ]) −
s
s
(Notice that the implied constants are independent of p.)
Let us point out that the same proof gives us also the local version of Proposition 43.
Proposition 43′ . Let Gs ⊆ (GLN0 )Q be a semisimple Qp -subgroup. Let ρ : Gs → GL(V) be a Qp homomorphism, where V is a Qp -vector group. Let P [pl ] := {g ∈ Gs (Qp ) ∩ GLN0 (Zp )| πpl (g) = 1}. Suppose
that V(Qp ) has no non-zero Gs (Qp ) fixed vector. Then for l ≫Gs ,ρ 1 we have
P
P
l
ΘGs ,ρ (l)
l
Zp [ρ(P ))].
ΘG ,ρ (1) ρ(P [p ]) ⊇ p
ΘG ,ρ (1) ρ(P [p ]) −
s
s
Lemma 44. Let K be a field (only in this Lemma), and G be a subgroup of GLN0 (K). Suppose V := K N0
is a completely reducible G-module, and V has no non-zero G-fixed vector. Then there is no a ∈ MN0 (K)
such that Tr(ag) = 1 for any g ∈ G.
Proof. Suppose to the contrary that there is such a ∈ MN0 (K). Let A := K[G] be the K-span of G. Since
A has a faithful semisimple finite dimensional A-module, A is a semisimple K-algebra. Let
a := hg − 1| g ∈ Gi
be the ideal generated by G − 1 in A. Since Tr(ag) = 1 for any g ∈ G, we have that Tr(ax) = 0 for any
x ∈ a. Thus a is a proper ideal. On the other hand, since A is semisimple, by the idempotent decomposition
there are α1 , α2 ∈ A such that
(1) α2i = αi ,
(2) α1 + α2 = 1,
(3) a = Aα1 .
SUPER-APPROXIMATION, II:
THE p-ADIC AND BOUNDED POWER OF SQUARE-FREE INTEGERS CASES.
43
Since a is a proper ideal, α1 6= 1. And so there is a non-zero vector v ∈ V in the kernel of α1 , which implies
that av = Aα1 v = 0. Therefore for any g ∈ G we have gv = v which contradicts the assumption that V does
not have a non-zero G-fixed point.
Proof of Proposition 43. Since Gs is a semisimple group, V(Qp ) is a completely reducible Qp [ρ(Gs (Qp ))]module, for any prime p. As Gs (Qp ) is Zariski-dense in Gs (Qp ), we have that Qp [ρ(Gs (Qp ))] = Qp [ρ(Gs (Qp ))].
Therefore Qp [ρ(Gs (Qp ))] has a faithful completely reducible module, which implies that Qp [ρ(Gs (Qp ))] is a
semisimple algebra. Hence the nilradical of Qp [ρ(Gs (Qp ))] is zero, too. Thus Qp [ρ(Gs (Qp ))] is a semisimple
algebra and V(Qp ) is a completely reducible Qp [ρ(Gs (Qp ))]-module.
Next we show that V(Qp ) does not have a non-zero Gs (Qp )-invariant vector. If not, V(Qp )Gs (Q) := {v ∈
P
V(Qp )| ρ(Gs (Q))(v) = v} is non-zero. For v0 ∈ V(Qp )Gs (Q) we have v0 = i ai vi where ai ∈ Qp are Qlinearly independent and vi ∈ V(Q). Since ρ is defined over Q, for g ∈ Gs (Q) we have ρ(g)(vi ) ∈ V(Q) for
any i. Hence ρ(g)(v0 ) = v0 implies that for any i we have ρ(g)(vi ) = vi . Since Gs (Q) is Zariski-dense in Gs ,
we get that ρ(Gs (Q))(vi ), which is a contradiction.
Hence Lemma 44 implies that the constant function 1 does not belong to the linear span of the analytic
functions ρij where ρij are the entries of ρ : Gs (Qp ) → GL(V(Qp )) with respect to a Qp -basis of V(Qp ). So
we get the desired result by Corollary 50.
Next we prove a corollary of Lemma 39.
Corollary 45. Let R ⊆ Mn0 (Zp ) be a Zp -subalgebra. Let
0 = Wk0 +1 ⊆ Wk0 ⊆ . . . ⊆ W1 = Qnp 0
be a composition series of Qnp 0 as an Qp [R]-module, where Qp [R] is the Qp -span of R. Suppose vi ∈ Mi :=
Wi ∩ Znp 0 and kvi k := inf{kvi + wkp | w ∈ Wi+1 } ≥ q −1 , where q is a power of p. Then
k0
X
Rvi ⊇ q
ΘR,{Wj } (1)
Znp 0 .
i=1
Proof. For any 1 ≤ i ≤ k0 , we get a ring homomorphism from R to EndQp (Wi /Wi+1 ). Let Ri be its image.
Notice that Mi /Mi+1 can be embedded into Wi /Wi+1 and for any v ∈ Mi we have Ri (v+Mi+1 ) = Rv+Mi+1 .
So by Lemma 39 for the ring Ri we have that,
q ΘR,Wi (1) Mi ⊆ Rvi + Mi+1 ⊆ Mi .
By induction on i one can easily see that, for any 1 ≤ i ≤ k0 , we have
q ΘR,{Wj } (1) Znp 0 = q ΘR,{Wj } (1) M1 ⊆
i
X
Rvj + Mi+1 .
j=1
Lemma 46. Let R ⊆ Mn0 (ZS ) be a ZS -subalgebra. Assume the Q-span Q[R] of R is a semisimple algebra.
Suppose p is a large prime number (depending on R), and W is a composition factor of Qnp 0 ; that means
there are two Qp [R]-submodules W1 ⊆ W2 such that W = W2 /W1 is a simple Qp [R]-module. Then for any
v ∈ WZp := (W2 ∩ Znp 0 ) + W1 /W1 we have Zp [R]v = WZp if v 6∈ pWZp .
Proof. By Wedderburn theorem and Artin-Brauer-Noether theorem, for large enough p we have that
L
(1) Qp [R] ≃ i Mni (Ki ) where Ki are finite extensions of QL
p,
(2) under the above identification Zp [R] gets identified with i Mni (O
L Ki ),ni
.
(3) under the above identification Znp 0 naturally gets identified with i OK
i
44
Hence WZp is isomorphic to
ni
.
we have Mni (OKi )v = OK
i
ALIREZA SALEHI GOLSEFIDY
ni
OK
i
ni
where Zp [R] acts via Mni (OKi ). It is clear that for any unit vector v ∈ OK
i
Corollary 47. Let R ⊆ Mn0 (ZS ) be a ZS -subalgebra. Assume the Q-span Q[R] of R is a semisimple algebra.
Let
0 = Wk0 +1 ⊆ Wk0 ⊆ . . . ⊆ W1 = Qnp 0
be a composition series of Qnp 0 as an Qp [R]-module, where Qp [R] is the Qp -span of R. Suppose vi ∈ Mi :=
Wi ∩ Zpn0 and kvi k := inf{kvi + wkp | w ∈ Wi+1 } = 1. Then, for large enough p, we have
Znp 0
=
k0
X
Rvi .
i=1
Proof. It is a direct corollary of Lemma 46.
5.3. Getting a p-adically large vector in a submodule in boundedly many steps. The main goal of
this short section is proving a key lemma (Lemma 48). Using this Lemma, we will be able to get a Qp -basis
of V(Qp ) consisting of large vectors in boundedly many steps.
Lemma 48. Let Ω, Γ and G be as above. In particular, G = Gs ⋉ V where Gs is a semisimple Qp group and V is a Qp -vector group; and for some non-negative integer kp ≪Ω 1 (which is zero for large
enough p) Qp [pkp ] = Kp ⋉ Vp where Qp is the closure of Γ in G(Qp ), Qp [pkp ] := Qp ∩ GLN0 (Zp )[pkp ],
Kp := Gs (Qp ) ∩ GLN0 (Zp )[pkp ], and Vp := V(Qp ) ∩ GLN0 (Zp )[pkp ] (see Section 2.6).
Let V′ ⊆ V be a non-zero Qp -subgroup which is Gs -invariant. For any 0 < ε ≪Ω,V′ 1, there are δ > 0 and a
positive integer C such that the following holds:
Suppose p is a prime number, Q = pn , and nε ≫Ω,V′ 1.
Q
If PQ (δ, A, l) holds and πQ (Γ[q2 ]) ⊆ πQ ( C A)πQ (Vp ∩ V′ (Qp )) where q2 = pn2 and n2 ≤ ε2 n, then
Q
(115)
{v ∈ V′ (Qp )| kvkp > Qε , πQ (v) ∈ πQ ( 4C A)}
is non-empty.
Proof. Since V′ is Gs -invariant and it commutes with V, V′ is a normal subgroup of G. On the other
hand, GLN0 (Zp )[pkp ] is a normal subgroup of GLN0 (Zp ) and Qp ⊆ GLN0 (Zp ); and so Vp ∩ V′ (Qp ) =
V′ (Qp ) ∩ GLN0 (Zp )[pkp ] is a normal subgroup of Qp . Let ηq1 : πq1 (Qp ) → πq1 (Qp )/πq1 (Vp ∩ V′ (Qp )) be the
projection map where q1 = pn1 and n1 = ⌊εn⌋. So by the assumption there is a section
s : ηq1 (πQ (Qp [q2 ])) → πq1 (Qp )
Q
such that the image of s is a subset of πq1 ( C A) (by section we mean ηq1 (s(x)) = x). Notice that
−1
−1
−1
−1
ηq1 (s(x−1
s(x1 )s(x−1
ηq1 (s(x1 ))ηq1 (s(x2−1 )) = (x−1
(x1 )(x−1
1 x2 )
2 )) = ηq1 (s(x1 x2 ))
1 x2 )
2 ) = 1.
Q
′
−1
Hence s(x−1
s(x1 )s(x−1
q1 (Vp ∩ V (Qp )). Now if we assume to the contrary that
1 x2 )
2 ) ∈ πq1 ( 3C A) ∩ πQ
the set in (115) is empty, then we get that πq1 ( 3C A) ∩ πq1 (Vp ∩ V′ (Qp )) = {1}. And so s is a group
homomorphism. So H := Im(s) is a subgroup of πq1 (Qp ), and [πq1 (Qp ) : H] ≥ |πq1 (Vp ∩ V′ (Qp ))| ≥ pΘV′ (nε) .
By Proposition 37 we have that
(2l)
Pπq (Ω) (H) ≤ [πq1 (Γ) : H]−ΘΩ (1) ≤ p−ΘΩ,V′ (nε) .
1
Q
On the other hand, if x ∈ πq1 ( 2C A) ∩ πq1 (Γ[q2 ]), then
Q
x−1 s(ηq1 (x))) ∈ πq1 ( 4C A) ∩ πq1 (Vp ∩ V′ (Qp )).
(116)
So by the contrary assumption we have that x ∈ H, which means
Q
(117)
πq1 ( 2C A) ∩ πq1 (Γ[q2 ]) ⊆ H.
SUPER-APPROXIMATION, II:
By the assumption we have
THE p-ADIC AND BOUNDED POWER OF SQUARE-FREE INTEGERS CASES.
(l)
PΩ (A)
45
> Q−δ (where δ is small positive number to be determined later). So
2
(l)
(x′ )) > Q−δ |πq2 (Γ)|−1 = p−ΘΩ (nε
PΩ (A ∩ πq−1
2
)
2
= p−n(δ+ΘΩ (ε
))
,
2
for some x′ ∈ πq2 (Γ). Hence P (2l) (A · A ∩ Γ[q2 ]) > p−nΘΩ (ε ) for small enough δ. Therefore by (117) and
(116) we have
−nΘΩ (ε2 ) < −nΘΩ,V′ (ε),
which is a contradiction if 0 < ε ≪Ω,V′ 1.
5.4. Proof of super-approximation: the p-adic case. Let us recall that G = Gs ⋉ V where Gs is a
semisimple Qp -group and V is a Qp -vector group; let ρ : Gs → GL(V) be the Qp -representation which
gives us the action of Gs on V. Let us recall that, since G is perfect, no non-zero vector of V(Qp ) is
ρ(Gs (Qp ))-invariant. Let Vi be Qp -subgroups of V such that
(1) Vi (Qp ) are irreducible Gs (Qp )-modules.
0
Vi .
(2) V = ⊕ki=1
Let P [pk ] := {g ∈ Gs (Qp )| kg − 1kp ≤ p−k } for any non-negative integer k.
Inductively we do the following simultaneously:
(1) construct a permutationL
σ : {1, . . . , k0 } → {1, . . . , k0 },
(2) find vectors vj ∈ Wj := i∈{1,...,k0 }\{σ(1),...,σ(j−1)} Vσ(i) (Qp ),
such that the following holds:
(1) for small enough δ, large enough C (depending on ε, Ω, and Vi ) and any 1 ≤ j ≤ k0 we have
Q
Θ (1)
(118)
πQ Γ[pn(ε G ) ] ⊆ πQ ( Θ(C) A)πQ (Vp ∩ Wj );
(2) vj ∈ Wj ;
(3) v 1 , . . . , v j generate W1 /Wj+1 as a Gs (Qp )-module, where v i is the projection of vi .
(4) logp kv j kp ≫ nεΘG (1) ;
Pj P
P
Θ (1)
Θ (1)
n(εΘG (1) )
(5)
])(vi ) − ΘG (1) ρ(P [pn(ε G ) ])(vi ) + Wj+1 ⊇ pn(ε G ) Vp .
i=1
ΘG (1) ρ(P [p
By [SG, Theorem 36], if δ is small enough and C is large enough, (118) holds for j = 1. Now we suppose
(118) holds for 1 ≤ j ≤ j0 and we have already found v1 , . . . , vj0 −1 as desired. We will define σ(j0 ) and find
vj0 so that the above properties hold.
Since (118) holds for j = j0 , by Lemma 48 there is vj0 ∈ Wj0 such that
Q
(119)
logp kvj0 kp ≫ nεΘG (1) , and πQ (vj0 ) ∈ πQ ( Θ(C) A).
Since logp kvj0 kp ≫ nεΘG (1) , the projection v j0 of vj0 to Vσ(j0 ) for some
σ(j0 ) ∈ {1, . . . , k0 } \ {σ(1), . . . , σ(j0 − 1)}
Θ (1)
has length at least ΘVi (pΘ(nε G ) ). Now by Proposition 43, and Corollaries 45 and 47 one gets all the
mentioned properties except the first one.
By [SG, Theorem 36] we have that for small enough δ and large enough C
Q
Θ(1)
(120)
πQ (P [pnε
]) ⊆ πQ ( C A)πQ (Vp ).
To simplify our notation let us drop the constant power of ε in the rest of the argument. Hence by (119) we
have
Q
(121)
πQ (pnε Vp ) ⊆ πQ ( Θ(C) A)πQ (Vp ∩ Wj ).
46
ALIREZA SALEHI GOLSEFIDY
For any g1 , g2 ∈ P [pnε ] and v1 , v2 ∈ Vp , we have
[(g1 , v1 ), (g2 , v2 )] ∈ P [p2nε ] ⋉ (pnε Vp ).
Hence by (120) and (121) we have
(122)
Q
πQ ({[g1 , g2 ]| g1 , g2 ∈ P [pnε ]}) ⊆ πQ ( Θ(C) A)πQ (Vp ∩ Wj ).
Let X := {[g1 , g2 ]| g1 , g2 ∈ P [pnε ]}. Using properties of the finite logarithmic maps (see [SG, Lemma 34]),
we have
Q
( dim Gs X ∩ P [pknε+Θ(1) ])P [p2knε ] = P [pknε+Θ(1) ],
Q
for any integer k ≥ 2. Therefore πQ ( Θ(− log ε) X) ⊇ πQ (P [pnΘ(ε) ]). Hence
Q
(123)
πQ (P [pnΘ(ε) ]) ⊆ πQ ( Θ(C) A)πQ (Vp ∩ Wj ).
And so by (121) and (123) we have
which finishes the proof.
Q
πQ (Γ[pnΘ(ε) ]) ⊆ πQ ( Θ(C) A)πQ (Vp ∩ Wj ),
6. Appendix A: quantitative open image for p-adic analytic maps.
The main goal of this appendix is to prove that one can get a large open set by adding the image of an
analytic function in controlled number of times (see Proposition 49).
Proposition 49. Let K be a characteristic zero non-Archimedean local field. Let O := OK be its ring
of integers, and p be a uniformizing element. Let U ⊆ K n0 be a neighborhood of the origin, and F :=
(f1 , . . . , fd0 ) : U → K d0 be an analytic function. Suppose the constant function 1 is not in the K-span of fi .
Then for any l ≫F 1 we have
P
P
ΘF (l) d0
l n0
l n0
O .
ΘF (1) F (p O ) ⊇ (K-span of F (U )) ∩ p
ΘF (1) F (p O ) −
In fact, we prove the following refinement of Proposition 49. In Proposition 49′ , we pinpoint how the implied
constants depend on the given analytic functions (with a few extra technical assumptions). This type of
control helps us to prove Corollary 50. Corollary 50 is the only result in the appendix that is needed in the
main part of the article. In Corollary 50, we deal with polynomial maps that are defined over a number
field, and prove that their images in all the non-Archimedean completions are uniformly large.
Proposition 49′ . Let K be a characteristic zero non-Archimedean local field. Let O := OK be its ring of
integers, and p be a uniformizing element. Let F := (f1 , . . . , fd0 ) : On0 → Od0 be an analytic function of the
form
X
F (x) =
(ci,1 xi , . . . , ci,d0 xi ),
i
i
where i = (i1 , . . . , in0 ) ranges over multi-indexes non-negative integers and xi = xi11 · · · xnn00 . Suppose
(1) |ci,j |p ≤ 1 for any i = (i1 , . . . , im ) and any 1 ≤ j ≤ d0 ,
′
(2) L : K d0 → K d0 is a linear embedding such that F (x) = L(fj1 (x), . . . , fjd′ (x)) for some indexes
0
1 ≤ ju ≤ d0 ,
(3) | det(cii ,ju )1≤i,u≤d′0 |p = |pk0 |p for some multi-indexes i1 , . . . , id′0 whose coordinates add up to a number
at most m0 and some indexes j1 , . . . , jd′0 .
Then for any l ≫d′0 ,n0 ,m0 ,k0 ,kLkp 1 we have
P
P
l n0
Θ(l) d0
l n0
O ,
Θ(1) F (p O ) −
Θ(1) F (p O ) ⊇ (K-span of F (U )) ∩ p
where all the implied constants depend on d0 , n0 , m0 and k0 .
SUPER-APPROXIMATION, II:
THE p-ADIC AND BOUNDED POWER OF SQUARE-FREE INTEGERS CASES.
47
Corollary 50. Let κ be a number field. Let f1 , . . . , fd0 ∈ κ[x1 , . . . , xn0 ]. Suppose 1, f1 , . . . , fd0 are linearly
independent where 1 is the constant polynomial one. Let F := (f1 , . . . , fd0 ). Then for any l ≫F 1 and any
p ∈ Vf (κ) we have
P
P
l n0
ΘF (l) d0
l n0
Op ,
ΘF (1) F (p Op ) ⊇ (κp -span of F (Op )) ∩ p
ΘF (1) F (p Op ) −
where κp is the completion of κ with respect to the finite place p ∈ Vf (κ), Op is the ring of integers of κp ,
and p also shows a uniformizing element of Op .
Let us recall the needed notation from p-adic analysis. If U is an open subset of K, then ∇k U :=
{(x1 , . . . , xk ) ∈ U × · · · × U | xi 6= xj if i 6= j}. If f : U → K, then Φk f : ∇k+1 U → K is defined
recursively
Φk−1 f (x1 , x3 , . . . , xk ) − Φk−1 f (x2 , . . . , xk )
,
Φk f (x1 , . . . , xk ) :=
x1 − x2
and Φ0 f = f . If f is an analytic function, then Φn f can be uniquely extended to a continuous function
n
i
Φ f : U n+1 → K for any n and f (i) (a) = i!Φ f (a, . . . , a). Similarly for open subsets Ui ⊆ K and a (multivariable) function f : U1 × · · · × Um → K we can define the kth order difference quotient Φki f of f with
respect to the ith variable. And then for a multi-index i := (i1 , . . . , im ) we set
Φi f := Φi11 ◦ Φi22 ◦ · · · ◦ Φimm f.
Φi f : ∇i1 +1 U1 × · · · × ∇im +1 Um → K,
im +1
If f is analytic, then Φi f can be uniquely extended to a continuous function Φi f : U1i1 +1 × · · · × Um
→ K,
and
∂i f (a1 , . . . , am ) = i!Φi f (θi (a1 , . . . , am )),
where i! := i1 !i2 ! · · · im ! and
θi (a1 , . . . , am ) := (a1 , . . . , a1 , a2 , . . . , a2 , . . . , am , . . . , am ).
|
{z
}
| {z } | {z }
i1 +1−times i2 +1−times
im +1−times
m
Let U := U1 × · · · × Um ⊆ K be an open subset, fi : U → K be analytic functions, and F := (f1 , . . . , fd ) :
U → K d (in computations, it is considered as a d-by-1 column matrix). Then dF (x) is a d-by-m matrix
dF (x) := [∂j fi (x)].
For any 1 ≤ i, j ≤ m, let Φei +ej F (•) be the column vector whose kth entry is Φei +ej fk (•). In the above
setting, by Taylor expansion we have
(124)
F (x0 + x) = F (x0 ) + dF (x0 )x + R2 F (x0 + x, x0 ),
P
where R2 F (x0 + x, x0 ) is of the form i,j xi xj Φei +ej F (•) (the entries are the entries of either x0 + x or x0 )
and x = (x1 , . . . , xm ).
For a d-by-m matrix X = [v1 · · · vm ] with entries in K, let
N (X) :=
max
1≤i1 ≤···≤id ≤m
| det[vi1 · · · vid ]|.
Let {ei }i be the standard basis of K n ; and for any I := {i1 < i2 <V · · · < im } ⊆ {1, . . . , n}, let eI :=
m
(K n ). For any
ei1 ∧ . . . ∧ eim . It is well-known that {eI }I⊆{1,...,n}, |I|=m is a basis of
X
V
xI eI ∈ m (K n ), we let kxk := max{|xI || I ⊆ {1, . . . , n}, |I| = m}.
x=
I⊆{1,...,n}, |I|=m
Lemma 51.
(1) If w1 , . . . , wd are the rows of X ∈ Md,m(K), then N (X) = kw1 ∧· · ·∧wd k; in particular
N (X) = 0 if rank(X) < d.
(2) If X ∈ Md,m (O) and N (X) ≥ |pk0 |, then for any positive integer l and y ∈ pl+k0 Od there is
x ∈ pl Om such that Xx = y.
(3) For any 1 ≤ i ≤ m and vectors vj we have
N [v1 · · · vi (vi+1 − vi ) · · · (vm − vi )] ≤ (d + 1)N [v1 · · · vm ].
48
ALIREZA SALEHI GOLSEFIDY
Proof. The first part is clear. Now suppose that N (X) ≥ |pk0 |; then there are d columns vi1 , . . . , vid such
that | det[vi1 · · · vid ]| ≥ |pk0 |. Let XI = [vi1 · · · vid ]. Then XI adj(XI )y = det(XI )y, and so there is x′ ∈ pl Od
such that XI x′ = y. Therefore there is x ∈ pl Om such that Xx = y.
For the third part, we get an upper-bound for the determinant of a d-by-d submatrix:
| det[vi1 · · · vij (vij+1 − v) · · · (vid − v)]| = kvi1 ∧ · · · ∧ vij ∧ (vij+1 − v) ∧ · · · ∧ (vid − v)k
= kvi1 ∧ · · · ∧ vid −
d
X
vij+1 ∧ · · · ∧ vik−1 ∧ v ∧ vik+1 ∧ · · · ∧ vid k
k=j+1
≤ (d + 1)N ([v1 · · · vm ]),
where v is either 0 or vi .
In what follows, a series of Lemmas are proved in pairs. In the first ones the constants depend on F , and
they are geared towards proving Proposition 49. In the second ones, we make the needed modification and
make them suitable for proving Proposition 49′ .
Lemma 52 (Hensel’s lemma). Let x0 ∈ K m and U ⊆ K m be an open neighborhood of x0 . Let fi : U → K
be analytic functions and F = (f1 , . . . , fd ). Suppose there is a positive integer k0 such that
(1)
(2)
(3)
(4)
x0 + pk0 Om ⊆ U ,
kdF (x0 )k ≤ 1,
for any i, j, if kx′ − θei +ej (x0 )k ≤ |pk0 |, then kΦei +ej F (x′ )k ≤ 1,
N (dF (x0 )) ≥ |pk0 |.
Then for any l ≥ k0 and any y ∈ Od there is x1 ∈ x0 + pl Om such that
kF (x1 ) − F (x0 ) − pl+k0 yk ≤ |p2l |.
Proof. By (124) we have that
(125)
kF (x0 + pl x) − F (x0 ) − pl dF (x0 )xk = k
X
(pl xi )(pl xj )Φei +ej F (xij )k,
i,j
where x = (x1 , . . . , xm ), kxk ≤ 1 and kxij − θei +ej (x0 )k ≤ |pl |. Now by Lemma 51 for any y ∈ Od there is
x ∈ Om such that dF (x0 )x = pk0 y. And so by (125)
kF (x0 + pl x) − F (x0 ) − pl+k0 yk ≤ |p2l |.
Lemma 53 (Quantitative open function theorem). Let x0 ∈ K m and U ⊆ K m be an open neighborhood of
x0 . Let fi : U → K be analytic functions and F = (f1 , . . . , fd ). Suppose
(1) kdF (x0 )k ≤ 1,
(2) k∂ij F (x0 )k ≤ 1, for any i, j,
(3) N (dF (x0 )) ≥ |pk0 |, for some positive integer k0 .
Then for any large enough integer l (depending on F and U ) we have
F (x0 + pl Om ) ⊇ F (x0 ) + pl+k0 Od .
Proof. By the continuity of dF and Φei +ej F , for large enough l (in particular l > k0 ), we have that
(1) kdF (x)k ≤ 1 if kx − x0 k ≤ |pl |,
(2) kΦei +ej F (x′ )k ≤ 1 if kx′ − θei +ej (x0 )k ≤ |pl |,
(3) N (dF (x)) ≥ |pk0 | if kx − x0 k ≤ |pl |.
SUPER-APPROXIMATION, II:
THE p-ADIC AND BOUNDED POWER OF SQUARE-FREE INTEGERS CASES.
49
By induction on i we prove that for any y ∈ Od there are xi ∈ x0 + pl Om and integers li ≥ 2l such that
(1) l1 := 2l and li+1 := 2(li − k0 ),
(2) kF (xi ) − F (x0 ) − pl+k0 yk ≤ |pli |.
Lemma 52 gives us the base of the induction. By induction hypothesis, there is xi ∈ x0 + pl Om such that
F (xi ) − F (x0 ) − pl+k0 y ∈ pli Od .
Hence by Lemma 52 there is xi+1 ∈ xi + pli −k0 Om ⊆ x0 + pl Om such that
kF (xi+1 ) − F (xi ) + (F (xi ) − F (x0 ) − pl+k0 y)k ≤ |p2(li −k0 ) | = |pli+1 |.
This proves the induction step. One can easily see that {li } is a strictly increasing integer sequence. So {xi }
is a Cauchy sequence. Therefore limi→∞ xi = x ∈ x0 + pl Om by the compactness of x0 + pl Om , and by the
continuity of F we have F (x) = F (x0 ) + pl+k0 y.
Lemma 53′ . Let F := (f1 , . . . , fd0 ) : On0 → Od0 , F (x) =
such that
P
i
i
i (ci,1 x , . . . , ci,d0 x )
be an analytic function
(1) |ci,j | ≤ 1 for any i and j,
(2) N (dF (0)) ≥ |pk0 | for some positive integer k0 .
Then for any l ≫k0 1 we have
F (pl O) ⊇ F (0) + pl+k0 Od0 .
P
Proof. Since F (x) := i (ci,1 xi , . . . , ci,d0 xi ) and |ci,j | ≤ 1, we have kdF (x)k ≤ 1 and kΦei +ej F (x′ )k ≤ 1 for
any kxk ≤ 1 and kx′ k ≤ 1.
Since |ci,j | ≤ 1 and N (dF (0)) ≥ |pk0 |, we have that N (dF (x)) ≥ |pk0 | for any x ∈ pk0 +1 O. One can finish
the argument as in the proof of Lemma 53.
Lemma 54. Let U be a neighborhood of 0 ∈ K, and fi : U → K be analytic functions. Suppose 1, f1 , . . . , fd
are linearly independent. Then for large enough m (depending on fi ) and large enough l (depending on fi
and m) we have that, for any x1 , . . . , xm ∈ pl O,
N ([fi′ (xj )]) ≫
Y
|(xi − xj )|,
1≤i<j≤m
where the implied constant depends on fi .
(j)
Proof. Notice that we can rescale, i.e. change fi to gi (x) := fi (pk x), and make sure that |fi (0)| ≤ 1 for
any 1 ≤ i ≤ d and any positive integer j.
(j)
Since fi are analytic and 1, f1 , . . . , fd are linearly independent, for large enough m we have that rank([fi (0)]) =
(j)
d, where 1 ≤ i ≤ d and 1 ≤ j ≤ m. So N ([fi (0)]) = |pk0 | for some non-negative integer k0 . Hence by the
j
j
continuity of Φ fi , for large enough l, we have that N ([Φ fi (xj )]) ≫ 1 if xj ∈ pl Oj+1 for any 1 ≤ j ≤ m.
50
ALIREZA SALEHI GOLSEFIDY
Let F (x) be the column vector [f1 (x) · · · fd (x)]T . Then by the repeated use of the third part of Lemma 51
we have
N ([fi′ (xj )]) = N [F ′ (x1 ) · · · F ′ (xm )]
≥ (d + 1)−1 N [F ′ (x1 )(F ′ (x2 ) − F ′ (x1 )) · · · (F ′ (xm ) − F ′ (x1 ))]
!
m
Y
1
1
−1
|xi − x1 | N [F ′ (x1 )Φ F ′ (x2 , x1 ) · · · Φ F ′ (xm , x1 )]
≥ (d + 1)
i=2
≥ ···
≥ (d + 1)−m
≫
Y
1≤i<j≤m
Y
|xi − xj |.
1
2
|xi − xj | N [F ′ (x1 )Φ F ′ (x2 , x1 )Φ F ′ (x3 , x2 , x1 ) · · · Φ
m−1
F ′ (xm , . . . , x1 )]
1≤i<j≤m
Lemma 54′ . Let F := (f1 , . . . , fd0 ) : O → Od0 , F (x) =
that
P
i
i
i (ci,1 x , . . . , ci,d0 x )
be an analytic function such
(1) |ci,j | ≤ 1 for any i and j,
(2) | det(cie ,j )| = |pk0 | for some indexes i1 , . . . , id0 ≤ m0 and some positive integer k0 .
Then for any l ≫k0 1 we have that, for any x1 , . . . , xm0 ∈ pl O,
Y
N ([fi′ (xj )]) ≥ |pΘk0 ,d0 ,m0 (1) |
|xi − xj |.
1≤i<j≤m0
j
ie
Proof. Since |ci,j | ≤ 1 and | det(ci,j )| = |pk0 |, we have that N (Φ fi (xj )) ≥ | det(Φ fj (xie ))| = |pk0 | for any
xj ∈ pk0 +1 Oj . Now one gets the claim as in the proof of Lemma 54.
Lemma 55. Let U be a neighborhood of 0 ∈ K, and fi : U → K be analytic functions. Suppose 1, f1 , . . . , fd
are linearly independent. Let F = (f1 , . . . , fd ). Then for any l ≫F 1 we have
P
P
l
ΘF (l) d
l
O .
ΘF (1) F (p O) ⊇ p
ΘF (1) F (p O) −
Proof. By Lemma 54 for large enough m0 (depending on F ) and large enough l (depending on F and m0 )
we have
Y
N (dFe (x)) ≫
|xi − xj |,
1≤i<j≤m0
where Fe(x) = F (x1 ) + · · · + F (xm0 ) and x = (x1 , . . . , xm0 ) has norm at most |pl |. Let x0 = (x1 , . . . , xm0 ) be
such that |xi | ≤ |pl | for any i and |xi − xj | ≥ |p2l | for any i 6= j. So N (dFe (x0 )) ≫ pΘF,m0 (l) . By rescaling, if
(j)
needed, we can assume that |fi (0)| ≤ 1 for any 1 ≤ i ≤ d and any positive integer j. And so kdFe (x0 )k ≤ 1
and k∂ij Fe(x0 )k ≤ 1. Therefore by Lemma 53 we have
and so
P
m0
Fe (pl Om0 ) = Fe(x0 + pl Om0 ) ⊇ Fe(x0 ) + pΘm0 ,F (l) Od ,
F (pl O) −
′
P
m0
F (pl O) = Fe(pl Om0 ) − Fe(pl Om0 ) ⊇ pΘm0 ,F (l) Od .
d0
Lemma 55 . Let F := (f1 , . . . , fd0 ) : O → O , F (x) =
that
(1) |ci,j | ≤ 1 for any i and j,
P
i
i (ci,1 x
i
, . . . , ci,d0 x ) be an analytic function such
SUPER-APPROXIMATION, II:
THE p-ADIC AND BOUNDED POWER OF SQUARE-FREE INTEGERS CASES.
51
(2) | det(cie ,j )| = |pk0 | for some indexes i1 , . . . , id0 ≤ m0 and some positive integer k0 .
Then for any l ≫k0 ,m0 1 we have
P
m0
F (pl O) −
P
m0
Pm0
Proof. As in the proof of Lemma 55, let Fe(x) =
l
l
m0
F (pl O) ⊇ pΘm0 ,d0 ,k0 (l) Od0 .
F (xi ). Hence by Lemma 54′ for l ≫k0 1 we have
Y
|xi − xj |,
N (dFe (x)) ≥ |pΘk0 ,d0 ,m0 (1) |
for any x ∈ p O . Let x0 ∈ p O
there is such x0 . Hence
m0
i=1
1≤i<j≤m0
be such that |xi − xj | ≥ |p2l | for any i 6= j. Notice that, since l ≫m0 1,
N (dFe (x0 )) ≥ |pΘm0 ,d0 ,k0 (l) |.
Let Fx0 (x) := Fe(x + x0 ). Hence Fx0 has a Taylor series expansion and its coefficients have norm at most
one. Moreover N (dFx0 (0)) ≥ |pΘm0 ,d0 ,k0 (l) |. Therefore by Lemma 53′ we have
Fx0 (pl O) ⊇ Fx0 (0) + pΘm0 ,d0 ,k0 (l) Od0 .
One can finish the proof as above.
n0
Lemma 56. Let U ⊆ K be a non-empty open subset. Let fi : U → K be analytic functions such that
1, f1 , . . . , fd are linearly independent. Then there is a polynomial curve r : K → K n0 such that 1, f1 ◦
r, . . . , fd ◦ r are linearly independent, and defined on a neighborhood of 0 ∈ K.
Proof. Let x0 be a point in U . After rescaling, if needed, we can assume that |∂i fj (x0 )/i!| ≤ 1 for any i
which is not zero. Since 1, f1 , . . . , fd are linearly independent, for large enough m we have that
rank[∂i fj (x0 )]1≤kik1 ≤m, 1≤j≤d = d.
Thus
N ([∂i fj (x0 )/i!]) = |pk0 |,
(126)
where k0 is a non-negative integer. Now let
2
n0 −1
r(t) := x0 + p(t, ts , ts , . . . , ts
)
where s is sufficiently large (to be determined later). By the Taylor expansion of fj we have
X (j)
Pn 0
k−1
fj (r(t)) =
ai pkik1 t k=1 ik s ,
i
where i = (i1 , . . . , in0 ) and
(j)
ai
(127)
= ∂i fj /i!. And so
∞
X
fj (r(t)) =
n=0
{i|
P
X
(j)
k−1 =n}
k ik s
ai pkik1 tn .
We make the following two observations:
P
P
(j)
(j) kik1
P
≡ {i|kik∞ <s, P ik sk−1 =n} ai pkik1 (mod ps ),
(1)
{i| k ik sk−1 =n} ai p
k
P
P
(2) If kik∞ , ki′ k∞ < s and k ik sk−1 = k i′k sk−1 , then i = i′ .
In particular we have
{i|
P
X
k−1 =n}
k ik s
(j)
ai pkik1
(
0
≡
(j)
ai(n) pki(n)k1
if n ≥ sn0
0 ≤ n ≤ sn0 − 1
(mod ps )
52
ALIREZA SALEHI GOLSEFIDY
P
where i(n) := (i1 , . . . , in0 ) gives us the s-adic digits of n; that means n = k ik sk−1 . Therefore we have
X
(j)
(j) ki(n)k1
(128)
N
ai pkik1
]) (mod ps ).
≡ N ([ai(n) p
{i|
P
k−1 =n}
k ik s
1≤n≤sn0 −1, 1≤j≤d
Now suppose s > dm + k0 ; in particular, if kik1 ≤ m, then i = i(n) for some n < sn0 . And so we have
(j)
N ([ai(n) pki(n)k1 ]) ≥ |pmd+k0 |.
(129)
Hence by (128), (129) and s > dm + k0 we have
X
(j)
N
ai pkik1
{i|
which implies that
P
k−1 =n}
k ik s
rank
{i|
P
1≤n≤sn0 −1, 1≤j≤d
X
(j)
k−1 =n}
k ik s
and 1, f1 ◦ r, . . . , fd ◦ r are linearly independent.
ai pkik1
s
6≡ 0 (mod p ),
1≤n≤sn0 −1, 1≤j≤d
= d,
Lemma 56′ . Let F := (f1 , . . . , fd0 ) : On0 → Od0 be an analytic function of the form
X
F (x) =
(ci,1 xi , . . . , ci,d0 xi ),
i
i
where i = (i1 , . . . , in0 ) ranges over multi-indexes non-negative integers and xi = xi11 · · · xnn00 . Suppose
(1) |ci,j | ≤ 1 for any i = (i1 , . . . , im ) and any 1 ≤ j ≤ d0 ,
(2) | det(cii ,j )| = |pk0 | for some multi-indexes i1 , . . . , id0 whose coordinates add up to a number at most
m0 and some positive integer k0 .
Then there is a polynomial curve r : O → On0 such that | det(c′ie ,j )| ≥ |pΘm0 ,d0 ,k0 (1) | for some indexes
P
ie ≪m0 ,n0 ,d0 ,k0 1 where fj ◦ r(x) = i c′i,j xi .
Proof. A close examination of proof of Lemma 56 yields that (128), (129) imply
N ([c′i,j ]1≤i≤(d0 m0 +k0 +1)n0 −1,1≤j≤d ) ≥ |pm0 d0 +k0 |,
which gives us the claim.
Corollary 57. Let U ⊆ K m0 be a non-empty open set. Let fj : U → K be analytic functions, and
F := (f1 , . . . , fd ). Suppose 1, f1 , . . . , fd are linearly independent. Then for any l ≫F 1 we have
P
P
l
l
ΘF (l) d
O .
ΘF (1) F (p O) −
ΘF (1) F (p O) ⊇ p
Proof. By Lemma 56 there is a polynomial curve r such that 1, f1 ◦ r, . . . , fd ◦ r are linearly independent and
defined on a neighborhood of 0 ∈ K. Hence by Lemma 54 we are done.
Corollary 57′ . Let F := (f1 , . . . , fd0 ) : On0 → Od0 be an analytic function of the form
X
F (x) =
(ci,1 xi , . . . , ci,d0 xi ),
i
where i = (i1 , . . . , in0 ) ranges over multi-indexes non-negative integers. Suppose
(1) |ci,j | ≤ 1 for any i = (i1 , . . . , im ) and any 1 ≤ j ≤ d0 ,
(2) | det(cii ,j )| = |pk0 | for some multi-indexes i1 , . . . , id0 whose coordinates add up to a number at most
m0 and some positive integer k0 .
SUPER-APPROXIMATION, II:
THE p-ADIC AND BOUNDED POWER OF SQUARE-FREE INTEGERS CASES.
Then for l ≫m0 ,n0 ,d0 ,k0 1 we have
P
Θ(1)
F (pl O) −
P
Θ(1)
53
F (pl O) ⊇ pΘ(l) Od0 ,
where all the implied constants depend on m0 , n0 , d0 , k0 .
Proof. This is direct corollary of Lemma 55′ and Lemma 56′ .
P
Proof of Proposition 49. Suppose {fi1 , . . . , fid } is a basis of
i Kfi . So 1, fi1 , . . . , fid are linearly independent. And moreover there is an injective linear map L : K d → K d0 such that F = L ◦ F , where
F (x) = (fi1 (x), . . . , fid (x)). By Corollary 57 we have that
P
P
l
l
ΘF (l) d
O .
Θ (1) F (p O) −
Θ (1) F (p O) ⊇ p
F
F
for any l ≫F 1. Applying L to the both sides and using the linearity of L we have
P
P
l
ΘF (l)
l
L(Od ) ⊇ pΘF (l) Od0 ∩ Im(L).
ΘF (1) F (p O) ⊇ p
ΘF (1) F (p O) −
Proof of Proposition 49′ . By a similar argument as in the proof of Proposition 49, Corollary 57′ implies the
desired result.
Proof of Corollary 50. Let Oκ be the ring of integers of κ. There is a ∈ Oκ such that for any i
Suppose f j1 , . . . , f jd′
0
f i (x) := fi (ax) ∈ Oκ [x1 , . . . , xn0 ].
Pd0
′
is a basis of i=1 κf i . So there is a linear embedding L : κd0 → κd0 such that
(f 1 (x), . . . , f d0 (x)) = L(f j1 , . . . , f jd′ ).
0
So the operator norm of the induced linear embedding Lp :
and zero.
d′
κ p0
→
κdp0
are uniformly bounded from infinity
Since f j1 , . . . , f jd′ are linearly independent, there are multi-indexes i1 , . . . , id′0 such that
0
where f j (x) =
P
0 6= det(cii ,ju ) ∈ Oκ ,
i
i ci,j x
. So | det(cii ,ju )|p = 1 for almost all p ∈ Vf (κ).
By the above discussion, one can easily finish the proof using Proposition 49′ .
References
[BY91]
C. Bernstein, A. Yger, Effective Bézout identities in Q[z1 , . . . , zn ], Acta Math. 166 (1991) 69–120.
[BG08-a] J. Bourgain, A. Gamburd, Uniform expansion bounds for Cayley graphs of SL2 (fp ), Annals of Mathematics 167
(2008) 625–642.
[BG08-b] J. Bourgain, A. Gamburd, Expansion and random walks in SLd (Z/pn Z):I, Journal of the European Mathematical
Society 10 (2008) 987–1011.
[BG09]
J. Bourgain, A. Gamburd, Expansion and random walks in SLd (Z/pn Z):II. With an appendix by J. Bourgain, Journal
of the European Mathematical Society 11, no. 5., (2009) 1057–1103.
[BGS10] J. Bourgain, A. Gamburd, P. Sarnak, Affine linear sieve, expanders, and sum-product, Inventiones mathematicae
179, no. 3., (2010) 559–644.
[BV12]
J. Bourgain, P. Varjú, Expansion in SLd (Z/qZ), q arbitrary, Inventiones mathematicae 188, no 1, (2012) 151–173.
[BO14]
E. Breuillard, H. Oh (editors), Thin groups and superstrong approximation, MSRI Publ. 61, Cambridge Univ. Press.,
Cambridge 2014.
[BS91]
M. Burger and P. Sarnak, Ramanujan duals II, Inventiones mathematicae 106, no. 1, (1991) 1–11.
[Cad]
A. Cadoret, An open adelic image theorem for abelian schemes, to appear in International Mathematics Research
Notices.
[CT12]
A. Cadoret, A. Tamagawa, A uniform open image theorem for l-adic representations, I, Duke Mathematical Journal
161 (2012) 2605–2634.
54
[CT13]
[CFW81]
[Din06]
[EHK12]
[EMO05]
[Fal91]
[Fer94]
[GS04]
[Gow08]
[Hel05]
[Hel11]
[HLW06]
[How77]
[Hui12]
[HL]
[Ioa14-a]
[Ioa14-b]
[Kes59]
[Kne64]
[LS74]
[Lub94]
[LM13]
[LW93]
[LZ03]
[MW]
[Maz00]
[Mc95]
[NP11]
[Nor89]
[OW80]
[Pin98]
[SG05]
[SG]
[SGS13]
[SGV12]
[SGZ]
[SX91]
ALIREZA SALEHI GOLSEFIDY
A. Cadoret, A. Tamagawa, A uniform open image theorem for l-adic representations, II, Duke Mathematical Journal
162 (2013) 2301–2344.
A. Connes, J. Feldman, B.Weiss, An amenable equivalence relations is generated by a single transformation, Ergodic
Theory and Dynamical Systems 1 (1981) 431–450.
O. Dinai, Poly-log diameter bounds for some families of finite groups, Proceeding of the American Mathematical
Society 134, no 11, (2006) 3137–3142.
J. Ellenberg, C. Hall, E. Kowalski, Expander graphs, gonality, and variation of Galois representations, Duke Mathematical Journal 161, no.7, (2012) 1233–1275.
A. Eskin, S. Mozes, H. Oh, On uniform exponential growth for linear groups, Inventiones mathematicae 160, no. 1,
(2005) 1–30.
G. Faltings, Diophantine approximation on abelian varieties, Annals of Mathematics (2) 133 (1991) 549–576.
G. Frey, Curves with infinitely many points of fixed degree, Israel J. Math. 85 (1994) 79–85.
A. Gamburd, M. Shahshahani, Uniform diameter bounds for some families of Cayley graphs, International Mathematics Research Notices 71 (2004) 3813–3824.
W.T. Gowers, Quasirandom Groups, Combinatorics, Probability and Computing 17 (2008) 363–387.
H. Helfgott, Growth and generation in SL2 (Z/pZ), Annals of Mathematics 167 (2008) 601–623.
H. Helfgott, Growth in SL3 (Z/pZ), Journal of the European Mathematical Society 13, no. 3, (2011) 761–851.
S. Hoory, N. Linial, A. Widgerson, Expander graphs and their application, Bull. American Mathematical Society 43,
no. 4, (2006) 439–561.
R. Howe, Kirillov theory for compact p-adic groups, Pacific Journal of Mathematics 73, no. 2, (1977) 365–381.
C.-Y. Hui, Specialization of monodromy and ℓ-independence, C. R. Acad. Sci. Paris, Ser. I 350 (2012) 5–7.
C.-Y. Hui, M. Larsen, Adelic openness without the Mumford-Tate conjecture, preprint. Available online:
http://arxiv.org/pdf/1312.3812v2.pdf
A. Ioana, Orbit equivalence and Borel reducibility rigidity for profinite actions with spectral gap, to appear in Journal
of the European Mathematical Society. Available online: http://arxiv.org/pdf/1309.3026v4.pdf
A. Ioana, Strong ergodicity, property (T), and orbit equivalence rigidity for translation actions, to appear in J. Reine
Angew. Math. Available online: http://arxiv.org/pdf/1406.6628.pdf
H. Kesten, Symmetric random walks on groups, Transactions of the American Mathematical Society 92 (1959)
336–354.
M. Kneser, Erzeugende und Relationen verallgemeinerter Einheitengruppen, J. Reine Angew. Math. 214-215 (1964)
345–349.
V. Landazuri, G. M. Seitz, On the minimal degrees of projective representations of the finite Chevalley groups,
Journal of Algebra 32 (1974) 418–443.
A. Lubotzky, Discrete Groups, Expanding Graphs and Invariant Measures, Birkhäuser, 1994.
A. Lubotzky, C. Meiri, Sieve methods in group theory I: powers in linear groups, Journal of the American Mathematical Society 25, no. 4, (2012) 1119–1148.
A. Lubotzky, B. Weiss, Groups and expanders, In Expanding graphs (Princeton, NJ, 1992) DIMACS Series 10,
Discrete Mathematics and Theoretical Computer Science 95–109. American Mathematical Society, Providence, RI,
1993.
A. Lubotzky, A. Żuk, Property(τ ), preliminary version, 2003. http://www.ma.huji.ac.il/~ alexlub/BOOKS/On%20property/On%20propert
D. W. Masser, G. Wüstholz, Fields of large transcendence degree generated by values of elliptic functions, Inventiones
mathematicae 72, no. 3 ,(1983) 407–464.
B. Mazur, Abelian varieties and the Mordell–Lang conjecture, 199-227; in D. Haskell, A. Pillay, C. Steinhorn (ed.),
Model theory, algebra, and geometry, MSRI Publ. 39, Cambridge Univ. Press., Cambridge, 2000.
M. McQuillan, Division points on semi-abelian varieties, Inventiones mathematicae 120 (1995) 143–159.
N. Nikolov, L. Pyber, Product decompositions of quasirandom groups and a Jordan type theorem, Journal of the
European Mathematical Society 13, no. 4, (2011) 1063–1077.
M. V. Nori, On subgroups of GLn (fp ), Inventiones mathematicae 88, no. 2, (1987) 257–275.
D. Ornstein, B.Weiss, Ergodic theory of amenable groups. I. The Rokhlin lemma., Bulletin of the American Mathematical Society (N.S.) 1 (1980) 161–164.
R. Pink, Compact subgroups of linear algebraic groups, Journal of Algebra 206 (1998) 438–504.
A. Salehi Golsefidy, Character degrees of p-groups and pro-p groups, Journal of Algebra 286 (2005) 476–491.
A. Salehi Golsefidy, Super-approximation, I: p-adic semisimple case, to appear in International Mathematics Research
Notices. https://doi.org/10.1093/imrn/rnw208
A. Salehi Golsefidy, P. Sarnak, Affine sieve, Journal of the American Mathematical Society 26, no. 4, (2013) 1085–
1105.
A. Salehi Golsefidy, P. Varjú, Expansions in perfect groups, Geometric And Functional Analysis 22, no. 6, (2012)
1832–1891.
A. Salehi Golsefidy, X. Zhang, Inducing super-approximation, preprint.
P. Sarnak and X. Xue, Bounds for multiplicities of automorphic representations, Duke Mathematical Journal 64,
no. 1, (1991) 207–227.
SUPER-APPROXIMATION, II:
[Sha09]
[Tao08]
[Tit79]
[Var12]
[Wei84]
THE p-ADIC AND BOUNDED POWER OF SQUARE-FREE INTEGERS CASES.
55
A. Shalev, Word maps, conjugacy classes, and a noncommutative Waring-type theorem, Annals of Mathematics (2)
170, no. 3, (2009) 1383–1416.
T. Tao, Product set estimates for non-commutative groups, Combinatorica 28, no. 5, (2008) 547–594.
J. Tits, Reductive groups over local fields, Proceedings of symposia in pure mathematics 33 (1979), part 1, 29–69.
P. Varjú, Expansion in SLd (OK /I), I square-free, Journal of the European Mathematical Society 14, no. 1, (2012)
273–305.
B. Weisfeiler, Strong approximation for Zariski-dense subgroups of semisimple algebraic groups, Annals of Mathematics 120 (1984) 271–315.
Mathematics Dept, University of California, San Diego, CA 92093-0112
E-mail address: [email protected]
| 4math.GR
|
Model Selection with Nonlinear Embedding for Unsupervised Domain Adaptation
Hemanth Venkateswara, Shayok Chakraborty, Troy McDaniel, Sethuraman Panchanathan
arXiv:1706.07527v1 [cs.AI] 23 Jun 2017
Center for Cognitive Ubiquitous Computing, Arizona State University, Tempe, AZ, USA
{hemanthv, shayok.chakraborty, troy.mcdaniel, panch}@asu.edu
Abstract
Domain adaptation deals with adapting classifiers trained on
data from a source distribution, to work effectively on data
from a target distribution. In this paper, we introduce the Nonlinear Embedding Transform (NET) for unsupervised domain
adaptation. The NET reduces cross-domain disparity through
nonlinear domain alignment. It also embeds the domainaligned data such that similar data points are clustered together. This results in enhanced classification. To determine
the parameters in the NET model (and in other unsupervised
domain adaptation models), we introduce a validation procedure by sampling source data points that are similar in distribution to the target data. We test the NET and the validation procedure using popular image datasets and compare the
classification results across competitive procedures for unsupervised domain adaptation.
Introduction
There are large volumes of unlabeled data available online,
owing to the exponential increase in the number of images
and videos uploaded online. It would be easy to obtain labeled data if trained classifiers could predict the labels for
unlabeled data. However, classifier models do not perform
well when applied to unlabeled data from different distributions, owing to domain-shift (Torralba and Efros 2011).
Domain adaptation deals with adapting classifiers trained
on data from a source distribution, to work effectively
on data from a target distribution (Pan and Yang 2010).
Some domain adaptation techniques assume the presence
of a few labels for the target data, to assist in training a domain adaptive classifier (Aytar and Zisserman 2011;
Duan, Tsang, and Xu 2012; Hoffman et al. 2013). However,
real world applications need not support labeled data in the
target domain and adaptation here is termed as unsupervised
domain adaptation.
Many of the unsupervised domain adaptation techniques
can be organized into linear and nonlinear procedures,
based on how the data is handled by the domain adaptation model. A linear domain adaptation model performs
linear transformations on the data to align the source
and target domains or, it trains an adaptive linear classifier for both the domains; for example a linear SVM
Copyright c 2017, Association for the Advancement of Artificial
Intelligence (www.aaai.org). All rights reserved.
Figure 1:
(Best viewed in color) Two-moon binary classification problem with
source data in blue and target data in red. We assume the target labels are unknown.
(a) Original data, (b) KPCA aligns the data along nonlinear directions of maximum
variance, (c) MMD aligns the two domains, (d) MMD+Similarity-based Embedding
aligns the domains and clusters the data to ensure easy classification.
(Bruzzone and Marconcini 2010). Nonlinear techniques are
deployed in situations where the source and target domains
cannot be aligned using linear transformations. These techniques apply nonlinear transformations on the source and
target data in order to align them. For example, Maximum
Mean Discrepancy (MMD) is applied to learn nonlinear
representations, where the difference between the source
and target distributions is minimized (Pan et al. 2011). Even
though nonlinear transformations may align the domains,
the resulting data may not be conducive to classification.
If, after domain alignment, the data were to be clustered
based on similarity, it can lead to effective classification.
We demonstrate this intuition through a binary classification
problem using a toy dataset. Figure (1a), displays the source
and target domains of a two-moon dataset. Figure (1b), depicts the transformed data after KPCA (nonlinear projection). In trying to project the data onto a common ‘subspace’, the source data gets dispersed. Figure (1c), presents
the data after domain alignment using Maximum Mean Discrepancy (MMD). Although the domains are now aligned, it
does not necessarily ensure enhanced classification. Figure
(1d), shows the data after MMD and similarity-based em-
bedding, where data is clustered based on class label similarity. Cross-domain alignment along with similarity-based
embedding, makes the data classification friendly.
In this work, we the present the Nonlinear Embedding
Transform (NET) procedure for unsupervised domain adaptation. The NET performs a nonlinear transformation to
align the source and target domains and also cluster the data
based on label-similarity. The NET algorithm is a spectral
(eigen) technique that requires certain parameters (like number of eigen bases, etc.) to be pre-determined. These parameters are often given random values which need not be optimal (Pan et al. 2011; Long et al. 2013; Long et al. 2014). In
this work, we also outline a validation procedure to finetune model parameters with a validation set created from
the source data. In the following, we outline the two main
contributions in our work:
• Nonlinear embedding transform (NET) algorithm for unsupervised domain adaptation.
• Validation procedure to estimate optimal parameters for
an unsupervised domain adaptation algorithm.
We evaluate the validation procedure and the NET algorithm
using 7 popular domain adaptation image datasets, including
object, face, facial expression and digit recognition datasets.
We conduct 50 different domain adaptation experiments to
compare the proposed techniques with existing competitive
procedures for unsupervised domain adaptation.
Related Work
For the purpose of this paper, we discuss the relevant literature under the categories linear domain adaptation methods and nonlinear domain adaptation methods. A detailed
survey on transfer learning procedures can be found in
(Pan and Yang 2010). A survey of domain adaptation techniques for vision data is provided by (Patel et al. 2015).
The
Domain
Adaptive
SVM
(DASVM)
(Bruzzone and Marconcini 2010), is an unsupervised
method that iteratively adapts a linear SVM from the source
to the target. In recent years, the popular unsupervised linear
domain adaptation procedures are Subspace Alignment
(SA) (Fernando et al. 2013), and the Correlation Alignment
(CA) (Sun, Feng, and Saenko 2015). The SA algorithm
determines a linear transformation to project the source and
target to a common subspace, where the domain disparity
is minimized. The CA is an interesting technique which
argues that aligning the correlation matrices of the source
and target data is sufficient to reduce domain disparity. Both
the SA and CA are linear procedures, whereas the NET is a
nonlinear method.
Although deep learning procedures are inherently highly
nonlinear, we limit the scope of our work to nonlinear transformation of data that usually involves a positive semidefinite kernel function. Such procedures are closely related to the NET. However, in our experiments, we do
study the NET with deep features also. The Geodesic
Flow Kernel (GFK) (Gong et al. 2012), is a popular domain adaptation method, where the subspace spanning the
source data is gradually transformed into the target subspace along a path on the Grassmann manifold of subspaces.
Spectral procedures like the Transfer Component Analysis (TCA) (Pan et al. 2011), the Joint Distribution Alignment (JDA) (Long et al. 2013) and Transfer Joint Matching
(TJM) (Long et al. 2014), are the most closely related techniques to the NET. All of these procedures involve a solution to a generalized eigen-value problem in order to determine a projection matrix to nonlinearly align the source and
target data. In these spectral methods, domain alignment is
implemented using variants of MMD, which was first introduced in the TCA procedure. JDA introduces joint distribution alignment which is an improvement over TCA that
only incorporates marginal distribution alignment. The TJM
performs domain alignment along with instance selection by
sampling only relevant source data points. In addition to domain alignment with MMD, the NET algorithm implements
similarity-based embedding for enhanced classification. We
also introduce a validation procedure to estimate the model
parameters for unsupervised domain adaptation approaches.
Domain Adaptation With Nonlinear
Embedding
In this section, we first outline the NET algorithm for unsupervised domain adaptation. We then describe a crossvalidation procedure that is used to estimate the model parameters for the NET algorithm.
We begin with the problem definition where we consider
two domains; source domain S and target domain T . Let
s
⊂ S be a subset of the source doDs = {(xsi , yis )}ni=1
t
⊂ T be the subset of the
main and Dt = {(xti , yit )}ni=1
s
target domain. Let XS = [x1 , . . . , xsns ] ∈ Rd×ns and XT =
[xt1 , . . . , xtnt ] ∈ Rd×nt be the source and target data points
respectively. Let YS = [y1s , . . . , yns s ] and YT = [y1t , . . . , ynt t ]
be the source and target labels respectively. Here, xsi and xti
∈ Rd are data points and yis and yit ∈ {1, . . . , C} are the
associated labels. We define X := [x1 , . . . , xn ] = [XS , XT ],
where n = ns +nt . The problem of domain adaptation deals
with the situation where the joint distributions for the source
and target domains are different, i.e. PS (X, Y ) 6= PT (X, Y ),
where X and Y denote random variables for data points and
labels respectively. In the case of unsupervised domain adaptation, the labels YT are unknown. The goal of unsupervised
domain adaptation is to estimate the labels of the target data
ŶT = [ŷ1t , . . . , ŷnt t ] corresponding to XT using Ds and XT .
Nonlinear Domain Alignment
A common procedure to align two datasets is to first project
them to a common subspace. Kernel-PCA (KPCA) estimates a nonlinear basis for such a projection. In this
case, data is internally mapped to a high-dimensional
(possibly infinite-dimensional) space defined by Φ(X) =
[φ(x1 ), . . . , φ(xn )]. φ : Rd → H is the mapping function and
H is a RKHS (Reproducing Kernel Hilbert Space). The dot
product between the mapped vectors φ(x) and φ(y), is estimated by a positive semi-definite (psd) kernel, k(x, y) =
φ(x)⊤ φ(y). The dot product captures the similarity between
x and y. The kernel similarity (gram) matrix consisting of
similarities between all the data points in X, is given by,
K = Φ(X)⊤ Φ(X) ∈ Rn×n . The matrix K is used to deter-
The (n × n) adjacency matrix W, captures the similarity
relationships between datapoints, where,
mine the projection matrix A, by solving,
max tr(A⊤ KHK⊤ A).
A⊤ A=I
(1)
Wij :=
1
1
n
Here, H is the n × n centering matrix given by H = I − ,
where I is an identity matrix and 1 is a n × n matrix of
1s. A ∈ Rn×k , is the matrix of coefficients and the nonlinear projected data is given by Z = [z1 , . . . , zn ] = A⊤ K ∈
Rk×n . Along with projecting the source and target data to
a common subspace, the domain-disparity between the two
datasets must also be reduced. We employ the Maximum
Mean Discrepancy (MMD) (Gretton et al. 2009), which is
a standard nonparametric measure to estimate domain disparity. We adopt the Joint Distribution Adaptation (JDA)
(Long et al. 2013), algorithm which seeks to align both the
the marginal and conditional probability distributions of the
projected data. The marginal distributions are aligned by estimating the coefficient matrix A, which minimizes:
min
A
ns
1
1 X ⊤
A ki −
ns i=1
nt
n
X
2
A⊤ kj
j=ns +1
= tr(A⊤ KM0 K⊤ A).
H
(2)
M0 , is the MMD matrix which given by,
(M0 )ij
1
ns ns , xi , xj ∈ Ds
= nt1nt , xi , xj ∈ Dt
−1 , otherwise,
ns nt
(3)
(Mc )ij
(4)
Here, Ds and Dt are the sets of source and target data points
(c)
respectively. Ds is the subset of source data points with
(c)
(c)
(c)
class label c and ns = |Ds |. Similarly, Dt is the subset
(c)
(c)
of target data points with class label c and nt = |Dt |.
Since the target labels being unknown, we use predicted tar(c)
get labels to determine Dt . We initialize the target labels
using a classifier trained on the source data and refine the labels over iterations. Combining both the marginal and conditional distribution terms leads us to the JDA model, which
is given by,
min
A
C
X
tr(A⊤ KMc K⊤ A).
1
yis = yjs or i = j
s
0 yi 6= yjs or labels unknown.
(6)
To ensure that the projected data is clustered based on data
similarity, we minimize the sum of squared distances between data points weighted by the adjacency matrix. This
can be expressed as a minimization problem,
min
Z
1 X zi
zj
√ −p
2 ij
di
dj
2
Wij = min tr(A⊤ KLK⊤ A).
A
(7)
P
P
Here, di = k Wik and dj = k Wjk . They form
√ the
diagonal
entries
of
D,
the
(n×n)
diagonal
matrix.
||z
/
di −
i
p
zj / dj ||2 , is the squared normalized distance between the
projected data points zi and zj , which get clustered together
when Wij = 1, (as they belong to the same category). The
normalized distance is a more robust clustering measure as
compared to the standard Euclidean distance ||zi − zj ||2 ,
(Chung 1997). Substituting Z = A⊤ K, yields the trace term,
where L, denotes the symmetric positive semi-definite graph
laplacian matrix with L := I − D−1/2 WD−1/2 , and I is an
identity matrix.
Optimization Problem
Likewise, the conditional distribution difference can also be
minimized by introducing matrices Mc , with c = 1, . . . , C,
defined as,
(c)
1
xi , xj ∈ Ds
(c) (c) ,
ns ns
(c)
1
n(c) n(c) , xi , xj ∈ Dt
t
t
(
(c)
(c)
=
xi ∈ Ds , xj ∈ Dt
−1
,
(c)
(c)
ns(c) nt(c)
xj ∈ Ds , xi ∈ Dt
0,
otherwise.
(
(5)
c=0
To arrive at the optimization problem, we consider the nonlinear projection in Equation (1), the joint distribution alignment in Equation (5) and the similarity based embedding
in Equation (7). Maximizing Equation (1) and minimizing
Equations (5) and (7) can also be achieved by maintaining
Equation (1) constant and minimizing Equations (5) and (7).
Minimizing the similarity embedding in Equation (7) can result in the projected vectors being embedded in a low dimensional subspace. In order to maintain the subspace dimensionality, we introduce a new constraint in place of Equation
(1). The optimization problem for NET is obtained by minimizing Equations (5) and (7). The goal is to estimate the
(n × k) projection matrix, A. Along with regularization and
the dimensionality constraint, we get,
min
α.tr(A⊤ K
A⊤ KDK⊤ A=I
Mc K⊤ A)
c=0
+ β.tr(A⊤ KLK⊤ A) + γ||A||2F .
(8)
The first term controls the domain alignment and is
weighted by α. The second term ensures similarity based
embedding and is weighted by β. The third term is the regularization (Frobenius norm) that ensures a smooth projection
matrix A and it is weighted by γ. The constraint on A (in
place of A⊤ KHK⊤ A = I), prevents the projection from
collapsing onto a subspace with dimensionality less than k,
(Belkin and Niyogi 2003). We solve Equation (8) by forming the Lagrangian,
Similarity Based Embedding
In addition to domain alignment, the NET algorithm ensures
that the projected data Z, is classification friendly (easily
classifiable). To this end we introduce laplacian eigenmaps
in order to cluster datapoints based on class label similarity.
C
X
L(A, Λ) =α.tr A⊤ K
C
X
c=0
Mc K⊤ A + β.tr(A⊤ KLK⊤ A)
+ γ||A||2F + tr((I − A⊤ KDK⊤ A)Λ),
(9)
Algorithm 1 Nonlinear Embedding Transform
target data. The model is tested on the validation set to estimate parameters yielding highest classification accuracies.
Require: X, YS , constants α, β, regularization γ and projection dimension k.
Ensure: Projection matrix A, projected data Z.
1: Compute kernel matrix K, for predefined kernel k(., .)
2: Define the adjacency matrix W (Eq. (6))
P
3: Compute D = diag(d1 , . . . , dn ), where di = j Wij
4: Compute normalized graph laplacian L = I − D−1/2 WD−1/2
5: Solve Eq (10) and select k smallest eigen-vectors as columns of A
6: Estimate Z ← A⊤ K
7: Train a classifier with modified data {[z1 , . . . , zns ], YS }
Experiments
In this section, we evaluate the NET algorithm and the model
selection proposition across multiple image classification
datasets and several existing procedures for unsupervised
domain adaptation.
Datasets
where the Lagrangian constants are represented by the diagonal matrix Λ = diag(λ1 , . . . , λk ). Setting the derivative
∂L
∂A = 0, yields the generalized eigen-value problem,
αK
C
X
c=0
Mc K⊤ + βKLK⊤ + γI A = KDK⊤ AΛ. (10)
The solution A in (8) are the k-smallest eigen-vectors of
Equation (10). The projected data points are then given by
Z = A⊤ K. The NET algorithm is outlined in Algorithm 1.
Model Selection
In unsupervised domain adaptation the target labels are
treated as unknown. Current domain adaptation methods
that need to validate the optimum parameters for their
models, inherently assume the availability of target labels (Long et al. 2013), (Long et al. 2014). However, in the
case of real world applications, when target labels are
not available, it is difficult to verify if the model parameters are optimal. In the case of the NET model,
we have 4 parameters (α, β, γ, k), that we want to predetermine. We introduce a technique using Kernel Mean
Matching (KMM) to sample the source data to create a
validation set. KMM has been used to weight source data
points in order to reduce the distribution difference between the source and target data (Fernando et al. 2013),
(Gong, Grauman, and Sha 2013). Source data points with
large weights have a similar marginal distribution to the target data. These data points are chosen to form the validation set. The KMM estimates the weights wi , i = 1, . . . , ns ,
P s
P t
2
t
by minimizing n1s ni=1
wi φ(xsi ) − n1t n
j=1 φ(xj ) H . In
P
nt
order to simplify, we define κi := nnst j=1
k(xsi , xtj ),
s
s
i = 1, . . . , ns and KSij = k(xi , xj ). The minimization is
then represented as a quadratic programming problem,
1
min = w⊤ KS w − κ⊤ w,
w
2
ns
X
s.t. wi ∈ [0, B],
wi − ns ≤ ns ǫ.
(11)
i=1
The first constraint limits the scope of discrepancy between source and target distributions, with B → 1, leading to an unweighted solution. The second constraint ensures the measure w(x)PS (x), is a probability distribution
(Gretton et al. 2009). In our experiments, we select 10% of
the source data with the largest weights to create the validation set. We fine tune the values of (α, β, γ, k), using the validation set. For fixed values of (α, β, γ, k), the NET model is
trained using the source data (without the validation set) and
We conduct our experiments across 7 different datasets.
Their characteristics are outlined in Table (1).
MNIST-USPS datasets: These are popular handwritten
digit recognition datasets. Here, the digit images are subsampled to 16 × 16 pixels. Based on (Long et al. 2014), we
consider two domains MNIST (2,000 images from MNIST)
and USPS (1,800 images from USPS).
CKPlus-MMI datasets: The CKPlus (Lucey et al. 2010),
and MMI (Pantic et al. 2005) are popular Facial Expression recognition datasets. They contain videos of facial
expressions. We choose 6 categories of facial expression,
viz., anger, disgust, fear, happy, sad, surprise. We create two domains, CKPlus and MMI, by selecting video
frames with the most intense expressions. We use a pretrained deep convolutional neural network (CNN), to extract features from these images. In our experiments,
we use the VGG-F model (Chatfield et al. 2014), trained
on the popular ImageNet object recognition dataset. The
VGG-F network is similar in architecture to the popular
AlexNet (Krizhevsky, Sutskever, and Hinton 2012). We extract the 4096-dimensional features that are fed into the
fully-connected f c8 layer. We apply PCA on the combined
source and target data to reduce the dimension to 500 and
use these features across all the experiments.
COIL20 dataset: It is an object recognition dataset consisting of 20 categories with two domains, COIL1 and COIL2.
The domains consist of images of objects captured from
views that are 5 degrees apart. The images are 32 × 32 pixels
with gray scale values (Long et al. 2013).
PIE dataset: The “Pose, Illumination and Expression” (PIE)
dataset consists of face images ( 32 × 32 pixels) of 68 individuals. The images were captured with different head-pose,
illumination and expression. Similar to (Long et al. 2013),
we select 5 subsets with differing head-pose to create 5 domains, namely, P05 (C05, left pose), P07 (C07, upward
pose), P09 (C09, downward pose), P27 (C27, frontal pose)
and P29 (C29, right pose).
Office-Caltech dataset: This is currently the most popular benchmark dataset for object recognition in the domain adaptation computer vision community. The dataset
consists of images of everyday objects. It consists of 4
domains; Amazon, Dslr and Webcam from the Office
dataset and Caltech domain from the Caltech-256 dataset.
The Amazon domain has images downloaded from the
www.amazon.com website. The Dslr and Webcam domains have images captured using a DSLR camera and a webcam respectively. The Caltech domain is a subset of the
Caltech-256 dataset that was created by selecting categories
common with the Office dataset. The Office-Caltech dataset
has 10 categories of objects and a total of 2533 images (data
points). We experiment with two kinds of features for the
Office-Caltech dataset; (i) 800-dimensional SURF features
(Gong et al. 2012), (ii) Deep features. The deep features are
extracted using a pre-trained network similar to the CKPlusMMI datasets.
Table 1: Statistics for the benchmark datasets
Dataset
Type
#Samples
#Features
#Classes
Subsets
MNIST
USPS
CKPlus
MMI
COIL20
PIE
Ofc-Cal SURF
Ofc-Cal Deep
Digit
Digit
Face Exp.
Face Exp.
Object
Face
Object
Object
2,000
1,800
1,496
1,565
1,440
11,554
2,533
2,505
256
256
4096
4096
1,024
1,024
800
4096
10
10
6
6
20
68
10
10
MNIST
USPS
CKPlus
MMI
COIL1, COIL2
P05, ..., P29
A, C, W, D
A, C, W, D
Existing Baselines
We compare the NET algorithm with the following baseline and state-of-the-art methods. Like NET, the TCA, TJM
Table 2: Baseline methods that are compared with the NET.
Method
SA
CA
GFK
TCA
TJM
JDA
Reference
Subspace Alignment (Fernando et al. 2013)
Correlation Alignment (Sun, Feng, and Saenko 2015)
Geodesic Flow Kernel (Gong et al. 2012)
Transfer Component Analysis (Pan et al. 2011)
Transfer Joint Matching (Long et al. 2014)
Joint Distribution Adaptation (Long et al. 2013)
and JDA are all spectral methods. While all the four algorithms use MMD to align the source and target datasets, the
NET, in addition, uses nonlinear embedding for classification enhancement. TCA, TJM and JDA, solve for A in a
setting similar to Equation (10). However, unlike NET, they
do not have the similarity based embedding term and α = 1,
is fixed in all the three algorithms. Therefore, these models have only 2 free parameters (γ and k), that need to be
pre-determined in contrast to NET, which has 4 parameters,
(α, β, γ, k). Since TCA, TJM and JDA, are all quite similar to each other, for the sake of brevity, we evaluate model
selection (estimating optimal model parameters) using only
JDA and NET. The SA, CA and GFK algorithms, do not
have any critical free model parameters that need to be predetermined.
In our experiments, NETv is a special case of the NET,
where model parameters (α, β, γ, k), have been determined
using a validation set derived from Equation (11). Similarly,
JDAv is a special case of JDA, where (γ, k), have been determined using a validation set derived from Equation (11).
In order to ascertain the optimal nature of the parameters determined with a source-based validation set, we estimate the
best model parameters using the target data (with labels) as
a validation set. These results are represented by NET in the
figures and tables. The results for the rest of the algorithms
Table 3: Recognition accuracies (%) for domain adaptation
experiments on the digit and face datasets. {MNIST(M),
USPS(U), CKPlus(CK), MMI(MM), COIL1(C1) and
COIL2(C2). M→U implies M is source domain and U is target domain. The best and second best results in every experiment (row) are in bold and italic respectively. The shaded
columns indicate accuracies obtained with model selection.
Expt.
M→U
U→M
C1→C2
C2→C1
CK→MM
MM→CK
P05→P07
P05→P09
P05→P27
P05→P29
P07→P05
P07→P09
P07→P27
P07→P29
P09→P05
P09→P07
P09→P27
P09→P29
P27→P05
P27→P07
P27→P09
P27→P29
P29→P05
P29→P07
P29→P09
P29→P27
Average
SA
67.39
51.85
85.97
84.17
31.12
39.75
26.64
27.39
30.28
19.24
25.42
47.24
53.47
26.84
23.26
41.87
44.97
28.13
35.62
63.66
72.24
36.03
23.05
26.03
27.76
30.31
41.14
CA
59.33
50.80
84.72
82.78
31.89
37.74
40.33
41.97
55.36
29.04
41.51
53.43
63.77
35.72
35.47
47.08
53.71
34.68
51.17
66.05
73.96
40.50
26.89
31.74
31.92
34.70
47.55
GFK
66.06
47.40
85.00
84.72
28.75
37.94
26.21
27.27
31.15
17.59
25.27
47.37
54.22
27.02
21.88
43.09
46.38
26.84
34.27
62.92
73.35
37.38
20.35
24.62
28.49
31.27
40.65
TCA
60.17
39.85
90.14
88.33
32.72
31.33
40.76
41.79
59.60
29.29
41.78
51.47
64.73
33.70
34.69
47.70
56.23
33.09
55.61
67.83
75.86
40.26
27.01
29.90
29.90
33.67
47.59
TJM
64.94
52.80
91.67
89.86
30.35
40.62
10.80
7.29
15.14
4.72
16.63
21.69
26.04
10.36
14.98
27.26
27.55
8.15
25.96
28.73
38.36
7.97
9.54
8.41
6.68
10.06
26.79
JDA
67.28
59.65
92.64
93.75
29.78
28.39
58.81
54.23
84.50
49.75
57.62
62.93
75.82
39.89
50.96
57.95
68.45
39.95
80.58
82.63
87.25
54.66
46.46
42.05
53.31
57.01
60.63
JDAv
71.94
59.65
95.28
93.89
25.82
26.79
77.53
66.42
90.78
52.70
74.70
79.66
81.14
63.73
77.16
78.39
84.92
65.93
92.83
90.18
90.14
72.18
60.20
71.39
74.02
76.66
72.85
NET
75.39
62.60
93.89
92.64
29.97
45.83
77.84
70.96
91.86
52.08
74.55
77.08
83.84
69.24
73.98
79.01
83.48
70.04
93.07
89.99
89.71
76.84
67.32
70.23
74.63
75.43
74.67
NETv
72.72
61.35
90.42
88.61
30.54
40.08
69.00
57.41
84.68
45.40
57.92
54.60
86.09
47.30
68.67
67.34
87.47
67.65
92.44
93.68
90.20
79.53
52.67
57.52
62.81
80.98
68.73
(SA, CA, GFK, TCA, TJM and JDA), are obtained with the
parameter settings described in their respective works.
Experimental Details
For fair comparison with existing methods, we follow
the same experimental protocol as in (Gong et al. 2012;
Long et al. 2014). We conduct 50 different domain adaptation experiments with the previously mentioned datasets.
Each of these is an unsupervised domain adaptation experiment with one source domain (data points and labels)
and one target domain (data points only). When estimating
Mc , we choose 10 iterations to converge to the predicted
test/validation labels. Wherever necessary, we use a Gaussian kernel for k(., .), with a standard width equal to the median of the squared distances over the dataset. We train a 1Nearest Neighbor (NN) classifier using the projected source
data and test on the projected target data for all the experiments. We choose a NN classifier as in (Gong et al. 2012;
Long et al. 2014), since it does not require tuning of crossvalidation parameters. The accuracies reflect the percentage
of correctly classified target data points.
Parameter Estimation Study
Here we evaluate our model selection procedure. The NET
algorithm has 4 parameters (α, β, γ, k), and the JDA has
2 parameters (γ, k), that need to be pre-determined. To
determine these parameters, we weight the source data
points using Equation (11) and select 10% of the source
data points with the largest weights. These source data
points have a distribution similar to the target and they are
Table 4: Recognition accuracies (%) for domain adaptation experiments on the Office-Caltech dataset with SURF and Deep
features. {Amazon(A), Webcam(W), Dslr(D), Caltech(C)}. A→W implies A is source and W is target. The best and
second best results in every experiment (row) are in bold and italic respectively. The shaded columns indicate accuracies
obtained with model selection.
Expt.
C→A
D→A
W→A
A→C
D→C
W→C
A→D
C→D
W→D
A→W
C→W
D→W
Average
SA
43.11
29.65
32.36
38.56
31.88
29.92
37.58
43.95
90.45
37.29
36.27
87.80
44.90
CA
36.33
28.39
31.42
33.84
29.56
28.76
36.94
38.22
85.35
31.19
29.49
83.39
41.07
(a) # bases k
GFK
45.72
26.10
27.77
39.27
30.45
28.41
34.40
43.31
82.17
41.70
35.59
79.66
42.88
SURF Features
TCA TJM JDA
44.47 46.76 44.78
31.63 32.78 33.09
29.44 29.96 32.78
39.89 39.45 39.36
30.99 31.43 31.52
32.15 30.19 31.17
33.76 45.22 39.49
36.94 44.59 45.22
85.35 89.17 89.17
33.90 42.03 37.97
32.88 38.98 41.69
85.42 85.42 89.49
43.07 46.33 46.31
JDAv
45.41
29.85
29.33
39.27
31.08
31.43
31.85
40.13
88.53
38.98
37.97
86.78
44.22
NET
46.45
39.67
41.65
43.54
35.71
35.89
40.76
45.86
89.81
44.41
44.41
87.80
49.66
NETv
46.24
35.60
39.46
43.10
34.11
32.77
36.31
36.31
91.72
35.25
33.56
90.51
46.24
(b) MMD weight α
SA
88.82
84.33
84.01
80.55
76.26
78.90
82.17
80.89
100.00
82.37
77.29
98.98
84.55
CA
91.12
86.63
82.76
82.47
75.98
74.98
87.90
82.80
100.00
80.34
79.32
99.32
85.30
GFK
90.60
88.40
88.61
81.01
78.63
76.80
82.80
77.07
100.00
84.41
78.64
98.31
85.44
Deep Features
TCA
TJM
JDA
89.13
91.01
90.07
88.19
88.72
91.22
86.21
88.09
91.43
75.53
78.08
83.01
74.43
76.07
80.09
76.71
79.18
82.74
82.17
87.26
89.81
75.80
82.80
89.17
100.00 100.00 100.00
76.61
87.12
87.12
78.31
88.48
85.76
97.97
98.31
98.98
83.42
87.09
89.12
(c) Embed weight β
JDAv
89.34
90.18
87.04
78.27
78.17
78.90
77.07
80.25
100.00
79.32
77.97
98.98
84.63
NET
92.48
91.54
92.58
83.01
82.10
82.56
91.08
92.36
99.36
90.85
90.85
99.66
90.70
NETv
90.70
91.43
91.95
82.28
83.38
82.28
80.89
90.45
100.00
87.46
84.07
99.66
88.71
(d) Regularization γ
Figure 2: NET Validation Study. Each figure depicts the accuracies over the source-based validation set. When studying a
parameter (say k), the remaining parameters (α, β, γ) are fixed at the optimum value. The legend is, Digit (Di), Coil (Cl),
MMI&CK+ Face (Fc), PIE (Pi), Office-Caltech SURF (O-S) and Office-Caltech Deep (O-D).
used as a validation set to determine the optimal values for
the model parameters (α, β, γ, k). The parameter space consists of k ∈ {10, 20, . . . , 100, 200} and α, β, γ from the set
{0, 0.0001, 0.0005, 0.001, 0.005, 0.01, 0.05, 0.1, 0.5, 1, 5, 10}.
For the sake of brevity, we present one set of parameters
for every dataset, although in practice, a unique set of
parameters can be evaluated for every domain adaptation
experiment. Given a set of model parameters, we conduct
the domain adaptation experiment using the entire source
data (data and labels) and the target data (data only). The
accuracies obtained are represented as shaded columns
JDAv and NETv in Tables (3) and (4).
In order to evaluate the validity of our proposed model selection method, we also determine the parameters using the
target data as a validation set. These results are represented
by the NET column in Tables (3) and (4). Since the NET
column values have been determined using the target data,
they can be considered as the best accuracies for the NET
model. The rest of the column values SA, CA, GFK, TCA,
TJM and JDA, were estimated with model parameters suggested in their respective papers. The recognition accuracies
for NETv is greater than that of the other domain adaptation
methods and is nearly comparable to the NET. In Table (3),
the JDAv has better performance than the JDA. This goes
to show that a proper validation procedure does help select
the best set of model parameters. It demonstrates that the
proposed model selection procedure is a valid technique for
evaluating an unsupervised domain adaptation algorithm in
the absence of target data labels. Figures (2) and (3), depict the variation of average validation set accuracies over
the model parameters. Based on these curves, the optimal
parameters are chosen for each of the datasets.
NET Algorithm Evaluation
The NET algorithm has been compared to existing unsupervised domain adaptation procedures across multiple
datasets. The results of the NET algorithm are depicted under the NET column in Tables (3) and (4). The parameters
used to obtain these results are depicted in Table (5). The accuracies obtained with the NET algorithm are consistently
better than any of the other spectral methods (TCA, TJM
and JDA). NET also consistently performs better compared
to non-spectral methods like SA, CA and GFK.
Discussion and Conclusions
The average accuracies obtained with JDA and NET using
the validation set are comparable to the best accuracies with
JDA and NET. This empirically validates the model selection proposition. However, there is no theoretical guarantee
that the parameters selected are the best. In the absence of
(a) # bases k
(b) Regularization γ
Figure 3: JDA Validation Study. Each figure depicts the accuracies over the source-based validation set. When studying a parameter (say k), the remaining parameter γ is fixed
at the optimum value. The legend is, Digit (Di), Coil (Cl),
MMI&CK+ Face (Fc), PIE (Pi), Office-Caltech SURF (OS) and Office-Caltech Deep (O-D).
Table 5: Parameters used for the NET model.
Dataset
α
β
γ
k
MNIST & USPS
MMI & CK+
COIL
PIE
Ofc-SURF
Ofc-Deep
1.0
0.01
1.0
10.0
1.0
1.0
0.01
0.01
1.0
0.001
1.0
1.0
1.0
1.0
1.0
0.005
1.0
1.0
20
20
60
200
20
20
theoretical validation, further empirical analysis is advised
when using the proposed technique for model selection.
In this paper, we have proposed the Nonlinear Embedding Transform algorithm and a model selection procedure
for unsupervised domain adaptation. The NET performs favorably compared to competitive visual domain adaptation
methods across multiple datasets.
This material is based upon work supported by the National Science Foundation (NSF) under Grant No:1116360.
Any opinions, findings, and conclusions or recommendations expressed in this material are those of the authors and
do not necessarily reflect the views of the NSF.
References
[Aytar and Zisserman 2011] Aytar, Y., and Zisserman, A.
2011. Tabula rasa: Model transfer for object category detection. In Intl. Conference on Computer Vision.
[Belkin and Niyogi 2003] Belkin, M., and Niyogi, P. 2003.
Laplacian eigenmaps for dimensionality reduction and data
representation. Neural computation 15(6):1373–1396.
[Bruzzone and Marconcini 2010] Bruzzone, L., and Marconcini, M. 2010. Domain adaptation problems: A dasvm
classification technique and a circular validation strategy.
Pattern Analysis and Machine Intelligence, IEEE Trans. on
32(5):770–787.
[Chatfield et al. 2014] Chatfield, K.; Simonyan, K.; Vedaldi,
A.; and Zisserman, A. 2014. Return of the devil in the
details: Delving deep into convolutional nets. In British Machine Vision Conference.
[Chung 1997] Chung, F. R. 1997. Spectral graph theory,
volume 92. American Mathematical Soc.
[Duan, Tsang, and Xu 2012] Duan, L.; Tsang, I. W.; and Xu,
D. 2012. Domain transfer multiple kernel learning. Pattern Analysis and Machine Intelligence, IEEE Trans. on
34(3):465–479.
[Fernando et al. 2013] Fernando, B.; Habrard, A.; Sebban,
M.; and Tuytelaars, T. 2013. Unsupervised visual domain
adaptation using subspace alignment. In IEEE Conference
on Computer Vision and Pattern Recognition, 2960–2967.
[Gong et al. 2012] Gong, B.; Shi, Y.; Sha, F.; and Grauman,
K. 2012. Geodesic flow kernel for unsupervised domain
adaptation. In IEEE Conference on Computer Vision and
Pattern Recognition.
[Gong, Grauman, and Sha 2013] Gong, B.; Grauman, K.;
and Sha, F. 2013. Connecting the dots with landmarks:
Discriminatively learning domain-invariant features for unsupervised domain adaptation. In Intl. Conference on Machine learning, 222–230.
[Gretton et al. 2009] Gretton, A.; Smola, A.; Huang, J.;
Schmittfull, M.; Borgwardt, K.; and Schölkopf, B. 2009.
Covariate shift by kernel mean matching. Dataset shift in
machine learning 3(4):5.
[Hoffman et al. 2013] Hoffman, J.; Rodner, E.; Donahue, J.;
Saenko, K.; and Darrell, T. 2013. Efficient learning of
domain-invariant image representations. In Intl. Conference
on Learning Representations.
[Krizhevsky, Sutskever, and Hinton 2012] Krizhevsky, A.;
Sutskever, I.; and Hinton, G. E. 2012. Imagenet classification with deep convolutional neural networks. In Neural
Information Processing Systems, 1097–1105.
[Long et al. 2013] Long, M.; Wang, J.; Ding, G.; Sun, J.; and
Yu, P. S. 2013. Transfer feature learning with joint distribution adaptation. In Intl. Conference on Machine Learning,
2200–2207.
[Long et al. 2014] Long, M.; Wang, J.; Ding, G.; Sun, J.; and
Yu, P. 2014. Transfer joint matching for unsupervised domain adaptation. In IEEE Conference on Computer Vision
and Pattern Recognition, 1410–1417.
[Lucey et al. 2010] Lucey, P.; Cohn, J. F.; Kanade, T.;
Saragih, J.; Ambadar, Z.; and Matthews, I. 2010. The extended cohn-kanade dataset (ck+): A complete dataset for
action unit and emotion-specified expression. In IEEE Conference on Computer Vision and Pattern Recognition, 94–
101.
[Pan and Yang 2010] Pan, S. J., and Yang, Q. 2010. A survey on transfer learning. Knowledge and Data Engineering,
IEEE Trans. on 22(10):1345–1359.
[Pan et al. 2011] Pan, S. J.; Tsang, I. W.; Kwok, J. T.; and
Yang, Q. 2011. Domain adaptation via transfer component
analysis. Neural Networks, IEEE Trans. on 22(2):199–210.
[Pantic et al. 2005] Pantic, M.; Valstar, M.; Rademaker, R.;
and Maat, L. 2005. Web-based database for facial expression analysis. In IEEE Conference on Multimedia and Expo.
IEEE.
[Patel et al. 2015] Patel, V. M.; Gopalan, R.; Li, R.; and
Chellappa, R. 2015. Visual domain adaptation: A sur-
vey of recent advances. IEEE Signal Processing Magazine
32(3):53–69.
[Sun, Feng, and Saenko 2015] Sun, B.; Feng, J.; and Saenko,
K. 2015. Return of frustratingly easy domain adaptation. In
Intl. Conference on Computer Vision, TASK-CV.
[Torralba and Efros 2011] Torralba, A., and Efros, A. A.
2011. Unbiased look at dataset bias. In IEEE Conference
on Computer Vision and Pattern Recognition, 1521–1528.
| 2cs.AI
|
MODEL MISSPECIFICATION AND BIAS FOR INVERSE
PROBABILITY WEIGHTING AND DOUBLY ROBUST ESTIMATORS
INGEBORG WAERNBAUM1 AND LAURA PAZZAGLI2
Abstract. In the causal inference literature a class of semi-parametric estimators are
arXiv:1711.09388v1 [math.ST] 26 Nov 2017
called robust if the estimator has desirable properties under the assumption that at least
one of the working models is correctly specified. A standard example is a doubly robust
estimator that specifies parametric models both for the propensity score and the outcome
regression. When estimating a causal parameter in an observational study the role of
parametric models are often not to be true representations of the data generating process, instead the motivation for their use is to facilitate the adjustment for confounding,
for example by reducing the dimension of the covariate vector, making the assumption of
at least one true model unlikely. In this paper we propose a crude analytical approach to
study the large sample bias of estimators when all models are assumed to be approximations of the true data generating process, i.e., all models are misspecified. We apply our
approach to three prototypical estimators, two inverse probability weighting estimators
(IPW), using a misspecified propensity score model, and a doubly robust (DR) estimator, using misspecified models for the outcome regression and the propensity score. To
compare the consequences of the model misspecifications for the estimators we show conditions for when using normalized weights leads to a smaller bias compared to a simple
IPW estimator. To analyze the question of when the use of two misspecified models are
better than one we derive necessary and sufficient conditions for when the DR estimator
has a smaller bias than the simple IPW estimator and when it has a smaller bias than
the IPW estimator with normalized weights. For most conditions in the comparisons,
the covariance between the propensity score model error and the conditional outcomes
plays an important role. The results are illustrated in a simulation study.
Keywords: Average causal effects, comparing biases, propensity score, robustness
1. Introduction
Identifying an average causal effect of a treatment with observational data requires
adjustment for background variables that affect both the treatment and the outcome
under study. Often parametric models are assumed for parts of the joint distribution of
the treatment, outcome and background variables (covariates) and large sample properties
of estimators are derived under the assumption that the parametric models are correctly
specified.
A class of semiparametric estimators called inverse probability weighting (IPW) estimators use the difference of the weighted means of the outcomes for the treatment groups
as an estimator of the average causal effect, see e.g. Lunceford and Davidian (2004). IPW
2
INGEBORG WAERNBAUM1 AND LAURA PAZZAGLI2
estimators reweights the observed outcomes to a random sample of all potential outcomes,
missing and observed, by letting each observed outcome account for itself and other individuals with similar characteristics proportionally to the probability of their outcome
being observed. IPW estimators are used in applied literature (Kwon, Jeong, et al. 2015)
and their properties have been studied in the missing data and causal inference literature,
see e.g. Vansteelandt, Carpenter, et al. (2010) and Seaman and White (2013) for reviews.
Earlier research have considered properties of IPW estimators for estimating the average
causal effect under the assumption that a parametric propensity score (PS) model is
correctly specified (Lunceford and Davidian 2004; Yao, Sun, et al. 2010). Properties of
IPW estimators using different weights, often referred to as stabilized (Hernán, Brumback,
et al. 2000; Hernán and Robins 2006) or normalized (Hirano and Imbens 2001; Busso,
DiNardo, et al. 2014) have been discussed together with the impact of violations to
an assumption of overlap (positivity) (Khan and Tamer 2010; Petersen, Porter, Gruber,
Wang, and van der Laan 2010).
To decrease the reliance on the choice of a parametric model of the PS an approach
with doubly or multiply robust estimators have emerged. An estimator is referred to as a
doubly robust estimator (DR) (Bang and Robins 2005; Tsiatis 2007) since it is a consistent
estimator of the average causal effect if either the model for the propensity score or the
outcome regression (OR) model is correct (Scharfstein, Rotnitzky, et al. 1999). The
efficiency of the DR estimator is a key property and its variance has been described under
correct specification of at least one of the models (Cao, Tsiatis, et al. 2009). When both
models are correct the estimator reaches the semiparametric efficiency bound described
in Robins and Rotnitzky (1994). The large sample properties of IPW estimators with
standard, normalized and variance minimized weights, together with a prototypical DR
estimator were studied and compared in Lunceford and Davidian (2004) under correct
specification of the PS and OR models.
There are few studies on doubly or multiply robust estimators under misspecification
of both the PS and the OR models. Kang and Schafer (2007) studied and compared
the performance of various DR and non-DR estimators under misspecification of both
models. They concluded that many DR methods perform better than simple inverse
probability weighting. However a regression-based estimator under a misspecified model
was not improved upon. The paper was commented and the relevance of the results were
discussed by several authors see e.g., Tsiatis and Davidian (2007), Tan (2007), Robins,
Sued, et al. (2007). In Waernbaum (2012) a matching estimator was compared to IPW and
DR estimators under misspecification of both the PS and OR models. Here, a robustness
class for the matching estimator under misspecification of the PS model was described.
In this paper we describe three commonly used semi-parametric estimators of the average causal effect under the assumption that none of the working models are correctly
specified. For this purpose we study the difference between the probability limit of the
estimator under model misspecification and the true average causal effect. The purpose
MODEL MISSPECIFICATION AND BIAS
3
of this definition of the bias is that the estimators under study converges to a well-defined
limit however not necessarily consistent for the true average causal effect. We study the
biases of two IPW estimators and a DR estimator and compare them under the same
misspecification of the PS-model. In the comparisons with the DR estimator the biases
provide a means to describe when two wrong models are better than one. To analyze the
consequences of the model misspecifications we compare the absolute values of the biases
in two parts separately, one for each of the two potential outcome means (µ1 , µ0 ). For the
comparisons we provide sufficient and necessary conditions for inequalities involving the
absolute value of the biases of the different estimators. We use a running example of a
data generating process with misspecified PS and OR models to illustrate the inequalities.
A simulation study is performed to investigate the biases for finite samples. The data generating processes and the misspecified models from the simulation designs are also used
for numerical approximations of the large sample properties derived in the paper.
In recent studies strategies for bias reduction under model misspecification have been
proposed by inclusion of additional conditions in the estimating equations for both IPW
(Imai and Ratkovic 2014) and DR estimators (Vermeulen and Vansteelandt 2015). However, the general approach for analyzing model misspecification provided in this paper
could also be used to study the biased reduced estimators.
The paper proceeds as follows. Section 2 presents the model and theory together with
the estimators and their properties when the working models are correctly specified. Section 3 presents a general approach and assumptions to study model misspecification. In
Section 4 the generic biases are derived and comparisons between the estimators are performed. We present a simulation study in Section 5 containing both finite sample properties of the estimators and numerical large sample approximations and thereafter we
conclude with a discussion.
2. Model and theory
The potential outcome framework defines a causal effect as a comparison of potential
outcomes that would be observed under different treatments (Rubin 1974). Let X be a
vector of pre-treatment variables, referred to as covariates, T a binary treatment, with
realized value T = 1 if treated and T = 0 if control. The causal effect of the treatment is
defined as a contrast between two potential outcomes, for example the difference, Y (1) −
Y (0), where Y (1) is the potential outcome under treatment and Y (0) is the potential
outcome under the control treatment. The observed outcome Y is assumed to be the
potential outcomes for each level of the observed treatment Y = T Y (1) + (1 − T )Y (0),
so that the data vector that we observe is (Ti , Xi , Yi ), where i = 1, . . . , n are assumed
independent and identically distributed copies. In the remainder of the paper we will drop
the subscript i for the random variables when not needed. Since each individual only
can be subject to one treatment either Y (1) or Y (0) will be missing. If the treatment
is randomized the difference of sample averages of the treated and controls will be an
INGEBORG WAERNBAUM1 AND LAURA PAZZAGLI2
4
unbiased estimator of the average causal effect ∆ = E [Y (1) − Y (0)], the parameter of
interest. In the following we will use the notation µ1 = E [Y (1)] and µ0 = E [Y (0)].
When the treatment is not assigned at random the causal effect of the treatment can be
estimated if all confounders are observed
Assumption 1. [No unmeasured confounding]
Y (t) ⊥
⊥ T |X, t = 0, 1.
and if the treated and controls have overlapping covariate distributions
Assumption 2. [Overlap]
η < P (T = 1|X) < 1 − η, for some η > 0.
Throughout the paper we assume that Assumptions 1 and 2 hold. Under these assumptions
we can estimate the average causal effect with the observed data by marginalizing over
the conditional means
(1)
∆ = E [E(Y | X, T = 1) − E(Y | X, T = 0)] .
For matching/stratification estimators (see Imbens and Wooldridge 2009 for a review)
the inner expectation in (1) is evaluated by grouping treated and controls in matched
pairs or strata formed by the resulting cells of the cross classification of the covariates.
Instead of comparing treated and controls on a high dimensional vector of the covariates
it is sufficient to condition on a scalar function of the covariates, e(X) = P (T = 1|X),
called the propensity score (Rosenbaum and Rubin 1983). Instead of conditioning on the
propensity score as in (1), the propensity score can be used as a weight
TY
(1 − T )Y
∆=E
−
= E [Y (1) − Y (0)] ,
e(X)
1 − e(X)
where the last equality follows from Assumption 1.
Usually it is assumed that the propensity score and the outcome regression follow parametric models.
Assumption 3. [Propensity score model]
The propensity score e(X) follows a model e(X, β) parametrized by, β = (β1 , . . . , βp ) and
ê(X) is the estimated propensity score e(X, β̂) with a n1/2 -consistent estimator of β
Assumption 4. [Outcome regression model]
The conditional expectation, µt (X) = E(Y (t)|X), t = 0, 1 follows a model µt (X, αt ),
t = 0, 1 parametrized by αt = (αt1 , . . . , αtqt ) and µ̂t (X) is the estimated outcome regression
µt (X, α̂t ) with a n1/2 -consistent estimator of αt .
Consider the example, e(X, β) = [1+exp(−X 0 β)]−1 and ê(X) are the fitted values of the
propensity score when β̂ is a maximum likelihood estimator of β. Similarly the outcome
MODEL MISSPECIFICATION AND BIAS
5
regression model could be a linear model µt (X, αt ) = X 0 αt where µ̂t (X), t = 0, 1 are the
fitted values when α̂t is the ordinary least squares estimator.
We study two IPW estimators and a DR estimator described in Lunceford and Davidian
ˆ IPW an estimator defined by:
(2004). We denote by ∆
1
n
n
i=1
i=1
X Ti Yi
1 X (1 − Ti )Yi
ˆ IPW = 1
∆
−
.
1
n
ê(Xi ) n
1 − ê(Xi )
(2)
ˆ IPW is
The variance of ∆
1
2
σIPW
= VIPW1 − aT I −1 a
1
(3)
where VIPW1 is the asymptotic variance when the propensity score is known
Y (1)2
Y (0)2
VIPW1 = E
− (µ1 − µ0 )2 ,
+
e(X)
1 − e(X)
a is a (p × 1) vector
a=E
Y (1)
Y (0)
+
e(X) 1 − e(X)
e0 (X)
for the partial derivatives e0 (X) = ∂/∂β {e(X, β)} and I is the p × p covariance matrix
ˆ IPW each observed treated individual is weighted
of the estimated propensity score. In ∆
1
by 1/e(X) and each control is weighted by 1/ [1 − e(X)]. Since the weights generate the
missing potential outcomes for each of the treatment and control groups respectively we
want to divide the weighted sum with the number of individuals in the generated sample
consisting of the observed and missing potential outcomes which may not be n for a given
ˆ IPW with normalized
sample (Hirano, Imbens, et al. 2003). This gives an IPW estimator ∆
2
weights
(4)
ˆ IPW =
∆
2
n
X
i=1
Ti
ê(Xi )
!−1
n
X
Ti Yi
−
ê(Xi )
i=1
n
X
i=1
1 − Ti
1 − ê(Xi )
!−1
n
X
(1 − Ti )Yi
i=1
1 − ê(Xi )
and variance
2
σIPW
= VIPW2 − bT I −1 b
2
(5)
where VIPW2 is the asymptotic variance when the propensity score is known
"
#
(Y (1) − µ1 )2 (Y (0) − µ0 )2
VIPW2 = E
+
,
e(X)
1 − e(X)
and b is a (p × 1) vector
b=E
Y (1) − µ1 Y (0) − µ0
+
e(X)
1 − e(X)
e0 (X)
.
,
INGEBORG WAERNBAUM1 AND LAURA PAZZAGLI2
6
Under Assumptions 1-3 the IPW estimators are consistent estimators of the average
√ ˆ
2
causal effect ∆ with asymptotic distribution n(∆
IPWk − ∆) ∼ N (0, σ IPWk ), k = 1, 2.
In addition we study a DR estimator (Lunceford and Davidian 2004; Tsiatis 2007)
ˆ DR = 1
∆
n
−
(6)
1
n
n
X
Ti Yi − (Ti − ê(Xi ))µ̂1 (Xi )
i=1
n
X
i=1
ê(Xi )
(1 − Ti )Yi + (Ti − ê(Xi ))µ̂0 (Xi )
.
1 − ê(Xi )
Under Assumptions 1-4 we have the large sample distribution
√ ˆ
2 )
n(∆DR −∆) ∼ N (0, σDR
where
2
σDR
= VIPW2 − d,
(7)
and
" s
d =E
1 − e(X)
(µ1 (X) − µ1 ) +
e(X)
s
!#2
e(X)
(µ0 (X) − µ0 )
,
1 − e(X)
2
2 ≤ σ2
with the property that σDR
IPW1 , σIPW2 which was shown by the theory of Robins and
colleagues (Robins and Rotnitzky 1994).
3. Model misspecification: a general approach
Our interest lies in the behaviors of the estimators when the propensity score and the
outcome regression models are misspecified. For this purpose we replace Assumptions 3
and 4 with two other assumptions defining the probability limit of the estimators under a
general misspecification. The misspecifications will further be used to define a general bias
of the IPW and DR-estimators. When the propensity score is misspecified an estimator,
e.g., a quasi maximum likelihood estimator (QMLE) is not consistent for β in Assumption
3. However, a probability limit for an estimator under model misspecification exists under
general conditions, see e.g. White (1982, Theorem 2.2) for QMLE or Wooldridge (2010,
Section 12.1) and Boos and Stefanski (2013, Theorem 7.1) for estimators that can be
written as a solution of an estimating equation (M-estimators).
In the following, and as an alternative to Assumptions 3 and 4, we will assume that
such limits exists. Below we define an estimator ê∗ (X) of the propensity score under a
misspecified model emis (X, β ∗ ).
Assumption 5. [Misspecified PS parameters]
p
Let β̂ ∗ be an estimator under model misspecification, emis (X, β ∗ ), then β̂ ∗ −→ β ∗ .
Under model misspecification the probability limit of βˆ∗ is generally well defined however
emis (X, β ∗ ) is not equal to the propensity score e(X). In the following we use the notation
ê∗ (X) = emis (X, β̂ ∗ ) as the estimated propensity score and e∗ (X) = emis (X, β ∗ ) under
MODEL MISSPECIFICATION AND BIAS
7
Assumption 5. Below we give an example for true and misspecified parametric models,
however, for Assumption 5 we do not need the existence of a true parametric model.
Example 1
β1 X − β2
X 2 )]−1
For one confounder X and a true PS model e(X, β) = [1 + exp(−β0 −
assume that we misspecify the propensity score with a probit model
emis (X, β ∗ ) = Ψ(−β0∗ − β1∗ X), i.e., we misspecify the link function and omit a second order
term. Let β̂ ∗ = (β̂0∗ , β̂1∗ ) be the QMLE estimator of the parameters in emis (X, β ∗ ) obtained
by maximizing the quasi-likelihood
ln L =
n
X
Ti ln emis (Xi , β ∗ ) + (1 − Ti ) ln(1 − emis (Xi , β ∗ )) ,
i=1
p
Then ê∗ (X) = Ψ(−β̂0∗ − β̂1∗ X), β̂ ∗ = (β̂0∗ , β̂1∗ ) −→ β ∗ = (β0∗ , β1∗ ) under Assumption 5 and
e∗ (X) = Ψ(−β0∗ − β1∗ X) .
When considering the existence of true and misspecified parametric models, as illustrated in Example 1, the parameters in β and the limiting parameters β ∗ under the
misspecified model need not to be of the same dimension. For instance, the true model
could contain higher order terms and interactions that are not present in the estimation
model.
The next assumption concerns overlap under model misspecification.
Assumption 6. [Overlap under misspecification]
ν < e∗ (X) < 1 − ν, for some ν > 0.
In addition to the PS model we also consider misspecified outcome regression models,
∗
∗
mis
∗
µmis
t (X, αt ), t = 0, 1. Denote by α̂t , t = 0, 1 the estimator of the parameters in µt (X, αt ).
Assumption 7. [Misspecified OR model parameters]
p
∗
∗
∗
Let α̂t∗ be an estimator under model misspecification µmis
t (X, αt ), t = 0, 1, then α̂t −→ αt ,
t = 0, 1.
∗
In the following we use the notation µ̂∗t (X) = µmis
t (X, α̂t ) as the estimated OR and
∗
∗
∗
µ∗t (X) = µmis
t (X, αt ) under Assumption 7 and µt for the expected value E [µt (X)], t = 0, 1.
Assumptions 5 and 7 are defined for misspecified PS and OR models for the purpose of
describing their influence on the estimation of ∆. The estimators (2), (4) and (6) can be
written by estimating equations where the equations solving for the PS and OR parameters
are set up below the main equation for the IPW and DR estimators, see e.g Lunceford
and Davidian (2004) and Williamson, Forbes, and White (2014). Assuming parametric
PS and OR models the IPW estimators correspond to solving 2 + p estimating equations
Pn
i=1 ψ(θ, Yi , Ti , Xi ) = 0 for the parameters θIPWk = (µ1 , µ0 , β), k = 1, 2 and for the DR
estimator 2 + p + q1 + q0 estimating equations for the parameters θDR = (µ1 , µ0 , β, α1 , α0 ).
Using the notation for the misspecified models in assumptions 5 and 7 the estimating
equations change according to the dimensions of the parameters β ∗ and αt∗ , t = 0, 1. A
INGEBORG WAERNBAUM1 AND LAURA PAZZAGLI2
8
key condition for Assumptions 5 and and 7 to hold is that the misspecification of the PS
and/or OR provides estimating equations that uniquely define the parameter although,
as a consequence of the misspecification, it will not be the true average causal effect. In
the next section we present the asymptotic bias for the IPW and DR estimators under
study with general expressions including the limits of the misspecified propensity score
and outcome regression models.
4. Bias resulting from model misspecification
ˆ DR
ˆ IPW and ∆
ˆ IPW , ∆
4.1. General biases. In order to study the large sample bias of ∆
2
1
ˆ ∗IPW , ∆
ˆ ∗IPW and ∆
ˆ ∗DR by replacing
under model misspecification we define the estimators ∆
1
2
ê(X) in Equations (2), (4) and (6) with ê∗ (X). For the DR-estimator we additionally
replace µ̂t (X) with µ̂∗t (X), t = 0, 1.
To assess the properties of the estimators we assume 1, 2, 5, 6 and 7 and regularity conditions for applying a weak law of large numbers for averages with estimated parameters,
see Appendix A.1. Note that Assumption 3 and 4 are no longer needed. We evaluate the
difference between the probability limits of the estimators under model misspecification
and the average causal effect ∆ for the IPW and DR estimators:
ˆ∗
Theorem 8 (Bias under model misspecification for ∆
IPW1 ). Under Assumptions 1-2 and
5-6
p
ˆ∗
∆
IPW1 − ∆ −→ E
e(X)
1 − e(X)
µ1 (X) − E
µ0 (X) − (µ1 − µ0 ).
e∗ (X)
1 − e∗ (X)
ˆ∗
Theorem 9 (Bias under model misspecification for ∆
IPW2 ). Under Assumptions 1-2 and
5-6
p
ˆ∗
∆
IPW2 − ∆ −→
E
h
e(X)
e∗ (X) µ1 (X)
E
h
e(X)
e∗ (X)
i
i
E
−
h
i
1−e(X)
1−e∗ (X) µ0 (X)
E
h
1−e(X)
1−e∗ (X)
i
− (µ1 − µ0 ).
ˆ ∗DR ). Under Assumptions 1-2 and
Theorem 10 (Bias under model misspecification for ∆
5-7
(e(X) − e∗ (X)) (µ1 (X) − µ∗1 (X))
∆DR − ∆ −→E
e∗ (X)
[e(X) − e∗ (X)] (µ0 (X) − µ∗0 (X))
+E
.
(1 − e∗ (X))
p
ˆ∗
See Appendix A.2 for proofs.
We refer to the limits in Theorem 8, 9, and 10 as the asymptotic biases of the respective
ˆ ∗IPW ), Bias(∆
ˆ ∗IPW ) and Bias(∆
ˆ ∗DR ) although they are the difference
estimators, i.e., Bias(∆
1
2
between the probablity limits of the estimators and the true ∆ and not the difference in
ˆ ∗ is displayed by Theorem 10 since if
expectations. The double robustness property of ∆
DR
p
ˆ ∗ −→
either e(X) = e∗ (X) or µt (X) = µ∗ (X), t = 0, 1 we have that ∆
∆.
t
DR
To provide an illustrative example of the biases of the estimators we obtain the misspecified models’ limits by misspecifying the link functions in generalized linear models.
MODEL MISSPECIFICATION AND BIAS
9
However, a nonparametric data generating process could also be used. For the propensity
score we use binary response models with logit link (true) and a cloglog link (misspecified), for the outcome regression models we use poisson models with log links (true) and
gaussian models (misspecified) with identity links. We use numerical approximations to
provide values on the parameters in e∗ (X) and µ∗t (X), t = 0, 1 under the given true and
misspecified models e(X), emis (X), µt (X) and µmis
t (X), t = 0, 1.
Example 2
[Bias from link misspecifications]
Let X ∼ U nif orm(−2, 2) and T ∼ Bernoulli(e(X)). Assume that
e(X) = [1 + exp (0.5 − X)]−1 ,
µ1 (X) = exp (2.3 + 0.14X),
µ∗1 (X) = 10.06 + 1.48X,
e∗ (X) = 1 − exp [− exp(−0.81 + 0.74X)] ,
µ0 (X) = exp (1.4 + 0.20X),
µ∗0 (X) = 4.14 + 0.79X.
The marginal means are µ1 = 10.11 and µ0 = 4.16, so that ∆ = 5.94. Here we have that,
ˆ ∗IPW ) = −0.16, Bias(∆
ˆ ∗IPW ) = 0.05, and Bias(∆
ˆ ∗DR ) = −0.02
Bias(∆
1
2
4.2. Comparisons. To analyze the impact of the model misspecification on the estimators’ biases in Section 4.1 we compare the biases for two parts separately. The first part
concerns the bias with respect to µ1 and the second part with respect to µ0 . The first
ˆ ∗IPW ) in Theorem 8 is
part of Bias(∆
1
(8)
E
e(X)
e(X)
e(X)
µ
(X)
−
µ
=
cov
,
µ
(X)
+
E
−
1
µ1 ,
1
1
1
e∗ (X)
e∗ (X)
e∗ (X)
ˆ ∗IPW ) in Theorem 9 is
and the first part of Bias(∆
2
E
(9)
h
i
e(X)
e∗ (X) µ1 (X)
E
h
e(X)
e∗ (X)
i
cov
− µ1 =
h
i
e(X)
e∗ (X) , µ1 (X)
E
h
e(X)
e∗ (X)
i
.
ˆ∗
For ∆
IPW1 we see in (8) that the mean difference between the expected value of the conditional outcome, scaled with the PS-model error, and the marginal outcome contributes
ˆ∗
to the bias. For ∆
the contribution (9) is the difference between the conditional and
IPW2
the marginal outcome with the same error scaling, but here, the expected value of the
ˆ ∗IPW we
PS-model error, E [e(X)/e∗ (X)], also enters the bias in the denominator. For ∆
1
see from the right hand side of 8 that the sign depends on the covariance of [e(X)/e∗ (X)]
ˆ ∗IPW we see from the right hand
and µ1 (X), and the sign of E [e(X)/e∗ (X) − 1] µ1 . For ∆
2
side of (9) that the sign of depends on the covariance only. Hence, the part of the biases
described above can be in different directions for the same model misspecification, see also
Example 2 for the bias in total.
In the sequel we will give results concerning the absolute values of the first part of the
biases in Theorems 8, 9 and 10 but the results can be directly applied for the second
INGEBORG WAERNBAUM1 AND LAURA PAZZAGLI2
10
part of the biases by replacing e(X)/e∗ (X), with (1 − e(X))/(1 − e∗ (X)) and µ1 (X) with
ˆ∗
ˆ∗
ˆ ∗IPW ), Bias1 (∆
µ0 (X), see Appendix A.3. We define Bias1 (∆
IPW ) and Bias1 (∆DR ) as
1
2
e(X)
µ1 (X) − µ1 ,
e∗ (X)
i
h
E ee(X)
∗ (X) µ1 (X)
ˆ ∗IPW ) =
h
i
− µ1 ,
Bias1 (∆
2
e(X)
E e∗ (X)
e(X)
∗
∗
ˆ
− 1 (µ1 (X) − µ1 (X)) .
Bias1 (∆DR ) = E
e∗ (X)
ˆ∗
Bias1 (∆
IPW1 ) = E
(10)
(11)
(12)
Below, we give a sufficient and necessary condition for when the absolute value of
ˆ ∗IPW ) is smaller than the absolute value of Bias1 (∆
ˆ ∗IPW ).
Bias1 (∆
2
1
ˆ ∗IPW ) and Bias1 (∆
ˆ ∗IPW ) ).
Theorem 11 (Comparing Bias1 (∆
2
1
h
i
e(X)
E e∗ (X) µ1 (X)
e(X)
h
i
− µ1 < E ∗
µ1 (X) − µ1 .
e (X)
E ee(X)
∗ (X)
if and only if
e(X)
e(X)
e(X)
e(X)
cov ∗
cov ∗
, µ1 (X) < E ∗
, µ1 (X) + E ∗
− 1 µ1 .
e (X)
e (X)
e (X)
e (X)
The proof follows directly from the multiplication rules of absolute values. The theorem
h displays
i that the comparison of the biases depends on the magnitude and sign of
e(X)
E e∗ (X) − 1 µ1 and the covariance. It is no surprise that the covariance of e(X)/e∗ (X)
and µ1 (X) (and similarly of [1 − e(X)] / [1 − e∗ (X)] and µ0 (X)) plays a role for the bias
of the estimators. If µ1 (X) was a constant it could be taken out of the expectations of the
first terms in (10) and (11) and the PS-model error would be cancelled by the denominator
E [e(X)/e∗ (X)]. In this case the bias in (11) would be 0, and thus smaller than (10).
ˆ ∗IPW ) and Bias1 (∆
ˆ∗
Example 2 1 [Example 2 revisited for Bias1 (∆
IPW1 )]
2
e(X)
e(X)
E ∗
= 0.98, cov ∗
, µ1 (X) = 0.065, µ1 = 10.11.
e (X)
e (X)
To apply Theorem 11 for the data generating process in Example 2 we evaluate
e(X)
cov ∗
, µ1 (X)
e (X)
e(X)
<E ∗
e (X)
e(X)
e(X)
cov ∗
, µ1 (X) + E ∗
− 1 µ1
e (X)
e (X)
|0.064| < 0.98 · |0.064 − 0.02 · 10.11|
= 0.13,
ˆ∗
ˆ∗
and we conclude that Bias1 (∆
IPW2 ) < Bias1 (∆IPW1 ).
MODEL MISSPECIFICATION AND BIAS
11
We proceed by investigating the difference between the bias of the IPW estimators (10),
(11) and the bias of the DR estimator (12). Hence, we analyze the question of when two
ˆ ∗ ) and Bias1 (∆
ˆ∗
wrong models are better than one. We start by comparing Bias1 (∆
DR
IPW1 ).
∗
ˆ
In the following theorem we show a necessary condition for Bias1 (∆ ) to be smaller than
DR
ˆ∗
Bias1 (∆
IPW1 ). In the sequel all proofs are provided in Appendix A.3.
ˆ ∗ ) smaller than Bias1 (∆
ˆ∗
Theorem 12 (Necessary condition for Bias1 (∆
DR
IPW1 )). If
e(X)
e(X)
∗
E
− 1 (µ1 (X) − µ1 (X)) < E
− 1 µ1 (X) ,
e∗ (X)
e∗ (X)
then
E
e(X)
e(X)
∗
− 1 µ1 (X) < 2 · E
− 1 µ1 (X) .
e∗ (X)
e∗ (X)
The theorem states that if the DR estimator improves upon the simple IPW-estimator
under misspecification of both the PS and the OR model, then, the absolute value of the
misspecified outcome model is less than double the absolute value of the true conditional
mean under the same scaling of the PS-model error.
Below we give two examples of sufficient conditions for the DR-estimator to have a
smaller bias than the simple IPW estimator.
ˆ ∗DR ) smaller than Bias(∆
ˆ ∗IPW )). If a) µ∗ =
Theorem 13 (Sufficient conditions for Bias(∆
1
1
µ1 and
cov
h
e(X)
e∗ (X) , µ1 (X)
i
− cov
h
e(X)
∗
e∗ (X) , µ1 (X)
i
< cov
h
i
e(X)
e∗ (X) , µ1 (X)
+E
h
e(X)
e∗ (X)
i
− 1 µ1 ,
or,
b)
e(X)
e(X)
∗
−
1
µ
(X)
<
2
·
E
−
1
µ
(X)
,
1
1
e∗ (X)
e∗ (X)
i
h
i
− 1 µ∗1 (X) and E ee(X)
−
1
µ
(X)
are either both positive or both
∗ (X)
1
E
and E
h
e(X)
e∗ (X)
negative, then
E
e(X)
e(X)
∗
−
1
(µ
(X)
−
µ
(X))
<
E
−
1
µ
(X)
.
1
1
1
e∗ (X)
e∗ (X)
One of the criteria in a), that µ1 = µ∗1 , is reasonable to assume when the corresponding
moment condition is used in the estimation of the misspecified outcome model. Also, we
have that criterion b) is the same as the necessary condition with the added assumption
that the expectation of the (PS-error scaled) conditional outcomes have the same sign.
ˆ ∗ ) and Bias1 (∆
ˆ∗
Example 2 2 [Example 2 revisited for Bias1 (∆
DR
IPW1 )]
For the data generating process in Example 2 we investigate the first part of the bias
INGEBORG WAERNBAUM1 AND LAURA PAZZAGLI2
12
E
e(X)
e(X)
∗
− 1 (µ1 (X) − µ1 (X)) < E
− 1 µ1 (X)
e∗ (X)
e∗ (X)
0.01 < 0.11
implies
E
e(X)
e(X)
∗
− 1 µ1 (X) < 2 · E
− 1 µ1 (X)
e∗ (X)
e∗ (X)
0.10 < 0.22
which is consistent with Theorem 12. Assessing the sufficient conditions, in Theorem 13
a) we have that
µ1 = 10.11 ≈
µ∗1
e(X) ∗
, µ (X) = 0.074,
cov ∗
e (X) 1
= 10.07,
and thereby,
e(X)
e(X) ∗
e(X)
e(X)
cov ∗
, µ1 (X) − cov ∗
, µ (X) < cov ∗
, µ1 (X) + E ∗
− 1 µ1
e (X)
e (X) 1
e (X)
e (X)
|−0.064 + 0.074| < |−0.064 − 0.02 · 10.11|
0.01 < 0.27.
h
i
h
i
e(X)
The conditions in b) cannot be applied since E ee(X)
−
1
µ
(X)
and
cov
,
µ
(X)
∗ (X)
1
1
e∗ (X)
do not have the same sign.
ˆ ∗ ) smaller than Bias1 (∆
ˆ∗
Theorem 14 (Necessary condition for Bias1 (∆
DR
IPW2 )). If
i
h
E ee(X)
∗ (X) µ1 (X)
e(X)
∗
i
h
−
1
(µ
(X)
−
µ
(X))
<
− µ1 ,
E
1
1
e∗ (X)
E e(X)
e∗ (X)
then
h
i
e(X)
cov
,
µ
(X)
∗
1
e (X)
e(X)
e(X)
∗
i
h
E
−
1
(µ
(X))
−
<
E
−
1
[µ
(X)]
1
1
e∗ (X)
e∗ (X)
E ee(X)
∗ (X)
h
i
e(X)
cov
,
µ
(X)
1
e∗ (X)
e(X)
h
i
<E
− 1 [µ1 (X)] +
.
∗
e (X)
E e(X)
e∗ (X)
From the theorem we see that for the DR estimator to improve upon the normalized
IPW estimator we need that the outcome misspecification is within an interval defined by
the true conditional outcome and the absolute value of the covariance. This means that
the smaller the covariance is, the more accuracy of the outcome model is required for the
ˆ ∗ to be less biased than ∆
ˆ∗
∆
DR
IPW2 .
Under similar assumptions as in Theorem 13 we describe sufficient conditions for the
ˆ ∗ ) and Bias1 (∆
ˆ∗
comparison of Bias1 (∆
).
DR
IPW2
MODEL MISSPECIFICATION AND BIAS
13
ˆ ∗ ) smaller than Bias1 (∆
ˆ∗
Theorem 15 (Sufficient conditions for Bias1 (∆
DR
IPW2 )). If a)
h
i
e(X)
h
i
h
i
cov e∗ (X) ,µ1 (X)
e(X)
∗
i
h
µ1 = µ∗1 and cov ee(X)
<
∗ (X) , µ1 (X) − cov e∗ (X) , µ1 (X)
e(X)
E
or h
if,
b) E
e(X)
e∗ (X)
e∗ (X)
i
h
i
− 1 µ1 (X) and cov ee(X)
,
µ
(X)
are either both positive or both negative,
∗ (X)
1
and
E
e(X)
− 1 µ∗1 (X)
e∗ (X)
< E
e(X)
− 1 µ1 (X) +
e∗ (X)
cov
h
e(X)
e∗ (X) , µ1 (X)
E
h
e(X)
e∗ (X)
i
i
then,
E
i
h
e(X)
µ
(X)
E
e∗ (X) 1
e(X)
h
i
− 1 (µ1 (X) − µ∗1 (X)) <
− µ1 .
∗
e (X)
E ee(X)
∗ (X)
ˆ ∗DR ) and Bias1 (∆
ˆ ∗IPW ) the numerical example again
For a comparison between Bias1 (∆
2
shows that the necessary conditions of Theorem 14 and the sufficient condition in Theorem
15 a) are satisfied.
5. Simulation study
In order to investigate the asymptotic biases and also the finite sample performance
ˆ∗
ˆ∗
ˆ∗
of ∆
IPW1 , ∆IPW2 and ∆DR under model misspecification of both the PS and the OR
models we perform a simulation study with three different designs A, B and C. The first
part of the simulations evaluate the finite sample performances of the estimators and
consist of 1000 replications of sample sizes 500, 1000 and 5000. We generate covariates
X1 ∼ Uniform(1,4), X2 ∼ Poisson(3) and X3 ∼ Bernoulli(0.4). We use generalized linear
models to generate a binary treatment T and potential outcomes Y (t), t = 0, 1 with second
order terms of X1 and X2 in both the PS and OR models. The PS-distributions for the
treated and controls are bounded away from zero and 1 under the true models and under
the model misspecifications. The PS and OR models (for the DR estimator) are stepwise
misspecified. We have three designs where:
A: a quadratic term X1 2 is omitted in the PS and OR models;
B: two quadratic terms, X1 2 and X2 2 , are omitted in the PS and OR models;
C: two quadratic terms are omitted and the both the OR and PS link functions are
misspecified.
The glm family and link functions together with the true parameter values are given
in Table 1 which also contains the details for the misspecified models. The simulation is
performed with the statistical software R (R Core Team 2015).
In Table 2 we give the simulation bias, standard error and MSE of the three estimators.
ˆ IPW , ∆
ˆ IPW and ∆
ˆ DR the
When using the true models, i.e, when studying the estimators ∆
1
2
bias is small and decreases when the sample size increases and the standard errors follow
ˆ DR followed by ∆
ˆ IPW and ∆
ˆ IPW (Lunceford
the expected order with the smallest for ∆
2
1
INGEBORG WAERNBAUM1 AND LAURA PAZZAGLI2
14
and Davidian 2004). Under misspecification the bias does not decrease with the sample
size but gets closer to the asymptotic biases, see Table 3. Under misspecification the
ˆ∗
standard errors follow the same pattern as under the true models. The bias of ∆
IPW1 is
∗
∗
∗
ˆ
ˆ
ˆ
the largest whereas the ∆
and ∆
are similar. The MSE of ∆
is the smallest,
IPW2
DR
DR
ˆ∗
ˆ∗
however for n = 5000, ∆
IPW2 and ∆DR are very similiar.
ˆ∗
In Table 3 we give numerical approximations for Bias(∆
IPW1 ),
ˆ∗
ˆ∗
Bias(∆
IPW2 ) and Bias(∆DR )
using a sample size of n = 1, 000, 000. We also show the same approximations for
ˆ∗
ˆ∗
ˆ∗
Biast (∆
IPW1 ), Biast (∆IPW2 ) and Biast (∆DR ), t = 0, 1. Here, we see that the total bias
ˆ ∗IPW in Design A but smaller for ∆
ˆ ∗DR in Design B and C. The absolute
is smallest for ∆
2
ˆ ∗DR in all designs. We also give the
value of the biases in the two parts are smallest for ∆
expectations and covariances that are used for the necessary and sufficient conditions in
Theorems 11-15. We immediately see that the necessary condition for the absolute valˆ ∗DR ) to be smaller than the absolute value of Biast (∆
ˆ ∗IPW ) holds for both
ues of Biast (∆
1
t = 0, 1. The means (µ0 , µ1 ) are close to the means under model misspecification (µ∗0 , µ∗1 )
which is an assumption needed in order to evaluate the sufficient conditions in Theorems
13 a) and 15 a). By inspecting the covariances and the additional critera of the Theorems
13 a) and 15 a) in Table 3 we can see that the resulting inequalities of the theorems are
in line with the results for the asymptotic biases.
Since we have that E [(e(X)/e∗ (X) − 1) µ1 (X)] and cov [e(X)/e∗ (X), µ1 (X)] and
E [((1 − e(X))/(1 − e∗ (X)) − 1) µ0 (X)] and cov [(1 − e(X))(1 − e∗ (X)), µ0 (X)] do not have
the same sign in Design A and C the sufficient conditions in Theorems 13 b) and 15 b)
cannot be applied.
6. Discussion
In this paper we investigate biases of two IPW estimators and a DR estimator under model misspecification. For this purpose, we use a generic probability limit, under
misspecification of the PS and OR models, which exists under general conditions. Since
the propensity score enters the estimator in different ways for the IPW estimators under
study the consequences of the model misspecification are different. The bias of the IPW
estimators depend on the covariance between the PS-model error and the conditional outcome in different ways and the resulting bias can be in different directions. Comparing
the bias of the DR estimator with a simple IPW estimator the necessary condition for the
DR estimator to have a smaller bias is that the expectation of the outcome model under
misspecification is less than twice the true conditional outcome, where the expectations
also includes a scaling with the PS-model error. For the comparison with the normalized
IPW estimator the (PS-error scaled) misspecified outcome involves an interval defined by
the true conditional outcome adding and subtracting the absolute value of the covariance
between the PS-model error and the conditional outcome.
To our knowledge, there are only simulation studies comparing DR-estimators with
other estimators (Kang and Schafer 2007; Waernbaum 2012) under the assumption that
References
15
all models are misspecified. In this paper we study the same problem with an analytical
approach although the comparisons are made between a DR estimator and IPW estimators.
Acknowledgement
The authors acknowledge the Royal Swedish Academy of Sciences and the Swedish
Research Council, (grant number 2016-00703) for financial support.
1
Department of Statistics, Umeå University, Sweden, Institute for Evaluation of Labour
Market and Education Policy, IFAU, Uppsala, Sweden,
2
Division of Statistics, Department
of Economics, University of Perugia, Italy,
E-mail address: [email protected]
Figure 1. Density plots of the propensity score distributions, ê(X) and
ê∗ (X) for treated and controls for Design A (top), B (middle), and C
(bottom).
Design A, e*(X) for treated and controls
2.0
1.5
Density
0.0
0.0
0.5
1.0
1.5
1.0
0.5
Density
2.0
2.5
2.5
Design A, e(X) for treated and controls
0.2
0.4
0.6
0.8
1.0
1.2
0.0
0.2
0.4
0.6
0.8
e(X)
e*(X)
Design B, e(X) for treated and controls
Design B, e*(X) for treated and controls
1.0
1.2
2.0
1.0
0.5
0.0
0.0
0.2
0.4
0.6
0.8
1.0
1.2
−0.2
0.0
0.2
0.4
0.6
e(X)
e*(X)
Design C, e(X) for treated and controls
Design C, e*(X) for treated and controls
0.8
1.0
0.8
1.0
1.5
Density
2.0
0.0
0.0
0.5
1.0
1.0
3.0
2.0
0.0
Density
1.5
Density
1.5
1.0
0.5
Density
2.0
2.5
2.5
3.0
0.0
0.0
0.2
0.4
0.6
0.8
1.0
1.2
−0.2
e(X)
0.0
0.2
0.4
0.6
1.2
e*(X)
References
Bang, H. and J. M. Robins (2005). Doubly robust estimation in missing data and causal
inference models. Biometrics 61 (4), 962–973.
Boos, D. D. and L. Stefanski (2013). M-estimation (estimating equations). In Essential
Statistical Inference, Volume 120, pp. 297–337. Springer New York.
16
References
Busso, M., J. DiNardo, et al. (2014). New evidence on the finite sample properties of
propensity score reweighting and matching estimators. Review of Economics and
Statistics 96 (5), 885–897.
Cao, W., A. A. Tsiatis, et al. (2009). Improving efficiency and robustness of the doubly
robust estimator for a population mean with incomplete data. Biometrika 96 (3),
723–734.
Hernán, M. A., B. Brumback, et al. (2000). Marginal structural models to estimate the
causal effect of zidovudine on the survival of hiv-positive men. Epidemiology 11 (5),
561–570.
Hernán, M. A. and J. M. Robins (2006). Estimating causal effects from epidemiological
data. Journal of Epidemiology and Community Health 60 (7), 578–586.
Hirano, K. and G. W. Imbens (2001). Estimation of causal effects using propensity score
weighting: An application to data on right heart catheterization. Health Services and
Outcomes Research Methodology 2 (3), 259–278.
Hirano, K., G. W. Imbens, et al. (2003). Efficient estimation of average treatment effects
using the estimated propensity score. Econometrica 71 (4), 1161–1189.
Imai, K. and M. Ratkovic (2014). Covariate balancing propensity score. Journal of the
Royal Statistical Society: Series B (Statistical Methodology) 76 (1), 243–263.
Imbens, G. W. and J. M. Wooldridge (2009). Recent developments in the econometrics
of program evaluation. Journal of Economic Literature 47 (1), 5–86.
Kang, J. D. and J. L. Schafer (2007). Demystifying double robustness: A comparison
of alternative strategies for estimating a population mean from incomplete data.
Statistical Science 22 (4), 523–539.
Khan, S. and E. Tamer (2010). Irregular identification, support conditions, and inverse
weight estimation. Econometrica 78 (6), 2021–2042.
Kwon, T., I. G. Jeong, et al. (2015). Adjuvant chemotherapy after radical cystectomy for
bladder cancer: a comparative study using inverse-probability-of-treatment weighting. Journal of Cancer Research and Clinical Oncology 141 (1), 169–176.
Lunceford, J. K. and M. Davidian (2004). Stratification and weighting via the propensity
score in estimation of causal treatment effects: a comparative study. Statistics in
Medicine 23 (19), 2937–2960.
Petersen, M. L., K. E. Porter, S. Gruber, Y. Wang, and M. J. van der Laan (2010).
Diagnosing and responding to violations in the positivity assumption. Statistical
Methods in Medical Research 21 (1), 31–54.
R Core Team (2015). R: A Language and Environment for Statistical Computing. Vienna, Austria: R Foundation for Statistical Computing.
Robins, J., M. Sued, et al. (2007). Comment: Performance of double-robust estimators
when “inverse probability” weights are highly variable. Statistical Science 22 (4),
544–559.
References
17
Robins, J. M. and A. Rotnitzky (1994). Estimation of regression coefficients when some
regressors are not always observed. Journal of the American Statistical Association 89 (427), 846–866.
Rosenbaum, P. R. and D. B. Rubin (1983). The central role of the propensity score in
observational studies for causal effects. Biometrika 70 (1), 41–55.
Rubin, D. B. (1974). Estimating causal effects of treatments in randomized and nonrandomized studies. Journal of Educational Psychology 66 (5), 688–701.
Scharfstein, D. O., A. Rotnitzky, et al. (1999). Adjusting for nonignorable drop-out
using semiparametric nonresponse models. Journal of the American Statistical Association 94 (448), 1096–1120.
Seaman, S. R. and I. R. White (2013). Review of inverse probability weighting for
dealing with missing data. Statistical Methods in Medical Research 22 (3), 278–295.
Tan, Z. (2007). Comment: Understanding OR, PS and DR. Statistical Science 22 (4),
560–568.
Tsiatis, A. A. (2007). Semiparametric theory and missing data. Springer Science &
Business Media.
Tsiatis, A. A. and M. Davidian (2007). Comment: Demystifying double robustness: A
comparison of alternative strategies for estimating a population mean from incomplete data. Statistical Science 22 (4), 569–573.
Vansteelandt, S., J. Carpenter, et al. (2010). Analysis of incomplete data using inverse
probability weighting and doubly robust estimators. Methodology 6 (1), 37–48.
Vermeulen, K. and S. Vansteelandt (2015). Bias-reduced doubly robust estimation. Journal of the American Statistical Association 110 (511), 1024–1036.
Waernbaum, I. (2012). Model misspecification and robustness in causal inference: comparing matching with doubly robust estimation. Statistics in Medicine 31 (15), 1572–
1581.
White, H. (1982). Maximum likelihood estimation of misspecified models. Econometrica:
Journal of the Econometric Society 50 (1), 1–25.
Williamson, E., A. Forbes, and I. White (2014). Variance reduction in randomised
trials by inverse probability weighting using the propensity score. Statistics in
Medicine 33 (1), 721–737.
Wooldridge, J. (2010). Econometric Analysis of Cross Section and Panel Data. Econometric Analysis of Cross Section and Panel Data. MIT Press.
Yao, L., Z. Sun, et al. (2010). Estimation of average treatment effects based on parametric propensity score model. Journal of Statistical Planning and Inference 140 (3),
806–816.
18
References
Appendix A.
A.1. Regularity conditions for applying a weak law of large numbers for averages of functions with estimated parameters. The convergence in probability of
ˆ∗
ˆ∗
ˆ ∗ to their corresponding expectations would follow directly from a
∆
,∆
and ∆
IPW1
IPW2
DR
WLLN for an iid sample of (Ti , Xi , Yi ) except for the estimated parameters β̂ ∗ in ê∗ (Xi )
and and α̂t∗ in µ̂∗t (Xi ), t = 0,h 1. To justify
the biases in Section 4 consider a general
i
p
representation of a function g T, Y, X, θ̂ where θ̂ −→ θ0 and
i p
1X h
g Xi , Ti , Yi , θ̂ −→ E [g(T, X, Y, θ0 )]
n
(13)
ˆ ∗IPW , ∆
ˆ ∗IPW and (α̂∗ , β̂ ∗ ) for ∆
ˆ ∗DR and under AsThe θ̂ in (13) corresponds to β̂ ∗ for ∆
1
2
sumptions 5 and 7 the consistency of θ̂ is ensured. Regularity conditions for the function
g can be given see e.g., Boos and Stefanski (2013, Theorem 7.3) who show that (13) holds
for differentiable
h
ifunctions with bounded derivatives (wrt θ). The regularity conditions
for g Xi , Ti , Yi , θ̂ , for the three estimators, imply conditions on the models e∗ (X, β ∗ ) and
µ∗t (X, αt∗ ) such that the regularity condition for g is satisfied. Under (13) we can insert
the limiting values β ∗ and α∗ and their corresponding e∗ (X) and µ∗t (X), t = 0, 1 when
taking a WLLN.
A.2. Biases. Under the regularity conditions and assumptions 5-6 (IPW) and 5-7 (DR)
ˆ ∗IPW :
we derive the bias in Theorem 8 below. For ∆
1
n
n
1 X (1 − Ti )Yi p
TY
(1 − T ) Y
1 X Ti Yi
−
−→ E ∗
−E
,
n
ê∗ (Xi ) n
1 − ê∗ (Xi )
e (X)
(1 − e∗ (X))
i=1
i=1
and
TY
(1 − T ) Y
TY
(1 − T ) Y
E ∗
−E
=E E ∗
X −E E
X
e (X)
(1 − e∗ (X))
e (X)
[1 − e∗ (X)]
(1 − e(X))
e(X)
µ1 (X) − E
µ
(X)
.
=E ∗
0
e (X)
(1 − e∗ (X))
subtracting with ∆ gives
ˆ∗
Bias(∆
IPW1 ) =E
e(X)
1 − e(X)
µ1 (X) − E
µ0 (X) − (µ1 − µ0 ).
e∗ (X)
1 − e∗ (X)
The biases of Theorems 9 and 10 are derived similarly.
A.3. Comparisons. To study the consequences of model misspecification for the estimators we compare each difference involving µ1 (X) and µ0 (X) separately. For example we
ˆ∗
study Bias(∆
IPW1 )
References
ˆ∗
∆IPW1
19
e(X)
1 − e(X)
− ∆ −→E ∗
µ1 (X) − E
µ0 (X) − (µ1 − µ0 )
e (X)
1 − e∗ (X)
1 − e(X)
e(X)
µ1 (X) − µ1 + µ0 − E
µ0 (X)]
=E ∗
e (X)
1 − e∗ (X)
p
ˆ ∗IPW )
and Bias(∆
2
p
ˆ∗
∆
IPW2 − ∆ −→
E
E
=
h
i
e(X)
e∗ (X) µ1 (X)
E
h
h
E
h
e(X)
e∗ (X)
E
h
−
i
i
1−e(X)
1−e∗ (X) µ0 (X)
E
h
1−e(X)
1−e∗ (X)
i
e(X)
e∗ (X) µ1 (X)
e(X)
e∗ (X)
i
E
h
− µ 1 + µ0 −
i
− (µ1 − µ0 )
i
1−e(X)
1−e∗ (X) µ0 (X)
h
E
1−e(X)
1−e∗ (X)
i
.
The inequalities concerning the biases are made with respect to the absolute values for
two parts separately, i.e, for Theorem 11
(14)
E
e(X)
µ1 (X) − µ1 <
e∗ (X)
E
h
i
e(X)
e∗ (X) µ1 (X)
E
h
e(X)
e∗ (X)
i
− µ1 .
Since | − a| = |a| the same comparison for the second part are
i
h
1−e(X)
µ
(X)
E
1−e∗ (X) 0
1 − e(X)
i
h
E
µ0 (X) − µ0 <
− µ0 ,
∗
1−e(X)
1 − e (X)
E 1−e
∗ (X)
(15)
and the conditions derived for the first part of the biases, (10) and (11), can be directly
applied to (15) replacing e(X)/e∗ (X), with (1−e(X))/(1−e∗ (X)) and µ1 (X) with µ0 (X).
Similarly, for the second part of the bias in Theorem 10 we have that
(1 − e(X))(1 − e∗ (X)) (µ0 (X) − µ∗0 (X))
E
(1 − e∗ (X))
(e(X) − e∗ (X)) (µ0 (X) − µ∗0 (X))
= E
(1 − e∗ (X))
and the conditions derived for the first part of the bias, (12), can be directly applied to
(15) additionally replacing µ∗1 (X) with µ∗0 (X).
Proof of Theorem 12. 1. Assuming that E
E
h
e(X)
e∗ (X)
i
− 1 µ1 (X) > 0:
e(X)
e(X)
∗
− 1 (µ1 (X) − µ1 (X)) < E
− 1 µ1 (X) ,
e∗ (X)
e∗ (X)
Here, we have
−E
e(X)
e(X)
e(X)
∗
− 1 µ1 (X) < E
− 1 (µ1 (X) − µ1 (X)) < E
− 1 µ1 (X) ,
e∗ (X)
e∗ (X)
e∗ (X)
20
References
and
e(X)
e(X)
∗
(16)
0<E
− 1 µ1 (X) < 2 · E
− 1 µ1 (X) .
e∗ (X)
e∗ (X)
h
i
2. Assuming that E ee(X)
−
1
µ
(X)
< 0:
∗ (X)
1
E
e(X)
e(X)
∗
− 1 (µ1 (X) − µ1 (X)) < −E
− 1 µ1 (X) ,
e∗ (X)
e∗ (X)
Here, we have
e(X)
e(X)
e(X)
∗
− 1 µ1 (X) < E
− 1 (µ1 (X) − µ1 (X)) < −E
− 1 µ1 (X) ,
e∗ (X)
e∗ (X)
e∗ (X)
e(X)
e(X)
∗
− 1 µ1 (X) < E
− 1 µ1 (X) < 0
e∗ (X)
e∗ (X)
E
(17)
2·E
and by (16) and (17)
e(X)
e(X)
∗
E
− 1 µ1 (X) ≤ 2 · E
− 1 µ1 (X) .
e∗ (X)
e∗ (X)
Proof of Theorem 13. a) We have that
e(X)
e(X)
e(X) ∗
∗
cov ∗
, µ1 (X) − cov ∗
, µ (X) = E
− 1 (µ1 (X) − µ1 (X)) .
e (X)
e (X) 1
e∗ (X)
by µ∗1 = µ1 . Also,
e(X)
e(X)
e(X)
cov ∗
, µ1 (X) + E ∗
− 1 E [µ1 (X)] = E
− 1 µ1 (X) .
e (X)
e (X)
e∗ (X)
Further, assuming that
e(X)
e(X)
e(X) ∗
e(X)
cov ∗
, µ1 (X) − cov ∗
, µ (X) < cov ∗
, µ1 (X) + E ∗
− 1 µ1 ,
e (X)
e (X) 1
e (X)
e (X)
it follows that
E
e(X)
e(X)
∗
E
<
.
−
1
(µ
(X)
−
µ
(X))
−
1
µ
(X)
1
1
1
e∗ (X)
e∗ (X)
c) see proof of Theorem 12.
References
21
Proof of Theorem 14. 1. Assuming that:
h
i
e(X)
E
µ
(X)
∗
1
e (X)
e(X)
h
i
cov ∗
, µ1 (X) > 0 ⇐⇒
− µ1 > 0
e(X)
e (X)
E e∗ (X)
Here, we have
h
i
e(X)
E
µ
(X)
1
∗
e
(X)
e(X)
h
i
h
i
−
+ µ1 < E
− 1 (µ1 (X) − µ∗1 (X)) <
− µ1 ,
e∗ (X)
E ee(X)
E ee(X)
∗ (X)
∗ (X)
e(X)
e(X)
e(X)
e(X)
∗
, µ1 (X) < E ∗
−
1
(µ
(X)
−
µ
(X))
<
cov
,
µ
(X)
,
−cov ∗
E
1
1
1
e (X)
e (X)
e∗ (X)
e∗ (X)
E
h
e(X)
µ (X)
e∗ (X) 1
i
and
(18)
h
i
h
i
cov e(X) , µ1 (X)
cov e(X) , µ1 (X)
e∗ (X)
e∗ (X)
e(X)
e(X)
e(X)
∗
h
i
h
i
E
− 1 µ1 (X) −
<E
− 1 µ1 (X) < E
− 1 µ1 (X) +
e∗ (X)
e∗ (X)
e∗ (X)
E e(X)
E e(X)
e∗ (X)
e∗ (X)
2. Assuming that:
h
i
e(X)
E
µ
(X)
e∗ (X) 1
e(X)
h
i
cov ∗
, µ1 (X) < 0 ⇐⇒
− µ1 < 0,
e (X)
E ee(X)
∗ (X)
then it follows that
E
h
i
E ee(X)
∗ (X) µ1 (X)
e(X)
h
i
h
i
− 1 (µ1 (X) − µ∗1 (X)) < −
− µ1 < E
+ µ1
e∗ (X)
E ee(X)
E ee(X)
∗ (X)
∗ (X)
e(X)
e(X)
e(X)
e(X)
∗
, µ1 (X) < E ∗
E
−
1
(µ
(X)
−
µ
(X))
<
−cov
,
µ
(X)
,
cov ∗
1
1
1
e (X)
e (X)
e∗ (X)
e∗ (X)
h
e(X)
µ (X)
e∗ (X) 1
i
and
(19)
h
i
h
i
cov e(X) , µ1 (X)
cov e(X) , µ1 (X)
∗
∗
e (X)
e (X)
e(X)
e(X)
e(X)
h
i
h
i
E
<E
,
− 1 µ1 (X) +
− 1 µ∗1 (X) < E
− 1 µ1 (X) −
∗ (X)
∗ (X)
e(X)
e∗ (X)
e
e
E
E e(X)
e∗ (X)
e∗ (X)
hence by (18) and (19) we conclude that
E
h
e(X)
e∗ (X)
i
− 1 µ1 (X) −
cov
h
i
e(X)
,µ (X)
e∗ (X) 1
h
i
e(X)
E e∗ (X)
<E
h
e(X)
e∗ (X)
i
h
i
− 1 µ∗1 (X) < E ee(X)
∗ (X) − 1 µ1 (X) +
cov
Proof of Theorem 15. For a) assuming µ∗1 = µ1 implies
e(X)
e(X) ∗
e(X)
∗
cov ∗
, µ1 (X) − cov ∗
, µ (X) = E
− 1 (µ1 (X) − µ1 (X)) .
e (X)
e (X) 1
e∗ (X)
h
i
e(X)
,µ (X)
e∗ (X) 1
h
i
e(X)
E e∗ (X)
,
22
References
Also,
cov
h
i
e(X)
e∗ (X) , µ1 (X)
E
h
e(X)
e∗ (X)
i
E
=
h
i
e(X)
e∗ (X) µ1 (X)
E
h
e(X)
e∗ (X)
i
− µ1 ,
and by
i
h
e(X)
,
µ
(X)
cov
∗
1
e (X)
e(X)
e(X) ∗
i
h
,
cov ∗
, µ1 (X) − cov ∗
, µ1 (X) <
e (X)
e (X)
E ee(X)
∗ (X)
it follows that
i
h
e(X)
µ
(X)
E
∗
1
e (X)
e(X)
h
i
E
− µ1 .
− 1 (µ1 (X) − µ∗1 (X)) <
∗
e (X)
E ee(X)
∗ (X)
h
i
h
i
e(X)
For b) we use that if E ee(X)
−
1
µ
(X)
and
cov
,
µ
(X)
are both positive
∗ (X)
1
1
e∗ (X)
then,
E
e(X)
e(X)
e(X)
e(X)
− 1 µ1 (X) + cov ∗
, µ1 (X) = E
− 1 µ1 (X) + cov ∗
, µ1 (X)
e∗ (X)
e (X)
e∗ (X)
e (X)
and
e(X)
e(X)
e(X)
e(X)
− E
− 1 µ1 (X) + cov ∗
, µ1 (X) < E
− 1 µ1 (X) + cov ∗
, µ1 (X) .
e∗ (X)
e (X)
e∗ (X)
e (X)
h
i
h
i
e(X)
If E ee(X)
∗ (X) − 1 µ1 (X) and cov e∗ (X) , µ1 (X) are both negative, then we have
E
e(X)
e(X)
e(X)
e(X)
− 1 µ1 (X) + cov ∗
, µ1 (X) < E
− 1 µ1 (X) + cov ∗
, µ1 (X) ,
e∗ (X)
e (X)
e∗ (X)
e (X)
and
− E
e(X)
e(X)
e(X)
e(X)
− 1 µ1 (X) + cov ∗
, µ1 (X) = E
− 1 µ1 (X) + cov ∗
, µ1 (X) ,
e∗ (X)
e (X)
e∗ (X)
e (X)
and the necessary condition from Theorem 14 follows.
Gaussian, identity
OR
Gaussian, identity
OR
Binomial, cauchit
Gamma, identity
PS
OR
Design C
Binomial, logit
PS
Design B
Binomial, logit
Class
PS
Design A
Models
Binomial, logit
Gaussian, identity
X1 , X2 , X12 , X22 , X3
Gaussian, identity
X1 , X2 , X12 , X22 , X3
Binomial, logit
X1 , X2 , X12 , X22 , X3
Gaussian, identity
X1 , X2 , X12 , X22 , X3
X1 , X2 , X12 , X22 , X3
Binomial, logit
Class
X1 , X2 , X3
X1 , X2 , X3
X1 , X2 , X3
X1 , X2 , X3
X1 , X2 , X22 , X3
X1 , X2 , X22 , X3
Linear predictor
MISSPECIFIED MODEL
X1 , X2 , X12 , X22 , X3
α1 = (4, 1.1, 0.1, 0.5, 0.3, 0.2)
β = (−1, 0.6, 0.1, 0.9, 0.1, 0.7), α0 = (3, 0.5, 0.2, 0.5, 0.2, 0.2)
Linear predictor and parameter values
TRUE MODEL
Table 1. Simulation Designs A, B and C
References
23
24
References
Table 2. Simulation results
ESTIMATORS
ˆ∗
∆
ˆ∗
∆
IPW1
n
Models
500
True
1000
5000
Design
ˆ∗
∆
DR
IPW2
Bias
SD
MSE
Bias
SD
MSE
Bias
SD
MSE
0.018
0.368
0.136
0.013
0.138
0.019
<0.001
0.11
0.012
False
A
0.118
0.390
0.166
0.019
0.136
0.019
0.021
0.119
0.015
False
B
0.262
0.324
0.174
0.033
0.126
0.017
0.024
0.110
0.013
False
C
0.229
0.349
0.174
-0.061
0.128
0.020
0.037
0.115
0.015
0.010
0.254
0.065
0.002
0.098
0.010
< 0.001
0.075
0.006
False
True
A
0.112
0.270
0.085
0.008
0.097
0.009
0.018
0.079
0.007
False
B
0.260
0.214
0.113
0.034
0.088
0.009
0.025
0.078
0.007
False
C
0.243
0.222
0.108
-0.056
0.092
0.012
0.043
0.081
0.008
0.007
0.107
0.011
0.005
0.044
0.002
0.003
0.035
0.001
False
True
A
0.112
0.116
0.026
0.013
0.044
0.002
0.025
0.037
0.002
False
B
0.260
0.090
0.076
0.033
0.039
0.003
0.025
0.036
0.002
False
C
0.226
0.096
0.060
-0.057
0.040
0.005
0.040
0.036
0.003
References
25
Table 3. Asymptotic approximations from Designs A, B and C.
Design
A
Parameter
B
C
µ1
11.127 11.127 12.130
µ∗1
11.092 11.098 12.097
µ0
8.628
8.628
9.633
µ∗0
8.578
8.564
9.582
ˆ∗
Bias(∆
IPW1 )
ˆ
Bias(∆∗
)
0.096
0.264
0.213
0.007
0.033
-0.057
ˆ∗
Bias(∆
0.017
0.025
0.037
ˆ∗
Bias1 (∆
IPW1 )
ˆ
Bias1 (∆∗
)
0.024
0.128
0.130
-0.028
0.013
-0.089
ˆ∗
(∆
0.009
0.009
0.029
1.005
1.010
1.019
-0.029
0.015
-0.095
-0.040
0.006
-0.121
0.060
0.113
0.224
0.030
0.128
0.130
0.019
0.119
0.102
ˆ ∗IPW )
Bias2 (∆
1
ˆ ∗IPW )
Bias2 (∆
0.076
0.137
0.083
0.039
0.022
0.031
ˆ∗
(∆
0.011
0.018
0.008
0.995
0.987
0.995
-0.037
-0.017
-0.033
-0.026
-0.004
-0.024
-0.038
0.111
-0.052
-0.074
-0.128
-0.085
-0.065
-0.114
-0.075
IPW2
DR )
IPW2
Bias1 DR )
h
i
E ee(X)
∗ (X)
h
i
cov ee(X)
∗ (X) , µ1 (X)
h
i
∗
cov ee(X)
∗ (X) , µ1 (X)
h
i
E ee(X)
−
1
µ1
∗
h (X)
i
E ee(X)
∗ (X) − 1 µ1 (X)
h
i
∗
E ee(X)
∗ (X) − 1 µ1 (X)
2
Bias2 DR )
h
i
1−e(X)
E 1−e
∗
h (X)
i
1−e(X)
cov 1−e∗ (X) , µ0 (X)
h
i
1−e(X)
∗
cov 1−e
∗ (X) , µ0 (X)
h
i
1−e(X)
E 1−e
−
1
µ0
∗
h (X)
i
1−e(X)
E 1−e
−
1
µ
(X)
∗ (X)
0
h
i
1−e(X)
E 1−e∗ (X) − 1 µ∗0 (X)
| 10math.ST
|
1
ND-Tree-based update: a Fast Algorithm for the
Dynamic Non-Dominance Problem
arXiv:1603.04798v2 [cs.DS] 7 Nov 2017
Andrzej Jaszkiewicz, Thibaut Lust
Abstract—In this paper we propose a new method called NDTree-based update (or shortly ND-Tree) for the dynamic nondominance problem, i.e. the problem of online update of a Pareto
archive composed of mutually non-dominated points. It uses a
new ND-Tree data structure in which each node represents a
subset of points contained in a hyperrectangle defined by its
local approximate ideal and nadir points. By building subsets
containing points located close in the objective space and using
basic properties of the local ideal and nadir points we can
efficiently avoid searching many branches in the tree. ND-Tree
may be used in multiobjective evolutionary algorithms and other
multiobjective metaheuristics to update an archive of potentially
non-dominated points. We prove that the proposed algorithm
has sub-linear time complexity under mild assumptions. We
experimentally compare ND-Tree to the simple list, Quad-tree,
and M-Front methods using artificial and realistic benchmarks
with up to 10 objectives and show that with this new method
substantial reduction of the number of point comparisons and
computational time can be obtained. Furthermore, we apply
the method to the non-dominated sorting problem showing that
it is highly competitive to some recently proposed algorithms
dedicated to this problem.
Index Terms—Multiobjective optimization, Pareto archive,
Many-objective optimization, Dynamic non-dominance problem,
Non-dominated sorting
I. I NTRODUCTION
I
N this paper we consider the dynamic non-dominance
problem [1], i.e. the problem of online update of a Pareto
archive with a new candidate point. The Pareto archive is
composed of mutually non-dominated points and this property
must remain fulfilled following the addition of the new point.
The dynamic non-dominance problem is typically used in
multiobjective evolutionary algorithms (MOEAs) and more
generally in other multiobjective metaheuristics (MOMHs),
whose goal is to generate a good approximation of the Pareto
front. Many MOEAs and other MOMHs use an external
archive of potentially non-dominated points, i.e. a Pareto
archive containing points not dominated by any other points
generated so far, see e.g. [2], [3], [4], [5], [6], [7], [8], [9]. We
consider here MOEAs that generate iteratively new candidate
points and use them immediately to update a Pareto archive.
Updating a Pareto archive with a new point y means that:
•
y is added to the Pareto archive if it is non-dominated
w.r.t. any point in the Pareto archive,
A. Jaszkiewicz is with Poznan University of Technology, Faculty of
Computing, Institute of Computing Science, ul. Piotrowo 2, 60-965 Poznan,
Poland, e-mail: [email protected].
T. Lust is with Sorbonne Universités, UPMC, Université Paris 06, CNRS,
LIP6, UMR 7606, F-75005, Paris, France, e-mail: [email protected].
all points dominated by y are removed from the Pareto
archive.
The time needed to update a Pareto archive, in general,
increases with a growing number of objectives and a growing
number of points. In some cases it may become a crucial
part of the total running time of a MOEA. The simplest data
structure for storing a Pareto archive is a plain list. When a
new point y is added, y is compared to all points in the Pareto
archive until either all points are checked or a point dominating
y is found. In order to speed up the process of updating a
Pareto archive some authors proposed the use of specialized
data structures and algorithms, e.g. Quad-tree [10]. However,
the results of computational experiments reported in literature
are not conclusive and in some cases such data structures may
in fact increase the update time compared to the simple list.
A frequently used approach allowing reduction of the time
needed to update a Pareto archive is the use of bounded
archives [11] where the number of points is limited and some
potentially non-dominated points are discarded. Please note,
however, that such an approach always reduces the quality of
the archive. In particular, one of the discarded points could be
the one that would be selected by the decision maker if the
full archive was known. Bounded archives may be especially
disadvantageous in the case of many-objective problems, since
with a growing number of dimensions it is more and more
difficult to represent a large set with a smaller sample of
points. The use of bounded archives may also lead to some
technical difficulties in MOEAs (see [11]). Summarizing, if an
unbounded archive can be efficiently managed and updated, it
is advantageous to use this kind of archive.
In this paper, our contribution is fourfold: firstly, we propose
a new method, called ND-Tree-based update, for the dynamic
non-dominance problem. The method is based on a dynamic
division of the objective space into hyperrectangles, which
allows to avoid many comparisons of objective function values. Secondly, we show that the new method has sub-linear
time complexity under mild assumptions. Thirdly, a thorough
experimental study on different types of artificial and realistic
sets shows that we can obtain substantial computational time
reductions compared to state-of-the-art methods. Finally, we
apply ND-Tree-based update to the non-dominated sorting
problem obtaining promising results in comparison to some
recently proposed dedicated algorithms.
The remainder of the paper is organized as follows. Basic
definitions related to multiobjective optimization are given in
Section II. In Section III, we present a state of the art of the
methods used for online updating a Pareto archive. The main
contribution of the paper, i.e. ND-Tree-based update method
•
2
is described in Section IV. Computational experiments are
reported and discussed in Section V. In sections VI, ND-Treebased is applied to the non-dominated sorting problem.
II. BASIC DEFINITIONS
A. Multiobjective optimization
We consider a general multiobjective optimization (MO)
problem with a feasible set of solutions X and p objective
functions yk (x) to minimize. The image of the feasible set
in the objective space is a set of points Y = y(X ) where
y(x) = (y1 (x), y2 (x), . . . , yp (x)).
In MO, points are usually compared according to the Pareto
dominance relation:
Definition 1. Pareto dominance relation: we say that a point
u = (u1 , ..., up ) dominates a point v = (v1 , ..., vp ) if, and only
if, uk ≤ vk ∀ k ∈ {1, . . . , p} ∧ ∃ k ∈ {1, . . . , p} : uk < vk . We
denote this relation by u v.
Definition 2. Non-dominated point: a point y ∗ corresponding
to a feasible solution is called non-dominated if there does
not exist any other point y ∈ Y such that y y ∗ . The set YN
of all non-dominated points is called Pareto front.
Definition 3. Coverage relation: we say that a point u covers
a point v if u v or u = v. We denote this relation by u v.
Please note that coverage relation is sometimes referred to
as weak dominance [12].
Definition 4. Mutually non-dominated relation: we say that
two points are mutually non-dominated or non-dominated w.r.t.
each other if neither of the two points covers the other one.
Definition 5. Pareto archive (YN ): set of points such that
any pair of points in the set are mutually non-dominated, i.e
∀y ∈ YN , @y 0 ∈ YN | y 0 y.
In the context of MOEAs, the Pareto archive contains the
mutually non-dominated points generated so far (i.e. at a given
iteration of a MOEA) that approximates the Pareto front YN .
In other words YN contains points that are potentially nondominated at a given iteration of the MOEA.
Please note that in MOEAs not only points but also representations of solutions are preserved in the Pareto archive, but
the above definition is sufficient for the purpose of this paper.
The new method ND-Tree-based update is based on the
(approximate) local ideal and nadir points that we define
below.
Definition 6. The local ideal point of a subset S ⊆ YN
denoted as z ∗ (S) is the point in the objective space composed
of the best coordinates of all points belonging to S, i.e.
zk∗ (S) = min{yk }, ∀ k ∈ {1, . . . , p}. A point zb∗ (S) such that
y∈S
zb∗ (S) z ∗ (S) will be called Approximate local ideal point.
Naturally, the (approximate) local ideal point covers all
points in S.
Definition 7. The local nadir point of a subset S ⊆ YN
denoted as z∗ (S) is the point in the objective space composed
of the worst coordinates of all points belonging to S, i.e.
z∗k (S) = max{yk }, ∀ k ∈ {1, . . . , p}. A point zb∗ (S) such
y∈S
that z∗ (S) zb∗ (S) will be called Approximate local nadir
point.
Naturally, the (approximate) local nadir point is covered by
all points in S.
B. Dynamic non-dominance problem
The problem of updating a Pareto archive (also called nondominance problem), can be divided into two classes: the static
non-dominance problem is to find the set of non-dominated
points YN among a set of points Y . The other class is the
dynamic non-dominance problem [1] that typically occurs in
MOEAs. We formally define this problem as follows. Consider
a candidate point y and a Pareto archive YN . The problem is
to update YN with y and consists in the following operations.
If y is covered by at least one point in YN , y is discarded
and YN remains unchanged. Otherwise, y is added to YN .
Moreover, if some points in YN are dominated by y, all these
points are removed from YN , in order to keep only mutually
non-dominated points (see Algorithm 1).
Algorithm 1 DynamicNonDominance
Parameter l: A Pareto archive YN
Parameter ↓: New candidate point y
if (@ y 0 ∈ YN | y 0 y) then
YN ← YN ∪ {y}
for each (y 0 ∈ YN | y y 0 ) do
YN ← YN \{y 0 }
In this work we consider only the dynamic non-dominance
problem. Note that in general static problems may be solved
more effectively than their dynamic counterparts since they
have access to richer information. Indeed, some efficient algorithms for static non-dominance problem have been proposed,
see [13], [14], [15], [16].
MOEAs and other MOMHs usually update the Pareto
archive using the dynamic version of the non-dominance
problem, i.e. the Pareto archive is updated with each newly
generated candidate point. In some cases it could be possible
to store all candidate points and then solve the static nondominance problem. The latter approach has, however, some
disadvantages:
• MOEAs need to store not only points in the objective
space but also full representations of solutions in the
decision space. Thus, storing all candidate points with
corresponding solutions may be very memory consuming.
• Some MOEAs use the Pareto archive during the run of the
algorithm, i.e. Pareto archive is not just the final output
of the algorithm. For example in [17], one of the parents
is selected from the Pareto archive. In [7] the success
of adding a new point to the Pareto archive influences
the probability of selecting weight vectors in further
iterations. The same applies to other MOMHs as well.
For example, the Pareto local search (PLS) method [18]
works directly with the Pareto archive and searches
3
neighborhood of each solution from the archive. In such
methods, computation of the Pareto archive cannot be
postponed till the end of the algorithm.
Note that as suggested in [19] the dynamic non-dominance
problem may also be used to speed up the non-dominated
sorting procedure used in many MOEAs. As the Pareto archive
contains all non-dominated points generated so far the first
front is immediately known and the non-dominated sorting
may be applied only to the subset of dominated points. Using
this technique, Drozdı́k et al. showed that their new method
called M-Front could obtain better performance than Deb’s fast
nondominated sorting [20] and Jensen-Fortin’s algorithm [21],
[22], one of the fastest non-dominated sorting algorithms.
III. S TATE OF THE ART
We present here a number of methods for the dynamic
non-dominance problem proposed in literature and used in
the comparative experiment. This review is not supposed to
be exhaustive. Other methods can be found in [23], [24] and
reviews in [25], [26]. We describe linear list, Quad-tree and
one recent method, M-Front [19].
A. Linear List
1) General case: In this structure, a new point is compared
to all points in the list until a covering point is found or
all points are checked. The point is only added if it is nondominated w.r.t. all points in the list, that is in the worst case
we need to browse the whole list before adding a point. The
complexity in terms of number of points comparison is thus
in O(N ) with N the size in the list.
2) Biobjective case: sorted list: When only two objectives
are considered, we can use the following specific property: if
we sort the list according to one objective (let’s say the first),
the non-dominated list is also sorted according to the second
objective. Therefore, roughly speaking, updating the list can
be efficiently done in the following way. We first determine
the potential position i of the new candidate point in the sorted
list according to its value of the first objective, with a binary
search. If the new point is not dominated by the preceding
one in the list (if there is one), the new point is not dominated
and can be inserted at position i. If the new point has been
added, we need to check if there are some dominated points:
we browse the next points in the list, until a point is found that
has a better evaluation according to the second objective. All
the points found that have a worse evaluation according to the
second objective have to be removed since they are dominated
by the new point.
The worst-case complexity is still in O(N ) since it can
happen that a new point has to be compared to all the other
points (in the special case where we add a new point in the
first position and all the points in the sorted list are dominated
by this new point). But on average, experiments show that the
behavior of this structure for handling biobjective updating
problems is much better than the simple list.
The algorithm of this method is given in Algorithm 2
(for the sake of clarity, we only present the case where the
candidate point y has a distinct value for the first objective
compared to all the other points in the archive YN ).
Algorithm 2 Sorted list
Parameter l: A biobjective Pareto archive YN
Parameter ↓: New candidate point y
if YN = ∅ then
YN ← YN ∪ {y}
else
- -| Looking for the position i of y in YN
i ← BinarySearch(YN , y1 )
(i−1)
if (i = 0) or (y2 < y2
) then
- -| y is added at position i in YN
Insert(YN , y, i)
j ←i+1
while (j < |YN |) and (y2 ≤ y2j ) do
- -| y j is dominated
YN ← YN \{y j }
j ←j+1
B. Quad-tree
The use of Quad-tree for storing potentially non-dominated
points was proposed by Habenicht [27] and further developed
by Sun and Steuer [28] and Mostaghim and Teich [10].
In Quad-tree, points are located in both internal nodes and
leaves. Each node may have p2 children corresponding to
each possible combination of results of comparisons on each
objective where a point can either be better or not worse. In
the case of mutually non-dominated points p2 − 2 children are
possible since the combinations corresponding to dominating
or covered points are not used. Quad-tree allows for a fast
checking if a new point is dominated or covered. A weak point
of this data structure is that when an existing point is removed
its whole sub-tree has to be re-inserted to the structure. Thus,
removal of a dominated point is in general costly.
C. M-Front
M-Front has been proposed relatively recently by Drozdı́k
et al. [19]. The idea of of M-Front is as follows. Assume that
in addition to the new point y a reference point ref relatively
close to y and belonging to the Pareto archive YN is known.
The authors define two sets:
RSU (y, ref ) = {z ∈ YN | ∃k ∈ {1, . . . , p} : zk ≥ refk ∧ zk ≤ yk }
RSL (y, ref ) = {z ∈ YN | ∃k ∈ {1, . . . , p} : zk ≥ yk ∧ zk ≤ refk }
and prove that if a point z ∈ YN is dominated by y then it
belongs to RSL (y, ref ) and if z dominates y then it belongs
to RSU (y, ref ). Thus, it is sufficient to compare the new
points to sets RSL (y, ref ) and RSU (y, ref ) only. To find all
points with objective values in a certain interval M-Front uses
additional indexes one for each objective. Each index sorts the
Pareto archive according to one objective.
To find a reference point close to y, M-Front uses the kd tree data structure. The k-d tree is a binary tree, in which
each intermediate node divides the space into two parts based
on a value of one objective. While going down the tree
the algorithm cycles over particular objectives, selecting one
4
objective for each level. Drozdı́k et al. [19] suggest to store
references to points in leaf nodes only, while intermediate
nodes keep only split values.
IV. ND-T REE - BASED UPDATE
A. Presentation
In this section we present the main contribution of the
paper. The new method for updating a Pareto archive is based
on the idea of recursive division of archive YN into subsets
contained in different hyperrectangles. This division allows to
considerably reduce the number of comparisons to be made.
More precisely, consider a subset S ⊆ YN composed of
mutually non-dominated points and a new candidate point
y. Assume that some approximate local ideal zb∗ (S) and
approximate local nadir points zb∗ (S) of S are known. In
other words, all points in S are contained in the axes-parallel
hyperrectangle defined by zb∗ (S) and zb∗ (S).
We can define the following simple properties that allow to
compare a new point y to the whole set S:
Property 1. If y is covered by zb∗ (S), then y is covered by
each point in S and thus can be rejected. This property
is a straightforward consequence of the transitivity of the
coverage relation.
Property 2. If y covers zb∗ (S), then each point in S is covered
by y. This property is also a straightforward consequence
of the transitivity of the coverage relation.
Property 3. If y is non-dominated w.r.t. both zb∗ (S) and
zb∗ (S), then y is non-dominated w.r.t. each point in S.
Proof. If y is non-dominated w.r.t. zb∗ (S) then there is at
least one objective on which y is worse than zb∗ (S) and
thus worse than each point in S. If y is non-dominated
w.r.t. zb∗ (S) then there is at least one objective on which
y is better than zb∗ (S) and thus better than each point in
S. So, there is at least one objective on which y is better
and at least one objective on which y is worse than each
point in S.
If none of the above properties holds, i.e. y is neither
covered by zb∗ (S), does not cover zb∗ (S), nor is non-dominated
w.r.t. both zb∗ (S) and zb∗ (S), then all situations are possible, i.e.
y may either be non-dominated w.r.t. all points in S, covered
by some points in S or dominate some points in S. This can be
illustrated by showing examples of each of the situations. Consider for example a set S = {(1, 1, 1), (0, 2, 2), (2, 2, 0)} with
zb∗ (S) = z ∗ (S) = (0, 1, 0) and zb∗ (S) = z∗ (S) = (2, 2, 2). A
new point (1, 1, 0) dominates a point in S, a new point (1,
1, 2) is dominated (thus covered) by a point in S, and points
(0, 3, 0) and (2, 0, 1) are non-dominated w.r.t. all points in S.
The properties are graphically illustrated for the biobjective
case in Figure 1. As can be seen in this figure, in the
biobjective case, if y is covered by zb∗ (S) and y is nondominated w.r.t. zb∗ (S) then y is dominated by at least one
point in S. Note, however, that this does not hold in the case of
three and more objectives as shown in the above example - the
point (0, 3, 0) is covered by zb∗ (S) = (0, 1, 0), non-dominated
w.r.t. zb∗ (S) = (2, 2, 2) and (0, 3, 0) is not dominated by any
points in S.
Fig. 1. Comparison of a new point to all points in set S based on comparisons
to zb∗ (S) and zb∗ (S) only.
In fact it is possible to distinguish more specific situations
when none of the three properties hold, e.g. a situation when a
new point may be covered but cannot dominate any point, but
since we do not distinguish them in the proposed algorithm
we do not define them formally.
The above properties allow in some cases to quickly compare a new candidate point y to all points in set S without the
need for further comparisons to individual points belonging to
S. Such further comparisons are necessary only if none of the
three properties hold. Intuitively, the closer the approximate
local ideal and nadir points the more likely it is that further
comparisons can be avoided. To obtain close approximate local
ideal and nadir points we should:
• Split the whole set of non-dominated points into subsets
of points located close in the objective space.
• Have good approximations of the exact local ideal and
nadir points. On the other hand calculation of the exact
points may be computationally demanding and a reasonable approximation may assure the best overall efficiency.
Based on these properties, we can now define the ND-Tree
data structure.
Definition 8. ND-Tree data structure is a tree with the
following properties:
1) With each node n is associated a set of points S(n).
2) Each leaf node contains a list L(n) of points and S(n) =
L(n).
3) For each internal node n, S(n) is the union of disjoint
sets associated with all children of n.
4) Each node n stores an approximate ideal point zb∗ (S(n))
and approximate nadir point zb∗ (S(n)).
5) If n0 is a child of n, then zb∗ (S(n)) zb∗ (S(n0 )) and
zb∗ (S(n0 )) zb∗ (S(n)).
The algorithm for updating a Pareto archive with ND-Tree is
given in Algorithm 3. The idea of the algorithm is as follows.
We start by checking if the new point y is covered or nondominated w.r.t. all points in YN by going through the nodes
5
of ND-Tree and skipping children (and thus their sub-trees)
for which Property 3 holds. This procedure is presented in
Algorithm 4.
The new point is first compared to the approximate ideal
point (zb∗ (S(n)) and nadir point (zb∗ (S(n)) of the current node.
If the new point is dominated by zb∗ (S(n) it is immediately
rejected (Property 1). If zb∗ (S(n) is covered, the node and
its whole sub-tree is deleted (Property 2). Otherwise if
zb∗ (S(n)) y or y zb∗ (S(n)), the node needs to be analyzed.
If n is an internal node we call the algorithm recursively for
each child. If n is a leaf node, y may be dominated by or
dominate some points of n and it is necessary to browse the
whole list L(n) of the node n. If a point dominating y is
found, y is rejected, and if a point dominated by y is found,
the point is deleted from L(n).
If after checking ND-Tree the new point was found to
be non-dominated it is inserted by adding it to a close leaf
(Algorithm 5). To find a proper leaf we start from the root and
always select a child with closest distance to y. As a distance
measure we use the Euclidean distance to the middle point,
i.e. a point lying in the middle of line segment connecting
approximate ideal and approximate nadir points.
Once we have reached a leaf node, we add the point y to
the list L(n) of the node and possibly update the ideal and
nadir points of the node n (Algorithm 7). However, if the size
of L(n) became larger than the maximum allowed size of a
leaf set, we need to split the node into a predefined number of
children. To create children that contain points that are more
similar to each other than to those in other children, we use
a simple clustering heuristic based on Euclidean distance (see
Algorithm 6).
The approximate local ideal and nadir points are updated
only when a point is added. We do not update them when
point(s) are removed since it is a more complex operation.
This is why we deal with approximate (not exact) local ideal
and nadir points.
Algorithm 3 ND-Tree-based update
Parameter l: A Pareto archive YN organized as ND-Tree
Parameter ↓: New candidate point y
if YN = ∅ then
Create a leaf node n with L(n) = {y} and use it as a
root
else
n ← root node
if UpdateNode(n l,y ↓) then
Insert(n l,y ↓)
B. Comparison to existing methods
Like other methods ND-Tree-based update uses a tree structure to speed up the process of updating the Pareto archive.
The tree and its use is, however, quite different from Quad-tree
or k-d tree used in M-Front. For example, both Quad-tree or
k-d tree partition the objective space based on comparisons on
particular objectives, while in ND-Tree the space is partitioned
Algorithm 4 UpdateNode
Parameter l: A node n
Parameter ↓: New candidate point y
Parameter ↑: Boolean (True if y is not dominated by any
points in the tree of root n, False otherwise)
if zb∗ (S(n)) y then
- -| Property 1, y is rejected
return False
else if y zb∗ (S(n)) then
- -| Property 2
Remove n and its whole sub-tree
else if zb∗ (S(n)) y or y zb∗ (S(n)) then
if n is a leaf node then
for each z ∈ L(n) do
if z y then
return False
else if y z then
L(n) ← L(n)\{z}
else
for each Child n0 of n do
if not UpdateNode (n0 l,y ↓) then
return False
else
if n0 became empty then
Remove n0
if there is only one child n0 remaining then
Remove node n and use n0 in place of n
else
- -| Property 3
Skip this node
return True
Algorithm 5 Insert
Parameter l: A node n
Parameter ↓: New candidate point y
if n is a leaf node then
L(n) ← L(n) ∪ {y}
UpdateIdealNadir (n l,y ↓)
if Size of L(n) became larger than maximum size of
a leaf set then
Split (n l)
else
Find child n0 of n being closest to y
Insert(n0 l,y ↓)
based on the distances of points. Both Quad-tree or k-d tree
have strictly defined degrees. In k-d tree it is always two
(binary tree) while in Quad-tree it depends on the number
of objectives. In ND-Tree the degree is a parameter. In Quadtree the points are kept in both internal nodes and leaves, while
ND-Tree keeps points in leaves only. In M-Front k-d tree is
used to find an approximate nearest neighbor and the Pareto
archive is updated using other structures (sorted lists for each
objective). In our case, ND-Tree is the only data structure used
6
Algorithm 6 Split
Parameter l: A node n
Find the point z ∈ L(n) with the highest average Euclidean
distance to all other points in L(n)
Create a new child n0 with list set L(n0 ) = {z}
L(n) ← L(n)\{z}
UpdateIdealNadir (n0 l,z ↓)
while The required number of children are not created do
Find the point z ∈ L(n) with the highest average
Euclidean distance to all points in all children of n
0
0
Create a new child n with an empty list set L(n )
L(n0 ) ← L(n0 ) ∪ {z}
UpdateIdealNadir (n0 l,z ↓)
L(n) ← L(n)\{z}
while L(n) is not empty do
z ← first point in L(n)
Find child n0 of n being closest to z
L(n0 ) ← L(n0 ) ∪ {z}
UpdateIdealNadir (n0 l,z ↓)
L(n) ← L(n)\{z}
Algorithm 7 UpdateIdealNadir
Parameter l: A node n
Parameter ↓: New candidate point y
Check in any component of y is lower than corresponding component in zb∗ (S(n)) or greater than corresponding
component in zb∗ (S(n)) and update the points if necessary
if zb∗ (S(n)) or zb∗ (S(n)) have been changed then
if n is not a root then
np ← parent of n
UpdateIdealNadir (np l,y ↓)
Fig. 2. Example of fully unbalanced ND-Tree.
In this case:
T (N ) = 4 + T (N − 1)
where N is the number of points in the archive and T (N ) is
the number of point comparisons needed to update an archive
of size N . Term 4 appears because we check two children
and for each child approximate ideal and nadir points are
compared. Solving the recurrence we get:
T (N ) = 4 + 4N
(2)
Thus, the algorithm has O(N ) time complexity.
We are not aware of any result showing that the worstcase time complexity of any algorithm for the dynamic nondominance problem may be lower than O(N ) in terms of point
comparisons. So, our method does not improve the worstcase complexity but according to our experiments performs
significantly better in practical cases.
2) Best case: Assume first that the candidate point is not
covered by any point in YN . In the optimistic case, at each
intermediate node the points are equally split into predefined
number of children and there is only one child that has to be
further processed (i.e. there is only one child for which none
of the three properties hold). In fact we could consider even
more optimistic distribution of points when the only node that
has to be processed contains just one point, but the equal split
is much more realistic assumption. In this case:
T (N ) = 2C + T (N/C) = Θ(logC N )
by the algorithm.
C. Computational complexity
1) Worst case: The worst case for the UpdateNode and
Insert algorithm is when we need to compare the new point
to each intermediate node of the tree. For example, consider
the following particular case: two objectives, ND-Tree with
maximum leaf size equal to 2, 2 children, and constructed by
processing the following list of points (with K ∈ N∗ ): (0, 0),
(1, -1), (2K , −2K ), (2K−1 , −2K−1 ),..., (4, -4), (2, -2). This
list of points is constructed in such a way that the first two
points are put in one leftmost leaf and the third point creates
a separate leaf. Then each further point is closer to the child
on the left side but finally after splitting the leftmost node the
new point creates a new leaf. The ND-Tree obtained is shown
in Figure 2.
Consider now that the archive is updated with point (0.5,
-0.5). This point will need to be compared to all N − 1
intermediate nodes and then to both points in leftmost leaf.
(1)
(3)
where C is the number of children. If the candidate point is
covered by a point in YN the UpdateNode algorithm may
stop even earlier and there will be no need to run Insert
algorithm.
3) Average case: Analysis and even definition of average
case for such complex algorithms is quite difficult. The simplest case to analyze is when each intermediate node has two
children which allows us to follow the analysis of well-known
algorithms like binary search or Quicksort. If a node has N
points then one of the two children may have 1, . . . , N − 1
points and the other child the remaining number of points.
Assuming that each split has equal probability and only one
child is selected:
T (N ) = 4 +
N −1
1 X
T (k)
N
(4)
k=1
Multiplying both sides by N :
N T (N ) = 4N +
N
−1
X
k=1
T (k)
(5)
7
Assuming that N ≥ 2:
(N − 1)T (N − 1) = 4(N − 1) +
N
−2
X
T (k)
(6)
k=1
Subtracting equations 5 and 6:
N T (N ) − (N − 1)T (N − 1) =
4N − 4(N − 1) +
N
−1
X
T (k) −
k=1
Simplifying:
T (N ) =
N
−2
X
T (k)
(7)
k=1
4
+ T (N − 1)
N
(8)
The main part of Split algorithm has constant time
complexity since it depends only on the maximum size of
a leaf set which is a constant parameter of the algorithm.
UpdateIdealNadir algorithm goes up the tree starting
from a leaf which is equivalent to going down the tree and
selecting just one child. So, its analysis is exactly the same as
of Insert algorithm.
We also need to consider the complexity of the operation
of removal of node n and its sub-tree. In the worst case,
the removed node is the root, and thus all N point need to
be removed. Such situation is very unlikely, since it happens
when the new point dominates all points in the current archive.
Typically, the new point will dominate only few points.
Solving this recurrence:
T (N ) = 4HN − 1
where HN is N-th harmonic number. Using well-known properties of harmonic numbers we get T (N ) = Θ(log N ).
We can expect, however, that in realistic cases more than
one child will need to be further processed in UpdateNode
algorithm because the candidate point may cover approximate
nadir points or may be covered by approximate ideal points of
more than one child. Assume that the probability of selecting
both children is equal to c1 . Then:
T (N ) = 4 +
1
(1 + c1 )
N
N
−1
X
T (k)
(10)
k=1
Following the above reasoning we get:
N + c1
2
+
T (N − 1)
N
N
Solving this recurrence, we obtain:
T (N ) =
T (N ) =
Γ(c1 + N + 1)
2
−
Γ(N + 1)
c1
(that can be checked by substituting (12) in (11)).
Since
Γ(N + α)
lim
=1
N →∞ Γ(N )N α
(11)
(12)
(13)
We have:
T (N ) = Θ(N c1 )
V. C OMPUTATIONAL EXPERIMENTS
(9)
(14)
and the algorithm remains sub-linear for any c1 < 1. In the
worst case, both children need to be selected, so c1 = 1 and
T(N ) = Θ(N ) which confirms the analysis presented above.
This analysis may give only approximate insight into behaviour of the algorithm since in UpdateNode algorithm c1
will not be constant at each level. We may rather expect that
while going down the tree from the root towards leaves the
probability that two children will need to be processed will
decrease because approximate nadir and ideal points of the
children will lie closer. Anyway, this analysis shows that the
performance of UpdateNode algorithm may be improved
by decreasing the probability that a child has to be processed.
This is why we try to locate in one node points lying close to
each other in Insert and Split algorithms.
In Insert algorithm always one child is processed, so the
time complexity remains Θ(log N ) in average case.
We will show results obtained with ND-Tree and other
methods in two different cases:
A) Results for artificially generated sets which allow us to
easily control the number of points in the sets and the
quality of the points.
B) Results for sets generated by a MOEA, namely
MOEA/D [2] for the multiobjective knapsack problem.
We compare the simple list, sorted list (biobjective case),
Quad-tree, M-Front and ND-Tree for these sets according to
the CPU time [ms]. To avoid the influence of implementation
details all methods were implemented from the scratch in C++
in as much homogeneous way as possible, i.e. when possible
the same code was used to perform the same operations like
Pareto dominance checks.
For the implementation of Quad-tree, we use Quad-tree2
version as described by Mostaghim and Teich [10].
For M-Front, we use as much as possible the description
found in the paper of Drozdı́k et al. [19]. However the authors
do not give the precise algorithm of k-d tree used in their
method. In our implementation of k-d tree, when a leaf is
reached a new division is made using the average value of
the current level objective. The split value is average between
the value of new point and the point in the leaf. Also like in
Drozdı́k et al. [19] the approximate nearest neighbor is found
exactly as in the standard exact nearest neighbor search, but
only four evaluations of the distance are allowed. Note that
at https://sites.google.com/site/ndtreebasedupdate/ we present
results of an additional experiment showing that for a higher
number of objectives the details of the implementation of k-d
tree do not have any substantial influence on the running time.
We also noticed that a number of elements of M-Front
can be improved to further reduce the running time. The
improvement is in our opinion significant enough to call
the new method M-Front-II. In particular in some cases MFront-II was several times faster than original M-Front in our
computational experiments. The modifications we introduced
are as follows:
•
In original M-Front the sets RSL (y, ref ) and
RSU (y, ref ) are built explicitly and only then the
points contained in these sets are compared to y.
In M-Front-II we build them only implicitly, i.e. we
8
immediately compare the points that would be added to
the sets to y.
• We analyze the objectives in such a way that we start
with objectives for which refk ≤ yk . In other words, we
start with points from set RSU (y, ref ). Since many new
points are dominated this allows to stop the search immediately when a point dominating y is found. Note that
a similar mechanism is in fact used in original M-Front
but only after the sets RSL (y, ref ) and RSU (y, ref ) are
explicitly built.
• The last modification is more technical. M-Front uses
linked lists (std::list in C++) to store the indexes
and a hash-table (std::unordered_map in C++) to
link points with their positions in these lists. We observed,
however, that even basic operations like iterating over the
list are much slower with linked lists than with static or
dynamic arrays (like std::vector in C++). Thus we
use dynamic arrays for the indexes. In this case, however,
there is no constant iterator that could be used to link the
points with their positions in these indexes. So, we use
a binary search to locate a position of a point in the
sorted index whenever it is necessary. The overhead of
the binary search is anyway smaller than the savings due
to the use of faster indexes.
For ND-Tree we use 20 as the maximum size of a leaf and
p+1 as the number of children. These values of the parameters
were found to perform well in many cases. We analyze the
influence of these parameters later in this section.
The code, as well as test instances and data sets, are available at https://sites.google.com/site/ndtreebasedupdate/. All of
the results have been obtained on an Intel Core i7-5500U CPU
at 2.4 GHz.
A. Artificial sets
1) Basic, globally convex sets: The artificial sets are composed of n points with p objectives to minimize. The sets
are created as follows. We generate randomly P
n points y i in
p
p
{0, . . . , Vmax } with the following constraint: k=1 (Vmax −
i 2
2
yk ) ≤ Vmax . With this constraint, all the non-dominated
points will be located inside the hypersphere with the center
at (Vmax , . . . , Vmax ) and with a radius of length equal to
Vmax . In order to control the quality P
of the generated points,
p
i 2
we also add a quality constraint:
≥
i=k (Vmax − yk )
2
(1 − ) ∗ Vmax . In this way, with a small , only high-quality
points will be generated. We believe that it is a good model
for points generated by real MOEAs since a good MOEA
should generate points lying close to the true Pareto front. The
hypersphere is a model of the true Pareto front and parameter
controls the maximum distance from the hypersphere. We
have generated data sets composed of 100 000 and 200 000
points, with Vmax = 10000, and for p = 2 to 10. In the main
experiment we use data sets with 100 000 points because for
the larger sets running times of some methods became very
long. Also because of very high running times for sets with
many objectives, in the main experiment we used sets with
up to 6 objectives. For each value of p, five different quality
levels are considered: quality q1, = 0.5 ; q2, = 0.25 ;
q3, = 0.1 ; q4, = 0.05 ; q5, = 0.01. The fraction of
non-dominated points grows both with increasing quality and
number of objectives and in extreme cases all points may be
non-dominated (see Table I).
TABLE I
N UMBERS OF NON - DOMINATED POINTS IN ARTIFICIAL SETS
p
Quality
2
2
2
2
2
3
3
3
3
3
4
4
4
4
4
5
5
5
5
5
6
6
6
6
6
q1
q2
q3
q4
q5
q1
q2
q3
q4
q5
q1
q2
q3
q4
q5
q1
q2
q3
q4
q5
q1
q2
q3
q4
q5
|YN |
convex
519
713
1046
1400
2735
4588
6894
12230
19095
53813
14360
21680
39952
64664
98283
28944
42246
77477
96002
99975
45879
65195
96687
99788
100000
|YN |
non-convex
379
613
1037
1454
2748
2587
5344
11497
18648
53255
6853
16420
37709
63140
98243
13437
34357
75796
95867
99975
22956
57966
96480
99786
100000
|YN |
clustered
449
552
785
1059
1781
3729
5514
9720
15502
44173
11963
18120
35460
57725
97137
23966
38028
71063
93842
98521
40483
61096
94978
99652
99999
2) Globally non-convex sets: In order to test whether the
global convexity of the above sets influences the behavior of
the tested methods we have also generated sets whose Pareto
fronts are globally non-convex. They were obtained by simply
changing the sign of each objective in the basic sets.
3) Clustered sets: In these sets, the points are located
in small clusters. We have generated sets composed of 100
clusters, where each cluster contains 1000 points (the sets are
thus composed of 100 000 points). The sets have been obtained
as follows: we start with the 200 000 points from the basic
convex sets. We select from the set one random point, and
we then select the 999 points closest to this point (using the
Euclidean distance). We repeat this operation 100 times to
obtain the 100 clusters of the sets.
The shapes of exemplary biobjective globally convex, globally non-convex and clustered data sets can be seen at https:
//sites.google.com/site/ndtreebasedupdate/.
Each method was run 10 times for each set, with the
points processed in a different random order for each run.
The average running times for basic sets are presented in
Figures 3 to 7. We use average values since the different values
were generally well distributed around the average with small
deviations. Please note that because of large differences the
running time is presented in logarithmic scale.
In addition, in Figure 8 we illustrate the evolution of the
running times according to the number of objectives for the
data sets of intermediate quality q3. In this case we use sets
with up to 10 objectives. With 7 and more objectives even
the sets of intermediate quality q3 contain almost only nondominated points (see Table I). This is why the running times
9
of the simple list are practically constant for 7 and more
objectives, because the simple list boils down in this case to
the comparison of each new point to all points in the list. The
running times of all other methods including ND-Tree increase
with a growing number of objectives, but ND-Tree remains 5.5
times faster than the second best method (Quad-tree) for sets
with 10 objectives.
Furthermore, we measured the number of comparisons
of points with the dominance relation for the data sets of
intermediate quality q3 with p = 2, ..., 10. For ND-Tree-based
Fig. 6. CPU time (logarithmic scale) for five-objective convex data sets.
Fig. 3. CPU time (logarithmic scale) for biobjective convex data sets.
Fig. 7. CPU time (logarithmic scale) for six-objective convex data sets.
Fig. 4. CPU time (logarithmic scale) for three-objective convex data sets.
Fig. 8. CPU time (logarithmic scale) vs. number of objectives for convex
data sets of quality q3.
TABLE II
P OINT COMPARISONS PER MS
Method
List
M-Front
M-Front-II
ND-Tree
Quad-tree
Fig. 5. CPU time (logarithmic scale) for four-objective convex data sets.
Comparisons per ms
26 752
3 476
8 370
17 733
9 040
10
update it includes also comparisons to the approximate ideal
and nadir points and for Quad-tree comparisons to sub-nodes.
The results are presented in Figure 9. Please note that the
results for the two versions of M-Front overlap in the Figure.
Fig. 10. CPU time (logarithmic scale) for three-objective non-convex data
sets.
Fig. 9. Number of point comparisons (logarithmic scale) vs. number of
objectives for convex data sets of quality q3.
The differences in running times cannot be fully explained
by the number of point comparisons because the methods
differ significantly in the number of point comparisons per milliseconds (see Table II). This ratio is highest for the simple list
because this method performs very few additional operations.
It is also relatively high for ND-Tree. Other methods perform
many other operations than point comparisons that strongly
influence their running times. This is particularly clear in
comparison of M-Front and M-Front-II. These method perform
the same number of point comparisons, but M-Front-II is
several times faster than M-Front because in the latter method
the sets RSL (y, ref ) and RSU (y, ref ) are built explicitly
and this method uses slower linked lists. Overall, ND-Tree
performs fewest number of point comparisons for data sets
with p ≥ 4. These results indicate that ND-Tree-based update
substantially reduces the number of comparisons with respect
to the simple list. For example for p = 10, all points in
the data set are non-dominated, thus on average each of the
100000 new points has to be compared to an archive composed
of 49999.5 points, while with ND-Tree-based update it only
requires 2029 point comparisons on average.
The results obtained for non-convex and clustered sets were
very similar to the results with basic sets. Thus, in Figures 10
to 13 we show only exemplary results for the three- and sixobjective cases. These results indicate that the running times of
the tested methods are not substantially affected by the global
shape of the Pareto front.
4) Discussion of the results for artificial sets: The main
observations from this experiment are:
• ND-Tree performs the best in terms of CPU time for
all test sets with three and more objectives. In some
cases the differences to other methods are of two orders
of magnitude and in some cases the difference to the
second best method is of one order of magnitude. NDTree behaves also very predictably, its running time
grows slowly with increasing number of objectives and
increasing fraction of non-dominated points.
Fig. 11. CPU time (logarithmic scale) for six-objective non-convex data sets.
Fig. 12. CPU time (logarithmic scale) for three-objective clustered data sets.
11
in this case and 10 runs for each method (see Figure 14). The
CPU time is cumulative time of processing a given number
of points. In addition, since the running time is much smaller
for ND-Tree its results are presented in Figure 15 separately.
We see that ND-Tree is the fastest method for any number of
points and its cumulative running time grows almost linearly
with the number of points. In other words, time of processing
a single point is almost constant. Please note that unlike in
other figures the linear scale is used in these two figures in
order to make them more informative.
Fig. 13. CPU time (logarithmic scale) for six-objective clustered data sets.
•
•
•
•
•
For biobjective instances sorted list is the best choice. In
this case, M-Front and M-Front-II also behave very well
since they become very similar to sorted list.
The simple list obtains its best performances for data
sets with many dominated points like p = 2 with lowest
quality. In this case the new point is often dominated by
many points, so the search process is quickly stopped
after finding a dominating point.
Quad-tree performs very badly for data sets with many
dominated points, e.g. in biobjective instances where it is
the worst method in all cases. In this case, many points
added to Quad-tree are then removed and the removal of
a point from Quad-tree is a costly operation. As discussed
above when an existing point is removed its whole subtree has to be re-inserted to the structure. On the other
hand, it is the second best method for most data sets with
six and more objectives.
M-Front-II is much faster than M-Front on data sets with
larger fraction of dominated points. In this case, M-FrontII may find a dominating point faster without building
explicitly the whole sets RSL (y, ref ) and RSU (y, ref ).
The performance of both M-Front and M-Front-II deteriorates with an increasing number of objectives. With six
and more objectives M-Front is the slowest method in all
cases. Intuitively this can be explained by the fact that
M-Front (both versions) uses each objective individually
to reduce the search space. In the case of two objectives
the values of one objective carry a lot of information
since the order on one objective induces also the order on
the other one. The more objectives, the less information
we get from an order on one of them. Furthermore,
sets RSL (y, ref ) and RSU (y, ref ) are in fact unions of
corresponding sets for particular objectives, which also
results in their growth. Finally, in many objective case, a
reference point close on Euclidean distance does not need
to be very close on each objective, since it will rather have
a good balance of differences on many coordinates.
In an additional experiment we analyzed the evolution of
the running times of all methods with increasing number of
points. We decided to use one intermediate globally convex
data set with p = 4 and quality q3. We used 200 000 points
Fig. 14. CPU time (linear scale) vs. the number of points for four-objective
convex data sets of quality q3.
Fig. 15. CPU time (linear scale) vs. the number of points for ND-Tree only
(four-objective convex data sets of quality q3).
ND-Tree has two parameters - the maximum size (number
of points) of a leaf, and the number of children, so the question
arises how sensitive it is to the setting of these parameters. To
study it we again use the intermediate data set with p = 4
and quality q3 and run ND-Tree with various parameters, see
Figure 16. Please note that number of children cannot be larger
than the maximum size of a leaf +1 since after exceeding
the maximum size the leaf is split into the given number of
children. We see that ND-Tree is not very sensitive to the two
parameters: the CPU time remains between about one and
three seconds regardless of the values of the parameters. The
best results are obtained with 20 for the maximum size of a
leaf and with 6 for the number of children.
12
TABLE III
N UMBERS OF NON - DOMINATED POINTS IN SETS GENERATED BY MOEA/D
p
2
3
4
5
6
|YN |
140
1789
5405
10126
16074
Fig. 16. ND-Tree CPU time for different values of parameters (four-objective
convex data sets of quality q3). Data series correspond to maximum leaf size.
CPU time in linear scale.
In our opinion the results confirm that ND-Tree performs
relatively well for a wide range of the values of the parameters.
In fact, it would remain the best method for this data set with
any of the parameters settings tested.
Fig. 17. CPU time (logarithmic scale) on MOEA/D sets.
B. Sets generated by MOEA/D
In order to test if the observations made for artificial sets
hold for sets generated by real evolutionary algorithms, we use
sets of points generated by well-known MOEA/D algorithm
[2] for multiobjective multidimensional knapsack problem
instances with 2 to 6 objectives. We used the code available
at http://dces.essex.ac.uk/staff/zhang/webofmoead.htm [2]. We
used the instances with 500 items available with the code with
2 to 4 objectives. The profits and weights of these instances
were randomly generated uniformly between 1 and 100. We
have generated ourselves the 5 and 6 objectives instances
by adding random profits and weights, between 1 and 100.
MOEA/D was run for at least 100 000 iterations and the first
100 000 points generated by the algorithm were stored for the
purpose of this experiment. The numbers of non-dominated
points are given in Table III.
Figure 17 presents running times of each of the tested
methods as well the running times of MOEA/D excluding
the time needed to update the Pareto archive. These results
confirm that the observations made for artificial sets also hold
in the case of real sets. ND-Tree is the fastest method for three
and more objectives. Quad-tree performs particularly badly
in the biobjective case. Both versions of M-Front relatively
deteriorate with a growing number of objectives. Furthermore,
these results show that the time of updating the Pareto archive
may be higher than the remaining running time of MOEA/D.
In particular, for the six-objective instance the running time
of M-Front is 5 times higher than the remaining running time
of MOEA/D. The running time of the simple list, Quad-tree
and M-Front-II are comparable to the remaining running time
of MOEA/D, and only the running time of ND-Tree is 10
times shorter. This confirms that the selection of an appropriate
method for updating the Pareto archive may have a crucial
influence on the running time of a MOEA.
We have also generated sets of points by applying another MOMH, namely Pareto local search [18] to solve
the multiobjective traveling salesman problem (MOTSP).
These results can be found at https://sites.google.com/site/
ndtreebasedupdate/ (the same conclusions apply).
VI. ND-T REE - BASED NON - DOMINATED SORTING
ND-Tree-based update may also be applied to the problem
of the non-dominated sorting. This problem arises in the nondominated sorting-based MOEAs, e.g. NSGA-III [29] where a
population of solutions needs to be assigned to different fronts
based on their dominance relationships.
We solve the non-dominated sorting problem in the very
straightforward way, i.e. we find the first front by updating
an initially empty Pareto archive with each point in the
population. Then the points from the first front are removed
from the population and the next front is found in the same
way. This process is repeated until all fronts are found.
We compare this approach to some recently proposed nondominated sorting algorithms, i.e. ENS-BS/SS [30] and DDANS [31]. ENS-BS/SS algorithm sorts the points lexicographically based on the values of objectives. Then it considers each
solution using this order to efficiently find the last front that
contains a point dominating the considered point. For each
front this method in fact solves the dynamic non-dominance
problem with the simple list and if the considered solution
is non-dominated within this front it needs to be compared
to each solution in this front. If there is just one front in the
population, the method boils down to solving the dynamic nondominance problem with the simple list and requires O(N 2 )
comparisons. DDA-NS algorithm sorts the population according to each objective which requires O(N log N ) objective
function comparisons and builds comparison matrices for each
13
objectives. Then it uses some matrix operations which in
general have O(N 2 ) complexity to build the fronts. We also
compare our algorithm to M-Front-II [19] applied in the same
way as ND-Tree-based update. Please note that similarly to
what was done in [31] we apply M-Front-II for finding each
front, while in [19] only the first front was found by M-Front.
Also, like [30], [31] we used populations of size 5000.
We used both random populations drawn with uniform probability from a hypercube, and populations composed of 5000
randomly selected points from our data sets of intermediate
quality q3. For each number of objectives 10 populations were
drawn.
The results are presented in Figures 18 to 21. We report
both CPU times and the number of comparisons of points.
We also show the number of fronts on the right y-axis. Please
note that we do not show the number of comparisons for
DDA-NS, since this algorithm does not explicitly compares
points with the dominance relation. The slowest algorithm is
DDA-NS. Our results are quite contradictory to the results
reported in [31] where DDA-NS is the fastest algorithm in
most cases. Please note, however, that in [31] the algorithms
were implemented in MATLAB which, as the authors note,
is very efficient in matrix operations. On the other hand, the
CPU times reported in [31] are of orders of magnitude higher
compared to our experiment which suggests rather that the
MATLAB implementation of M-Front and ENS-BS/SS is quite
inefficient.
Fig. 19. #comparisons in non-dominated sorting algorithms for artificial sets
with quality q3.
Fig. 20. CPU times of non-dominated sorting algorithms for random
populations.
Fig. 18. CPU times of non-dominated sorting algorithms for artificial sets
with quality q3.
ENS-BS/SS performs well for populations with many fronts
but its performance deteriorates when the number of fronts
is reduced. As it was mentioned above if there is just one
front in the population, the method boils down to the dynamic
non-dominance problem solved with the simple list. This is
why in the case of populations drawn from our sets which
contain points of higher quality than random populations, and
for higher numbers of objectives, where the populations often
contain just one front, the number of comparisons saturates at
a constant level.
In most cases ND-Tree-based non-dominated sorting is the
most efficient method in terms of both CPU time and the
number of comparisons. These results are very promising
but only preliminary and further experiments, especially with
Fig. 21. #comparisons in non-dominated sorting algorithms for random
populations.
14
populations generated by real MOEAs, are necessary. Please
note that as suggested in [19] the practical efficiency of the
non-dominated sorting in the context of a MOEA may be
further improved by maintaining the first front in a Pareto
archive.
VII. C ONCLUSIONS
We have proposed a new method for the dynamic nondominance problem. According to the theoretical analysis the
method remains sub-linear with respect to the number of
points in the archive under mild assumptions and the time
of processing a single point observed in the computational
experiments are almost constant. The results of computational
experiments with both artificial data sets of various global
shapes, as well as results with sets of points generated by two
different multiobjective methods, i.e. MOEA/D evolutionary
algorithm and Pareto local search metaheuristic, indicate that
the proposed method outperforms competitive methods in the
case of three- and more objective problems. In biobjective case
the best choice remains the sorted list.
We believe that with the proposed method for updating
a Pareto archive, new state-of-the art results could be obtained for many multiobjective problems with more than two
objectives. Indeed, results of our computational experiment
indicate that the choice of an appropriate method for updating
the Pareto archive may have crucial influence on the running
time of multiobjective evolutionary algorithms and other metaheuristics especially in the case of higher number of objectives.
Interesting directions for further research are to adapt NDTree to be able to deal with archives of a relatively large but
bounded size or to solve the static non-dominance problem.
We have also obtained promising results by applying NDTree-based update to the non-dominated sorting and comparing it to some state-of-the-art algorithms for this problem.
Further experiments are, however, necessary especially with
populations generated by real MOEAs. Furthermore, good
results of ENS algorithm for populations with many fronts
suggest that a combination of ENS with ND-Tree may be an
interesting direction for further research.
ACKNOWLEDGMENT
The research of Andrzej Jaszkiewicz was funded by
the the Polish National Science Center, grant no. UMO2013/11/B/ST6/01075.
R EFERENCES
[1] O. Schütze, “A new data structure for the nondominance problem in
multi-objective optimization,” in Proceedings of the 2nd International
Conference on Evolutionary Multi-criterion Optimization, ser. EMO’03.
Berlin, Heidelberg: Springer-Verlag, 2003, pp. 509–518.
[2] Q. Zhang and H. Li, “MOEA/D: A multiobjective evolutionary algorithm
based on decomposition,” Evolutionary Computation, IEEE Transactions
on, vol. 11, no. 6, pp. 712–731, 2007.
[3] E. Zitzler, M. Laumanns, and L. Thiele, “SPEA2: Improving the Strength
Pareto Evolutionary Algorithm,” Computer Engineering and Networks
Laboratory (TIK), Swiss Federal Institute of Technology (ETH) Zurich,
Zurich, Switzerland, Tech. Rep. 103, May 2001.
[4] C. A. C. Coello, G. T. Pulido, and M. S. Lechuga, “Handling multiple
objectives with particle swarm optimization,” IEEE Transactions on
Evolutionary Computation, vol. 8, no. 3, pp. 256–279, June 2004.
[5] S. Agrawal, B. K. Panigrahi, and M. K. Tiwari, “Multiobjective particle
swarm algorithm with fuzzy clustering for electrical power dispatch,”
IEEE Transactions on Evolutionary Computation, vol. 12, no. 5, pp.
529–541, Oct 2008.
[6] B. Li, K. Tang, J. Li, and X. Yao, “Stochastic ranking algorithm
for many-objective optimization based on multiple indicators,” IEEE
Transactions on Evolutionary Computation, vol. 20, no. 6, pp. 924–938,
Dec 2016.
[7] X. Cai, Y. Li, Z. Fan, and Q. Zhang, “An external archive guided multiobjective evolutionary algorithm based on decomposition for combinatorial optimization,” IEEE Transactions on Evolutionary Computation,
vol. 19, no. 4, pp. 508–523, Aug 2015.
[8] S. Yang, M. Li, X. Liu, and J. Zheng, “A grid-based evolutionary
algorithm for many-objective optimization,” IEEE Transactions on Evolutionary Computation, vol. 17, no. 5, pp. 721–736, Oct 2013.
[9] L. Tang and X. Wang, “A hybrid multiobjective evolutionary algorithm
for multiobjective optimization problems,” IEEE Transactions on Evolutionary Computation, vol. 17, no. 1, pp. 20–45, Feb 2013.
[10] S. Mostaghim and J. Teich, “Quad-trees: A data structure for storing
Pareto sets in multiobjective evolutionary algorithms with elitism,” in
EMO, Book Series: Advanced Information and Knowledge, A. Abraham,
L. Jain, and R. Goldberg, Eds. Springer-Verlag, 2004, pp. 81–104.
[11] J. Fieldsend, R. Everson, and S. Singh, “Using unconstrained elite
archives for multiobjective optimization,” IEEE Transactions on Evolutionary Computation, vol. 7, no. 3, pp. 305–323, June 2003.
[12] W. Brockhoff, Theory of Randomized Search Heuristics: Foundations
and Recent Developments. World Sc. Publ. Comp., 2010, ch. Theoretical aspects of evolutionary multiobjective optimization, pp. 101–139.
[13] H. T. Kung, F. Luccio, and F. P. Preparata, “On finding the maxima of
a set of vectors,” J. ACM, vol. 22, no. 4, pp. 469–476, Oct. 1975.
[14] H. N. Gabow, J. L. Bentley, and R. E. Tarjan, “Scaling and related
techniques for geometry problems,” in Proceedings of the Sixteenth
Annual ACM Symposium on Theory of Computing, ser. STOC ’84. New
York, NY, USA: ACM, 1984, pp. 135–143.
[15] F. P. Preparata and M. I. Shamos, Computational Geometry: An Introduction. New York, NY, USA: Springer-Verlag New York, Inc., 1985.
[16] P. Gupta, R. Janardan, M. Smid, and B. Dasgupta, “The rectangle enclosure and point-dominance problems revisited,” International Journal
of Computational Geometry & Applications, vol. 7, no. 5, pp. 437–455,
1997.
[17] K. Deb, M. Mohan, and S. Mishra, “Evaluating the -domination
based multi-objective evolutionary algorithm for a quick computation
of pareto-optimal solutions,” Evol. Comput., vol. 13, no. 4, pp. 501–
525, Dec. 2005.
[18] L. Paquete, M. Chiarandini, and T. Stützle, “Pareto local optimum sets
in the biobjective traveling salesman problem: an experimental study,” in
Metaheuristics for Multiobjective Optimisation, X. Gandibleux, M. Sevaux, K. Sörensen, and V. T’kindt, Eds. Berlin: Springer. Lecture Notes
in Economics and Mathematical Systems Vol. 535, 2004, pp. 177–199.
[19] M. Drozdı́k, Y. Akimoto, H. Aguirre, and K. Tanaka, “Computational
cost reduction of nondominated sorting using the M-front,” IEEE Transactions on Evolutionary Computation, vol. 19, no. 5, pp. 659–678, 2015.
[20] K. Deb, A. Pratap, S. Agarwal, and T. Meyarivan, “A fast and elitist
multiobjective genetic algorithm: NSGA-II,” IEEE Transactions on
Evolutionary Computation, vol. 6, no. 2, pp. 182–197, Apr 2002.
[21] M. Jensen, “Reducing the run-time complexity of multiobjective EAs:
The NSGA-II and other algorithms,” IEEE Transactions on Evolutionary
Computation, vol. 7, no. 5, pp. 503–515, Oct 2003.
[22] F.-A. Fortin, S. Grenier, and M. Parizeau, “Generalizing the improved
run-time complexity algorithm for non-dominated sorting,” in Proceedings of the 15th Annual Conference on Genetic and Evolutionary
Computation. New York, NY, USA: ACM, 2013, pp. 615–622.
[23] O. Schütze, S. Mostaghim, M. Dellnitz, and J. Teich, “Covering Pareto
sets by multi-level evolutionary subdivision techniques,” in Proceedings
of the Second Int. Conf. on Evolutionary Multi-Criterion Optimization.
Berlin: Springer, 2003, pp. 118–132.
[24] J. E. Fieldsend, R. M. Everson, and S. Singh, “Using unconstrained
elite archives for multiobjective optimization,” IEEE Transactions on
Evolutionary Computation, vol. 7, no. 3, pp. 305–323, 2003.
[25] N. Altwaijry and M. Bachir Menai, “Data structures in multi-objective
evolutionary algorithms,” Journal of Computer Science and Technology,
vol. 27, no. 6, pp. 1197–1210, 2012.
[26] S. Mostaghim, J. Teich, and A. Tyagi, “Comparison of data structures
for storing Pareto-sets in MOEAs,” in Evolutionary Computation, 2002.
CEC ’02., vol. 1, May 2002, pp. 843–848.
[27] W. Habenicht, Essays and Surveys on MCDM: Proceedings of the
Fifth International Conference on Multiple Criteria Decision Making,
15
[28]
[29]
[30]
[31]
Mons, Belgium, August 9–13, 1982. Springer, 1983, ch. Quad Trees, a
Datastructure for Discrete Vector Optimization Problems, pp. 136–145.
M. Sun and R. Steuer, “Quad trees and linear list for identifying nondominated criterion vectors,” INFORM Journal on Computing, vol. 8,
no. 4, pp. 367–375, 1996.
K. Deb and H. Jain, “An evolutionary many-objective optimization
algorithm using reference-point-based nondominated sorting approach,
part i: Solving problems with box constraints,” IEEE Transactions on
Evolutionary Computation, vol. 18, no. 4, pp. 577–601, Aug 2014.
X. Zhang, Y. Tian, R. Cheng, and Y. Jin, “An efficient approach
to nondominated sorting for evolutionary multiobjective optimization,”
IEEE Transactions on Evolutionary Computation, vol. 19, no. 2, pp.
201–213, April 2015.
Y. Zhou, Z. Chen, and J. Zhang, “Ranking vectors by means of
the dominance degree matrix,” IEEE Transactions on Evolutionary
Computation, vol. 21, no. 1, pp. 34–51, Feb 2017.
| 8cs.DS
|
LOCAL ASYMPTOTIC EQUIVALENCE OF PURE QUANTUM STATES ENSEMBLES
AND QUANTUM GAUSSIAN WHITE NOISE
arXiv:1705.03445v1 [math.ST] 9 May 2017
CRISTINA BUTUCEA1 , MĂDĂLIN GUŢĂ2 , MICHAEL NUSSBAUM3
Abstract. Quantum technology is increasingly relying on specialised statistical inference methods for analysing
quantum measurement data. This motivates the development of “quantum statistics”, a field that is shaping up
at the overlap of quantum physics and “classical” statistics. One of the less investigated topics to date is that
of statistical inference for infinite dimensional quantum systems, which can be seen as quantum counterpart
of non-parametric statistics. In this paper we analyse the asymptotic theory of quantum statistical models
consisting of ensembles of quantum systems which are identically prepared in a pure state. In the limit of large
ensembles we establish the local asymptotic equivalence (LAE) of this i.i.d. model to a quantum Gaussian
white noise model. We use the LAE result in order to establish minimax rates for the estimation of pure states
belonging to Hermite-Sobolev classes of wave functions. Moreover, for quadratic functional estimation of the
same states we note an elbow effect in the rates, whereas for testing a pure state a sharp parametric rate is
attained over the nonparametric Hermite-Sobolev class.
Keywords and phrases: Le Cam distance, local asymptotic equivalence, quantum Gaussian process,
quantum Gaussian sequence, quantum states ensemble, nonparametric estimation, quadratic functionals, nonparametric sharp testing rates.
1. Introduction
A striking insight of quantum mechanics is that randomness is a fundamental feature of the physical world at
the microscopic level. Any observation made on a quantum system such as an atom or a light pulse, results in a
non-deterministic, stochastic outcome. The study of the direct map from the system’s state or preparation to the
probability distribution of the measurement outcomes, has been one of the core topics in traditional quantum
theory. In recent decades the focus of research has shifted from fundamental physics towards applications at
the interface with information theory, computer science, and metrology, sharing the paradigm that individual
quantum systems as carriers of a new type of information [52].
In many quantum protocols, the experimenter has incomplete knowledge and control of the system and its
environment, or is interested in estimating an external field parameter which affects the system dynamics. In this
case one deals with a statistical inverse problem of inferring unknown state parameters from the measurement
data obtained by probing a large number of individual quantum systems. The theory and practice arising
from tackling such questions is shaping up into the field of quantum statistics, which lies at the intersection of
quantum theory and statistical inference [39, 37, 36, 56, 6, 1].
One of the central problems in quantum statistics is state estimation: given an ensemble of identically prepared,
independent systems with unknown state, the task is to estimate the state by performing appropriate measurements and devising estimators based on the measurement data. A landmark experiment aimed at creating
multipartite entangled states [34] highlighted the direct practical relevance of efficient estimation techniques
for large dimensional systems, the complexity of estimating large dimensional states, and the need for solid
statistical methodology in computing reliable “error bars”. This has motivated the development of new methods
such as compressed sensing and matrix ℓ1 -minimisation [29, 28, 22], spectral thresholding for low rank states
[14], confidence regions [17, 18, 65, 62, 21].
Another important research direction is towards developing a quantum decision theory as the overall mathematical framework for inference involving quantum systems seen as a form of “statistical data”. Typically, the
route to finding the building blocks of this theory starts with a decision problem (e.g. testing between two
states, or estimating certain parameters of a state) and the problem of finding optimal measurement settings
and statistical procedures for treating the (classical, random) measurement data. For instance, in the context
1
CREST, ENSAE, Université Paris-Saclay, 3, ave. P. Larousse 92245 MALAKOFF Cedex, FRANCE,
University of Nottingham, School of Mathematical Sciences, University Park, NG7 2RD Nottingham,
UK, 3 Department of Mathematics, Malott Hall, Cornell University, Ithaca, NY 14853, USA
2
1
2
LAE OF PURE STATES ENSEMBLES AND QUANTUM GAUSSIAN WHITE NOISE
of asymptotic binary hypothesis testing, two key results are the quantum Stein lemma [38, 55] and the quantum
Chernoff bound [2, 53, 3, 50]. As in the classical case, they describe the exponential decay of appropriate
error probabilities for optimal measurements, and they provide operational interpretations for quantum relative
entropy, and respectively quantum Chernoff distance. Similarly, an important problem in state estimation is to
identify measurements which allow for the smallest possible estimation error. A traditional approach has been
to establish a “quantum Cramér-Rao bound” (QCRB) [39, 37, 10] for the covariance of unbiased estimators,
where the right side is the inverse of the “quantum Fisher information matrix”, the latter depending only on
the structure of the quantum statistical model. However, while the QCRB is achievable asymptotically for onedimensional parameters, this is not the case for multi-parameter models due to the fact that the measurements
which are optimal for different one-dimensional components, are generally incompatible with each other.
These difficulties can be overcome by developing a fundamental theory of comparison and convergence of quantum statistical models, as an extension of its classical counterpart [63, 48]. While classical “data processing” is
described by randomisations, physical transformations of quantum systems are described by quantum channels
[52]. Following up on this idea, Petz and Jencova [58] have obtained a general characterisation of equivalent
models, as families of states that are related by quantum channels in both directions. This naturally leads to
the notion of Le Cam distance between quantum statistical models as the least trace-norm error incurred when
trying to map one model into another via quantum channels [43]. In this framework, the asymptotic theory
of state estimation can be investigated by adopting ideas from the classical local asymptotic normality (LAN)
theory [48]. Quantum LAN theory [31, 33, 43] shows that the sequence of models describing large samples of
identically prepared systems can be approximated by a simpler quantum Gaussian shift model, in the neighbourhood of an interior point of the parameter space. The original optimal state estimation problem is then
solved by combining LAN theory with known procedures for estimation of Gaussian states [30, 32, 25].
In this paper we extend the scope of the quantum LAN theory to cover non-parametric quantum models; more
precisely we will be interested in the set of pure states (one-dimensional projections) on infinite dimensional
Hilbert spaces. Infinite dimensional systems such as light pulses, free particles, are commonly encountered in
quantum physics, and their estimation is an important topic in quantum optics [49]. The minimax results
derived in this paper can serve as a benchmark for the performance of specific methods such as for instance
quantum homodyne tomography [1, 13], by comparing their risk with the minimax risk derived here.
The paper is organised as follows. In Section 2 we review the basic notions of quantum mechanics needed
for understanding the physical context of our investigation. In particular, we define the concepts of state,
measurement and quantum channel which can loosely be seen as quantum analogs of probability distribution and
Markov kernels, respectively. We further introduce the formalism of quantum Gaussian states, the Fock spaces
and second quantisation, which establish the quantum analogs of Gaussian distributions, Gaussian sequences
and Gaussian processes in continuous time. In Section 3.1 we review results in classical statistics on nonparametric asymptotic equivalence which serve as motivation and comparison to our work. In Section 3.2 we
introduce the general notion of a quantum statistical model and the Le Cam distance between two models. In
particular, in Section 3.3 we define the i.i.d. and Gaussian quantum models which are analysed in the remainder
of the paper.
One of the main results is Theorem 4.1 giving the local asymptotic equivalence (LAE) between the non-parametric
i.i.d. pure states model and the Gaussian shift model. This extends the existing local asymptotic normality
theory from parametric to non-parametric (infinite dimensional) models. Section 5 details three applications of
the LAE result in Theorem 4.1. In Section 5.1 we derive the asymptotic minimax rates and provide concrete
estimation procedures for state estimation with respect to the trace-norm and Bures distances, which are
analogues of the norm-one and Hellinger distances respectively. The main results are Theorems 5.1 and 5.3
which deal with the upper and respectively lower bound for a model consisting of ensemble of n independent
identically prepared systems in a pure state belonging to a Hermite-Sobolev class S α (L) of wave functions. In
Theorem 5.1 we describe a specific measurement procedure which provides an estimator whose risk attains the
nonparametric rate n−α/(2α+1) . The lower bound follows by using the LAE result to approximate the model
with a Gaussian one, combined with the lower bound for the corresponding quantum Gaussian model derived in
Theorem 5.2. In Section 5.2 we consider the estimation of a state functional corresponding to the expectation of
a power N 2β of the number operator. Theorems 5.4 and 5.5 establish the upper and lower bounds for functional
estimation for the Hermite-Sobolev class S α (L). The minimax rates are n−1/2 (parametric) if α ≥ 2β, and
n−1+β/α if β < α < 2β. In Section 5.3 we investigate non-parametric testing between a single state and a
composite hypothesis consisting of all states outside a ball of shrinking radius. Surprisingly, we find that the
minimax testing rates are parametric, in contrast to the non-parametric estimation rates. This fact is closely
related to the fact that the optimal estimation and testing measurements are incompatible with each other, so
LAE OF PURE STATES ENSEMBLES AND QUANTUM GAUSSIAN WHITE NOISE
3
that no single measurement strategy can allow for minimax estimation and testing in the same time. Results on
the minimax optimal rate for testing and the sharp asymptotics are given in Theorems 5.6 and 5.7 respectively.
Notation. Following physics convention, the vectors of a Hilbert space H will be denoted by the “ket” |vi,
so that the inner product of two vectors is the “bra-ket” hu|vi ∈ C which is linear with respect to the right
entry and anti-linear with respect to the left entry. Similarly, M := |uihv| is the rank one operator acting as
M : |wi 7→ M |wi = hv|wi|ui. We denote by L(H) the space of bounded linear operators on H which is a C∗ algebra with respect to the operator norm kAk := supψ6=0 kAψk/kψk. Additionally, T1 (H) ⊂ L(H) is the space
of Hilbert-Schmidt (or trace-class) operators equipped with the norm-one kτ k1 := Tr(|τ |), where the operator
|τ | := (τ ∗ τ )1/2 is the absolute value of τ , and τ ∗ is the adjoint of τ . Finally, we denote by T2 (H) ⊂ L(H) the
space of Hilbert-Schmidt operators equipped with the norm-two kτ k22 := Tr(|τ |2 ), which is a Hilbert space with
respect to the inner product (τ, σ) := Tr(τ ∗ σ).
2. Quantum mechanics background
In this section we review some basic notions of quantum mechanics (QM), in as much as it is required for
understanding the subsequent results of the paper. Since QM is a probabilistic theory of quantum phenomena, it
is helpful to approach the formalism from the perspective of analogies and differences with “classical” probability.
We refer to [52] for more details on the quantum formalism.
2.1. States, measurements, channels. The QM formalism assigns to each quantum mechanical system (e.g.
an atom, light pulse, quantum spin) a complex Hilbert space H, called the space of states. For instance, the
finite dimensional space Cd is the Hilbert space of a system with d “energy levels”, while L2 (R) is the space
of “wave functions” of a particle moving in one dimension, or of a monochromatic light pulse. The state of a
quantum system is represented mathematically by a density matrix.
Definition 1. Let H be the Hilbert space of a quantum system. A density matrix (or state) on H is a linear
operator ρ : H → H which is positive (i.e. it is selfadjoint and has non-negative eigenvalues), and has trace
one.
We denote by S(H) the convex space of states on H. Its linear span is the space of trace class operators T1 (H),
which is the non-commutative analogue of the space of absolutely integrable functions on a probability space
L1 (Ω, Σ, P). For any states ρ1 or ρ2 , the convex combination λρ1 + (1 − λ)ρ2 is also a state which corresponds
to randomly preparing the system in either the state ρ1 or ρ2 with probabilities λ and respectively 1 − λ. The
extremal elements of the convex set S(H) are the one dimensional projections Pψ = |ψihψ| where |ψi is a
normalised vector, i.e. kψk = 1. Such states are called pure (as opposed to mixed states which are convex
combinations of pure ones), and are uniquely determined by the vector |ψi. Conversely, the vector |ψi is fixed
by the state up to a complex phase factor, i.e. |ψi and |ψ ′ i := eiφ |ψi represent the same state.
Although the quantum state encodes all information about the preparation of the system, it is not a directly
observable property. Instead, any measurement produces a random outcome whose distribution depends on the
state, and thus reveals in a probabilistic way a certain aspect of the system’s preparation. The simplest type of
measurement is determined by an orthonormal basis (ONB) {|ii}dimH
and a set of possible outcomes {λi }dimH
i=1
i=1
in the following way: the outcome is a random variable X taking the value λi with probability given by the
diagonal elements of ρ in this particular basis
Pρ ([X = λi ]) = ρii = hi|ρ|ii.
More generally, a measurement M with outcomes in a measurable space (Ω, Σ) is determined by a positive
operator valued measure.
Definition 2. A positive operator valued measure (POVM) is a map M : Σ → L(H) having the following
properties
1) positivity: M (E) ≥ 0 for all events
E∈Σ
P
2) sigma additivity: M (∪i Ei ) = i M (Ei ) for any countable set of mutually disjoint events Ei
3) normalization: M (Ω) = 1.
The outcome of the corresponding measurement associated to M has probability distribution
Pρ (E) = Tr(ρM (E)),
E ∈ Σ.
4
LAE OF PURE STATES ENSEMBLES AND QUANTUM GAUSSIAN WHITE NOISE
The most important example of a POVM, is that associated to the measurement of an observable, the latter
being represented mathematically by a selfadjoint operator A : H → H. The Spectral Theorem shows that such
operators can be “diagonalised”, i.e. they have a spectral decomposition
Z
xP (dx)
A=
σ(A)
where σ(A) is the spectrum of A, and {P (E) : E ∈ Σ} is the collection of spectral projections of A. The
corresponding measurement has outcome a ∈ σ(A) with probability distribution Pρ [a ∈ E] = Tr(ρP (E)).
Unlike “classical” systems which can be observed without disturbing their state, quantum systems are typically
perturbed by the measurement, so the system needs to be reprepared in order to obtain more information
about the state. In this sense, the system can be seen as a “quantum sample” which it can be converted into
a “classical” sample only by performing a measurement. Thus, a measurement can be seen as a “quantum-todP
classical randomisation”, i.e. a linear map M which sends a state ρ to the probability density M(ρ) ≡ pρ := dPρ
with respect to a reference measure P. The latter can be taken to be Pρ0 for a strictly positive density matrix
ρ0 . The following lemma summarises this perspective on measurements.
Lemma 2.1. Let H be a Hilbert space, and let (Ω, Σ) be a measurable space. For any fixed state ρ0 > 0 on H,
there is a one-to-one correspondence between POVMs M over (Ω, Σ) and quantum-to-classical randomisations,
i.e. linear maps
M : T1 (H) → L1 (Ω, Σ, P)
which are positive and normalised (maps states into probability densities). The correspondence is given by
Z
dPρ
.
M(ρ) ≡ pρ :=
pρ (ω)Pρ0 (dω),
Pρ (E) = Tr(M (E)ρ) =
dP
A
For comparison, recall that a linear map R : L1 (Ω′ , Σ′ , P′ ) → L1 (Ω, Σ, P) is a stochastic operator if it maps
probability densities into probability densities [61]. Typically such maps arise from Markov kernels and describe
randomizations of dominated statistical experiments (models).
While a measurement is a quantum-to-classical randomization, a “quantum-to-quantum randomization” describes how the system’s state changes as a result of time evolution or interaction with other systems. The
maps describing such transformations are called quantum channels.
Definition 3. A quantum channel between systems with Hilbert spaces H1 and H2 is a trace preserving, completely positive linear map
T : T1 (H1 ) → T1 (H2 ).
The two properties mentioned above are similar to those of a classical randomization, so in particular T maps
states into states. However, unlike the classical case, T is required to satisfy a stronger positivity property:
T is completely positive if Idm ⊗ T is positive for all m ≥ 1, where Idm is the identity map on the space of
m dimensional matrices. This ensures that when the system is correlated with an ancillary system Cm , and
the latter undergoes the identity transformation, the final joint state is still positive, as expected on physical
grounds.
The simplest example of a quantum channel is a unitary transformation ρ 7→ U ρU ∗ , where U is a unitary
operator on H. More generally, if |ϕi ∈ K is a pure state of an ancillary system, and V is a unitary on H ⊗ K,
then
ρ 7→ T (ρ) := TrK (V (ρ ⊗ |ϕihϕ|)V ∗ )
is a quantum channel describing the system state after interacting with the ancilla. By computing the partial
trace TrK over K with respect to an orthonormal basis {|fi i}dimK
i=1 we obtain the following expression
X
Ki ρKi∗
(1)
T (ρ) =
i
where Ki are operators on H definedP
by hψ|Ki |ψ i := hψ ⊗ fi |U |ψ ′ ⊗ ϕi. Note that by definition, these operators
satisfy the normalisation condition i Ki∗ Ki = 1. Conversely, the Kraus Theorem shows that any quantum
channel is of the form (1) with operators Ki respecting the normalisation condition.
′
2.2. Continuous variables, Fock spaces and Gaussian states. In this section we look at the class of
“continuous variables” (cv) systems, which model a variety of physical systems such as light pulses, or free
particles. Such systems play an important role in this work as “carriers” of quantum Gaussian states, and in
particular in the local asymptotic equivalence result. We refer to [49] for further reading.
LAE OF PURE STATES ENSEMBLES AND QUANTUM GAUSSIAN WHITE NOISE
5
2.2.1. One mode systems. We start with the simplest case of a “one-mode” cv system, after which we show how
this construction can be extended to more general “multi-mode” cv systems. The Hilbert space of a one-mode
system is L2 (R), i.e. the space of square integrable wave functions on the real line. On this we define the
selfadjoint operators acting on appropriately defined domains as
dψ(q)
(Qψ)(q) = qψ(q),
(P ψ)(q) = −i
dq
which satisfy the “canonical commutation relations” QP − P Q = i1. To better understand the meaning of
the observable Q, let us consider its measurement for a pure state ρ = |ψihψ| with wave function |ψi. The
outcome takes values in R, and its probability distribution has density with respect to the Lebesgue measure
2
P
2
pQ
ρ (x) = |ψ(x)| . Similarly, the probability density of the observable P is given by pρ (x) = |ψ̃(x)| , where
2
ψ̃ ∈ L (R) is the Fourier transform of the function ψ(·). When the system under consideration is the free
particle, Q and P are usually associated to the position and momentum observables, while for a monochromatic
light mode they correspond to the electric and magnetic fields. Note that the distributions of P and Q are not
sufficient to identify the state, even in the case of a pure state. However, it turns out that the state is uniquely
determined by the collection of probability distributions of all quadrature observables Xφ := cos(φ)·Q+sin(φ)·P
for angles φ ∈ [0, 2π]. To understand this, it is helpful to think of the state of the one-mode cv system as a
quantum analogue of a joint distribution of two real valued variables, i.e. a 2D distribution. Indeed, in the
latter case, the distribution is determined by collection of marginals along all directions in the plane (its Radon
transform); this fact is exploited in PET tomography which aims at estimating the 2D distribution from samples
of its Radon transform. In the quantum case, since Q and P do not commute with each other, they cannot be
measured simultaneously and cannot be assigned a joint distribution in a meaningful way. However, the “quasidistribution” defined below has some of the desired properties, and is very helpful in visualising the quantum
state.
Definition 4. For any state ρ ∈ T1 (L2 (R)) we define the quantum characteristic function of ρ
fρ (u, v) := Tr(exp(−iuQ − ivP )ρ).
W
fρ with respect to both variables is called Wigner function Wρ , or quasiThe inverse Fourier transform of W
distribution associated to ρ:
Z Z
1
fρ (u, v)dudv.
exp(iuq + ivp)W
Wρ (q, p) =
(2π)2
A consequence of this definition is that the marginal of Wρ (q, p) along an arbitrary direction with angle φ is
the probability density of the quadrature Xφ introduced above. This is the basis of a quantum state estimation
scheme called “quantum homodyne tomography” [49, 1], where the Wigner function plays the role of the 2D
distribution from “classical” PET tomography. One of the important differences however, is that the Wigner
functions need not be positive in general, and satisfy other constraints which are specific to the quantum setting
and can be exploited in the estimation procedure.
The Wigner function representation offers an intuitive route to defining the notion of Gaussian state.
Definition 5. A state ρ of a one-mode cv system is called Gaussian if its Wigner function Wρ is a Gaussian
probability density, or equivalently if it has the quantum characteristic function
V
T
f
· exp(iuq0 + ivp0 ).
Wρ (u, v) = exp −(u, v) (u, v)
2
where (q0 , p0 ) ∈ R2 and V (a real positive 2 × 2 matrix) are the mean and variance of Wρ , respectively.
In particular, all the quadratures Xφ of a Gaussian state have Gaussian distribution. As consequence of
the commutation relation QP − P Q = i1 the observables Q and P cannot have arbitrarily small variance
simultaneously; in particular, the covariance matrix V must satisfy the “uncertainty principle” Det(V ) ≥ 1/4,
where the equality is achieved if and only if the state is a pure Gaussian state.
We will be particularly interested in coherent states |G(z)i which are pure Gaussian states whose Wigner
functions have covariance matrix V = I2 /2, where I2 is the 2×2 identity matrix. To give a concrete Hilbert space
representation, it is convenient to introduce a special orthonormal basis of L2 (R), consisting of the eigenvectors
√
{|0i, |1i, . . . } of√the number operator N = a∗ a, with N |ki = k|ki. Here, the operators a∗ = (Q − iP )/ 2 and
a = (Q + iP )/ 2 are called creation and annihilation operators and act as “ladder operators” on the number
basis vectors (or Fock states)
√
√
a|ki = n|k − 1i,
a∗ |ki = k + 1|k + 1i.
6
LAE OF PURE STATES ENSEMBLES AND QUANTUM GAUSSIAN WHITE NOISE
The coherent states denoted by |G(z)i are obtained by applying the unitary Weyl (displacement) operators to
the vacuum state |0i
∞
X
zk
√ |ni,
(2)
|G(z)i = exp (za∗ − z̄a) |0i = exp(−|z|2 /2)
k!
k=0
where z ∈ C is the eigenvalue
√ of the annihilation operator√a|G(z)i = z|G(z)i; in particular, the quadrature
means are hG(z)|Q|G(z)i = 2Re(z) and hG(z)|P |G(z)i = 2Im(z), and the Wigner function is given by
√
√
1
W|zi (q, p) = exp −(q − 2x)2 − (p − 2y)2 , q, p ∈ R.
(3)
π
Equation (2) implies that the number operator N has a Poisson distribution with mean |z|2 . Additionally, it can
be seen from the Fourier expansion in the second equality that the unitary Γ(φ) = exp(iφN ) acts by rotating
the coherent states by an angle φ in the complex plane, i.e. Γ(φ)|G(z)i = |G(eiφ z)i.
Another important class of Gaussian states are the mixed diagonal states
∞
X
Φ(r) = (1 − r)
rk |kihk|, 0 < r < 1
(4)
k=0
which are also called thermal states, cf. section 3.3 in [49]. The corresponding Wigner function is a centred
Gaussian
2
1
q + p2
WΦ(r) (q, p) =
.
(5)
exp
−
2πσ 2 (r)
2σ 2 (r)
1+r
with covariance matrix V = σ 2 (r) · I2 where σ 2 (r) = 12 1−r
.
Proposition 2.2. Consider the family of coherent states {|G(z)ihG(z)|, z ∈ C}, with random displacement
(location) z distributed
according to Π(dz), having a Gaussian law with covariance matrix σ 2 · I2 . Then, the
R
mixed state Φ = |G(z)ihG(z)|Π(dz) is a thermal state
Φ(r), with r =
2σ 2
.
2σ 2 + 1
Proof. Consider the corresponding Wigner function
Z
1
1
dxdy
WΦ (q, p) =
W|G(z)i (q, p) exp − 2 (x2 + y 2 )
2σ
2πσ 2
Z
Z
√
√
1
dx
dy
x2
y2
2
2
√
=
· exp −(p − 2 y) − 2 √
exp −(q − 2 x) − 2
πσ 2
2σ
2σ
2π
2π
1
q 2 + p2
=
.
exp −
π(4σ 2 + 1)
2(2σ 2 + 1/2)
Therefore, the state Φ is identical to the thermal state Φ(r) with 2σ 2 + 21 =
1 1+r
2 1−r ,
or equivalently r =
2σ2
1+2σ2 .
(6)
This fact will be used later on in in section 5 in applications to functional estimation and testing.
2.2.2. Fock spaces and second quantisation. The above construction can be generalised to multimode systems
by tensoring several one-mode systems. Thus, the Hilbert space of a k-mode system is L2 (R)⊗k ∼
= L2 (Rk ),
upon which we define “canonical pairs” (Qi , Pi ) acting on the i-th tensor as above, and as identity on the
other tensors. Similarly we define the one-mode operators ai , a∗i , Ni . The number basis consists now of tensor
products |ni := ⊗ki=1 |ni i indexed by the sequences of integers n = (n1 , . . . , nk ). A multimode coherent state is
a tensor product of one-mode coherent states
!
k
∞
Y
X
zin
k
†
†
2
√
|ni ∈ L2 (R)⊗k
(7)
|G(z)i = ⊗i=1 |G(zi )i = exp za − az |0i = exp(−|z| /2)
n
!
i
n=0 i=1
where z = (z1 , . . . , zk ) is the vector of means, a = (a1 , . . . , ak ), and † denotes the transposition and adjoint
(complex conjugation) of individual entries.
We will now extend this construction to systems with infinitely many modes. One way to do this is by defining
an infinite tensor product of one-mode spaces, as completion of the space spanned by tensors in which all but
a finite number of modes are in the vacuum state. Instead, we will present an equivalent but more elegant
construction called second quantisation which will be useful for later considerations.
LAE OF PURE STATES ENSEMBLES AND QUANTUM GAUSSIAN WHITE NOISE
Definition 6. Let K be a Hilbert space. The Fock space over K is the Hilbert space
M
F (K) =
K ⊗s n
7
(8)
n≥0
where K⊗s n denotes the n-fold symmetric tensor product, i.e. the subspace of K⊗n consisting of vectors which
are symmetric under permutations of the tensors. The term K⊗s 0 =: C|0i is called the vacuum state.
In this definition the space K should be regarded as the “space of modes” rather than physical states. As we
will see below, by fixing an orthonormal basis in K, we can establish an isomorphism between the Fock space
F (K) and a tensor product of one-mode cv spaces, one for each basis vector. In particular, if K = C, then
F (C) ∼
= L2 (R) so that the one-dimensional subspaces in the direct sum in (8) correspond to the number basis
vectors |0i, |1i, · · · ∈ L2 (R) of a one-mode cv system.
We now introduce the general notion of coherent state on a Fock space.
Definition 7. Let F (K) be the Fock space over K. For each |vi ∈ K we define an associated coherent state
M 1
2
√ |vi⊗n ∈ F (K).
|G(v)i := e−kvk /2
n!
n≥0
The coherent vectors form a dense subspace of F (K). This fact can be used to prove the following factorisation
property, and to define the annihilation operators below. Let K = K0 ⊕ K1 be a direct sum decomposition of K
into orthogonal subspaces, and let |vi = |v0 i ⊕ |v1 i be the decomposition of a generic vector |vi ∈ K. Then the
map
U : F (K)
U : |G(v)i
→ F (K0 ) ⊗ F (K1 )
7→ |G(v0 )i ⊗ |G(v1 )i
is unitary. We will use this correspondence to identify F (K) with the tensor product F (K0 ) ⊗ F (K1 ). By the
same argument, for any orthonormal basis {|e1 i, |e2 i, . . . } of K, the Fock space F (K) is isomorphic with the
tensor product of one mode spaces Fi := F (C|ei i) and the coherent states factorise as
O
Fi
F (K) ∼
=
i
|G(u)i
so that we recover the formula (7).
∼
=
O
i
|G(ui )i,
ui = hei |ui.
(9)
We define the annihilation operators through their action on coherent states as follows: for each mode |ui ∈ K
the associated annihilator a(u) : F (K) → F (K) is given by
a(u) : |G(v)i = hu|vi|G(v)i,
|vi ∈ K.
Then the annihilation and (their adjoint) the creation operators satisfy the commutation relations
a(u)a∗ (w) − a∗ (w)a(u) = hu|vi1.
For each mode we can also define the canonical operators Q(u), P (u) and the number operator N (u) in terms
of a(u), a∗ (u) as in the one-mode case. Moreover, if |ui = |u0 i ⊕ |u1 i is the decomposition of |ui as above,
then a(u0 ) acts as a(u0 ) ⊗ 1F (K1 ) , when the Fock space is represented in the tensor product form. Similar
decompositions hold for a∗ (u0 ), N (u0 ), a(u1 ), a∗ (u1 ), N (u1 ).
The second quantisation has the following functorial properties which will be used later on.
Definition 8. Let W : K → K be a unitary operator. The quantisation operator Γ(W ) is the unitary defined
by Γ(W ) : F (K) → F (K) by
M
Γ(W ) :=
W ⊗n
(10)
n≥0
where W
⊗n
acts on the n-th level of the Fock space K
⊗s n
.
From the definition it follows that the action of Γ(W ) on coherent states is covariant in the sense that
Γ(W ) : F (K)
Γ(W ) : |G(v)i
→ F (K)
7
→
|G(W v)i.
8
LAE OF PURE STATES ENSEMBLES AND QUANTUM GAUSSIAN WHITE NOISE
In particular, it follows from the definitions that Γ(eiφ 1) = exp(iφN ), where N is the total number operator,
whose action on the n-th level of the Fock space is N |vi⊗n = n|vi⊗n . Note that while |vi and eiφ |vi differ only
by a phase and hence represent the same state, the corresponding coherent states |G(v)i and Γ(eiφ )|G(v)i =
|G(eiφ v)i are linearly independent and represent different states.
As in the single mode case, the coherent states can be obtained by acting with the unitary displacement (or
Weyl) operators onto the vacuum
|G(u)i = exp(a∗ (u) − a(u))|0i
Moreover, the coherent states |G(u)i are Gaussian with respect to all coordinates. The means of annihilation operators are given by√ hG(u)|a(w)|G(u)i
= hw|vi, from which we can deduce that the the coordinates
√
(Q(w), P (w)) have means ( 2Rehw|ui, 2Imhw|ui). The covariance of coherent states is constant (independent
of the displacement u), and is given by h0|a(w)a∗ (v)|0i = hw|vi. This implies that orthogonal modes (i.e.
hw|vi = 0) have independent pairs of coordinates.
2.3. Metrics on the space of states. For future reference we review here the states space metrics used in
the paper. Recall that the space of states S(H) on a Hilbert space H is the cone of positive, trace one operators
in T1 (H). The norm-one (or trace-norm) distance between two states ρ0 , ρ1 ∈ S(H) is given by
kρ0 − ρ1 k1 := Tr(|ρ0 − ρ1 |)
√
where |τ | := τ ∗ τ denotes the absolute value of τ . The norm-one distance can be interpreted as the maximum
difference between expectations of bounded observables
kρ0 − ρ1 k1 =
sup
A:kAk≤1
|Tr(ρ0 A) − Tr(ρ1 A)|.
Another interpretation is in terms of quantum testing. Let M = (M0 , M1 ) be a binary POVM used to test
between hypotheses H0 := {measured state is ρ0 } and H1 := {measured state is ρ1 }. The sum of error probabilities is
PM
e = Tr(M0 ρ1 ) + Tr(M1 ρ0 ).
By optimizing over all possible POVM we obtain [37] the optimal error probability sum
1
(11)
P∗e := inf PM
e = 1 − kρ0 − ρ1 k1 .
M
2
In the special case of pure states, the norm-one distance is given by
p
k|ψ0 ihψ0 | − |ψ1 ihψ1 |k1 = 2 1 − |hψ0 |ψ1 i|2 ,
(12)
as proven e.g. in [44]. The previous formula becomes for coherent states
p
k|G(ψ0 )ihG(ψ0 )| − |G(ψ1 )ihG(ψ1 )|k1 = 2 1 − exp(−kψ0 − ψ1 k2 ).
The second important metric is the Bures distance whose square is given by
q
√
√
2
ρ0 ρ1 ρ0 )
db (ρ0 , ρ1 ) := 2(1 − Tr
and is a quantum extension of the Hellinger distance. In the case of pure states the Bures distance becomes
so for coherent states it is given by
d2b (|ψ0 ihψ0 | , |ψ1 ihψ1 |) = 2(1 − |hψ0 |ψ1 i|)
(13)
1
.
d2b (|G(ψ0 )ihG(ψ0 )| , |G(ψ1 )ihG(ψ1 )|) := 2 1 − exp − kψ0 − ψ1 k2
2
Similarly to the classical case, the following inequality holds for arbitrary states [23]
d2b (ρ0 , ρ1 ) ≤ kρ0 − ρ1 k1 ≤ 2db (ρ0 , ρ1 ).
Moreover, since |hψ0 |ψ1 i| ≤ |hψ0 |ψ1 i|, the additional inequality holds for pure states
√
kρ0 − ρ1 k1 ≥ 2db (ρ0 , ρ1 ).
2
(14)
(15)
This means that for pure states, the trace and Bures distances are equivalent (up to constants).
Finally, we will be using the fact that both the norm-one and the Bures distance are contractive under quantum
channels. T : T1 (H) → T1 (H′ ), i.e.
kT (ρ0 ) − T (ρ1 )k1 ≤ kρ0 − ρ1 k1 ,
d2b (T (ρ0 ), T (ρ1 )) ≤ d2b (ρ0 , ρ1 ).
LAE OF PURE STATES ENSEMBLES AND QUANTUM GAUSSIAN WHITE NOISE
9
3. Classical and quantum statistical models
In this section we review key elements of quantum statistics, and introduce the quantum statistical models
which will be analysed later on. For comparison, we review certain asymptotic equivalence results for related
classical statistical models.
3.1. Classical models. Here we review several asymptotic normality results for classical models which are
analogous to the quantum models investigated in the paper.
A classical statistical model is defined as a family of probability distributions Q = {Pf : f ∈ W} on a measurable
space (X , A), indexed by an unknown, possibly infinite dimensional parameter f to be estimated, which belongs
to a parameter space W. In the asymptotic framework considered here we assume that we are given a (large)
number n of independent, identically distributed samples X1 , . . . , Xn from Pf , from which we would like to
estimate f . If d : W × W → R+ is a chosen loss function, then the risk of an estimator fˆn = fˆn (X1 , . . . , Xn ) is
i
h
R(fˆn , f ) = Ef d(fˆn , f )2 .
In nonparametric statistics, the parameter of the model f is often a function that belongs to a smoothness class.
We consider two classes W: the periodic Sobolev class S α (L) of functions on [0, 1] with smoothness α > 1/2,
and the Hölder class Λα (L), with smoothness α > 0. For any f ∈ L2 [0, 1], let {fj , j ∈ Z} be the set of Fourier
coefficients with respect to the standard trigonometric basis. The classes are defined as
XZ
S α (L) := f : [0, 1] → R :
|fj |2 |j|2α du ≤ L .
j∈Z
and
Λα (L) := {f : [0, 1] → R : |f (x) − f (y)| ≤ L|x − y|α , x, y ∈ [0, 1]} .
In addition, when densities f are considered, we will assume that W includes an additional restriction to a class
)
(
Z
f (x)dx = 1
Dε = f : [0, 1] → [ε, ∞) :
[0,1]
for some ε > 0.
Density model. The classical density model consists of n observations X1 , . . . , Xn which are independent,
identically distributed (i.i.d.) with common probability density f
n
o
Pn = P⊗n
: f ∈W .
f
Gaussian regression model with fixed equidistant design. In this model, we observe Y1 , ..., Yn such that
i
1/2
+ ξi , i = 1, ..., n,
Yi = f
n
where the errors ξ1 , ..., ξn are i.i.d., standard Gaussian variables. Denote the Gaussian regression model by
)
( n
O
i
1/2
,1 : f ∈ W .
N f
Rn =
n
i=1
Gaussian white noise model. In this model the square-root density f 1/2 is observed with Gaussian white noise
of variance n−1 , i.e.
1
dYt = f 1/2 (t)dt + √ dWt , t ∈ [0, 1].
(16)
n
If we denote by Qf the probability distribution of {Y (t) : t ∈ [0, 1]}, the corresponding model is
Fn := {Qf : f ∈ W} .
Gaussian sequence model. In this model we observe a sequence of Gaussian random variables with means equal
to the coefficients of f 1/2 in some orthonormal basis of L2 [0, 1] for f ∈ F
1
yj = θj (f 1/2 ) + √ ξj ,
n
i = 1, 2, . . .
(17)
10
LAE OF PURE STATES ENSEMBLES AND QUANTUM GAUSSIAN WHITE NOISE
where {ξi }i≥1 are Gaussian i.i.d. random variables. We denote this model
O
1
:f ∈W .
Nn =
N θj f 1/2 ,
n
j≥1
In [54] it was shown that the sequences of models Pn and Fn are asymptotically equivalent in the sense that
their Le Cam distance converges to zero as n → ∞ when W = Λα (L) ∩ Dε with α > 1/2; in [12], a similar
result was established for Rn and Fn (more precisely, with f 1/2 any real valued function f 1/2 ∈ Λα (L)). Later,
[60] showed that models Fn and Nn are asymptotically equivalent over periodic Sobolev classes f 1/2 ∈ S α (L)
with smoothness α > 1/2. Among many other results [27] considered generalized linear models, [11] regression
models with random design and [59] multivariate and random design, [26] compared the stationary Gaussian
process with the Gaussian white noise model Fn .
In all classical results, the underlying nonparametric function was assumed to belong to a smoothness class in
order to establish asymptotic equivalence of models. In the quantum setup of pure states and Gaussian states
that we discuss later on, no such smoothness assumption is needed.
3.2. Quantum models, randomisations and convergence. In this subsection we introduce the basic notions of a theory of quantum statistical models which is currently still in its early stages, cf. [33, 25] for more
details. We will focus on the notions of quantum-to-classical randomisation carried out through measurements,
and quantum-to-quantum randomisations implemented by quantum channels, which allow us to define the
equivalence and the Le Cam distance between models.
In analogy to the classical case, we make the following definition.
Definition 9. A quantum statistical model over a parameter space Θ consists of a family of quantum states
Q = {ρθ : θ ∈ Θ} on a Hilbert space H, indexed by an unknown parameter θ ∈ Θ.
A simple example is a family of pure states {ρθ = |ψθ ihψθ | : θ ∈ R} with |ψθ i := exp(iθH)|ψi, where H
is a selfdajoint operator generating the one-dimensional family of unitaries exp(iθH), and |ψi ∈ H is a fixed
vector. Physically, the parameter θ could be for instance time, a phase, or an external magnetic field. Another
example is that of a completely unknown state of a finite dimensional system, which can be parametrised in
terms of its density matrix elements, or the eigenvalues and eigenvectors. In order to increase the estimation
precision one typically prepares a number n of identical and independent copies of the state ρθ , in which case
the corresponding model is Qn := {ρ⊗n
: θ ∈ Θ}. Our work deals with non-parametric quantum statistical
θ
models for which the underlying Hilbert space is infinite dimensional, as we will detail below.
In order to obtain information about the parameter θ, we need to perform measurements on the system prepared
in ρθ . Using the random measurement data, we then employ statistical methods to solve specific decision
problems. For instance, the task of estimating an unknown quantum state (also known as quantum tomography)
is a key component of quantum engineering experiments [34]. In particular, the estimation of large dimensional
states has received significant attention in the context of compressed sensing [29, 22], and estimation of low rank
states [14]. Suppose that we perform a measurement M on the system in state ρθ , and obtain a random outcome
O ∈ Ω with distribution PM
θ (E) := Tr(ρθ M (E)), cf. section 2. The measurement data is therefore described by
the classical model P M := {PM
θ : θ ∈ Θ}, and the estimation problem can be treated using “classical” statistical
methods. The measurement map
M
M
: T1 → L1 (Ω, Σ, P)
dPθ
: ρθ 7→ pθ :=
dP
can be seen as a randomisation from a classical to a quantum model, which intuitively means that Q is more
informative that P M for any measurement M . Here P can be chosen to be the distribution corresponding
to an arbitrary full rank (strictly positive) state ρ which insures the existence of all probability densities pθ .
One of the distinguishing features of quantum statistics is the possibility to choose appropriate measurements
for specific statistical problems (e.g. estimation, testing) and the fact that optimal measurements for different
problems may be incompatible with each other. In the applications section we will discuss specific instances of
this phenomenon.
LAE OF PURE STATES ENSEMBLES AND QUANTUM GAUSSIAN WHITE NOISE
11
Beside measurements, the quantum model Q can be transformed into another quantum model Q′ := {ρ′θ : θ ∈ Θ}
on a Hilbert space H′ by means of a quantum randomisation, i.e. by applying a quantum channel
T
T
: T1 (H) → T1 (H′ )
: ρθ 7→ ρ′θ .
The model Q′ is less informative than Q in the sense that for any measurement M ′ on H′ one can construct
′
the measurement M := M ′ ◦ T on H such that PM
= PM
θ
θ for all θ. If there exists another channel S such
′
that S(ρθ ) = ρθ for all θ we say (in analogy to the classical case) that the models Q and Q′ are equivalent ; in
particular, for any statistical decision problem, one can match a procedure for one model with a procedure with
the same risk, for the other model. A closely related concept is that of quantum sufficiency whose theory was
developed in [58]. More generally, we define the Le Cam distance in analogy to the classical case [48].
Definition 10. Let Q and Q′ be two quantum models over Θ. The deficiency between Q and Q′ is defined by
δ (Q, Q′ ) := inf sup kT (ρθ ) − ρ′θ k1
T θ∈Θ
where the infimum is taken over all channels T . The Le Cam distance between Q and Q′ is defined as
∆ (Q, Q′ ) := max (δ (Q, Q′ ) , δ (Q′ , Q)) .
(18)
Its interpretation is that models which are “close” in the Le Cam distance have similar statistical properties.
In practice, this metric is often used to approximate a sequence of models by another sequence of simpler
models, providing a method to establish asymptotic minimax risks. In particular, the approximation of i.i.d.
quantum statistical models by quantum Gaussian ones has been investigated in [31, 33, 43], in the case of finite
dimensional systems with arbitrary mixed states. Our goal is to extend these results to non-parametric models
consisting of pure states on infinite dimensional Hilbert spaces. The following lemma will be used later on.
P
Lemma 3.1. Let Q, Q′ be two quantum
models as defined above. Let ρi = i µi,j ρθi,j be two arbitrary mixtures
P
(i = 1, 2) of states in Q and let ρ′i = i µi,j ρ′θi,j be their counterparts in Q′ . Then
kρ′1 − ρ′2 k1 − 2∆(Q, Q′ ) ≤ kρ1 − ρ2 k1 ≤ kρ′1 − ρ′2 k1 + 2∆(Q, Q′ ).
Proof. Since quantum channels are contractive with respect to the norm-one
and by the triangle inequality we get
kS(ρ′1 ) − S(ρ′2 )k1 ≤ kρ′1 − ρ′2 k1
kρ1 − ρ2 k1 ≤ kρ1 − S(ρ′1 )k1 + kS(ρ′1 ) − S(ρ′2 )k1 + kS(ρ′2 ) − ρ2 k1 ≤ 2∆(Q, Q′ ) + kρ′1 − ρ′2 k1
The second inequality can be shown in a similar way.
3.3. The i.i.d. and the quantum white noise models. We now introduce the non-parametric quantum
models investigated in the paper, and discuss the relationship with the classical models described in section 3.1.
Let H be an infinite dimensional Hilbert space and let B := {|e0 i, |e1 i, P
. . . } be a fixed orthonormal basis in H.
∞
The Fourier decomposition of an arbitrary vector is written as |ψi = j=0 ψj |ej i. Since most of the models
will consist of pure states, we will sometimes define them in terms of the Hilbert space vectors rather than the
density matrices, but keep in mind that the vectors are uniquely defined only up to a complex phase.
Let us consider the general problem of estimating an unknown pure quantum state in H. For finite dimensional
systems, the risk with respect to typical rotation invariant loss functions scales linearly with the number of
parameters [24], hence with the dimension of the space. Therefore, since H is infinite dimensional, it is not
possible to develop a meaningful estimation theory without any prior information about the state. Motivated
by physical principles and statistical methodology we introduce the following Hermite-Sobolev classes [8] and
[9] of pure states characterised by an appropriate decay of the coefficients with respect to the basis B:
∞
X
|ψj |2 j 2α ≤ L, and kψk = 1 ,
α > 0, L > 0.
(19)
S α (L) := |ψihψ| :
j=0
To gain some intuition about the meaning of this class, let us assume that B is the Fock basis of a one-mode
cv system. Then the constraint translates into the moment condition for the number operator hψ|N 2α |ψi ≤ L;
this is a mild assumption considering that all experimentally feasible states have finite moments to all orders.
Even more, the coefficients of typical states such as coherent, squeezed, and Fock states decay exponentially
with the photon number.
12
LAE OF PURE STATES ENSEMBLES AND QUANTUM GAUSSIAN WHITE NOISE
Our first model describes n identical copies of a pure state belonging to the Sobolev class
Qn := {|ψihψ|⊗n : |ψihψ| ∈ S α (L)}.
(20)
In section 5.1 we show that the minimax rate of Qn for the norm-one and Bures distance loss functions is
n−α/(2α+1) . This is identical to the minimax rate of the classical i.i.d. model described in section 3.1.
We now introduce
the corresponding quantum Gaussian model. Let F :=
√ F (H) be the Fock space over H,
√
and let |G(√ nψ)i ∈ F be the coherent state with “displacement” vector nψ. As discussed in section 2.2.2,
the vector nψ should be seen now as the expectation of the infinite dimensional Gaussian state rather than a
quantum state in itself, for which reason we have omitted the ket notation. We define the coherent states model
√
√
(21)
Gn = G( nψ) G( nψ) : |ψi ∈ H, such that |ψihψ| ∈ S α (L) .
Using the factorisation property (9) with respect to the orthonormal basis B,√we see that the model is equivalent
to the product of independent one-mode coherent Gaussian states of mean nψi
∞
O
√
√
∼
G( nψi )
G( nψ) =
i=1
which is analogous to the classical Gaussian sequence model Nn defined in equation (17).
Similarly, we can draw an analogy with the white noise model Fn by realising H as L2 ([0, 1]). Let us define the
quantum stochastic process [57] on F (L2 ([0, 1]))
B(t) := a χ[0,t] + a∗ χ[0,t]
and note that [B(t), B(s)] = 0 for all t, s ∈ [0, 1] so that {B(t) : t ∈ [0, 1]} is a commutative family of operators.
This implies that {B(t) : t ∈ [0, 1]} have a joint probability distribution which is uniquely determined by
the quantum state, and can be regarded as a classical stochastic process. If the state is the vacuum |0i,
the process is
√ has the same distribution as the Brownian motion. Consider now the process
√ Gaussian and
X(t) := W ( nψ)∗ B(t)W ( nψ). which is obtained by applying a unitary Weyl transformation to B(t). In
physics terms we work here in the “Heisenberg picture” where the transformation acts on operators while √
the
state is fixed. Using quantum stochastic calculus one can derive the following differential equation for X(t)/ n
1
1
√ dX(t) = ψ(t)dt + √ dB(t).
n
n
√
Therefore, X(t)/ n is similar to the process (16) with the exception that it has a complex rather than real
valued drift function. Note that in this correspondence ψ(t) plays the role of f 1/2 , which agrees with the
intuitive interpretation of the wave function
√ as square
√ root of the state |ψihψ|. Alternatively, one can use the
Schrödinger picture, where the state is | nψi = W ( nψ)|0i, such that the process B(t) has the same law as
X(t) under the vacuum state.
In section 5.1 we show that the minimax rate of Gn for loss functions based on the norm-one and the Bures
distance, is n−α/(2α+1) . Although the rate is identical to that of the corresponding classical model, the result
does not follow from the classical case but relies on an explicit measurement strategy for the upper bounds,
and on the quantum local asymptotic equivalence Theorem 4.1 for the lower bound. Furthermore, the minimax
rate for the estimation of certain quadratic functionals are established in section 5.2, and the minimax testing
rates are derived in section 5.3. While the former are similar to the classical ones, the quantum testing rates
are parametric as opposed to non-parametric in the classical case. This reflects the fact that in the quantum
case, the optimal measurements for different statistical problems are in general incompatible with each other
and in some cases they differ significantly from what is expected on classical basis.
4. Local asymptotic equivalence for quantum models
In this section we prove that the sequence (20) of non-parametric pure states models is locally asymptotically
equivalent (LAE) with the sequence (21) of quantum Gaussian models, in the sense of the Le Cam distance.
This is one of the main results of the paper and will be subsequently used in the applications. Throughout the
section |ψ0 i is a fixed but arbitrary state in an infinite dimensional Hilbert space H. We let H0 := {|ψi ∈ H :
hψ0 |ψi = 0} denote the orthogonal complement of C|ψ0 i. Any vector state |ψi ∈ H decomposes uniquely as
p
|ui ∈ H0
(22)
|ψi = |ψu i := 1 − kuk2 |ψ0 i + |ui,
where the phase has been chosen such that the overlap hψ|ψ0 i is real and positive. Therefore, the pure states
are uniquely parametrised by vectors |ui ∈ H0 .
LAE OF PURE STATES ENSEMBLES AND QUANTUM GAUSSIAN WHITE NOISE
13
Further to the i.i.d. and Gaussian models Qn and Gn defined in (20) and respectively (21), we now introduce
their local counterparts which are parametrised by the local parameter |ui rather than by |ψi. Let γn be a
sequence such that γn = o(n−1/4 ), and define the pure state models
Qn (ψ0 , γn ) :=
Gn (ψ0 , γn ) :=
{|ψu⊗n i ∈ H⊗n : |ui ∈ H0 , kuk ≤ γn }
√
{|G( nu)i ∈ F (H0 ) : |ui ∈ H0 , kuk ≤ γn }.
(23)
(24)
The LAE Theorem below shows that these local models are asymptotically equivalent. An interesting fact is
that LAE holds without imposing global restrictions such as defined by the Sobolev classes, rather it suffices
that the local balls shrink at rate γn which is faster than n−1/4 . This contrasts with the classical case where
both types of conditions are needed, as explained in section 3.1. However, since the state cannot be “localised”
without any prior knowledge, in applications we need to make additional assumptions which allow us to work
in a small neighbourhood and make use of local asymptotic equivalence. In particular, the convergence holds
for the restricted models where the Sobolev condition is imposed on top of the local one. This will be used in
establishing the estimation, testing, and functional estimation results.
Theorem 4.1. Let Qn (ψ0 , γn ) and Gn (ψ0 , γn ) be the models defined in (23) and respectively (24). Then the
following convergence holds uniformly over states |ψ0 i:
lim sup sup ∆(Qn (ψ0 , γn ), Gn (ψ0 , γn )) = 0
(25)
n→∞ |ψ0 i∈H
where ∆(·, ·) is the quantum Le Cam distance defined in equation (18).
Proof. The direct map channel Tn is defined as an isometric embedding
Tn : T1 (H⊗s n ) → T1 (F (H0 ))
ρ 7→ Vn ρVn∗ .
where Vn : H⊗s n → F (H0 ) is an isometry defined below. Since we deal with pure states, it suffices to prove
that
√
lim sup sup sup Vn ψu⊗n − G( nu) = 0.
n→∞ |ψ0 i∈H kuk≤γn
We now define the isometric embedding Vn by showing its explicit action on the vectors of an ONB. For any
permutation σ ∈ Sn , let
Uσ : |u1 i ⊗ · · · ⊗ |un i 7→ |uσ−1 (1) i ⊗ · · · ⊗ |uσ−1 (n) i
P
1
⊗n
be the unitary action on H
by tensor permutations. Then Ps := n!
σ∈Sn Uσ is the orthogonal projector
onto the subspace of symmetric tensors H⊗s n . We construct an orthonormal basis in H⊗s n as follows.
Let B0 := {|e1 i, |e2 i, .P
. . } be an orthonormal basis in H0 . Let ñ = (n0 , n) = (n0 , n1 , . . . ) be an infinite sequence
of integers such that i≥0 ni = n, and note that only a finite number of ni s are different from zero. Then the
symmetric vectors
r
O
n!
Ps |ψ0 i⊗n0 ⊗
|ei i⊗ni
|ñi = |n0 , n1 , n2 , . . . i :=
n0 ! · n1 ! · . . .
i≥1
form an ONB of H
⊗s n
.
As discussed N
in section 2.2.2 the Fock space F (H0 ) can be identified with the infinite tensor product of one-mode
Fock spaces i≥1 F (C|ei i) which has an orthonormal number basis (or Fock basis) consisting of products of
number basis vectors of individual modes
O
|ni :=
|ni i
i≥1
where ni 6= 0 only for a finite number of indices. We define Vn : H⊗s n → F (H0 ) as follows
Vn : |ñi 7→ |ni.
Its image consists of states with at most n “excitations”, with |ψ0 i⊗n being mapped to the vacuum state |0i.
√ We
would like to show that the embedded state Vn |ψu i⊗n are well approximated by the coherent states |G( nu)i
uniformly over the local neighbourhood kuk ≤ γn . For this we will make use of the covariance and functorial
properties of the second quantisation construction in order to reduce the non-parametric LAE statement to the
corresponding one for 2-dimensional systems.
Let |ui ∈ H0 be a fixed unit vector. Let j : C2 7→ H be the isometric embedding
j : |0i 7→ |ψ0 i,
j : |1i 7→ |ui
14
LAE OF PURE STATES ENSEMBLES AND QUANTUM GAUSSIAN WHITE NOISE
and let j0 : C|1i → H0 be the restriction of j to the one dimensional subspace C|1i. Since second quantisation
is functorial under contractive maps, there is a corresponding isometric embedding J0 = Γ(j0 ) satisfying
J0 : F (C|1i) → F (H0 )
|G(α)i
2 ⊗s n
(26)
7→ |G(j0 (α))i = |G(αu)i.
Let Ṽn : C
→ F (C|1i) be the isometry constructed in the same way as Vn , where |0i plays the role of
p
|ψ0 i and C|1i is the analogue of H0 . As before, let |ψα i = 1 − |α|2 |0i + α|1i, with |α| ≤ 1. Then by the
properties of the embedding map Vn we have
From equations (26) and (27) we find
sup
|α|≤γn
(27)
J0 Ṽn |ψα i⊗n = Vn |ψαu i⊗n .
√
⊗n
− G( nαu) = sup
Vn ψαu
|α|≤γn
√
Ṽn ψα⊗n − G( nα)
Since the right-hand side of the above equality is independent of |ui the same equality holds with supremum
on the left side taken over all |ui ∈ H0 with kuk ≤ γn . Therefore the LAE for the non-parametric models has
been reduced to that of a two-dimensional (qubit) model which has already been established in [31]. Therefore
we obtain a first version of LAE in which the i.i.d and Gaussian models are expressed in terms of the local
parameter |ui
√
lim sup sup sup Vn ψu⊗n − G( nu) = 0.
n→∞ |ψ0 i∈H kuk≤γn
Conversely, we define the reverse channel Sn : T1 (F (H)) as follows. Let Pn denote the orthogonal projection in
F (H) onto the image space of Vn , i.e. the subspace with total excitation number at most n
X
F≤n (H0 ) := Lin{|n1 , n2 , . . . i :
ni ≤ n}.
i≥1
Let Rn : F (H0 ) → H
⊗s n
be a right inverse of Vn , i.e. Rn Vn = 1. Then the reverse channel is defined as
Sn (ρ) = Rn Pn ρPn Rn∗ + Tr(ρ(1 − Pn ))|ψ0 ihψ0 |⊗n .
Operationally, the action of Sn consists of two steps. We first perform a projection measurement with projections
Pn and (1 − Pn ); if the first outcome occurs the conditional state of the system is Pn ρPn /Tr(Pn ρn ) , while if the
second outcome occurs the state is (1 − Pn )ρ(1 − Pn )/Tr((1 − Pn )ρn ). In the second stage, if the first outcome
was obtained we map the projected state through the map Rn into a state in H⊗n , while if the second outcome
was obtained, we prepare the fixed state |ψ0 ihψ0 |⊗n .
√
When applied to the pure Gaussian states |G( nψu )i, the output of Sn is the mixed state
√
√
Sn (|G( nψu )ihG( nψu )|) = pnu |φnu ihφnu | + (1 − pnu )|ψ0 ihψ0 |⊗n
where
√
√
Rn Pn |G( nψu )i
:=
, pnu = kPn G( nψu )k2
√ n
pu
The key observation is that the Gaussian states are almost completely supported by the subspace F≤n (H0 ),
uniformly with respect to the ball kuk ≤ γn . This means that
|φnu i
lim sup sup sup (1 − pnu ) = 0
n→∞ |ψ0 i kuk≤γn
which together with the fact that Rn is the inverse of Vn implies
√
√
lim sup sup sup Sn (|G( nψu )ihG( nψu )|) − |ψu ihψu |⊗n
n→∞ |ψ0 i kuk≤γn
1
= 0.
This completes the proof of (25).
5. Applications
In this section we discuss three major applications of the local asymptotic equivalence result in Theorem 4.1,
namely to the estimation of pure states, estimation of a physically meaningful quadratic functional, and finally
to testing between pure states. We stress that local asymptotic equivalence allows us to translate these problems
into similar but easier ones involving Gaussian states. This strategy has already been successfully employed [31]
in finding asymptotically optimal estimation procedures for finite dimensional mixed states, which otherwise
appeared to be a difficult problem due to the complexity of the set of possible measurements.
LAE OF PURE STATES ENSEMBLES AND QUANTUM GAUSSIAN WHITE NOISE
15
As discussed in section 3.3, we will assume that we are given n independent systems, each prepared in a
state |ψi ∈ H belonging to the Sobolev ellipsoid S α (L) defined in equation (19). The corresponding quantum
statistical model Qn was defined in equation (20), and the Gaussian counterpart model Gn was defined in
equation (21).
Here is a summary of the results. In Theorem 5.2 we show that the estimation rates over such ellipsoids are
n−α/(2α+1) ; this is similar to the well-known rates, e.g. for density estimation, in nonparametric statistics
(see [64]). The estimation of the quadratic functional
X
F (ψ) =
|ψj |2 j 2β , for some fixed β > 0
j≥0
of the unknown pure state presents two regimes: a parametric rate n−1 for the MSE is attained when the
unknown state has enough "smoothness" (that is α ≥ 2β), whereas a nonparametric rate n−2(1−β/α) is obtained
when β < α < 2β. This double regime is known in nonparametric estimation for the density model, with
different values for both the rates and the values of the parameters where the phase-transition occurs, cf [15],
[45] and references therein.
Parametric rates and sharp asymptotic constants are obtained for the testing problem of a pure state against
an alternative described by the Sobolev-type ellipsoid with an L2 -ball removed. In the classical density model
only nonparametric rates for testing of order n−2α/(4α+1) can be obtained for the L2 norm. In our quantum
i.i.d. model, parametric rate n−1/2 is shown to be minimax for testing H0 : ψ = ψ0 , for some ψ0 in S α (L) over
the nonparametric set of alternatives:
H1 : ψ ∈ S α (L) is such that k|ψihψ| − |ψ0 ihψ0 |k1 ≥ cn−1/2 .
The sharp asymptotic constant we obtain for testing is specific for ensembles of pure states. As we discuss in
the sequel, quantum testing of states allows us to optimize over the measurements, and thus to obtain the most
distinguishable likelihoods for the underlying unknown quantum state.
5.1. Estimation. We consider the problem of estimating an unknown pure state belonging to the HermiteSobolev class S α (L) given an ensemble of n independent, identically prepared systems. The corresponding
sequence of statistical models Qn was defined in equation (20). We first describe a specific measurement
procedure which provides an estimator whose risk attains the nonparametric rate n−2α/(2α+1) . We prove the
lower bounds for estimating a Gaussian state in the model Gn defined in (21). Subsequently we use LAE to
establish a lower bound showing that the rate is optimal in the i.i.d. model as well.
Before deriving the bounds we briefly review the definitions of the loss functions used here and the relations
between them, cf. section 2.3. Recall that the trace norm distance between states ρ and ρ′ is given by
kρ − ρ′ k1 := Tr(|ρ − ρ′ |), and is the quantum analogue of the norm-one
p√ √ distance between probability densities.
ρρ′ ρ)), and is a quantum extension of the
The square of the Bures distance is given by d2b := 2(1 − Tr(
Hellinger distance. These distances satisfy the inequalities (14).
In the case of pure states (i.e. ρ = |ψihψ|, and ρ′ = |ψ ′ ihψ ′ |) these metrics become (cf. (12) and (13)),
p
d2b (ρ , ρ′ ) = 2(1 − |hψ|ψ ′ i|).
kρ − ρ′ k1 = 2 1 − |hψ|ψ ′ i|2 ,
Since vectors are not uniquely defined by the states, the distances cannot be expressed directly in terms of the
length kψ − ψ ′ k. However if we consider a reference vector |ψ0 i and define the representative vector |ψi such
that hψ0 |ψi ≥ 0, then we can write (as in section 4)
p
p
|ψu′ i = 1 − ku′ k2 |ψ0 i + |u′ i, |ui, |u′ i ⊥ |ψ0 i
|ψu i = 1 − kuk2 |ψ0 i + |ui,
and the distances have the same (up to a constant) quadratic approximation
d2b (ρu , ρu′ ) = ku − u′ k2 + O(max(kuk, ku′k)4 ),
kρu − ρu′ k21 = 4ku − u′ k2 + O(max(kuk, ku′k)4 ),
(28)
where the correction terms are of order 4 as kuk and ku k tend to 0. Below we show that asymptotically with n
the estimation risk for norm-one square and Bures distance square will have the same rate as that of estimating
the local parameter u with respect to the Hilbert space distance.
′
5.1.1. Upper bounds. We first describe a two steps measurement procedure, which provides an estimator whose
risk has rate n−2α/(2α+1) .
16
LAE OF PURE STATES ENSEMBLES AND QUANTUM GAUSSIAN WHITE NOISE
Theorem 5.1. Consider the i.i.d. quantum model Qn given by equation (20). There exists an estimator
ρbn := |ψbn ihψbn | such that
lim sup sup n2α/(2α+1) Eρ d2 (ρ̂n , ρ) ≤ C,
n→∞ |ψi∈S α (L)
where ρ := |ψihψ|, d(ρ̂n , ρ) denotes either the trace-norm distance, or the Bures distance, and C > 0 is a
constant depending only on α > 0 and L > 0.
Proof of Theorem 5.1. According to inequalities (14) and (15) the two distances are equivalent on pure states,
so it suffices to prove the upper bound for the trace-norm distance.
Firstly, a projective operation is applied to each of the n copies separately, whose aim is to truncate the
state to a finite dimensional subspace of dimension dn = [n1/(2α+1) ] + 1. Let Pn be the projection onto the
subspace Hn spanned by the first dn basis vectors {|e0 i, . . . , |edn −1 i}. For a given state |ψi the operation
consists of randomly projecting the state with Pn or (1 − Pn ), which produces i.i.d. outcomes Oi ∈ {0, 1} with
P(Oi = 1) = pn = kPn ψk2 . The posterior state conditioned on the measurement outcome is
Pn |ψihψ|Pn
(n)
(n)
with probability
pn
|ψ ihψ | :=
pn
|ψihψ| 7→
(1−Pn )|ψihψ|(1−Pn )
with probability
1 − pn
1−pn
Since |ψihψ| ∈ S α (L), the probability 1 − pn is bounded as
1 − pn =
∞
X
i=dn
|ψi |2 =
∞
X
i=dn
i−2α i2α |ψi |2 ≤ d−2α
n
∞
X
i=1
i2α |ψi |2 = n−2α/(2α+1) L.
(29)
Pn
Let ñ = i=1 Oi be the number of systems for which the outcome was equal to 1, so that ñ has binomial
distribution Bin(n, pn ). Then E(ñ/n) = pn and Var(ñ/n) = pn (1 − pn )/n = O(1/n). Therefore ñ/n →1 in
probability.
In the second step we discard the systems for which the outcome was 0, and we collect those with outcome 1, so
that the joint state is |ψ (n) ihψ (n) |⊗ñ which is supported by the symmetric subspace Hn⊗s ñ . In order to estimate
the truncated state |ψ (n) i (and by implication |ψi), we perform a covariant measurement Mn [35] whose space
of outcomes is the space of pure states ρ̂n = |ψ̂n ihψ̂n | over Hn , and the infinitesimal POVM element is
ñ + dn − 1 ⊗n
Mn (dρ̂) =
ρ̂ dρ̂.
(30)
dn − 1
The covariance property means that the unitary group has a covariant action on states and their corresponding
probability distributions
∗
∗
Mn
n
PM
UρU ∗ (dρ̂) = Tr(U ρU · dρ̂) = Pρ (d(U ρ̂U )).
Recall that the trace-norm distance squared for pure states is given by d21 (ρ, ρ′ ) := kρ − ρ′ k21 = 4(1 − |hψ|ψ ′ i|2 ).
In [35] it has been shown that, conditionnally on ñ, the risk of the estimator ρ̂ with respect to the trace-norm
square distance is1
h
i 4(d − 1)
n
.
Eñ d21 (ρ̂n , ρ(n) ) =
dn + ñ
√
Using the triangle inequality we have d21 (ρ̂n , ρ) ≤ 2(d21 (ρ̂n , ρ(n) ) + d21 (ρ, ρ(n) )). Since |ψ (n) i = Pn |ψi/ pn , the
bias term is d21 (ρ, ρ(n) ) = 4(1 − pn ), which by (29) is bounded by 4n−2α/(2α+1) L. Therefore
2
(dn − 1)
+ 8n−2α/(2α+1) L.
E db (ρ̂n , ρ) ≤ 8E
dn + ñ
For an arbitrary small ε > 0, we have
(dn − 1)
dn
ñ
ñ
(dn − 1)
1
E
≤P
+C .
<1−ε +E
· I( ≥ 1 − ε) ≤ O
dn + ñ
n
dn + n · ñ/n
n
n
n
Putting together the last two upper bounds concludes the proof.
1Reference [35] uses a fidelity distance erroneously called “Bures distance" , which for pure states coincides with the trace-norm
distance up to a constant
LAE OF PURE STATES ENSEMBLES AND QUANTUM GAUSSIAN WHITE NOISE
17
5.1.2. Lower bounds - Unimprovable rates. We will first consider the Gaussian model Gn given by equation (21)
which is indexed by Hilbert √
space vectors ψ ∈ H in the Sobolev class S α (L), playing the role of means of
quantum Gaussian states |G( nψ)i. In Theorem 5.2 we find a lower bound for the mean square error of any
estimator ψ̂. This is then used in conjunction with the local asymptotic equivalence Theorem 4.1 to obtain a
lower bound for the risk of the i.i.d. model Qn , with respect to the norm-one and Bures distances.
Theorem 5.2. Consider the quantum Gaussian model Gn given by equation (21). There exists some constant
c > 0 depending only on α and L such that
i
h
lim inf inf sup n2α/(2α+1) Eψ kψbn − ψk22 ≥ c,
n→∞ ψ
bn ψ∈S α (L)
where the infimum is taken over all estimators ψbn , understood as combination of measurements and classical
estimators.
The proof is given in Section 6.
We now proceed to consider the i.i.d. model Qn defined in (20). We are given n copies of an unknown pure state
|ψihψ|, with ψ in the Sobolev class S α (L). The goal is to find an asymptotic lower bound for the estimation risk
(with respect to the Bures or norm-one loss functions) which matches the upper bound derived in section 5.1.1.
Since both loss functions satisfy the triangle inequality, it can be shown that by choosing estimators which are
mixed states, rather than pure states, one can improve the risk by at most a constant factor 2. Therefore we
consider estimators which are pure states. In order to fix the phase of the vector representing the true and the
estimated state, we will assume that hψ|e0 i ≥ 0 and hψ̂|e0 i ≥ 0.
Theorem 5.3. Consider the i.i.d. quantum model Qn given by equation (20). There exists some constant c > 0
depending only on α > 1/2 and L > 0 such that
lim inf inf
sup n2α/(2α+1) Eρ d2 (ρ̂n , ρ) ≥ c,
n→∞ |ψ
bn i |ψi∈S α (L)
where ρ := |ψihψ|, the infimum is taken over all estimators ρbn := |ψbn ihψbn | (defined by a combination of
measurement and a classical estimator), and the loss function d(ρ̂, ρ) is either the norm-one or the Bures
distance.
The proof is given in Section 6.
5.2. Quadratic functionals. This section deals with the estimation of the quadratic functional
X
F (ψ) =
|ψj |2 · j 2β , for some fixed 0 < β < α,
j≥0
which is well defined for all pure states |ψi in the ellipsoid S α (L). If the Hilbert space H is represented as
L2 (R) and {|ji : j ≥ 0} is the Fock basis (cf. section 2.2.1) then F (ψ) is the moment of order 2β of the number
operator N :
F (ψ) = Tr(|ψihψ| · N 2β ).
Below we derive upper and lower bounds for the rate of the quadratic risk for estimating F (ψ), which is of order
n−1 if α ≥ 2β, and n−2(1−β/α) if β < α < 2β.
5.2.1. Upper bounds. Let us describe an estimator Fbn of F (ψ) in the quantum i.i.d. model. We
P consider the
measurement of the number operator with projections {|jihj|}j≥0 . For a pure state |ψi =
j≥0 ψj |ji, we
obtain an outcome X taking values j ∈ N with probabilities pj := Pψ (X = j) = |ψj |2 , for j ≥ 0. By measuring
each quantum sample |ψi separately, we obtain i.i.d. copies X1 , . . . , Xn of X, allowing us to estimate each pj
empirically, by
n
1X
p̂j =
I(Xk = j), j ≥ 0.
n
k=1
which is an unbiased estimator of pj with variance pj (1 − pj )/n. The estimator of the quadratic functional is
defined as
N
X
p̂j · j 2β
(31)
Fbn =
j=1
for an appropriately chosen truncation parameter N defined below. The next theorem, shows that a parametric
rate can be attained for estimating the quadratic functional F (ψ) if α ≥ 2β, whereas a nonparametric rate is
attained if β < α < 2β.
18
LAE OF PURE STATES ENSEMBLES AND QUANTUM GAUSSIAN WHITE NOISE
Theorem 5.4. Consider the i.i.d. quantum model Qn given by equation (20). Let Fbn be the estimator (31) of
F (ψ) with N ≍ n1/4(α−β) , for α ≥ 2β, respectively N ≍ n1/2α , for β < α < 2β. Then
−1
2
n ,
if α ≥ 2β
2
2
where ηn =
sup Eψ F̂n − F (ψ) = O(1) · ηn ,
(32)
−2(1−β/α)
n
,
if
β < α < 2β.
α
ψ∈S (L)
Proof of Theorem 5.4. The usual bias-variance decomposition yields
2
2
Eψ Fbn − F (ψ) = Eψ Fbn − F (ψ) + V arψ Fbn .
The bias can be upper bounded as
F (ψ) − Eψ Fbn
=
F (ψ) −
N
X
j=1
pj · j 2β =
X
j≥N +1
pj · j 2β ≤ N −2(α−β)
X
j≥N +1
pj · j 2α ≤ LN −2(α−β) .
For the variance, let us note that the vector
Vb = n · (p̂1 , . . . , p̂N , p̂∗N +1 ),
with pb∗N +1 = n−1
n
X
k=1
I(Xk ≥ N + 1),
P
has a multinomial distribution with parameters n and probability vector V := (p1 , . . . , pN , p∗N +1 = j≥N +1 pj )⊤ .
The covariance matrix of a multinomial vector writes n·(Diag(V )−V ·V ⊤ ), where Diag(V ) denotes the diagonal
matrix with entries from V . In particular, if pb := (p̂1 , ..., p̂N )⊤ , p := (p1 , ..., pN )⊤ and B := (1, 22β , ..., N 2β )⊤
then
1
Covψ (Fbn ) = Covψ (B ⊤ · pb) = B ⊤ · Covψ (b
p) · B = · B ⊤ · (Diag(p) − p · p⊤ ) · B.
n
This gives
N
1
1X
pj · j 4β .
Covψ (Fbn ) ≤ · B ⊤ · Diag(p) · B =
n
n j=1
The bound of this last term and the resulting bound of the risk is treated separately for the two cases.
a) Case α ≥ 2β. In that case,
N
X
j=1
pj · j 4β ≤
N
X
L
pj · j 2α ≤ L implying that V ar(Fbn ) ≤ .
n
j=1
The upper bound of the risk is, in this case,
2
L
Eψ Fbn − F (ψ) ≤ L2 N −4(α−β) + .
n
If we choose N ≍ n1/(4(α−β)) or larger, then the parametric rate is attained for the risk:
2
Eψ Fbn − F (ψ) = O(1) · n−1 .
b) Case β < α < 2β. Here we have,
N
Covψ (Fbn ) ≤
N
1X
N 4β−2α
1X
pj · j 4β ≤
pj · j 4β−2α j 2α pj ≤
L.
n j=1
n j=1
n
The upper bound of the risk becomes
2
N 4β−2α
Eψ Fbn − F (ψ) ≤ L2 N −4(α−β) +
L.
n
The optimal choice of the parameter N that balances the two previous terms is N ≍ n1/(2α) , giving the attainable
rate for the quadratic risk
2
Eψ Fbn − F (ψ) = O(1) · n−2(1−β/α) .
Cases a) and b) together prove that the rate ηn2 is attainable.
LAE OF PURE STATES ENSEMBLES AND QUANTUM GAUSSIAN WHITE NOISE
19
5.2.2. Lower bounds. The next Theorem proves the optimality of the previously attained rate for the estimation
of quadratic functionals.
Theorem 5.5. Consider the i.i.d. quantum model Qn given by equation (20). Then, there exists some constant
c > 0 depending only on α, β (with α > β > 0), and L > 0 such that
lim inf inf
sup
n→∞ F
bn ψ∈S α (L)
2
ηn−2 · Eψ Fbn − F (ψ) ≥ c,
where the infimum is taken over all measurements and resulting estimators Fbn of F (ψ).
The proof is given in Section 6.
5.3. Testing. In the problem of testing for signal in classical Gaussian white noise, over a smoothness class with
an L2 -ball removed, minimax rates of convergences (separation rates) are well known [41]; they are expressed
in the rate of the ball radius tending to zero along with noise intensity, such that a nontrivial asymptotic power
is possible. We will consider an analogous testing problem here for pure states. Accordingly, let ρ = |ψihψ|
denote pure states, let ρ0 = |ψ0 ihψ0 | be a fixed pure state to serve as the null hypothesis, and let
(33)
B (ϕ) = {kρ − ρ0 k1 ≥ ϕ}
be the complement of a trace norm ball around ρ0 . We want to test in the i.i.d. quantum model Qn given by
equation (20) the following hypotheses about a pure state ρ :
H0 :
ρ = ρ0
H1 (ϕn ) : ρ ∈ S α (L) ∩ B (ϕn )
(34)
for {ϕn }n≥1 a decreasing sequence of positive real numbers. Consider a binary POVM M = (M0 , M1 ), acting
on the product states ρ⊗n , cf. Definition 2. We denote the testing risk between two fixed hypotheses by the
sum of the two error probabilities
⊗n
⊗n
RnM = RnM (ρ⊗n
, M ) = Tr(ρ⊗n
· M0 ).
0 ,ρ
0 · M1 ) + Tr(ρ
In the minimax α-testing approach which dominates the literature on the classical Gaussian white noise case, one
⊗n
would require Tr(ρ⊗n
·
0 · M1 ) ≤ α while trying to minimize the worst case type 2 error supρ∈S α (L)∩B(ϕn ) Tr(ρ
M0 ). However we will consider here the so-called detection problem [40] where the target is the worst case total
error probability
PM
e (ϕn ) =
sup
ρ∈S α (L)∩B(ϕn )
⊗n
RnM (ρ⊗n
, M ) = Tr(ρ0⊗n · M1 ) +
0 ,ρ
sup
ρ∈S α (L)∩B(ϕn )
Tr(ρ⊗n · M0 ).
The minimax total error probability is then obtained by optimizing over T :
P∗e (ϕn ) =
inf
M binary POVM
PM
e (ϕn ) .
5.3.1. Separation rate. A sequence {ϕ∗n }n≥1 is called a minimax separation rate if any other sequence {ϕn }n≥1
fulfills
P∗e (ϕn ) → 1 if ϕn /ϕ∗n → 0 and P∗e (ϕn ) → 0 if ϕn /ϕ∗n → ∞.
(35)
Below we establish that ϕ∗n = n−1/2 is a separation rate in the current problem, even though the alternative
H1 (·) in (34) is a nonparametric set of pure states. Recall relations (11), (12) describing the total optimal error
for testing between simple hypotheses given by two pure states.
Theorem 5.6. Consider the i.i.d. quantum model Qn given by equation (20), and the testing problem (34).
Assume that ρ0 is in the interior of S α (L), i.e ρ0 ∈ S α (L′ ) for some L′ < L. Then ϕ∗n = n−1/2 is a minimax
separation rate.
The proof is given in the Section 6.
20
LAE OF PURE STATES ENSEMBLES AND QUANTUM GAUSSIAN WHITE NOISE
5.3.2. Sharp asymptotics. Having identified the optimal rate of convergence in the testing problem (i.e. the
minimax separation rate), we will go a step further and aim at a sharp asymptotics for the minimax testing
error. We will adopt the approach of [20], extended in [41], where testing analogs of the Pinsker-type sharp
risk asymptotics in nonparametric estimation were obtained. The result will be framed as follows: if the radius
is chosen ϕn ∼ cn−1/2 for a certain c > 0, then the minimax testing error behaves as P∗e (ϕn ) ∼ exp −c2 /4 .
Thus the sharp asymptotics is expressed as a type of scaling result: a choice of constant c in the radius implies
a certain minimax error asymptotics depending on c.
To outline the problem, consider the upper and lower error bounds obtained in the proof of the separation rate.
In (62) we obtained the bound
n
(ϕn ) ≤ exp −c2n /4
(36)
PM
e
⊗n
if ϕn = cn n−1/2 , where Mn was the sequence of projection tests Mn = (ρ⊗n
0 , I − ρ0 ). The lower risk bound
obtained in the course of proving Theorem 5.6 was
q
inf
PM
(ϕ
)
≥
1
−
1 − (1 − c2n n−1 /4)n .
n
e
M binary POVM
If cn = c we can summarize this as
p
1 − 1 − exp (−c2 /4) + o (1) ≤ P∗e (ϕn ) ≤ exp −c2 /4 .
Our result will be that the upper bound is sharp and represents the minimax risk asymptotics.
Theorem 5.7. Consider the i.i.d. quantum model Qn given by equation (20), and the testing problem (34).
Assume that ρ0 ∈ S α (L′ ) for some L′ < L. At the minimax separation rate for the radius, i.e. for ϕn ≍ n−1/2
we have
lim nϕ2n log P∗e (ϕn ) = −1/4.
n
The proof is given in the Section 6.
5.4. Discussion.
State estimation.
Tomography and optimal rates. Consider a model where the Sobolev-type assumption ρ ∈ S α (L) about the
pure state ρ = |ψi hψ| (cf. (19)) is replaced by a finite dimensionality assumption: ρ ∈ Hd where
Hd = {|ψi hψ| : ψj = 0, j ≥ d}
and d is known. One observes n identical copies of the pure state ρ = |ψi hψ|, with possibly d = dn → ∞, i. e.
the model Qn of (20) is replaced by
Qn := ρ⊗n : ρ ∈ Hd .
Since Hd can be written Hd = S1,d where
Sr,d := {ρ : hei |ρ|ej i = 0, i, j ≥ d, rank(ρ) = r} ,
the model is effectively a special case of the d × d density matrices of rank(ρ) = r considered in [47]. In [47]
however, it is not known in advance that r = 1 but ρ is a density matrix of possibly low rank r, and the aim
is estimation of ρ using quantum state tomography performed on n identical copies of ρ. Data are obtained by
defining an observable ⊗ni=1 Ei where E1 , . . . , En are i.i.d. uniformly selected elements of the Pauli basis of the
linear space of d × d Hermitian matrices, and applying the corresponding measurement to ρ⊗n . Let ρ̂∗n denote
an arbitrary estimator of ρ based on that measurement. A lower asymptotic risk bound for norm-one risk is
established; in the special case d2 r2 = o (n) it reads as
h
i
r2 d2
2
∗
kρ̂
−
ρk
≥
c
inf
sup
E
(37)
ρ
n
1
ρ̂∗
n
n ρ∈Sr,d
for some c > 0 (Theorem 10 in [47]). It is also shown in [47] that (37) is attained, up to a different constant
and logarithmic terms, by an entropy penalized least squares type estimator based on measurement of ⊗ni=1 Ei ,
even when the rank r is unknown. Analogous optimal rates for d × d mixed states ρ with Pauli measurements,
but under sparsity assumptions on the entries of the matrix ρ have been obtained in [16].
Returning to our setting of pure states, where r = 1 is known, with an infimum over all measurements of ρ⊗n
and corresponding estimators ρ̂n , according to [35] one has
h
i 4 (d − 1)
2
inf sup Eρ kρ̂n − ρk1 =
(38)
ρ̂n ρ∈S1,d
d+n
and the bound is attained by an estimator of the pure state ρ based on the covariant measurement (cf. equation
(30) ). Comparing (37) for r = 1 and dn → ∞, dn = o (n) with (38), we find that the latter bound is of
LAE OF PURE STATES ENSEMBLES AND QUANTUM GAUSSIAN WHITE NOISE
21
order dn /n whereas the former is of order d2n /n. It means that for estimation of finite dimensional pure states,
estimators based on the Pauli type measurement ⊗ni=1 Ei do not attain the optimal rate when dn → ∞. It
may be conjectured that the same holds for the optimal rate over ρ ∈ S α (L), i.e. our rate of Theorem 5.1.
We emphasize again that our results establish lower asymptotic risk bounds over all quantum measurements
and estimators, whereas lower risk bounds within one specific measurement scheme [46] [47] [16] are essentially
results of non-quantum classical statistics.
Separate measurements. A notable fact is also that ⊗ni=1 Ei is a separate (or local) measurement, i.e. produces
independent random variables (or random elements) Y1 , . . . , Yn each based on a measurement of a copy of
ρ, whereas the covariant measurement (cp. (30)) we used for attainment our risk bound of Theorem 5.1 is of
collective (or joint) type with regard to the product ρ⊗n . Separate measurements are of interest from a practical
point of view since collective measurements of large quantum systems may be unfeasible in implementations
[51]. In [5] it is shown that for fixed d = 2, the bound (38) can be attained asymptotically as n → ∞ (up to a
factor 1 + o (1)) by a separate measurement of ρ⊗n ; it is an open question whether in our infinite dimensional
setting, the optimal rate of Theorem 5.1 can be attained by a separate measurement. For mixed qubits (d = 2),
an asymptotic efficiency gap between separate and collective measurements is known to exist [4].
Quadratic functionals.
The elbow phenomenon. The change of regime which occurs in the optimal MSE rate ηn2 in (32) has been
described as the elbow phenomenon in the literature [15]. In the classical Gaussian sequence model, it takes
the following shape. Consider observations introduced in (17):
yj = ϑj + n−1/2 ξj , j = 1, 2, . . . ,
P∞ 2α 2
where {ξj } are i.i.d. standard normal, and the parameter ϑ = (ϑj )∞
j=1 j ϑj ≤ L
j=1 satisfies a restriction
P
∞
2β 2
for some α > 0. For estimation of the quadratic functional F̃ (ϑ) = j=1 j ϑj with β < α, the minimax MSE
rate of convergence is
(
n−1 if α ≥ 2β + 1/4
1 4 (α − β)
−2r̃
2
4(α−β)
=
n
for
r̃
=
min
η̃n =
,
2 4α + 1
n−2 4α+1 if β < α < 2β + 1/4
(cf [45] and references cited therein). The same rate holds for estimation of the squared L2 -norm of the β-th
2
derivative of a density inan α-Hölder
class, cf. [7]. Comparing with our rate ηn in (32) which can be written
4(α−β)
1
, we see that both rates exhibit the elbow phenomenon, but at different
ηn2 = n−2r for r = min 2 , 4α
critical values for (α, β), and the rate for the quantum case is slightly faster in the region α < 2β + 1/4.
A tail functional of a discrete distribution. Our method of proof for the optimal rate ηn2 = n−2r shows that
∞
it is also the optimal rate in the following non-quantum problem: suppose P = {pj }j=0 is a probability
P∞ 2α
measure on the nonnegative integers, satisfying a restriction j=0 j pj ≤ L, and the aim is to estimate the
P∞
linear functional F0 (P ) = j=0 j 2β pj (which might be called a linear tail functional) on the basis of n i.i.d.
PN
observations X1 , . . . , Xn having law P . Indeed, Theorem 5.4 shows that the estimator F̂n = j=0 j 2β p̂j with
Pn
p̂j = n−1 i=1 I (Xi = j) attains the rate ηn2 for mean square error, for an appropriate choice of N . On the
other hand, the observations X1 , . . . , Xn are obtained from one specific measurement in the quantum model
2
(20), in such a way that pj = |ψj | for j ≥ 0 and F0 (P ) = F (ψ). If the rate ηn2 is unimprovable in the quantum
model then it certainly is in the present derived (less informative) classical model. In the latter model, we
note that since F0 (P ) is linear and the law P is restricted to a convex body, optimality of the rate ηn2 can be
confirmed by standard methods, e.g. based on the concept of modulus of continuity [19]. The current problem is
thus an example where the elbow phenomenon is present for estimation of a linear functional; a specific feature
here is that the probability measure P is discrete.
Fuzzy quantum hypotheses. Our method of proof of the lower bound for quadratic functionals, which works in
the approximating quantum Gaussian model, utilizes the well-known idea of setting up two prior distributions
and then invoking a testing bound between simple hypotheses. This has been described as the method of fuzzy
hypotheses in the literature [64]. A summary of the present quantum variant could be as follows. First, the
Gaussian quantum model is represented in a fashion analogous to the classical sequence model (17) where the
ϑj correspond to the displacement parameter uj in certain Gaussian pure states (the coherent states). These
displacement parameters are then assumed to be random as independent, non-identically distributed normal,
for j = 1, . . . , N where N = o(n). Now Gaussian averaging over the displacements uj leads to certain non-pure
Gaussian states, i.e. the thermal states as the alternative, which happen to commute with the vacuum pure state
(corresponding to uj = 0) as the null hypothesis. Even though both are again Gaussian states, by commutation
22
LAE OF PURE STATES ENSEMBLES AND QUANTUM GAUSSIAN WHITE NOISE
the problem is reduced to testing between two ordinary discrete probability distributions, i.e. the point mass
at 0 and a certain geometric distribution with parameter rj , depending on j = 1, . . . , N . The combined error
probability for this classical testing problem with N independent observations gives the lower risk bound.
Nonparametric testing.
The separation rate n−1/2 . Recall that for the classical Gaussian sequence model (17), for the testing problem
H0 :
ϑ
0
P=
∞
2α 2
H1 (ϕn ) :
j=1 j ϑj ≤ L and kϑk2 ≥ ϕn
(39)
(Sobolev ellipsoid with an L2 -ball removed), the separation rate is ϕn = n−2α/(4α+1) [41]. We established that
ϕn = n−1/2 is the separation rate for the quantum nonparametric testing problem (34) involving a pure state
ρ. While this “parametric” rate for a nonparametric problem is somewhat surprising, it should be noted that
there also exist testing problems for classical i.i.d. data with nonparametric alternative where that separation
rate applies; cf [41], sec. 2.6.2.
In our case, the rate n−1/2 appears to be related to the fast rate ϕ2n = n−1 in the following nonparametric
∞
classical problem: given n i.i.d. observations X1 , . . . , Xn having law P = {pj }j=0 on the nonnegative integers,
the hypotheses are
H0 :
P = δ0 (the degenerate law at 0)
(40)
H1 (ϕn ) : kP − δ0 k1 ≥ ϕ2n .
For that, note first that
∞
X
pj = 2 (1 − p0 ) .
kP − δ0 k1 = 1 − p0 +
j=1
The likelihood ratio test for δ0 against any P ∈ H1 (ϕn ) rejects if max1≤j≤n Xj > 0, thus it does not depend on
P . The pertaining sum of error probabilities is
n
n
1
1
≤ 1 − ϕ2n
P max Xj = 0 = pn0 = 1 − kP − δ0 k1
1≤j≤n
2
2
and with a supremum over P ∈ H1 (ϕn ), the upper bound is attained. This means that for ϕn = cn−1/2 ,
the minimax sum of error probabilities tends to exp −c2 /2 , so that ϕ2n = n−1 is the separation rate here as
claimed.
In fact there is a direct connection to the quantum nonparametric testing problem (34): in the latter, for n = 1,
∞
consider a measurement defined as follows. Let {|ẽj i}j=0 be an orthonormal basis in H such that ρ0 = |ẽ0 i hẽ0 |
∞
and consider the POVM {|ẽj i hẽj |}j=0 ; the corresponding measurement yields a probability measure P on the
∞
nonnegative integers. Here the state ρ0 is mapped into δ0 and an alternative state ρ is mapped into P = {pj }j=0
such that p0 = Tr (ρ0 ρ). Condition (33) on the distance of the two states implies (cp (12))
q
p
p
ϕn ≤ kρ − ρ0 k1 = 2 1 − Tr (ρ0 ρ) = 2 1 − p0 = 2 kP − δ0 k1
so that up to a constant, the testing problem (40) is obtained.
In the quantum problem (34), we noted that the optimal
test between ρ0 and a specific alternative ρ depends
⊗n
on ρ, but found that the test (binary POVM) Mn = ρ⊗n
is minimax optimal in the sense of the rate
0 , I − ρ0
and also in the sense of a sharp risk asymptotics. The sharp minimax optimality seems to be a specific result
∞
for the quantum case. We note that the optimal test Mn can be realized via a measurement {|ẽj i hẽj |}j=0 as
⊗n
described above, applied separately to each component of ρ , resulting in independent identically distributed
r.v.’s X1 , . . . , Xn . The test Mn then amounts to rejecting H0 if max1≤j≤n Xj > 0. Note that this measurement
is incompatible with the one (30) providing the optimal rate for state estimation.
Other separation rates. In our proof of the lower bound for quadratic functionals, we formulated
the nonparaP
metric testing problem for pure states (50) where the alternative includes the restriction j≥0 |ψj |2 j 2β ≥ ηn ,
and established that the rate ηn = n−1+β/α is unimprovable there. Introduce a seminorm
1/2
X
2
kψk =
|ψj | j 2β
2,β
j≥1
LAE OF PURE STATES ENSEMBLES AND QUANTUM GAUSSIAN WHITE NOISE
23
(excluding the term for j = 0) and write the restriction as
kψk2,β ≥ ϕn = ηn1/2 ;
(41)
then the case β = 0 gives (cp (12))
X
1
2
2
2
2
ϕ2n ≤
|ψj | = 1 − |ψ0 | = 1 − |hψ|e0 i| = k|e0 i he0 | − |ψi hψ|k1 ,
4
j≥1
in other words, for ρ0 = |e0 i he0 | and ρ = |ψi hψ|, the restriction (41) is equivalent to kρ − ρ0 k1 ≥ 2ϕn . In
that sense, the testing problems (34) and (50) are equivalent up to a constant, if β = 0 and ρ0 = |e0 i he0 |. For
β > 0, the testing problem (50) is a quantum pure state analog of the generalization of the classical problem
(39) where kϑk2 ≥ ϕn is replaced by kϑk2,β ≥ ϕn (α-ellipsoid with a β-ellipsoid removed); the separation rate
in the latter is ϕn = n−2(α−β)/(4α+1) , cf. [41], sec. 6.2.1. In (50) the separation rate is ϕn = n−1/2+β/2α , i.e.
of the more typical nonparametric form as well.
References
[1] Artiles, L, Gill, R., and Guţă, M. An invitation to quantum tomography. J. Royal Statist. Soc. B (Methodological), 67:109–134, 2005.
[2] K. M. R. Audenaert, J. Calsamiglia, R. Muñoz Tapia, E. Bagan, Ll. Masanes, A. Acin, and F. Verstraete.
Discriminating states: The quantum Chernoff bound. Phys. Rev. Lett., 98:160501, 2007.
[3] K. M. R. Audenaert, M. Nussbaum, A. Szkola, and F. Verstraete. Asymptotic error rates in quantum
hypothesis testing. Commun. Math. Phys., 279:251–283, 2008.
[4] E. Bagan, M. A. Ballester, R. D. Gill, R. Muñoz Tapia, and O. Romero-Isart. Separable measurement
estimation of density matrices and its fidelity gap with collective protocols. Phys. Rev. Lett., 97:130501,
Sep 2006.
[5] E. Bagan, A. Monras, and R. Muñoz Tapia. Comprehensive analysis of quantum pure-state estimation for
two-level systems. Phys. Rev. A, 71:062318, Jun 2005.
[6] O. E. Barndorff-Nielsen, Gill, R., and Jupp, P. E. On quantum statistical inference (with discussion). J.
R. Statist. Soc. B, 65:775–816, 2003.
[7] P. J. Bickel and Y. Ritov. Estimating integrated squared density derivatives: sharp best order of convergence
estimates. Sankhyā Ser. A, 50(3):381–393, 1988.
[8] B. Bongioanni and J. L. Torrea. Sobolev spaces associated to the harmonic oscillator. Proc. Indian Acad.
Sci. Math. Sci., 116(3):337–360, 2006.
[9] Bruno Bongioanni. Sobolev spaces diversification. Rev. Un. Mat. Argentina, 52(2):23–34, 2011.
[10] S. L. Braunstein and Caves C. M. Statistical distance and the geometry of quantum states. Phys. Rev.
Lett., 72:3439–3443, 1994.
[11] Lawrence D. Brown, T. Tony Cai, Mark G. Low, and Cun-Hui Zhang. Asymptotic equivalence theory
for nonparametric regression with random design. Ann. Statist., 30(3):688–707, 2002. Dedicated to the
memory of Lucien Le Cam.
[12] Lawrence D. Brown and Mark G. Low. Asymptotic equivalence of nonparametric regression and white
noise. Ann. Statist., 24(6):2384–2398, 1996.
[13] C. Butucea, M. Guţă, and L. Artiles. Minimax and adaptive estimation of the Wigner function in quantum
homodyne tomography with noisy data. Annals of Statistics, 35:465–494, 2007.
[14] C. Butucea, M. Guţă, and T. Kypraios. Spectral thresholding quantum tomography for low rank states.
New Journal of Physics, 17:113050, 2015.
[15] T. Tony Cai and Mark G. Low. Nonquadratic estimators of a quadratic functional. Ann. Statist.,
33(6):2930–2956, 2005.
[16] Tony Cai, Donggyu Kim, Yazhen Wang, Ming Yuan, and Harrison H. Zhou. Optimal large-scale quantum
state tomography with Pauli measurements. Ann. Statist., 44(2):682–712, 2016.
[17] A. Carpentier, J. Eisert, D. Gross, and R. Nickl. Uncertainty Quantification for Matrix Compressed Sensing
and Quantum Tomography Problems. arXiv:1504.03234.
[18] M. Christandl and R. Renner. Reliable quantum state tomography. Phys.Rev. Lett., 109:120403, 2012.
[19] David L. Donoho and Richard C. Liu. Geometrizing rates of convergence ii,iii. Ann. Statist., 19(2):633–667,
668–701, 1991.
[20] M. S. Ermakov. Minimax detection of a signal in Gaussian white noise. Teor. Veroyatnost. i Primenen.,
35(4):704–715, 1990.
[21] P. Faist and R. Renner. Practical and reliable error bars in quantum tomography. Phys. Rev. Lett.,
117:010404, 2016.
[22] S. T. Flammia, D. Gross, Y.-K. Liu, and J. Eisert. Quantum Tomography via Compressed Sensing: Error
Bounds, Sample Complexity, and Efficient Estimators. New. J. Phys., 14:095022, 2012.
24
LAE OF PURE STATES ENSEMBLES AND QUANTUM GAUSSIAN WHITE NOISE
[23] C. A. Fuchs and J. van de Graaf. Cryptographic distinguishability measures for quantum-mechanical states.
IEEE Transactions Information Theory, 45, 1999.
[24] R. D. Gill and S. Massar. State estimation for large ensembles. Phys. Rev. A, 61:042312, 2000.
[25] Richard D. Gill and Mădălin I. Guţă. On asymptotic quantum statistical inference. In From probability to
statistics and back: high-dimensional models and processes, volume 9 of Inst. Math. Stat. (IMS) Collect.,
pages 105–127. Inst. Math. Statist., Beachwood, OH, 2013.
[26] Georgi K. Golubev, Michael Nussbaum, and Harrison H. Zhou. Asymptotic equivalence of spectral density
estimation and Gaussian white noise. Ann. Statist., 38(1):181–214, 2010.
[27] Ion Grama and Michael Nussbaum. Asymptotic equivalence for nonparametric generalized linear models.
Probab. Theory Related Fields, 111(2):167–214, 1998.
[28] D. Gross. Recovering Low-Rank Matrices From Few Coefficients in Any Basis. IEEE Transactions on
Information Theory, 57:1548–1566, 2011.
[29] D. Gross, Y.-K. Liu, S.T. Flammia, S. Becker, and J. Eisert. Quantum State Tomography via Compressed
Sensing. Physical Review Letters, 105:150401, 2010.
[30] M. Guţă, B. Janssens, and J. Kahn. Optimal estimation of qubit states with continuous time measurements.
Commun. Math. Phys., 277:127–160, 2008.
[31] M. Guţă and J. Kahn. Local asymptotic normality for qubit states. Phys. Rev. A, 73(5):052108, 2006.
[32] M. Guta and J. Kahn. Local asymptotic normality and optimal estimation for d-dimensional quantum
systems. In Quantum Stochastics and Information: statistics, filtering and control, pages 300–322. World
Scientific, 2008.
[33] Mădălin Guţă and Anna Jenčová. Local asymptotic normality in quantum statistics. Comm. Math. Phys.,
276(2):341–379, 2007.
[34] H. Häffner, W. Hänsel, C. F. Roos, J. Benhelm, D. Chek-al kar, M. Chwalla, T. Körber, U. D. Rapol,
M. Riebe, P. O. Schmidt, C. Becher, O. Gühne, W. Dür, and R. Blatt. Scalable multiparticle entanglement
of trapped ions. Nature, 438:643, 2005.
[35] M. Hayashi. Asymptotic estimation theory for a finite-dimensional pure state model. Journal of Physics
A: Mathematical and General, 31(20), 1998.
[36] Masahito Hayashi, editor. Asymptotic theory of quantum statistical inference: selected papers. World
Scientific, 2005.
[37] Carl W. Helstrom. Quantum Detection and Estimation Theory. Academic Press, New York, 1976.
[38] F. Hiai and D. Petz. The proper formula for relative entropy and its asymptotics in quantum probability.
Commun. Math. Phys., 143:99–114, 1991.
[39] A. S. Holevo. Probabilistic and Statistical Aspects of Quantum Theory. North-Holland, 1982.
[40] Yuri Ingster and Natalia Stepanova. Estimation and detection of functions from anisotropic Sobolev classes.
Electron. J. Stat., 5:484–506, 2011.
[41] Yuri I. Ingster and Irina A. Suslina. Nonparametric goodness-of-fit testing under Gaussian models, volume
169 of Lecture Notes in Statistics. Springer-Verlag, New York, 2003.
[42] P. Ji and M. Nussbaum. Sharp minimax adaptation over Sobolev ellipsoids in nonparametric testing. To
appear, Electron. J. Stat.
[43] J. Kahn and M. Guţă. Local asymptotic normality for finite dimensional quantum systems. Commun.
Math. Phys., 289:597–652, 2009.
[44] Vladislav Kargin. On the Chernoff bound for efficiency of quantum hypothesis testing. Ann. Statist.,
33(2):959–976, 2005.
[45] Jussi Klemelä. Sharp adaptive estimation of quadratic functionals. Probab. Theory Related Fields,
134(4):539–564, 2006.
[46] Vladimir Koltchinskii. Von Neumann entropy penalization and low-rank matrix estimation. Ann. Statist.,
39(6):2936–2973, 2011.
[47] Vladimir Koltchinskii and Dong Xia. Optimal estimation of low rank density matrices. J. Mach. Learn.
Res., 16:1757–1792, 2015.
[48] L. Le Cam. Asymptotic Methods in Statistical Decision Theory. Springer Verlag, New York, 1986.
[49] Ulf Leonhardt. Measuring the Quantum State of Light. Cambridge University Press, 1997.
[50] K. Li. Discriminating quantum states: The multiple Chernoff distance. Ann. Statist., 44:1661–1679, 2016.
[51] Ranjith Nair, Saikat Guha, and Si-Hui Tan. Realizable receivers for discriminating coherent and multicopy
quantum states near the quantum limit. Phys. Rev. A, 89:032318, Mar 2014.
[52] Michael A. Nielsen and Isaac L. Chuang. Quantum Computation and Quantum Information. Cambridge
University Press, Cambridge, 2000.
[53] M. Nussbaum and A. Szkola. The Chernoff lower bound for symmetric quantum hypothesis testing. Ann.
Statist., 37:1040–1057, 2006.
[54] Michael Nussbaum. Asymptotic equivalence of density estimation and Gaussian white noise. Ann. Statist.,
24(6):2399–2430, 1996.
LAE OF PURE STATES ENSEMBLES AND QUANTUM GAUSSIAN WHITE NOISE
25
[55] T. Ogawa and H. Nagaoka. Strong converse and Stein’s lemma in quantum hypothesis testing. IEEE
Transactions on Information Theory, 46:2428–2433, 2000.
[56] M. G. A. Paris and J. Řeháček, editors. Quantum State Estimation, 2004.
[57] K. R. Parthasarathy. An Introduction to Quantum Stochastic Calculus. Modern Birkhäuser Classics.
Birkhäuser/Springer Basel AG, Basel, 1992.
[58] D. Petz and Jencova, A. Sufficiency in quantum statistical inference. Commun. Math. Phys., 263:259 –
276, 2006.
[59] Markus Reiß. Asymptotic equivalence for nonparametric regression with multivariate and random design.
Ann. Statist., 36(4):1957–1982, 2008.
[60] Angelika Rohde. On the asymptotic equivalence and rate of convergence of nonparametric regression and
Gaussian white noise. Statist. Decisions, 22(3):235–243, 2004.
[61] H. Strasser. Mathematical Theory of Statistics. de Gruyter, 1985.
[62] K. Temme and F. Verstraete. Quantum chi-squared and goodness of fit testing. J. Math. Phys., 56:012202,
2015.
[63] E. Torgersen. Comparison of Statistical Experiments. Cambridge University Press, 1991.
[64] Alexandre B. Tsybakov. Introduction to Nonparametric Estimation. Springer Series in Statistics. Springer,
New York, 2009.
[65] M. Walter and J. M. Renes. Lower bounds for quantum parameter estimation. IEEE Transactions on
Information Theory, 60:8007–8023, 2014.
6. Proofs
h
i
Proof of Theorem 5.2. Let us denote by RnE = inf ψbn supψ∈S α (L) Eψ kψbn − ψk22 the minimax risk.
α
The first step is to reduce the set of states S α (L) to a finite hypercube denoted S1:N
(L) consisting of certain
P
1/(2α+1)
non-zero coefficients with respect to the
“truncated” vectors |ψi = 1≤i≤N ψi |ei i which have N ≍ n
standard basis. This will provide a lower bound to the minimax risk. The coefficients are chosen as
σj
ψj = ± √ , σj2 = λ(1 − (j/N )2α ), j = 1, . . . , N, for some fixed λ > 0
n
and we check that they satisfy the ellipsoid constraint
X
j≥1
N
|ψj |2 j 2α
=
N 2α+1
2αλ
λ X 2α
(j − j 4α N −2α ) ≤
(1 + o(1)) ≤ L
n j=1
n (2α + 1)(4α + 1)
for an appropriate choice of λ > 0.
Using the factorisation property
(9) we can identify the corresponding Gaussian states with the N -mode state
√
defined by |φi = ⊗N
|G(
nψ
)i,
where the remaining modes are in the vacuum state and can be ignored.
j
j=1
Thus
RnE
≥
inf
=
inf
sup
b ψ∈S α (L)
ψ
1:N
sup
b ψ∈S α (L)
ψ
1:N
i
h
Eψ kψb − ψk22
N
X
|ψbj − ψj |2 .
Eψ
j=1
α
The supremum over the finite hypercube S1:N
(L) is bounded from below by the average over all its elements.
This turns the previous maximal risk into a Bayesian risk, that we can further bound from below as follows:
RnE
≥
inf
=
inf
≥
1
2N
b
ψ
b
ψ
X
α (L) j=1
ψ∈S1:N
N
X
1
N
2
j=1
N
X
inf
b
j=1 ψj
N
X
1
2N
X
α (L)
ψ∈S1:N
X
α (L)
ψ∈S1:N
i
h
Eψ |ψbj − ψj |2
i
h
Eψ |ψbj − ψj |2
i
h
Eψ |ψbj − ψj |2 .
(42)
26
LAE OF PURE STATES ENSEMBLES AND QUANTUM GAUSSIAN WHITE NOISE
√
In the second line ψb is the result of an arbitrary measurement and estimation procedure of the state |G( nψ)i.
In the third line each infimum is over procedures for estimating the component ψj only; since such procedure
may not be compatible with a single measurement, the third line is upper bounded by the second.
The second major step in the proof of the lower bounds is to reduce the risk over all measurements, to testing
two simple hypotheses. Let us bound from below the term (42) for arbitrary fixed j between 1 and N :
i
h
X
1
bj − ψj |2
|
ψ
E
ψ
2N
ψ∈S α (L)
1:N
h
h
X
X
√ 2i
√ 2 i
1
1 1
Eψ |ψbj − σj / n| + N −1
Eψ |ψbj − (−σj / n)|
=
2 2N −1
2
α
α
=
1n
2
h
ψ∈S(j+) (L)
√
Eρ+ |ψbj − σj / n|2
j
i
ψ∈S(j−) (L)
h
io
√
+ Eρ− |ψbj − (−σj / n)|2 ,
j
(43)
√
α
where the sum over ψ ∈ S(j±)
(L) means that the j th coordinate is fixed to ±σj / n and all k th coordinates,
√
√
for k 6= j, take values in {σk / n, −σk / n}. In the third line, we denote by ρ±
j the average state over states in
α
S(j±)
(L).
−
Let us define the testing problem of the two hypotheses H0 : ρ = ρ+
j against H1 : ρ = ρj . For a given estimator
ψbj we construct the test
σj
σj
∆ = I ψbj − √ > ψbj − (− √ ) ,
n
n
and decide H1 or H0 , if ∆ equals 1 or 0, respectively. By the Markov inequality, we get that
#
"
2
σj2
σj
σ
σj
j
.
≥
Eρ± ψbj − (± √ )
Pρ± ψbj − (± √ ) ≥ √
j
n
n j
n
n
On the one hand,
√
σj
b
≥ Pρ+ (∆ = 1).
Pρ+ |ψj − σj / n| ≥ √
j
j
n
(44)
σ
σ
Indeed, under Pρ+ , the event ∆ = 1 implies that |ψbj − √jn | > |ψbj + √jn |, which further implies by the triangular
j
inequality that
σj
2σj
2σj
σj
σj
ψbj − √ ≥ √ − ψbj + √ ≥ √ − ψbj − √ ,
n
n
n
n
n
σ
giving |ψbj − ψj | ≥ √j . By a similar reasoning for the P − distribution we get
n
ρj
√
σj
b
√
≥ Pρ− (∆ = 0).
Pρ− |ψj + σj / n| ≥
j
j
n
By using (44) and (45) in (43)
σj2
√ 2
√ 2
1
Eρ+ ψbj − σj / n
Pρ+ (∆ = 1) + Pρ− (∆ = 0) .
+ Eρ− ψbj − (−σj / n)
≥
j
j
j
j
2
2n
(45)
To summarise, we have lower bounded the MSE by the probability of error for testing between the states ρ±
j .
At closer inspection, these states are of the form |G(σj )ihG(σj )| ⊗ ρ and |G(−σj )ihG(−σj )| ⊗ ρ where ρ is a
fixed state obtained by averaging the coherent states of all the modes except j. Recall that the optimal testing
error in (11) gives a further bound from below
1
− ρ−
Pρ+ (∆ = 1) + Pρ− (∆ = 0) ≥ 1 − kρ+
j k1 .
j
j
2 j
Moreover, the state ρ can be dropped without changing the optimal testing error
We conclude that
−
2
kρ+
j − ρj k1 = k|G(σj )ihG(σj )| − |G(−σj )ihG(−σj )|k1 = 2(1 − exp(−2σj )).
inf
bj
ψ
1
2
σj2
√ 2
√ 2
· exp(−2σj2 )
Eρ+ ψbj − σj / n
+ Eρ− ψbj − (−σj / n)
≥
j
j
2n
LAE OF PURE STATES ENSEMBLES AND QUANTUM GAUSSIAN WHITE NOISE
27
and we further use this in (43) to get
RnE
N
N
X
σj2
j 2α
j 2α
N λ X
N
2
1−( )
exp −2 · λ(1 − ( ) ) ≥ c .
· exp(−2σj ) =
·
≥
2n
n
2N
N
N
n
j=1
j=1
Indeed, the average over j is the Riemann sum associated to the integral of a positive function and can be
bounded from below by some constant c > 0 depending on α. Moreover, N/n ≍ n−2α/(2α+1) and thus we finish
the proof of the theorem.
Proof of Theorem 5.3. Let R̃nE = inf |ψbn i sup|ψi∈S α (L) Eρ d(ρ̂n , ρ)2 be the minimax risk for Qn .
We bound from below the risk by restricting to (pure) states in a neigbourhood Σn (e0 ) of the basis vector |e0 i
defined as follows. As in (22) we write the state and the estimator in terms of their corresponding local vectors
p
p
|ψi = 1 − kuk2 |e0 i + |ui,
|ψ̂i = 1 − kûk2 |e0 i + |ûi,
|ui, |ûi ⊥ |e0 i.
Then the neighbourhood is given by Σn (e0 ) := {|ψu i : kuk ≤ γn }, where γn = (n log n)−1/4 . Such states are
described by the local model Qn (e0 , γn ), cf. equation (23). The risk is bounded from below by
R̃nE ≥ inf
sup
Eρ d(ρ̂n , ρ)2 .
bn i |ψi∈S α (L)∩Σn (e0 )
|ψ
By using the triangle inequality we can assume that ψ̂ ∈ Σn (e0 ), while incurring at most a factor 2 in the risk.
By using the quadratic approximation (28) we find that
d2 (b
ρn , ρ) = kku − ûk2 + o(n−1 )
(46)
where k = 1 or k = 4 depending on which distance we use. Since n−1 decreases faster than n−2α/(2α+1) , the
second term does not contribute to the asymptotic rate and can be neglected, so that the problem has been
reduced to that of estimating the local parameter u with respect to the Hilbert space distance. To study the
latter, we further restrict the set of states to a hypercube similar
to the one in the proof of Theorem 5.2,
P
α
(L). As before,
consisting of states |ψu i with “truncated” local vectors |ui = 1≤i≤N ui |ei i belonging to S1:N
1/(2α+1)
there are N ≍ n
non-zero coefficients of the form
σj
uj = ± √ , σj2 = λ(1 − (j/N )2α ), j = 1, . . . , N.
n
It has been already shown that such vectors belong to the ellipsoid S α (L). Additionally, we show that they also
belong to the local ball Σn (e0 ). Indeed
kuk2 =
N
X
j=1
2
N
N
1X
1X 2
2α
σj =
λ 1 − (j/N )
n j=1
n j=1
N
N 1 X
N
2α
=
λ 1 − (j/N )
≤ C1 ,
n N j=1
n
|uj | =
where we used that as N → ∞ the expression between the parentheses tens to a finite integral. As N scales as
n1/(2α+1) , the upper bound becomes
ke0 − ψu k2 ≤ C2 n−2α/(2α+1) = o(γn2 )
and the state |ψu i belongs to the local ball Σn (e0 ). Taking into account (46) the risk is therefore lower bounded
as
R̃nE ≥ inf sup Eρu ku − ûk2 + o(n−1 ).
u
b u∈S α (L)
1:N
where ρu = |ψu ihψu |, and the infimum is now taken over the local component |ûi of an estimator |ψ̂i =
p
1 − kûk2 |e0 i + |ûi. The first term is further lower bounded by passing to the Bayes risk for the uniform
α
distribution over S1:N
(L), similarly to the proof of Theorem 5.2
R̃nE ≥
N
X
j=1
By following the same steps we get
X
1
Eρu |b
u j − u j |2
2N
α
inf
u
bj
=
u∈S1:N (L)
≥
1
2N
X
α (L)
u∈S1:N
Eψu |b
uj − uj |2 + o(n−1 ).
h
h
io
√ i
√
1n
Eτ + |ψbj − σj / n|2 + Eτ − |ψbj − (−σj / n)|2 ,
j
j
2
σ2
σj2
1
j
Pτ + (∆ = 1) + Pτ − (∆ = 0) ≥
· (1 − kτj+ − τj− k1 ),
j
2n
2n
2
(47)
28
LAE OF PURE STATES ENSEMBLES AND QUANTUM GAUSSIAN WHITE NOISE
α
where we denote by τj± the average state over states |ψu ihψu |⊗n with u ∈ S(j±)
(L), and ∆ is a test for the
+
−
hypotheses H0 : τ = τj and H1 : τ = τj . In the last inequality we used the Helstrom bound [37] which
expresses the optimal average error probability for two states discrimination in terms of the norm-one distance
between states.
We now make use of the local asymptotic equivalence result in Theorem 4.1. From (25) we know that there
exist quantum channels Sn such that
√
√
|ψu ihψu |⊗n − Sn |G( nu)ihG( nu)| 1 ≤ ∆(Qn , Gn ) = o(1).
δn := max
α
u∈S1:N (L)
By Lemma 3.1 we get
−
kτj+ − τj− k1 ≤ kρ+
j − ρj k1 + 2δn
where ρ±
j are the corresponding mixtures in the Gaussian model as defined in the proof of Theorem 5.2. From
(47) we then get
1
2N
X
α (L)
u∈S1:N
σj2
σj2
1
−
u j − u j |2 ≥
Eρu |b
· (1 − kρ+
−
ρ
k
−
δ
)
≥
· (exp(−2σj2 ) − δn )
1
n
j
2n
2 j
2n
The rest of the proof follows as in the proof of Theorem 5.2, with the additional remark that
min exp(−2σj2 ) = λ(1 − N −α ) ≍ λ
j
and infinitely larger than δn , for n large enough.
2
Proof of Theorem 5.5. Denote by RnF = inf Fbn supψ∈S α (L) ηn−2 · Eψ Fbn − F (ψ) the minimax risk.
The case a) where α ≥ 2β reduces to the Cramér-Rao bound that proves that the parametric rate 1/n is always
a lower bound for the mean square error for estimating F (ψ).
We prove that in the case b) where β < α < 2β, this bound from below increases to n−2(1−β/α) (up to
constants). By the Markov inequality,
2 1
ηn
.
(48)
ηn−2 · Eψ Fbn − F (ψ) ≥ · Pψ |Fbn − F (ψ)| ≥
4
2
α
Let us restrict the set
n (e0 , γn ) (see equation
p of pure states S (L) to its intersection with the local model Q
2
(23)) where |ψu i = 1 − kuk · |e0 i + |ui is such that kuk ≤ γn , with γn = (n log n)−1/4 . In other words, u
belongs to the set
X
sα (L, γn ) = u ∈ ℓ2 (N∗ ) :
|uj |2 j 2α ≤ L and kuk ≤ γn .
j≥1
Using the fact that F (e0 ) = 0, we have
1
ηn
sup
· Pψ |Fbn − F (ψ)| ≥
2
ψ∈S α (L) 4
)
(
1
ηn
ηn
b
b
≥ max Pe0 |Fn | ≥
,
sup
Pψu |Fn − F (ψu )| ≥
4
2
2
u∈sα (L,γn ),F (ψu )≥ηn
(
)
1
η
η
n
n
≥
Pe0 |Fbn | ≥
+
sup
Pψu |Fbn − F (ψu )| ≥
8
2
2
u∈sα (L,γn ),F (ψu )≥ηn
(
)
ηn
ηn
1
Pe0 |Fbn | ≥
+
sup
Pψu |Fbn | <
≥
8
2
2
u∈sα (L,γn ),F (ψu )≥ηn
(49)
where in the last inequality we used that |Fbn | < ηn /2 and F (ψu ) ≥ ηn imply |Fbn − F (ψu )| ≥ ηn /2. Note also
that F (ψu ) = F (u) for |ui ∈ H0 ; we now consider the testing problem with hypotheses
H0 :
|ui = |0i
(50)
H1 (α, L, γn , ηn ) : |ui, with u ∈ sα (L, γn ) and F (u) ≥ ηn .
Let ∆ = ∆(ηn ) = I(|Fbn | ≥ ηn /2) be the test that accepts the null hypothesis when ∆ = 0 and rejects the
null hypothesis when ∆ = 1. Then the right-hand side of (49) is lower bounded by the sum of the error
LAE OF PURE STATES ENSEMBLES AND QUANTUM GAUSSIAN WHITE NOISE
29
probability of type I and of the maximal error probability of type II of ∆. We can describe ∆ as a binary
POVM M = (M0 , M1 ), depending on ηn : M (ηn ) = (M0 (ηn ), M1 (ηn )). Thus,
ηn
= Tr(|e0 ihe0 |⊗n · M1 )
(51)
Pe0 |Fbn | ≥
2
and
ηn
Pψu |Fbn | <
= Tr(|ψu ihψu |⊗n · M0 ).
2
By putting together (48)-(52), we get that the minimax risk has the lower bound
1
RnF ≥ inf
8 M
⊗n
he⊗n
0 |M1 |e0 i +
sup
u∈sα (L,γn ),F (u)≥ηn
(52)
!
hψu⊗n |M0 |ψu⊗n i .
Now, using the local asymptotic equivalence Theorem 4.1 with respect to the state |ψ0 i := |e0 i we map the
i.i.d. ensemble |ψu i⊗n to the Gaussian state |G(u)i ∈ F (H0 ). The lower bound becomes
!
√
√
1
F
Rn ≥ inf h0|Q1 |0i +
(53)
sup
hG( nu)|Q0 |G( nu)i + o(1)
8 Q
u∈sα (L,γn ),F (u)≥ηn
where the infimum is taken over tests Q = (Q0 , Q1 ) and the o(1) terms stems from the vanishing Le Cam
distance ∆(Qn (e0 , γn ), Gn (e0 , γn )). The lower bound has been transformed into a testing problem for the
Gaussian model.
In order to bound from below the maximal error probability of type II, we define a prior distribution on the
set of alternatives and average over the whole set with respect to this a priori distribution. Similarly to the
classical proofs of lower bounds, our construction will lead to a test of simple hypotheses: the former null and
the constructed averaged state. Assume that {uj }j≥1 are all independently distributed, such that uj has a
complex (bivariate) Gaussian distribution N2 (0, 12 σj2 · I2 ) for all j from 1 to N , and that uj = 0 for all j > N ,
where I2 is the 2 × 2 identity matrix. The σj2 are defined as
2α !
j
2
,
(54)
σj = λ 1 −
N
+
where λ, N > 0 are selected such that
X
X
j 2α σj2 = L(1 − ε) and
j 2β σj2 = n−1+β/α (1 + ε),
j≥1
(55)
j≥1
for an arbitrary ε > 0. Let us denote by Π the joint prior distribution of {uj }j≥1 .
Such a choice of the prior distribution was first introduced by Ermakov [20] for establishing sharp minimax risk
bounds for nonparametric testing in the Gaussian white noise model. This construction represents an analog of
the prior distribution used in Pinsker’s theory for sharp estimation of functions. In our case, using a Gaussian
prior as an alternative hypothesis leads to the well-known Gaussian thermal state.
The essence of this construction is that the random vectors u = {uj }j≥1 concentrate asymptotically, with
probability tending to 1, on the spherical segment
{u ∈ ℓ2 (N) : C n−1 ≤ kuk2 ≤ C n−1 (1 + 2ε′ )},
for ε′ > 0 depending on ε and some constant C > 0 depending on α and β described later on, and on the
alternative set of hypothesis, H1 (α, L, γn , ηn ). Note that the spherical segment is included in the set kuk ≤ γn ,
as γn = (n log n)−1/4 ≫ n−1/2 . This is proven by the following lemma.
Lemma 6.1. A unique solution (λ, N ) of (54), (55), exists for n large enough and admits an asymptotic
expansion with respect to n
((2β + 1)(2β + 2α + 1))(α+1/2)/(α−β)
(1 + ε)(α+1/2)/(α−β)
, Cλ =
(β+1/2)/(α−β)
(1 − ε)
2α(L(2α + 1)(4α + 1))(β+1/2)/(α−β)
1/(2(α−β))
1/(2(α−β))
1−ε
L(2α + 1)(4α + 1)
1/2α
N ∼n
CN
, CN =
.
1+ε
(2β + 1)(2β + 2α + 1)
λ ∼ n−1−1/2α Cλ
(56)
30
LAE OF PURE STATES ENSEMBLES AND QUANTUM GAUSSIAN WHITE NOISE
The independent complex Gaussian random variables uj ∼ N2 (0, 21 σj2 I2 ), with σj ’s and (λ, N ) given in (54),
(55), are such that, for an arbitrary ε > 0,
N
X
2
|uj | ≤ C n−1 (1 + 2ε′ ) → 1,
(57)
P C n−1 ≤
j=1
N
X
2
P
j 2α |uj | ≤ L → 1,
(58)
j=1
N
X
2
j 2β |uj | ≥ n−1+β/α → 1,
P
(59)
j=1
where C = Cλ · CN · 2α/(2α + 1) is a positive constant depending on α and β, and ε′ > 0 depends only on ε.
Proof of Lemma 6.1. The solution of the problem (54), (55) can be found in [20] (see also [42], Lemma 8 ) for
β = 0; a similar reasoning applies here. Let us prove that the random variables {uj }j=1,...,N satisfy (57) to
(59). We have
2α !
N
N
X
X
j
2α
2
1−
σj = λ
∼ λN
N
2α + 1
j=1
j=1
∼ Cλ CN
2α
n−1 (1 + ε)α/(α−β) (1 − ε)−β/(α−β) = C n−1 (1 + ε′ ),
2α + 1
where we denote ε′ = (1 + ε)α/(α−β) (1 − ε)−β/(α−β) − 1 which is positive for all ε ∈ (0, 1).
2
2
Note that E |uj | = σj2 and V ar |uj | = σj4 . We have
N
N
X
X
2
2
|uj | ≤ C n−1 (1 + 2ε′ ) = 1 − P
|uj | < C n−1 − P |uj |2 > C n−1 (1 + 2ε′ ) .
P C n−1 ≤
j=1
j=1
Now, by the Markov inequality,
N
N
X
X
2
2
|uj | < C n−1 = P (|uj | − σj2 ) < C n−1 − C n−1 (1 + ε′ + o(1))
P
j=1
j=1
N
X
2
≤ P (σj2 − |uj | ) > C n−1 (ε′ + o(1))
j=1
≤
≍
Moreover,
PN
j=1 V ar(|uj |
C 2 n−2 ε′2 /2
2
)
≤
PN
4
j=1 σj
C 2 n−2 ε′2
2
λ2 N
≍ n−1/2α = o(1).
C 2 n−2 ε′2
N
N
X
X
2
2
P
|uj | > C n−1 (1 + 2ε′ ) = P (|uj | − σj2 ) > C n−1 (ε′ + o(1)) = o(1),
j=1
j=1
which finishes the proof of (57).
Also, in view of (55), we have
N
N
X
X
2
2
j 2α |uj | > L = P
P
j 2α (|uj | − σj2 ) > L ε
j=1
j=1
≤
PN
j=1
2
≍
2
j 4α V ar |uj |
L 2 ε2
4α+1
λ N
L 2 ε2
=
PN
≍ n−1/2α = o(1),
4α 4
j=1 j σj
L 2 ε2
(60)
LAE OF PURE STATES ENSEMBLES AND QUANTUM GAUSSIAN WHITE NOISE
31
proving (58). Also,
N
N
X
X
2
2
P
j 2β |uj | < n−1+β/α ≤ P
j 2β (|uj | − σj2 ) < −n−1+β/α ε
j=1
j=1
≤
≍
PN
2
4β
j=1 j V ar(|uj | )
n−2+2β/α ε2
=
PN
4β 4
j=1 j σj
n−2+2β/α ε2
λ2 N 4β+1
≍ n−1/2α = o(1),
n−2+2β/α ε2
proving (59).
Let us go back to (53) and bound from below the maximal error probability of type II by the averaged risk,
with respect to our prior measure Π:
Z
√
√
√
√
Tr(|G( n u)ihG( n u)| · Q0 )Π(du)
sup
hG( n u)|Q0 |G( n u)i ≥
u∈sα (L),F (u)≥ηn
= Tr
≥ Tr
Z
Z
H1 (α,L,γn ,ηn )
√
√
|G( n u)ihG( n u)|Π(du) · Q0
√
√
|G( n u)ihG( n u)|Π(du) · Q0
−
Z
H1 (α,L,γn ,ηn )C
√
√
Tr(|G( n u)ihG( n u)| · Q0 )Π(du)
− Π(H1 (α, L, γn , ηn )C ).
√
√
In the last inequality we used that Tr(|G( n u)ihG( n u)|·Q0 ) ≤ 1. By Lemma 6.1, Π(H1 (α, L, γn , ηn )C ) = o(1)
and thus we deduce from (53) that
Z
√
√
1
F
+ o(1).
Rn ≥ inf Tr (|G(0)ihG(0)| · Q1 ) + Tr
|G( n u)ihG( n u)|Π(du) · Q0
8 Q
We recognize in the previous line the sum of error probabilities of type I and II for testing two simple quantum
hypotheses, i.e. the underlying state is either |G(0)i or the mixed state
Z
√
√
Φ := |G( n u)ihG( n u)|Π(du).
As a last step of the proof, we characterize more precisely the previous mixed Gaussian state as a thermal state
and use classical results from quantum testing of two simple hypotheses to √
give the bound from below of the
testing risk. Recall from Section 2.2.2, equation (9) that coherent√states |G( n u)i factorize as tensor product
of one-mode coherent states with displacements uj , i.e. ⊗j≥1 |G( nuj )i. A coherent state with displacement
z = x + iy with x, y ∈ R is fully characterized by its Wigner function given by equation (3). Since the prior is
Gaussian, our mixed state Φ is Gaussian and can be written
Z
N Z
O
O
√
√
√
√
|G( n u)ihG( n u)|Π(du) =
|G( n uj )ihG( n uj )|Πj (duj ) ⊗
|0ih0|
j=1
:=
N
O
j=1
Φj ⊗
O
j≥N +1
j≥N +1
|0ih0|
where Πj represents the bivariate centred Gaussian distribution with covariance matrix σj2 /2·I2 over the complex
plane uj = xj + iyj . Using equation (6), and setting σ 2 = nσj2 /2 there, we find that the individual modes with
index j ≤ N are centred Gaussian thermal states Φj = Φ(rj ) (cf. definition (4)) with rj = nσj2 /(nσj2 + 1).
In order to bound from below the right-hand side term in (53) we use the theory of quantum testing of two
simple hypotheses
H0 : ⊗j≥1 Φ(0) against H1 : ⊗N
j=1 Φ(rj ) ⊗j≥N +1 Φ(0).
Using (11), it is easy to see that this testing problem is equivalent to
H0 : (Φ(0))⊗N
against H1 : ⊗N
j=1 Φ(rj ).
As the vacuum and the thermal state are both diagonalized by the Fock basis, they commute, which reduces
⊗N
the problem to a classical test between the N -fold products of discrete distributions H0 : {G(0)}
and H1 :
∞
N
{⊗j=1 G(rj )}. In view of the form (4) of the thermal state, G(rj ) is the geometric distribution (1 − rj )rjk k=0
and G(0) is the degenerate distribution concentrated at 0. The optimal testing error is given by the maximum
32
LAE OF PURE STATES ENSEMBLES AND QUANTUM GAUSSIAN WHITE NOISE
likelihood test which decides H0 if and only if all observations are 0. The type I error is 0 and the type II error
is
N
N
N
Y
Y
X
1
(1 − rj ) =
σj2 ≥ exp(−c),
2 + 1 ≥ exp −n
nσ
j
j=1
j=1
j=1
for some c > 0, where in the last inequality we used (60). Using this in (53), we get as a lower bound
RnF ≥ exp(−c) + o(1) ≥ c0 ,
where c0 > 0 is some constant depending on c. This finishes the proof.
⊗n
Proof of Theorem 5.6. Let ϕn = cn n−1/2 for a positive sequence cn . Let Mn = (ρ⊗n
0 , I −ρ0 ) be the well-known
projection test for the problem (34). Then
⊗n
⊗n
RnMn = Tr(ρ⊗n · ρ⊗n
0 ) + Tr(ρ0 · (I − ρ0 ))
= (Tr(ρ · ρ0 ))n = |hψ|ψ0 i|2n .
Let us recall that for any pure states ρ = |ψihψ| and ρ0 = |ψ0 ihψ0 |, we have
p
kρ − ρ0 k1 = 2 1 − |hψ|ψ0 i|2 ,
thus |hψ|ψ0 i|2 = 1 − 41 kρ − ρ0 k21 and hence
RnMn =
(61)
n
1
1 − kρ − ρ0 k21 .
4
For any ρ satisfying the alternative hypothesis H1 (ϕn ), we have kρ − ρ0 k1 ≥ ϕn and consequently
n
n
c2n −1
1 2
n
=
1
−
ϕ
n
(ϕ
)
≤
1
−
PM
n
e
4 n
4
2
2
n
c
c
= exp − n .
≤ exp − n n−1
4
4
(62)
n
(ϕn ) → 0, so that the second relation in (35) is fulfilled.
If now ϕn /ϕ∗n → ∞ then cn → ∞ and PM
e
Consider now the case ϕn /ϕ∗n → 0 so that cn → 0. For any vector v ∈ H define
2
kvkα =
∞
X
j=0
2
|hej |vi| j 2α ;
(63)
2
then kvkα is a seminorm on the space of v fulfilling kvkα < ∞. The assumption that ρ0 = |ψ0 ihψ0 | ∈ S α (L′ )
2
means that kψ0 kα ≤ L′ < L. For some N > 0, consider the linear space
H0,N = {u ∈ H : hu|ψ0 i = 0, hu|ej i = 0, j > N } ;
it is nonempty if N ≥ 1. Let u ∈ H0,N , kuk = 1 be an unit vector; and for ε > 0 consider
p
ψu,ε = ψ0 1 − ε2 + εu.
(64)
Then kψu,ε k = 1, ρu,ε = |ψu,ε ihψu,ε | is a pure state, and
According to (61) we then have
|hψu,ε |ψ0 i|2 = 1 − ε2 .
q
kρu,ε − ρ0 k1 = 2 1 − |hψu,ε |ψ0 i|2 = 2ε
so for a choice ε = cn n−1/2 /2 it follows kρu,ε − ρ0 k1 = ϕn and ρu,ε ∈ B (ϕn ). On the other hand, by (64) and
the triangle inequality
p
kψu,ε kα ≤ 1 − ε2 kψ0 kα + ε kukα .
Now kukα < ∞ for u ∈ H0,N , and by assumption kψ0 kα < L1/2 , so for sufficiently large n
kψu,ε kα ≤ L1/2
and thus ρu,ε ∈ S α (L). Thus ρu,ε ∈ S α (L) ∩ B (ϕn ) for sufficiently large n. By (11) the optimal error
probability for testing between states ρu,ε and ρ0 fulfills
inf
M binary POVM
⊗n
RnT (ρ⊗n
0 , ρu,ε , M ) = 1 −
1 ⊗n
ρ − ρ⊗n
u,ε
2 0
1
LAE OF PURE STATES ENSEMBLES AND QUANTUM GAUSSIAN WHITE NOISE
q
q
⊗n 2
1 − |hψ0⊗n |ψu,ε
i| = 1 − 1 − |hψ0 |ψu,ε i|2n
q
q
n
n
= 1 − 1 − (1 − ε2 ) = 1 − 1 − (1 − c2n n−1 /4) .
n
Obviously if c2n → 0 then 1 − c2n n−1 /4 → 1 so that
33
=1−
inf
M binary POVM
(65)
⊗n
RnM (ρ⊗n
0 , ρu,ε , M ) ≥ 1 + o (1) .
But since ρu,ε ∈ S α (L) ∩ B (ϕn ) we have
P∗e (ϕn ) ≥
inf
M binary POVM
⊗n
RnM (ρ⊗n
0 , ρu,ε , M ) ≥ 1 + o (1) ,
so that the first relation in (35) is shown.
Proof of Theorem 5.7. It suffices to prove that if ϕn = cn n−1/2 with cn → c > 0 then P∗e (ϕn ) → exp −c2 /4 .
In view of the upper bound (36), if suffices to prove
P∗e (ϕn ) ≥ exp −c2 /4 (1 + o (1)) .
(66)
Recall (cf. (61)) that for any pure states ρ = |ψihψ| and ρ0 = |ψ0 ihψ0 |, the condition kρ − ρ0 k1 ≥ ϕn in H1 (ϕn )
is equivalent to a condition for the fidelity F 2 (ρ, ρ0 ) = |hψ|ψ0 i|2 ≤ 1 − ϕ2n /4.
Let H0 ⊂ H be the orthogonal complement of C|ψ0 i in H. Consider the vector
p
ψu = 1 − kuk2 · ψ0 + u, u ∈ H0
and the corresponding pure state |ψu ihψu | defined in terms of the local vector u. We restrict the alternative
hypothesis to a smaller set of states such that kuk ≤ γn , with γn = n−1/4 (log n)−1 . Since the fidelity is given
by F 2 (ρ0 , |ψu ihψu |) = |hψu |ψ0 i|2 = 1 − kuk2 , the restricted hypothesis is characterised by
1 − γn2 ≤ F 2 (ρ0 , |ψu ihψu |) ≤ 1 − ϕ2n /4,
or ϕ2n /4 ≤ kuk2 ≤ γn2 .
and additionally by kψu kα ≤ L where k·kα is given by (63).
Consider again the linear space H0,N defined in the proof of Theorem 5.7 for a choice N = Nn ∼ log n. Since
H0,N ⊂ H0 , we can further restrict the local vector u to u ∈ H0,N . Note that for u ∈ H0,N and kuk ≤ γn we
have
2
kukα =
N
X
j=0
2
|hej |ui| j 2α ≤ N 2α kuk2 ≤ N 2α γn2
∼ (log n)2α n−1/2 (log n)−2 = o (1) .
It follows that
q
kψu kα ≤ 1 − kuk2 kψ0 kα + kukα ≤ L
for sufficiently large n, thus ψu ∈ S α (L). We can now write the test problem with restricted alternative as
H0 :
ρ = ρ0
H1′ (ϕn ) : ρ = |ψu ihψu |: u ∈ H0,N , ϕn /2 ≤ kuk ≤ γn .
By the strong approximation proven in Theorem 4.1 we get that the models
√
√
{|ψu ihψu |⊗n , kuk ≤ γn } and {|G( nu)ihG( nu)|, kuk ≤ γn }
√
are asymptotically equivalent, where G( nu) is the coherent vector in the Fock space Γs (H0 ) pertaining to
√
nu. Note that this proof is very similar to the previous proofs of lower bounds, with a major difference: the
reduced set of states under the alternative hypothesis is defined with repect to ρ0 given by the null hypothesis
H0 instead of an arbitrary state previously.
In the asymptotically equivalent Gaussian white noise model, the modified hypotheses concern Gaussian states
which can be written in terms of their coherent vectors as
H0 :
|G(0)i
√
H1 (ϕn ) : |G( nu)i: u ∈ H0,N , ϕn /2 ≤ kuk ≤ γn .
(67)
34
LAE OF PURE STATES ENSEMBLES AND QUANTUM GAUSSIAN WHITE NOISE
In order to prove the theorem it is sufficient to prove that
inf
sup
Mn ϕn /2≤kuk≤γn , u∈H0,N
as n → ∞.
≥ exp −c2 /4 + o (1)
√
√
RnM (|G(0)ihG(0)|, |G( nu)ihG( nu)|, Mn )
(68)
(69)
PN
Note that dim H0,N = N ; let {gj , j = 1, . . . , N } be an orthogonal basis of H0,N and let |ui = j=1 uj |gj i.
√
The quantum Gaussian white noise model {|G( nu)i, u ∈ H0,N , kuk ≤ γn } is then equivalent to the quantum
√
Gaussian sequence model {⊗N
j=1 |G( nuj )i, kuk ≤ γn }. From now on |G(z)i denotes the coherent vector in the
Fock space F (C) pertaining to z := x + iy ∈ C. Recall that such a state is fully characterized by its Wigner
function WG(z) , which in the case of coherent states is the density fuction of a bivariate Gaussian distribution.
We shall bound from below the maximal type 2 error probability in the risk RnM in (68)
√
√
sup
Tr |G( nu)ihG( nu)| · Mn0
(70)
ϕn /2≤kuk≤γn , u∈H0,N
by an average over u, where the average is taken with respect to a prior distribution defined as follows. Assume
that uj , j = 1, . . . , N are independently distributed following a complex centered Gaussian law with variance
c2 1+ε
σ2
2
2 I2 , where σ = 4n N , for some fixed and arbitrary small ε > 0, and I2 is the 2 by 2 identity matrix.
Lemma 6.2. Let Π be the distribution of independent complex random variables uj , for j = 1, ..., N , each one
distributed as
σ2
c2 1 + ε
N 0, I2 , σ 2 =
,
2
4n N
for fixed ε > 0 and m = log n. Then as n → ∞
2
c2
cn
≤ kuk2 ≤ n (1 + ε)2 → 1,
P
4n
4n
and in particular P (ϕn /2 ≤ kuk ≤ γn ) → 1,
as n → ∞,
as n → ∞.
Proof. We have
N
2
2
2
X
c
c
1
+
ε
c
P kuk2 < n = P (|uj |2 − σ 2 ) < n − N
4n
4n
4n
N
j=1
≤
PN
Var( j=1 |uj |2 )
2
N σ4
=
(c2 ε + o(1))2 /16n2
2
1+ε
1
N c4 (1 + ε)2 /16n2N 2
=
= o(1),
=
2
ε + o(1)
N
(c2 ε + o(1)) /16n2
(c2n − c2 (1 + ε)) /16n2
since N ∼ log n → ∞. Similarly, as (1 + ε)2 > 1 + ε, one shows that
c2
P kuk2 > n (1 + ε)2 → 0,
4n
as n → ∞ and thus we get
P
c2
c2n
≤ kuk2 ≤ n (1 + ε)2
4n
4n
→ 1.
As γn2 = n−1/2 (log n)−2 decays slower than c2n /n, and ϕn /2 = cn n−1/2 /2, we deduce that P (ϕn /2 ≤ kuk ≤ γn ) →
1 as n → ∞ which ends the proof of the lemma.
Let us denote by Π the prior distribution introduced in Lemma 6.2. Let us go back to (70) and bound the
expression from below as follows:
√
√
sup
Tr |G( nu)ihG( nu)| · Mn0
ϕn /2≤kuk≤γn , u∈H0,N
≥
Z
ϕn /2≤kuk≤γn
√
√
Tr(|G( nu)ihG( nu)|Mn0 )Π(du)
LAE OF PURE STATES ENSEMBLES AND QUANTUM GAUSSIAN WHITE NOISE
≥
≥
Z
Z
√
√
Tr(|G( nu)ihG( nu)|Mn0 )Π(du) −
Z
{ϕn /2≤kuk≤γn }c
35
√
√
Tr(|G( nu)ihG( nu)|Mn0 )Π(du)
√
√
c
Tr(|G( nu)ihG( nu)|Mn0 )Π(du) − Π ({ϕn /2 ≤ kuk ≤ γn } ) .
By Lemma 6.2, we get for (68)
sup
ϕn /2≤kuk≤γn , u∈H0,N
≥ Tr(|G(0)ihG(0)|Pn1 ) + Tr
Z
√
RnM (G(0), G( nu), Mn )
√
√
|G( nu)ihG( nu)|Π(du) · Mn0 + o(1).
(71)
The integral on the right side is a mixed state which can be written as
Z
N Z
O
√
√
√
√
Φ := |G( nu)ihG( nu)|Π(du) =
|G( nuj )ihG( nuj )| · Πj (duj ).
j=1
Similarly to the proof of Theorem 5.5 we use equation (6) to show that each of the Gaussian integrals above
produces a thermal (Gaussian) state
∞
X
nσ 2
Φ(r) = (1 − r)
rk |kihk|,
r=
.
nσ 2 + 1
k=0
Since |G(0)ihG(0)| = Φ(0), the main terms in (71) are the sum of error probabilities for testing two simple
hypothesis H0 : Φ(0)⊗N against H1 : Φ(r)⊗N . Moreover, we have two commuting product states under the
two simple hypotheses, which reduces the problem to a classical test between the N -fold products
of discrete
∞
distributions H0 : {G(0)}⊗N and H1 : {G(r)}⊗N . Here G(r) is the geometric distribution (1 − r)rk k=0 ; in
particular s G(0) is the degenerate distribution concentrated at 0. The optimal testing error is given by the
maximum likelihood test which decides H0 if and only if all observations are 0. The type 1 error is 0 and the
type 2 error is
(1 − r)N = (nσ 2 + 1)−N ≥ exp(−N · nσ 2 )
2
c (1 + ε)
c2 1 + ε
= exp −
.
= exp −N n
4n N
4
Since ε > 0 was arbitrary, this establishes the lower bound (69) and thus (66).
| 10math.ST
|
PARAMETRIZING FILTERS OF A CNN
WITH A
GAN
Yannic Kilcher∗, Gary Bécigneul∗, Thomas Hofmann
Department of Computer Science
ETH Zurich
{yannic.kilcher,gary.becigneul,thomas.hofmann}@inf.ethz.ch
arXiv:1710.11386v1 [cs.LG] 31 Oct 2017
A BSTRACT
It is commonly agreed that the use of relevant invariances as a good statistical
bias is important in machine-learning. However, most approaches that explicitly
incorporate invariances into a model architecture only make use of very simple
transformations, such as translations and rotations. Hence, there is a need for
methods to model and extract richer transformations that capture much higherlevel invariances. To that end, we introduce a tool allowing to parametrize the
set of filters of a trained convolutional neural network with the latent space of a
generative adversarial network. We then show that the method can capture highly
non-linear invariances of the data by visualizing their effect in the data space.
1
I NTRODUCTION
In machine-learning, solving a classification task typically consists of finding a function f : X → Y,
from a rather large data space X to a much smaller space of labels Y. Such a function will therefore necessarily be invariant to a lot of transformations of its input data. It is now clear that being
able to characterize such transformations can greatly help the learning procedure, one of the most
striking examples being perhaps the use of convolutional neural networks (CNN) for image classification (Krizhevsky et al., 2012), with built-in translation invariance via convolutions and subsequent
pooling operations. But as a convolutional layer is essentially a fully connected layer with a constraint tying some of its weights together (LeCun et al., 1995), one could expect other invariances
to be encoded in its weights after training. Indeed, from an empirical perspective, CNNs have been
observed to naturally learn more invariant features with depth (Goodfellow et al., 2009; Lenc &
Vedaldi, 2015), and from a theoretical perspective, it has been proven that under some conditions
satisfied by the weights of a convolutional layer, this layer could be re-indexed as performing a
convolution over a bigger group of transformations than only translations (Mallat, 2016).
It is exciting to note that there has recently been a lot of interest in theoretically extending such successful invariant computational structures to general groups of transformations, notably with group
invariant scattering operators (Mallat, 2012), deep symmetry networks (Gens & Domingos, 2014),
group invariant signal signatures (Anselmi et al., 2015), group invariant kernels (Mroueh et al.,
2015) and group equivariant convolutional networks (Cohen & Welling, 2016). However, practical
applications of these have mostly remained limited to linear and affine transformations. Indeed, it is
a challenge in itself to parametrize more complicated, non-linear transformations preserving labels,
especially as they need to depend on the dataset. In this work, we seek to answer this fundamental
question:
What invariances in the data has a CNN learned during its training on a classification task and
how can we extract and parameterize them?
The following is a brief summary of our method: Considering an already trained CNN on a labeled
dataset, we train a generative adversarial network (GAN) (Goodfellow et al., 2014) to produce filters
of a given layer of this CNN, such that the filters’ convolution output be indistinguishable from the
one obtained with the real CNN. We combine this with an InfoGAN (Chen et al., 2016) discriminator to prevent the generator from producing always the same filters. As a result, the generator
provides us with a smooth, data-dependent, non-trivial parametrization of the set of filters of this
CNN, characterizing complicated transformations irrelevant for this classification task. Finally, we
∗
contributed equally
1
describe how to visualize what these smooth transformations of the filters would correspond to in
the image space.
2
2.1
BACKGROUND
G ENERATIVE A DVERSATIAL N ETWORKS
A Generative Adversarial Network (Goodfellow et al., 2014) consists of two players, the generator
G and the discriminator D, playing a minimax game in which G tries to produce samples indistinguishable from some given true distribution pdata , and D tries to distinguish between real and
generated samples. G typically maps a random noise z to a generated sample G(z), transforming
the noise distribution into a distribution pg supposed to match pdata . The objective function of this
minimax game with maximum likelihood is given by
min max V (D, G) = Ex∼pdata [log D(x)] + Ez [log(1 − D(G(z)))].
G
D
The noise space, input of the generator, is also called its latent space.
2.2
I NFORMATION M AXIMIZING G ENERATIVE A DVERSARIAL N ETS
In an InfoGAN (Chen et al., 2016), the generator takes as input not only the noise z but also another
variable c, called the latent code. The aim is to make the generated samples G(z, c) depend on
cQ:= (c1 , ..., cn ) in a structured way, for instance by choosing independent ci ’s, modelling p(c) as
i p(ci ). In order to avoid a trivial correspondence between c and G(z, c), the InfoGAN procedure
maximizes the mutual information I(c, G(z, c)).
The mutual information I(X, Y ) between two random variables, defined with an entropy H as
I(X, Y ) := H(X) − H(X|Y ) = H(Y ) − H(Y |X) is symmetric, measures the amount of information that is known about the value of one random variable when the value of the other one is
known, and is equal to zero when they are independent. Hence, maximizing the mutual information
I(c, G(z, c)) prevents G(z, c) from being independent of c.
In practice, I(c, G(z, c)) is indirectly maximized using a variational lower bound
LI (G, Q) ≤ I(c, G(z, c)),
where Q(c|x) approximates P (c|x) and
LI (G, Q) := Ec∼p(c),x∼G(z,c) [log Q(c|x)] + H(c).
The minimax game becomes
min max VI (D, G, Q) = V (D, G) − λLI (G, Q),
G,Q
D
where λ is a hyperparameter controlling the mutual information regularization.
3
E XTRACTING I NVARIANCES BY L EARNING F ILTERS
Let CNN be an already trained CNN, whose `th -layer representation of an image I will be denoted by
CNN` (I), with CNN0 (I) = I. As our goal is to learn what kind of filters such a CNN would learn at
layer `, it could be tempting to simply train a GAN to match the distribution of filters of this CNN’s
layer. However, this set of filters is way too small of a dataset to train a GAN (G, D), which would
cause the generator G to massively overfit instead of extracting the smooth, hidden structure lying
behind our discrete set of filters. To cope with this problem, instead of feeding the discriminator
D alternatively with filters produced by G and real filters from CNN` , we propose to feed D with
the activations of these filters caused by the data passing through the CNN, i.e. alternatively with
CNN` (I) or Conv(CNN`−1 (I), G(z)), corresponding respectively to real and fake samples. Here, I is
an image sampled from data, z is sampled from the latent space of G and Conv(CNN`−1 (I), G(z))
is the activation obtained by passing I through each layer of CNN but while replacing the filters of
the `th -layer by G(z).
2
In short, in each step, the generator G produces a set of filters for the `th -layer of the CNN. Next,
different samples of data are passed through one CNN using its real filters and through the same
CNN, but having its `th -layer filters replaced by the fake filters produced by G. The discriminator
D will then try to guess if the activation it is fed was produced using real or generated filters at
the `th -layer, while the generator G will try to produce filters making the subsequent activations
indistinguishable from some obtained with real filters.
However, even though this formulation allows us to train the GAN on a dataset of reasonable size,
saving us from an otherwise unavoidable overfitting, G could a priori still always produce the same
set of filters to fool D. Ideally it simply reproduces the real filters of CNN` . To overcome this
problem, we augment our model by an InfoGAN discriminator whose goal will be to predict which
noise z was used to produce Conv(CNN`−1 (I), G(z)). This prevents G from always producing the
same filters, by preventing z and G(z) from being indenpendent random variables. Note that, just
as the GAN discriminator, the InfoGAN discriminator does not act directly on the output of G - the
filters - but on the activation output that these filters produce.
In this setting, the noise z of the generator G plays the role of the latent code c of the InfoGAN. As
in the original InfoGAN paper (Chen et al., 2016), we train a neural network Q to predict the latent
code z, which shares all its layers but the last one with the discriminator D. Finally, by modelling
the latent codes as independent gaussian random variables, the term LI (G, Q) in the variational
bound being a log-likelihood, it is actually given by an L2 -reconstruction error. The joint training
of these three neural networks is described in Algorithm 1 and illustrated in Figure 1.
3
Real Layer L
Predicts
real / fake
signal
Layer L-1
Layer 2
Layer 1
Data
D
Discriminator
Fake Layer L
Fake
Filters
Predicts z
from signal
InfoGAN Loss
Generator
z
Q
G
Figure 1: Illustration of how the different neural networks interact with each other. CNN layers are
depicted in light gray. The flow of data is shown in green, while the generation of the filters by the
generative model is shown in red. The discriminator part of the GAN is shown in blue. Note that
the discriminator does not have direct access to the generated filters, but can only observe the data
after it has passed through them. The CNN is fixed, while G, D and Q are trained jointly.
4
Algorithm 1 Minibatch stochastic gradient descent training of D, G and Q.
1: for number of training iterations do
2:
• Sample minibatch of m noise samples {z (1) , ..., z (m) } from noise prior pnoise (z).
3:
• Generate the filters Gθg (z (1) ), ..., Gθg (z (m) ).
4:
• Sample minibatch of m examples {I (1) , ..., I (m) } from data distribution pdata (I).
5:
• Pass the data through the CNN with the real and generated filters, to obtain the CNN` (I (i) )’s
and Conv(CNN`−1 (I (i) ), Gθg (z (i) ))’s respectively.
6:
• Feed these to Dθd and Qθq , letting Dθd guess if it was fed CNN` (I (i) ) or
Conv(CNN`−1 (I (i) ), Gθg (z (i) )), and letting Qθq recover the z (i) .
7:
• Update the discriminator by ascending its stochastic gradient:
m
∇θ d
8:
1 X
[log Dθd (CNN` (I (i) )) + log(1 − Dθd (Conv(CNN`−1 (I (i) ), Gθg (z (i) ))))].
m i=1
• Update the generator by descending its stochastic gradient:
m
∇θ g
1 X
[log(1 − Dθd (Conv(CNN`−1 (I (i) ), Gθg (z (i) ))))
m i=1
+ λkz (i) − Qθq (Conv(CNN`−1 (I (i) ), Gθg (z (i) )))k22 ].
9:
• Update the InfoGAN discriminator by descending its stochastic gradient:
!
m
λ X (i)
(i)
(i)
2
∇θ q
kz − Qθq (Conv(CNN`−1 (I ), Gθg (z )))k2 .
m i=1
end for
10: The gradient-based updates can use any standard gradient-based learning rule. We used RM-
Sprop in our experiments.
5
4
V ISUALIZING THE LEARNED TRANSFORMATIONS
Using our method, we can parameterize the filters of a trained CNN and thus characterize its learned
invariances. But in order to assess what has actually been learned, we need a way to visualize these
invariances once the GAN has been trained. More specifically, given a data sample x, we would like
to know what transformations of x the CNN regards as being invariant. We do this in the following
manner:
We take some latent noise vector z and obtain its generated filters G(z). Using those filters, we pass
the data sample x through the network to obtain Conv(CNN`−1 (x), Gθg (z)) =: a(x|z) , which we
call the activation profile of x given z.
Next we select two dimensions i and j of z at random and construct a grid of noise vectors {zk |zki ∈
[−ξ, ξ], zkj ∈ [−ξ, ξ]}k by moving around z in the dimensions i and j in a small neighborhood.
For each zk , we use Gradient Descent to start from x and find the data point x0k that gives the same
activation profile for the filters generated using zk as the data point x gave for the filters generated
using z. Formally, for each zk we want to find x0k , s.t.
x0k = argmin ka(x0 |zk ) − a(x|z) k22 + Ψ(x0 ),
x0
0
where Ψ(x ) is a regularizer corresponding to a natural image prior. Specifically, we use the loss
function proposed in (Mahendran & Vedaldi, 2015).
By using Gradient Descent and starting from the original data point x, we make sure that the solution
we find is likely in the neighborhood of x, i.e. can be obtained by applying a small transformation
to x.
As a result, from our grid of z-vectors, we obtain a grid of x-points. This grid in data space represents
a manifold traversal in a small neighborhood on the manifold of learned invariances. If our method
is successful, we expect to see sensible continuous transformations of the original data point along
the axes of this grid.
5
E XPERIMENTAL R ESULTS
We apply our method of extracting invariances on a convolutional neural network trained on the
MNIST dataset. In particular, we train a standard architecture featuring 5 convolutional layers with
ReLU nonlinearities, max-pooling and batch normalization for 10 epochs on the 10-class classification task.
Once converged, we use our GAN approach to learn the filters of the 4th convolutional layer in the
CNN. Since this is one of the last layers in the network, we expect the invariances that we extract to
be very high-level and highly nonlinear.
5.1
V ISUALIZING THE LEARNED INVARIANCES
The results can be seen in Figure 3 and a sample of the learned filters themselves can be seen in
Figure 2. Our expectations are clearly met as the resulting outputs are in fact an ensemble of highly
nonlinear and high-level transformations of the data. Even more visualizations can be found in the
Appendix.
We further hypothesize that if we apply the same method to the filters of one of the first layers in the
network, the transformations that we learn will be much more low-level and more pixel-local. To
test this, we use our method on the same CNN’s second convolutional layer. The results can be seen
in Figure 4. As expected, the transformations are much more low-level, such as simple brightness
changes or changes to the stroke width.
6
Figure 2: Learned filters of the CNN’s 4th layer. We summed one third of the orignal channels
together in order to visualize the learned filters.
Figure 3: Invariance transformations extracted from the CNN’s 4th layer. The middle sample of
each grid represents the original data sample, while the rest of the grid are found by matching the
original sample’s activation profile.
7
Figure 4: Invariance transformations extracted from the CNN’s 2nd layer. The middle sample of
each grid represents the original data sample, while the rest of the grid are found by matching the
original sample’s activation profile.
8
5.2
A SSESSING THE QUALITY OF THE GENERATOR
In order to assess the quality of the generator, we need to be sure that: (i) filters produced by the
generator would yield a good accuracy on the original classification task of our CNN, and (ii) the
generator can produce a variety of different filters for different latent noises.
For the first part, we randomly drew 10 noise vectors z (1) , ..., z (10) , computed the corresponding
set of filters G(z (i) ) for each of them, and then each data sample x, after going through CNNl−1 , is
passed through each of these 10 `th -layer and averaged over them, so that the signal fed to the next
layer becomes:
10
1 X
Conv(CNNl−1 (x), G(z (i) )),
10 i=1
all the next layers being re-trained. This averaging can be seen as an average pooling w.r.t. the
transformations defined by the generator, which, if the transformations we learned were indeed
irrelevant for the classification task, should not induce any loss in accuracy. Our expectations are
confirmed, as the test accuracy obtained by following the above procedure is of 0.982, against a test
accuracy of 0.971 for the real CNN.
As for the second part, Figure 5 shows a Multi-Dimensional Scaling (MDS) of both the original
set of filters of CNN` , and of generated filters for randomly sampled noise vectors. We observe that
different noise vectors produce a variety of different filters, which confirms that the generator has
not overfitted on the set of real filters. Further, since the generator has learned to produce a variety of
filters for each real filter, all the while retaining its classification accuracy, this means that we have
truly captured the invariances of the data with regard to the CNN’s classification task.
Figure 5: Multi-Dimensional Scaling for the filters produced by the GAN. Individual colors represent different samples for the same filter of the true CNN. The large cluster sizes shows that the
GAN is producing a wide variety of different filters for each corresponding real filter.
6
C ONCLUSION AND FUTURE WORK
Introducing an invariance to irrelevant transformations for a given task is known to constitute a
good statistical bias, such as translations in CNNs. Although a lot of work has already been done
regarding how to implement known invariances into a computational structure, practical applications
of these mostly include very simple linear or affine transformations. Indeed, characterizing more
complicated transformations seems to be a challenge in itself.
In this work, we provided a tool allowing to extract transformations w.r.t. which a CNN has been
trained to be invariant to, in such a way that these transformations can be both visualized in the
image space, and potentially re-used in other computational structures, since they are parametrized
by a generator. The generator has been shown to extract a smooth hidden structure lying behind the
discrete set of possible filters. It is the first time that a method is proposed to extract the symmetries
learned by a CNN in an explicit, parametrized manner.
9
Applications of this work are likely to include transfer-learning and data augmentation. Future work
could apply this method to colored images. As suggested by the last subsection, the parametrization
of such irrelevant transformations of the set of filters could also potentially define another type of
powerful pooling operation.
R EFERENCES
Fabio Anselmi, Joel Z Leibo, Lorenzo Rosasco, Jim Mutch, Andrea Tacchetti, and Tomaso Poggio.
Unsupervised learning of invariant representations in hierarchical architectures. Theoret. Comput.
Sci., dx.doi.org/10.1016/j.tcs.2015.06.048, 2015.
Xi Chen, Yan Duan, Rein Houthooft, John Schulman, Ilya Sutskever, and Pieter Abbeel. Infogan:
Interpretable representation learning by information maximizing generative adversarial nets. In
Advances in Neural Information Processing Systems, pp. 2172–2180, 2016.
Taco Cohen and Max Welling. Group equivariant convolutional networks. In International Conference on Machine Learning, pp. 2990–2999, 2016.
Robert Gens and Pedro M Domingos. Deep symmetry networks. In Advances in neural information
processing systems, pp. 2537–2545, 2014.
Ian Goodfellow, Honglak Lee, Quoc V Le, Andrew Saxe, and Andrew Y Ng. Measuring invariances
in deep networks. In Advances in neural information processing systems, pp. 646–654, 2009.
Ian Goodfellow, Jean Pouget-Abadie, Mehdi Mirza, Bing Xu, David Warde-Farley, Sherjil Ozair,
Aaron Courville, and Yoshua Bengio. Generative adversarial nets. In Advances in neural information processing systems, pp. 2672–2680, 2014.
Alex Krizhevsky, Ilya Sutskever, and Geoffrey E Hinton. Imagenet classification with deep convolutional neural networks. In Advances in neural information processing systems, pp. 1097–1105,
2012.
Yann LeCun, Yoshua Bengio, et al. Convolutional networks for images, speech, and time series.
The handbook of brain theory and neural networks, 3361(10):1995, 1995.
Karel Lenc and Andrea Vedaldi. Understanding image representations by measuring their equivariance and equivalence. In Proceedings of the IEEE conference on computer vision and pattern
recognition, pp. 991–999, 2015.
Aravindh Mahendran and Andrea Vedaldi. Understanding deep image representations by inverting
them. In Proceedings of the IEEE conference on computer vision and pattern recognition, pp.
5188–5196, 2015.
Stéphane Mallat. Group invariant scattering. Communications on Pure and Applied Mathematics,
65(10):1331–1398, 2012.
Stéphane Mallat. Understanding deep convolutional networks. Phil. Trans. R. Soc. A, 374(2065):
20150203, 2016.
Youssef Mroueh, Stephen Voinea, and Tomaso A Poggio. Learning with group invariant features:
A kernel perspective. In Advances in Neural Information Processing Systems, pp. 1558–1566,
2015.
10
A
M ORE I NVARIANCE V ISUALIZATIONS
Figure 6: Invariance transformations extracted from the CNN’s 4th layer. The middle sample of
each grid represents the original data sample, while the rest of the grid are found by matching the
original sample’s activation profile.
11
Figure 7: Invariance transformations extracted from the CNN’s 4th layer. The middle sample of
each grid represents the original data sample, while the rest of the grid are found by matching the
original sample’s activation profile.
12
| 2cs.AI
|
Optimizing Active Cyber Defense
Wenlian Lu1,2 , Shouhuai Xu3 , and Xinlei Yi1
arXiv:1603.08312v1 [cs.CR] 28 Mar 2016
1
3
School of Mathematical Sciences, Fudan University
Shanghai, P. R. China, 200433
Emails: {wenlian,11210180008}@fudan.edu.cn
2
Department of Computer Science, University of Warwick
Coventry CV4 7AL, UK
Department of Computer Science, University of Texas at San Antonio
San Antonio, Texas 78249, USA
Email: [email protected]
Abstract. Active cyber defense is one important defensive method for
combating cyber attacks. Unlike traditional defensive methods such as
firewall-based filtering and anti-malware tools, active cyber defense is
based on spreading “white” or “benign” worms to combat against the
attackers’ malwares (i.e., malicious worms) that also spread over the
network. In this paper, we initiate the study of optimal active cyber
defense in the setting of strategic attackers and/or strategic defenders.
Specifically, we investigate infinite-time horizon optimal control and fast
optimal control for strategic defenders (who want to minimize their cost)
against non-strategic attackers (who do not consider the issue of cost).
We also investigate the Nash equilibria for strategic defenders and attackers. We discuss the cyber security meanings/implications of the theoretic
results. Our study brings interesting open problems for future research.
Keywords: cyber security model, active cyber defense, optimization,
epidemic model
1
Introduction
The importance of cyber security is well recognized now. However, our understanding of cyber security is still at its infant stage. In general, the attackers are
constantly escalating their attack power and sophistication, while the defenders
largely lag behind. To be specific, we mention the following asymmetry between
cyber attack and cyber defense: The effect of malware-like attacks is automatically amplified by the network connectivity, while the defense effect is not. This
phenomenon had been implied by many previous results (e.g., [28, 9, 6, 26, 34]),
but was not explicitly pointed out until very recently [35]. The asymmetry is fundamentally caused by that the defense is reactive, including intrusion detection
systems, firewalls and anti-malware tools. The asymmetry can be eliminated by
the idea of active cyber defense [35], where the defender also aims to take advantage of the network connectivity. The concept of active cyber defense is not
completely new because researchers have proposed for years the idea of using the
defender’s “white” or “benign” worms to combat against the attackers’ malwares
[5, 1, 29, 23, 16, 18, 13, 30]. In a sense, active cyber defense already happened in
practice; for example, the Welchia worm attempted to “kill” the Blaster worm in
compromised computers [23, 20]. It appears that full-fledged active cyber defense
is perhaps inevitable in the near future according to some recent reports [18, 24,
31]. It is therefore more imperative than ever to systematically characterize the
effectiveness of active cyber defense. This motivates the present study.
1.1
Our Contributions
This paper is inspired by the recent mathematical model of active cyber defense dynamics [35], which characterizes the effect of various model parameters
(including the underlying complex network structures) in the setting where neither the attacker nor the defender is strategic (i.e., both the attacker and the
defender do not consider the issue of cost). Here we study a new perspective of
active cyber defense, namely the strategic interaction between the attacker and
the defender. On one hand, our study moves a step beyond [35] because we incorporate control-theoretic and game-theoretic models to accommodate strategic
interactions. On the other hand, our study assumes away the underlying complex
network structures that are explicitly investigated in [35]. This means that our
study is essentially based on the homogeneous (or well-mixed) assumption that
each compromised computer can attack the same portion of computers. Tackling
the problem of strategic attack-defense interactions with explicit complex network structures is left for future research. Therefore, we deem the present paper
as a significant first step toward ultimately understanding the effectiveness of
strategic active cyber defense. Specifically, we make the following contributions.
First, we investigate two flavors of optimal control for strategic defenders
against non-strategic attackers: infinite-time horizon optimal control and fast
optimal control. In the setting of infinite-time horizon optimal control for the
defender, we characterize the conditions under which the defender should adjust
its active cyber defense power in a certain quantitative fashion. For example, we
identify a condition under which the defender should give up using active cyber
defense alone, and instead should resort to other defense methods as well (e.g.,
proactive defense). In the setting of fast optimal control, where the defender
wants to occupy a certain portion of the network as soon as possible and at the
minimal cost, there is a significant difference between the case that the active
defense cost is linear and the case that the active defense cost is quadratic.
Second, we identify the Nash equilibrium strategies when both the defender
and the attacker are strategic. The findings are interesting. For example, when
the defender (or attacker) is reluctant to use/expose its advanced active cyber
defense tools (or zero-day exploits), it will give up escalating its active defense
(or attack) power; otherwise, there are three scenarios: (i) If the defender (or
attacker) initially occupies only a certain small portion of the network, it will
give up escalating its active defense (or attack). (ii) If the defender (or attacker)
initially occupies a certain significant portion of the network, it will escalate its
active defense (or attack) as much as possible. (iii) If the defender (or attacker)
initially occupies a certain large portion of the network, it will not escalate its
active defense (or attack) — a sort of diminishing returns.
The rest of the paper is structured as follows. Section 2 briefly reviews the
related prior work. Section 3 describes the basic active cyber defense model under
the homogeneous assumption. Section 4 investigates optimal control for strategic
defenders against non-strategic attackers. Section 5 studies Nash equilibria for
strategic defenders and attackers. Section 6 concludes the paper with some open
problems. Lengthy proofs are deferred to the Appendix. The main notations
used in the paper are listed below:
αB , αR defender B’s defense power αB and attacker R’s attack power αR
iB (t), iR (t) portions of the nodes occupied respectively by the defender and the
attacker at time t, where iB (t) + iR (t) = 1
πB , πB (t) πB is control variable and πB (t) is control function
π̂B solution in the infinite-time horizon optimal control case
∗
∗∗
πB
, πB
solutions in the case of fast optimal control with linear and
quadratic cost functions, respectively
z discount rate
kB normalization ratio between the defender’s detection cost and recovery cost
λ normalization ratio between the unit of time and the defender’s
active defense cost
kR normalization ratio between the attacker’s maintenance cost and
penetration cost
2
Related Work
Our investigation is built on recent studies in mathematical computer malware
models. These models originated in the mathematical biological epidemic models
introduced in the 1920’s [19, 12], which were first adapted to study the spreading of computer virus in the 1990’s [10, 11]. All these models made the homogeneous assumption that each individual (e.g., computer) in the population has
equally infection effect on the other individuals in the population, and the assumption that the infected individuals recover because of reactive defense (e.g.,
anti-malware tools). In the past decade, there were many studies that aim to
eliminate the aforementioned homogeneous assumption, by explicitly incorporating the heterogeneous network structures [28, 9, 6, 26, 34, 32]. The mathematical
tools used for these studies are Dynamical Systems in nature. These studies
demonstrated that the attack effect of malware spreading against reactive defense is automatically amplified by the largest eigenvalue of the adjacency matrix, which represents the underlying complex network structure. This is the
attack-defense asymmetry phenomenon mentioned above.
The attack-defense asymmetry phenomenon motivated the study of mathematical models of active cyber defense [35], which is a relatively new sub-field
in cyber security [18, 24, 31] as previous explorations were mainly geared toward
legal and policy issues [5, 1, 29, 23, 16, 18, 13, 30]. One real-life incident of the flavor of active cyber defense is that the Welchia worm attempted to “kick out”
another kind of worms (e.g., the Blaster worm) [23, 20]. In the first mathematical characterization of active cyber defense [35], neither the attacker nor the
defender is strategic (i.e., they do not consider the issue of cost), albeit the
model accommodates the underlying complex network structure. In the present
paper, we move a step toward ultimately understanding optimal active cyber
defense, where the attacker and/or the defender are/is strategic (i.e., they want
to minimize their cost). Finally, we note that automatic patching [27] is not active cyber defense because automatic patching aims to prevent attacks, whereas
active cyber defense aims to identify and possibly clean up infected computers.
There have been many studies (e.g., [33, 21, 8, 4, 14, 22, 15, 25]) on applying
Control Theory and Game Theory to understand various issues related to computer malware spreading. Our study is somewhat inspired by the botnet-defense
model investigated in [4]. All the studies mentioned above only considered reactive defense; whereas we investigate how to optimize active cyber defense. For
general information about the applications of Control Theory and Game Theory
to cyber security, we refer to [2, 17] and the references therein.
3
The Basic Active Cyber Defense Model
Consider a population of nodes, which can abstract computers in a cyber system.
At any point in time, a node is either occupied by defender B (i.e., the node is
secure), or occupied by attacker R (i.e., the node is compromised). Denote by
iB (t) the portion of nodes that are occupied by the defender at time t, and by
iR (t) the portion of nodes that are occupied by the attacker at time t, where
iB (t)+iR (t) = 1 for any t ≥ 0. In the interaction between cyber attack and active
cyber defense, the defender and the attacker can “grab” nodes from each other
in the same fashion. Let αB abstract defender B’s power in grabbing attackeroccupied nodes using active cyber defense, and αR abstract attacker R’s power in
compromising defender-occupied nodes using malware-like cyber attacks. Under
the homogeneous assumption that (i) each secure node has the same power in
“grabbing” the attacker-occupied nodes and (ii) each compromised node has
the same power in compromising the defender-occupied nodes, we obtain the
following Dynamical System model:
( diB (t)
= αB iB (t)iR (t) − αR iR (t)iB (t)
dt
diR (t)
dt
= αR iR (t)iB (t) − αB iB (t)iR (t),
where iB (t) + iR (t) = 1, iB (t) ≥ 0, and iR (t) ≥ 0 for all t ≥ 0. Due to the
symmetry, we only need to consider
diB (t)
= αB iB (t)(1 − iB (t)) − αR iB (t)(1 − iB (t)).
dt
(1)
If neither the attacker nor the defender is strategic (i.e., they do not consider
the issue of cost), the dynamics of system (1) can be characterized as follows.
– If the attacker is more powerful than the defender, namely αR > αB , the
attacker will occupy the entire network in the fashion of the Logistic equation
(i.e., when iR is small, iR increases slowly; when iR is around a threshold
value, iR increases exponentially; when iR is large, iR increases slowly).
– If the defender is more powerful than the attacker, namely αB > αR , the
defender will occupy the network in the same fashion as in the above case.
– If the attacker and the defender are equally powerful, namely αR = αB , the
system state is in equilibrium. In other words, iB (t) = iB (0) and iR (t) =
iR (0) = 1 − iB (0) for any t > 0.
The above model accommodates non-strategic attackers and non-strategic defenders, and is the starting point for our study of optimal active cyber defense.
4
4.1
Optimal Control for Strategic Defender Against
Non-Strategic Attacker
Infinite-time Horizon Optimal Control
In this setting, the non-strategic attacker R maintains a fixed degree of attack
power αR , while the defender B is strategic. That is, the strategic defender aims
to minimize its cost (specified below) by adjusting its defense power αB via
αB = b + πB (a − b),
while obeying the dynamics of (1), where πB ∈ [0, 1] is the control variable and
αB ∈ [b, a] is the defender’s defense power with 1 ≥ a > b ≥ 0. The cost to the
defender consists of two parts.
– The recovery cost for recovering the compromised nodes to secure states
(e.g., re-installing the operating systems and updating the backup data files,
interference with the computers’ routine functions). We represent this cost by
fB (iB (t)) for some real-valued function fB (·). We assume fB′ (·) < 0 because
the more nodes the defender occupies, the lower the cost for the defender to
recover the compromised nodes.
– The detection cost for detecting (or recognizing) compromised nodes via
active cyber defense, which partly depends on the attack’s evasiveness. We
represent this cost by kB ·πB (·), where kB is the normalization ratio between
the detection cost and the recovery cost, and πB (·) is the control function
that specifies the adjustable degree of active cyber defense power. This is
plausible because using more powerful active defense mechanisms (e.g., more
sophisticated/advanced “white” worms) causes a higher cost but allows the
defender to fight against the attacks more effectively.
The above definition of cost accommodates at least the following family of active cyber defense: The defender uses “white” worms to detect the compromised
nodes, then possibly manually recovers the compromised nodes. This is perhaps
the most probable scenario because for example, the attacker’s malware may
have corrupted or deleted some data files in the compromised computers. Note
that the detection cost highlights the difference between (i) active-cyber-defense
based detection, where the defender’s detection tools (i.e., “white” worms) do
not reside on the compromised computers, and (ii) reactive-cyber-defense based
detection such as the current generation of anti-virus software, where the detection tools do not spread over the network.
Assuming that the attacker maintains a fixed degree of attack power αR ,
the defender’s optimization goal is to minimize the total cost with a constant
discount rate z over an infinite-time horizon, namely
inf
0≤πB (·)≤1
Z
JB (πB (·)) =
0
∞
e−zt (fB (iB (t)) + kB · πB (t))dt ,
(2)
where fB′ (·) < 0, πB (·) ∈ [0, 1], and the attacker’s fixed degree of attack power αR
is treated as a constant. Now the optimization problem reduces to identifying
the optimal defense strategy π̂B . To solve the minimization problem, we use
Pontryagin’s Minimum Principle to find the Hamiltonian associated to (2):
HB (iB , πB , p)
= fB (iB ) + kB πB + p[αB iB (1 − iB ) − αR iB (1 − iB )]
= (kB + piB (1 − iB )(a − b))πB + fB (iB ) + pbiB (1 − iB ) − pαR iB (1 − iB ),(3)
where p is the adjoint equation
′
B
ṗ
= − ∂H
∂iB + zp = −fB (iB ) + p[z − (αB − αR )(1 − 2iB )]
p1 (∞) = 0.
(4)
The optimal strategy π̂B is obtained by minimizing the Hamiltonian HB (iB , πB , p).
Since HB (iB , πB , p) is linear in πB , the optimal control strategy π̂B takes the
following bang-bang control form:
1
if
π̂B = uB (0 < uB < 1, to be determined) if
0
if
∂HB
∂πB
∂HB
∂πB
∂HB
∂πB
B
where ∂H
∂πB = kB + piB (1 − iB )(a − b). In the singular form
period of time, we have
p=
−kB
.
iB (1 − iB )(a − b)
<0
=0
(5)
>0
∂HB
∂πB
= 0 and for a
(6)
Further differentiating
∂HB
∂πB
with respect to t, we have
d ∂HB
= ṗiB (1 − iB )(a − b) + p(1 − 2iB )i̇B (a − b)
dt ∂πB
= iB (1 − iB )(a − b) − fB′ (iB ) + p[z − (αB − αR )(1 − 2iB )]
+p(1 − 2iB )(a − b) αB iB (1 − iB ) − αR iB (1 − iB )
= −iB (1 − iB )(a − b)fB′ (iB ) − kB z
Define FB (iB ) = −iB (1 − iB )(a − b)fB′ (iB ) − kB z. Then we need to study the
roots of FB (·) = 0.
B
0.2
0.15
B
y=i (1−i )(a−b)
0.25
0.1
y=k z
B
0.05
i
i
2
1
0
0
0.2
0.4
i
B
0.6
0.8
1
Fig. 1. Illustration of the roots of FB (iB ) = 0 with fB (iB ) = 1−iB , a−b = 1 and kB z =
1/8, where the x-axis represents iB and the y-axis represents y(iB ) = iB (1 − iB )(a − b).
The arrows indicate the directions the outcome under optimal control will head for.
Before presenting the results, we discuss the ideas behind them. In this paper,
we focus on case fB (iB ) = 1 − iB , which can be easily extended to any linear
recovery-cost function. If kB z < 14 (a − b), then FB (iB ) = 0 has two roots:
1−
i1 =
q
Bz
1 − 4 ka−b
2
1+
and i2 =
q
Bz
1 − 4 ka−b
2
with 0 < i1 < i2 < 1. As illustrated in Figure 1, this implies
if iB < i1
FB (iB ) < 0
FB (iB ) > 0
if i1 < iB < i2
FB (iB ) < 0
if iB > i2 .
Then, the optimal strategy π̂B of the singular form can be obtained by solving
i̇B |iB =i1 or iB =i2 = 0.
Theorem 1. Suppose the non-strategic attacker maintains a fixed degree of attack power αR , fB (iB ) = 1 − iB and kB z < 14 (a − b). Let i1 < i2 be the roots of
R −b
. The optimal control strategy for defender B is:
FB (iB ) = 0. Let uB = αa−b
0
if iB < i1
if iB = i1
uB
if i1 < iB < i2 .
π̂B = 1
(7)
u
if
i
=
i
B
B
2
0
if iB > i2
Proof of Theorem 1 is deferred to Appendix A. In practice, i1 and i2 can
be obtained numerically. Theorem 1 (also as illustrated in Figure 1) shows that
the outcome of the infinite-time horizon optimal control, namely limt→∞ iB (t),
depends on the initial system state iB (0) as follows:
– If 1 > iB (0) > i2 , the defender should use the least powerful/costly active
defense mechanisms (i.e., αB = b) because π̂B = 0. Moreover, the outcome
of the optimal defense is that the defender will occupy i2 portion of the
network, namely limt→∞ iB (t) = i2 . This suggests a sort of diminishing
returns in active cyber defense: It is more cost-effective to pursue “good
enough” security (i.e., limt→∞ iB (t) = i2 < 1) than to pursue “perfect”
security (i.e., limt→∞ iB (t) = 1) even if it is possible.
– If 0 = iB (0) < i1 , the defender should use the least powerful/costly active
defense mechanisms (i.e., αB = b) because π̂B = 0. Moreover, the outcome
of the optimal defense is that the defender should give up (using active cyber
defense as the only defense methods), as the attacker will occupy the entire
network, namely limt→∞ iB (t) = 0. In other words, the defender should
resort to other defense methods as well (e.g., proactive defense).
– If iB (0) ∈ (i1 , i2 ), the defender should use the most powerful/costly active
defense mechanisms (i.e., αB = a) because π̂B = 1. Moreover, the outcome
of the optimal defense is that the defender will occupy i2 portion of the
network, namely limt→∞ iB (t) = i2 . This also suggests a sort of diminishing
returns mentioned above.
– If iB (0) = i1 or iB (0) = i2 , the defender should adjust its deployment
R −b
of active cyber defense mechanisms according to uB = αa−b
, which means
αB = αR . Moreover, the outcome of the optimal defense is that iB (t) = iB (0)
for all t > 0.
Now we consider the degenerated scenarios of kB z ≥ 1/4(a − b). The proof
is similar to, but much simpler than, the proof of Theorem 1, and thus omitted.
Theorem 2. Suppose the non-strategic attacker maintains a fixed degree of attack power αR and fB (iB ) = 1 − iB .
– If kB z = 1/4(a − b), then FB (iB ) = 0 has only one root, i1 = i2 = 21 . The
optimal control strategy is
if iB < i1
0
R −b
if iB = i1
π̂B = uB = αa−b
(8)
0
if iB > i1 .
– If kB z > 1/4(a − b), then FB (iB ) = 0 has no root. The optimal control
strategy is π̂B = 0.
The cyber security implications of Theorem 2 are the following. In the case
kB z = 41 (a − b), the outcome under the optimal control depends on the initial
system state as follows:
– If 1 > iB (0) > i1 , the defender should use the least powerful/costly active
cyber defense mechanisms because π̂B = 0. The outcome is that the defender
will occupy i1 portion of the network, namely limt→∞ iB (t) = i1 .
– If 0 = iB (0) < i1 , the defender should use the least powerful/costly active
cyber defense mechanisms because π̂B = 0. The outcome is that the defender
will give up using active cyber defense alone, as the attacker will occupy the
entire the network, namely limt→∞ iB (t) = 0. In other words, the defender
should resort to other defense methods as well (e.g., proactive defense).
– If iB (0) = i1 , the defender will adjust its degree of active cyber defense power
R −b
, which means αB = αR . The outcome is that
according to π̂B = uB = αa−b
iB (t) = iB (0) for all t > 0.
In the case kB z > 1/4(a−b), the defender should use the least powerful/costly active cyber defense mechanisms because π̂B = 0. The outcome is that limt→∞ iB (t) =
0, meaning that the defender should give up using active cyber defense alone and
resort to other defense methods as well (e.g., proactive defense).
By considering Theorems 1 and 2 together, we draw some deeper insights.
Specifically, for a given z, different kB ’s suggest different optimal active defense
1
(a − b), meaning that the cost of optimal
strategies. More specifically, if kB > 4z
control is dominating, then defender B should use the least powerful/costly
active cyber defense mechanisms because π̂B (t) = 0 for all t and the outcome is
limt→∞ iB = 0. In other words, the defender should give up using active cyber
defense alone, and resort to other kinds of defense methods as well (e.g., proactive
1
(a − b), meaning that the cost of control is not dominating,
defense). If kB < 4z
the defender should enforce optimal control according to the initial state iB (0).
In particular, if kB = 0, meaning that the special case that the cost of control
is not counted, defender B should use the most powerful/costly active defense
mechanisms as π̂B (t) = 1 for all t, and the outcome is that limt→∞ iB = 1,
namely that the defender will occupy the entire network.
4.2
Fast Optimal Control for Strategic Defenders against
Non-Strategic Attackers
Now we consider fast optimal control for strategic defenders against non-strategic
attackers, as motivated by the following question: Suppose the attacker maintains
a fixed degree of attack power αR and the defender initially occupies iB (0) =
i0 < ie portions of the nodes, how can the defender use optimal control to
occupy the desired ie portions of the nodes as soon as possible? More precisely,
the optimization is to minimize the sum of active defense cost and time (after
appropriate normalization), which can be described by the following functional:
Z T
h(πB (t))dt
JF (πB (·)) = T + λ
0
where h(·) is the cost function with respect to the control function πB (·). We
consider two scenarios of cost functions: linear and quadratic. In both scenarios,
we need to identify defender B’s optimal strategy with respect to the dynamics
of (1) and a given objective ie > i0 for some hitting time T that is to be identified.
Scenario I: Fast optimal control with linear cost functions. In this scenario, we have h(πB ) = πB . The optimization task is to minimize the active
defense cost plus the time T :
)
(
Z T
πB (t)dt
(9)
JF (πB (·)) = T + λ
inf
0≤πB (·)≤1
0
di (t)
Bdt = αB iB (t)(1 − iB (t)) − αR iB (t)(1 − iB (t))
subject to
i (0) = i0
B
iB (T ) = ie
where λ > 0 Ris the normalization ratio between the unit of time and the active
T
defense cost 0 πB (t)dt, and i0 < ie . That is, λ, i0 and ie are given, but T is
RT
free. Note that the active defense cost 0 πB (t)dt includes both detection and
recovery cost, where πB (t) is the control function.
Theorem 3. The solution to the fast optimal control problem (9) is
where T1 =
1
a−αR
∗
(πB
, T ∗ ) = (1, T1 ),
ie 1−i0
.
ln 1−i
i0
e
(10)
Proof of Theorem 3 is deferred to Appendix B. The cyber security implication
of Theorem 3 is the following. In order to achieve fast optimal control, the
defender should use the most powerful/costly active cyber defense mechanisms,
namely πB (t) = 1 for t < T ∗ , until the system state becomes iB (T ∗ ) = ie at time
T ∗ . After time T ∗ , if the defender continues enforcing πB (t) = 1 for t > T ∗ , then
limt→∞ iB (t) = 1, meaning that the defender will occupy the entire network.
Scenario II: Fast optimal control with quadratic cost functions. In this
2
scenario, we have h(πB ) = πB
. The optimization task is to minimize the following
sum of active defense cost and time, which differs from the linear cost (9) in that
2
the cost function πB is replaced with cost function πB
:
(
)
Z
T
inf
0≤πB (·)≤1
JF (πB (·)) = T + λ
0
2
(t)dt
πB
di (t)
Bdt = αB iB (t)(1 − iB (t)) − αR iB (t)(1 − iB (t))
subject to
i (0) = i0
B
iB (T ) = ie
(11)
where
R T 2 λ > 0 is the ratio between the unit of time and the active defense cost
0 πB (t)dt (including both recovery cost and detection cost), and i0 < ie . That
is, λ, i0 and ie are given, but T is free.
Theorem 4. The solution to the fast optimal control problem (11) is
(
a−b
(u∗ , T2 ), if λ ≥ a+b−2α
and a − b > 2(αR − b),
∗∗
∗∗
R
(πB , T ) =
(1, T3 ), otherwise
(12)
where
r
αR − b
b − αR 2 1
+ ,
+
u =
a−b
a−b
λ
1
ie 1 − i0
,
ln
T2 =
b + (a − b)u∗ − αR
1 − ie i0
1
ie 1 − i0
T3 =
.
ln
a − αR
1 − ie i0
∗
Proof of Theorem 4 is deferred to Appendix C. It cyber security implication is:
Unlike in the setting of linear cost function (Theorem 3), the defender should not
necessarily enforce the most powerful/costly active cyber defense mechanisms as
∗∗
πB
is not always equal to 1. If the defender continues enforcing πB (t) = 1 for
t > T ∗∗ after the system reaches state iB (T ∗∗ ) = ie at time T ∗∗ , the defender
will occupy the entire network, namely limt→∞ iB (t) = 1.
5
Nash Equilibria for Strategic Attacker and Defender
Now we ask the question: What if the attacker is also strategic? Analogous to the
way of modeling strategic defenders, we assume αR ∈ [b, a]. (It is straightforward
to extend the current setting αB , αR ∈ [b, a] to the setting αB ∈ [bB , aB ] and
αR ∈ [bR , aR ].) A strategic attacker can adjust its attack power
αR = b + πR (a − b),
via control variable πR (·) ∈ [0, 1]. That is, the attacker can launch more sophisticated attacks (i.e., greater πR leading to greater αR ), which however incurs
higher cost (e.g., the investment for obtaining more powerful attack tools).
Since both the defender and the attacker are strategic, we naturally consider
a game-theoretic model. Specifically, the defender B’s optimization task is
Z ∞
−zt
JB (πB (·), πR (·)) =
e (fB (iB (t)) + kB · πB (iB (t)))dt ,
φB (iB ) =
inf
0≤πB (·)≤1
0
and the attacker R’s optimization task is
Z
JR (πB (·), πR (·)) =
φR (iB ) =
inf
0≤πR (·)≤1
0
∞
e−zt (fR (iB (t)) + kR · πR (iB (t)))dt ,
where πB (·), πR (·) ∈ [0, 1], fB′ (·) < 0 (as in the infinite-time horizon optimal
control case investigated above), fR′ (·) > 0 because fR (iB (t)) represents the
maintenance cost to the attacker, kR is the normalization ratio between the attacker’s maintenance cost and penetration cost (which depends on the capability
of the attack tools), and kR · πR (·) is the penetration cost. Note that fR′ (·) > 0
is relevant because the attacker may need to conduct some costly (or risky)
activities after “grabbing” a node from the defender (e.g., downloading attack
payloads from some remote server, while this downloading operation may increase the chance that the compromised node is detected by active defense).
Since fR′ (·) > 0 implies dfR /diR < 0, the attacker’s optimization task for πR is
in parallel to the optimization for πB . The Hamiltonians associated to defender
B’s and attacker R’s optimization problems are:
HB (iB , πB (iB ), πR (iB ), p1 )
= fB (iB ) + kB πB + p1 [αB iB (1 − iB ) − αR iB (1 − iB )]
= (kB + p1 iB (1 − iB )(a − b))πB + fB (iB ) + p1 biB (1 − iB ) − p1 αR iB (1 − iB );
HR (iB , πB (iB ), πR (iB ), p2 )
= fR (iB ) + kR πR + p2 [αB iB (1 − iB ) − αR iB (1 − iB )]
= (kR − p2 iB (1 − iB )(a − b))πR + fR (iB ) + p2 αB iB (1 − iB ) − p2 biB (1 − iB ).
The adjoint equation is
′
B
ṗ1
= − ∂H
∂iB + zp1 = −fB (iB ) + p1 [z − (αB − αR )(1 − 2iB )]
p1 (∞) = 0
′
R
ṗ
= − ∂H
∂iB + zp2 = −fR (iB ) + p2 [z − (αB − αR )(1 − 2iB )]
2
p2 (∞) = 0.
Theorem 5. Suppose fB (iB ) = 1 − iB , fR (iB ) = iB . Then, the Nash equilibria
under various scenarios are listed in Table 1, where FB (iB ) = −iB (1 − iB )(a −
b)fB′ (iB ) − kB z and FR (iB ) = iB (1 − iB )(a − b)fR′ (iB ) − kR z.
Proof of Theorem 5 is similar to the proof of Theorem 1 and omitted due
to space limitation. Its cyber security implication is: The outcome of playing
the Nash equilibrium strategies also depends on the initial system state and
the relationship between kB and kR . As illustrated in Figure 2, if kB < kR with
kR z < 41 (a−b), meaning that the attacker is more concerned with its control cost
(e.g., reluctant to use/expose its advanced attack tools such as zero-day exploits)
than the defender, then FB (iB ) = 0 has two roots i1 , i2 and FR (iB ) = 0 has two
roots i3 , i4 . Then, we have i1 < i3 < i4 < i2 (the only possibility under the given
conditions). Therefore, the outcomes under the Nash equilibrium strategies are
summarized as follows:
– If iB (0) < i1 , then iB (t) = iB (0) and iR (t) = iR (0) for all t > 0 because
π̂B = π̂R = 0 are the Nash equilibrium strategies.
– If i3 > iB (0) > i1 , then π̂B = 1 and π̂R = 0 until iB = i3 , which implies that
iB (t) strictly increases until iB = i3 . When iB (t) = i3 at some point in time
t = t1 , π̂B = π̂R = 1 implies iB (t) = i3 for t > t1 .
Table 1. Nash equilibrium strategies for defender and attacker in various cases.
kB
Roots of FB (iB ) = 0 Roots of FR (iB ) = 0
kR
kB z <
1
4 (a
− b) kR z <
1
4 (a
− b)
kB z <
1
4 (a
− b) kR z =
1
4 (a
− b)
kB z <
1
4 (a
− b) kR z >
1
4 (a
− b)
kB z =
1
4 (a
− b) kR z <
1
4 (a
− b)
kB z =
1
4 (a
− b) kR z =
1
4 (a
− b)
kB z =
1
4 (a
− b) kR z >
1
4 (a
− b)
kB z >
1
4 (a
− b) kR z <
1
4 (a
− b)
kB z >
kB z >
1
4 (a
1
4 (a
− b) kR z =
− b) kR z >
1
4 (a
1
4 (a
− b)
− b)
Nash equilibria
0 if iB (0) ≤ i1
1 if i1 < iB (0) < i2
π̂B =
0 if iB (0) ≥ i2
0 < i1 < i2 < 1
0 < i3 < i4 < 1
0 if iB (0) < i3
1 if i3 ≤ iB (0) ≤ i4
π̂R =
0 if i (0) > i
B
4
0 if iB (0) ≤ i1
1 if i1 < iB (0) < i2
π̂B =
0 if i (0) ≥ i
B
2
0 < i1 < i2 < 1
i3 = i4 = 21
0 if iB (0) < i3
1 if iB (0) = i3
π̂R =
0 if iB (0) > i3
0 if iB (0) ≤ i1
1 if i1 < iB (0) < i2
π̂ =
0 < i1 < i2 < 1
No real-valued roots B
0 if i (0) ≥ i
B
2
π̂R = 0
0 if iB (0) < i1
1 if iB (0) = i1
π̂B =
0 if iB (0) > i2
0 < i1 = i2 = 21
0 < i3 < i4 < 1
0 if iB (0) ≤ i3
1 if i3 < iB (0) < i4
π̂R =
0 if i (0) ≥ i
B
4
0 if iB (0) < i1
πR if iB (0) = i1
π̂B =
0 if iB (0) > i2
0 < i1 = i2 = 12
i3 = i4 = 12
0 if iB (0) < i3
πB if iB (0) = i3
π̂R =
0 if i (0) > i
B
3
0 < i1 = i2 = 21
No real-valued roots
π̂B = 0, π̂R = 0
π̂B = 0
0 if iB (0) ≤ i3
No real-valued roots
0 < i3 < i4 < 1
1 if i3 < iB (0) < i4
π̂R =
0 if i (0) ≥ i
B
4
π̂B = 0, π̂R = 0
No real-valued roots
i3 = i4 = 21
π̂B = 0, π̂R = 0
No real-valued roots No real-valued roots
B
0.2
0.15
B
y=i (1−i )(a−b)
0.25
0.1
y=k z
R
y=k z
B
0.05
i
0
0
1
0.2
i4
i
3
0.4
i
B
0.6
i
2
0.8
1
Fig. 2. Illustration of the roots of FB (iB ) = 0 with fB (iB ) = 1 − iB , and the roots of
FR (iB ) = 0 with fR (iB ) = iB , where a − b = 1, kB z = 1/8 and kR z = 1/6. The x-axis
represents iB and the y-axis represents y(iB ) = iB (1 − iB )(a − b). Arrows indicate
the directions the outcome under the Nash equilibrium heads for. Black-colored bars
indicate that the trajectory under the Nash equilibrium stays static.
– If i4 > iB (0) > i3 , then iB (t) = iB (0) and iR (t) = iR (0) for all t > 0 because
π̂B = π̂R = 1.
– If i2 > iB (0) > i4 , then π̂B = 1 and π̂R = 0 until iB = i2 , which implies that
iB (t) strictly increases until iB = i2 . When iB (t) = i2 at some point in time
t = t2 , π̂B = π̂R = 1 implies iB (t) = i3 for t > t2 .
– If iB (0) > i2 , then iB (t) = iB (0) and iR (t) = iR (0) for all t > 0 because
π̂B = π̂R = 1.
If kR > 14 (a − b) > kB , meaning that the attacker is extremely concerned
with its control cost (e.g., not willing to easily use/expose its advanced attack
tools such as zero-day exploits) but the defender is not, then it always holds
that π̂R = 0 because FR (iB ) = 0 has no root but FB (iB ) = 0 has two roots
i1 < i2 . From Table 1, we see that limt→∞ iB (t) = 1 always holds, namely that
the attacker gives up using its advanced attack tools.
If both kB > 14 (a− b) and kR > 41 (a− b), meaning that both the defender and
the attacker are extremely concerned with their control costs (i.e., neither the
defender wants to easily use/expose its advanced active defense tools, nor the
attacker wants to use/expose its advanced attack tools such as zero-day exploits),
then it always holds that π̂B = π̂R = 0 because FB (iB ) = 0 and FR (iB ) = 0
have no real-valued roots. As a result, iB (t) = iB (0) for any t > 0.
The scenarios that one or both FB (iB ) = 0 and FR (iB ) = 0 have one root can
be regarded as degenerated cases of the above. Moreover, the cases of kB > kR
(i.e., the defender is more concerned about its control cost, such as not willing
to easily use/expose its advanced active defense tools), the outcomes under the
Nash equilibria can be derived analogously.
6
Conclusion
We have investigated how to optimize active cyber defense, by presenting optimal control solutions for strategic defenders against non-strategic attackers, and
identifying Nash equilibrium strategies for strategic defenders and attackers. We
have discussed the cyber security implications of the theoretic results.
This paper brings interesting problems for future research. First, it is interesting to extend the models to accommodate nonlinear fB (·) and fR (·). Second,
the models are geared toward active cyber defense. A comprehensive defense
solution, as hinted in our analysis, should require the optimal integration of reactive, active, and proactive cyber defense. Therefore, we need to extend the
models to accommodate reactive defense and proactive cyber defense. Moreover,
it is interesting to investigate how to extend the models to accommodate moving
target defense, which has not be systematically evaluated yet [7]. Third, how to
extend the models to accommodate the underlying network structures?
Acknowledgement. Wenlian Lu was jointly supported by the Marie Curie
International Incoming Fellowship from the European Commission (no. FP7PEOPLE-2011-IIF-302421), the National Natural Sciences Foundation of China
(no. 61273309), the Shanghai Guidance of Science and Technology (SGST) (no.
09DZ2272900) and the Laboratory of Mathematics for Nonlinear Science, Fudan
University. Shouhuai Xu was supported in part by ARO Grant #W911NF-12-10286 and AFOSR Grant FA9550-09-1-0165. Any opinions, findings, and conclusions or recommendations expressed in this material are those of the author(s)
and do not necessarily reflect the views of any of the funding agencies.
References
1. D. Aitel.
Nematodes – beneficial worms.
http://www.immunityinc.com/
downloads/nematodes.pdf, Sept. 2005.
2. T. Alpcan and T. Başar. Network Security: A Decision and Game Theoretic Approach. Cambridge University Press, 2011.
3. M. Bardi and I. Capuzzo-Dolcetta. Optimal control and viscosity solutions of
Hamilton-Jacobi-Bellman equations. Birkhauser, 2008.
4. Alain Bensoussan, Murat Kantarcioglu, and SingRu Hoe. A game-theoretical
approach for finding optimal strategies in a botnet defense model. In Proc.
GameSec’10, pages 135–148, 2010.
5. F. Castaneda, E. Sezer, and J. Xu. Worm vs. worm: preliminary study of an active
counter-attack mechanism. In Proc. ACM WORM’04, pages 83–93, 2004.
6. D. Chakrabarti, Y. Wang, C. Wang, J. Leskovec, and C. Faloutsos. Epidemic
thresholds in real networks. ACM Trans. Inf. Syst. Secur., 10(4):1–26, 2008.
7. M. Collins. A cost-based mechanism for evaluating the effectiveness of moving
target defenses. In Proc. GameSec’12, pages 221–233, 2012.
8. Neal Fultz and Jens Grossklags. Blue versus Red: Towards a Model of Distributed
Security Attacks. In Proc. Financial cryptography and data security (FC’99), pages
167–183. 2009.
9. A. Ganesh, L. Massoulie, and D. Towsley. The effect of network topology on the
spread of epidemics. In Proc. of IEEE Infocom 2005, 2005.
10. J. Kephart and S. White. Directed-graph epidemiological models of computer
viruses. In Proc. IEEE Symposium on Security and Privacy, pages 343–361, 1991.
11. J. Kephart and S. White. Measuring and modeling computer virus prevalence. In
Proc. IEEE Symposium on Security and Privacy, pages 2–15, 1993.
12. W. Kermack and A. McKendrick. A contribution to the mathematical theory of
epidemics. Proc. of Roy. Soc. Lond. A, 115:700–721, 1927.
13. J. Kesan and C. Hayes. Mitigative counterstriking: Self-defense and deterrence
in cyberspace. Harvard Journal of Law and Technology (forthcoming, available at
SSRN: http: // ssrn. com/ abstract= 1805163 ).
14. M. Khouzani, S. Sarkar, and E. Altman. A dynamic game solution to malware
attack. In Proc. IEEE INFOCOM, pages 2138–2146, 2011.
15. M. Khouzani, S. Sarkar, and E. Altman. Saddle-point strategies in malware attack.
IEEE Journal on Selected Areas in Communications, 30(1):31–43, 2012.
16. H. Lin. Lifting the veil on cyber offense. IEEE Security & Privacy, 7(4):15–21,
2009.
17. M. Manshaei, Q. Zhu, T. Alpcan, T. Basar, and J. Hubaux. Game theory meets
network security and privacy. ACM Computing Survey, to appear.
18. W. Matthews. U.s. said to need stronger, active cyber defenses. http://www.
defensenews.com/story.php?i=4824730, 1 Oct 2010.
19. A. McKendrick. Applications of mathematics to medical problems. Proc. of Edin.
Math. Soceity, 14:98–130, 1926.
20. R.
Naraine.
’friendly’
welchia
worm
wreaking
havoc.
http://www.internetnews.com/ent-news/article.php/3065761/
Friendly-Welchia-Worm-Wreaking-Havoc.htm, August 19, 2003.
21. J. Omic, A. Orda, and P. Van Mieghem. Protecting against network infections: A
game theoretic perspective. In Infocom’09, pages 1485–1493, 2009.
22. R. Pı́bil, V. Lisý, C. Kiekintveld, B. Bosanský, and M. Pechoucek. Game theoretic
model of strategic honeypot selection in computer networks. In Proc. GameSec’12,
pages 201–220, 2012.
23. B. Schneier. Benevolent worms. http://www.schneier.com/blog/archives/
2008/02/benevolent_worm_1.html, February 19, 2008.
24. L. Shaughnessy. The internet: Frontline of the next war? http://www.cnn.com/
2011/11/07/us/darpa/, November 7, 2011.
25. George Theodorakopoulos, Jean-Yves Le Boudec, and John S. Baras. Selfish response to epidemic propagation. IEEE Trans. Aut. Contr., 58(2):363–376, 2013.
26. Piet Van Mieghem, Jasmina Omic, and Robert Kooij. Virus spread in networks.
IEEE/ACM Trans. Netw., 17(1):1–14, February 2009.
27. M. Vojnovic and A. Ganesh. On the race of worms, alerts, and patches. IEEE/ACM
Trans. Netw., 16:1066–1079, October 2008.
28. Y. Wang, D. Chakrabarti, C. Wang, and C. Faloutsos. Epidemic spreading in real
networks: An eigenvalue viewpoint. In Proc. IEEE SRDS’03, pages 25–34, 2003.
29. N. Weaver and D. Ellis. White worms don’t work. ;login: The USENIX Magazine,
31(6):33–38, 2006.
30. Homeland Security News Wire.
Active cyber-defense strategy best deterrent against cyber-attacks.
http://www.homelandsecuritynewswire.com/
active-cyber-defense-strategy-best-deterrent-against-cyber-attacks, 28
June 2011.
31. J. Wolf. Update 2-u.s. says will boost its cyber arsenal. http://www.reuters.com/
article/2011/11/07/cyber-usa-offensive-idUSN1E7A61YQ20111107, November
7, 2011.
32. S. Xu, W. Lu, and L. Xu. Push- and pull-based epidemic spreading in arbitrary
networks: Thresholds and deeper insights. ACM Transactions on Autonomous and
Adaptive Systems (ACM TAAS), 7(3):32:1–32:26, 2012.
33. S. Xu, W. Lu, L. Xu, and Z. Zhan. Adaptive epidemic dynamics in networks:
Thresholds and control. ACM Transactions on Autonomous and Adaptive Systems
(ACM TAAS), to appear.
34. S. Xu, W. Lu, and Z. Zhan. A stochastic model of multivirus dynamics. IEEE
Trans. Dependable Sec. Comput., 9(1):30–45, 2012.
35. Shouhuai Xu, Wenlian Lu, and Hualun Li. A stochastic model of active cyber
defense dynamics. Internet Mathematics, to appear.
A
Proof of Theorem 1
Proof. By the Dynamic Programming (DP) argument [3], we know that defender
B’s value function of the optimal solution can be defined as:
Z ∞
−zt
JB (πB (·)) =
e (fB (iB (t)) + kB · πB (t))dt . (13)
φ(iB ) =
inf
0≤πB (·)≤1
0
This leads to the following Bellman equation:
′
fB (iB ) + kB πB (t) + φ (iB )[αB iB (1 − iB ) − αR iB (1 − iB )]
zφ(iB ) =
inf
0≤πB (·)≤1
=
=
inf
HB (iB , πB (t), φ′ (iB ))
inf
HB (iB , πB (t), p), where p = φ′ (iB ).
0≤πB (·)≤1
0≤πB (·)≤1
(14)
From (5), we know that the optimal strategy π̂B takes the form:
π̂B = 1kB +piB (1−iB )(a−b)<0 + uB 1kB +piB (1−iB )(a−b)=0 ,
(15)
where 1 is the indicator function. The infimum of Hamiltonian (3) is:
inf
0≤πB (·)≤1
HB (iB , πB , p)
= fB (iB ) + [kB + piB (1 − iB )(a − b)]1kB +piB (1−iB )(a−b)<0 + p(b − αR )iB (1 − iB ).
Hence, we have
zφ(iB )
= fB (iB ) + [kB + piB (1 − iB )(a − b)]1kB +piB (1−iB )(a−b)<0 + p(b − αR )iB (1 − iB )
= fB (iB ) + [kB + φ′ (iB )iB (1 − iB )(a − b)]1kB +φ′ (iB )iB (1−iB )(a−b)<0 +
φ′ (iB )(b − αR )iB (1 − iB ).
(16)
Let y(iB ) = kB + φ′ (iB )iB (1 − iB )(a − b). In what follows, we are to verify that
(7) satisfies (15) with φ(iB ) defined by (16), which means that (7) minimizes the
Hamiltonian in the term of (13). This completes the proof.
In order to verify that (7) satisfies (15) with φ(iB ) defined by (16), we differentiate (16) with respect to iB to obtain
[(b − αR ) + (a − b)1y<0 ]iB (1 − iB )y ′ − zy − FB (iB ) = 0,
(17)
which can be rewritten as
y′ −
z
FB (iB )
y−
= 0.
[(b − αR ) + (a − b)1y<0 ]iB (1 − iB )
[(b − αR ) + (a − b)1y<0 ]iB (1 − iB )
(18)
If kB + φ′ (iB )iB (1 − iB )(a − b) < 0 namely y(iB ) < 0, (18) should be
R
R
z
z
FB (x)
d
− x
dξ
− x
dξ
y(x)e 0 (a−αR )ξ(1−ξ)
−
e 0 (a−αR )ξ(1−ξ) = 0. (19)
dx
(a − αR )x(1 − x)
If kB + φ′ (iB )iB (1 − iB )(a − b) > 0 namely y(iB ) > 0, (18) should be:
R1
R1
z
z
d
FB (x)
dξ
dξ
x (b−αR )ξ(1−ξ)
y(x)e
−
e x (b−αR )ξ(1−ξ) = 0.
dx
(b − αR )x(1 − x)
(20)
Therefore, we only need to prove that the optimal defense strategy (7) satisfies
(17), namely (19) or (20). The proof is split into cases, depending on x residing
in interval (i2 , 1) or (0, i1 ) or (i1 , i2 ), or x = i1 , or x = i2 .
Case 1: i2 < x < 1. By (20), we have
y(x)e
R1
z
dξ
x (b−αR )ξ(1−ξ)
−
Z
x
i2
R1
z
FB (ζ)
dξ
e ζ (b−αR )ξ(1−ξ) dζ = 0.
(b − αR )ζ(1 − ζ)
Hence, we have
y(x) =
Z
x
i2
Rx
FB (ζ)
eζ
(b − αR )ζ(1 − ζ)
z
(b−αR )ξ(1−ξ) dξ
dζ.
(21)
Since iB > i2 , we have y(x) > 0. Therefore, we have π̂B = 0 for iB ∈ (i2 , 1).
Case 2: 0 < x < i1 . By (20), we have
y(x)e
−
Rx
0
z
dξ
(b−αR )ξ(1−ξ)
−
Z
x
i1
R
FB (ζ)
− ζ
e 0
(b − αR )ζ(1 − ζ)
z
dξ
(b−αR )ξ(1−ξ)
dζ = 0.
Hence,
x
Rx
Rx
z
FB (ζ)
dξ
e ζ (b−αR )ξ(1−ξ) dζ + kB e 0
0 (b − αR )ζ(1 − ζ)
Z x
Rx
z
a−b ′
dξ
fB (ζ)e ζ (b−αR )ξ(1−ξ) dζ
=kB −
0 b − αR
y(x) =
Z
z
dξ
(b−αR )ξ(1−ξ)
(22)
Since iB < i1 , we have y(x) > 0. Therefore we have π̂B = 0 for iB ∈ (0, i1 ).
Case 3: i1 < x < i2 . By (19), we have
y(x)e
R1
z
dξ
x (a−αR )ξ(1−ξ)
−
Z
x
i2
R1
z
FB (ζ)
dξ
e ζ (a−αR )ξ(1−ξ) dζ = 0.
(a − αR )ζ(1 − ζ)
Hence
y(x) =
Z
x
i2
Rx
FB (ζ)
eζ
(a − αR )ζ(1 − ζ)
z
dξ
(a−αR )ξ(1−ξ)
dζ.
(23)
Since iB ∈ (i1 , i2 ), we have y(x) < 0. This implies π̂B = 1.
Cases 4 & 5: x = i1 or x = i2 . By (21,22,23), we have y(x) = 0. If x = i1 or
−kB
x = i2 , we can derive φ′ (iB ) = i∗ (1−i
∗ )(a−b) from the definition of y(·). According
R −b
. Differentiating with respect to i∗ ,
to (16), we have: zφ(i∗ ) = fB (i∗ ) + kB αa−b
∗
∗
′
∗
we have −i (1 − i )(a − b)fB (i ) − kB z = FB (i∗ ) = 0. Consider the singular
form iB (t) = i∗ for a period of time. We obtain that i̇B |iB =i∗ = 0 and thus
R −b
π̂B = uB = αa−b
, where i∗ = i1 or i∗ = i2 .
⊓
⊔
B
Proof of Theorem 3
Proof. To solve the minimization problem, we formulate the current value Hamiltonian associated with (9):
HF (iB , πB , q) = λπB + q(αB − αR )iB (1 − iB )
= [λ + q(a − b)iB (1 − iB )]πB + q(b − αR )iB (1 − iB ).
F
The adjoint equation is q̇ = − ∂H
∂iB = −q(αB − αR )(1 − 2iB ), with the boundary
condition
∗
HF (i∗B (T ∗ ), πB
(T ∗ ), q(T ∗ )) + 1 = 0,
(24)
∗
where T ∗ denotes the optimal hitting time that iB (T ∗ ) = ie , πB
(·) denotes the
∗
optimal feedback control, and iB (·) denotes the corresponding trajectory.
∗
The optimal control πB
is obtained by minimizing Hamiltonian HF (iB , πB , q).
∗
Since HF (iB , πB , q) is linear in πB , the optimal control πB
takes the following
bang-bang form:
F
1
if ∂H
∂πB < 0
∗
F
πB
= u∗B (0 < u∗B < 1, to be determined) if ∂H
∂πF = 0
0
if ∂HF > 0
∂πB
F
where ∂H
∂πB = λ + q(a − b)iB (1 − iB ). From (24), there are two possibilities: (i).
∂HF
∂HF
a−b
R ∂HF
If ∂πB ≥ 0, then 0 = b−α
a−b ( ∂πB − λ) + 1, which implies ∂πB = αR −b + λ is a
b−αR ∂HF
F
F
< 0, then 0 = ∂H
positive constant. (ii). If ∂H
∂πB + a−b ( ∂πB − λ) + 1, which
h ∂πB
i
b−αR
a−b
F
implies ∂H
∂πB = a−αR
a−b λ − 1 is a negative constant. It can be seen that
only under the above (ii), the constraint iB (T ) = ie can be obtained for some T .
Then, the solution to the optimal fast control should be πB (t) = 1 for all time.
∗
So (πB
, T ∗ ) = (1, T1 ), where T1 satisfies
iB (T1 ) =
that is, T1 =
C
1
a−αR
ln
ie 1−i0
1−ie i0
1
i0
(a−αR )T1
1−i0 e
i0
+ 1−i0 e(a−αR )T1
= ie ,
. This completes the proof.
⊓
⊔
Proof of Theorem 4
Proof. To solve the optimization problem, we formulate the current value Hamiltonian associated with (11):
2
HF (iB , πB , q) = λπB
+ q(αB − αR )iB (1 − iB )
2
= λπB
+ q(a − b)iB (1 − iB )πB + q(b − αR )iB (1 − iB ).
F
The adjoint equation is q̇ = − ∂H
∂iB = −q(αB − αR )(1 − 2iB ), and the boundary
condition is
∗∗
HF (iB (T ∗∗ ), πB
(T ∗∗ ), q(T ∗∗ )) + 1 = 0,
(25)
∗∗
where T ∗∗ denotes the optimal final time, πB
(·) denotes the optimal feedback
control, iB (·) denotes the corresponding trajectory, and iB (T ∗∗ ) = ie . Let D =
q(a − b)iB (T ∗∗ )(1 − iB (T ∗∗ )). From (25) we have
b − αR
D + 1 = 0. (26)
a−b
∗∗
The optimal control, πB
, is obtained by minimizing the Hamiltonian HF (iB , πB , q).
Because the Hamiltonian HF (iB , πB , q) is quadratic in πB , the optimal control,
∗∗
πB
, takes the following form:
D
>1
1
if − 2λ
∗∗
D
D
πB
= − 2λ
(0 < u∗B < 1, to be determined) if 0 ≤ − 2λ
≤1
0
D
if −
< 0.
∗∗
∗∗ 2
∗∗
HF (iB (T ∗∗ ), πB
(T ∗∗ ), q(T ∗∗ )) + 1 = λ(πB
) + DπB
+
2λ
D
From (26), we know there are three possibilities. (i). If − 2λ
< 0, then 0 =
b−αR
a−b
D
D
+
1,
namely
that
D
=
is
a
positive
constant.
(ii)
If
− 2λ
> 1, then
a−b
αR −b
b−αR
a−b
0 = a−b D + 1, namely that D = − a−αR (λ + 1) is also a constant. Note that
a−b
D < −2λ if and only if a − b ≤ 2(αR − b), or if and only if λ < a+b−2α
and
R
D
a − b > 2(αR − b). (iii). If 0 ≤ − 2λ ≤ 1, then
2
2
D
D
b − αR
D
0=λ −
+ 1,
− 2λ −
−
2λ −
2λ
2λ
a−b
2λ
namely that
b − αR
λ−
D=2
a−b
r
b − αR 2
4
λ + 4λ
a−b
a−b
D
∈ (0, 1) if and only if λ ≥ a+b−2α
and a − b >
is a constant. Note that − 2λ
R
2(αR − b).
In term of minimizing the Hamiltonian HF under the above case (i), we
∗∗
have πB
= 0 for all time, which is impossible to obtain iB (T ) = ie ; under the
∗∗
above case (ii), we have πB
= 1 for all time; under the above case (iii), we have
D
∗∗
πB = −2λ for all time. To sum up, we have
∗
a−b
(u , T2 ) if λ ≥ a+b−2α
and a − b > 2(αR − b)
∗∗
R
(πB
, T ∗∗ ) =
(27)
(1, T3 ) otherwise
r
αR −b
a−b
b−αR
a−b
where u∗ =
D
−2λ
iB (T2 ) =
i0
(b+(a−b)u∗ −αR )T2
1−i0 e
i0
+ 1−i
e(b+(a−b)u∗ −αR )T2
0
1
=
+
2
+ λ1 , and T2 and T3 satisfy
= ie ,
respectively. This completes the proof.
iB (T3 ) =
1
i0
(a−αR )T3
1−i0 e
i0
+ 1−i
e(a−αR )T3
0
= ie ,
⊓
⊔
| 3cs.SY
|
Concurrent Pump Scheduling and Storage Level Optimization
Using Meta-Models and Evolutionary Algorithms1
Morad Behandish1 and Zheng Yi Wu2
arXiv:1711.04988v1 [cs.NE] 14 Nov 2017
1
Department of Computer Science and Engineering, University of Connecticut, USA
2
Applied Research Group, Bentley Systems, Incorporated, USA
Abstract
In spite of the growing computational power offered by the commodity hardware, fast pump scheduling of
complex water distribution systems is still a challenge. In this paper, the Artificial Neural Network (ANN)
meta-modeling technique has been employed with a Genetic Algorithm (GA) for simultaneously optimizing
the pump operation and the tank levels at the ends of the cycle. The generalized GA+ANN algorithm has
been tested on a real system in the UK. Comparing to the existing operation, the daily cost is reduced by
about 10−15%, while the number of pump switches are kept below 4 switches-per-day. In addition, tank
levels are optimized ensure a periodic behavior, which results in a predictable and stable performance over
repeated cycles.
Keywords. Pump Scheduling; Meta-Modeling; Graphics Processing Unit; Artificial Neural Network; Genetic Algorithm.
1. Introduction
Drinking water and waste water utilities account for about 3 − 4% of the total energy use in the United
States, and are responsible for more than 45 million tons of greenhouse gas emission annually, as reported
by the US Environmental Protection Agency (EPA). According to the same report, these systems account
for 30 − 40% of total energy consumption of municipal governments, and the energy-related operating costs
are expected to increase as much as 20% in the next fifteen years due to population growth and tightening
drinking water regulations (EPA (2012)). These facts pronounce the increasing need for the water industry
to improve water management strategies.
The problem of operation optimization is most frequently addressed by pump scheduling, i.e., predicting
a set of either implicit control rules or explicit time-based specifications on when to turn pumps on and off,
such that the supply or disposal service requirements are met with minimal energy cost. Pump scheduling
has been extensively researched over the past few decades, using a variety of optimization techniques, the
most popular of which has been Genetic Algorithms (GAs), including the studies by Mackle et al. (1995);
Beckwith and Wong (1995); Engelbrecht and Haarhoff (1996); Nitivattananon et al. (1996); De Schaetzen
et al. (1998); Wu et al. (2001); Kelner and Léonard (2003) and Van Zyl et al. (2004). The earliest studies
were mostly conducted based on a single objective function (i.e., operation energy or cost). Multi-objective
evolutionary optimization algorithms have also been used by Savic et al. (1997); Sotelo et al. (2002); Barán
et al. (2005); López-Ibáñez et al. (2005), and Wang et al. (2009). In either case, pump scheduling with the
1 This article was presented in the 12th International Conference on Computing and Control for the Water Industry
(CCWI’2013) and published on 04/23/2014 in Procedia Engineering. For citation, please use:
Behandish, Morad and Wu, Zheng Yi, 2014. “Concurrent Pump Scheduling and Storage Level Optimization Using
Meta-Models and Evolutionary Algorithms.” Procedia Engineering, 70, pp.103–112.
Preprint submitted to Procedia Engineering
Technical Report No. BENT-TR-14-04, April 23, 2014
Figure 1: Evolutionary optimization of water distribution systems: (a) the multi-ANN meta-model used to predict the hydraulic
response; (b) the Genetic Algorithm (GA) used to optimize the pump operation schedule; (c) and (d) reproduction operators
of the GA.
direct application of hydraulic solvers is computationally intensive when applied to models of large utilities.
To overcome this difficulty, parallel computing technology has been utilized by von Lücken et al. (2004); Wu
and Zhu (2009) and Wu and Behandish (2012a,b) to speed up the optimization. The speed-up factors scale
well by increasing the number of physical computing cores; however, expensive multiprocessor systems are
required, especially for real-time applications demanding continuous updates of the pump control policies.
In addition to the pump schedule, there are other hydraulic parameters that are indirectly related to
the energy consumption. One such set of parameters are the operation ranges of storage tank levels, which
can introduce an independent set of additional decision variables alongside pump control settings. In real
systems, a tank is not necessarily designed to result in optimal energy consumption when its entire capacity
is utilized, nor is necessarily placed at optimal locations. As a result, some tanks might have more impacts
on energy saving than the others if they are filled when electricity is inexpensive and drained during the
peak-demand periods.
Different techniques have been used for hydraulic system modeling, ranging from empirical models to
simplified hydraulic models and state-of-the-art hydraulic simulation packages, as reviewed by Rao and
Alvarruiz (2007). An alternative effective solution to reduce the computation time is offered by machine
learning. Artificial Neural Networks (ANNs) have been extensively utilized for predicting the hydraulic
state of water distribution and disposal systems in the recent years by Broad et al. (2005) and Tabach
et al. (2007), and for the specific problem of real-time pump operation control by Jamieson et al. (2007);
Rao and Alvarruiz (2007), and Rao and Salomons (2007). These techniques were applied to real systems
in Salomons et al. (2007); Martinez et al. (2007); Shamir and Salomons (2008) and Behandish and Wu
(2012). Furthermore, the ANN training process lends itself well to high-throughput parallel processing on
the Graphics Processing Unit (GPU), as demonstrated by Wu and Eftekharian (2011) and Behandish and
Wu (2012). The multi-ANN meta-modeling was later improved and generalized by Behandish and Wu
(2013) to encompass a wider spectrum of systems with diverse state variable combinations to be predicted
with significantly improved accuracy and robustness.
In this article, the generalized multi-ANN meta-modeling technique developed by the authors is utilized
in combination with a modified Genetic Algorithm (GA). The modified GA is employed to search for the
decision space made of both discrete variables (e.g., pump/valve statuses) and continuous variables (e.g.,
tank levels). The results on a case study system are provided and compared with those of the earlier studies
in terms of energy saving, number of required pump switches, and periodic tank water level variations.
2
Figure 2: Multi-ANN meta-model construction: (a) sensitivity analysis results; (b) sample sub-ANN structures to predict the
hydraulic response.
2. Generalized Meta-Modeling
Fig. 1 (a) is a schematic illustration of the Extended Period Simulation (EPS) in a succession of time,
with each snapshot being replicated by a set of independent neural networks (Behandish and Wu (2013)).
The initial conditions such as the storage tank levels are specified at time t0 , while control variables including
pump and valve statuses (on or off), pump speed settings (when applicable), etc., are explicitly controlled
for each time interval. In addition, time patterns are pre-specified for nodal demands, electricity tariffs,
reservoir heads, etc. Each snapshot of the simulation relates the state variables at time ti to those at
ti+1 = ti + δt, where typically δt = 1 hour. Before setting up such a network, a sensitivity analysis was
carried out to find out which state variables to select as the inputs of each sub-ANN to predict a particular
output (Behandish and Wu (2013)). Fig. 2 (a) shows the results of the sensitivity analysis for a Demand
Monitoring Zone (DMZ) system in the UK. The corresponding sub-ANN constructions are depicted in Fig.
2 (b), where the inputs with the highest impacts on each output are maintained. Therefore, 12 sub-ANN
structures are formed based on these input/output couplings, together with an additional sub-ANN that is
used to predict the energy consumption rates from the pump statuses. The sub-ANNs are trained one at
a time or in parallel on the GPU with NVIDIAs Compute Unified Device Architecture (NVIDIA (2012)),
using large datasets obtained with the hydraulic model. The reader is referred to Behandish and Wu (2013)
for further details on the generalized meta-modeling.
3. Optimization Model
The trained and verified multi-ANN meta-model was integrated with an evolutionary optimization algorithm to evaluate the fitness of a trial solution for the system operation control, and to search for a
near-optimal solution. Two distinct sets of decision variables are considered in general:
1. Control variables: The control variables include the statuses and settings of a control element (i.e.,
pump or valve) at every control interval. The statuses are defined as the fraction of the time interval δt
during which the pump or valve has been kept open (e.g., 0 for closed, 1 for open) and the settings can
represent pump speeds, valve flow/pressure settings, etc., as detailed by Behandish and Wu (2013).
2. Initial Conditions: The initial storage tank levels, which are typically constrained to be recovered at
the end of the operation cycle, are also included in the decision variable set. The tank levels at t = t0
are decided by the optimizer, and those at t > t0 are predicted successively by the meta-model.
3
Using the convention of Behandish and Wu (2013), all state variables are normalized with the following
linear equation, to make them dimensionless and with similar orders of magnitude:
s̄(t) =
si (t) − si,min
,
si,max − si,min
i = 1, 2, · · · , n.
(1)
The first set of decision variables can be represented in two matrices, one for statuses and the other for
settings. The former is a (nP +nV )×m matrix [Ai,k ] made of nP pump statuses 0 ≤ AP,i (tk ) ≤ 1 (1 ≤ i ≤ nP ),
and nV valve statuses 0 ≤ AV,i ≤ 1(tk ) (1 ≤ i ≤ nV ) over m control intervals starting at tk = t0 + kδt (0 ≤
k ≤ m − 1). Similarly, the latter is a (n0P + n0V ) × m matrix [S̄i,k ] that represents n0P normalized pump
settings 0 ≤ S̄P,i (tk ) ≤ 1 (1 ≤ i ≤ n0P ), and n0V normalized valve settings 0 ≤ S̄V,i (tk ) ≤ 1 (1 ≤ i ≤ n0V )
defined over the same m control intervals. The second set of decision variables, on the other hand, can be
represented by a vector [P̄j,0 ] made of nT normalized tank levels P̄T,j (t0 ) at the beginning of the cycle.
As schematically depicted in Fig. 1 (b), the Genetic Algorithm (GA) generates the trial solution set SGA
of the aforementioned n = (nP + nV + n0P + n0V ) × m + nT variables. The trial solution together with other
parameters Spar , that may include the demand patterns, the reservoir head patterns, the pre-specified pump
and valve settings, etc., all in normalized forms, are passed to the multi-ANN meta-model. For each GA
trial, the extended period simulation is replicated by the successive calls to the trained ANNs. The ANN
outputs may include the pump energy rates, the tank levels, and possibly other dependent state variables
such as the junction pressures and the pipe flow rates. Part of the output is passed to the fitness computing
routine, where the object function and penalty function are evaluated.
The objective function is defined as the total pumping energy cost, normalized in the following form:
F̄ (SGA ; Spar ) =
m−1
1 X
C̄(tk )Ētot (tk ),
m
(2)
k=0
where C̄(t) = C(t)/Cmax is the normalized electricity tariff, Ētot (t) = Etot (t)/Emax is the normalized energy
consumption rate aggregated over all pumps, ∆t = topt − t0 is the operation time, δt = ∆t/m is the control
interval, and tk = t0 + kδt is the discrete time step 0(≤ k ≤ m − 1).
The water distribution service requirements are quite diverse among different systems. In this article,
three different classes of generalized constraints are defined and implemented as follows:
1. Time-Based Constraints: The time-based constraints specify all of the requirements on the selected
hydraulic state responses over the control horizon, ranging from the tank levels and the junction
pressures to the pipe flow rates, etc. These constraints are typically expressed as si,min (t) ≤ si (t) ≤
si,max (t), where si (t) is any state variable that is dependent on the decision variables, and si,min (t) and
si,max (t) are the prescribed lower-bound and upper-bound. The most common time-based constraints
are:
(a) Lower-bounds on tank levels PT,j (t) ≥ PT,min , e.g., PT,min = 30% of capacity, maintained for
emergency.
(b) Upper-bounds on tank levels PT,j (t) ≤ PT,max , e.g., PT,max = 95% of capacity, to avoid water
overtopping.
(c) Requirements on storage at early morning hours PT,j (t) ≥ PT,AM (t), e.g., PT,AM (t) = 80% of
capacity at early morning t = 6:00 to 8:00 AM, and PT,AM (t) = −∞ at all other times (i.e., no
lower-bound).
(d) Lower-bounds on node pressures PJ,j (t) ≥ PJ,min , e.g., PJ,min = minimum pressure required at
consumer end.
(e) Upper-bounds on node pressures PJ,j (t) ≤ PJ,max , e.g., PJ,max = maximum pressure required to
avoid leakage, or the maximum pressure that the junctions can endure, whichever is smaller.
(f) Lower-bounds on pipe flow rates QI,j (t) ≥ QI,min , e.g., QI,min = minimum flows required to avoid
stagnation and maintain water quality.
(g) Upper-bounds on pipe flow rates QI,j (t) ≤ QI,max , e.g., QI,max = maximum flows that pipes can
endure.
4
2. Periodicity Constraints: Regardless of the time-variant constraints on the values, some state variables
are constrained to return to their initial values at the end of the cycle as an operational requirement,
so that the operation would repeat itself periodically under similar conditions at the subsequent cycles.
For instance, a tank level at the end of the cycle can be constrained to be in a tolerance range of its
initial water level, e.g., |PT,j (topt ) − PT,j (topt )| ≤ ∆PT,j .
3. Switch Constraints: A pump scheduling scenario is of practical significance only if the number of pump
switches per day (i.e., the number of times that each pump is turned on and off) is restricted, e.g., to
4 or 6 switches per day for each pump. This is because numerous pump switches are detrimental to
the pump’s life-cycle, resulting in prohibitively large maintenance and replacement costs.
The violation of each inequality constraint written in the standard form of g(·) ≤ 0, is quantified with
hg(·)i := min{0, g(·)}. For a trial solution SGA , the violation measured over the cycle ∆t = topt − t0 is
formulated as:
Ḡ(SGA ; Spar ) =
+
m−1
n
X
c1,i X
i=1
n
X
m
c2,j
[hs̄i (tk ) − s̄max,i (tk )i + hs̄min,i (tk ) − s̄i (tk )i] ,
nP
D
E X
D
E
topt
t
∆s̄j |0 − ∆s̄max,j +
c3,p ηp |0opt − ηmax,p .
(4)
p=1
j=1
t
(3)
k=0
t
where |∆s̄j |0opt := |s̄j (topt ) − s̄j (t0 )|, and ηp |0opt is the actual number of pump switches per cycle, e.g.,
24 hours. The first term on the right is the sum of violations of the time-base constraints, integrated for
each state variable over the time interval ∆t = topt − t0 . The second term measures the violation of the
periodicity requirements with the normalized tolerances ∆s̄max,j , and the last term quantifies the violation
of pump switch constraints with the maximum allowable number of switches ηmax,p . The state variables
s̄i (t) in the first two terms can be the normalized tank levels s̄i (t) := P̄T,i (t) (1 ≤ i ≤ nT ), the normalized
junction pressures s̄i (t) := P̄J,i (t) (1 ≤ i ≤ nJ ), the normalized pipe flow rates s̄i (t) := Q̄I,i (t) (1 ≤ i ≤ nI ),
etc. The weight factors c1,i , c2,j , and c3,p are set to 1 by default for constrained variables, and to 0 for the
unconstrained variables.
The objective function and the violation function formulated in Eqs. (2) and (4) are combined using
additive penalty method into the following penalized objective function:
F̄ ∗ (SGA ; Spar ) = F̄ (SGA ; Spar ) + P × Ḡ(SGA ; Spar ),
(5)
where P is the penalty factor, typically selected in the order of P ∼ 102 − 103 depending on how strictly
the constraints are being enforced. The penalized objective function F ∗ (SGA ; Spar ) is used as a measure of
fitness of the trial decision set SGA generated by the GA. The lower the value of this function is, the fitter
the trial scenario is, hence more likely it is to survive or pass its properties to the next generations of the
evolutionary optimization.
4. Optimization Algorithm
The optimization problem is solved by combining the classic binary operators with those of a modified
Genetic Algorithm (GA) (Mundo and Yan (2007)). The hybrid search algorithm is schematically illustrated
in Fig. 1 (b). The algorithm iterates over ngen generations, and at each generation g where 0 ≤ g ≤ ngen − 1,
a population of npop individuals (i.e., chromosomes) are maintained. Each individual is represented by a set
of normalized decision variables SGA,g composed of the control variables (i.e, the matrices of pump/valve
statuses and normalized settings) and initial conditions (i.e., the vectors of normalized initial storage levels)
as explained in Section 3. The normalized decision variables SGA,0 of the initial population are randomly
assigned with values in [0, 1], corresponding to random values in the physical domain in each state variable’s
min/max range. The subsequent generations SGA,g (1 ≤ g ≤ ngen − 1) are descended from their parents in
the previous generation SGA,g−1 , through a series of evolutionary operations. The modified GA utilizes a
5
combination of evolutionary operators that are designed for binary-coded chromosomes (e.g., for pump/valve
statuses that are constrained to a small number of switches) as well as real-valued variables (e.g., tank levels
or pump speed settings, if applicable) without a need to use binary-coding for the latter.
At the beginning of each iteration, the fitness of the individuals are computed using the multi-ANN
meta-model and sorted in descending order of the penalized objective function F ∗ (SGA ; Spar ). The algorithm
follows an elitist strategy, hence a constant fraction felit ∼ 1 − 5% of the population consisting of the fittest
individuals are directly sent to the next generation. Another fraction frand ∼ 5 − 15% of the population
is randomized at every generation, to enhance the exploration and avoid local minima. The rest of the
population consisting of 1 − (felit + frand ) ∼ 80 − 95% are generated from reproduction (i.e., breeding)
operations. The selection of the parents for reproduction can be based on different probability distributions.
In this algorithm, the method of normalized geometric ranking selection is used (Mundo and Yan (2007)),
in which the probability of an individual to be selected is calculated from its rank in the sorted array based
on fitness:
P0 (1 − P0 )r
, r = 0, 1, · · · , npop − 1,
(6)
Prep (r) =
1 − (1 − P0 )npop
where 0 < P0 < 1 is a constant and is proportional to the probability of selecting the fittest individual, r is
the rank of the individual (r = 0 for the fittest individual, and r = npop − 1 for the least fit individual of
that generation).
Once the parents are selected, different reproduction operators are employed to produce one offspring at
a time:
1. Linear Combination: With a probability of 0 ≤ Pcom ≤ 1, the offspring genes are generated using a
linear combination of the parents’ genes:
s̄off,i = s̄A,i + Rcom (s̄B,i − s̄A,i ),
Rcom ∈ − , 1 + ,
(7)
where the subscripts “off”, A, and B refer to the offspring and the two parents, respectively (Mundo
and Yan (2007)), as exemplified in Fig. 1 (c) with Rcom = 0.6. The combination factor Rcom is a
random number in [−, 1 + ) where ∼ 0 − 0.2 is the overshoot. It is evident that this reproduction
scheme applies to the real-valued decision variables only. Therefore, for the pump/valve statuses
restricted to binary values, the linear combination is replaced with a simple random selection, i.e., for
every gene s̄off,i := s̄A,i , or s̄off,i := s̄B,i , each with 50% probability.
2. Single Split Cross-Over: With a probability of 0 ≤ Pcrs ≤ 1, the offspring genes are generated from a
single point cross-over operation on the parents. For the pump/valve statuses and settings arranged
into matrices, the splitting can occur along the rows or columns, each with a 50% probability. For
the initial conditions arranged into a vector, on the other hand, it is a simple splitting along the
one-dimensional array, as illustrated in Fig. 1 (d).
3. Direct Transfer: With a probability 1 − (Pcom + Pcrs ), one of the two parents are directly transferred
to the next generation, each with 50% probability. This could be replaced with an elitist procedure,
in which the fitter parent is more likely to be transferred.
With the above operations, one offspring is created by a pair of parents. The selection of the parents and
the reproduction have to be repeated 1 − (felit + frand ) times to generate enough number of offsprings for
the next generation.
The produced offsprings are subjected to mutation with a small probability Pmut ∼ 1 − 2% to avoid
local minima and promote global exploration of the search landscape. For the real-valued decision variables,
a uniform random mutation operation is employed, which means a reassignment of one of the normalized
decision variables into a random value in [0, 1], corresponding to a random value in the physical min/max
range, as illustrated in Fig. 1 (c). For binary-coded decision variables, on the other hand, the mutated gene
is changed from 0 to 1 and from 1 to 0, as shown in Fig. 1 (d).
As illustrated in Fig. 1 (b), once the next generation of npop ∼ 100 − 300 new individuals are produced,
the procedure is repeated for enough number of generations ngen = 5, 000, until a near-optimal set of control
variables and initial tank levels is found. In addition to the selection, reproduction, and mutation, every
6
several generations nres ∼ 100 − 500, the entire population except the elite fraction are completely reset to
random decision variables to simulate several optimization sessions following one another, always keeping
the best results obtained so far.
5. Results & Discussion
The generalized GA+ANN technique was applied to a Demand Monitoring Zone (DMZ) in the UK
(Wu et al. (2009)). The hydraulic model for this system is composed of 3, 537 junctions, 3, 273 pipes, 5
reservoirs, 12 storage tanks, 19 constant-speed pumps, and 420 valves. The optimization was carried out
over ∆t = 24 hours with time steps δt = 1 hour. Two different optimization scenarios are reported here,
one with pump scheduling of 9 active pumps (the same pumps utilized in the existing operation), which
involves (9 × 24) = 216 decision variables, and the other with simultaneous pump scheduling and storage
optimization of all 12 tanks, which gives rise to 216 + 12 = 238 decision variables. For this particular case
study, the three categories of constraints defined in Section 3 are specified, including: (1) for 10 out of 12
tank levels, PT,j ≥ 30% of capacities for emergency, and PT,j ≤ 95% of capacities to avoid overtopping; (2)
the periodicities of the same tank levels are enforced as |PT,j (24) − PT,j (0)| ≤ 10cm, which is less than 2.5%
of the smallest tank capacity. (3) For every active pump, a maximum of 4 switches per 24 hours are allowed.
No explicit constraints on junction pressures or pipe flow rates are specified. The GA parameters are set
as: P = 1, 000, ngen = 5, 000, nres = 100, npop = 300, felit = 0.01, frand = 0.10, P0 = 0.05, Pcom = 0.40,
Pcrs = 0.50, and Pmut = 0.01.
The GA+ANN solutions are compared with the existing operation and the pump scheduling study by
Wu et al. (2009) on the same system. The existing operation uses simple controls that define when the
pumps are turned on and off based on tank levels; for instance, the pump “PILWTH” is turned on if the
tank “BUTa2” level falls below 5.30 meters, and turned off if the level starts to exceed 5.73 meters. The
thresholds are based on experience and does not necessarily ensure cost-effective operation. The study by
Wu et al. (2009), on the other hand, used rule-based controls, but additional rules were specified with greater
thresholds on the minimum storage at night when the electricity is inexpensive; e.g., at clock-times between
10:00 PM and 8:00 AM, the pump “PILWTH” is turned on if tank “BUTa2” level falls below 5.73 meters,
and turned off otherwise. Both simple and rule-based controls secure the controlled tank levels between the
limits, but they rely on the known relationships of pumps with tanks, and do not guarantee any bound on
the number of pump switches that is required for this purpose. The cost can be significantly reduced by
optimizing the rules in the latter method, but as shown in Table 1, this comes at the expense of numerous
pump switches enforced by the control rules. For instance, the utilization of pump “PILWTH” is decreased
by about 60% with the optimized rules, saving around ₤190 per day. However, as Fig. 3 shows, the reduced
pump utilization is made possible with numerous pump switches at high frequencies to satisfy the rules,
which not desirable in practice.
Table 1 compares the pump utilization characteristics of the 4 different operations, obtained with
EPANET using a time step of 5 minutes for every one of the 4 scenarios. It is observed that the rulebased control strategy saves about 30% of the daily cost, but the number of pump switches is prohibitively
large (see Fig. 3). The near-optimal GA+ANN solution, on the other hand, can results in 10 − 15% cost
saving with a maximum of 4 pump switches per day. Furthermore, both of the existing simple and rulebased controls result in large differences between the initial and the final tank levels, making it difficult to
anticipate the operating cost and performance of subsequent cycles. The GA+ANN solution, on the other
hand, guarantees repeatability of the operation by bringing the tank levels back to the initial conditions at
the end of each cycle. It is also worthwhile noting that the existing simple and rule-based controls are on
7 out of 12 tanks, hence emptied or overtopped levels are observed for the other 5 tanks; while GA+ANN
studies constrain 10 tanks to be periodic.
Fig. 4 (a) shows the difference between the initial and the final tank levels for the two GA+ANN scenarios
where the shaded area represents the feasible space |PT,j (24)−T, j(0)| ≤ 10cm. The energy and cost of both
scenarios are plotted in panel (b) and compared with those of the current operation. The solution of pump
scheduling and storage optimization is obtained with F̄ = 0.4199 and Ḡ = 0.0000 (no violation), resulting
in saving of around ₤130 per day. However, the result of sole pump scheduling is obtained with F̄ = 0.4687
7
Table 1: Comparison of pump utilization, energy cost, number of pump switches (#PS), and tank level periodicities for 4
operation scenarios.
Case Studies
Pump ID
Simple Controls
Rule-Based Controls
Near-Optimal GA+ANN
Near-Optimal GA+ANN
(Existing Operation)
(Based on Wu et al. (2009))
(Pump Scheduling Only)
(Pump Sch. + Storage Opt.)
Util.
Cost
100 %
100 %
0.00%
36.0%
36.0%
13.6%
52.7%
93.6%
0.00%
₤226.3
₤350.4
₤00.00
₤24.39
₤24.39
₤04.14
₤46.79
₤276.6
₤00.00
#PS
0
0
0
2
2
2
3
2
0
Total
N/A
₤953.0
N/A
Cost Savings
No. Switches
Periodicities
N/A
ηmax < 4 per day (X)
|∆max | 10.0 cm (5)
X2420052
X2420014
X2420075
X2410361
X2419963
X241998C
X2450024
PILWTH
NEWMRKT
Util.
Cost
100 %
41.7%
21.5%
41.7%
41.7%
26.8%
23.0%
36.0%
41.7%
₤179.1
₤127.2
₤51.48
₤22.54
₤22.54
₤04.23
₤15.21
₤84.91
₤177.6
#PS
0
28
26
3
3
48
56
116
3
N/A
₤684.8
N/A
₤268.2 (28.15%)
ηmax 4 per day (5)
|∆max | 10.0 cm (5)
Util.
Cost
79.2%
41.7%
37.5%
58.3%
41.7%
25.0%
37.5%
79.2%
12.5%
₤178.1
₤127.2
₤192.3
₤39.41
₤27.34
₤05.69
₤34.80
₤236.3
₤59.48
#PS
4
2
4
4
2
4
4
4
2
N/A
₤900.7
N/A
₤52.36 (5.49%)
ηmax = 4 per day (X)
|∆max | = 15.3 cm (5)
Util.
Cost
83.3%
37.5%
41.7%
12.5%
75.0%
25.0%
37.5%
75.0%
8.33%
₤183.5
₤122.6
₤173.7
₤06.36
₤47.81
₤05.67
₤27.53
₤211.7
₤32.89
#PS
4
4
4
4
4
2
2
4
2
N/A
₤821.8
N/A
₤131.29 (13.78%)
ηmax = 4 per day (X)
|∆max | = 10.3 cm (X)
Figure 3: Pump flows over 24 hours for 4 operation scenarios. The results are obtained with EPANET using a simulation time
step of 5 minutes.
Figure 4: GA+ANN optimization results: (a) tank level difference between the two ends of cycle; (b) comparison of the
energy consumptions; and (c) comparison of the energy costs. The post-processing results are generated with EPANET using
a simulation time step of 1 hour.
8
Figure 5: Storage level variations over 24 hours for 4 operation scenarios. For GA+ANN results, both meta-model prediction
and hydraulic model post-processing results are shown. The post-processing results are generated with EPANET using a
simulation time step of 1 hour.
and Ḡ = 0.1966, showing some violation of the periodicity constraints depicted in Fig. 4 (a), and less saving
of ₤50 per day. Although this result after 5, 000 GA generations does not necessarily mean that no feasible
solution exists for the latter case, nevertheless it shows that starting from non-optimal initial conditions of
the existing operation introduces difficulties to the GA optimizer to find a feasible solution with significant
cost reduction. This demonstrates the importance of a proper choice of the initial combination of tank levels
that would be able to recover in 24 hours, and also explains some of the difficulties faced in the previous
pump scheduling studies in Wu and Behandish (2012a,b). Fig. 5 compares the tank level variations for the
four operation scenarios. For the GA+ANN solutions, both meta-model predictions (used within GA) and
hydraulic model post-processing results are illustrated. The meta-model accuracy was validated with the
accumulated errors of less than ±5cm per 24 hours, which is a significant improvement over the 20 − 30cm
errors of earlier development (Behandish and Wu (2012) and Wu and Behandish (2012a,b)). It is observed
that when the initial tank levels are confined to the existing operation values, the GA+ANN results do
not deviate much from the existing operation over the first few hours. The solution with concurrent pump
scheduling and storage optimization, on the other hand, utilizes a larger fraction of tank capacities and
recovers all of the constrained tank levels to ±10cm of initial values. The saving of around 835KWH (₤130)
per day corresponds to an annual saving of around 300MWH (₤480,000), which is significant for a DMZ
system of this size.
6. Conclusions
The study has demonstrated that in addition to the pump scheduling policy, the decisions on tank
operation range can play a significant role in water distribution operation cost and storage utilization, and
to guarantee the repeatability of the operation policy with predictable behavior over subsequent cycles. The
GA+ANN algorithm is generalized to represent a wide range of complex systems and their requirements for
pump and valve operation control, carried out concurrently with the optimization of tank operation ranges.
Finally, the set of near-optimal tank levels obtained in this offline optimization with a typical demand profile
can be useful information for the implementation of the real-time pump operation optimization.
9
References
References
EPA, Energy Efficiency for Water and Wastewater Utilities, URL water.epa.gov/infrastructure/sustain/energyefficiency.
cfm, 2012.
G. Mackle, G. A. Savic, G. A. Walters, Application of Genetic Algorithms to Pump SCheduling for Water Supply, in: 1st
International Conference on Genetic Algorithms in Engineering Systems: Innovations and Applications (GALESIA), IET,
400–405, 1995.
S. P. Beckwith, K. P. Wong, A Genetic Algorithm Approach for Electric Pump Scheduling in Water Supply Systems, in: IEEE
International Conference on Evolutionary Computation, vol. 1, IEEE, 21, 1995.
R. J. Engelbrecht, J. Haarhoff, Optimization of Variable-Speed Centrifugal Pump Operation with a Genetic Algorithm, in:
The 3rd International Conference on Computer Methods and Water Resources (CMWR III), 1996.
V. Nitivattananon, E. Sadowski, R. Quimpo, Optimization of Water Supply System Operation, Journal of Water Resources
Planning and Management 122 (5) (1996) 374–384.
W. B. F. De Schaetzen, D. A. Savic, G. A. Walters, A Genetic Algorithm Approach to Pump Scheduling in Water-Supply
Systems, in: Hydroinformatics, vol. 98, 897–900, 1998.
Z. Wu, P. Boulos, W. de Schaetzen, C.-H. Orr, M. Moore, M. Soft, Using Genetic Algorithms for Water Distribution System
Optimization, in: Proceedings of the ASCE Environmental and Water Resources Institute’s (EWRI’s) World Water &
Environmental Resource Congress, 20–24, 2001.
V. Kelner, O. Léonard, Optimal Pump Scheduling for Water Supply Using Genetic Algorithms, in: International Congress on
Evolutionary Methods for Design, Optimization and Control with Applications to Industrial Problems (EUROGEN), 2003.
J. E. Van Zyl, D. A. Savic, G. A. Walters, Operational Optimization of Water Distribution Systems Using a Hybrid Genetic
Algorithm, Journal of Water Resources Planning and Management 130 (2) (2004) 160–170.
D. A. Savic, G. A. Walters, M. Schwab, Multiobjective Genetic Algorithms for Pump Scheduling in Water Supply, in: Evolutionary Computing Lecture Notes in Computer Science, vol. 1305, Springer, 227–235, 1997.
A. Sotelo, C. von Lücken, B. Barán, Multiobjective Evolutionary Algorithms in Pump Scheduling Optimisation, in: Proceedings
of the 3rd International Conference on Engineering Computational Technology, Civil-Comp press, 175–176, 2002.
B. Barán, C. von Lücken, A. Sotelo, Multi-Objective Pump Scheduling Optimisation Using Evolutionary Strategies, Advances
in Engineering Software 36 (1) (2005) 39–47.
M. López-Ibáñez, T. Devi Prasad, B. Paechter, Multi-Objective Optimisation of the Pump Scheduling Problem Using SPEA2,
in: IEEE Congress on Evolutionary Computation, vol. 1, IEEE, 435–442, 2005.
J.-Y. Wang, T.-P. Chang, J.-S. Chen, An Enhanced Genetic Algorithm for Bi-Objective Pump Scheduling in Water Supply,
Expert Systems with Applications 36 (7) (2009) 10249–10258.
C. von Lücken, B. Barán, A. Sotelo, Pump Scheduling Optimization Using Asynchronous Parallel Evolutionary Algorithms,
CLEI Electronic Journal 7 (2).
Z. Y. Wu, Q. Zhu, Scalable Parallel Computing Framework for Pump Scheduling Optimization, in: World Water and Environmental Resource Congress, 2009.
Z. Y. Wu, M. Behandish, Comparing Methods of Parallel Genetic Optimization for Pump Scheduling Using Hydraulic Model
and GPU-Based ANN Meta-Model, in: 14th Water Distribution Systems Analysis Conference (WDSA), Engineers Australia,
1088, 2012a.
Z. Y. Wu, M. Behandish, Real-Time Pump Scheduling Using Genetic Algorithm and Artificial Neural Network Based on
Graphics Processing Unit, in: 14th Water Distribution Systems Analysis Conference (WDSA), Engineers Australia, 1088,
2012b.
Z. Rao, F. Alvarruiz, Use of an Artificial Neural Network to Capture the Domain Knowledge of a Conventional Hydraulic
Simulation Model, Journal of Hydroinformatics 9 (1) (2007) 15–24.
D. R. Broad, G. C. Dandy, H. R. Maier, Water Distribution System Optimization Using Metamodels, Journal of Water
Resources Planning and Management 131 (3) (2005) 172–180.
E. E. Tabach, L. Lancelot, I. Shahrour, Y. Najjar, Use of Artificial Neural Network Simulation Metamodelling to Assess
Groundwater Contamination in a Road Project, Mathematical and Computer Modelling 45 (7) (2007) 766–776.
D. Jamieson, U. Shamir, F. Martinez, M. Franchini, Conceptual design of a Generic, Real-time, Near-Optimal Control System
for Water-Distribution Networks, Journal of Hydroinformatics 9 (1) (2007) 3–14.
Z. Rao, E. Salomons, Development of a Real-Time, Near-Optimal Control Process for Water-Distribution Networks, Journal
of Hydroinformatics 9 (1) (2007) 25–37.
E. Salomons, A. Goryashko, U. Shamir, Z. Rao, S. Alvisi, Optimizing the Operation of the Haifa-A Water-DistributionNetwork,
Journal of Hydroinformatics 9 (1) (2007) 51–64.
F. Martinez, V. Hernandez, J. Alonso, Z. Rao, S. Alvisi, Optimizing the Operation of the Valencia Water-Distribution Network,
Journal of Hydroinformatics 9 (1) (2007) 65–78.
U. Shamir, E. Salomons, Optimal Real-Time Operation of Urban Water Distribution Systems Using Reduced Models, Journal
of Water Resources Planning and Management 134 (2) (2008) 181–185.
M. Behandish, Z. Y. Wu, GPU-based Artificial Neural Network Configuration and Training for Water Distribution System
Analysis, in: ASCE Annual World Environmental and Water Resources Congress, Albuquerque, NM, U.S.A., 2012.
Z. Y. Wu, A. A. Eftekharian, Parallel Artificial Neural Network Using CUDA-Enabled GPU for Extracting Hydraulic Domain
Knowledge of Large Water Distribution Systems, in: ASCE Annual World Environmental and Water Resources Congress,
2011.
10
M. Behandish, Z. Y. Wu, Generalized GPU-based Artificial Neural Network Surrogate Model for Extended Period Hydraulic
Simulation, in: ASCE Annual World Environmental and Water Resources Congress, 2013.
NVIDIA, CUDA C Programming Guide, URL docs.nvidia.com/cuda/pdf/CUDA_C_Programming_Guide.pdf, 2012.
D. Mundo, H.-S. Yan, Kinematic Optimization of Ball-Screw Transmission Mechanisms, Mechanism and Machine Theory 42 (1)
(2007) 34–47.
Z. Y. Wu, K. Woodward, T. Allen, Pump scheduling study on a demand monitoring zone, in: International Conference on
Computing and Control for Water Industry (CCWI), 2009.
11
| 9cs.NE
|
On sequential hypotheses testing via convex optimization
Anatoli Juditsky
∗
Arkadi Nemirovski
†
April 12, 2018
arXiv:1412.1605v2 [math.ST] 24 Feb 2017
Abstract
We propose a new approach to sequential testing which is an adaptive (on-line) extension
of the (off-line) framework developed in [10]. It relies upon testing of pairs of hypotheses
in the case where each hypothesis states that the vector of parameters underlying the distribution of observations belongs to a convex set. The nearly optimal under appropriate
conditions test is yielded by a solution to an efficiently solvable convex optimization problem. The proposed methodology can be seen as a computationally friendly reformulation of
the classical sequential testing.
1
Introduction
Let ωi ∈ Ω, i = 1, 2, ... be a sequence of independent and identically distributed (iid) observations
with a common density w.r.t. a given measure P on the observation space Ω, known to belong
to a given parametric family {pµ (·) : µ ∈ M}. Assume that we are given a (finite) number I of
subsets X i , 1 ≤ i ≤ I, of the parameter space M; these sets define hypotheses Hi on the density
p(·) underlying our observation, Hj stating that the parameter of this density belongs to Hj :
p(·) = pµ (·) with µ ∈ X i . We are interested in the problem of multiple testing of composite
hypotheses: our goal is, given observations ωi to decide on the hypotheses H1 , ..., HI . When
the size K of the observation ω K = (ω1 , ..., ωK ) is fixed, the performance of a test T K (·) – a
measurable mapping of ΩK into {1, ..., I} – can be quantified by the maximal error probability
of rejecting the true hypothesis:
Risk(T K ) = max sup Probµ,K {ω K : T K (ω K ) 6= i}
1≤i≤I µ∈X i
(1)
where Probµ,K is the probability w.r.t. the distribution of the observation ω K corresponding
to the “true” parameter µ ∈ M. Then the problem of optimal testing can be approached
through minimization of the risk over the class of tests. Yet, in many practical applications,
especially in those where observations come at a price, such approach may be too conservative.
The sequential approach to testing, introduced in the pioneering papers of Barnard [2] and
Wald [19, 20] is now a reach area of statistical theory and offers many strong results (see also
∗
LJK, Université Grenoble Alpes, B.P. 53, 38041 Grenoble Cedex 9, France, [email protected]
Georgia Institute of Technology, Atlanta, Georgia 30332, USA, [email protected]
Research of the first author was supported by the CNRS-Mastodons project GARGANTUA, and the LabEx
PERSYVAL-Lab (ANR-11-LABX-0025). Research of the second author was supported by NSF grants CMMI1232623, CMMI-1262063, CCF-1415498.
†
1
[8, 9, 1] for references or [13] for a recent review). For the problem of interest this approach
can be summarized as follows: given an upper bound of the risk the problem is reexamined
sequentially, when new observations are available. The test terminates when either a decision
with the risk ≤ is possible, or when maximal allowed observation count K is reached (in which
case the test produces no decision). Usually, the sequential test is based on the Generalized
Likelihood or Mixture Likelihood statistics, and its performance is evaluated by its asymptotical,
as → +0 and K → ∞. The approach we pursuit in this paper is of completely different spirit.
It originates from [12] and was applied to the testing problem in [10] in the case where the size
of the observation sample is fixed. The main “building block” of this approach is a construction,
based on Convex Programming (and thus computationally efficient) allowing, under appropriate
assumptions, to build a provably nearly optimal test for deciding, given observation ω K , between
a pair of composite hypotheses of the sort H1 : µ ∈ X and H2 : µ ∈ Y where X and Y are convex
compact subsets of M. This approach is applicable in several important situations, namely
where (a) pµ is Gaussian density on Rn with expectation µ and fixed covariance matrix, (b) pµ
is the distribution of the Poisson vector in Rm with independent components with parameters
µi , and (c) is a distribution of a discrete random variable taking values in {1, ..., m} with µ ∈ R+
being the vector of probabilities of the corresponding distribution. As a compensation for rather
restrictive assumptions on the families of densities pµ , the approach in question is extremely
permissive as far the structure of the sets X and Y is concerned: what we require (apart
from compactness and convexity) is the efficient tractability of X and Y . While the analytical
expressions for the test characteristics are not available in the proposed approach, rather detailed
information about performance guarantees of the resulting decisions can be obtained by efficient
situation-oriented computation. In [10] we introduced a calculus of pairwise tests to design
nearly optimal testing procedures in the situation where the sets X i corresponding to different
hypotheses H i are unions of (not too large number of) convex sets with the risk of the test
defined as in (1) (in the minimax setting).1 What follows can be seen as an adaptive version
of the testing procedure from [10] – when the “true distribution” of observations corresponds
to the value µ of the parameter which is “deeply inside” of some X i , the correct decision (Hi is
accepted) can be taken much faster (using a smaller observation sample) than in the case of µ
close to some “wrong” H i ’s. On the other hand, it is nothing but a “computationally friendly”
application of the classical Sequential Analysis methodology to the problem in question.
The paper is organized as follows. In section 2 we give a summary of the results of [10] on
hypothesis testing in good observation schemes and discuss the construction of quasi-optimal
“off-line” test. Then in section 3 we introduce the sequential problem setting and construct a
generic test aggregation procedure which allows to reduce multiple testing to pairwise testing.
Finally, in section 4 we consider in detail a particular implementation of the proposed approach
and present some very preliminary simulation results. The proofs missing in the main body of
the paper are postponed to the appendices.
1
It should be mentioned that what we call below “simple tests” were used to test composite hypotheses
represented by convex sets in the white noise model (a) in [6, 7, 11] and in the distribution model (c) in [14, 15,
3, 4, 5, 16].
2
2
Preliminaries
What follows is the summary of the approach of [12] as applied to hypotheses testing; for detailed
presentation of the constructions and results of this section, same as for the related proofs, see
[10].
2.1
Good observation schemes
We start with introducing good observation schemes, those with which we intend to work. Recall
that we are interested to make inferences from a random observation ω taking values in a given
observation space Ω and obeying probability density w.r.t. a given measure P on Ω; this density
is known to belong to a given parametric family {pµ (·) : µ ∈ M} of probability densities,
taken w.r.t. P , on Ω. We intend to work under the following assumptions on our “observation
environment:”
1. M ⊂ Rm is a convex set which coincides with its relative interior;
2. Ω is a Polish (i.e., separable complete metric) space equipped with a Borel σadditive σ-finite measure P , supp(P ) = Ω, and distributions Pµ ∈ P possess
densities pµ (ω) w.r.t. P . We assume that
• pµ (ω) is continuous in µ ∈ M, ω ∈ Ω and is positive;
• the densities pµ (·) are “locally uniformly summable:” for every
compact set
R M
M
M ⊂ M, there exists a Borel function p (·) on Ω such that Ω p (ω)P (dω) <
∞ and pµ (ω) ≤ pM (ω) for all µ ∈ M , ω ∈ Ω;
3. We are given a finite-dimensional linear space F of continuous functions on Ω
containing constants such that ln(pµ (·)/pν (·)) ∈ F whenever µ, ν ∈ M.
Note that the latter assumption implies that distributions Pµ , µ ∈ M, belong
to an exponential family.
R
4. For every φ ∈ F, the function Fφ (µ) = ln Ω exp{φ(ω)}pµ (ω)P (dω) is well
defined and concave in µ ∈ M.
In the just described situation, where assumptions 1-4 hold, we refer to the collection O =
((Ω, P ), {pµ (·) : µ ∈ M}, F) as good observation scheme (o.s.).
2.1.1
Basic examples
Basic examples of good o.s.’s are as follows:
Gaussian o.s. Here Ω = Rm , P is the Lebesque measure on Ω, M = Rm and pµ (ω) =
N (µ, Im ) is the density of Gaussian random vector with the unit covariance matrix2 and expectation µ. The space F is comprised of all affine functions on Ω = Rm , and because
Z
aT ω+b
ln
e
pµ (ω)dω = b + aT µ + 21 aT a,
Rm
Gaussian o.s. is good.
2
of course, we could replace the unit covariance with any other positive definite covariance matrix, common
for all distributions from the family in question
3
Poisson o.s. Here Ω = Zm
+ is the discrete set of m-dimensional vectors with nonnegative
integer entries, P is the counting measure on Ω, M = Rm
++ is the set of positive m-dimensional
vectors, and pµ (·), µ = [µ1 ; ...; µm ] > 0, is the probability distribution of a random vector
ω = [ω1 ; ...; ωm ], where ωi is Poisson random variable with parameter µi , and with independent
of each other ω1 , ..., ωm . The space F is comprised of all affine functions on Ω. Note that
m
X
X
ln
exp(aT ω + b)pµ (ω) =
(eai − 1)µi + b
ω∈Zm
+
i=1
is concave in µ, and thus Poisson o.s. is good.
Discrete o.s. Here Ω = {1, ..., m} is a finite set, P is a counting measure on Ω, and M is
comprised of all non-vanishing probability densities taken w.r.t. P , that is, the parameter space
M is comprised of all m-dimensional vectors µ = [µ1 ; ...; µm ] > 0 with entries summing up to
1, and the random variable ω distributed according to pµ (·) takes value ω ∈ Ω with probability
µω . The space F is comprised of all real-valued functions on Ω = {1, ..., m}. Since for φ ∈ Rm ,
!
!
m
X
X
φ(ω)
φω
ln
e
pµ (ω) = ln
e µω
ω=1
ω∈Ω
is concave in µ ∈ M, the Discrete o.s. is good.
We have seen that Gaussian, Poisson, and Discrete o.s.’s are good. More examples of good
o.s.’s can be obtained by taking direct products.
Direct products of good o.s.’s. Let Ot = ((Ωt , Pt ), {pµt ,t (·) : µt ∈ Mt }, Ft ), 1 ≤ t ≤ K,
be good o.s.’s. We can associate with this collection a new o.s. – their direct product, which,
informally, describes the situation where our observation is ω K = (ω1 , ..., ωK ), with ωt drawn,
independently across t, from the o.s.’s Ot . Formally, the direct product is the o.s.
OK
=
(ΩK , P K ) = (
K
Q
Ωt , P K = P1 × .. × PK ),
t=1
{pµK (ω1 , ..., ωK ) = pµ1 ,1 (ω1 )...pµK ,K (ωK ) : µK = [µ1 ; ...; µK ] ∈ MK = M1 × ... × MK },
PK
F K = {f (ω1 , ..., ωK ) = t=1 ft (ωt ) : ft ∈ Ft , 1 ≤ t ≤ K} ,
and this o.s. turns to be good provided all factor Ot are so. Note that with this definition, the
parameter µK underlying the distribution of observation ω K is the collection of parameters µt ,
t = 1, ..., K, underlying distributions of the components ω1 ,...,ωK of ω K .
Now consider the special case of this construction where all factors Ot are identical to each
other:
Ot = ((Ω, P ), {pµ (·) : µ ∈ M}, F), 1 ≤ t ≤ K.
In this case we can “shrink” the direct product OK of our K identical o.s.’s to the o.s. O(K) by
passing to observations ω K = (ω1 , ..., ωK ) with ω1 , ..., ωK drawn, independently of each other,
from the same density pµ (·), rather than being drawn from their own densities pµt ,t (·). The
formal description of O(K) is as follows:
• for O(K) , the observation space ΩK = Ω×...×Ω and the reference measure P K = P ×...×P
are exactly the same as for OK ;
4
(K)
• the family of probability densities {p(·)} for O(K) is {pµ (ω1 , ..., ωK ) =
K
Q
pµ (ωt ) : µ ∈
t=1
M};
• the family F for O(K) is F (K) = {f (K) (ω1 , ..., ωK ) = f (ω1 ) + ... + f (ωK ) : f ∈ F}.
Note that O(K) is a good o.s., provided O is so. We shall refer to the just defined O(K) as to
stationary K-repeated observations associated with O.
2.2
Pairwise hypothesis testing
Detectors and their risks. Let Ω be a Polish space and X1 , X2 be two nonempty sets of
Borel probability distributions on Ω. Given a detector – a real-valued Borel function φ on Ω,
we define the risk (φ|X1 , X2 ) of this detector w.r.t. (X1 , X2 ) as the smallest such that
R
ΩRexp{−φ(ω)}p(dω) ≤ ∀p ∈ X1 ,
(2)
Ω exp{φ(ω)}p(dω) ≤ ∀p ∈ X2 ,
The sets X1 , X2 give rise to two hypotheses, H1 , H2 , on the distribution of a random observation
ω ∈ Ω, with Hχ stating that the distribution p(·) of this observation belongs to Xχ , χ = 1, 2,
while a detector φ gives rise to the test Tφ which, given an observation ω, accepts H1 (and rejects
H2 ) when φ(ω) ≥ 0, and rejects H1 (and accepts H2 ) otherwise. We define the risk of a test T
deciding between H1 and H2 as a maximal p(·)-probability of the test rejecting the hypothesis
Hχ , χ = 1, 2, when it is true:
"
#
Risk = max
sup p({ω : T (ω) = −1}), sup p({ω : T (ω) = 1}) ,
p(·)∈X1
p(·)∈X2
where T (ω) = 1 when the test T , as applied to observation ω, accepts H1 , and T (ω) = −1
otherwise. Clearly, the risk of the test Tφ associated with detector φ is ≤ := (φ|X1 , X2 ).
Indeed, denoting by p(·) the distribution of the observation, when H1 is true, the test rejects
this hypothesis when φ(ω) < 0, and p(·)-probability of this event, by the first inequality in (2)
and due to the definition of := (φ|X1 , X2 ), is at most ; when H2 is true, the test rejects H2
and accepts H1 when φ(ω) ≥ 0, and p(·)-probability of the latter event is ≤ due to the second
inequality in (2).
Note that if we have at our disposal a test, say T , which decides between H1 and H2 with
the risk bounded with ¯ ∈ (0, 1/2), we can associate with it the detector
1 − ¯
1
φ̄(ω) = 2 ln
T (ω)
¯
One can easily see that the risk of φ̄(·) satisfies the bounds of (2) with
p
= 2 ¯(1 − ¯).
As we will see in an instant, tests associated with detectors satisfying (2) allow for a simple
calculus – one can “propagate” the tests properties to the case of repeated observations and
multiple testing. Our first goal is to describe a systematic construction of detectors satisfying
(2) in the situation where the underlying o.s. is good.
5
Near-optimal tests. As far as testing pairs of hypotheses is concerned, the main result of
[10] – and the starting point of our developments in this paper – is as follows:
Theorem 1 [10, Theorem 2.1] Let O = ((Ω, P ), {pµ (·) : µ ∈ M}, F) be a good o.s., and let
X1 , X2 be two nonempty convex compact subsets of M. Consider the optimization problem
Z q
ψ(µ, ν) := ln
pµ (ω)pν (ω)P (dω)
(3)
Opt = Opt(X1 , X2 , O) := max
µ∈X1 ,ν∈X2
Ω
The function ψ(µ, ν) : M × M → R is concave and continuous, so that (3) is a convex optimization problem. This problem is solvable, and every optimal solution (µ∗ , ν∗ ) to this problem
gives rise to the detector
1
φ(ω) = ln(pµ∗ (ω)/pν∗ (ω)) : Ω → R
(4)
2
with the following properties:
(i) [risk bound] Denoting by Xχ the set of all probability distributions on Ω with densities,
w.r.t. P , of the form pµ (·) with µ ∈ Xχ , χ = 1, 2, the risk of the detector φ w.r.t. (X1 , X2 ) is
(φ|O, X1 , X2 ) := ? (O, X1 , X2 ) = exp{Opt};
consequently, the risk of the test Tφ induced by the detector φ, when deciding on the hypotheses
H1 , H2 associated with X1 , X2 , as at most exp{Opt}.
(ii) [near-optimality] Assume that for some ∈ (0, 1/4), “in the nature” there exists a test
T (a deterministic Borel function T (ω) of ω ∈ Ω taking values 1 (“H1 accepted, H2 rejected”)
and -1 (“H2 accepted, H1 rejected”) with risk in deciding on H1 , H2 based on an observation ω
at most . Then
√
∗ (X1 , X2 ) ≤ 2 .
We interpret (ii) as a statement of near-optimality of the test T defined in (i) – whenever the
hypotheses H1 , H2 associated with X1 , X2 can be decided upon with small risk, the risk of
our test Tφ also is small. Note that (ii) also suggests that (maximal) degradation of the test
performance when passing from the optimal test to the near-optimal one associated with detector
(4) may be significant in the setting of Theorem 1, when the decision is taken on the basis of
one observation. When repeated observations are available, the sub-optimality of the proposed
tests is expressed by a moderate absolute factor, when measured in terms of the length of the
observation sample necessary to attain the desired testing accuracy.
Corollary 1 [10, Proposition 2.1] Let O = ((Ω, P ), {pµ (·) : µ ∈ M}, F), X1 , X2 be as in
Theorem 1, and K be a positive integer. Assume we have at our disposal stationary K-repeated
observation ω K stemming from O, and let Hχ , χ = 1, 2, be the hypotheses on the density
Q
K stating that µ ∈ X . Then
pµ (ω1 , ..., ωK ) = K
χ
t=1 pµ (ωt ) of ω
(i) The optimal solution (µ∗ , ν∗ ) to the problem (3) associated with O, X1 , X2 , is optimal for
the same problem associated with O(K) , X1 , X2 , and
? (O(K) , X1 , X2 ) = [? (O, X1 , X2 )]K ;
moreover, the detectors φ and φ(K) associated with this optimal solution by Theorem 1 as applied
to (O, X1 , X2 ) and (OK , X1 , X2 ), respectively, are linked by the relation
φ(K) (ω1 , ..., ωK ) =
K
X
t=1
6
φ(ωt ).
(ii) Assume that for some ∈ (0, 1/4) and some positive integer K̄, there exists a test which
decides on the hypotheses H1 , H2 via stationary K̄-repeated observation ω K̄ with risk ≤ .
Setting
2K̄
,
K = Ceil
2 ln(2)
1 − ln(1/)
we ensure that the risk of the test based on detector φ(K) (·) when deciding on the hypotheses Hχ ,
χ = 1, 2, is ≤ as well. Note that K/K̄ → 2 as → +0.
2.3
Testing multiple hypotheses
The completing “building block” from [10] we need is a simple technique for passing from pairwise
tests to tests deciding on N ≥ 2 hypotheses.
The situation we are interested in is as follows. We are given a Polish observation space Ω
along N ≥ 2 nonempty families Xj , 1 ≤ j ≤ N , of Borel probability distributions on Ω, and
have at our disposal pairwise detectors – real-valued Borel functions φij (ω) : Ω → R and risk
bounds ij ∈ (0, 1], 1 ≤ i, j ≤ N , such that
φ
R ij (·) = −φji (·), ji = ij , 1 ≤ i, j ≤ N
Ω exp{−φij (ω)}p(dω) ≤ ij ∀p(·) ∈ Xi , 1 ≤ i, j ≤ N.
(5)
Our goal is, given a positive integer K, to decide via stationary K-repeated observation ω K =
M
S
(ω1 , ..., ωK ), with ω1 , ..., ωK drawn, independently of each other, from a distribution p(·) ∈
Xj ,
j=1
between hypotheses Hj , 1 ≤ j ≤ N , with Hj stating that p(·) ∈ Xj .
Note that the present setting is a straightforward extension of the situation considered in
section 2.2. In particular, given what was called in this section “a detector φ with (X1 , X2 )-risk
≤ ,” and setting φ1,2 = φ, φ2,1 = −φ, φ1,1 ≡ φ2,2 ≡ 0, 1,2 = 2,1 = , 1,1 = 2,2 = 1, we
meet the requirements (5) corresponding to the case N = 2. Vice versa, given detectors and
risks satisfying (5) for the case N = 2 and setting φ = φ1,2 , = 1,2 , we get a detector φ with
(X1 , X2 )-risk ≤ .
We are about to “aggregate” detectors φij into a testing procedure deciding on the hypotheses
H1 , ..., HN via stationary K-repeated observations ω K . It makes sense to consider a slightly more
general problem, specifically, as follows: assume that on the top of the setup data Xj , φij (·), ij ,
we are given an N × N symmetric (proximity) matrix C with zero-one entries and zero diagonal.
We interpret the relation Cij = 0 as closeness of hypotheses Hi and Hj 3 , and we refer to indices
i, j (and hypotheses Hi , Hj ) such that Cij = 0 as C-close. We expect our testing procedure
not to reject the true hypothesis, while rejecting all hypotheses which are not not C-close to
it. On the other hand, we do not care about distinguishing the true hypothesis from C-close
alternatives.
3
In the general case considered in [10], the closeness (i.e., the zero-one matrix C with zero diagonal) not
necessary is symmetric; here we restrict ourselves with symmetric closeness only.
7
The construction.
set
Given Xj , φij (·), ij satisfying (5) along with positive integer K, let us
K
φK
ij (ω ) =
K
X
(K)
φij (ωt ), ij
= K
ij ,
t=1
and let XjK , j = 1, ..., N be a family of probability distributions of ω K = (ω1 , ..., ωK ) where ωt
are i.i.d. with distribution p(·) ∈ Xj . Clearly, (5) implies that
(K)
(K)
φK (·) = −φK
ji (·), ji = ij , 1 ≤ i, j ≤ N
R ij
(K)
K
K
K
∀p(·) ∈ XiK , 1 ≤ i, j ≤ N.
ΩK exp{−φij (ω )}p(dω ) ≤ ij
Now, let α = [αij ] be a skew-symmetric N × N matrix, and let
K
φij (ω K ) = φK
ij (ω ) − αij .
We associate with φ the test TK which, given observation ω K , builds the N × N matrix with
the entries φij (ω K ) and accepts all hypotheses Hi which satisfy the condition
φij (ω K ) > 0 ∀(j : Cij = 1),
and rejects all remaining hypotheses. Note that TK can accept no hypotheses at all, or can
accept more than one hypothesis.
The properties of TK are summarized in the following simple statement (see [10, section
2.3.1]):
Proposition 1 Let
ε = max
1≤i≤N
X
K
ij exp{−αij }.
(6)
j:Cij =1
Let ω K = (ω1 , ..., ωK ) be sampled independently from the distribution p∗ (·) ∈ Xi∗ , for some
i∗ ∈ {1, ..., N }. Then the p∗ (·)-probability of the event “TK does not accept the true hypothesis
Hi∗ or accepts a hypothesis Hi which is not C-close to Hi∗ ” is ≤ ε.
The risk bound ε, as given by (6), depends on the “shifts” αij = −αji , and we would like to use
the shifts resulting in as small as possible value of ε in (1). It is shown in [10, section 3] that
the corresponding shifts solve the convex optimization problem
X
T
Opt = min
f (α) = max
K
exp{α
}
:
α
=
−α
,
(7)
ij
ij
1≤i≤N
α∈RN ×N
j:Cij =1
the optimal ε in (6) being exactly Opt. Moreover, from [10, Proposition 3.3] it immediately follows (see below), that Opt is nothing but the spectral norm kDk2,2 of the entry-wise nonnegative
symmetric matrix
D = [dij = K
ij Cij ]1≤i,j≤N .
Moreover, it is immediately seen (cf. [10, section 3.2]) that if the Perron-Frobenius eigenvector
g of the entrywise nonnegative symmetric matrix D is positive, an optimal solution to (7) is
given by
αij = ln(gj ) − ln(gi ).
(8)
8
In the general case the Perron-Frobenius eigenvector of D can have zero entries, and, moreover,
(7) may happen not to have optimal solutions at all; we, however, can easily find δ-optimal,
with a whatever small δ > 0, solutions to the problem by utilizing in (8) in the role of g the
Perron-Frobenius eigenvector of a matrix D0 with the entries d0ij > dij close to dij , specifically,
such that kD0 k2,2 ≤ kDk2,2 + δ.
In fact, Proposition 3.3 from [10] states that if E = [eij ] is a symmetric N × N matrix with
zero diagonal and positive off-diagonal entries, then the optimal value in the optimization
problem
N
X
OptE = min max
eij exp{αij } : α = −αT
α 1≤i≤N
j=1
is equal to the spectral norm kEk2,2 . Let now eij = dij (= K
ij Cij ) for 1 ≤ i, j ≤ N such
that Cij = 1, and let eij = δ > 0 for i 6= j such that Cij = 0, so that all off-diagonal entries
of E are positive. Note that when δ ↓ 0, both the spectral norms of kEk2,2 and kDk2,2 ,
and the optimal values Opt and OptE become arbitrarily close to each other. Since for the
“perturbed” matrix E the spectral norm kEk2,2 coincides with the optimal value OptE , the
same holds true for the “unperturbed” matrix D.
3
3.1
Sequential Hypothesis Testing
Problem setting
Let us consider the situation of section 2.3. Specifically, assume that we are given N ≥ 2
nonempty families Xj of Borel probability distributions on a Polish observation space Ω. Let,
further,
I
[
J := {1, 2, ..., N } =
Ji
i=1
be a partition of the set of indicesSof Xj ’s into I ≥ 2 non-overlapping nonempty groups J1 , ..., JI .
We associate with the sets X i = j∈Ji Xj the hypotheses Hi , stating that the elements ω1 , ..., ωK
of the K-repeated stationary observation sample ω K are drawn independently from the common
distribution p ∈ X i , and our goal is to decide from observation ω K on the hypotheses H1 , ..., HI .
It is convenient to think about the (values of the) indices i = 1, ..., I of the sets X i as of the
colors of these sets.4
We also assume that we have at our disposal pairwise detectors – Borel functions φij : Ω → R
– for the sets Xj , j = 1, ..., N which satisfy relations (5). From now on, we make the following
assumption:
A: When the indices j and j 0 ∈ J are of different colors (in other words, do not
belong to the same group Ji ), the risk jj 0 (= j 0 j ) of the detector φjj 0 (same as the
detector φj 0 j ) satisfies jj 0 < 1.
Note that assumption A implies that for j and j 0 of different colors the sets of distributions
Xj and Xj 0 are at positive Hellinger distance from each other. This implies that our color
4
Let us agree that these colors are inherited by different entities associated with X i ’s such as index groups Ji ,
indices j ∈ Ji , corresponding sets Xj and distributions p ∈ X i .
9
assignments are unambiguous, and our goal may be reformulated as that of identifying the color
of the distribution underlying the observations.
From the results of section 2.3 it easily follows that if assumption A holds then for any given
> 0, we can decide on the hypotheses H1 , ..., HI with risk ≤ (meaning that the probability to
reject the true hypothesis, same as the probability to accept a wrong one, is ≤ ), provided that
the number K of observations is large enough. This being said, the “large enough” K could be
indeed quite large if there are pairs j ∈ Ji , j 0 ∈ Ji0 , i 6= i0 , with close to 1 values of jj 0 . As a
tradeoff, we can switch from decision rules based on K observations to sequential decision rules,
for which the decision is made on the basis of on-line adjustable number of observations. We can
expect that if we are lucky and the distribution p∗ underlying our observation is “deeply inside”
of some X i∗ and thus is “far” from all X i , i 6= i∗ , the true hypothesis Hi∗ will be accepted much
sooner than in the case when p∗ is close to some of “wrong” X i ’s. Our objective now is to build
sequential tests utilizing the results of section 2.
3.2
Sequential test: construction
The setup
for our “generic” sequential test is given by
1. required risk ∈ (0, 1);
2. positive integer S – number of stages, along with the following entities, defined for 1 ≤
s ≤ S and forming s-th component of the setup:
P
(a) positive reals s , 1 ≤ s ≤ S, such that Ss=1 s = ;
ιS
js
(b) representations Xj =
Xjιs , j ∈ J , where Xjιs are nonempty subsets of Xj ;
ι=1
(c) tolerances δs ∈ (0, 1).
To avoid messy notation, we enumerate, for every s ≤ S, the sets Xjιs , 1 ≤ j ≤ N, 1 ≤ ι ≤ ιjs ,
and call the resulting sets Z1s , ..., ZLs s . Thus, Zqs is one of the sets Xjιs , and we assign to Zqs
and the index q the same color as that of j.
Detectors. We suppose for every s ≤ S and every pair (q, q 0 ), 1 ≤ q, q 0 ≤ Ls , we are given
pairwise detectors φqq0 ,s and reals qq0 ,s ∈ (0, 1] associated with Zqs and Zq0 s and satisfying the
relations
0
φ
R qq0 ,s (·) = −φq0 q,s (·), qq0 ,s = q0 q,s , 1 ≤ q, q ≤ Ls
0
0
0
Ω exp{−φqq ,s (ω)}p(dω) ≤ qq ,s ∀p(·) ∈ Zqs , 1 ≤ q, q ≤ Ls .
s-closeness. Let Hqs be the hypotheses on the distribution p(·) of an observation stating
p(·) ∈ Zqs ; by convention, Hqs is of the same color as Zqs . Let us say that hypothesis Hq0 s is
s-close to hypothesis Hqs (same as q 0 is s-close to q), if either Hqs and Hq0 s are of the same color,
or qq0 ,s > δs if they are of different colors. Let C = C s be the Ls × Ls matrix with (q, q 0 )-entry
equal to 0 if and only if q is s-close to q 0 , and equal to 1 otherwise. This matrix clearly meets
the requirements imposed on C in section 2.3: it is a symmetric matrix (recall that qq0 ,s = q0 q,s )
with 0/1 entries and zero diagonal.
10
Tests Ts . Now let us apply to the collection of hypotheses {Hqs : 1 ≤ q ≤ Ls }, detectors
{φqq0 ,s (·) : 1 ≤ q, q 0 ≤ Ls } and the just defined matrix C s the construction from section 2.3,
assuming that when deciding upon hypotheses H1s , ..., HLs s , we have at our disposal k-repeated
observation ω k , with a given k. Specifically, consider the optimization problem
X
k
T
Opt(k, s) = min
fs (α) := max
qq0 ,s exp{αqq0 } : α = −α
;
(9)
q
α∈RLs ×Ls
q 0 : C s 0 =1
qq
as was shown in section 2.3, the value Opt(k, s) is nothing but the spectral norm of the entry-wise
s .
nonnegative symmetric matrix with the entries kqq0 ,s Cqq
0
s = 1, Opt(k, s) goes to 0 as k → ∞, so that the smallest
Since qq0 ,s ≤ δs ∈ (0, 1) when Cqq
0
k = k(s) such that Opt(k, s) < s is well defined. And since Opt(k(s), s) < s , problem (9)
with k = k(s), whether solvable or not, admits a feasible solution ᾱ(s) such that fs (ᾱ(s) ) ≤ s .
(s)
Applying to detectors φqq0 ,s (·) = φqq0 ,s (·) − ᾱqq0 and to C = C s the construction from section
2.3, we get a test Ts deciding on the hypotheses Hqs , 1 ≤ q ≤ Ls , via k(s)-repeated observation
ω k(s) , with properties as follows:
Let ω1 , ..., ωk(s) be drawn, independently of each other, from common distribution
p∗ (·) obeying the hypothesis Hq∗ s for some 1 ≤ q∗ ≤ Ls . Then the p∗ (·)-probability
of the event “Ts does not accept the true hypothesis Hq∗ s or accepts a hypothesis Hqs
with q not s-close to i∗ ” is at most s .
S-th setup component. We assume that when s = S, the partition of Xj is trivial: ιjS = 1
and Xj1S = Xj for all j ≤ N . Furthermore, we define K = k(S) such that δS ≥ qq0 ,S whenever
q, q 0 are of different colors, implying that q and q 0 are S-close if and only if q and q 0 are of the
same color.5 For the reasons which will become clear in a moment, we are not interested in those
components of our setup for which k(s) > K; if components with this property were present in
our original setup, we can just eliminate them, reducing S accordingly. Finally, we can reorder
the components of our setup to make k(s) nondecreasing in s. Thus, from now on we assume
that
k(1) ≤ k(2) ≤ ... ≤ k(S) =: K.
Sequential test T corresponding to the outlined setup when applied to the observation ω K
works by stages s = 1, 2, ..., S: at stage s we apply test Ts to the initial fragment ω k(s) of ω K . If
the outcome of the latter test is acceptance of a nonempty set of hypotheses Hqs and all these
hypotheses are of the same color i, test T accepts the hypothesis Hi and terminates, otherwise
it proceeds to stage s + 1 (when s < S) or terminates without accepting any hypothesis (s = S).
Note that by construction T never accepts more than one of the hypotheses H1 , ..., HI .
3.3
Sequential test: analysis
For a distribution p ∈ X =
N
S
Xj , let s[p] ∈ {1, 2, ..., S} be defined as follows: for every s, p
j=1
belongs to (perhaps, several) of the sets Zqs , 1 ≤ q ≤ Ls . Further, let Qs [p] be the set of all
We can do so because by assumption A all quantities qq0 ,s , 1 ≤ q, q 0 ≤ LS = N , with q, q 0 of different colors
are less than 1.
5
11
q’s such that p ∈ Zqs . Note that the color of every q ∈ Qs [p] (recall that the sets Zqs and the
corresponding values of q have already been assigned colors) is the same as the color of p. Now,
given p ∈ X , for some s ≤ S it may happen that
∃q ∈ Qs [p] : all q 0 s-close to q are of the same color as q.
(10)
In particular, the latter condition definitely takes place when s = S, since, as we have already
seen, q and q 0 are S-close if and only if q and q 0 are of the same color. Hence for s = S the
conclusion in (10) is satisfied for all q ∈ QS [p]. Now let s[p] be the smallest s such that (10)
takes place, and let the corresponding q be denoted by q[p]. Thus for all p ∈ X ,
s[p] ∈ {1, 2, ..., S}, q[p] ∈ {1, ..., Ls[p] }, p ∈ Zq[p]s[p] ;
whenever q 0 ∈ {1, ..., Ls[p] } is s[p]-close to q[p], q 0 and q[p] are of the same color.
The main result of this section is as follows.
Proposition
2 Let ω1 , ..., ωK be drawn, independently of each other, from a distribution p∗ ∈
S
X (= N
X
),
so that p∗ ∈ Xj∗ for some j∗ ≤ N , and let i∗ be such that j∗ ∈ Ji∗ (i.e., i∗ is the
j
j=1
color of p∗ ). Then p∗ -probability of the event
T , as applied to ω K , terminates not later than at the stage s[p∗ ]
K
E= ω :
and accepts upon termintation the true hypothesis Hi∗
is at least 1 − .
4
Implementing sequential test
Observe that in order for the just described sequential test to recover the “non-sequential” test
from section 2.3, it suffices to utilize setup with S = 1 (recall that our construction fully specifies
the setup component with s = S). This being said, with our sequential test the running time (the
number of observations used to make the inference) depends on the true distribution underlying
the observations, and we can use several degrees of freedom in our setup in order to save on the
number of observations when the true distribution is “deeply inside” the true hypothesis. We
are about to illustrate some of the options available for the basic good o.s.’s presented in section
2.1.
4.1
Preliminaries
Assume that our o.s. is either Gaussian, or Poisson, or Discrete, see section 2.1.1. We denote
by n the dimension of the associated parameter vector µ: µ = [µ1 ; ...; µn ].
Assume that we are given the risk level ∈ (0, 1) and a collection of J nonempty compact
convex sets Xj ⊂ M painted in I ≥ 2 colors (i.e., the set of indices {1, ..., J} is split into I ≥ 2
non-overlapping nonempty sets J1 , ..., JI , i being the common color of all sets Xj , j ∈ Ji ). Sets
Xj give rise to the sets Xj of probability distributions defined by the corresponding densities
pµ (·), µ ∈ Xj . Note that for all considered o.s.’s different values of the parameter µ ∈ M
correspond to different probability densities pµ . We assume that the sets Xj , Xj 0 of different
colors do not intersect thus giving rise to non-intersecting sets of distributions Xj and Xj 0 .
12
Given this input, we intend to specify the setup for sequential test deciding on the associated
hypotheses Hi , 1 ≤ i ≤ I.
Let ψ(µ, ν) : M × M → R be the rate function
Z q
ψ(µ, ν) = ln
pµ (ω)pν (ω)P (dω) ,
Ω
associated with the o.s. in question. An immediate computation shows that
2.3])
1
− νk22 , µ, ν ∈ M = Rn ,
− 8 kµ
√
√
1 Pn
− 2 P `=1 ( µ` − ν` )2 , µ, ν ∈ M = {x ∈ Rn : x >
ψ(µ, ν) =
P0},
√
n
n
n
ln
ω=1 µω νω , µ, ν ∈ M = {x ∈ R : x > 0,
ω=1 xω = 1},
(cf. [10, section
Gaussian case
Poisson case
Discrete case
Theorem 1 states that for the considered o.s.’s, given two nonempty convex compact subsets
X, Y of M and setting
ΨXY = max ψ(µ, ν),
µ∈X,ν∈Y
the quantity exp{ΨXY } is exactly the risk of the detector φ(·) yielded by Theorem 1 as applied
to X, Y and the o.s. in question. Besides this, ψ(µ, ν) clearly is a smooth concave and symmetric
(ψ(µ, ν) = ψ(ν, µ)) function on its domain, and ψ(µ, ν) < 0 whenever µ 6= ν.
For j ∈ {1, ..., J}, we set
ψj (µ) = max ψ(µ, ν) : M → R.
ν∈Xj
Since ψ(µ, ν) is concave in µ, ν ∈ M, and Xj is a convex compact set, the functions ψj (·) are
concave and continuous on M.
Let now R be the set of all ordered pairs (j, j 0 ), 1 ≤ j, j 0 ≤ J, with j, j 0 of different colors;
note that (j, j 0 ) ∈ R if and only if (j 0 , j) ∈ R.
For a pair j, j 0 , 1 ≤ j, j 0 ≤ J, let
ψjj 0 = ΨXj Xj 0 :=
max
µ∈Xj ,ν∈Xj 0
ψ(µ, ν) = max ψj 0 (µ).
µ∈Xj
(11)
Note that for (j, j 0 ) ∈ R the convex compact sets Xj , Xj 0 do not intersect, thus ψjj 0 < 0, and
the objective in the optimization problem on the right hand side of (11) is negative on the
compact feasible set of the problem. We set
d = min [−ψjj 0 ],
(j,j 0 )∈R
(12)
so that d > 0.
Finally, for (j, j 0 ) ∈ R and a nonnegative r we say that a linear inequality `(µ) ≤ 0, µ ∈ M,
defines a (jj 0 r)-cut if for all µ ∈ Xj such that `(µ) ≤ 0,
ψjj 0 (µ) ≤ −r.
Example: default cuts. For (j, j 0 ) ∈ R, let (µjj 0 , νjj 0 ) be (µ, ν)-components of an optimal solution
to the optimization problem (11). Setting
ejj 0 = ∇µ ψ(µjj 0 , νjj 0 ), fjj 0 = ∇ν ψ(µjj 0 , νjj 0 ), (j, j 0 ) ∈ R,
13
and invoking optimality conditions for (11) along with concavity of ψ(·, ·), we get for all (j, j 0 ) ∈
R:
T
(a)
ejj 0 [µ − µjj 0 ] ≤ 0,
T
f 0 [ν − νjj 0 ] ≤ 0,
(b)
∀(µ ∈ Xj , ν ∈ Xj 0 ) :
(13)
jj
T
T
0
0
0
ψ(µ, ν) ≤ ψjj + ejj 0 [µ − µjj ] + fjj 0 [ν − νjj ]. (c)
We conclude that setting
`jj 0 (µ) = ψ(µjj 0 , νjj 0 ) + eTjj 0 (µ − µjj 0 ) − r,
we get an affine function of µ ∈ M which upper-bounds ψjj 0 (·) − r on Xj .
Indeed, for any ν ∈ Xj 0 , µ ∈ Xj
ψ(µ, ν) ≤
≤
T
ψ(µjj 0 , νjj 0 ) + eTjj 0 (µ − µjj 0 ) + fjj
0 (ν − νjj 0 )
ψ(µjj 0 , νjj 0 ) + eTjj 0 (µ − µjj 0 )
(we have used (13.c,b)). Taking in the resulting inequality the supremum over ν ∈ Xj 0 , we
arrive at `jj 0 (µ) ≥ ψj 0 (µ) − r, µ ∈ Xj .
The bottom line is that `jj 0 (µ) ≤ 0 is a (jj 0 r)-cut; we shall refer to this cut as default.
Default cuts: refinement in Gaussian case. As it will become clear from the sequel, for our
purposes the larger is the set {µ ∈ Xj : `(µ) ≤ 0}, the better is (jj 0 r)-cut `(µ) ≤ 0. It is
immediately seen that in the Gaussian case the default cuts as defined above can be easily
improved. Indeed, let Gaussian case take place, and let (j, j 0 ) ∈ R. In this case, in the above
notation we have ψ(µ, ν) = − 18 kµ − νk22 , (ajj 0 , bjj 0 ) is a pair of k · k2 -closest to each other points
in Xj and Xj 0 , ejj 0 = − 41 (ajj 0 − bjj 0 ) is nonzero and separates Xj and Xj 0 :
1 2
∀(µ ∈ Xj , ν ∈ Xj 0 ) : eTij (µ − ν) ≤ eTij (ajj 0 − b0jj ) = − δjj
0 , δjj 0 = kajj 0 − bjj 0 k2 ,
4
whence, setting ηjj 0 = −ejj 0 /kejj 0 k2 ,
ψj 0 (µ) = −
1
1
2
T
T
min kµ − νk22 ≤ − min |ηjj
0 [µ − ν]| & (∀µ ∈ Xj ) : ηjj 0 (µ − ajj 0 ) ≥ 0.
8 ν∈Xj 0
8 ν∈Xj 0
(14)
It follows that
µ ∈ Xj ⇒ ψj 0 (µ) ≤ −
2
1
1 T
T
2
min |ηjj
ηjj 0 (µ − ajj 0 ) +δjj 0 ,
0 [µ − ν]| = −
8 ν∈Xj 0
8 |
{z
}
≥0
implying that the linear inequality
√
T
`jj 0 r (µ) := 2 2r − δjj 0 − ηjj
0 (µ − ajj 0 ) ≤ 0
(15)
is a (jj 0 r)-cut. On the other hand, the default (jj 0 r) cat as defined by our general construction
in Gaussian case is
δjj 0 T
1 2
r − δjj
η 0 (µ − ajj 0 ) ≤ 0.
0 −
8
4 jj
T (µ − a 0 ) ≥ 0}, it is
Taking into account that, as we have already mentioned, Xj ⊂ {µ : ηjj
0
jj
immediately seen that the validity of the latter inequality at µ ∈ Xj implies the validity of (15),
that is, the default cut cuts off Xj a smaller set than the cut (15). By this reason, from now on,
when in the Gaussian case, we refer to the (jj 0 r)-cut (15) as to the default one.
14
4.2
Specifying the setup
The setup for our sequential test is as follows.
1. We select a sequence of positive integers {k̄(s)}∞
s=1 satisfying
k̄(1) = 1, k̄(s) < k̄(s + 1) ≤ 2k̄(s), s = 1, 2, ...
and specify S as the smallest positive integer such that
k̄(S) > d−1 ln SJ 2 / ;
(16)
(17)
(S is well defined due to k̄(s) ≥ s).
For 1 ≤ s ≤ S, we set
s =
r(s) = k̄(s)−1 ln(SJ 2 /), δs = exp{−r(s)}.
S,
(18)
2. For every j ∈ {1, ..., J} and every s ∈ {1, ..., S}, we specify closed convex subsets Xjιs ,
1 ≤ ι ≤ ιjs , of Xj as follows. For every pair (j, j 0 ) ∈ R, we select somehow a (jj 0 r(s))-cut
`jj 0 s (·) ≤ 0 and set
0
j
Xjs
= {µ ∈ Xj : `jj 0 s (µ) ≥ 0}, j 0 ∈ J j ,
j
Xjs
= {µ ∈ Xj : `jj 0 s (µ) ≤ 0, j 0 ∈ J j },
(19)
where for 1 ≤ j ≤ J the set J j contains all indices 1 ≤ j 0 ≤ J of the color different from
that of j. Eliminating form this list all sets which are empty, we end up with a number
ιjs ≤ J of nonempty convex compact sets Xjιs , 1 ≤ ι ≤ ιjs , with Xj being their union.
Observe that r(S) < d by (17), and for (j, j 0 ) ∈ R we clearly have
max ψj 0 (µ) = ψjj 0 ≤ max ψjj 0 = −d
(j,j 0 )∈R
µ∈Xj
implying that affine functions `jj 0 S (µ) ≡ −1 specify legitimate (jj 0 r(S))-cuts. These are
exactly the cuts we use when s = S.
Let us agree that for a subset of the parameter space M denoted by a capital Latin letter, the
script version of the letter denotes the set of all densities pµ with parameters µ from the original
set; e.g., if A denotes a subset of M, then A = {pµ : µ ∈ A}. This agreement works in both
directions: if, say, A is a subset of {pµ : µ ∈ M }, then A = {µ ∈ M : pµ ∈ A}. With this
j0
j
j0
j
convention, the above sets Xjs
, j ∈ J j , Xjs
, Xjιs give rise to families Xjs
, j ∈ J j , Xjs
, Xjιs
of probability distributions on Ω; these families (and densities from the families) inherit colors
from their indices j. We claim that the resulting entities form a legitimate setup for a sequential
test. All we need in order to justify this claim is to verify that the S-component of our setup is
as required, that is, that (a) for every j ∈ {1, ..., J}, ιjS = 1, whence Xj1S = Xj and LS = J,
and that (b) q, q 0 ∈ {1, ..., LS } (recall that LS = J) are S-close if and only if q and q 0 are of the
same color.
15
0
j
To verify (a), note that `jj 0 S (·) ≡ −1 whenever (j, j 0 ) ∈ R, implying that XjS
= ∅ when
j
0
j ∈ J j and XjS = Xj , as claimed in (a). To verify (b), note that as it was already mentioned,
for j, j 0 of different colors, the risk of the detector yielded by Theorem 1 as applied to the
sets X = Xj , Y = Xj 0 , is exp{ψjj 0 }, that is, this risk is ≤ exp{−d}. Invoking the already
verified (a), we conclude that qq0 ,S ≤ exp{−d} whenever 1 ≤ q, q 0 ≤ LS = J and q, q 0
are of different colors. As we have seen, r(S) < d, whence δS = exp{−r(S)} > exp{−d}.
The bottom line is that whenever 1 ≤ q, q 0 ≤ J and q, q 0 are of different colors, we have
qq0 ,S < δS ; this observation combines with the definition of S-closeness to imply that q, q 0
are S-close if and only if q, q 0 are of different colors, as claimed in (b).
The legitimate setup we have presented induces a sequential test, let it be denoted by T . We
are about to analyse the properties of this test.
4.3
Analysis
Our first observation is that for the sequential setup we have presented one has k(s) ≤ k̄(s),
1 ≤ s ≤ S. To verify this claim, we need to check that when k = k̄(s), we have
Opt(k, s) < s =
.
S
As we have already mentioned (see the comment after the definition (9) of Opt(·, ·)), Opt(k, s)
is the spectral norm of the entrywise nonnegative symmetric matrix Dks of the size Ls × Ls
with entries not exceeding δsk . Since, by construction, Ls ≤ J 2 and the diagonal of D is zero,
the spectral norm of Dks does not exceed (J 2 − 1)δsk . The latter quantity indeed is < s when
k = k̄(s), see (18).
Worst-case performance. In the analysis to follow, we assume that ∈ (0, 14 ).
By Proposition 2, the sequential test T always accepts at most one of the hypotheses
H1 , ..., HI , and the probability not to accept the true hypothesis is at most ; moreover, the
number of observations used by T never exceeds k(S) ≤ K := k̄(S). On the other hand, from
the definition (12) of d and Corollary 1 it follows that in order for a whatever test to decide
on the hypotheses H1 , ..., HI with risk via stationary repeated observations, the size of the
observation sample should be at least6
1
ln(1/)
+
2 ln(1/) − ln(2) ln(1/)
K =
≥
.
(20)
ln(1/)
d
4d
As a result, unless d is “astronomically small”, K is within logarithmic factor of K + , implying
quasi-optimal worst-case performance of the test T . The precise statement is as follows:
Proposition 3 Let d > 0, J ≥ 2 and ∈ (0, 41 ) satisfy, for some κ ≥ 1, the relation
Then
ln(1/d) ≤ κ ln(J 2 /).
(21)
κ ln(J 2 /)
K ≤ max 1, 5
.
d
(22)
6
Indeed, this is exactly the smallest number of observations which is necessary, according to Corollary 1, to
separate with the risk ≤ the pair of hypotheses corresponding to (j, j 0 ) ∈ R for which ψjj 0 = −d.
16
For proof, see the appendix.
For all practical purposes we can assume that d ≥ 10−6 , otherwise the lower bound K +
on the number of observations required by (1 − )-reliable test would be impractically large.
Assuming d ≥ 10−6 , (21) is satisfied with κ = 5 (recall that ≤ 41 and J ≥ 2). Thus, for all
practical purposes we may treat the quantity κ from the premise of Proposition 3 as a moderate
absolute constant, implying that the upper bound K on the worst-case observation time of our
(1 − )-reliable sequential test T indeed is within a logarithmic factor O(1) ln(J/)
ln(1/) of the lower
+
bound K on the worst-case observation time of an “ideal” (1 − )-reliable test.
Remark 1 It is easily seen that when k̄(s) grows with s as rapidly as allowed by (16): k̄(s) =
2s−1 , the result completely similar to the one of Proposition 3 holds true in a much wider than
(21) range of values of d, specifically, in the range ln(1/d) ≤ CJ 2 /, for a whatever constant
2
C ≥ 1; and in this range, one has K ≤ C 0 max[1, ln(Jd /) ], with C 0 depending solely on C.
S
Actual performance. For µ ∈ X := Jj=1 Xj let s∗ (µ) be the smallest s ≤ S such that for
j
some j ≤ J it holds µ ∈ Xjs
(see (19)). Equivalently:
s∗ (µ) = min s : ∃j ≤ J : µ ∈ Xj & `jj 0 s (µ) ≤ 0 ∀j 0 ∈ J j .
(23)
0
j
= ∅ whenever j 0 ∈ J j , so that
Note that s∗ (µ) is well defined – we have already seen that XjS
s = S is feasible for the right hand side problem in (23).
S
Recall that in Section 3.3 we have associated with a distribution p ∈ j Xj a set of indices
Qs [p], 1 ≤ s ≤ S, and indices s[p] and
S q[p]. Now
S we are in the situation where there is one-to-one
correspondence µ 7→ pµ between j Xj and j Xj ; with this in mind (and allowing for slight
abuse of notation), in what follows we
S set s[µ] = s[pµ ], q[µ] = q[pµ ], Qs [µ] = Qs [pµ ]. In other
words, by construction, for µ ∈ X := j Xj , Qs [µ] is the set of indices q of those of the sets Zqs
(the latter sets are obtained by linear ordering of the Ls sets Xjιs ) which contain µ, s[µ] is the
smallest s such that for some q ∈ Qs [µ], all q 0 ≤ Ls which are s-close to q are of the same color
as the one of q, and the latter property is shared by q[µ] ∈ Qs[µ] [µ].
Proposition 4 One has
s[µ] ≤ s∗ (µ).
S
Assume that the observations are drawn from density pµ (·), µ ∈ j Xj . By Proposition 2, with
pµ -probability ≥ 1 − our sequential test T terminates in no more than s[µ] ≤ s∗ (µ) steps and
upon termination recovers correctly the color i[µ] of µ (i.e., accepts the true hypothesis Hi[µ] ,
and only this hypothesis). Thus, if µ is “deeply inside” one of the sets Xj , meaning that s∗ (µ)
is much smaller than S, our sequential test will, with reliability 1 − , identify correctly the true
hypothesis Hi[µ] much faster than in S stages.
4.4
Selecting the cuts
Defining the cuts `jj 0 s (·) is one of principal “degrees of freedom” of the just described construcj0
tion. Informally speaking, we would like these cuts to result in as small sets Xjs
, j0 ∈ J j
as possible, thus increasing chances for the probability density underlying our observations to
j
belong to Xjs
for a small value of s (such a value, as we remember, with probability 1 −
upper-bounds the number of stages before termination). Thus, to improve over the basic option
17
represented by the default cuts one may look, for instance, for cuts which minimize, given s
j0
and (j, j 0 ) ∈ R, the mj -dimensional volume of the set Xjs
, where mj is the dimension of Xj .
Setting m = mj , X = Xj and Y = {x ∈ X : ψj 0 (x) ≥ −r(s)}, this goal boils down to solving
the following geometric problem:
(*): Given a convex compact set X ⊂ Rm with a nonempty interior and a nonempty
convex compact subset Y of X, find an affine function `(x) such that the linear
inequality `(x) ≥ 0 is valid on Y and minimizes, under this requirement, the average
linear size
1/m
Size(X`(·) ) := mesm (X`(·) )
of the set X`(·) = {x ∈ X : `(x) ≥ 0}.
Problem (*) seems to be heavily computationally intractable. We are about to present a crude
suboptimal solution to (*).
We can assume w.l.o.g. that Y intersects with the interior of X.7 Let us equip X with a ϑ-selfconcordant barrier F (·).8 Then the minimizer x̄ of F on Y is uniquely defined, belongs to int X
and can be found efficiently by solving the (solvable) convex optimization problem minx∈Y F (x).
¯
By optimality conditions for the latter problem, the affine function `(x)
= h∇F (x̄), x − x̄i is
nonnegative on Y and thus is a feasible solution to (*). Further, from the basic facts of the
theory of self-concordant barriers it follows that
A. The Dikin ellipsoid of F at x̄ – the set D = {x ∈ X : hx − x̄, ∇2 F (x̄)[x − x̄]i ≤ 1} – is
contained in X;
9
B. The set X`(·)
¯ = {x ∈ X : hx − x̄, ∇F (x̄)i ≥ 0} is contained in the set
√
D+ = {x ∈ Rm : hx − x̄, ∇2 F (x̄)[x − x̄]i ≤ ρ2 := (ϑ + 2 ϑ)2 , hx − x̄, ∇F (x̄)i ≥ 0}.
From B it follows that
+
0
0
Size(X`(·)
¯ ) ≤ Size(D ) = ρSize(D ), D = {x ∈ D : hx − x̄, ∇F (x̄)i ≥ 0}.
(24)
On the other hand, if `(·) is a feasible solution to (*), then the set X`(·) contains Y and thus
contains x̄, implying that `(x̄) ≥ 0. Consequently, by A we have
D̄ := {x ∈ D : `(x) ≥ 0} ⊂ {x ∈ X : `(x) ≥ 0} = X`(·) ,
whence
Size(X`(·) ) ≥ Size(D̄).
7
Indeed, otherwise we can specify `(·) as an affine function separating Y and X, so that `(x) ≥ 0 when x ∈ Y
and `(x) ≤ 0 on X and `(·) is nonconstant on X. Clearly, `(·) is a feasible solution to (*) with Size(X`(·) ) = 0.
8
That is, F is a three times continuously differentiable strictly convex function on int X which is an interior
penalty for X (i.e., diverges to +∞ along every sequence of interior points of X converging to a boundary point
of X), and, in addition, satisfies specific differential inequalities; for precise definition and related facts to be used
in the sequel, see [18]. A reader will not lose much when assuming that X is a convex P
compact set given by a
strictly feasible system fi (x) ≤ bi , 1 ≤ i ≤ m, of convex quadratic inequalities, F (x) = − m
i=1 ln(bi − fi (x)) and
ϑ = m.
√
√
9
What is ϑ + 2 ϑ below is 3ϑ in [18]; refinement 3ϑ → ϑ + 2 ϑ is due to F. Jarre, see [17, Lemma 3.2.1].
18
Since `(x̄) ≥ 0 and `(·) is affine, we have D̄ ⊃ D00 := {x ∈ D : `(x) ≥ `(x̄)}, and the mdimensional volume of D00 is at least half of the m-dimensional volume of D0 (since every one of
the sets D00 , D0 is either the entire ellipsoid D, or is the intersection of D with half-space with
the boundary hyperplane passing through the center of D). It follows that
Size(D̄) ≥ Size(D00 ) ≥ 2−1/m Size(D0 ).
Thus, for every feasible solution `(·) to (*) it holds
Size(X`(·) ) ≥ Size(D̄) ≥ 2−1/m Size(D0 ) ≥ 2−1/m ρ−1 Size(X`(·)
¯ ),
¯ to (*) (which can be
where the concluding ≥ is due to (24). We see that the feasible solution
`(·)
√
1/m
1/m
found efficiently) is optimal within the factor 2
ρ=2
[ϑ + 2 ϑ]. This factor is moderate
when X is an ellipsoid (or the intersection of O(1) ellipsoids), and can be unpleasantly large
when ϑ is large. However, this is the best known to us computationally tractable approximation
to the optimal solution to (*).
4.5
Application in the Gaussian case
Comparing cut policies. To get impression on the effect of different cut policies (default
cuts vs. cuts based on self-concordant barriers), consider the Gaussian case with J = 2, X1 =
{x ∈ Rn : δ ≤ x1 ≤ 1 + δ, 0 ≤ xi ≤ 1, 2 ≤ i ≤ n}, X2 = {x ∈ Rn : −1 ≤ xi ≤ 0, 1 ≤ i ≤ n},
I1 = {1}, J2 = {2} (i.e., color of X1 is 1, and color of X2 is 2), and let k̄(s) = 2s−1 , s = 1, 2, ....
We have
P
ψ(µ, ν) = − 18 kµ − νk22 , ψ1 (ν) = − 18 (ν1 − δ)2 + ni=2 νi2 , ψ2 (µ) = − 81 kµk22 ;
2
ψ1,2 = ψ2,1 = − 81 δ 2 , d := min[−ψ1,2 , −ψ2,1 ] = δ8 ;
a1,2 = b2,1 = e := [1; 0; ...; 0], b1,2 = a2,1 = 0 ∈ Rn , e1,2 = f2,1 = e, f1,2 = e2,1 = 0.
Further,
2 ln(4S/)
, 1 ≤ s ≤ S,
2s
where is the target risk, and S is the smallest positive integer such that r(S) as given by the
2
above formula is < χ = δ8
Now, the default cuts are (see (15))
√
(1, 2, r) : `1,2,r (µ) := −µ1 + 2 2r√≤ 0
(2, 1, r) : `2,1,r (ν) := −δ + ν1 + 2 2r ≤ 0
r(s) =
and the sets Zqs associated with the default cuts are:
(
p
1 = {µ : max[δ, 2 2r(s)] ≤ µ ≤ 1 + δ & 0 ≤ µ ≤ 1, 2 ≤ i ≤ n},
Z1s = X1,s
1
i
p
color # 1:
2
Z
= X1,s = {µ : δ ≤ µ1 ≤ 2 2r(s) & 0 ≤ mui ≤ 1, 2 ≤ i ≤ n};
( 2s
p
2 = {ν : −1 ≤ ν ≤ min[0, δ − 2 2r(s)] & − 1 ≤ ν ≤ 0, 2 ≤ i ≤ n},
Z3s = X2,s
1
i
p
.
color # 2:
1 = {ν : δ − 2 2r(s) ≤ ν ≤ 0 & − 1 ≤ ν ≤ 0, 2 ≤ i ≤ n}.
Z4s = X2,s
1
i
The “good” sets here are Z1s and Z3s , meaning that if a single observation is distributed according to N (µ, Id ) with µ belonging to Z1s or Z3s , our sequential test T with probability at least
19
Figure 1: X1 , X2 (blue squares) and cuts (red – default, magenta – smart) for s = 11
[n = 2, = 0.01, δ = 0.1, S = 14, r(11) = 0.0092].
cuts
default
smart
n=2
0.32
3.8e-2
n=3
0.32
7.6e-3
n=4
0.32
1.4e-3
n=5
0.32
2.2e-4
n=6
0.32
3.4e-5
n = 20
0.32
<1.3e-18
Table 1: n-dimensional volumes of “bad” sets Zqs for s = 11 [ = 0.01, δ = 0.1, S = 14, r(11) = 0.0092].
1 − terminates in course of the first s stages with correct conclusion on the color of µ. We see
also that when δ 1 and s is “moderate,” meaning
that r(s) δ 2 , the bad sets Z2s , Z4s , in
p
terms of their n-dimensional volume, form O(1) r(s)-fractions of the respective boxes X1 , X2 ,
provided default cuts are used. When using “smart” cuts – those induced by self-concordant
barriers for X1 , X2 – it is immediately seen that in the range 1 δ r(s) δ 2 (what exactly
means, depends on n), the bad sets Z2s and Z4s become simplexes:
p
Pn
ci λi ≤ r(s)n},
Z2s = {µ = [δ; 0; ...; 0] + [λ1 ; ...; λn ] : λ ≥ 0, p
i=1
P
Z4s = {ν = −[λ1 ; ...; λn ] : λ ≥ 0, ni=1 ci λi ≤ r(s)n},
where ci = ci (n, s) are of order of 1; the good sets Z1s , Z3s are the closures of the complements
of these simplexes to the respective boxes X1 , X2 . The new bad sets are much smaller than
p the
old ones; e.g., their volumes are of order of rn/2 (s) – much smaller than the volumes O( r(s))
of the old bad sets, see Figure 1 and Table 1.
Upper bounding s[µ]. In Gaussian case with default cuts, the quantity s∗ (µ) which, as we
have just seen, is a (1 − )-reliable upper bound on the number of stages in which T recognizes
(1 − )-reliably the true hypothesis Hi[µ] provided the observations are drawn from pµ (·), admits
a transparent geometric upper bound. Observe, first, that in Gaussian case we have
χ=
1
min
min ka − bk22 .
8 (j,j 0 )∈R a∈Xj ,b∈Xj 0
20
(25)
Now let ρ(µ) be the largest ρ such the k · k2 -ball of radius ρ centered at µ is contained in certain
Xj . We claim that
p
p
s[µ] ≤ s∗ (µ) ≤ s̄(µ) := min{s ≤ S : 2 2r(s) ≤ 2 2χ + ρ(µ)},
(26)
meaning that the deeper µ is “inside” one of Xj (the larger is ρ(µ)), the smaller is the number
of observations needed for T to identify correctly the color of µ.
Justification of (26) is as follows. The first inequality in (26) has already been proved. To
prove the second inequality, observe, first, that s̄ := s̄(µ) is well defined (indeed, as we have
seen, r(S) < χ). Let j be such that the k · k2 -ball B of radius ρ(µ) centered at µ is contained
in Xj , and let j 0 ∈ Jjo . In the notation from the description of the Gaussian case default
T
0
0
0
cuts we have, by (14), ηjj
0 [µ − aj∗ j ] ≥ 0 for all µ ∈ Xj and thus for all µ ∈ B, and therefore
T
ηjj
0 [µ − ajj 0 ] ≥ ρ(µ), since by construction kηjj 0 k2 = 1. Further, δjj 0 is the k · k2 -distance
√
between Xj and Xj 0 , whence δjj 0 ≥ 2 2χ due to (25) combined with j 0 ∈ Cjo . Thus,
p
T
∀(j 0 ∈ Jjo ) : δjj 0 + ηjj
2χ + ρ(µ).
0 (µ − ajj 0 ) ≥ 2
(27)
Invoking the definition of s̄ := s̄(µ) and the description (15) of (jj 0 r(s̄))-cuts, we conclude
from (27) that µ indeed satisfies all these cuts and therefore µ ∈ Xjjs̄ , that is, s∗ (µ) ≤ s̄
(recall the definition of s∗ (µ)), as claimed.
Numerical illustration. The following numerical experiment highlights the power of sequential testing in the Gaussian case with default cuts. In this experiment, we are given J = 4 sets
Xj ⊂ R2 ; X1 is the square {0.01 ≤ x1 , x2 ≤ 1}, X2 , X3 , X4 are obtained from X1 by reflections
w.r.t. the coordinate axes and the origin. The partition of the index set into groups Ji is trivial
– these groups are the elements of J = {1, 2, 3, 4}, so that our goal is to recognize which of the
sets Xj contains the mean µ of the observation. Figure 2 presents the graph of the logarithm
of the 0.99-reliable upper bound on the number of observations used by the sequential test with
S = 20, k̄(s) = 2s−1 and d = 5.0e−5 as a function of the mean µ of the observation. We see that
the savings from sequential testing are quite significant. The related numbers are as follows:
4
S
when selecting µ in
Xj at random according to the uniform distribution, the empirical averj=1
age of the number of observations before termination is as large as 1.6 · 105 , reflecting the fact
that Xj are pretty close to each other. At the same time, the median number of observations
before termination is just 154, reflecting the fact that in our experiment µ, with reasonably high
probability, indeed is “deeply inside” the set Xj containing µ.
References
[1] R. Bakeman. Observing interaction: An introduction to sequential analysis. Cambridge
University Press, 1997.
[2] G. A. Barnard. Sequential tests in industrial statistics. Supplement to the Journal of the
Royal Statistical Society, pages 1–26, 1946.
[3] L. Birgé. Approximation dans les spaces métriques et théorie de l’estimation: inégalités de
Cràmer-Chernoff et théorie asymptotique des tests. PhD thesis, Université Paris VII, 1980.
21
Figure 2: “Power” of sequential test, Gaussian case. Left display: sets X1 , X2 , X3 , X4 ; right
4
S
display: ln(k(s∗ (µ))) as a function of µ ∈
Xi , = 0.01.
i=1
[4] L. Birgé. Sur un théorème de minimax et son application aux tests. Probab. Math. Stat.,
3:259–282, 1982.
[5] L. Birgé. Robust testing for independent non identically distributed variables and Markov
chains. In Specifying Statistical Models, pages 134–162. Springer, 1983.
[6] M. Burnashev. On the minimax detection of an imperfectly known signal in a white noise
background. Theory Probab. Appl., 24:107–119, 1979.
[7] M. Burnashev. Discrimination of hypotheses for gaussian measures and a geometric characterization of the gaussian distribution. Math. Notes, 32:757–761, 1982.
[8] H. Chernoff. Sequential analysis and optimal design, volume 8. Siam, 1972.
[9] B. K. Ghosh. A brief history of sequential analysis. Handbook of Sequential Analysis, pages
1–19, 1991.
[10] A. Goldenshluger, A. Juditski, and A. Nemirovski. Hypothesis testing by convex optimization. arXiv preprint arXiv:1311.6765, 2013.
[11] Y. Ingster and I. A. Suslina. Nonparametric goodness-of-fit testing under Gaussian models,
volume 169 of Lecture Notes in Statistics. Springer, 2002.
[12] A. B. Juditsky and A. S. Nemirovski. Nonparametric estimation by convex programming.
The Annals of Statistics, 37(5a):2278–2300, 2009.
[13] T. L. Lai. Sequential analysis: some classical problems and new challenges. Statistica
Sinica, 11(2):303–350, 2001.
[14] L. Le Cam. Convergence of estimates under dimensionality restrictions. The Annals of
Statistics, pages 38–53, 1973.
22
[15] L. Le Cam. On local and global properties in the theory of asymptotic normality of experiments. Stochastic processes and related topics, 1:13–54, 1975.
[16] L. Le Cam. Asymptotic Methods in Statistical Decision Theory. Springer Series in Statistics.
Springer, 1986.
[17] A. Nemirovski. Interior point polynomial time methods in convex programming, 2004.
http://www.isye.gatech.edu/~nemirovs/Lect_IPM.pdf.
[18] A. Nemirovskii and Y. Nesterov. Interior-point polynomial algorithms in convex programming. Society for Industrial and Applied Mathematics, Philadelphia, 1994.
[19] A. Wald. Sequential tests of statistical hypotheses. The Annals of Mathematical Statistics,
16(2):117–186, 1945.
[20] A. Wald and J. Wolfowitz. Optimum character of the sequential probability ratio test. The
Annals of Mathematical Statistics, pages 326–339, 1948.
A
Proof of Proposition 2
Let p∗ , j∗ , i∗ be the entities from the proposition. For 1 ≤ s ≤ S, there exist qs ∈ {1, ..., Ls }
such that p∗ ∈ Zqs s ; without loss of generality, we can assume that qs[p∗ ] = q[p∗ ]. For 1 ≤ s ≤ S
let
E1s = {ω K : Ts as applied to ω k(s) does not accept the hypothesis Hqs s },
E2s = {ω K : Ts as applied to ω k(s) accepts a hypothesis Hqs with q not C s -close to qs },
E s = E1s ∪ E2s .
From the just outlined properties of Ts it follows that p∗ -probability of E s does not exceed s .
Now let
E ∗ = {ω K : no one of the events E1s , E2s , 1 ≤ s ≤ S, takes place.},
P
so that p∗ -probability of E ∗ is at least 1 − 2 Ss=1 s = 1 − . All we need is to verify that E ⊃ E ∗ ,
which is immediate. Indeed, let s∗ = s[p∗ ], q∗ = q[p∗ ] (so that q∗ = qs∗ due to qs[p∗ ] = q[p∗ ]),
and let ω K ∈ E ∗ . By the latter inclusion, E1s∗ does not take place, what implies that
(a) Ts∗ as applied to ω k(s∗ ) accepts Hq∗ s∗ .
Besides this, since ω K ∈ E ∗ , we have ω k(s∗ ) 6∈ E2s∗ , implying, by definition of E2s∗ , that all
hypotheses Hqs∗ accepted by Ts∗ as applied to ω k(s∗ ) are such that q is C s∗ -close to q∗ = q[p∗ ].
By definition of s∗ = s[p∗ ] and q∗ = q[p∗ ], for all q which are C s∗ -close to q∗ , the hypotheses Hqs∗
are of the same color as the accepted by Ts∗ , by (a), hypothesis Hq∗ s∗ . Thus, when ω K ∈ E ∗ ,
we have
(b) the set of hypotheses Hqs∗ accepted by Ts∗ is nonempty, and all these hypotheses are of
the same color, equal to the color of µ.
Invoking the description of T , we conclude that when ω K ∈ E ∗ , the test T terminates not later
than at step s∗ = s[p∗ ] and the termination is productive – some of the hypotheses Hi indeed
are accepted (in fact, in this case exactly one of the hypotheses Hi is accepted, since, as it was
already mentioned, T never accepts more than one hypothesis).
Now let ω K ∈ E ∗ , and let s̄ be the step at which T terminates. As we have already seen,
s̄ ≤ s∗ = s[p∗ ] and T terminates accepting exactly one hypothesis Hī , where ī is a deterministic
23
function of ω k(s̄) . Since p∗ obeys hypothesis Hqs̄ s̄ (by definition of qs̄ ) and ω K 6∈ E1s̄ , the test
Ts̄ as applied to ω k(s̄) accepts the hypothesis Hqs̄ s̄ (which, by construction, has the same color
i∗ as p∗ ). The latter observation combines with the termination rule for T to imply that the
outcome of T in the case of ω K ∈ E ∗ is the hypothesis Hi of the same color i as Hqs̄ s̄ , that is,
the hypothesis i with the same color as the color i∗ of p∗ , that is, the true hypothesis Hi∗ , and
that the outcome is obtained not later than at the stage s[p∗ ].
B
Proof of Proposition 3
Let d > 0, J ≥ 2, ∈ (0, 14 ), κ > 2, and κ ≥ 1 satisfy (21). Relation (22) is trivially true when
S = 1 and thus K = k̄(1) = 1. We denote ϑ = J 2 /, and assume from now on that S > 1, so
that (17) is not satisfied when S = k̄(S) = 1 and therefore d ≤ ln(ϑ).
10 .
We claim that
4κ ln(ϑ)
.
S ≤ S̄ := σ =
3d
(28)
To justify this claim, it suffices to verify that
ln(ϑS̄)/k̄(S̄) < d.
(29)
Indeed, we have σ ≥ 4/3, whence σ ≤ S̄ ≤ 32 σ. Using the bound k̄(S) ≥ 2S̄ ≥ 2σ, we conclude
that the left hand side in (29) does not exceed Cd, with C given by
C=
ln (2κϑ ln(ϑ)/d)
ln (2κϑ ln(ϑ)) + κ ln(ϑ)
≤
,
8
8
3 κ ln(ϑ)
3 κ ln(ϑ)
where the concluding ≤ is due to ln(1/d) ≤ κ ln(ϑ) (the latter is nothing but (21)). From the
relations ϑ = J 2 / > 16 and κ ≥ 1 it immediately follows that C < 1, implying the validity of
(29) and thus – the validity of (28).
20 .
Since 1 < S ≤ S̄, we have
d≤
ln([S − 1]ϑ)
ln([S̄ − 1]ϑ)
ln(σϑ)
≤
≤
.
k̄(S − 1)
k̄(S − 1)
k̄(S − 1)
Thus
2 ln( 34 κϑ ln(ϑ)/d)
2 ln(σϑ)
K = k̄(S) ≤ 2k̄(S − 1) ≤
=
d
d
2 ln( 43 κϑ ln(ϑ)) + κ ln(ϑ)
κ ln(ϑ)
[by (21)] ≤
= C0
,
d
d
2[ln( 43 κ ln(ϑ)) + (κ + 1) ln(ϑ)]
C0 =
.
κ ln(ϑ)
Since ϑ > 16 and κ ≥ 1 we have C 0 ≤ 5, and we arrive at (22).
24
C
Proof of Proposition 4
Let s∗ = s∗ (µ), and let j∗ be such that µ ∈ Xjj∗∗s∗ , see the definition of s∗ (µ). The set Xjj∗∗s∗ 3 µ
is some Zq∗ s∗ with q∗ ∈ Qs∗ [µ]. We claim that
(!) all q 0 which are s∗ -close to q∗ are of the same color as q∗ ;
note that the validity of (!) means that setting s = s∗ , q = q∗ , we meet the requirements in
(10), implying, by definition of s[µ], our target relation s[µ] ≤ s∗ .
To verify (!), let q 0 be s∗ -close to q∗ , so that by definition of s-closeness either
(a) q 0 and q∗ are of the same color, or
(b) q 0 and q∗ are of different colors and q0 q∗ ,s∗ > δs∗ = exp{−r(s∗ )};
all we need to prove is that (b) in fact is impossible. Assume, on the contrary, that (b) takes
place, and let us lead this assumption to a contradiction. As it was already mentioned,
q0 q∗ ,s∗ = exp{λ} where λ =
max
µ∈Zq∗ s∗ ,ν∈Zq0 s∗
ψ(µ, ν).
Thus, we are in the case when the color i of q 0 differs from the color i∗ of q∗ and
λ :=
max
µ∈Zq∗ s∗ ,ν∈Zq0 s∗
ψ(µ, ν) > −r(s∗ ).
(30)
Since q 0 is of color i, we have Zq0 s∗ ⊂ Xj 0 for some j 0 ∈ Ji , and since i 6= i∗ , j 0 and j∗ are of
different colors, or, equivalently, j 0 ∈ J j∗ . Now, by the definition of q∗ we have
Zq∗ s∗ = Xjj∗∗s∗ = {µ ∈ Xj∗ : `j∗ js∗ (µ) ≤ 0 ∀j ∈ J j∗ }.
Hence, taking into account that j 0 ∈ J j∗ and that `j∗ j 0 s∗ (·) ≤ 0 is (j∗ j 0 s∗ )-cut, µ ∈ Zq∗ s∗ implies
that
ψj 0 (µ) ≤ −r(s∗ ).
In its turn, the latter relation, due to Zq0 s∗ ⊂ Xj 0 and that ψj 0 (·) = maxν∈Xj 0 ψ(µ, ν), implies
that
ψ(µ, ν) ≤ −r(s∗ ),
max
µ∈Zq∗ s∗ ,ν∈Zq0 s∗
which contradicts (30).
25
| 10math.ST
|
There and Back Again:
A General Approach to Learning Sparse Models
arXiv:1706.08146v1 [cs.LG] 25 Jun 2017
Vatsal Sharan
Stanford University
[email protected]
Kai Sheng Tai
Stanford University
[email protected]
Peter Bailis
Stanford University
[email protected]
Gregory Valiant
Stanford University
[email protected]
Abstract
We propose a simple and efficient approach to learning sparse models. Our approach consists of (1) projecting the data into a lower dimensional space, (2)
learning a dense model in the lower dimensional space, and then (3) recovering
the sparse model in the original space via compressive sensing. We apply this
approach to Non-negative Matrix Factorization (NMF), tensor decomposition and
linear classification—showing that it obtains 10× compression with negligible loss
in accuracy on real data, and obtains up to 5× speedups. Our main theoretical
contribution is to show the following result for NMF: if the original factors are
sparse, then their projections are the sparsest solutions to the projected NMF problem. This explains why our method works for NMF and shows an interesting new
property of random projections: they can preserve the solutions of non-convex
optimization problems such as NMF.
1
Introduction
In what settings can sparse models be trained more efficiently than their dense analogs? Sparse models
commonly arise in feature selection (e.g., diagnostic tasks [1, 2, 3, 4]), in problems with few, highly
indicative features (e.g., time-series prediction [5, 6]), and in resource-constrained environments (i.e.,
a sparse model is needed for more efficient inference [7, 8]). In these tasks, sparsity can reduce the
data requirements for a desired predictive power, and improve interpretability and inference speed;
and a range of popular techniques provide means of learning these sparse models from dense and/or
high-dimensional data [9, 10]. In this work, we study an additional dimension of sparse statistical
models: the extent to which a model’s sparsity can be leverage for a computational advantage.
We present a general and efficient approach to learning sparse models that explicitly leverages the
target model sparsity. We adapt now-classic techniques from compressive sensing [11, 12] in a
simple framework for learning sparse models from high-dimensional data: (1) project data to a lowerdimensional subspace using a random projection, (2) learn a dense model in this lower-dimensional
subspace, then (3) recover a sparse model in the original dimensionality using compressive sensing.
This approach has several benefits. First, it allows us to perform the bulk of model training in an
often much lower-dimensional dense subspace, offering in significant efficiency gains when data is
high-dimensional and the target model is sparse. Second, it is useful in settings where the dataset is
very large and cannot be fit in memory; our approach allows us to work with a much smaller, low
dimensional projection but still recover the original model. Third, it provides a conceptually simple
approach to learning sparse models, combining powerful ideas from dimensionality reduction and
compressive sensing.
Although low-dimensional projections are extremely popular tools for speeding up a variety of
optimization and learning problems [13, 14, 15], our framework is conceptually different, as we
are interested in being able to recover the sparse solution in the original space from the (possibly
dense) solution in the lower-dimensional space. Note that this is not even information theoretically
possible for many tasks even if the original model or solution is sparse: if the solution in the lowerdimensional space is not a projection of the original solution, then performing a compressive sensing
based recovery has no hope of recovering the original solution (even if the original solution is sparse).
Our contributions. We develop a simple framework for learning sparse models and investigate
it in the context of Non-negative Matrix Factorization (NMF), tensor decomposition, and linear
classification. We show that our approach recovers the true solution with a negligible error across
several synthetic and real datasets. Empirically, sparse learning allows 10× compression with
negligible loss in accuracy on real data, and obtains up to 5× speedups in terms of wall clock time to
obtain the same accuracy.
Theoretically establishing why such speed-ups and high fidelity reconstructions are possible and
justified requires care: we establish novel uniqueness guarantees for NMF which demonstrate that if
the original factors are sparse, then the projection of the factors of the original, high-dimensional
NMF problem is the sparsest possible factorization for the low-dimensional projected NMF problem.
As NMF algorithms implicitly [16] or explicitly [17] learn sparse factorizations, our results explain
why the factorization is preserved by projection, explaining the empirical success of our procedure.
Our results do not assume separability [18] of the NMF factors, a commonly used—but very strong—
assumption for showing uniqueness of NMF. We also derive simple conditions for uniqueness and
provable recovery for tensor decomposition in the projected space. These theoretical insights, backed
by our empirical findings, illustrate the benefits of random projections to learn sparse models—
learning is faster, without compromising model quality.
Our results on the uniqueness of NMF after random projection—and our empirical findings—open a
new perspective on the properties of random projections. It is well known that random projections
can preserve distances and other geometric properties (e.g., [19]), but our results show that random
projections can also preserve the solutions to non-convex optimization problems such as NMF. Our
approach successfully leverages this property to efficiently learn sparse models, and we believe this
aproach has several promising applications beyond our target settings.
We briefly outline the three main classes of sparse models we consider in this paper:
Non-negative Matrix Factorization (NMF). The goal of NMF is to approximate a non-negative
matrix M ∈ Rm×n with a product of two non-negative matrices W ∈ Rm×r and H ∈ Rr×n . The
matrix W is regarded as the dictionary, and H stores the combination weights of each column of M
in terms of the dictionary W . Applications of NMF include document clustering [20] and image
analysis [16, 21]. A crucial property of NMF that enables it to learn good representation of the data is
that the coefficients H are typically sparse [16, 17], so that each document or image is expressed as a
combination of few basis vectors in the dictionary. In settings such as document clustering or image
classification, the columns of M correspond to samples from the data set (e.g., documents or images).
In many scenarios, n m, as the number of samples is large for big datasets. We are interested in
this setting, where n is very large but the factor H is sparse.
Tensor Decomposition. The goal of tensor decomposition
is to decompose a tensor T ∈
Pr
Rm1 ×m2 ×n with rank r in terms of its factors T = i=1 Ai ⊗ Bi ⊗ Ci . Here, Ai denotes the ith column of the factor A ∈ Rm1 ×r and ⊗ denotes the outer product: a, b, c ∈ Rn then a⊗b⊗c ∈ Rn×n×n
and (a ⊗ b ⊗ b)ijk = ai bj ck . As in NMF, we consider settings where the third mode corresponding
to the factor C represents data points and is very large—but sparse. The first two modes could
correspond to features for each data point; for example in a topic modeling setup [22], the first two
modes could represent word co-occurrences, with the third mode representing different documents.
Linear Classification. We also consider the problem of finding the linear classifier to classify m
data points in n dimensions. In modern datasets, the number of features n is often very large, containing numerous irrelevant features [1, 2, 3, 4]. In these settings, the best (and/or most interpretable)
classifier is sparse, as it only relies on the few features which are predictive/indicative.
2
Related Work
Sparse Recovery. In the compressive sensing or sparse recovery framework, there is a sparse signal
x ∈ Rn , and we are given measurements Ax, A ∈ Rd×n , where d n. The goal is to recover
x using the measurements Ax, and the property that x is sparse. Compressive sensing has been
2
extensively studied by several different communities, and many algorithms have been proposed which
have a high compression rate and allow fast decoding [23, 24, 11]. Particularly relevant to our work
is the work on sparse recovery using sparse projection matrices [25, 26, 27, 12], and we use these
results to efficiently recover back the solution in the original space from the solution in the projected
space. As is evident, the key difference of our approach from compressive sensing is that we do not
have direct access to the measurements Aw∗ of the solution w∗ in the projected space, and instead
hope to recover an approximation to Aw∗ by solving the optimization problem in the projected space.
Non-negative Matrix Factorization. Many algorithms have been proposed for NMF, the most
popular ones are multiplicative updates [16] and Alternating Least Squares (ALS). As mentioned
in the introduction, for many modern datasets n m as the number of samples is very large.
Approaches such as ALS and multiplicative weights do not scale to this online setting where the data
set is very large and does not fit in memory. For the online setting, many algorithms [28, 29, 30] have
been proposed for the dictionary learning problem, where the goal is to only learn the matrix W ,
but these approaches are unable to learn the coefficient matrix H. In the context of using random
projection for NMF, Wang and Li [31] obtained speedups compared to ALS by projecting down the
matrix via a random projection. However, they do not attempt to directly recover the factors in the
original space by using the factors in the projected space and exploiting properties of the original
solution such as its sparsity.
Tensor Decomposition. The most popular algorithm for tensor decomposition is Alternating Least
Squares (ALS), which proceeds by fixing two of the factor matrices and solving for the third one
via a least squares problem, and continuing this till convergence. Recently, there has been a burst of
exciting work on speeding up ALS via sampling and sketching approaches [32, 33, 34, 35]. These
approaches obtain significant speedups over the naive ALS algorithm, but they do not leverage the
fact that the factors of the tensor are sparse, which is what we aim to exploit in our paper.
Random Projections for Classification. Say we have data points x ∈ Rn with labels y ∈ {0, 1}
which are separated by a linear classifier w∗ ∈ Rn , therefore y = sign(xT w∗ ). If we perform
a random projection of the points to d dimensions, are they still separable? A number of papers
[36, 37, 38] have examined this basic question, and a one line summary of the results is that if the
original data points are separable with a large margin, then a random projection preserves the margin.
In our framework however, the goal is not to just classify accurately in the projected space, but to
recover the optimal classifier in the projected space. Zhang et al. [39] also consider the question of
obtaining the optimal classifier in the original space from the optimal classifier in the projected space.
They show that a novel approach based on the dual of the optimization problem in the projected space
can recover the original classifier if the data matrix is approximately low rank.1 In our framework,
we instead aim to employ the sparsity of the original classifier w∗ in order to be able to recover it.
3
Our Approach
We instantiate our framework in the context of NMF, tensor decomposition and linear classification.
In each case, we obtain a lower-dimensional problem via a random projection (described in detail
in Sec. 3.1). Denote the random projection matrix as P ∈ Rd×n , where n is the original dimension
and d < n. After solving the projected problem, we obtain an approximation to the solution in the
original space by solving a sparse recovery problem via the following linear program (LP):
ŵ∗ = minimize kwk1 , subject to P w = w̃∗
w
(1)
where w ∈ Rn and w̃∗ ∈ Rd is the solution of the projected problem2 . In the compressive sensing
literature, this technique is also known as basis pursuit [40].
We now describe the specifics of each application of our method.
Non-negative Matrix Factorization. Consider the problem of factorizing a non-negative matrix
M ∈ Rm×n into a product of non-negative matrices M = W H. We are interested in the setting
where matrix H is sparse and n m. We first project M to M̃ = M P T . This corresponds to
projecting each row of H from n dimensions to d dimensions. We then factorize M̃ = W̃ H̃ by
Zhang et al. [39] also obtain guarantees when the true classifier w∗ is sparse, but require that the data points
have most of their mass on the non-zero entries of w∗ in this case, and hence are actually close to low rank.
2
In settings like NMF where the desired solution is non-negative, we can add the additional constraint x ≥ 0
to Eq. 1. In practice, we find that projecting the solutions of Eq. 1 post-hoc onto the non-negative orthant is
often an effective heuristic. Omitting the additional constraint typically yields faster solution times.
1
3
solving the NMF problem for M̃ . We set our estimate Ŵ of W to be W̃ . To recover H from H̃, we
solve the compressive sensing problem in Eq. 1 for each row of Hi of H. Therefore, for a rank-r
approximation, we solve r instances of Eq. 1. The outputs of the compressive sensing step are our
estimates of the rows of H.
Tensor
The goal here is to decompose a tensor T ∈ Rm1 ×m2 ×n in terms of r factors:
Pdecomposition.
r
T = i=1 Ai ⊗ Bi ⊗ Ci where n m1 , m2 and C is sparse. We project the tensor T along the 3rd
Pr
mode to a projected tensor T̃ ∈ Rm1 ×m2 ×d , which has factorization T̃ = i=1 Ai ⊗ Bi ⊗ C̃i , where
C̃i = P Ci . To understand this projection, denote T(n) as the mode n matricization of the tensor T ,
which is the flattening of the tensor T along the nth mode obtained by stacking the matrix slices
together as columns of T(n) . To project T , we project its mode 3 matricization T(3) ∈ Rm1 m2 ×n to
obtain the projection of the mode 3 matricization T̃(3) ∈ Rm1 m2 ×d of T̃ by computing T̃(3) = T(3) P T .
We decompose T̃ in the projected space to obtain the factors Ã, B̃, C̃. Our estimates Â, B̂, Ĉ for the
original factors are  = Ã, B̂ = B̃ and Ĉ is obtained by solving the compressive sensing problem in
Eq. 1 for each column Ci . As above, this yields r instances of Eq. 1.
Linear classification. For the problem of classifying data points x(i) ∈ Rn , 1 ≤ i ≤ m, we first
perform a random projection to project the points into d dimensions using the projection matrix P .
Hence the projected points x̃(i) ∈ Rd are given by x̃(i) = P x(i) . We then solve the classification
problem in the projected space to learn a classifier w̃∗ . Finally, we use compressive sensing to recover
an estimate of the true classifier ŵ∗ from w̃∗ .
3.1
Choosing the projection matrix P
There are several requirements which our projection matrix P must satisfy. The first, more obvious,
requirement is that we should be able to solve the compressive sensing problem to recover the
sparse solution in the original space from the solution in the projected space. A slightly more subtle
requirement is that P should be chosen such that if w∗ is the optimal solution in the original space,
then the solution in the projected space is w̃∗ = P w∗ , so that compressive sensing based recovery can
then be carried out. This requires uniqueness of the optimization problem in the projected space, the
requirement for which could depend on the original optimization problem. In Section 4 we analyze
these requirements for NMF and tensor decomposition. Finally, for computational efficiency, we
want projections which are efficient to compute, and which have the property that the sparse recovery
problem can be solved efficiently using them.
Consider a sparse projection matrix P ∈ {0, 1}d×n such that every column of P has exactly p
non-zero entries, we refer to p as the number of buckets, because each coordinate in the original n
dimensional space contributes to p locations in the projected d dimensional space. Let the support of
the non-zero entries of each column of P be chosen independently and uniformly at random. The
key property of P which enables compressive sensing is that P is the adjacency matrix of a bipartite
expander [41]. Consider a bipartite graph G with n nodes on the left part and d nodes on the right
part such that every node in the left part has degree p. We call G a (γn, α) expander if every subset of
at most t ≤ γn nodes in U has at least αtp neighbors in V . It is not difficult to show that a randomly
chosen matrix P with p non-zero entries per column is the adjacency matrix of a (γn, 4/5) expander
for γn = d/(pe5 ) with high probability if p = O(log n). We formally prove this as Lemma 5 in
Appendix D. Note that as P is sparse, the projection step can be carried out efficiently.
For the basis pursuit problem described in (1), seminal results in compressive sensing [23, 24, 11]
show that using dense projection P guarantees recovery with d = O(k log n) if the original solution is
k-sparse. This property was subsequently extended to sparse projection matrices, and we will leverage
this fact. Theorem 3 of Berinde et al. [12] shows that if x is k-sparse, then with P as defined above
x can be recovered from P x using the LP if P is the adjacency matrix of a (2k, 4/5) expander. A
randomly chosen P with d = O(k log n) will be a (2k, 4/5) expander with high probability (Lemma
5), and hence can be used for sparse recovery of k-sparse signals. Efficient recovery algorithms
[26, 27] which are much faster than solving the LP are also known for sparse projection matrices.
Hence using P defined as above ensures that the sparse recovery step is computationally efficient and
provably works if the original solution w∗ is k-sparse and the solution w̃∗ in the projected space is
w̃∗ = P w∗ .
However, we still need to ensure that the solution w̃∗ = P w∗ . We analyze when this is the case for
NMF and tensor decomposition next in Section 4. We summarize these conditions in Table 1 for
4
Table 1: Parameters for the projection matrix P ∈ {0, 1}d×n suggested by our theoretical analysis
for NMF and tensor decomposition (TD). p is the number of non-zero entries per column of P .
Task
Parameter
Number of buckets p Projection dimension d
NMF
r: rank of matrix
O(log n)
O((r + k) log n)
TD
r: rank of tensor
O(log n)
max{r, O(k log n)}
NMF and tensor decomposition. Note that the bounds suggested by our theoretical results are tight
up to logarithmic factors, as projecting the matrix or the tensor to dimension below the rank will not
preserve uniqueness of the decomposition.
4
Guarantees for our approach
In this section, we will show uniqueness guarantees for performing non-negative matrix factorization
and tensor decomposition on the projected matrix or tensor. This will establish that the factors in the
projected space correspond to a projection of the factors in the original space, a prerequisite for being
able to use compressive sensing to recover the factors in the original space.
4.1
Guarantees for non-negative matrix factorization
Donoho and Stodden [18] showed that non-negative matrix factorization is unique under the separability condition. However, separability is a very strong assumption. Instead of using separability,
we will show uniqueness results for NMF using sparsity of the weight matrix H. Our uniqueness
results for NMF in the original space (Part (a) of Theorem 1) are similar in principle to uniqueness
results obtained for sparse coding by Spielman et al. [42]. We show that if a matrix M has a true
rank r factorization M = W H and H is sparse, then for any other rank r factorization M = W 0 H 0 ,
H 0 has strictly more non-zero entries than H. As NMF has been found to recover sparse solutions
[16, 17], and many algorithms explicitly enforce sparsity [17], it is reasonable to hope that the
recovery procedure will be successful if the true solution is the sparsest possible solution. The main
contribution of Theorem 1 is Part (b), which shows uniqueness results in the projected space. This is
significantly more challenging than showing uniqueness in the original space, because in the original
space we could use the fact the entries of H are drawn independently at random to simplify our
arguments. However we lose access to this independence after the random projection step as the
entries of the projected matrix are no longer independent. Instead, the proof in the projected space
uses the expansion properties of the projection matrix P to control the dependencies, enabling us to
prove uniqueness results in the projected space.
Theorem 1 only assumes that W is full column rank, and that the non-zero entries of H are independent Gaussian random variables. The Gaussian assumption is not necessary and is stated for
simplicity—any continuous probability distribution for the non-zero entries of H suffices. Theorem 1
shows that if the rows of H are k-sparse, then projecting into Ω(p(r + k)) dimensions preserves
uniqueness, with failure probability at most re−βk /n for some constant β > 0.
Theorem 1. Consider a rank r matrix M ∈ Rm×n which has factorization M = W H where
W ∈ Rm×r is full column rank, H ∈ Rr×n and H = B Y where each row of B has exactly k
non-zero entries chosen uniformly at random, each entry of Y is drawn independently from N (0, 1)
and denotes the element wise product. Assume k > C, where C is a fixed constant. Consider
the projection matrix P ∈ {0, 1}d×n where each column of P has exactly p non-zero entries,
p = O(log n) and P is a (γn, 4/5) expander with γ = d/(npe5 ) . Assume d = Ω(p(r + k)). Let
M̃ = M P T . Note that M̃ has a factorization M̃ = W H̃ where H̃ = HP T . Then,
(a) For any other rank r factorization M = W 0 H 0 , kHk0 < kH 0 k0 with failure probability at
most re−βk /n for some β > 0.
(b) For any other rank r factorization M̃ = W 0 H 0 , kH̃k0 < kH 0 k0 with failure probability at
most re−βk /n for some β > 0.
Proof. We prove Part (a) of Theorem 1 in Appendix B. Here, we provide the outline of the proof
for the more challenging Part (b), moving proof details to Appendix B. We will argue about the row
space of the matrix H̃. We first claim that the row space of M̃ equals the row space of H̃. To verify,
note that the row space of M̃ is a subspace of the row space of H̃. As W is full column rank, the rank
of the row space of M̃ equals the rank of the row space of H̃. Therefore, the row space of M̃ equals
5
the row space of H̃. By the same argument, for any alternative factorization M̃ = W 0 H 0 , the row
space of H 0 must equal the row space of M̃ – which equals the row space of H̃. As the row space of
H 0 equals the row space of H̃, therefore H 0 must lie in the row space of H̃. Therefore, our goal will
be to prove that the rows of H̃ are the sparsest vectors in the row space of H̃, which implies that for
any other alternative factorization M̃ = W 0 H 0 , kH̃k0 < kH 0 k0 .
The outline of the proof is as follows. First, we argue that the if we take any subset S of the rows
of H̃, then the number of columns which have non-zero entries in at least one of the rows in S is
large. This implies that taking a linear combination of all the S rows will result in a vector with a
large number of non-zero entries—unless the non-zero entries cancel in many of the columns. But
using the properties of the projection matrix P and the fact that the non-zero entries of the original
matrix H are drawn from a continuous distribution, we show this happens with zero probability.
Lemma 1 shows that the number of columns which are have at least one zero entry in a subset S of
the rows of H grows proportionately with the size of S. The proof proceeds by showing that choosing
B such that each row has k randomly chosen non-zero entries ensures expansion for B with high
probability, and the fact that P is given to be an expander.
Lemma 1. For any subset S of the rows of H̃, define N (S) to be the subset of the columns of H̃
which have a non-zero entry in at least one of the rows in S. Then for every subset S of rows of H̃,
|N (S)| ≥ min{16|S|kp/25, d/200} with failure probability re−βk /n.
We now prove the second part of the argument–that any linear combination of rows in S cannot have
much fewer non-zero entries than N (S), as the probability that many of the non-zero entries get
canceled is zero. Lemma 2 is the key to showing this. Define a vector x as fully dense if all its entries
are non-zero.
Lemma 2. For any subset S of the rows of H̃, let U be the submatrix of H̃ corresponding to the S
rows and N (S) columns. Then with probability one, every subset of the columns of U of size at least
|S|p does not have any fully dense vector in its left null space.
Proof. Without loss of generality, assume that S corresponds to the first |S| rows of H̃, and N (S)
corresponds to the first |N (S)| columns of H̃. We will partition the columns of U into |S| groups
{G1 , . . . , G|S| }. Each group will have size at most p. To select the first group, we choose any entry y1
of Y which appears in the first column of U . For example, if the first row of H has a one in its first
column, and P (1, 1) = 1, then the random variable Y1,1 appears in the first column of U . Say we
choose y1 = Y1,1 . We then choose G1 to be the set of all columns where y1 appears. We then remove
the set of columns G1 from U . To select the second group, we pick any one of the remaining columns,
and choose any entry y2 of Y which appears in that column of U . G2 is the set of all columns where
y2 appears. We repeat this procedure to obtain |S| groups, each of which will have size at most p as
every variable appears in p columns.
Let Nj be the left null space of the first j groups of columns. We define N0 = R|S| . We will now
show that either rank(Ni ) = |S| − i or Ni does not contain a fully dense vector. We prove this
by induction. Consider the jth step, at which we have j groups {G1 , . . . , Gj }. By the induction
hypothesis, either Nj does not contain any fully dense vector, or rank(Nj ) = |S| − j. If Nj does not
contain any fully dense vector, then we are done as this implies that Nj+1 also does not contain any
fully dense vector. Assume that Nj contains a fully dense vector x. Choose any column Uc which
has not been already been assigned to one of the sets. By the following elementary proposition, the
probability that x is orthogonal to Uc is zero. We provide a proof in the appendix.
Proposition 1. Let v = (v1 , . . . , vn ) ∈ Rn be a vector of n independent N (0, 1) random variables.
For any subset S ⊆ {1, . . . , n}, let v(S) ∈ R|S| refer to the subset of v corresponding to the indices in
S. Consider t such subsets S1 , . . . , St . Let each set Si defines some linear relation αSTi v(Si ) = 0, for
some αSi ∈ R|Si | . Assume that the variable vi appear in the set Si . Then the probability distribution
of the set of variables {vt+1 , . . . , vn } conditioned on the linear relations defined by S1 , . . . , St is
still continuous and its density function has full support. In particular, any linear combination of the
set of variables {vt+1 , . . . , vn } has zero probability of being zero.
If Nj contains a fully dense vector, then with probability one, rank(Nj+1 ) = rank(Nj )−1 = n−j−1.
This proves the induction argument. Therefore, with probability one, either rank(N|S| ) = 0 or N|S|
does not contain a fully dense vector and Lemma 2 follows.
6
Figure 1: Pearson correlation of recovered factors for the Enron dataset with NMF (left) and EEG
data with non-negative tensor factorization (right). ‘Recovered coefficients’ here refers to the factors
obtained along the larger mode of the matrix/tensor which was projected down. The ‘dictionary’
refers to the factors along the modes which were not projected down.
Table 2: Frobenius errors and runtimes (mins) for NMF on original and compressed data.
Dataset
n×d
orig. err. orig. time
comp. err.
comp. time
MNIST [46]
RCV1 [47]
Enron [48]
NYTimes [48]
60, 000 × 786
21, 531 × 18, 758
39, 861 × 28, 102
131, 688 × 102, 659
1.42E3
6.41E3
4.51E3
1.04E4
3.1
14.8
18.8
300.2
1.78E3 (+25.3%)
6.50E3 (+ 1.4%)
4.97E3 (+10.0%)
1.07E4 (+ 3.3%)
1.5 (48%)
2.0 (14%)
5.3 (28%)
72.1 (24%)
Rank r = 10; m = b0.1nc; compressed times include projection and LP recovery
We now complete the proof for the second part of Theorem 1. Note that the rows of H̃ have at most
kp non-zero entries. Consider any set S of rows of H̃. Consider any linear combination v ∈ Rd
of the set S rows, such that all the combination weights x ∈ R|S| are non-zero. By Lemma 1,
|N (S)| ≥ min{16|S|kp/25, d/200} with failure probability re−βk /n. We claim that v has more
than |N (S)| − |S|p non zero entries. We prove by contradiction. Assume that v has |N (S)| − |S|
or fewer non zero entries. Consider the submatrix U of H̃ corresponding to the S rows and N (S)
columns. If v has |N (S)| − |S|p or fewer non zero entries, then there must be a subset S 0 of the
|N (S)| columns of U with |S 0 | = |S|p, such that each of the columns in S 0 has at least one non-zero
entry, and the fully dense vector x lies in the left null space of S 0 . But by Lemma 2, the probability
of this happening is zero. Hence v has more than |N (S)| − |S|p non zero entries. Lemma 3 obtains a
lower bound on |N (S)| − |S|p using simple algebra.
Lemma 3. |N (S)| − |S|p ≥ 6kp/5 for |S| > 1 for d ≥ 400p(r + k).
Hence any linear combination of more than one row of H̃ has at least 6kp/5 non-zero entries with
failure probability re−βk /n. Hence the rows of H̃ are the sparsest vectors in the row space of H̃
with failure probability re−βk /n.
4.2
Guarantees for tensor decomposition
Showing uniqueness for tensor decomposition after random projection is not difficult, as tensor
decompsition is unique for every general condition on the factor matrices [43, 44]. For example,
tensors have a unique decomposition if all the factor matrices are full rank. The following elementary
fact follows from this property whenever the projection matrix P is full column rank, which is true
with high probability over the randomness in choosing P .
Pr
Fact 1. Consider a n-dimensional rank r tensor T = i=1 wi Ai ⊗ Bi ⊗ Ci . If A, B and C are full
column rank, then the tensor has a unique decomposition after projecting into any dimension d ≥ r.
Hence tensors have a unique decomposition on projecting to dimension at least the rank r. For
tensor decomposition, we can go beyond uniqueness and prove that efficient recovery of the factors
is also possible in the projected space. We show this using known guarantees for the tensor power
method [45], a popular tensor decomposition algorithm. Our results for provable recovery require
projection dimension d = O(r4 log r). This is higher than the uniqueness requirement, possibly due
to gaps in the analysis of the tensor power method. We state these results in Appendix C.
5
Experiments
7
Figure 2: Visualization of a sample factor from the decomposition of the EEG signal tensor. The top
row is from the decomposition in the original space. The bottom row is from the decomposition of a
compressed version of the tensor where the temporal dimension is reduced by 10x. Notice that we
are able to recover the major features of all 3 factors, including the one compressed by 10x.
We evaluate the practical feasibility of sparse model recovery through a series of experiments on real and synthetic
data. Our experiments were run on a server equipped with
a Xeon E7-4870 CPU at 2.4GHz with 1TB of RAM. All
timing measurements were taken under single-threaded
execution.
Non-negative Matrix Factorization. We evaluate the
performance of compressed NMF on a number of standard non-negative datasets. In our experiments, we used
MATLAB’s built-in nnmf with Alternating Least Squares.
Unless otherwise stated, the optimization was run to convergence with a tolerance of 10−5 . We set the rank to
10. In Table 2 we give normalized Frobenius errors along
with wall clock runtimes for NMF in the original and
compressed space. We use LP-based reconstruction for
the recovery step. Table 2 shows that the compressed
NMF runs achieve achieve up to 5× speedup at the cost
of slightly increased error.
Figure 3: Sparse recovery of a linear classifier on the C ENSUS I NCOME
dataset (original d = 123) with 2500 additional distractor features. The dashed
line shows the F1 score of an L1regularized logistic regressor trained in
the uncompressed space. Plotted scores
As we are usually interested in learning interesting fac- are medians of 5 runs. Observe that we
tors via NMF rather than just minimizing squared error, can obtain a 6.5× compression with a
the linear correlation between the recovered factors and 6% loss in accuracy.
ground-truth factors gives a finer-grained evaluation of
recovery quality than Frobenius error alone. We use the
factors obtained in the original space at convergence as a proxy for ground-truth in the absence of
cluster labels for evaluation. Since the two sets of factors may not be in the same order, we compute
correlation scores via a maximum bipartite matching. In Fig. 1, we plot the correlation of the factors
from the Enron word-document count matrix for various compression factors. At a fixed quality
level of 0.5 (as measured by correlation), we are able to achieve 4× speedups by operating in the
compressed space, and are able to recover back the solution in the original space to small error using
compressive sensing, even for 10× compression.
Tensor Decomposition. The CHB-MIT Scalp EEG Database [49] is a collection of scalp electroencephalogram (EEG) recordings of children susceptible to epileptic seizures. Each EEG recording
consists of 23 real-valued time series; each channel corresponds to a measurement of electrical activity
in the different region of the brain. We preprocess one EEG trace from the dataset (CHB01) by computing the spectrogram of each channel. This yields a non-negative tensor of size 27804 × 303 × 23
(time×frequency×channel) corresponding to approximately 40 hours of data. We expect that this
data should exhibit the desired sparsity property since EEG signals typically exhibit well-localized
time-frequency structure. In our experiments, we set the rank to 10. We compute a non-negative
decomposition of this tensor using projected CP-ALS [50]. The tensor was projected along the first
(temporal) mode using the random projection described in Section 3.1. As with NMF, we treat the
factors obtained in the original space at convergence as ground truth and evaluate the mean correlation
of the factors via a maximum bipartite matching.
In Fig. 1 (right), we plot the correlation of each factor over time for three compression ratios. As in
our NMF experiments, we observe that for the uncompressed dictionary factors (second and third
figures of the EEG data), our approach converges to a good solution with 0.5 correlation 5× faster
8
than running the decomposition along the original dimension. For the compressed factors (first figure
of the EEG data) we get a more modest 2× speedup for d/n ≥ 0.1. Fig 2 shows that we are able to
accurately recover the original factors even with a 10× compression. We provide additional details
regarding this experiment and a qualitative analysis of the resulting decomposition in the Appendix.
Linear Classification. We present empirical results on recovering sparse linear classifiers trained
on a compressed feature space. We use the 2-class C ENSUS I NCOME dataset [48] augmented with
2,500 random binary features sampled iid from Bernoulli(0.5). Including the distractor features, the
total dimensionality is n = 2623. We use the standard train-test split of 32,561/16,281. For a range of
projection dimensions d, we trained logistic regressors on the compressed data and recovered linear
classifiers in the original space using sparse recovery. The sparsity of the random projections was
fixed at s = dm/50e. In addition to performing sparse recovery by solving the equality-constrained
`1 -minimization problem (Eq. 1), we also experimented with less-stringent `2 -ball constraints [51]:
kP w − w̃∗ k2 ≤ ε. In our experiments, we used the heuristic setting ε = 12 kP T w̃∗ k∞ . We also
compare the test performance of the recovered classifiers against the test performance obtained by
the naive approach of projecting the test data using the random projection P and classifying directly
with the weights w̃∗ learnt in the compressed space. Our results for this experiment are summarized
in Fig. 3. The `2 -ball constraint slightly outperforms the stricter equality constraint across all values
of d. At d = 400, we achieve ∼ 6.5× compression of the feature space at a loss of 3.7 F1 points
versus a classifier trained in the uncompressed space (i.e. 94% of the baseline score). Both recovery
techniques significantly outperform the naive approach, highlighting the importance of the denoising
effect realized by sparse recovery.
6
Conclusion
We proposed a simple framework for learning sparse models by combining ideas from dimension
reduction and compressive sensing, and theoretical and empirical results illustrate the viability of our
approach for NMF and tensor decomposition as well as preliminary results for linear classifiers. A
natural direction of future work is to extend these theoretical and experimental results to more models
and applications that require learning sparse models.
References
[1] Robert Tibshirani. Regression shrinkage and selection via the lasso. Journal of the Royal
Statistical Society. Series B (Methodological), pages 267–288, 1996.
[2] Isabelle Guyon and André Elisseeff. An introduction to variable and feature selection. Journal
of machine learning research, 3(Mar):1157–1182, 2003.
[3] Myunghwan Kim, Roshan Sumbaly, and Sam Shah. Root cause detection in a service-oriented
architecture. In ACM SIGMETRICS Performance Evaluation Review, volume 41, pages 93–104.
ACM, 2013.
[4] Peter Bailis, Edward Gan, Samuel Madden, Deepak Narayanan, Kexin Rong, and Sahaana Suri.
MacroBase: Prioritizing attention in fast data. In SIGMOD. ACM, 2017.
[5] Jushan Bai and Serena Ng. Forecasting economic time series using targeted predictors. Journal
of Econometrics, 146(2):304–317, 2008.
[6] Ling Huang, Jinzhu Jia, Bin Yu, Byung-Gon Chun, Petros Maniatis, and Mayur Naik. Predicting
execution time of computer programs using sparse polynomial regression. In Advances in neural
information processing systems, pages 883–891, 2010.
[7] Jeremy Fowers, Kalin Ovtcharov, Karin Strauss, Eric S Chung, and Greg Stitt. A high memory
bandwidth FPGA accelerator for sparse matrix-vector multiplication. In FCCM, 2014.
[8] Song Han, Junlong Kang, Huizi Mao, Yiming Hu, Xin Li, Yubin Li, Dongliang Xie, Hong Luo,
Song Yao, Yu Wang, et al. Ese: Efficient speech recognition engine with sparse lstm on fpga.
In FPGA, 2017.
[9] Trevor Hastie, Robert Tibshirani, and Martin Wainwright. Statistical Learning with Sparsity:
The Lasso and Generalizations. CRC Press, 2016.
[10] Irina Rish and Genady Grabarnik. Sparse Modeling: Theory, Algorithms, and Applications.
CRC Press, 2014.
9
[11] Emmanuel J Candes and Terence Tao. Near-optimal signal recovery from random projections:
Universal encoding strategies? IEEE transactions on information theory, 52(12):5406–5425,
2006.
[12] Radu Berinde, Anna C Gilbert, Piotr Indyk, Howard Karloff, and Martin J Strauss. Combining
geometry and combinatorics: A unified approach to sparse signal recovery. In Communication,
Control, and Computing, 2008 46th Annual Allerton Conference on, pages 798–805. IEEE,
2008.
[13] David P. Woodruff. Sketching as a tool for numerical linear algebra. Foundations and Trends
in Theoretical Computer Science, 10(1–2):1–157, 2014. ISSN 1551-305X. doi: 10.1561/
0400000060. URL http://dx.doi.org/10.1561/0400000060.
[14] Petros Drineas and Michael W. Mahoney. Randnla: Randomized numerical linear algebra.
CACM, 59(6):80–90, May 2016. ISSN 0001-0782. doi: 10.1145/2842602. URL http:
//doi.acm.org/10.1145/2842602.
[15] Roman Vershynin. Introduction to the non-asymptotic analysis of random matrices. arXiv
preprint arXiv:1011.3027, 2010.
[16] Daniel D Lee and H Sebastian Seung. Learning the parts of objects by non-negative matrix
factorization. Nature, 401(6755):788–791, 1999.
[17] Patrik O Hoyer. Non-negative matrix factorization with sparseness constraints. Journal of
machine learning research, 5(Nov):1457–1469, 2004.
[18] David Donoho and Victoria Stodden. When does non-negative matrix factorization give a
correct decomposition into parts? In Advances in neural information processing systems, pages
1141–1148, 2004.
[19] Daniel M Kane and Jelani Nelson. Sparser johnson-lindenstrauss transforms. Journal of the
ACM (JACM), 61(1):4, 2014.
[20] Wei Xu, Xin Liu, and Yihong Gong. Document clustering based on non-negative matrix
factorization. In Proceedings of the 26th annual international ACM SIGIR conference on
Research and development in informaion retrieval, pages 267–273. ACM, 2003.
[21] Stan Z Li, Xin Wen Hou, Hong Jiang Zhang, and Qian Sheng Cheng. Learning spatially
localized, parts-based representation. In Computer Vision and Pattern Recognition, 2001. CVPR
2001. Proceedings of the 2001 IEEE Computer Society Conference on, volume 1, pages I–I.
IEEE, 2001.
[22] Animashree Anandkumar, Rong Ge, Daniel J Hsu, Sham M Kakade, and Matus Telgarsky.
Tensor decompositions for learning latent variable models. Journal of Machine Learning
Research, 15(1):2773–2832, 2014.
[23] David L Donoho. Compressed sensing. IEEE Transactions on information theory, 52(4):
1289–1306, 2006.
[24] Emmanuel J Candes. The restricted isometry property and its implications for compressed
sensing. Comptes Rendus Mathematique, 346(9-10):589–592, 2008.
[25] Anna Gilbert and Piotr Indyk. Sparse recovery using sparse matrices. Proceedings of the IEEE,
98(6):937–947, 2010.
[26] Radu Berinde, Piotr Indyk, and Milan Ruzic. Practical near-optimal sparse recovery in the l1
norm. In Communication, Control, and Computing, 2008 46th Annual Allerton Conference on,
pages 198–205. IEEE, 2008.
[27] Piotr Indyk and Milan Ruzic. Near-optimal sparse recovery in the l1 norm. In Foundations of
Computer Science, 2008. FOCS’08. IEEE 49th Annual IEEE Symposium on, pages 199–207.
IEEE, 2008.
[28] Julien Mairal, Francis Bach, Jean Ponce, and Guillermo Sapiro. Online learning for matrix
factorization and sparse coding. Journal of Machine Learning Research, 11(Jan):19–60, 2010.
[29] Naiyang Guan, Dacheng Tao, Zhigang Luo, and Bo Yuan. Online nonnegative matrix factorization with robust stochastic approximation. IEEE Transactions on Neural Networks and
Learning Systems, 23(7):1087–1099, 2012.
10
[30] Fei Wang, Chenhao Tan, Ping Li, and Arnd Christian König. Efficient document clustering
via online nonnegative matrix factorizations. In Proceedings of the 2011 SIAM International
Conference on Data Mining, pages 908–919. SIAM, 2011.
[31] Fei Wang and Ping Li. Efficient nonnegative matrix factorization with random projections.
In Proceedings of the 2010 SIAM International Conference on Data Mining, pages 281–292.
SIAM, 2010.
[32] Yining Wang, Hsiao-Yu Tung, Alexander J Smola, and Anima Anandkumar. Fast and guaranteed
tensor decomposition via sketching. In Advances in Neural Information Processing Systems,
pages 991–999, 2015.
[33] Zhao Song, David Woodruff, and Huan Zhang. Sublinear time orthogonal tensor decomposition.
In Advances in Neural Information Processing Systems, pages 793–801, 2016.
[34] Casey Battaglino, Grey Ballard, and Tamara G Kolda. A practical randomized cp tensor
decomposition. arXiv preprint arXiv:1701.06600, 2017.
[35] Evangelos Papalexakis, Christos Faloutsos, and Nicholas Sidiropoulos. Parcube: Sparse parallelizable tensor decompositions. Machine Learning and Knowledge Discovery in Databases,
pages 521–536, 2012.
[36] Rosa I Arriaga and Santosh Vempala. An algorithmic theory of learning: Robust concepts and
random projection. In Foundations of Computer Science, 1999. 40th Annual Symposium on,
pages 616–623. IEEE, 1999.
[37] Maria-Florina Balcan, Avrim Blum, and Santosh Vempala. Kernels as features: On kernels,
margins, and low-dimensional mappings. Machine Learning, 65(1):79–94, 2006.
[38] Qinfeng Shi, Chunhua Shen, Rhys Hill, and Anton Hengel. Is margin preserved after random
projection? In Proceedings of the 29th International Conference on Machine Learning (ICML12), pages 591–598, 2012.
[39] Lijun Zhang, Mehrdad Mahdavi, Rong Jin, Tianbao Yang, and Shenghuo Zhu. Random
projections for classification: A recovery approach. IEEE Transactions on Information Theory,
60(11):7300–7316, 2014.
[40] Scott Shaobing Chen, David L Donoho, and Michael A Saunders. Atomic decomposition by
basis pursuit. SIAM review, 43(1):129–159, 2001.
[41] Shlomo Hoory, Nathan Linial, and Avi Wigderson. Expander graphs and their applications.
Bulletin of the American Mathematical Society, 43(4):439–561, 2006.
[42] Daniel A Spielman, Huan Wang, and John Wright. Exact recovery of sparsely-used dictionaries.
In COLT, pages 37–1, 2012.
[43] Joseph B Kruskal. Three-way arrays: rank and uniqueness of trilinear decompositions, with
application to arithmetic complexity and statistics. Linear algebra and its applications, 18(2):
95–138, 1977.
[44] Tamara G Kolda and Brett W Bader. Tensor decompositions and applications. SIAM review, 51
(3):455–500, 2009.
[45] Anima Anandkumar, Prateek Jain, Yang Shi, and Uma Naresh Niranjan. Tensor vs. matrix methods: Robust tensor decomposition under block sparse perturbations. In Artificial Intelligence
and Statistics, pages 268–276, 2016.
[46] Yann LeCun. The mnist database of handwritten digits. http://yann. lecun. com/exdb/mnist/,
1998.
[47] Massih Amini, Nicolas Usunier, and Cyril Goutte. Learning from multiple partially observed
views-an application to multilingual text categorization. In Advances in neural information
processing systems, pages 28–36, 2009.
[48] M. Lichman. UCI machine learning repository, 2013. URL http://archive.ics.uci.edu/
ml.
[49] Ali H Shoeb and John V Guttag. Application of machine learning to epileptic seizure detection.
In Proceedings of the 27th International Conference on Machine Learning (ICML-10), pages
975–982, 2010.
[50] Tamara G Kolda and Brett W Bader. Tensor decompositions and applications. SIAM review, 51
(3):455–500, 2009.
11
Figure 4: Visualization of a factor that correlates with the onset of seizures in the patient. The top
set of plots shows the temporal mode while the bottom set shows the channel mode. 10× and 20×
labels refer to the level of compression. Red dotted lines indicate the actual onset times of seizures.
The 10× compression level preserves most of the peaks, but at 20× compression these peaks are lost.
The channel dictionary is well-preserved across these levels of compression.
[51] Emmanuel J Candes, Justin K Romberg, and Terence Tao. Stable signal recovery from incomplete and inaccurate measurements. Communications on pure and applied mathematics, 59(8):
1207–1223, 2006.
[52] Vatsal Sharan and Gregory Valiant. Orthogonalized ALS: A theoretically principled tensor
decomposition algorithm for practical use. arXiv preprint arXiv:1703.01804, 2017.
[53] Animashree Anandkumar, Rong Ge, and Majid Janzamin. Guaranteed non-orthogonal tensor
decomposition via alternating Rank-1 updates. arXiv preprint arXiv:1402.5180, 2014.
A
A.1
Supplementary Experimental Results
Tensor decomposition of EEG data
Preprocessing. Each channel is individually whitened with a mean and standard deviation estimated
from segments of data known to not contain any periods of seizure. The spectrogram is computed
with a Hann window of size 512 (corresponding to two seconds of data). The window overlap is set to
64. In order to capture characteristic sequences across time windows, we transform the spectrogram
by concatenating groups of sequential windows, following Shoeb and Guttag [49]. We concatenate
groups of size three.
Detection of seizures. The original goal of the work by [49] was to train patient-specific classifiers
to detect the onset of epileptic seizures. We find that a tensor decomposition of the time series yields
a factor that correlates with the onset of seizures, as shown in Fig. 4. This example illustrates the
tradeoff between the compression factor and the fidelity of the recovered modes.
A.2
NMF experiments on synthetic data
We perform experiments on NMF using synthetic data where we can control the sparsity of the true
factors to verify how our approach performs if the factors are very sparse. We take n = m and
r = 20. We vary n from 1000 to 10000. For M = W H, we set W to be the absolute value of a
12
500
1
Full NMF
Compressed NMF
Quality of solution
Time in seconds
400
300
200
100
0
0
0.98
Full NMF
Compressed NMF
0.96
0.94
0.92
2000
4000
6000
8000
N:original dimension
10000
0.9
0
12000
2000
4000
6000
8000
N:original dimension
10000
12000
(a) Time taken for solving NMF in original (b) Quality of solution for solving NMF in
space and compressed space
original space and compressed space
Figure 5: Synthetic experiments on NMF. The total number of non zero entries per row of H is
fixed to be k = 20 and rank r = 20. We project down to k log n dimensions. We obtain significant
asymptotic speedups with negligible change in accuracy in this case.
matrix G ∈ Rn×r such that the entries of G are independent Gaussian random variables. For each
row of H, we fix the number of non-zero entries to be k = 20. We project down to the theoretical
limit of d = O(k log n). We can see from Fig. 5a and 5b that we can get significant speedups in this
setting, at negligible degradation in accuracy. Of course, this setting is highly artificial, but serves as
a sanity check that compressed NMF works.
B
B.1
Additional proofs for Section 4.1: Uniqueness for NMF
Proof of Part (a) of Theorem 1
Proof. As in Part (b), we will argue about the row space of the matrix H. By the same argument as
in Part (b) ,for any alternative factorization M = W 0 H 0 , the row space of H 0 must equal the row
space of M , which equals the row space of H. As the row space of H 0 equals the row space of H,
therefore H 0 must lie in the row space of H. Hence our goal we will prove that the rows of H are the
sparsest vectors in the row space of H, which implies that for any other factorization M = W 0 H 0 ,
kHk0 < kH 0 k0 . Lemma 4 shows that the number of columns which are have at least one zero entry
in a subset S of the rows of H grows proportionately with the size of S.
Lemma 4. For any subset S of the rows of H, define N (S) to be the subset of the columns of H
which have a non-zero entry in at least one of the rows in S. Then for every subset S of rows of H,
|N (S)| ≥ min{4|S|k/5, n/200} with high probability.
Proof. Consider a bipartite graph G with r nodes on the left part U corresponding to the r rows of
H, and n nodes on the right part V corresponding to the n columns or indices of each factor. The ith
node in U has an edge to k nodes in V corresponding to the non-zero indices of the ith row of H.
Note that |N (S)| is the neighborhood of the set of nodes S in G. From Part 1 of Lemma 5, the graph
G is a (γ1 r, 4/5) expander with failure probability re−βk /n for γ1 = n/(rke5 ) and a fixed constant
β > 0.
Lemma 5. Randomly choose a bipartite graph G with n1 vertices on the left part U and n2 vertices
on the right part V such that every vertex in U has degree D. Then,
1. For every n1 , n2 , n1 < n2 , G is a (γn1 , 4/5) expander for D ≥ c for some fixed constant c
n2
−βD
/n2 for a fixed constant β > 0.
and γn1 = De
5 except with probability n1 e
2. For every n1 , n2 , n2 < n1 , G is a (γn1 , 4/5) expander for D ≥ c log n1 for some fixed
n2
constant c and γn1 = De
5 except with probability 1/n2 .
As G is a (γ1 r, 4/5) expander, every set of |S| ≤ γ1 r nodes has at least 4|S|k/5 neighbors. A set of
size |S| > γ1 r nodes, must include a subset of size γ1 r which has 4n/(5e5 ) ≥ n/200 neighbours,
and hence every set of size |S| > γ1 r has at least n/200 neighbors. Therefore, for every subset of S
rows, |N (S)| ≥ min{4|S|k/5, n/200} with failure probability re−βk /n.
13
We now show the second part of the proof—that for any subset S of the rows, every linear combination
of all S rows must have |N (S)| − |S| non-zero entries. The following simple Lemma is a restatement
of a result from from Spielman et al. [42]. We provide a proof in the appendix for completeness.
Define a vector x as fully dense if all its entries are non-zero.
Lemma 6. Let Ω ∈ {0, 1}n×n be any binary matrix with at least one nonzero in each column. Let
U = Ω Y where each entry of Y is drawn independently from N (0, 1). Then with probability one
in the random matrix V , the left nullspace of U does not contain a fully dense vector.
Proof. Let Ni be the null space of the first i columns of U , we define N0 = Rn . We will show that
with probability one, either Ni does not contain a fully dense vector, or rank(Ni ) = n − i. The
proof proceeds by induction. Consider i = 1. Consider any full dense vector x ∈ Rn . As every
column of U has at least one non-zero entry, the probability that x is orthogonal to the first column
U1 is exactly zero. This is because the non-zero entries of U1 are independent Gaussian random
variables, hence the probability that any fixed linear combination is zero is exactly zero. Hence
rank(N1 ) = rank(N0 ) − 1 = n − 1 with probability 1. Consider the jth step. By the induction
hypothesis, either Nj does not contain any fully dense vector, or rank(Nj ) = n − j. If Nj does not
contain any fully dense vector, then we are done as this implies that Nj+1 also does not contain a
fully dense vector. Let Nj contain a fully dense vector x. The (j + 1)th column Uj+1 of U has at
least one non-zero entry, and the entries are independent of the previous columns. The probability
that x is orthogonal to Uj+1 is zero, by the previous argument. Hence with probability one, either
Nj+1 does not contain a fully dense vector, or rank(Nj+1 ) = rank(Nj+1 ) − 1 = n − j − 1. This
proves the induction argument. Therefore, with probability 1, either rank(Nn ) = 0 or Nn does not
contain a fully dense vector. Hence the left null space of U does not contain a fully dense vector.
We now complete the proof for the first part of Theorem 1. Consider any set S of the rows of H.
Consider any linear combination v ∈ Rn of the set S rows, such that all the combination weights x ∈
R|S| are non-zero. By Lemma 4, |N (S)| ≥ min{4|S|k/5, n/200} with failure probability re−βk /n.
We claim that v has more than |N (S)| − |S| non zero entries. We prove by contradiction. Assume
that v has at most |N (S)| − |S| non zero entries. Consider the submatrix U of H corresponding to
the S rows and N (S) columns. If v has |N (S)| − |S| or fewer non zero entries, then there must be
a subset S 0 of the |N (S)| columns of U with |S 0 | = |S|, each of which has at least one non-zero
entry, and such that the fully dense vector x lies in the left null space of S 0 . But by Lemma 6, the
probability of this happening is zero. Hence v has more than |N (S)| − |S| non zero entries. Note that
|N (S)| − |S| ≥ 7k/5. Hence any linear combination of more than one row of H has more than 7k/5
non-zero entries with failure probability re−βk /n. Hence the rows of H are the sparsest vectors in
the row space of H with failure probability re−βk /n.
B.2
Additional proofs for Part (b) of Theorem 1
Lemma 1. For any subset S of the rows of H̃, define N (S) to be the subset of the columns of H̃
which have a non-zero entry in at least one of the rows in S. Then for every subset S of rows of H̃,
|N (S)| ≥ min{16|S|kp/25, d/200} with failure probability re−βk /n.
Proof. Recall that in the proof of Lemma 4, we considered a bipartite graph corresponding to the r
rows and the n indices. After the projection step, the n indices are projected to d dimensions, and the
projection matrix is a (γ2 n, 4/5) expander with γ2 = d/(nke5 ). We can now consider a tripartite
graph, by adding a third set W with d nodes. We add an edge from a node i in V to node j in W if
P (j, i) = 1. For any subset S of rows of H̃, N (S) are the set of nodes in W which are reachable
from the nodes S in U . By Lemma 4, the number of neighbors in V of any set S of nodes in U is of
size at least min{4|S|k/5, n/200} with failure probability re−βk /n. As the projection matrix P is a
(γ2 n, 4/5) expander with γ2 = d/(npe5 ), every subset of size t in V has at least min{4tp/5, d/200}
neighbors in W . By combining this argument with Lemma 4, it follows that for every subset of S
rows of H, |N (S)| ≥ min{16|S|kp/25, d/200} with failure probability re−βk /n.
Proposition 1. Let v = (v1 , . . . , vn ) ∈ Rn be a vector of n independent N (0, 1) random variables.
For any subset S ⊆ {1, . . . , n}, let v(S) ∈ R|S| refer to the subset of v corresponding to the indices in
S. Consider t such subsets S1 , . . . , St . Let each set Si defines some linear relation αSTi v(Si ) = 0, for
14
some αSi ∈ R|Si | . Assume that the variable vi appear in the set Si . Then the probability distribution
of the set of variables {vt+1 , . . . , vn } conditioned on the linear relations defined by S1 , . . . , St is
still continuous and its density function has full support. In particular, any linear combination of the
set of variables {vt+1 , . . . , vn } has zero probability of being zero.
Proof. We prove by induction. For the base case, note that without any linear constraints, the set of
n random variables {v1 , · · · , vn } is continuous and has full support as the random variables vi are
independent Gaussian. Consider the jth step, when linear constraints defined by the sets S1 , · · · , Sj
have been imposed on the variables. We claim that the distribution of the set of random variables
{vj+1 , · · · , vn } is continuous and has full support after imposition of the constraints S1 , · · · , Sj . By
the induction hypothesis, the distribution of the set of random variables {vj , · · · , vn } is continuous
and has full support after imposition of the constraints S1 , · · · , Sj−1 . Note that the linear constraint
Sj can be satisfied for any assignment to the subset of variables {vj+1 , · · · , vn } which appear in
the constraint Sj , as vj can be chosen appropriately because by the induction hypothesis it has full
support conditioned on the previous constraints S1 , · · · , Sj−1 . Hence the probability distribution of
the set of variables {vj+1 , · · · , vn } is still continuous and has full support after adding the constraint
Sj .
Lemma 3. |N (S)| − |S|p ≥ 6kp/5 for |S| > 1 for d ≥ 400p(r + k).
Proof. For 2 ≤ |S| ≤ d/(128kp),
|N (S)| − |S|p ≥ (16kp/25)|S| − p|S| = 30kp/25 + kp(16|S| − 30)/25 − p|S|
≥ 6kp/5 + p k(16|S| − 30) − |S|
For |S| ≥ 2 and k ≥ 2, k(16|S| − 30) − |S| ≥ 0, hence |N (S)| − |S|p ≥ 6kp/5 for 2 ≤ |S| ≤
d/(128kp). For |S| > d/(128kp), |N (S)| ≥ d/200. Therefore, |N (S)|−|S|p ≥ d/200−rp ≥ 2kp
for d ≥ 400p(r + k).
C
Guaranteed recovery for tensors in the projected space
We can prove a stronger result for symmetric, incoherent tensors and guarantee accurate recovery in
the compressed space using the tensor power method. The tensor power method is the tensor analog
of the matrix power method for finding eigenvectors. It is equivalent to finding a rank 1 factorization
using the Alternating Least Squares (ALS) algorithm. Incoherent tensors are tensors for which the
factors have small inner products with other. We define the incoherence µ = maxi6=j {|ATi Aj |}. Our
guarantees for tensor decomposition follow from the analysis of the tensor power method by Sharan
and Valiant [52]. Proposition 2 shows guarantees for recovering one of the true factors, multiple
random initializations can then be used for the tensor power method to recover back all the factors
(see Anandkumar et al. [53]).
Pr
Proposition 2. Consider a n-dimensional rank r tensor T = i=1 wi Ai ⊗ Ai ⊗ Ai . Let cmax =
max
maxi6=j |ATi Aj | be the incoherence between the true factors and γ = w
wmin be the ratio of the
−2
largest and smallest weight. Assume γ is a constant and µ ≤ o(r ). Consider a projection matrix
P ∈ {0, ±1}n×d where every row has exactly p non-zero entries, chosen uniformly and independently
at random and the non-zero entries have uniformly and independently distributed signs. We take
d = O(r4 log r) and p = O(r2 log r). Let à = AP and T̃ be the d dimensional projection of T ,
Pk
hence T̃ = i=1 wi Ãi ⊗ Ãi ⊗ Ãi . Then,
1. For the original tensor decomposition problem, if the initialization x0 ∈ Rn is chosen
uniformly at random from the unit sphere, then with high probability the tensor power method
converges to one of the true factors of A (say the first factor A1 ) in O(r(log r + log log n))
steps, and the estimate A01 satisfies k A1 − A01 k22 ≤ O(r max{µ2 , 1/n2 }).
2. For the projected tensor decomposition problem, if the initialization x0 ∈ Rd is chosen
uniformly at random from the unit sphere, then with high probability the tensor power method
converges to one of the true factors of T̃ (say the first factor Ã1 ) in O(r(log r + log log d))
steps, and the estimate Ã0 satisfies k Ã1 − Ã01 k22 ≤ O(r max{µ2 , 1/d2 }).
15
Proof. Our proof relies on Theorem 3 of Sharan and Valiant [52] and sparse Johnson Lindenstrauss
transforms due to Kane and Nelson [19]. Claim 1 of Proposition 2 is Theorem 3 of Sharan and Valiant
[52]. To show Claim 2 we need to ensure that the incoherence parameter in the projected space is
small. We use the Johnson Lindenstrauss property of our projection matrix to ensure this. A matrix
M is regarded as a Johnson Lindenstrauss matrix if it preserves the norm of a randomly chosen unit
vector x up to a factor of (1 ± ), with failure probabilty δ.
Px [(1 − ) <k M x k22 < (1 + )] > 1 − δ
We use the results of Kane and Nelson [19] who show that with high probability a matrix P ∈
{0, ±1}n×d where every row has p non-zero entries, chosen uniformly and independently at random
and the non-zero entries have uniformly and independently distributed signs, preserves pairwise
distances to within a factor for d = O(−2 log(1/δ)) and p = Θ(−1 log(1/δ)).
It is easy to verify that inner-products are preserved to within an additive error if the pairwise
distances are preserved to within a factors of (1 ± ). By choosing δ = 1/r3 and doing a union
bound over all the r2 pairs of factors, the factors are (µ ± ) incoherent in the projected space with
high probability if they were µ incoherent in the original space. Setting = r−2 log−1 r ensures that
µ + = o(r−2 ). Claim 2 now again follows from Theorem 3 of Sharan and Valiant [52].
D
Proof of expander property: Lemma 5
In this section, we provide a proof of our claim that randomly chosen projection matrix is an expander
with the desired expansion. This is Part 2 of Lemma 5. Part 1 is used in the proof of uniqueness for
NMF (Lemma 4).
Lemma 5. Randomly choose a bipartite graph G with n1 vertices on the left part U and n2 vertices
on the right part V such that every vertex in U has degree D. Then,
1. For every n1 , n2 , n1 < n2 , G is a (γn1 , 4/5) expander for D ≥ c for some fixed constant c
n2
−βD
and γn1 = De
/n2 for a fixed constant β > 0.
5 except with probability n1 e
2. For every n1 , n2 , n2 < n1 , G is a (γn1 , 4/5) expander for D ≥ c log n1 for some fixed
n2
constant c and γn1 = De
5 except with probability 1/n2 .
Proof. Consider any subset S ⊂ U with |S| ≤ γn1 . Let P(N (S) ⊆ M ) denote the probability of
D|S|
|
the event that the neighborhood of S is entirely contained in M ⊂ V . P(N (S) ⊆ M ) ≤ |M
.
n2
We will upper bound the probability of G not being an expander by upper-bounding the probability
of each subset S ⊂ U with |S| ≤ γn1 not expanding. Let P(S̄) denote the probability of the
neighborhood of S being entirely contained in a subset M ⊂ V with M < α|S|D. By a union
bound,
X
P(G is not a (γn1 , α) expander) ≤
P(S̄)
S⊂U
|S|≤γn1
X
≤
X
P(N (S) ⊆ M )
S⊂U
M ⊂V
|S|≤γn1 M =α|S|D
=
γn1 X
X
X
s=1 S⊂U
M ⊂V
|S|=s M =α|S|D
≤
α|S|D D|S|
n2
γn1
X
n1
n2 αDs Ds
s
αDs
n2
s=1
16
By using the relation
n
k
≤
ne
k
k
, we get:
γn1
X
n1 e s n2 e αDs αDs Ds
s
αDs
n2
s=1
"
#s
γn
1
X n1 e n2 e αD αDs D
=
s
αDs
n2
s=1
P(G is not a (γn1 , α) expander) ≤
≤
γn1
X
xss
s=1
where xs =
n1 e
s
n2 e
αDs
αD
αDs
n2
D
. xs can be bounded as follows-
n e αDse1/(1−α) (1−α)D
1
s
n2
e αDγn e1/(1−α) (1−α)D
1
≤
γ
n2
n e1+1/(1−α)
1
≤
Dα(1−α)D
n2
n e6
1
≤
De−D/25 = x
n2
xs =
where in the last step we set α = 4/5. Hence we can upper bound the probability of G not being an
expander as follows—
P(G is not a (γn1 , α) expander) ≤
∞
X
s=1
xs ≤
x
1−x
The two parts of Lemma 5 follow by plugging in the respective values for n1 , n2 and D.
17
| 2cs.AI
|
1
Energy-Aware Wireless Relay Selection in
Load-Coupled OFDMA Cellular Networks
Lei You1 , Di Yuan1,2, Nikolaos Pappas1, and Peter Värbrand1
1 Department
of Science and Technology, Linköping University, Sweden
of Information Technology, Uppsala University, Sweden
{lei.you; di.yuan; [email protected]} {[email protected]}
A. Network Model
We consider a heterogeneous cellular network (HetNet)
with macro cells (MCs), user equipments (UEs), and relay
cells (RCs). Denote by B = {1, 2, . . . , nB } the set of MCs,
T = {1, 2, . . . , nT } the set of UEs, and R = {nT + 1, nT +
2, . . . , nT + nR } the set of RCs. We focus on downlink
transmission in this paper. For any UE j ∈ T, the set of j’s
candidate serving cells is denoted by Cj . For any RC k ∈ R,
denote by Ck the set of k’s candidate MCs for establishing
the backhaul link. The relay selection aims at 1) choosing a
serving cell out of Cj for all j ∈ T, and 2) finding for each
RC k ∈ R an MC out of Ck to establish the backhaul link, so
as to minimize the network transmission energy.
We assume in-band wireless relay transmission [11], which
implies no explicit splitting of available time-frequency resource between the backhaul links and the access links. To
avoid the loop interference [7], the backhaul and access links
should operate on orthogonal resources, meaning that, within
the area of each MC, the time-frequency resource units (RUs)
utilized by the two types of links do not overlap. Thus some
of the links preserve orthogonality with each other. We refer
to Figure 1 for an illustration.
Frequency
Relay techniques provide coverage extension, alleviate fading effects in wireless channels, and lead to more rapid
network roll-out to improve the overall system energy efficiency [1]–[3]. In meeting the fast growing demand of mobile
communication and the increase of user density, wireless
relaying is viewed as a promising technique for the upcoming
5G [4]. It is shown that wireless backhaul technologies have
competitive advantages over the fiber-based solution [5]. In
5G, outdoor relays are likely to be densely deployed in urban
areas, which may cause the cost of installing fiber-based
relay nodes to reach an unacceptable level. For the indoor
scenarios, wireless backhauling may provide better flexibility
and cost-efficiency, compared to a fiber-based solution [6].
In addition, though fiber-based backhauling has advantage in
capacity, reliability, and robustness for transmission, there are
cases in which wired backhauling is impossible (e.g. shortterm links for emergency/disaster relief), hence making the
wireless solution to be the only option for such scenarios [5].
There are two types of relaying modes in terms of wireless
backhauling [6], [7]. One is called “out-band” mode, in which
the backhaul and access links operate on different carriers.
The other is “in-band” mode, meaning that there is no explicit
splitting in frequency resource between backhaul links and
access links [6], [7]. Compared to the former, the latter
does not require a pre-defined separation in the frequency
domain. Moreover, if relays are required to operate on a single
carrier, then there is no possibility to make separation for
implementing out-band relay mode, and thus in-band relay
would be the only option in this case [7].
Recently, studies [8], [9] investigated energy minimization
in orthogonal-frequency-division multiple access (OFDMA)
networks, under an interference model proposed in [10].
This model characterizes the coupling relationship among
the load of cells, which is defined to be the proportion of
consumed time-frequency resource in each cell. The model is
therefore named as a “load-coupling” model [10]. However,
understanding and analyzing load coupling for relays with
wireless backhauling is not straightforward. In this paper, we
provide significant extensions of the model to wireless relay
II. S YSTEM M ODEL
Access
Time
UE 1
RC
Time
Backhaul
Access
Time
MC
Frequency
I. I NTRODUCTION
scenarios, following the LTE-advanced standard of wireless
relays in [6]. We formulate the energy-aware relay selection
problem, named MinE, and prove its computational hardness.
Moreover, we derive an optimality condition, based on which
a relay selection algorithm is proposed for solving MinE.
Numerical results show significant improvement on network
energy consumption, compared to the standard strategy of
strongest-cell association.
Frequency
Abstract—We investigate transmission energy minimization
via optimizing wireless relay selection in orthogonal-frequencydivision multiple access (OFDMA) networks. We take into account the impact of the load of cells on transmission energy.
We prove the NP-hardness of the energy-aware wireless relay
selection problem. To tackle the computational complexity, a
partial optimality condition is derived for providing insights
in respect of designing an effective and efficient algorithm.
Numerical results show that the resulting algorithm achieves high
energy performance.
Frequency
arXiv:1610.00982v2 [cs.IT] 21 Jun 2017
2 Department
UE 2
Time
Figure 1. Illustration of the HetNet model. In this example, there are one
MC, one RC, and two UEs. UE 1 and UE 2 are served by the RC and
MC, respectively. The RC is doing a wireless backhauling with the MC. The
used resource for access links and backhaul links is marked by yellow and
blue colors, respectively. In time and frequency, the resource allocated to
backhauling has orthogonality to the resource used by the access links.
In the remaining parts of this paper, we use the term
“orthogonal links” to refer to those links using orthogonal
time-frequency resource for transmission. Such links are said
2
to be orthogonal to each other. Below, we discuss the characterization of orthogonality. For the sake of presentation,
consider given association of UE access and RC backhauling,
and denote by Ri the set of RCs with a wireless backhaul
connected to MC i for each i ∈ B. The set of UEs served
by any cell (MC or RC) i is represented by Ti . Denote by
tuple hi, ji any (backhaul or access) link from i to j. For
any access link hi, ji with i ∈ B and j ∈ Ti , denote by
Lij = {hi, vi : v ∈ Ri ∪ Ti } the set of links that preserve
orthogonality to link hi, ji. For k ∈ R, suppose that it is
connected with some MC i with a backhaul link. We define
Lkj = {hk, vi : v ∈ Tk } ∪ {hi, ki} to be the set of links
having the intra-cell orthogonality to the access link hk, ji
with j ∈ Tk . And for the backhaul link hi, ki, we define
Lik = {hi, vi : v ∈ Ri ∪ Ti } ∪ {hk, vi : v ∈ Tk } the set
Lik . We denote by L the set of all backhaul and access links
in the network.
in (1) and (2). Let n = nT + nR . The proportions of RU consumption for all potential links in the network are represented
by the vector x = [ [x11 , . . . , x|C1 |1 ], . . . , [x1n , . . . , x|Cn |n ] ].
By plugging (2) in (1), we get the function of the proportion of consumed RUs by hi, ji in (3) below. For vector x
satisfying the cell-load coupling relation in the system model
xij = Fij (x) holds for all hi, ji ∈ L.
rij
(3)
Fij (x) =
pij gij
P
MB log2 (1 +
puv guj xuv +σ2 )
B. Load-Coupling Model
For any UE j ∈ T, we use a variable aj to indicate the
UE’s serving cell, i.e. aj = i if UE j is currently served
by cell i. Similarly, for any RC k ∈ R, we use ak = i
to indicate that RC k is connected to MC i with a wireless
backhaul. For any j, aj ∈ Cj for all j ∈ T ∪ R. The vector
a then denotes the association among MCs, RCs and UEs.
The energy-aware relay selection problem, a.k.a. MinE, is
n
X
[MinE] min M
pa j j x a j j
(4a)
Let rij be the bit rate demand on the link hi, ji. Denote by
γij the signal-to-interference-and-noise ratio (SINR) from i to
j. Without loss of generality, we use an (RU) to refer to the
minimum unit for resource allocation. The bandwidth per RU
is denoted by B. In the denominator in (1), B log2 (1 + γij )
computes the achievable bit rate per RU. We assume that there
are M RUs in total, such that MB log2 (1 + γij ) is the total
achievable bit rate for UE j. In (1), xij is then defined to be the
proportion of RUs used by the transmission link hi, ji, among
all RUs in cell i. The
P sum of the proportion of allocated RUs
be the load of cell
in any cell i, i.e., j∈Ti xij , is defined to P
i, which is bounded by the full load, i.e. j∈Ti xij 6 1 i ∈
B ∪ R.
rij
(1)
xij =
MB log2 (1 + γij )
The SINR on any RU allocated to hi, ji is given by (2). In the
nominator, pij is the transmission power of an RU of link hi, ji
in cell i. The value of gij is the power gain from i to j. In the
denominator of (2), recall that xvu represents the proportion
of occupied RU by hv, ui in cell v. The value of xvu is then
interpreted as the likelihood that hi, ji receives interference
from hv, ui on the RU. Note that hv, ui ∈ L\Lij , which is the
set of all links that are not required to be orthogonal to hi, ji.
pij gij
(2)
γij = P
2
hv,ui∈L\Lij pvu gvj xvu + σ
By (1) and (2), one can observe that a change on xuv
for any link hu, vi may cause a variation after the SINR
of some link hi, ji, thus leading to a new value of xij , i.e.,
the required resource consumption for link hi, ji. Thus, the
levels of resource consumption are inherently coupled. This
relationship, as characterized by (1) and (2), is called loadcoupling.
hv,ui∈L\Lij
It can be verified by observing the concavity of function
Fij (x) that Fij is a standard interference function (SIF) in
respect of x [8], [12]. An SIF has the following property:
starting from an arbitrary positive x(0) , if the fixed point
of function Fij exists, then it is unique, can be iteratively
(k)
computed by xij = Fij (x(k−1) ) (k > 1).
III. P ROBLEM F ORMULATION
x,a,r
j=1
s.t.
x = F(x, a, r)
(4b)
raj j = dj j ∈ T
X
dj k ∈ R
ra k k =
(4c)
(4d)
j∈Tk
X
k∈R:ak =i
xa k k +
X
xik +
xij 6 1
i∈B
(4e)
j∈T:aj =i
X
xkj 6 1
k∈R
(4f)
j∈T:aj =k
aj ∈ Cj
j∈T∪R
(4g)
formulated in (4). The objective of minimizing the energy on
all links is given in (4a). Constraint (4b) ensures that x satisfies
the coupling relationship in the system model. Constraint (4c)
guarantees that the bit rate demand of any UE j ∈ T is
satisfied. Constraint (4d) ensures sufficient bit rate on each
backhaul link. Constraint (4e) and (4f) are imposed to limit
the proportion of consumed RUs in each cell to be no more
than 1, corresponding to the full load constraint for MCs and
RCs, respectively. Constraint (4g) is imposed such that the
selected cell for a backhaul/access link is within the candidate
set.
IV. C OMPLEXITY A NALYSIS
C. Computation of Transmission Energy
Theorem 1. MinE is NP-hard.
Recall that xij represents the proportion of consumed RU of
link hi, ji. Hence, the number of RUs that are used for transmission by hi, ji is Mxij . On each RU, the transmit power is
pij . Then the energy consumption on link hi, ji is Mpij xij . We
now focus on how to compute xij in the load-coupling model
Proof. We reduce the Maximum Independent Set (MIS) problem to MinE. We construct a specific HetNet scenario. For
each UE, there is one potential MC and one potential RC as
candidate serving cells. Correspondingly, for any undirected
graph instance G with N nodes (N > 2) in the MIS problem,
3
we define N UEs. Thus, for any node in G, we have one UE,
one MC and one RC. We use 1, 2, . . . , N to index the nodes
in graph G. We use the term “neighboring” to refer to the
relationship of any two entities that are associated respectively
to two neighboring nodes in G.
For any node i in G, we set the gain from MC i to UE i
to 1.0, the gain from RC i to UE i to 6.0, and the gain from
MC i to RC i to 3.0, respectively. For any two neighbouring
nodes i and k (meaning that there is an edge between node i
and node k) in graph G, we set the gain from RC i to UE k to
a small positive value ǫ. The gain values other than the above
three cases are set to be negligible, treated as zero. The noise
σ2 is set to 1.0. The values of gain and noise can be scaled
without affecting the validity of the proof. The transmit powers
of MCs and RCs are set to 1.0 and 0.5, respectively.The bit
rate demand for any UE is set to 1.0.
Due to space limit, we give a sketch of the line of arguments.
For any UE i, if we have it served by RC i, then all the
resource in RC i is in use. If any RC k neighboring to UE i is
activated, then UE i would receive the interference from RC k,
leading to that UE i’s demand cannot be satisfied anymore by
the access link from RC i to UE i. Hence in a feasible solution,
any pair of two neighboring UEs cannot be simultaneously
served by their corresponding RCs. In addition, one can verify
that it is always better to serve any UE i with RC i rather
than MC i for energy saving. Thus we finish the reduction by
concluding that, to solve this constructed problem instance is
to maximize the number of activated RCs, subject to that at
most one RC of any pair of neighboring RCs can be in use.
Hence the conclusion.
V. E NERGY M INIMIZATION
VIA
O PTIMALITY C ONDITION
This section aims to seek for an effective strategy to deal
with the combinatorial nature of MinE. A partial optimality
condition is derived below, based on which we propose the
relay selection algorithm.
A. Optimality Condition
(1)
construction that x̂(0) = xt , we have x̂âj j = Fâj j (x̂(0) , â) =
(0)
/ t,
Gâj j (x̂(0) , â, t ) = Gâj j (xt , â, t ) = xtâj j = x̂âj j . For j ∈
(1)
we have x̂âj j = Fâj j (x̂(0) , â). By the construction that x̂(0) =
(1)
xt and condition 2), x̂âj j = Fâj j (x̂(0) , â) = Fâj j (xt , â) 6
xtâj j =
(0)
x̂âj j
holds. Therefore, we have
x̂(1) 6 x̂(0) = xt .
(6)
We first consider ê. By the monotonicity of F(x, â) in x, we
have the following property. For any k > 0, if x(k) 6 x(k−1) ,
then F(x(k) , â) 6 F(x(k−1) , â) holds, which would directly
lead to x(k+1) 6 x(k) . According to the discussion above,
we have x(k+1) 6 x(k) for k = 0. We therefore conclude
by mathematical induction that x̂ 6 · · · 6 x̂(1) 6 x̂(0) = xt .
Thus, we have
n
X
ê 6
pâj j xtâj j
(7)
j=1
We then consider ě. For any j ∈
/ t , we have xtâj j =
t
Gâj j (x , â, t ). Since l ⊆ t , we conclude j ∈
/ l for any
j ∈
/ t . Therefore, according to the definition of l , we have
âj = ǎj for j ∈
/ t . Note that in condition 1), x̌ is the starting
point for the fixed-point iterations of the function G(x, ǎ, t ).
According to the definition of the function Gij in (5), for
any j ∈
/ t , we have xtâj j = x̌âj j = x̌ǎj j . Thus we conclude
P
P
t
. For any j ∈ t , condition 1)
= j/
jj
∈t pǎj j xǎP
j/
∈t pâj j xâ
jj
P
t
p
x
<
shows that
â
j
j
j∈t pǎj j x̌ǎj j . Therefore, we
j∈t
âj j
obtain
n
X
ě >
pâj j xtâj j
(8)
j=1
We introduce some notations for deriving the optimality
condition. Denote by â and ǎ any two associations, such
that ∃j âj 6= ǎj . Denote l = {j : âj 6= ǎj , j ∈ T ∪ R}.
Denote by x̂ and x̌ the fixed points of the function F under
â and ǎ, respectively. Denote by ê the
transmission
Ptotal
n
energy with association â, i.e. ê =
j=1 pâj j x̂âj j , and
by ěPthe total transmission energy with association ǎ, i.e.
ě = n
j=1 pǎj j x̌ǎj j .
Definition 1. We define the following function for any i ∈ Cj ,
and any j ∈ T ∪ R, where t is a non-empty subset of T ∪ R.
Gij (x, a, t ) =
Proof. The necessity can be proved straightforwardly by
letting t = T ∪ R. For the sufficiency, the basic idea is
to prove x̂ 6 xt by using Condition 2), and then combine it with Condition 1) to compute respectively ê and
ě. Suppose there exists some set t (l ⊆ t ⊆ T ∪ R)
satisfying 1) and 2). We consider the fixed-point iterations
x̂(k) = F(x̂(k−1) , â) (k > 0). Let x̂(0) = xt . For j ∈ t ,
since xtâj j = Gâj j (xt , â, t ) = Fâj j (xt , â), combined with the
Fij (x, a) j ∈ t
xij
otherwise
(5)
Theorem 2. (Optimality Condition) ê < ě if and only if for
some set t (l ⊆ t ⊆ T ∪ R) such that:
P
P
t
t
1)
j∈t pâj j xâj j <
j∈t pǎj j x̌ǎj j where any xâj j with
t
t
j ∈ T ∪ R is an element of x , and x is the fixed point
of G(x, â, t ), with x̌ being the starting point.
2) Fâj j (xt , â) 6 x̌ǎj j for any j ∈
/ t.
Hence the conclusion ê < ě.
Given any two associations â and ǎ, Theorem 2 serves as
a sufficient and necessary condition for checking whether â
leads to a better energy performance than ǎ. When t ⊂ T ∪ R,
it is called asynchronous fixed-point iterations [12], which is
rather faster if |t | ≪ |T ∪ R| and can be implemented in a
distributed manner.
B. Algorithm Design
We use the function Aj (x) = arg mini∈Cj pij xij to assign
each j ∈ T∪R to the cell with lowest energy for transmitting to
j. Algorithm 1 below takes an initial association ǎ as the input,
and outputs the optimized association â. The pre-defined
parameter η indicates the maximum number of loop rounds.
The vectors a and x are iteratively updated by functions
A and F, respectively. Once a(k) = a(k−1) holds for any
round k, meaning that there is no update on vector a(k) , the
algorithm terminates and returns the optimized association â.
In each round, the set l (k) records the positions of all the
different elements between a(k) and ǎ. In Lines 9 and 10, the
4
asynchronous fixed-point iterations are applied with respect
to set t , with t ⊇ l (k) . Lines 11 and 12 use the optimality
condition to check if the new association a(k) would improve
the transmission energy, with numerical tolerances ǫ1 and ǫ2 .
Algorithm 1: Relay Selection
1
2
3
4
5
6
7
8
9
10
11
a(0) ← ǎ, â ← ǎ;
x(0) ← fixed point of F(x, a(0) );
for k ← 1 to η do
a(k) ← A(x(k−1) );
if a(k) = a(k−1) then
break;
x(k) ← F(x(k−1) , a(k) );
l (k) ← {j : a(k)
6= ǎj , j ∈ T ∪ R};
j
choose a set t such that l (k) ⊆ t ;
(k)
xt P
← fixed point of G(x,
Pâ, t ), starting from x ;
if j∈t pa(k) j xt (k) < j∈t pǎj j x̌ǎj j + ǫ1
aj
j
12
j
∧ Fa(k) j (xt , â) 6 x̌ǎjj + ǫ2 then
j
13
14
â ← a(k)
return â;
VI. N UMERICAL AND S IMULATION R ESULTS
For simulation, 7 MCs are deployed at the center of a
hexagonal region, each with the distance of 500 meters to its
neighbor MC. In each hexagonal region, 2 or 4 RCs as well
as 20 UEs are randomly placed. The network operates at 2
GHz. Each RU is set to 180 kHz bandwidth and the bandwidth
for each cell is 20 MHz. The noise power spectral density is
set to −174 dBm/Hz. We remark that the simulation settings
follow the 3GPP standardization document [6], to be consistent
with expected 5G network scenarios in terms of bandwidth
and network density. Also, the path loss of MCs and RCs
follow the standard 3GPP urban macro and micro models [13],
respectively. With the simulation setup and based on one
thousand instances, the peak rate can achieve 1 Gbps; this
is consistent with [6] and the expectation of 5G. The average
rate, which is naturally lower than the peak one, depends on
user density and resource sharing The shadowing coefficients
are generated by the log-normal distribution with 6 dB and
3 dB standard deviation [13], for MCs and RCs, respectively.
The maximum transmit power levels for MCs and RCs are set
to 800 mW and 50 mW per RU, respectively. Simulations are
run over multiple data sets and averaged afterwards.
Figure 2. User demand versus energy cost.
In Figure 2, there are two network scenarios, in which 2 and
4 RCs are deployed in each hexagonal region, respectively. As
references for comparison, each UE (or RC) j is associated
with the cell in Cj with the best received power. As expected,
the 4-RCs case benefits more on energy performance via relay
selection, compared to the 2-RCs case, since that each UE
has more options for choosing its serving cell in the former.
For these two cases, the improvements by using the proposed
algorithm are 34% and 47%, respectively. Furthermore, the
improvement becomes larger, with the increase of the user
demand, which indicates that an appropriate relay selection is
crucial for a network with heavy data traffic. We remark that
for the best-received power based relay selection, the network
can still also benefit from deploying more RCs on energy cost.
In other words, the energy cost can be reduced by deploying
more RCs, without optimizing the relay selection. However,
one can see from the numerical results that the corresponding
gain is far less compared to optimizing the relay selection.
VII. C ONCLUSION
This paper has provided insights as well as an algorithm for
energy-aware relay selection in load-coupled OFDMA cellular
networks. The algorithm exhibits good performance for energy
saving.
ACKNOWLEDGEMENT
This work has been supported by the Swedish Research
Council and the Linköping-Lund Excellence Center in Information Technology (ELLIIT), Sweden, and the European
Union Marie Curie project MESH-WISE (FP7-PEOPLE2012-IAPP: 324515), DECADE (H2020-MSCA-2014-RISE:
645705), and WINDOW (FP7-MSCA-2012-RISE: 318992).
The work of D. Yuan has been carried out within European
FP7 Marie Curie IOF project 329313.
R EFERENCES
[1] J. G. Andrews, S. Buzzi, W. Choi, S. V. Hanly, A. Lozano, A. C. K.
Soong, and J. C. Zhang, “What will 5G be?” IEEE Journal on Selected
Areas in Communications, vol. 32, no. 6, pp. 1065–1082, 2014.
[2] D. S. Michalopoulos, H. A. Suraweera, and R. Schober, “Relay selection
for simultaneous information transmission and wireless energy transfer:
A tradeoff perspective,” IEEE Journal on Selected Areas in Communications, vol. 33, no. 8, pp. 1578–1594, 2015.
[3] Z. Sheng, J. Fan, C. H. Liu, V. C. M. Leung, X. Liu, and K. K. Leung,
“Energy-efficient relay selection for cooperative relaying in wireless
multimedia networks,” IEEE Transactions on Vehicular Technology,
vol. 64, no. 3, pp. 1156–1170, 2015.
[4] “5g radio access,” Ericsson, Tech. Rep., 2014.
[5] M. Dohler, T. Nakamura, A. Osseiran, J. F. Monserrat, O. Queseth,
and P. Marsch, 5G Mobile and Wireless Communications Technology.
Cambridge University Press, 2016.
[6] “3gpp tr 36.116,” 3GPP, Tech. Rep. V13.0.1, 2016.
[7] J. Gora and S. Redana, “In-band and out-band relaying configurations
for dual-carrier LTE-advanced system,” in Proceedings of IEEE PIMRC,
2011, pp. 1820–1824.
[8] R. L. G. Cavalcante, S. Stanczak, M. Schubert, A. Eisenblaetter, and
U. Tuerke, “Toward energy-efficient 5G wireless communications technologies: Tools for decoupling the scaling of networks from the growth
of operating power,” IEEE Signal Processing Magazine, vol. 31, no. 6,
pp. 24–34, 2014.
[9] C. K. Ho, D. Yuan, L. Lei, and S. Sun, “Power and load coupling
in cellular networks for energy optimization,” IEEE Transactions on
Wireless Communications, vol. 14, no. 1, pp. 509–519, 2015.
[10] I. Siomina and D. Yuan, “Analysis of cell load coupling for lte network
planning and optimization,” IEEE Transactions on Wireless Communications, vol. 11, no. 6, pp. 2287–2297, 2012.
[11] “3gpp tr 36.913,” 3GPP, Tech. Rep. V13.0.0, 2016.
[12] R. D. Yates, “A framework for uplink power control in cellular radio
systems,” IEEE Journal on Selected Areas in Communications, vol. 13,
no. 7, pp. 1341–1347, 1995.
[13] “3gpp tr 36.814,” 3GPP, Tech. Rep. V9.0.0, 2010.
| 7cs.IT
|
arXiv:1612.03408v1 [math.AC] 11 Dec 2016
COHEN-MACAULAY PROPERTIES UNDER THE AMALGAMATED
CONSTRUCTION
Y. AZIMI, P. SAHANDI, AND N. SHIRMOHAMMADI
A BSTRACT. Let A and B be commutative rings with unity, f : A → B a ring homomorphism and
J an ideal of B. Then the subring A ⊲⊳ f J := {(a, f (a) + j)|a ∈ A and j ∈ J} of A × B is called the
amalgamation of A with B along J with respect to f . In this paper, we study the property of CohenMacaulay in the sense of ideals which was introduced by Asgharzadeh and Tousi, a general notion
of the usual Cohen-Macaulay property (in the Noetherian case), on the ring A ⊲⊳ f J. Among other
things, we obtain a generalization of the well-known result that when the Nagata’s idealization is
Cohen-Macaulay.
1. I NTRODUCTION
The theory of Cohen-Macaulay rings is a major area of study in commutative algebra and algebraic geometry. From the appearance of the notion of Cohen-Macaulayness, this notion admits
a rich theory in commutative Noetherian rings. There have been attempts to extend this notion
to commutative non-Noetherian rings, since Glaz raised the question that whether there exists
a generalization of the notion of Cohen-Macaulayness with certain desirable properties to nonNoetherian rings [13], [14]. In order to provide an answer to the question of Glaz [14, Page 220],
recently several notions of Cohen-Macaulayness for non-Noetherian rings and modules were introduced in [15], [16], and [2]. Among those is the Cohen-Macaulay in the sense of A , introduced
by Asgharzadeh and Tousi [2], where A is a non-empty subclass of ideals of a commutative ring
(the definition will be given later in Section 2).
In [7] and [8], D’Anna, Finocchiaro, and Fontana have introduced the following new ring construction. Let A and B be commutative rings with unity, let J be an ideal of B and let f : A → B be
a ring homomorphism. The amalgamation of A with B along J with respect to f is the following
subring
A ⊲⊳ f J := {(a, f (a) + j)|a ∈ A and j ∈ J}
of A × B. This construction generalizes the amalgamated duplication of a ring along an ideal (introduced and studied in [6], [10]). Moreover, several classical constructions such as the Nagata’s
idealization (cf. [19, page 2], [17, Chapter VI, Section 25]), the A + X B[X ] and the A + X B[[X ]]
constructions can be studied as particular cases of this new construction (see [7, Examples 2.5 and
2.6]).
Below, we review briefly some known results about the behavior of Cohen-Macaulayness under
the amalgamated construction and its particular cases.
Let M be an A-module. In 1955, Nagata introduced a ring extension of A called the trivial
extension of A by M (or the idealization of M in A), denoted here by A ⋉ M. Now, assume that A is
Noetherian local and that M is finitely generated. It is well known that the trivial extension A ⋉ M
is Cohen-Macaulay if and only if A is Cohen-Macaulay and M is maximal Cohen-Macaulay, see
[1, Corollary 4.14].
2010 Mathematics Subject Classification. 13A15, 13C14, 13C15.
Key words and phrases. Amalgamated algebra, Cohen-Macaulay ring, Koszul grade, Non-Noetherian ring.
1
2
Y. AZIMI, P. SAHANDI, AND N. SHIRMOHAMMADI
Let A be a Noetherian local ring and I be an ideal of A. Consider the amalgamated duplication
A ⊲⊳ I := {(a, a + i)|a ∈ A and i ∈ I} as in [6], [10]. The properties of being Cohen-Macaulay,
generalized Cohen-Macaulay, Gorenstein, quasi-Gorenstein, (Sn ), (Rn ) and normality under the
construction of amalgamated duplication were studied further in many research papers such as
[6], [9], [3], and [21].
In [9], under the condition that A is Cohen-Macaulay (Noetherian local) and J is a finitely
generated A-module, it is observed that A ⊲⊳ f J is a Cohen-Macaulay ring if and only if it is a
Cohen-Macaulay A-module if and only if J is a maximal Cohen-Macaulay module. Then, in
[22], assuming (A, m) is Noetherian local, J is contained in the Jacobson radical of B such that
depthA J < ∞ and that f −1 (q) 6= m, for each q ∈ Spec(B)\V (J), it is shown that A ⊲⊳ f J is CohenMacaulay if and only if A is Cohen-Macaulay and J is a big Cohen-Macaulay A-module (i.e.
depthA J = dim A).
The next natural step is to seek when the amalgamated algebra A ⊲⊳ f J is Cohen-Macaulay
without Noetherian assumption.
In this paper, we investigate the property of Cohen-Macaulay in the sense of ideals (resp. maximal ideals, finitely generated ideals) on the amalgamation. More precisely, in Section 2, we
recall some essential definitions and results on which we base our approach. In Section 3, we
fix our notation and give some elementary results on the behavior of the Koszul grade with respect to amalgamation. In Section 4, we classify some necessary and sufficient conditions for the
amalgamated algebra A ⊲⊳ f J to be Cohen-Macaulay in the sense of ideals (resp. maximal ideals,
finitely generated ideals) (Theorems 4.1, 4.6 and 4.11). Among the applications of our results are
the classification of when the trivial extension A ⋉ M and the amalgamated duplication A ⊲⊳ I are
Cohen-Macaulay in the sense of ideals (Corollaries 4.8 and 4.16).
2. P RELIMINARIES
To facilitate the reading of the paper, we recall in this section some preliminary definitions and
properties to be used later.
Let b be a finitely generated ideal of a commutative ring A and M be an A-module. Assume
that b is generated by the sequence x = x1 , . . . , xℓ . We denote the Koszul complex related to x by
K• (x). The Koszul grade of b on M is defined by
K. gradeA (b, M) := inf{i ∈ N ∪ {0}|H i (HomA (K• (x), M)) 6= 0}.
It follows from [5, Corollary 1.6.22] and [5, Proposition 1.6.10(d)] that this does not depend on
the choice of generating sets of b.
Let a be an arbitrary ideal of A. One can then define the Koszul grade of a on M by setting
K. gradeA (a, M) := sup{K. gradeA (b, M)|b is a finitely generated subideal of a}.
In view of [5, Proposition 9.1.2(f)], this definition coincides with the original one for finitely generated ideals. In particular, when (A, m) is local Noetherian, depthA M was defined by K. gradeA (m, M)
in [5, Section 9.1].
The Čech grade of b on M is defined by
Č. gradeA (b, M) := inf{i ∈ N ∪ {0}|Hxi (M) 6= 0}.
Here Hxi (M) denotes the i-th cohomology of the Čech complex of M related to x. It follows from
[16, Proposition 2.1(e)] that Hxi (M) is independent of the choice of sequence of generators for b.
One can then define
Č. gradeA (a, M) := sup{Č. gradeA (b, M)|b is a finitely generated subideal of a}.
By virtue of [16, Proposition 2.7], one has Č. gradeA (a, M) = K. gradeA (a, M).
COHEN-MACAULAY
3
Let p a prime ideal of R. By htM p, we mean the Krull dimension of the Rp -module Mp . Also,
htM a := inf{htM p|p ∈ SuppA (M) ∩V (a)}.
Let A be a non-empty subclass of the class of all ideals of the ring A and M be an A-module.
We say that M is Cohen-Macaulay in the sense of A if htM (a) = K. gradeA (a, M) for all ideals a
in A , see [2, Definition 3.1]. The classes we are interested in are the class of all maximal ideals,
the class of all ideals and the class of all finitely generated ideals. Assume that A is Noetherian. It
is well-known that A is Cohen-Macaulay (in the sense of the original definition in the Noetherian
setting) if and only if it is Cohen-Macaulay in the sense of ideals (resp. maximal ideals, finitely
generated ideals) see [5, Corollary 2.1.4].
3. T HE KOSZUL
GRADE ON AMALGAMATION
Let us fix some notation which we shall use frequently throughout the paper: A, B are two
commutative rings with unity, f : A → B is a ring homomorphism, and J denotes an ideal of B. So
that J is an A-module via the homomorphism f . In the sequel, we consider the contraction and
extension with respect to the natural embedding ιA : A → A ⊲⊳ f J defined by ιA (x) = (x, f (x)), for
every x ∈ A. In particular, for every ideal a of A, ae means a(A ⊲⊳ f J).
This section is devoted to prove some lemmas on the behavior of the Koszul grade on amalgamation. These lemmas provide the key for some crucial arguments later in this paper. In the proof
of the next lemma, we use Hi (x, M) to denote the ith Koszul homology of an A-module M with
respect to a finite sequence x ⊂ A.
Lemma 3.1. Let the notation and hypotheses be as in the beginning of this section. Then
(i) for any finitely generated ideal b of A, one has the equality
K. gradeA⊲⊳ f J (be , A ⊲⊳ f J) = min{K. gradeA (b, A), K. gradeA (b, J)}.
(ii) for any ideal a of A, one has the inequality
K. gradeA⊲⊳ f J (ae , A ⊲⊳ f J) ≤ min{K. gradeA (a, A), K. gradeA (a, J)}.
Proof. Assume that b is a finitely generated ideal of A and that b is generated by a finite sequence
x of length ℓ. Then, using [2, Proposition 2.2(iv)] together with [16, Proposition 2.7], we have
K. gradeA⊲⊳ f J (be , A ⊲⊳ f J)
= K. gradeA (b, A ⊲⊳ f J)
= sup{k ≥ 0|Hℓ−i (x, A ⊲⊳ f J) = 0 for all i < k}
= sup{k ≥ 0|Hℓ−i (x, A) ⊕ Hℓ−i(x, J) = 0 for all i < k}
= min{K. gradeA (b, A), K. gradeA (b, J)}.
For the third equality, one notices that the amalgamation A ⊲⊳ f J, as an A-module, is isomorphic
to the direct sum of A ⊕ J using [7, Lemma 2.3(4)]. This proves (1). To obtain (2), assume that
a is an ideal of A. Let Σ be the class of all finitely generated subideals of a. It follows from the
definition that
K. gradeA (a, A ⊲⊳ f J)
= sup{K. gradeA (b, A ⊲⊳ f J)|b ∈ Σ}
= sup{min{K. gradeA (b, A), K. gradeA (b, J)}|b ∈ Σ}
≤ min{sup{K. gradeA (b, A)|b ∈ Σ}, sup{K. gradeA (b, J)|b ∈ Σ}}
= min{K. gradeA (a, A), K. gradeA (a, J)}.
4
Y. AZIMI, P. SAHANDI, AND N. SHIRMOHAMMADI
Again, using this in conjunction with [2, Proposition 2.2(iv)], one deduces that
K. gradeA⊲⊳ f J (ae , A ⊲⊳ f J) = K. gradeA (a, A ⊲⊳ f J)
≤ min{K. gradeA (a, A), K. gradeA (a, J)}.
Lemma 3.2. Assume that A is Cohen-Macaulay in the sense of (finitely generated) ideals and
K. gradeA (a, J) ≥ ht a for every (finitely generated) ideal a of A. Then
K. gradeA⊲⊳ f J (ae , A ⊲⊳ f J) = K. gradeA (a, A) ≤ K. gradeA (a, J)
for any (finitely generated) ideal a of A.
Proof. Assume that a is a (finitely generated) ideal of A and let Σ be the class of all finitely
generated subideals of a. Then, as in the proof of Lemma 3.1, again, using [2, Proposition 2.2(iv)],
we have
K. gradeA⊲⊳ f J (ae , A ⊲⊳ f J)
= K. gradeA (a, A ⊲⊳ f J)
= sup{K. gradeA (b, A ⊲⊳ f J)|b ∈ Σ}
= sup{min{K. gradeA (b, A), K. gradeA (b, J)}|b ∈ Σ}
= sup{K. gradeA (b, A)|b ∈ Σ}
= K. gradeA (a, A).
The forth equality follows from [2, Lemma 3.2] and our assumption. This completes the proof.
The following lemma is a slight modification of [2, Lemma 3.2].
Lemma 3.3. Let a be an ideal of A and M be an A-module.
(i) Let A be quasi-local with the maximal ideal m. If K. gradeA (m, M) < ∞, then K. gradeA (m, M) ≤
dim A.
(ii) If, for every minimal prime ideal p over a, K. gradeA (pRp , Mp ) < ∞ (e.g. when M is finitely
generated), then K. gradeA (a, M) ≤ ht a.
Proof. (1) Using [16, Proposition 2.7], it is enough for us to show that Č. gradeA (m, M) ≤ dim A. In
order to prove this, assume that dim A < ∞ and let x be a finite sequence of elements in m. It follows
from [16, Proposition 2.4] that Č. gradeA (x, M) ≤ dim A. Therefore Č. gradeA (m, M) ≤ dim A. (2)
Notice, by [2, Proposition 2.2(iii)], that K. gradeA (a, M) < ∞. Then, by [2, Proposition 2.2(ii) and
(iii)], one may assume that A is quasi-local with the maximal ideal m. Now (1) completes the
proof.
4. M AIN
RESULTS
Assume that A is Noetherian local, and that J is contained in the Jacobson radical of B and it is
a finitely generated A-module. Recall that a finitely generated module M over A is called a maximal Cohen-Macaulay A-module if depthA M = dim A. Note that, in this circumstance, depthA M
equals the common length of the maximal M-regular sequences in the maximal ideal of A. In [22,
Corollary 2.5], it is shown that A ⊲⊳ f J is Cohen-Macaulay if and only if A is Cohen-Macaulay
and J is a maximal Cohen-Macaulay A-module. Our first main result improves this corollary by
removing the Noetherian assumption.
The reader should be aware that when we say A ⊲⊳ f J is Cohen-Macaulay in the sense of a
non-empty class of ideals, we mean A ⊲⊳ f J is Cohen-Macaulay as a ring.
COHEN-MACAULAY
5
Theorem 4.1. Assume that (A, m) is quasi-local such that m is finitely generated. Assume that J
is contained in the Jacobson radical of B and it is finitely generated as an A-module. Then A ⊲⊳ f J
is Cohen-Macaulay (ring) in the sense of maximal ideals if and only if A is Cohen-Macaulay in
the sense of maximal ideals and K. gradeA (m, J) = dim A.
Proof. Assume that m is generated by the sequence a = a1 , . . . , an and that J is generated by the
sequence b = b1 , . . . , bm . Hence m′ f = m ⊲⊳ f J, the unique maximal ideal of A ⊲⊳ f J [9, Corollary
2.7(3)], is generated by the sequence c = (a1 , f (a1 )), p
. . . , (an , f (an )), (0, bp
1 ), . . . , (0, bm ). Notice,
f
by [9, Corollary 3.2 and Remark 3.3], that one has ιA (a)(A ⊲⊳ J) = m(A ⊲⊳ f J) = m′ f =
c(A ⊲⊳ f J). Therefore
K. gradeA⊲⊳ f J (m′ f , A ⊲⊳ f J) = Č. gradeA⊲⊳ f J (m′ f , A ⊲⊳ f J)
= inf{i|Hci (A ⊲⊳ f J) 6= 0}
= inf{i|HιiA (a) (A ⊲⊳ f J) 6= 0}
= inf{i|Hai (A ⊲⊳ f J) 6= 0}
= inf{i|Hai (A) ⊕ Hai (J) 6= 0}
= min{Č. gradeA (m, A), Č. gradeA (m, J)}
= min{K. gradeA (m, A), K. gradeA (m, J)}.
The first equality obtains by p
[16, Proposition 2.7], the third equality follows from [16, Proposition
2.1(e)] in conjunction with ιA (a)(A ⊲⊳ f J) = c(A ⊲⊳ f J), the forth equality deduces from [16,
Proposition 2.1(f)], and the fifth equality holds since as an A-module A ⊲⊳ f J ∼
= A ⊕ J [7, Lemma
2.3(4)].
Consequently, the conclusion yields by the equality
K. gradeA⊲⊳ f J (m′ f , A ⊲⊳ f J) = min{K. gradeA (m, A), K. gradeA (m, J)}
together with dim A ⊲⊳ f J = dim A. This last equality holds true, since A ⊲⊳ f J is integral over A
(see [8, Proposition 4.2]).
Corollary 4.2. (See [22, Corollary 2.5]) Assume that A is Noetherian local, and that J is contained
in the Jacobson radical of B and it is finitely generated as an A-module. Then A ⊲⊳ f J is CohenMacaulay (ring) if and only if A is Cohen-Macaulay and J is a maximal Cohen-Macaulay Amodule.
The key to the next theorem is given by the following elementary lemmas. Their proofs are
straightforward; so that we omit them. Recall from [9, Corollary 2.5] that the prime ideals of
A ⊲⊳ f J are of the type q f or p′ f , for q varying in Spec(B)\V (J) and p in Spec(A), where
p′ f :=p ⊲⊳ f J := {(p, f (p) + j)|p ∈ p, j ∈ J},
q f :={(a, f (a) + j)|a ∈ A, j ∈ J, f (a) + j ∈ q}.
Lemma 4.3. Assume that a is an ideal of A, p is a prime ideal of A and that q is a prime ideal of
B. Then
(i) ae ⊆ p′ f if and only if a ⊆ p.
(ii) ae ⊆ q̄ f if and only if f (a) ⊆ q.
In the sequel, we use Nil(B) to denote the nil radical of the ring B.
Lemma 4.4. Assume that a is an ideal of A, J ⊆ Nil(B) and that p is a prime ideal of A. Then
6
Y. AZIMI, P. SAHANDI, AND N. SHIRMOHAMMADI
(i) p ∈ Min(a) if and only if p′ f ∈ Min(ae ).
(ii) ht a = ht ae .
(iii) Min(pe ) = {p′ f }. In particular ht pe = ht p′ f .
Proposition 4.5. Let A be a non-empty class of ideals of A. Assume that ht ae ≥ ht a for each
a ∈ A . If A ⊲⊳ f J is Cohen-Macaulay (ring) in the sense of A e := {ae |a ∈ A }, then A is CohenMacaulay in the sense of A and K. gradeA (a, J) ≥ ht a for each a ∈ A .
Proof. Assume that a ∈ A . Then, by Lemma 3.1(2), we have
K. gradeA (a, A) ≥ K. gradeA⊲⊳ f J (ae , A ⊲⊳ f J)
= ht ae
≥ ht a
≥ K. gradeA (a, A).
Thus K. gradeA (a, A) = ht a. This means that A is Cohen-Macaulay in the sense of A . Similarly,
one obtains K. gradeA (a, J) ≥ ht a.
It is not clear for us whether, in general, the inequality ht ae ≥ ht a holds for each a ∈ A .
However, under the assumption J ⊆ Nil(B), for each ideal a, one has the equality ht ae = ht a by
Lemma 4.4.
The second main result of the paper is the following theorem.
Theorem 4.6. Assume that J ⊆ Nil(B). Then A ⊲⊳ f J is Cohen-Macaulay (ring) in the sense of
ideals if and only if A is Cohen-Macaulay in the sense of ideals and K. gradeA (a, J) ≥ ht a for every
ideal a of A.
Proof. One implication follows from Proposition 4.5 and Lemma 4.4(2). Then, to prove the converse, assume that A is Cohen-Macaulay in the sense of ideals and K. gradeA (a, J) ≥ ht a for every
ideal a of A. Let a be an ideal of A. First observe, by Lemmas 3.2 and 4.4(2), that
K. gradeA⊲⊳ f J (ae , A ⊲⊳ f J) = K. gradeA (a, A)
= ht a
= ht ae .
Now, let I be an arbitrary proper ideal of A ⊲⊳ f J. Then, by [20, Theorem 16 of Chapter 5], there exists a prime ideal P of A ⊲⊳ f J containing I such that K. gradeA⊲⊳ f J (I, A ⊲⊳ f J) = K. gradeA⊲⊳ f J (P, A ⊲⊳ f
J). Notice that P = p′ f for some prime ideal p of A by [9, Corollaries 2.5 and 2.7]. Hence, by
Lemma 4.4(3), one has
ht I ≥ K. gradeA⊲⊳ f J (I, A ⊲⊳ f J)
= K. gradeA⊲⊳ f J (p′ f , A ⊲⊳ f J)
≥ K. gradeA⊲⊳ f J (pe , A ⊲⊳ f J)
= ht pe
= ht p′ f
≥ ht I.
Therefore A ⊲⊳ f J is Cohen-Macaulay in the sense of ideals.
The next example shows that, if, in the above theorem, the hypothesis J ⊆ Nil(B) is dropped,
then the corresponding statement is no longer always true.
COHEN-MACAULAY
7
Example 4.7. Let k be a field and X ,Y are algebraically independent indeterminates over k. Set
A := k[[X ]], B := k[[X ,Y ]] and let J := (X ,Y ). Let f : A → B be the inclusion. Note that A is CohenMacaulay and K. gradeA (a, J) = ht a for every ideal a of A. Indeed, if a is a non-zero proper ideal
of A, and a is a non-zero element of a, then one has
1 ≤ K. gradeA (aA, J) ≤ K. gradeA (a, J) ≤ htJ a ≤ ht a ≤ 1.
The first and second inequalities follow from [5, Proposition 9.1.2(a),(f)], respectively. While the
third inequality follows from Lemma 3.3(ii), the others are obvious. However, A ⊲⊳ f J which is
isomorphic to k[[X ,Y, Z]]/(Y, Z) ∩ (X −Y ) is not Cohen-Macaulay.
Let M be a A-module. Then A ⋉ M denotes the trivial extension of A by M. It should be
noted that 0 ⋉ M is an ideal in A ⋉ M and (0 ⋉ M)2 = 0. As in [7, Example 2.8], if B := A ⋉ M,
J := 0 ⋉ M, and f : A → B be the natural embedding, then A ⊲⊳ f J ∼
= A ⋉ M. Hence the next
result follows from Theorem 4.6. With it, we not only offer an application of Theorem 4.6, but we
also provide a generalization of the well-known characterization of when the trivial extension is
Cohen-Macaulay in the Noetherian (local) case, see [1, Corollary 4.14].
Corollary 4.8. Let M be an A-module. Then A ⋉ M is Cohen-Macaulay (ring) in the sense of
ideals if and only if A is Cohen-Macaulay in the sense of ideals and K. gradeA (a, M) ≥ ht a for
every ideal a of A.
Assume that A is Noetherian. In [22, Corollary 2.7], the authors showed that A is CohenMacaulay if A ⊲⊳ f J is Cohen-Macaulay provided that f −1 (q) 6= m for each q ∈ Spec(B)\V (J) and
each m ∈ Max(A). In the following corollary we improve the conclusion of the mentioned result
in the circumstance that J ⊆ Nil(B).
Assume that A is Noetherian and M is a finitely generated A-module. It can be seen that
ht a ≤ gradeA (a, M)(= K. gradeA (a, M)) for every ideal a of A if and only if Mp is maximal CohenMacaulay for every prime ideal p ∈ SuppA (M). Indeed, assume that Mp is maximal CohenMacaulay for every prime ideal p ∈ SuppA (M), and a is an ideal of R. There is nothing to
prove if aM = M, since in this case gradeA (a, M) = ∞. So assume that aM 6= M. Then using
[5, Proposition 1.2.10(a)], there is a prime ideal p containing a such that gradeA (a, M) = depth Mp .
Hence by assumption one has gradeA (a, M) = depth Mp = dim Rp = ht p ≥ ht a. To prove the
converse assume that p ∈ SuppA (M). Then again in view of [5, Proposition 1.2.10(a)], one has
dim Rp = ht p ≤ gradeA (p, M) ≤ depth Mp . Thus Mp is maximal Cohen-Macaulay.
Corollary 4.9. Assume that A is Noetherian, and that J ⊆ Nil B is finitely generated as an Amodule. Then A ⊲⊳ f J is Cohen-Macaulay if and only if A is Cohen-Macaulay and Jp is maximal
Cohen-Macaulay for every prime ideal p ∈ SuppA (J).
The next proposition provides other sufficient and necessary condition for A ⊲⊳ f J to be CohenMacaulay in the sense of ideals.
Proposition 4.10. With the notation and hypotheses of the beginning of Section 3, one has
(i) Let A be a non-empty class of ideals of A. Assume that ht f −1 (q) ≤ ht q for every q ∈
Spec(B)\V (J). If A ⊲⊳ f J is Cohen-Macaulay (ring) in the sense of A e := {ae |a ∈ A },
then A is Cohen-Macaulay in the sense of A and K. gradeA (a, J) ≥ ht a for every a ∈ A .
(ii) Assume that ht P ≤ ht P c for every P ∈ Spec(A ⊲⊳ f J), where the contraction P c is
given with respect to ιA . If A is Cohen-Macaulay in the sense of ideals and K. gradeA (a, J) ≥
ht a for every ideal a of A, then A ⊲⊳ f J is Cohen-Macaulay (ring) in the sense of ideals.
Proof. (1) Assume that A ⊲⊳ f J is Cohen-Macaulay ring in the sense of A e . In order to prove the
assertion, by Proposition 4.5, it is enough for us to show that ht ae ≥ ht a for each ideal a ∈ A . To
8
Y. AZIMI, P. SAHANDI, AND N. SHIRMOHAMMADI
this end, assume that a ∈ A and that P is a prime ideal of A ⊲⊳ f J containing ae . In view of [9,
Corollaries 2.5 and 2.7], one has the following three cases to consider.
Case 1. If P = p′ f for some prime ideal p of A such that f −1 (J) * p, then
ht P = ht p′ f = dim(A ⊲⊳ f J)p′ f = dim Ap = ht p ≥ ht a,
by [9, Proposition 2.9] and Lemma 4.3(1).
Case 2. If P = p′ f for some prime ideal p of A such that f −1 (J) ⊆ p, then
ht P = ht p′ f
= dim(A ⊲⊳ f J)p′ f
= dim(Ap ⊲⊳ fp JSp )
= max{dim Ap , dim( fp (Ap ) + JSp )}
≥ dim Ap
= ht p
≥ ht a,
by [9, Proposition 2.9], [8, Proposition 4.1] and Lemma 4.3(1), where Sp := f (A\p) + J.
Case 3. If P = q̄ f for some prime ideal q of B, then
ht P = ht q̄ f
= dim(A ⊲⊳ f J)q̄ f
= dim Bq
= ht q
≥ ht f −1 (q)
≥ ht a.
The third equality follows by [9, Proposition 2.9], the first inequality holds by assumption, and the
second one follows by Lemma 4.3. This completes the proof of the first assertion.
(2) Assume that A is Cohen-Macaulay in the sense of ideals and that K. gradeA (a, J) ≥ ht a for
every ideal a of A. As indicated by [2, Theorem 3.3], it is enough to show that
K. gradeA⊲⊳ f J (P, A ⊲⊳ f J) = ht P
for every prime ideal P of A ⊲⊳ f J. Let P be a prime ideal of A ⊲⊳ f J. Then
ht P ≤ ht P c
= K. gradeA (P c , A)
= K. gradeA⊲⊳ f J (P ce , A ⊲⊳ f J)
≤ K. gradeA⊲⊳ f J (P, A ⊲⊳ f J)
≤ ht P.
The first inequality holds by assumption, the second inequality is by [5, Proposition 9.1.2(f)], and
the last one is by Lemma 3.3(2), and the second equality follows from Lemma 3.2.
We are now in a position to present our third main result.
Theorem 4.11. With the notation and hypotheses of the beginning of Section 3, the following
statements hold:
COHEN-MACAULAY
9
(i) Let A be a non-empty class of ideals of A. Assume that the homomorphism f : A → B
satisfies the going-down property. If A ⊲⊳ f J is Cohen-Macaulay (ring) in the sense of
A e := {ae |a ∈ A }, then A is Cohen-Macaulay in the sense of A and K. gradeA (a, J) ≥
ht a for every a ∈ A .
(ii) Assume that ιA : A → A ⊲⊳ f J is an integral ring extension. If A is Cohen-Macaulay in the
sense of ideals and K. gradeA (a, J) ≥ ht a for every ideal a of A, then A ⊲⊳ f J is CohenMacaulay (ring) in the sense of ideals.
Proof. It is well-known that ht f −1 (q) ≤ ht q for every q ∈ Spec(B) if the homomorphism f : A → B
satisfies the going-down property by [18, Exercise 9.9]. In the light of Proposition 4.10, this proves
(1). To prove (2), keeping in mind Proposition 4.10, notice that, for every P ∈ Spec(A ⊲⊳ f J), the
inequality ht P ≤ ht P c holds since ιA : A → A ⊲⊳ f J is an integral ring extension [18, Exercise
9.8], where the contraction P c is given with respect to ιA .
Note that Example 4.7 also shows that we can not neglect the integral assumption in part two
of the above theorem.
Example 4.12.
(i) Assume that A is an integral domain with dim A ≤ 1 and that B is an
integral domain containing A. Assume that J is an ideal of B which is finitely generated
A-module. Hence, as in Example 4.7, one has K. gradeA (a, J) = ht a for every proper
ideal a of A. Notice that A is Cohen-Macaulay in the sense of ideals by [2, Page 2305].
Therefore one obtains that A ⊲⊳ f J is Cohen-Macaulay in the sense of ideals by Theorem
4.11.
(ii) To construct a concrete example for (1), set A := Q + X R[X ], where Q is the field of
rational numbers, R is the field of real numbers and X is an indeterminate over R. √
It is
easy to see that A is a one dimensional non integrally closed domain. Put B := A[ 2],
which is finitely generated as an A-module. Let J be a finitely generated ideal of B.
Consequently, by (1), A ⊲⊳ f J is Cohen-Macaulay in the sense of ideals.
(iii) Assume that A is a valuation domain, B an arbitrary integral domain containing A and that
J is an ideal of B. Then by [11, Corollary 4] and [12, Theorem 1], the inclusion homomorphism f : A ֒→ B satisfies the going-down property. Also notice, by [2, Proposition
3.12], that A is Cohen-Macaulay in the sense of ideals if and only if dim A ≤ 1. Further,
assume that dim A > 1. Then A ⊲⊳ f J can never be Cohen-Macaulay in the sense of ideals
by Theorem 4.11. In particular, the composite ring extensions A + X B[X ] and A + X B[[X ]]
can never be Cohen-Macaulay in the sense of ideals.
Note that if J is finitely generated as an A-module, then ιA : A → A ⊲⊳ f J is an integral ring
extension, and that, in this case, K. gradeA (a, J) ≤ ht a for every ideal a of A by Lemma 3.3. Hence
we can make the following corollaries right away.
Corollary 4.13. Assume that the homomorphism f : A → B satisfies the going-down property and
that J is finitely generated as an A-module. Then A ⊲⊳ f J is Cohen-Macaulay (ring) in the sense
of ideals if and only if A is Cohen-Macaulay in the sense of ideals and K. gradeA (a, J) = ht a for
every ideal a of A.
Corollary 4.14. Assume that f : A → B is a monomorphism of integral domains, and A is integrally
closed and that B is integral over A. Then A ⊲⊳ f J is Cohen-Macaulay (ring) in the sense of ideals
if and only if A is Cohen-Macaulay in the sense of ideals and K. gradeA (a, J) ≥ ht a for every ideal
a of A.
Proof. By [18, Theorem 9.4], f : A → B satisfies the going-down property. Also, ιA : A → A ⊲⊳ f J
is an integral ring extension by assumption and [8, Lemma 3.6].
10
Y. AZIMI, P. SAHANDI, AND N. SHIRMOHAMMADI
Corollary 4.15. Assume that f : A → B is a flat and integral homomorphism. Then A ⊲⊳ f J is
Cohen-Macaulay (ring) in the sense of ideals if and only if A is Cohen-Macaulay in the sense of
ideals and K. gradeA (a, J) ≥ ht a for every ideal a of A.
Proof. By [18, Theorem 9.5], f : A → B satisfies the going-down property. Also, ιA : A → A ⊲⊳ f J
is an integral ring extension by assumption and [8, Lemma 3.6].
In concluding, we apply Corollary 4.15 on amalgamated duplication. Recall that if f := idA
is the identity homomorphism on A, and J is an ideal of A, then A ⊲⊳ J := A ⊲⊳idA J is called the
amalgamated duplication of A along J. Assume that (A, m) is Noetherian local. In [6, Discussion
10], assuming that A is Cohen-Macaulay, D’Anna showed that A ⊲⊳ J is Cohen-Macaulay if and
only if J is maximal Cohen-Macaulay. Next in [21, Corollary 2.7], the authors improved D’Anna’s
result as A ⊲⊳ J is Cohen-Macaulay if and only if A is Cohen-Macaulay and J is maximal CohenMacaulay. Our final corollary generalizes these results.
Corollary 4.16. Let J be an ideal of A. Then A ⊲⊳ J is Cohen-Macaulay (ring) in the sense of
ideals if and only if A is Cohen-Macaulay in the sense of ideals and K. gradeA (a, J) ≥ ht a for
every ideal a of A.
Proof. This immediately follows from Corollary 4.15, since f = idA : A → A is flat and integral.
Acknowledgements. The authors is deeply grateful to the referee for a very careful reading of
the manuscript and many valuable suggestions.
R EFERENCES
[1] D. D. Anderson, M. Winders, Idealization of a module, J. Commut. Algebra, 1, (2009), 3–56.
[2] M. Asgharzadeh and M. Tousi, On the notion of Cohen-Macaulayness for non-Noetherian rings, J. Algebra, 322,
(2009), 2297–2320.
[3] A. Bagheri, M. Salimi, E. Tavasoli and S. Yassemi, A construction of quasi-Gorenstein rings, J. Algebra Appl. 11,
No. 1, (2012), 1250013, (9 pages).
[4] M. P. Brodmann and R. Y. Sharp, Local Cohomology: An Algebraic Introduction with Geometric Applications,
Cambridge Studies in Advanced Mathematics, 136, Cambridge University Press, Cambridge, 2013.
[5] W. Bruns and J. Herzog, Cohen-Macaulay rings. Rev. ed. Cambridge Studies in Advanced Mathematics 39, Cambridge, Cambridge University Press 1998.
[6] M. D’Anna, A construction of Gorenstein rings, J. Algebra, 306, (2006), 507–519.
[7] M. D’Anna, C. A. Finocchiaro, and M. Fontana, Amalgamated algebras along an ideal, in: Commutative Algebra
and Applications, Proceedings of the Fifth International Fez Conference on Commutative Algebra and Applications, Fez, Morocco, 2008, W. de Gruyter Publisher, Berlin, 2009, pp. 155–172.
[8] M. D’Anna, C. A. Finocchiaro, and M. Fontana, Properties of chains of prime ideals in an amalgamated algebra
along an ideal, J. Pure Appl. Algebra, 214, (2010), 1633–1641.
[9] M. D’Anna, C. A. Finocchiaro, and M. Fontana, New algebraic properties of an amalgamated algebra along an
ideal, Commun. Alg. 44, (2016), 1836–1851.
[10] M. D’Anna and M. Fontana, An amalgamated duplication of a ring along an ideal: the basic properties, J. Algebra
Appl. 6, No.3, (2007), 443–459.
[11] D. E. Dobbs, On going-down for simple overrings, Proc. Amer. Math. Soc. 39, (1973), 515–519.
[12] D. E. Dobbs and I. J. Papick, On going-down for simple overrings III, Proc. Amer. Math. Soc. 54, (1976), 35–38.
[13] S. Glaz, Coherence regularity and homological dimensions of commutative fixed rings, in: Ngô Viêt Trung, Aron
Simis, Guiseppe Valla (Eds.), Commutative Algebra, World Scientific, Singapore, 1992, pp. 89–106.
[14] S. Glaz, Homological dimensions of localizations of polynomial rings, in: Zero-Dimensional Commutative Rings,
Knoxville, TN, 1994, in: Lect. Notes Pure Appl. Math., vol. 171, Marcel Dekker, New York, 1995, pp. 209–222.
[15] T. D. Hamilton, Weak Bourbaki unmixed rings: A step towards non-Noetherian Cohen-Macaulayness, Rocky
Mountain J. Math. 34, (2004), 963–977.
[16] T. D. Hamilton and T. Marley, Non-Noetherian Cohen-Macaulay rings, J. Algebra, 307, (2007), 343–360.
[17] J. Huckaba, Commutative Rings with Zero Divisors, M. Dekker, New York, 1988.
COHEN-MACAULAY
11
[18] H. Matsumura, Commutative Ring Theory, Cambridge Stud. Adv. Math., vol. 8, Cambridge University Press,
1986.
[19] M. Nagata, Local Rings, Interscience, New York, 1962.
[20] D. G. Northcott, Finite Free Resolutions, Cambridge Tracts in Math., vol. 71, 1976.
[21] P. Sahandi and N. Shirmohammadi, Notes on amalgamated duplication of a ring along an ideal, Bull. Iranian
Math. Soc. 41, (2015), 749–757.
[22] P. Sahandi, N. Shirmohammadi and S. Sohrabi, Cohen-Macaulay and Gorenstein properties under the amalgamated construction, Commun. Alg. 44, (2016), 1096–1109.
D EPARTMENT OF M ATHEMATICS , U NIVERSITY OF TABRIZ , TABRIZ , I RAN .
E-mail address: [email protected]
D EPARTMENT OF M ATHEMATICS , U NIVERSITY OF TABRIZ , TABRIZ , I RAN .
E-mail address: [email protected]
D EPARTMENT OF M ATHEMATICS , U NIVERSITY OF TABRIZ , TABRIZ , I RAN .
E-mail address: [email protected], [email protected]
| 0math.AC
|
Subgraph Counting: Color Coding Beyond Trees
Venkatesan T. Chakaravarthy1 , Michael Kapralov2 , Prakash Murali1 ,
Fabrizio Petrini3 , Xinyu Que3 , Yogish Sabharwal1 , and Baruch Schieber3
arXiv:1602.04478v2 [cs.DC] 2 Apr 2016
1,3
1
IBM Research
{vechakra, prakmura, ysabharwal}@in.ibm.com
3
{fpetrin, xque, sbar}@us.ibm.com
2
EPFL
2
[email protected]
April 5, 2016
Abstract
The problem of counting occurrences of query graphs in a large data graph, known
as subgraph counting, is fundamental to several domains such as genomics and social
network analysis. Many important special cases (e.g. triangle counting) have received
significant attention. Color coding is a very general and powerful algorithmic technique for subgraph counting. Color coding has been shown to be effective in several
applications, but scalable implementations are only known for the special case of tree
queries (i.e. queries of treewidth one).
In this paper we present the first efficient distributed implementation for color
coding that goes beyond tree queries: our algorithm applies to any query graph of
treewidth 2. Since tree queries can be solved in time linear in the size of the data
graph, our contribution is the first step into the realm of colour coding for queries that
require superlinear running time in the worst case. This superlinear complexity leads
to significant load balancing problems on graphs with heavy tailed degree distributions. Our algorithm structures the computation to work around high degree nodes
in the data graph, and achieves very good runtime and scalability on a diverse collection of data and query graph pairs as a result. We also provide theoretical analysis of
our algorithmic techniques, showing asymptotic improvements in runtime on random
graphs with power law degree distributions, a popular model for real world graphs.
1
Introduction
Graphs serve as common abstractions for real world data, making graph mining primitives
a critical tool for analyzing real-world networks. Counting the number of occurrences of a
query graph in a large data graph (subgraph counting, often referred to as motif counting)
is an important problem with applications in a variety of domains such as bioinformatics,
social sciences and spam detection (e.g. [8, 10, 23]).
Subgraph counting and its variants have received a lot of attention in the literature.
Substantial progress has been achieved for the case of small queries such as triangles or
1
Figure 1: Illustration of a match (left) and a colorful match (right)
4-vertex subgraphs: not only have very efficient algorithms been developed (e.g. [15, 20,
27, 31]), but also theoretical explanation of their performance on popular graph models
has been obtained (see [?] and references therein).
Some of the recent work has addressed larger queries [29, 30, ?, 26, 7], but our understanding here is far from complete. Even for reasonably large graphs (a million edges)
and small queries (e.g. 5-cycles), the number of solutions tend to be enormous, running
into billions. This explosion in the search space makes the subgraph counting problem
very hard even for moderately large queries. Theoretically, the fastest known algorithm
for counting occurrences of a k-vertex subgraph in an n-vertex data graph runs in time
nωk/3 , where O(nω ) is the time complexity of matrix multiplication (currently ω ≈ 2.38).
This improves upon the trivial algorithm with runtime nk , but is prohibitively expensive
even for moderate size queries.
To address the above issue, Alon et al. [2] proposed the color coding technique. Here,
given a k-node query, we assign random colors between 1 and k to the vertices of the
data graph, and count the number of occurrences of the query that are colorful, meaning
the vertices matched to the query have distinct colors. See Figure 1. The count is scaled
up appropriately to get an estimate on the actual number of occurances. The accuracy is
then improved by repeating the process over multiple random colorings and taking the
average. Restricting the search to colorful matches leads to pruning of the search space
and improved efficiency. Using this method, Alon et al. obtained faster algorithms for
cetain queries such as paths, cycles, trees and bounded treewidth graphs.
The power of color coding as a very general counting technique together with the importance of subgraph counting in various applications (as mentioned above) makes it important to design practically efficient and scalable implementations. In a different work,
Alon et al. [1] applied the color coding technique for counting the occurrences of treelets
(tree queries) in biological networks. Color coding allowed them to handle tree queries
up to size 10 in protein interaction networks, extending beyond the reach of previously
known approaches [25, 18, 17]. Recently, Slota and Madduri [28, 30] presented FASCIA, an
efficient and scalable distributed implementation of subgraph counting (via color coding),
again for the case of treelet queries. However, despite considerable interest in non-tree
queries from several application domains (see the experimental section for details), the
technque has not been explored for more general settings. In this work we present the first
efficient distributed implementation of color coding beyond tree queries.
2
As part of their original color coding solution, Alon et al. [2] presented faster algorithms
for certain special classes of queries. They showed that if the query is a tree, then colorful
subgraph counting can be solved in time O(2k m), i.e. in time linear in the size of the data
graph. They extended the algorithm to show that if the query is close to a tree, specifically
has (small) treewidth t, a running time of O(2k nt+1 ) can be achieved. Treewidth [9] is a
widely adopted measure of the intrinsic complexity of a graph. Intuitively, it measures
how close the topology of a given graph is to being a tree: tree queries have treewidth 1,
and a cycle is the simplest example of a treewidth 2 query. The above algorithm, restricted
to trees, forms the basis for the previously-mentioned treelet counting implementations
[28, 30, 1].
While the runtime of the above algorithm is linear for the case of trees (i.e. acyclic
queries), it becomes at least quadratic for query graphs of treewidth 2 and beyond. This
phenomenon also manifests itself in practice: on real world graphs with even moderately
skewed degree distribution load imbalance is observed and the running time tends to have
quadratic dependence on the maximum degree of the graph. Thus, even triangles (the
smallest cyclic query) are harder to handle, and have received considerable attention from
the research community (as mentioned earlier).
The goal of this paper is to study the colorful subgraph counting problem on queries of
treewidth 2, taking the first step in the realm of color coding with cyclic queries. The class of
queries of treewidth 2 is quite rich. In particular, it contains all trees, cycles, series-parallel
graphs and beyond. Figure 8 shows treewidth 2 queries (used in our experimental evaluation) drawn from real-world studies on biological, social and collaboration networks
[22, 32, 4].
To the best of our knowledge, the previously-mentioned algorithm [1] is the best known
algorithm for treewidth 2 queries, and we use it as our baseline. We rephrase this algorithm
within our framework and devise a distributed implementation. The rephrased algorithm
becomes a recursive procedure that decomposes the query into simpler path subqueries,
which are then solved to get the overall count. We thus refer to our baseline as the Path
Splitting algorithm (PS).
Our Contributions
1. Building on the PS algorithm, we develop novel strategies that lead to significant
performance gains in terms of runtime, scalability, and the size of graphs and queries handled.
2. Our algorithm works by decomposing the query to cycles and leaves, thereby reducing the problem of colorful subgraph counting on treewidth 2 queries to counting (annotated) cycles.
3. The decomposition in terms of cycles enables us to exploit the so-called degree ordering approach (e.g., M IN B UCKET algorithm for triangle enumeration [?]) Specifically, we
show how to force the computation process to (mostly) work around high degree vertices,
leading to substantial speedups and scalability gains.
4. We present a detailed experimental evaluation of the algorithms on real-world graphs
having more than million edges and real-world queries of size up to 10 nodes. The results
show that our strategies offer improvements of up to 28x in terms of running time and
3
exhibit improved scalability.
5. Finally, we complement our experimental evalutation by a theoretical analysis of the
runtime of our degree ordering approach for cycle queries, on a popular class of random
power law graphs (Chung-Lu graphs [14]). Our analysis provides justification for empirically observed performance gains of the approach.
Related Work
Subgraph counting has received significant attention in the fields of computational biology [25, 18, 17] and social network analysis [21, 13, 27, ?, 20]. We give an overview of prior
work on the problem (both theoretical and empirical) as well as techniques for making
subgraph counting scalable, and explain how our contributions relate to this prior work.
Color Coding and Approximate Subgraph Counting: Color coding was introduced in an
influential paper by Alon et al. [2] as a fast algorithm for finding occurrences of a query
in a data graph and counting the number of such occurrences. In a different work, Alon
et al. [1] explored its applications to approximate subgraph counting (most commonly
known as motif counting) in computational biology. They were motivated by the fact that
subgraph counting is an important primitive for characterizing biological networks [?].
Color coding allowed Alon et al. to count occurrences of treelets (tree queries) up to
size 10 in protein interaction networks, extending beyond the reach of previously known
approaches [25, 18, 17]. A scalable distributed implementation of color coding for trees
has been reported by Slota and Madduri [29, 30], but no principled solutions beyond tree
queries are known. ParSE [33] extends beyond tree queries, by considering query graphs
that can partitioned into subtemplates via edge cuts of size 1. However, the only class of
query graphs that can be perfectly partitioned using this method is trees; ParSE resorts to
brute force enumeration for other cases. Our work provides the first principled approach
to implementing color coding in a scalable way beyond trees queries. Further, our analysis
of the runtime of our cycle counting subroutine on a random graphs with a power law
degree distribution provides a theoretical justification of our algorithmic techniques.
While our work and the above-mentioned prior work [1, 29, 30] count non-induced sugraphs, some other prior work [25, 18, 17] addressed the case of counting induced subgraphs.
The search space of non-induced subgraphs is larger and furthermore, these counts are
more robust with respect to perturbations of the data graph [1].
Degree Based Approaches: Designing scalable subgraph counting algorithms turns out
to be hard even for the simple case of triangle counting. A naive approach lets each vertex enumerate pairs of neighbors and check if they are connected. This leads to wasteful
computations and also runs into load balancing issues on graphs with heavy tailed degree
distributions [31]. The above issue has been addressed using a simple, but efficient solution (referred to as the M IN B UCKET algorithm [15, 31]): each vertex enumerates pairs of
neighbors with degree no smaller than its own (with arbitrary tie breaking) and checks
they are connected. It is not hard to see that this gives a correct count, and it has been
empirically observed that this algorithm does not run into load balancing issues even on
heavy tailed graphs [31]. The M IN B UCKET heuristic has also been shown to give polynomial runtime improvement over the naive method when the input is a random graph
with a power law degree distribution [?]. A recent work by Jha et al. [20] applies the
4
degree based technique for coutning 4-vertex queires. There are a few prior approaches
for arbitrary queries of [7, 3, 26], but algorithms do not use degree information, and are
comparable to the baseline algorithm used in our study.
To the best of our knowledge, prior to our work there has not been a systematic study of
how M IN B UCKET generalizes to larger subgraph counting problems. In this work we generalize the method for counting occurrences of treewidth 2 graphs, perform a thourough
experimental evaluation and provide a theoretical runtime analysis of our technique in the
random power law graph model. Our paper improves upon prior work along three axes:
generality of queries handled, scalability of the proposed solution and theoretical analysis of the main algorithmic primitive on a class of graphs often used to model real world
networks.
2
Preliminaries
Subgraph counting problem. The subgraph counting problem is defined as follows. The
input consists of a query graph Q = (VQ , EQ ) over a set of k nodes and a data graph G =
(VG , EG ) over a set of n vertices and m edges. The task is to count the number of (not
necessarily induced) subgraphs of G that are isomorphic to Q. Formally, count the number
of injective mappings π : VQ → VG such that for any pair of query nodes a, b ∈ VQ , if
hq1 , q2 i ∈ EQ , then hπ(q1 ), π(q2 )i ∈ EG . We refer to such mappings π as matches.
Color coding and colorful matches. A coloring is a function χ : VG → {1, 2, . . . , k}, where
for every vertex u ∈ VG , χ(u) denotes its color. A match π from VQ toSVG is colorful if
the vertices of Q are mapped to k distinctly colored vertices in G, i.e. a∈VQ χ(π(a)) =
{1, 2, 3, . . . , k}. The main idea is that instead of counting all possible matches of the k
vertices of the query graph to the vertices of the data graph, one first colors the vertices of
the data graph uniformly at random using k colors, and then searches for colorful matches.
Colorful subgraph counting problem. In the colorful subgraph counting problem the task is
to count the number of colorful matches of the query Q in VG .
Our setting counts the number of colorful matches or mappings from Q to the data
vertices. Alternatively, we may want to count the number of colorful subgraphs that are
isomorphic to Q. The latter quantity can be obtained by dividing the former by aut(Q), the
number of automorphisms of Q. While it is computationally hard to compute aut(Q) for
an arbitrary query graph, the quantity can be computed quickly for queries of relatively
small size (say about 10 nodes). Given the above discussion, we focus on counting the
number of colorful matches.
Treewidth. Intuitively, if the query graph Q = (VQ , EQ ) has treewidth t then Q can be
decomposed into subgraphs Q1 , Q2 , . . . such that each subgraph Qi is also of treewidth t,
and each Qi has no more than t nodes that belong also to other subgraphs. We call such
nodes the boundary nodes of Qi . In addition, the total number of distinct boundary nodes
in all subgraphs Q1 , Q2 , . . . is at most t + 1. Note that the decomposition can be done
recursively as each Qi has treewidth t, until we are left only with subgraphs that have
at most t + 1 nodes. This results in a treewidth decomposition tree denoted TQ . A formal
definition is givne below.
A tree decompsition of a query |Q| is a tree T = (VT , ET ), wherein each node p ∈ VT is
5
associated with a subset of query nodes S(p) ⊆ VQ , called pieces, such that the following
properties are true: (i) for every query edge (a, b) ∈ EQ , there exists a piece S(p) (for some
p ∈ VT ) that contains both a and b; (ii) for every query node a ∈ VQ , the set of nodes
whose pieces contain a induce a connected subtree. Alternatively, the second property
states that if a belongs to pieces S(p1 ) and S(p2 ) for some p1 and p2 , then a must also
belong to the piece S(p) for any node p found on the (unique) path connecting p1 and p2 in
T . The width of the tree decomposition is the maximum cardinality overall pieces minus
one, i.e., maxp |S(p)| − 1. The treewidth t of the query is the minimum width over all its
tree decompositions.
Approximate subgraph counting via color coding. Counting the number of colorful
matches turns out to be easier than counting the actual (not necessarily colorful) matches.
The price to pay is that the algorithm is randomized. We color the graph randomly and
obtain the number of colorful matches, and repeat the process independently at random a
few times. Then, an estimate for the number of matches (occurances of the query) can be
obtained by taking the average.
For a given input graph G and query Q let n(G, Q) denote the number of matches π
from Q to G. For a (random) coloring χ of vertices of G let ncolorf ul (G, Q, χ) denote the
number of colorful matches of Q to G under coloring χ. It was shown [2, 1] that with
proper normalization the colorful count ncolorf ul (G, Q, χ) is an unbiased estimator of the
actual count. Specifically, the right normalization factor is k k /k!, i.e. we have (k k /k!) ·
Eχ [ncolorf ul (G, Q, χ)] = n(G, Q). The variance of the estimator can also be bounded (see [1],
section 2.1). Thus, taking the average of ncolorf ul (G, Q, χ) under a few independently chosen colorings χ converges to the right answer, i.e. n(G, Q). Thus, in order to obtain an
approximate subgraph counting algorithm it suffices to solve the colorful subgraph counting problem. The rest of the paper is devoted to designing a scalable solution to colorful
subgraph counting.
3
Overview
The work of Alon et al. [2] yields a natural algorithm for the colorful subgraph counting
problem on bounded treewidth query graphs. This algorithm is based on the following
intuition. Suppose that we have found a colorful match π for a subgraph Q of the input
query graph Q, and we wish to extend it into a colorful match π 0 for Q by additionally
fixing the mapping of the nodes outside Q. For this we do not need to know the mapping of
the non-boundary nodes of Q, since they do not share edges with nodes outside Q. Instead,
it suffices to know the mapping of the boundary nodes (i.e., the nodes that share edges with
nodes outside Q) and the set of colors used by π. The mapping of the boundary nodes is
needed to ensure that for any edge from a boundary node to outside, the corresponding
data vertices share an edge in the data graph; and the set of colors is needed to avoid
repeating a color already used by π. Analogously, in the setting of counting, in order to
count the number of colorful matches for Q, we do not need a complete listing of colorful
matches of Q. Instead, we can group the colorful matches based on the set of colors used
and the mappings for the boundary nodes and it suffices to know the count per group.
Based on the above intuition, we apply dynamic programming to count the number
6
colorful matches of Q. Let TQ be the tree decomposition of Q with treewith t. The algorithm processes TQ in a bottom-up manner and a creates a hash table (that we call a
projection table) for each tree node. The subgraph graph Q associated with a node has at
most t boundary nodes and these nodes can be mapped to the data vertices in at most nt
ways. In addition, we need to record the colors of the data vertices to which the nodes
of Q are mapped. Since we focus
on colorful matches, the set of colors used (that we call
“signature”) can be at most kt ≤ 2k (where k is the size of the query graph). For each
combination of mappings to the boundary nodes and the signature, we record the number
of colorful matches of Q consistent with the combination. The number of entries in the
table is at most nt 2k . The projection table for a tree node can be computed from those of
its children. We get the total number of colorful matches by performing an aggregation on
the projection table of the root node.
Working in the realm of motif counting, Slota and Madduri [30] described an efficient
distributed implementation of the above algorithm for the case of tree queries and presented an experimental evaluation. Trees have treewidth one hence, the size of projection
tables is linear in the number of vertices and the overall computation can be carried out in
time linear in the graph size. Our goal is to address a more general class of queries (beyond
trees) in a distributed setting and we focus on the case of queries of treewidth 2. Treewidth
2 queries are more challenging since in the worst case, the tables can be of size quadratic
in the number vertices and the computation time also gets quadratic.
The construcion of our algorithm is motivated by the fact that real life data graphs
tend to exhibit variations in the degree distribution. A naive implementation that treats all
data vertices in the same manner would result in a lot of entries in the projection tables of
the high degree vertices that do not lead to colorful matches for the overall input query.
Moreover, in a distributed setting the processors owning such vertices perform more computation leading to load imbalance.
Our algorithm is based on a crucial observation that any treewidth 2 query can be recursively decomposed into (annotated) cycles or leaves. The core component of the algorithm
is an efficient procedure for handling cycles that employs a strategy based on degree based
ordering of vertices. This leads to reduction in wasteful computation, as well as improved
load balancing. The procedure is inspired by a similar strategy used in prior work [?] for
handling triangles. The overall algorithm uses the above decomposition and the improved
procedure for handling cycles.
4
Overall Algorithm
In this section we describe the overall structure of our subgraph counting algorithm that
proceeds in two steps. In the first step, we decompose the query into cycles and leaves
(called blocks) and construct a decomposition tree for the input query Q which is essentially
a carefully chosen treewidth decomposition tree; each node of the tree represents a block
and encodes a convenient subquery. This step is independent of the data graph and can
be viewed as a preprocessing phase for the query. Then in the second step we traverse the
tree in a bottom up manner, performing primitive counting operations over the data graph
prescribed by the internal nodes and combining the results. The final count is produced
7
by the root of the tree.
4.1
Decomposition Tree
For an input query graph Q = (VQ , EQ ), construct the decomposition tree T (Q) by iteratively applying one of two primitive operations: contraction of a leaf edge or a cycle. As
these operations are applied the number of nodes in the query Q decreases. At the same
time new edges may appear in Q to represent contracted structures, and edges as well as
nodes may get annotated with the identity of the contracted structures that they represent.
Before defining the tree construction algorithm we need to introduce two definitions. First,
we say that a cycle C in Q is contractible if (a) C = (a0 , a1 , . . . , aL−1 ) is induced (i.e. there are
no edges between nodes a0 , a1 , . . . , aL−1 except the edges of C) and (b) cycle C has most
two boundary nodes (i.e., nodes that share edges with nodes outside of C). Second, a leaf
edge is an edge L = (a, b), where b is a leaf node (has degree one); a is called the boundary
node of the leaf edge. We use the common term block to refer to leaf edges and contractible
cycles.
For example, consider the query named Satellite in Fig 2. The cycle (i, j, k) is contractible with a single boundary node i, the cycle (a, b, c, d, e) is contractible with two
boundary nodes a and c, and (f, h) is a leaf edge. The cycle (i, f, g) is not contractible
since it has three boundary nodes.
We construct the decomposition tree T (Q) starting with an empty tree. The tree is
built bottom-up starting from the leaf level and hence, the structure may be a forest with
multiple roots in the intermediate stages. Each iteration adds a new node and may make
some of the existing roots as its children, culminating in a tree.
In the construction process we iteratively perform the following operations until Q
contains a single node: find a block B (a leaf edge or a contractible cycle) in Q and remove
it from Q (while possibly adding an edge to Q), and add a corresponding node to T (Q).
We iterate until Q contains a single node. We distinguish 3 cases.
Case 1: B is a contractible cycle C with exactly one boundary node a ∈ VQ : Remove the nodes
and edges of C from Q, except for node a. Erase any annotation found on a in Q and
annotate it with the block name B.
Case 2: B is a contractible cycle C with two boundary nodes a, b ∈ VQ : Remove the nodes and
edges of C from Q, except for the nodes a and b. Add an edge (a, b) in Q and annotate it
with B. Erase any annotation found on a and b in Q.
Case 3: B is a leaf edge L = (a, b): Remove b and the edge from Q. Erase any annotation
found on node a ∈ Q and annotate it with the block name B.
The nodes and edges of B inherit the annotations from Q, as they were before Q was
transformed (this ensures that the annotations on the boundary nodes that got erased get
captured by the new annotation).
Next we add a new node B to the tree T (Q). If any node or edge in B has an annotation B 0 , make B 0 a child of B in T (Q). This completes the construction of T . We show
below that the process can find a block in each iteration and terminate successfully on every query of treewidth 2. Assuming termination, it is not difficult to see that the process
produces a tree. During contraction, every block B 0 annotates a particular node or an edge
of Q, recording the way in which it has been contracted. The annotation gets inherited by
8
Figure 2: Illustration of the decomposition process. the top row shows the sequence of
queries considered in the process (the original query is on the left), the bottom row shows
the blocks that were contracted in each step.
some other block B in a subsequent iteration. The block B becomes the parent of B 0 . The
annotation is erased in Q, ensuring that no other block becomes a parent of B 0 .
Taking Satellite as the input query Q, Figure 2 provides an illustration process, along
with the output decomposition tree. The bottom row shows the blocks being contracted
and the top row shows the transformed Q. The first iteration contracts the cycle B1 =
(a, b, c, d, e). A new edge (a, c) is added to Q, along with the annotation B1 , and B1 is added
to the tree. The second iteration contracts the leaf block B2 = (f, h). Node f is annotated as
B2 and the B2 is added to the tree. The third iteration contracts B3 = (a, f, g, c), by adding
an edge (f, g) with the annotation B3 . The block is added to the tree and it is made the
parent of B1 and B2 . In the fourth iteration, the cycle B4 = (i, j, k) is contracted. Node i
gets annotated as B4 and B4 is added to the tree. Finally, the query Q4 is contracted leaving
Q empty. We add Q4 as the root of the tree, making it the parent of B3 and B4 .
The following lemma guarantees that for any treewidth 2 query Q, the tree construction
procedure will always find a block (a leaf edge or a contractible cycle) in each iteration and
terminate successfully. The proof relies on prior work on nested ear decompositions of
treewidth 2 queries [16].
9
Lemma 4.1 (i) Any treewidth 2 query Q contains a block; (ii) the transformed query resulting
from the contraction process is also a treewidth 2 query.
Proof: We first prove part (ii) of the lemma. If the contracted block has one boundary
node then no new edges are added to Q, in which case the tree TQ for the updated Q
is given by deleting all the nodes not in the updated VQ from the subsets SQ (t). If the
contracted block has two boundary nodes a and b then the edge (a, b) is added to Q. In
this case we get the tree for the updated Q by replacing each occurrence of the nodes not
in the updated VQ by b. Note that the size of each subset is still at most 3, nodes associated
with subsets that contain b form a connected component, and for at least one subset SQ (t),
{a, b} ⊆ SQ (t).
We now prove part (i). First, Root the tree TQ at an arbitrary non-leaf node. This induces an ancestor-descendant relationship on the nodes in VT . Note that if there are two
nodes {t, t0 } ⊆ VT , such that SQ (t0 ) ⊆ SQ (t), node t0 can be omitted and all its children connected to t. Thus from now on we assume that no subset SQ (t) is contained (or identical)
to another subset.
We need the following definition and claim.
Definition 4.1 For a node t ∈ VT , let Qt be the subgraph of Q induced by the nodes that are in
the union of the subsets associated with the nodes of TQ in the subtree rooted at t.
Claim 4.1 For every node t ∈ VT , either Qt contains a block, or Qt is a path whose endpoints are
in the subset associated with the parent of t (if such exists).
Before proving the claim we show how it implies the lemma. Since the claim holds also for
the root of TQ then either Q contains a block or it is a path in which case it also contains a
leaf block.
Proof of Claim 4.1: We prove the claim by induction. The base of the induction is a leaf
node. Consider a leaf node t ∈ VT . There are two possibilities: (i) SQ (t) = {x, y}, and (ii)
SQ (t) = {x, y, z}. If SQ (t) = {x, y}, then at least one node, say y, is only connected to x
and thus (x, y) is a leaf edge.
If SQ (t) = {x, y, z}, then consider the subgraph induced by {x, y, z}. If this subgraph
is a triangle then it must be a contractible cycle. The only remaining case is the subgraph
induced by {x, y, z} forms a path. Assume that the endpoints of this path are x and z. If
one of these endpoints, say z, is not in the subset associated with the parent of t then (y, z)
is a leaf edge. Otherwise, let t0 be the parent of t, we have SQ (t) ∩ SQ (t0 ) = {x, z}.
For the inductive step consider a non-leaf node t ∈ VT . If Qt0 for any child t0 of t
contains a block then we are done. Assume that this is not the case. Consider first the case
that t has a single child t0 . By the inductive hypothesis Qt0 is a path whose endpoints x and
y are in SQ (t). Let SQ (t) = {x, y, z}. If z is connected to both x and y then the cycle closed
by z is a contractible cycle. If z is connected to only one endpoint, say y, then we get a path
with endpoints x and z. If either x or z are not in the subset associated with the parent of
t, then the missing endpoint is leaf node. If both x and z are in the subset associated with
the parent of t then the inductive claim follows.
Next, Consider the case that t has several children. If two of the children of t, say t0 and
t00 , share endpoints then the cycle formed by Qt0 and Qt00 is contractible. Otherwise, t must
have exactly two children, say t0 and t00 , with endpoint {x, y} and {y, z}, forming a path
with endpoints x and z. If z is connected also to x then the cycle closed by the edge (x, z)
10
is a contractible cycle. If either x or z are not in the subset associated with the parent of t,
then the missing endpoint is a leaf node. If both x and z are in the subset associated with
the parent of t then the inductive claim follows.
An input query may admit multiple decomposition trees and the choice of the tree influences the performance of our algorithm. In Section 6, we present a heuristic for finding
a good decomposition. Each node of the tree represents a block and it will be convenient
to view to the node simply as the block represented by it.
At this point, it is interesting to consider tree queries studied by Slota and Madduri [30].
Given a tree query, their algorithm fixes a suitable query node as the root and iteratively
processes the tree in a bottom-up manner. The algorithm implicitly uses a decomposition
tree. However, since trees do not have cycles, the decomposition tree consists of only leaf
edge blocks. In contrast, the decomposition trees of treewidth two queries involve the
more challenging case of cycles as well.
4.2
Tree Traversal
Here, we describe the second step of the algorithm that traverses the decomposition tree
in a bottom-up manner and computes the number of colorful matches of the blocks in the
data graph. For this purpose, we define the notion of subqueries represented by blocks.
A subquery Q of the input query Q refers to any induced subgraph of Q. Consider a
block B and let U be the union of nodes found in the block B and its descendant blocks
in the tree. The subquery represented by B, denoted SQ(B), refers to the subquery induced
by U . For example, Figure 2 shows the subquery represented by the block B4 . The decomposition tree yields a nested hierarchy of subqueries: the root block represents the whole
input query and for any block B with the parent B 0 , the subquery SQ(B) is contained
within SQ(B 0 ).
Let B be a block. A node a ∈ SQ(B) is said to be a boundary node, if a shares an edge
with a node outside SQ(B). It is not hard to see that these boundary nodes are the same
as the boundary nodes of B (identified during the tree construction process). Thus, SQ(B)
can have at most two boundary nodes.
Before describing the counting algorithm we extend the notion of colorful matches
to subqueries: a colorful match for a subquery Q = (VQ , EQ ) is an injective mapping
π : VQ → VG , such that for any edge (a, b) ∈ EQ , (π(a), π(b)) ∈ EG , and the vertices of Q
are mapped to distinctly colored vertices of G.
The algorithm traverses the tree in a bottom-up manner. For each block B, it outputs a
succinct synopsis of the set of colorful matches of the subquery SQ(B), using a projection
table and signature (as outlined in Section 3). that we now define precisely.
Signature: Let K = {1, 2, . . . , k} denote the set of colors used in the data graph, where
k is the size of the input query Q. The term signature refers to any subset α ⊆ K. For a
subquery Q and a colorful match π of Q, the signature of π refers to the set of colors of the
data vertices used by π and it is denoted sig(π), i.e., sig(π) = ∪a∈Q {χ(π(a))}.
Projection Tables: Let Q be subquery with two boundary nodes a and b. For a pair
of data vertices u and v and a signature α ⊆ K let cnt(u, v, α|Q) denote the number of
colorful matches of Q wherein the boundary nodes a and b are mapped to u and v and the
11
Overall Algorithm
1. Compute a decomposition tree T (Q) for the input query Q.
2. Traverse the tree bottom-up.
For each non-root block B:
Use the projection tables of the children blocks of B and
compute the projection table for B
3. Output the number of colorful mathes of the subquery
represented by the root-block.
Figure 3: Overall Algorithm
signature of π is α:
cnt(u, v, α|Q) = |{π ∈ Π : π(a) = u and π(b) = v
and sig(π) = α}|,
where Π is the set of all the colorful matches of Q.
These counts can be conveniently represented in the form a hash table with (u, v, α)
forming the key and the count forming the value. We refer to any encoding of the above
counts (such as the hash table above) as the projection table of Q. In the worst case, the table
may have size quadratic in the input data graph. However, a significant fraction of the
triplets will have a count of zero and we maintain only the non-zero counts.
The projection table for subqueries having a single boundary node a is defined in a
similar manner. For a data vertex u and a signature α ⊆ K, define
cnt(u, s|Q) = |{π ∈ Π : π(q) = u and sig(π) = α}|.
4.3
Computing the Counts
Given a decomposition tree, the algorithm works based on the fact that the projection table
for a block can be computed by joining the projection table of its children blocks.
As an illustration of the idea, consider the block B3 having boundary nodes f and
g, and the subquery represented by it (Figure 2). For a pair of vertices u and v, and a
signature α, the projection count cnt(u, v, α|B3 ) can be computed as follows. The block
consists of the path (a, f, g, c), and any match π for the subquery must map these nodes
to vertices (x, u, v, y) that form a path in the data graph. The block is annotated by its
children blocks B1 with boundary nodes a and c, and B2 with boundary node f . Any pair
of matches π1 and π2 for SQ(B1 ) and SQ(B2 ) can be extended as matches for SQ(B3 ), as
long as their signatures α1 and α2 are disjoint (since the blocks do not share any node)
and are contained within α. Therefore, we can derive the desired count by performing the
following aggregation over all quadruples (x, y, α1 , α2 ) satisfying the properties: (x, u, v, y)
forms a path in the data graph; α1 , α2 ⊆ α; (α1 ∩ α2 ) is empty. The aggregation is:
X X
cnt(u, v, α|B3 ) =
cnt(x, y, α1 |B1 ) × cnt(u, α2 |B2 ).
x,y α1 ,α2
12
We can express the projection counts for any block in the above manner. However,
as the number of children increases, the cartesian product involved in the aggregation
would be prohibitively expensive. Our procedures efficiently simulate the aggregation by
performing a sequence of join operations involving the projection tables of children blocks.
Given a decomposition tree, the algorithm traverses the decomposition tree in a bottomup manner, computing the projection tables for all the blocks and culminates in the rootblock representing the whole input query. At this step, instead of producing a projection
table, the algorithm simply computes the number of colorful matches. The pseudo-code is
shown in Figure 3.
5
Solving Blocks
The main step of the algorithm is the construction of the projection tables of a block from
its children blocks. In this section we develop efficient procedures for handling cycles.
For the sake of highlighting the main ideas, we first focus on the case of cycles found at a
leaf level of the decomposition tree (such as the cycle B1 in Figure 2); these cycles do not
have other blocks annotating them. General cycles are handled by extending these ideas
as discussed later.
5.1
Solving Cycles at the Leaf Level
Consider a cycle block C = (a0 , . . . , aL−1 ) of length L without annotations. The cycle may
have at most two boundary nodes. We discuss the more interesting case where the number
of boundary nodes is exactly two; the other cases are handled in a similar fashion. Let the
two boundary nodes of the cycle be ap and aq , for some 0 ≤ p, q ≤ L − 1. We present two
procedures for computing the projection table of C: a baseline procedure that uses a path
splitting strategy and an efficient procedure guided by a degree based ordering of vertices.
+
−
Path Splitting Algorithm (PS). For two nodes as and at on the cycle, let Ps,t
and Ps,t
be the
paths obtained by traversing the cycle from as to at in the clockwise and counter-clockwise
−
+
directions, respectively, i.e., Ps,t
= (as , as⊕1 , . . . , at ) and Ps,t
= (as , as 1 , . . . , at ), where ⊕
and refer to addition and subtraction modulo L.
+
+
Let cnt(·, ·, ·|Ps,t
) denote the projection counts for path Ps,t
taking as and at as the
+
boundary nodes. Namely, for a triple (u, v, α), let cnt(u, v, α|Ps,t ) denote the number of
+
colorful matches for Ps,t
wherein π(as ) = u, π(at ) = v and sig(π) = α. A similar notion is
−
defined for the paths Ps,t .
+
The procedure splits the cycle into two paths along the boundary nodes, given by Pp,q
− ; we refer to these special paths as P + and P − . See Fig 5 (a) for an illustration.
and Pp,q
The projection table for P + is constructed iteratively, by building the tables for the
+
paths Pp,j
, for each node aj found along the path. This is accomplished by extending
+
the projection table for the prior path Pp,j
1 via a join with the edges of the data graph.
The pseudocode is given in Figure 4 (Procedure 1). We assume that all the counts are initialized to zero. The first iteration is handled by directly reading the edges of the data
graph. In the subsequent iterations, we extend every triple (u, v, α) with non-zero count
+
cnt(u, v, s|Pp,j
1 ), with any edge (v, w), provided the resulting match is colorful. The
13
Procedure 1: Computing Projection Table for P +
For each edge (u, v) in the data graph G
+
cnt(u, v, α|Pp,p⊕1
) ← 1, where α = {χ(u), χ(v)}.
For j = p ⊕ 2, p ⊕ 3, . . . , q
+
For each triple (u, v, α) with cnt(u, v, α|Pp,j
1 ) 6= 0
For each edge (v, w) in G such that χ(w) 6∈ α do:
Let α0 = α ∪ {χ(w)}.
+
+
Increment cnt(u, w, α0 |Pp,j
) by cnt(u, v, α|Pp,j
1 ).
Procedure 2: Computing Projection Table for C
For each entry (u, v, α1 ) with cnt(u, v, α1 |P + ) 6= 0
For each entry (u, v, α2 ) with cnt(u, v, α2 |P − ) 6= 0
If α1 ∩ α2 = {χ(u), χ(v)}
α0 ← α1 ∪ α2
val1 ← cnt(u, v, α1 |P + ); val2 ← cnt(u, v, α2 |P − )
Increment cnt(u, v, α0 |C) by val1 × val2 .
Figure 4: PS Algorithm
counts for P − are constructed analogously. Finally, the projection table for the cycle C
is obtained by joining the counts of P + and P − , as shown in Procedure 2. Here, a pair of
triples (u, v, α1 ) and (u, v, α2 ) are joined, if the resulting match is colorful.
Discussion of baseline. As discussed below (Section 5.2), the PS procedure can be extended to handle general cycles with annotations, and yields an algorithm for handling
treewidth 2 queries. The resultant PS algorithm is equivalent to the original color coding
algorithm of Alon et al. [2]. Prior work [30, 1] on colorful subgraph counting utilize the
algorithm of Alon et al. as the basis for counting tree queries (treelets). We developed a
distributed implementation of the PS algorithm, and use it as the baseline in our experimental study. Known techniques for subgraph counting with large queries (e.g. [7, 26])
employ similar graph traversal techniques, making PS consistent with the state of the art
for subgraph counting as well as color coding.
We develop an procedure, called Degree Based (DB) algorithm, that outperforms the PS
algorithm for practical graphs and queries. It is motivated by the following observations.
First, the paths P + and P − may have uneven lengths (for instance, in Figure 5), |P + | = 6
and |P − | = 2) and the processing of the longer path dominates the overall running time.
Second, in real-graphs with skewed degree distributions, high degree vertices tend to have
more paths passing through them, which populate the projection tables of P + and P − .
However, significant fraction of these paths do not find appropriate counterparts in the
other table to complete a match, leading to wasteful computations. Third, in a distributed
setting, the above phenomenon manifests as higher load on processors owning high degree
vertices, leading to load imbalance.
It is not difficult to address the first issue alone. The only intricacy is that when the
paths are split evenly, the boundary nodes may appear internally on the the paths (see
Figure 5 with a split across nodes denoted h and d). This can be handled by recording the
mapping for the boundary nodes as part of the projection counts. We implemented the
above algorithm as well and noticed that the issue of wasteful computations and load im14
Figure 5: PS and DB Illustrations.
balance still persists. And furthermore, performance of the PS algorithm and the modified
implementations does not differ significantly on our benchmark graphs and queries.
Degree Based Algorithm (DB). The DB algorithm addresses all the three issues by using
the strategy of building the paths from high degree vertices.
Arrange the data vertices in the increasing order of their degree; if two vertices have
the same degree, the tie is broken arbitrarily, say by placing the vertex having the least
id first. We say that a vertex u is higher than a vertex v, if u appears after v in the above
ordering and this is denoted “u v”.
Consider the input cycle C = (a0 , a1 , . . . , aL−1 ) with boundary nodes ap and aq and let
π be a colorful match for C that maps the above nodes to data vertices u0 , u1 , . . . , uL−1 ,
respectively. Among these data vertices, let uj be the highest vertex. We refer to the corresponding node aj as the highest node of π.
The idea is to partition the set of colorful matches into L groups based on their highest
node ah and compute the projection table for each group separately. For a pair of data
vertices u and v, and a signature α, let cnt(u, v, α|C, hi = h) denote the number of colorful
matches of π for C, wherein π(ap ) = u, π(aq ) = v, sig(π) = α and ah is the highest node
of π. The projection table for C can be obtained by aggregating the above counts: for any
triple (u, v, α),
cnt(u, v, α|C) =
L−1
X
cnt(u, v, α|C, hi = h).
(1)
h=0
We next describe an efficient procedure for computing the counts cnt(u, v, α|C, hi = h).
The concept of high starting matches plays a crucial role in the procedure. Let ad be the
node diagonally opposite to ah on the cycle, i.e., d = h ⊕ bL/2c. The procedure splits the
+
−
cycles into two paths Ph,d
and Ph,d
; Figure 5 (b) shows the paths for two sample values
+
+
of h. Let aj be a node found on the path Ph,d
, A colorful match π for Ph,j
is said to be
high-starting, if the data vertex π(ah ) is higher than all the other data vertices used by π,
15
+
Procedure 1: Compute cnt∗ (u, v, α|Ph,d
)
For each edge (u, v) in the data graph G with u v
+
cnt∗ (u, v, α|Ph,h⊕1
) ← 1, where α = {χ(u), χ(v)}.
For j = h ⊕ 2, a ⊕ 3, . . . , d
+
For each triple (u, v, α) with cnt∗ (u, v, α|Ph,j
1 ) 6= 0
For each edge (v, w) in G s.t. u w and χ(w) 6∈ α:
Let α0 = α ∪ {χ(w)}.
+
+
Incr. cnt∗ (u, w, α0 |Ph,j
) by cnt∗ (u, v, α|Ph,j
1 ).
Procedure 2: Compute cnt∗ (x, y, α|C, hi = h) for Config. (A)
+
For each entry (u, v, x, α1 ) with cnt∗ (u, v, x, α1 |Ph,d
) 6= 0
−
∗
For each entry (u, v, y, α2 ) with cnt (u, v, y, α2 |Ph,d
) 6= 0
If α1 ∩ α2 = {χ(u), χ(v)}
α0 ← α1 ∪ α2
+
val1 ← cnt∗ (u, v, x, α1 |Ph,d
);
−
∗
val2 ← cnt (u, v, y, α2 |Ph,d )
Incr. cnt∗ (x, y, α0 |C, hi = h) by val1 × val2 .
Figure 6: DB Algorithm
+
. For a pair of vertices u and v, and a
i.e., π(ah ) π(ai ), for all nodes ai on the path Ph,j
+
∗
signature α, let cnt (u, v, α|Ph,j ) denote the number of high-starting colorful matches for
+
the path Ph,j
wherein π(ah ) = u, π(aj ) = v and sig(π) = α.
We then count the high-starting colorful matches for the two paths, which can be accomplished via edge extensions, as in the PS algorithm. However, the current setting offers
a crucial advantage: we can dictate that the starting node ah is the highest node, meaning
whenever an entry (u, v, α) gets extended by an edge (v, w), we can impose the condition
that u is higher than w in the degree based ordering. Imposing the condition leads to a
significant pruning of the tables. The pseudo-code is given in Figure 6 (Procedure 1).
While the degree based strategy is more efficient, we need to address an intricacy regarding the projection aspects. In contrast to the PS algorithm, the DB algorithm splits
at the highest node and consequently, the boundary nodes p and q may appear inside the
paths. Thus, in order to get the projection counts on p and q, we also need to explicitly
record the mappings for the boundary nodes.
+
−
The two nodes ap and aq may occur on either Ph,d
or Ph,d
. Six different configurations
are possible, of which two are shown in Figure 5 (b). In Configuration (A), the paths
include one boundary each, whereas in the second configuration, the same path includes
both the boundary nodes. The other four configurations are symmetric: the boundary
nodes may swap the paths in which they occur and in Configuration (B) can also reverse
the order in which they occur. We discuss the two configurations shown in the figure; the
other configurations are handled in a similar fashion.
Consider configuration (A). In order to record the mappings of the boundary node
ap , we introduce an additional field in the projection counts. For a triple of data vertices
+
u, v and x, and a signature α, let cnt∗ (u, v, x, α|Ph,d
) denote the number of high-starting
+
matches π for Ph,d with π(ah ) = u, π(ad ) = v, π(ap ) = x and sig(π) = α. These counts
16
+
Compute Projection Table for Ph,d
Let B be the block annotating the edge (ah , ah⊕1 )
+
cnt∗ (·, ·, ·|Ph,h⊕1
) = cnt∗ (·, ·, ·|B)
For j = h ⊕ 1, h ⊕ 2, . . . , d
Execute NodeJoin(aj )
Execute EdgeJoin(aj )
Execute NodeJoin(ad )
NodeJoin(aj ):
If aj is annotated by a block B
+
For each (u, v, α1 ) with cnt∗ (u, v, α1 |Ph,j
) 6= 0
For each (v, α2 ) with cnt(v, α2 |B) 6= 0
If (α1 ∩ α2 = {χ(v)}
α ← α1 ∪ α2
+
val1 ← cnt∗ (u, v, α1 |Ph,j
); val2 ← cnt(v, α2 |B)
+
Incr. cnt∗ (u, v, α|Ph,j
) by val1 × val2
EdgeJoin(aj )
+
For each entry cnt∗ (u, v, α1 |Ph,j
) 6= 0
For each entry cnt(v, w, α2 |B) 6= 0 and u w
If (α1 ∩ α2 = {χ(v)}
α ← α1 ∪ α2
+
val1 ← cnt∗ (u, v, α1 |Ph,j
); val2 ← cnt(v, w, α2 |B)
+
Incr. cnt∗ (u, w, α|Ph,j⊕1 ) by val1 × val2
Figure 7: DB Procedure for General Cycle Blocks
are computed in a manner similar to the base procedure shown in Figure 6 (Procedure
1); however, when the process encounters the boundary node p (namely, the initialization step or j = p), the mapped vertex (v or w, respectively) is recorded in the additional field. The analogous counts for P − can derived in a similar manner. The value
of cnt∗ (u, v, α|C, hi = h) is obtained by joining the two; see Procedure (2) in Figure 6. Configuration (B) is handled in a similar fashion, except that we need two additional fields
to record the mappings for both the boundary nodes. Namely, we maintain counts having keys of the form (u, v, x, y) representing the mapping of the nodes h, d, q and p to the
vertices u, v, x and y. Procedure (2) is also adjusted accordingly. Finally, we can get the
projection table cnt(u, v, α|C) via aggregation, as in Equation 1.
5.2
Solving General Blocks
In this section, we present procedures for handling generic blocks. We first consider the
case of cycle blocks with two boundary nodes.
Consider a generic cycle C = (a0 , a1 , . . . , aL−1 ) having two boundary nodes ap and aq ,
whose nodes and edges may be annotated with other blocks (children of C in the decomposition tree). All these blocks have at most two boundary nodes and these are found on
C. For such any block B, the subquery represented by B has the same boundary nodes as
that of B. Thus, we can get the projection table for C by joining the projection tables of the
17
subqueries represented by the above blocks, as described below.
As before, we consider each possible choice for the highest node ah and split the cycle
+
−
+
into two paths Ph,d
and Ph,d
. The path segment Ph,d
also represents a subquery (induced
by the union of the nodes found in the path and the blocks annotating path). Thus, we
can extend the notion of projection tables for these segments as well. The procedure for
+
computing the projection table for Ph,d
is similar that the one discussed in previous section
(Procedure 1 in Figure 6), and works by extending one edge in each step. However, two
aspects need to be addressed. Firstly, in contrast to the prior procedure, the edge being
extended may be annotated with a block or un-annoated (and correspond to an original
edge found in input query Q). In the former case, we perform a join operation with the
edges of the data graph (as before), whereas in the latter case the join operation involves
the projection table of the block B. For the sake of uniformity, it will be convenient to view
the former edges as blocks as well, denoted BG , and associate with them a projection table
derived from the graph edges, as follows. For each edge (u, v) ∈ G, set cnt(u, v, α) as 1, for
α = {χ(u), χ(v)}; all other entries of the table are set to a count of zero. The second aspect
is that the nodes of the cycles may also be annotated, and these get included as part of the
sequence of joins being performed. The two aspects are addressed by procedures called
NodeJoin and EdgeJoin. The pseudo-code is shown in Figure 7.
The procedure starts with an initial table representing the first edge (ah , ah⊕1 ) and performs a sequence of join operation with the blocks annoatating the nodes and edges of the
cycle. At this juncture, two intricacies must be highlited. Firstly, the endpoint ah and/or
+
+
ad may be annotated by a block B, which must be joined by either Ph,d
or Ph,d
, but not
+
by both (to avoid double counting). For this purpose, we adopt the convention that Ph,d
−
and Ph,d include only the block annotating ad and ah (if found), respetively. Secondly, for
a block with two boundary nodes p and q, the projection table views one of them as the
first boundary node and the other as the second (corresponding to the two components
of the keys of the form (u, v, α)). Thus, the boundary nodes are ordered and the projection tables need not be symmetric: taking q as the first boundary node and p as the second
boundary node would produce a different boundary tables. However, the boundary tables
are transpose of each other (cnt(u, v, α) = cnt(v, u, α)). Our algorithm maintains both the
tables and uses the appropriate one as dictated by the nodes of the cycle. The pseudo-code
reflects the first aspect, but, for the sake of clarity, ignores the second.
The projection counts obtained by the above process are joined using a proecure similar
to Figure 6, taking into account the configuration in which the boundary nodes occur.
These are aggregated over all possible choices of the high node ah .
Cycles with a single boundary node are handled in a similar manner by considering
+
each possible choice for the highest node ah and splitting the cycle into two paths Ph,d
and
−
Ph,d . The setting is simpler with only two configurations possible on how the boundary
nodes may appear on the paths: the (single) boundary node may appear in P + or P − .
Thus, the prior procedures can be applied here as well.
The case of leaf blocks are also handled via join operations. Any leaf block (a, b) is
processed by joining the projection table for the blocks annotating the nodes a, the edge
(a, b) and the node b (if found).
At the end of the traversal process, the root block is solved, which is either a cycle or a
18
singleton node. In the former case, the block is treated as a cycle without boundary nodes.
Instead of computing its projection table, we simply count the number of colorful matches,
via a procedure similar to that of two-boundary cycles. In the latter case, we consider the
projection table of the block annotating the singleton node and output the sum of counts
across all entries of the table. The process yields the number of colorful matches of the
input query Q.
6
Finding Good Decomposition Trees
In each step of the decomposition process, multiple blocks may be available for contraction. Each sequence of choices leads to a unique decomposition tree, and hence, multiple
trees are possible for a given query. For example, the query brain1 (Figure 8) admits two
decomposition trees: (i) contract the 4-cycle first and then the 6-cycle, and (ii) vice versa.
We conducted an experimental study involving a number of real-world data graphs and
queries. For each query, we enumerated all the possible decomposition trees and evaluated the execution time on each graph. We observed a maximum difference of 13x in the
execution times of two decomposition trees for the same graph-query combination. However, we noted that in most cases the optimal tree is independent of the data graph and is
mainly determined by the structure of the query. These observations show that we need a
procedure for selecting a good tree, but in this process, we need not analyze the large data
graph; rather, it suffices to focus on the structural properties of the small query graph.
Our study also showed that the following factors, in the decreasing order of importance, determine the execution time: (i) length of the longest cycle block; (ii) number of
boundary nodes; (iii) number of node/edge annotations. Armed with the above observations, we designed a simple heuristic procedure. Enumerate all possible trees for the given
query and pick the best using the above factors for comparison. In our experimental setting, barring a few exceptions, the heuristic picked the optimal tree in majority of the cases
and a near-optimal tree for the rest. Since the queries are of small size (about 10 nodes),
even a sequential implementation of the heuristic takes insignificant amount of running
time.
7
Distributed Implementation
In this section, we present a brief sketch of the distributed implementation of the two
algorithms, highlighting their main aspects. The distributed implementation consists of
three layers. The first layer, called the planner, finds a good decomposition tree for the
given query a fast sequential implementation the heuristic discussed in Section 6. The
second layer, called the plan solver, takes the data graph and the decomposition tree and
implements the PS and DB algorithms presented in Section 5. It accomplishes the above
task by using efficient join routines supported by the third layer, called engine. The engine
has three functionalities. The first is to store the data graph in a distributed manner. This
is achieved via a 1D decomposition, wherein the vertices are equally distributed among
the processors using block distribution, and each vertex is owned by some processor. The
second is to maintain projection tables. These tables are of two types: unary projection
19
Graph
Table 1: Real Data Graphs
Domain
Nodes Edges
brightkite
condMat
astroph
enron
hepph
slashdot
epinions
orkut
roadNetCA
brain
Geo loc.
Collab.
Collab.
Commn.
Citation
Soc. net.
Soc. net.
Soc. net.
Road net.
Biology
58K
23K
18K
36K
34K
82K
131K
524K
2M
400K
214K
93K
198K
180K
421K
900K
841K
1.3M
2.7M
1.1M
Avg
Deg
4
4
11
5
12
11
6
3
1.3
3
Max
Deg
1135
281
504
1385
848
2554
3558
1634
14
286
tables having single-vertex keys of the form (u, α) associated with blocks having single
boundary nodes; binary projection tables having two-vertex keys of the form (u, v, α). The
binary tables also have variants involving additional fields for storing the mappings for
the boundary vertices. The engine provides a convenient abstraction to the plan solver
for all these types of tables. All the tables are maintained as distributed hash tables which
use open addressing to resolve collisions. Every entry (u, v, α) is stored on the processor
owning v; the degree of v is packed as part of the entry for enforcing the degree constraint
in the join operations (of the form u w in Procedure 1 of Figure 6). Signatures are
maintained as bitmaps. The third functionality is to support two types of join operations
on the projection tables. The first type of join is used for extending a path segment an
edge; this involves a join with either the graph edges or the projection table of the block
annotating the edge. In the former case, the extension of an entry with a key (u, v, α)
with an edge (v, w) will be performed at the owner of v. The result is an entry with a key
(u, w, α0 ); this entry is communicated to the owner of w, where it gets stored. The latter case
involves join of two entries with keys (u, v, α1 ) and (v, w, α2 ). Since the first entry is stored
at the owner of v and the second, at the owner of w, a communication is performed to
bring the two entries to a common processor. The second type of join is used for merging
the projection tables of two path segments (for example, Procedure 2 in Figure 6) and it
is implemented in a similar way. The two operations are implemented using a standard
sort-merge join procedure with signature compatibility checks performed via fast bitwise
operations.
8
Experimental Study
We present an extensive experimental evaluation of the algorithms presented in the paper.
Our experiments include a comparison of the algorithms on execution time, strong and
weak scaling studies for our algorithm, and studies to evaluate the quality of our query
plan generation heuristic and the efficacy of color coding for treewidth two queries.
20
Figure 8: Real world queries used in our study.
Figure 9: Average execution time (seconds).
8.1
Experimental Setup
System. The experiments were conducted on an IBM Blue Gene/Q system [12]. Each
BG/Q node has 16 cores and 16 GB memory; multiple nodes are connected using a 5D
torus interconnect. Our implementation is based on MPI2 with gcc 4.4.6 with the number of ranks varying from 32 to 512. Each MPI rank was mapped to a single core. The
number of MPI ranks mapped to a node was adjusted based on the memory requirements
of individual experiments.
Graphs. The experiments involved nine real world graphs obtained from the SNAP
dataset collection and the human brain network from the Open Connectome Project (http:
//snap.stanford.edu, http://www.openconnectomeproject.org/). Our benchmark includes representative graphs from different domains in SNAP. The graphs and
their characteristics are presented in Table 1. We also used synthetic R-MAT graphs [11],
for the purpose of studying the weak scaling behavior of our algorithms.
Queries. Our query benchmark consists of the ten real world queries shown in Figure 8. The queries were derived from prior network analysis work spanning diverse domains: dros, ecoli1, ecoli2, brain1, brain2, brain3 - biological networks
[22, 19]; glet1, glet2 - graphlets [7]; wiki - collaboration networks [32]; youtube spam networks [24].
Algorithms. We study two algorithms: PS, which serves as the baseline, and our
degree based DB algorithm. Recall that PS is equivalent to the dyamic programming based
algorithm of Alon et al. [2].
21
8.2
Graph-Query Characteristics
The characteristics of the input graph and query strongly influence the running time of
query counting algorithms. To obtain an overall characterization of the phenomenon,
we measured the execution time of the DB algorithm on each of the 100 real graph and
query combinations using 512 MPI ranks. Figure 9 shows the average running time for
each graph across the ten queries and the average running time of each query across the
ten graphs. The wide variations in execution time across graphs and queries is indicative of their relative difficulty in practice. For example, although roadNetCA is a larger
graph than epinions, the average running time of the former is smaller than the latter by
an order of magnitude. We can understand this behaviour by studying the skew in underlying degree distribution. In general, counting colorful occurrences of a query on a
graph with high skew (indicated by high maximum degree in Table 1) tends to be computationally expensive. Similarly, the queries also exhibit large variations in running time,
ranging from sub-second for youtube, glet1 and glet2 to more than a minute for brain2 and
brain3. These variations can be accounted for by studying the differences in the size and
the sub-structures of the queries. We observed that queries with longer cycles are more
challenging. As an extreme case, a 12-vertex complete binary tree query requires 2 seconds on average, in contrast to the 10-vertex brain3 query which requires nearly 2 minutes
on average, exemplifies our observation.
8.3
Performance Comparison of PS and DB Algorithms
We study the performance of the PS and DB algorithms on 100 graph-query combinations
obtained by selecting a graph from Table 1 and a query from Figure 8. For our DB algorithm, we used plans supplied by the heuristic described in Section 6. In contrast, for the
PS algorithm, we enumerated all the possible plans and obtained the optimal plan. Thus,
we compare our algorithm to the best possible scenario for the baseline algorithm.
We compute the improvement factor (IF ) of DB over PS as the ratio of the execution
time of PS to DB. Figure 10 shows IF at 32 and 512 ranks. The combinations where DB
outperforms PS (IF > 1) are highlighted in green. The blank entries represent cases where
PS (or DB) did not complete execution, due to lack of available memory. At 32 ranks, we
can see that DB outperforms PS on 84% of the graph-query combinations with IF being as
high as 9.1x (average 2.4x). At 512 ranks, DB outperforms the baseline on 89% of the cases,
with IF becoming as high as 28.7x (average 5.0x).
We can see that the relative performance of the two algorithms is dependent on the
graph-query pair. For instance, the average IF on enron and condmat graphs are 8.4
and 3.1 on 512 ranks, respectively, correlating well with their skew in the degree distribution (see Table 1). Similarly, the improvement factors is higher on complex queries such
as brain1 where the average improvement is 13.1x, compared to youtube where the average improvement is only 4.1x. The phenomenon becomes extreme in the case of road
networks that have very low skew and exhibit sub-second average running time across
queries.
Our DB algorithm scales better than PS, as demonstrated by the increase in IF at higher
ranks. For different graph-query combinations, we computed the ratio of IF at 512 ranks
to that of 32 ranks and found that IF increases by a factor of up to 4.7x (average 1.7x).
22
(a) 32 Ranks
(b) 512 Ranks
Figure 10: Improvement factor of the DB algorithm over the PS algorithm.
23
DB
PS
DB
PS
1.00
Normalized Time
Normalized MaxLoad
1.00
0.75
0.75
0.50
0.50
0.25
0.25
0.00
0.00
brain1 brain2 dros ecoli1 ecoli2 glet1 glet2
wiki youtube
brain1 brain2 dros ecoli1 ecoli2 glet1 glet2
(a) Time
wiki youtube
(b) Max. Load
DB
PS
Normalized AvgLoad
1.00
0.75
0.50
0.25
0.00
brain1 brain2 dros ecoli1 ecoli2 glet1 glet2
Real World Queries
wiki youtube
(c) Avg. Load
Figure 11: Normalized execution time, average load and maximum load on enron graph.
To understand this trend further, we compute the load (number of projection function operations) for both algorithms for processing different queries on the enron graph at 512
ranks. For different queries, Figure 11 shows the execution time and the average and maximum load. We can see that DB has lesser average load than PS, since DB avoids wasteful computations. Furthermore, the improvement obtained by DB over PS on execution
time correlates well with improvement obtained on the maximum load. For example, on
ecoli1 query, even though PS outperforms DB at 32 ranks, the perforamance is reversed
at 512 ranks (see Fig 10), because of superior load balancing characteristic of DB.
8.4
Scalability Characteristics of DB Algorithm
We studied the scaling of DB across the 100 graph-query combinations. For each combination, we computed the ratio of the execution time at 512 ranks to that of 32 ranks. Figure 12 summarizes the above information by providing the averge of the above speedup
for each query across graphs and the same for each graph across queries. As against an
24
Figure 12: Avg. speedup of DB at 512 ranks compared to 32 ranks.
●
brain1
brain2
dros
ecoli1
ecoli2 ● glet2
glet1 ● wiki
●
youtube
Ideal
●
●
Speedup
8
●
●
4
●
●
2
1
brain3
dros
ecoli1 glet1 ● wiki
ecoli2 ● glet2 youtube
10000
Time (Seconds)
16
brain1
brain2
●
●
●
32
●
●
●
●
●
●
●
●
●
●
●
●
●
●
●
32
64
128
256
512
1000
100
10
1
64
128
Ranks
256
512
Ranks
Figure 13: Strong and weak scaling
ideal speedup of 16x, we see that the algorithms obtains speedups in the range of 7.4x to
15.8x.
We studied the strong scaling behavior of our algorithm, using enron as a representative graph. Taking 32 ranks as the baseline, Figure 13 shows the speedup up to 512 ranks
for different queries. The algorithm scales well across queries, with an average speedup of
8.2x and maximum speedup of 9.9x at 512 ranks (as against an ideal speedup of 16x).
To study weak scaling, we use R-MAT synthetic graphs with parameters A = 0.5,
B = 0.1, C = 0.1 and D = 0.3 and edge factor 16, suggested in a Graph 500 benchmark
specification (http://www.cc.gatech.edu/˜jriedy/tmp/graph500/). The number of vertices was fixed at 1K per rank and the number of ranks was varied from 32 to
512. We report the execution times each query-rank combination in Figure 13. We see excellent weak scaling behavior with the execution times at 512 ranks remaining close to that
of the baseline 32 ranks.
8.5
Evaluation of Plan Generation Heuristic
We studied the quality of our plan generation heuristic for the DB algorithm at 512 ranks.
For each graph-query combination, we determined the optimal plan via an exhaustive
enumeration. We compared the execution time of the heuristic plan to the optimal plan and
measured the percentage difference. These results are reported in Figure 14. We can see
that in 90% of the case, the heuristic generated the optimal plan, whereas in the remaninig
cases, the difference was at most 15%.
25
Figure 14: Error % of the execution time of the plan proposed by the plan heuristic with
reference to the optimal plan for each graph-query combination.
Figure 15: Coefficient of variation with 50 trials of color coding for each graph-query combination.
8.6
Precision of Color Coding
We evaluated the precision of color coding on our benchmark by performing independent
trials and computing the empirical variance of the sample (see Section 2). Specifically, for a
given graph-query combination we performed a sequence of trials, where in each trial the
colorful count ncolorf ul (G, Q, χ) was computed for a fresh random coloring. We performed
10 random trials for each of the 100 graph-query combinations in our test set and evaluated
the empirical mean and variance of the number of colorful matches. For each graph-query
combination, we computed the coefficient of variation, which is the ratio of the empirical
variance to the mean. The results are shown in Figure 15. A value close to 0 indicates the
convergence of our estimate to the true mean n(G, Q). We observed that with only three
trials, 82% of the graph-query combinations had coefficient of variation at most 0.1; when
the number of trials was increased to 10, it increases to 91%. Hence, using 512 ranks, for
a majority of the input graph-query combinations in our benchmark, we require less than
a minute to count the actual number of matches of the query, with ≈ 10% accuracy. We
conclude that our DB algorithm enables fast approximate counting of treewidth 2 queries
for data graphs spanning various real domains.
26
9
Cycle queries on random power law graphs
In this section we concentrate on cycle queries of constant size and analyze the expected
runtime of a variant of the PS and DB procedures on a certain class of random data graphs.
We prove a lower bound on the expected runtime of the PS procedure and an upper bound
on the expected runtime of the DB procedure. Both bounds are functions of the (expected)
degree sequence of the graph. We show that our upper bound on the runtime of the DB
procedure is never worse (up to constant factors) than the lower bound on the PS procedure. Moreover, if the random graphs satisfy a natural power law property then we prove
that the expected runtime of the DB procedure is polynomially better. Recall that the most
complicated blocks in our general decomposition of the query graph are (annotated) cycles. Thus, we postulate that the better performance of our variant of the DB procedure
on cycle queries explains the better performance of the DB procedure studied in the main
body of the paper on queries of treewidth 2.
The class of random graphs considered is a certain variant of the Chung-Lu graphs [14],
a popular model for random graphs that captures several properties of real world social
networks, as defined precisely below.
9.1
The procedures analyzed
Consider a cycle query C of constant size k. Since the query graph is a cycle, the decomposition tree consists of only a root node which represents this cycle with a single boundary
node.
Recall that procedure PS computes cnt(u, {1, . . . , k}|C), for each node u. For this
it first computes cnt(u, v, α1 |P + ) and cnt(u, v, α2 |P − ) for all of nodes v and all signatures. Since k is constant the number of signatures is also constant and thus computing
cnt(u, {1, . . . , k}|C) given cnt(u, ·, ·|P + ) and cnt(u, ·, ·|P − ) can be done in linear time. In
the PS procedure the computation of cnt(u, ·, ·|P + ) and cnt(u, ·, ·|P − ) is done by iteratively recomputing cnt(u, v, ·|P + ) for all u and v. This recomputation can be viewed as a
optimized version of enumerating all paths starting at u, where instead of storing paths
explicitly, we only store the two endpoints, the signature, and a count. Our simplified
variant of the PS procedure, which is more amenable to analysis, computes cnt(u, ·, ·|P + )
and cnt(u, ·, ·|P − ) by enumerating all possible paths, instead of performing this optimized
enumeration. Thus, its complexity is linear in the number of possible paths of lengths
1, . . . , d 12 ke. We will refer to this version of PS procedure as the PS procedure throughout
this section to simplify notation.
To increase the efficiency of our PS procedure when applied to cycle queries we can
break symmetry using the id of the nodes and count only colorful matches where node
u has the largest id among all data nodes in the image of π. Consequently, cnt(u, ·, ·|P + )
and cnt(u, ·, ·|P − ) need only to count paths with the property that node u has the largest
id among nodes on the path.
For an integer q ≥ 3 let
Y (q) = |{(u1 , . . . , uq ) is a simple path and id(u1 ) > id(uj ), j ∈ [2..q]}| .
It follows that the expected complexity of procedure PS is linear in
27
Pdk/2e
q=1
(2)
E[Y (q)]. Later,
we derive lower bounds on E[Y (q)] for any constant q, and our bounds are monotone
increasing in q. Thus, the dominant term in the complexity of procedure PS is provided
by our bound on Y (d 21 ke).
Procedure DB also computes cnt(u, {1, . . . , k}|C), but does it by computing high-starting
colorful matches. Again, to increase the efficiency of procedure DB when applied to cycle
queries we can break symmetry and only compute cnt(u, {1, . . . , k}|C, hi = u), namely,
colorful matches where node u is the highest in the degree based ordering among all data
+
nodes in the image of π. It follows that for this we need only to compute cnt∗ (u, ·, ·|Ph,d
)
−
∗
and cnt (u, ·, ·|Ph,d ), namely, the number of high-starting colorful matches for the paths
+
−
Ph,d
and Ph,d
, wherein π(ah ) = u, π(ad ) = v and sig(π) = α, for all possible v and α. Note
that similarly to the PS procedure, the recomputation in the DB procedure considered in
the main body of the paper can be viewed as a optimized enumeration, where instead of
storing paths explicitly, we only store the two endpoints, the signature, and a count. Our
simplified variant of the DB procedure, which is more amenable to analysis, computes the
counts by enumerating all these paths, instead of performing this optimized enumeration.
For an integer q ≥ 3 let
X(q) = |{(u1 , . . . , uq ) is a simple path and u1 uj , j ∈ [2..q]}| .
(3)
Pdk/2e
It follows that the expected complexity of procedure DB is linear in q=1 E[X(q)]. Later,
we will derive bounds on E[X(q)], and since our bounds will be monotone increasing
in q, the dominant term in complexity of procedure DB is provided by our bound on
E[X(d 21 ke)].
9.2
The random data graphs
We analyze our algorithm on Chung-Lu graphs whose expected degree sequence has some
additional property.
Chung-Lu distribution The Chung-Lu distribution on random graphs is defined as follows. First choose a degree sequence
d = (d1 , . . . , dn ), where V = [n]. We assume that
P
√
du ≥ 1 for all u ∈ V . Let m = 21 u∈V du . We assume that m ≥ n and maxu∈V du ≤ n. To
generate the graph G = (V, E) we include each edge (u, v) independently
with probability
P
du dv /(2m). Note that the expected degree of every node u ∈ V is v∈[n] du dv /(2m) = du
as required. We write deg(u) to denote the actual degree of u in G.
We require the degree sequence d = (d1 , . . . , dn ) to be λ-balanced.
Balanced degree sequence A degree
d =
. , dn ) is λ-balanced for some
Psequence
P(d1a, . .P
a+b
λ ∈ (0, 1) if for any integers a, b ≥ 1, u du ≤ λ · ( u du )( u dbu ). Intuitively, a degree
sequence is balanced if it is not too concentrated on the high degree nodes.
For some of the claims we require the sequence to satisfy the stronger truncated power
law.
28
Truncated power law distribution A degree sequence d = (d1 , . . . , dn ) satisfies the truncated power law for a constant α ∈ (1, 2) if for each 0 ≤ j ≤ 21 log n, the number of nodes
with degree between 2j and 2j+1 is Θ(n/2αj ). We show later that if a sequence satisfies the
power law for α then it is λ-balanced for λ = O(nα/2−1 ).
9.3
Main theorem
Theorem 9.1 Let G be sampled according to the Chung-Lu distribution on n vertices with an
n−δ -balanced degree sequence d, for some constant δ > 0, and let q ≥ 3 be a constant. Then,
(1) The expected number of paths (u1 , . . . , uq ) of length q for which id(ui ) is with the highest id is
lower bounded by
1
E[Y (q)] ≥ (1 − o(1)) · (2m)−q+3
q
!q−2
X
d2u
.
u
(2) The expected number of high-starting paths of length q is upper bounded by
!q−1
E[X(q)] ≤ C(2m)−q+2
X
d2−1/(q−1)
u
u∈V
for some constant C > 1.
(3) Based on the above inequalities
1. E[X(q)] = O(E[Y (q)]).
2. If the degree sequence satisfies the truncated power law with parameter α, for any constant α ∈ (1, 2), then E[X(q)] is polynomially smaller than E[Y (q)].
Proof:
(1) follows by Lemma 9.5. (2) follows by Lemma 9.6. (3) follows by putting
together Lemma 9.7 and Lemma 9.8.
P
P 2−1/(q−1)
P
P
Remark 9.2 Note that if u d2u ≥
≥
u du = 2m and
u du
u du = 2m, since
du ≥ 1 by our assumption on the degree sequence. Thus both E[Y (q)] and E[X(q)] are monotone
in q. It follows that for any constant k, the dominant terms in the complexity of our PS and DB
procedures for query cycles of length k is indeed determined by E[Y (d 12 kee)] and E[X(d 12 kee)],
respectively, as claimed in section 9.1.
The rest of this section is devoted to proving Lemmas 9.5, 9.6, 9.7 and 9.8. The
analysis uses the approach of [5]. It turns out, however, that new ingredients are necessary
for handling cycles due to the presence of multiple intermediate nodes.
9.4
Useful facts
In this section we state some simple claims and known results that will be useful in the
rest of the proof. The following claim specifies the probability that a fixed path exists in a
random graph drawn from the Chung-Lu distribution:
29
Claim 9.1 Let G be drawn from the Chung-Lu distribution with degree sequence d. Then for any
q ≥ 2 and a vector (u1 , . . . , uq ) ∈ V q of distinct nodes
q−1
du1 duq Y d2uj
Pr[(u1 , . . . , uq ) is a path in G] =
·
.
2m
2m
j=2
Proof: Since the input graph G is drawn from the Chung-Lu distribution, for each j =
1, . . . , q − 1 we have
du du
Pr[(uj , uj+1 ) ∈ E] = j j+1 .
2m
Since edges are included in E independently, we have
Pr[(u1 , . . . , uq ) is a path in G] =
q−1
Y
j=1
q−1
duj duj+1
du du Y d2uj
= 1 q ·
2m
2m
2m
j=2
We will also use
Theorem 9.3 (Chernoff
bound) Let X1 , . . . , Xn be independent 0/1 Bernoulli random variP
ables. Let X = ni=1 Xi , and let µ := E[X]. Then for any ∈ (0, 1)
2 µ)
Pr[X ≤ (1 − )µ] < e−Ω(
and for every λ > 0
Pr[X ≥ (1 + λ)µ] < (e/(1 + λ))(1+λ)µ
A simple corollary of the bound is
P
Corollary 9.4 Let X1 , . . . , Xn be independent 0/1 Bernoulli random variables. Let X = ni=1 Xi ,
and let µX := E[X]. Let Y1 , . . . , Yn be independent 0/1 Bernoulli random variables. Let Y =
P
n
−cµX , for some constant
i=1 Yi , and let µY := E[Y ]. If µY ≤ µX /20, then Pr[Y ≥ X] < 2e
c > 0.
Proof: We have
Pr[Y ≥ X] ≤ Pr[X ≤ (µX + µY )/2] + Pr[Y ≥ (µX + µY )/2].
Since (µX + µY )/2 ≤ (3/4)µX , we have Pr[X ≤ (µX + µY )/2] < e−cµX , for some constant
c > 0 by the first bound from Theorem 9.3 invoked with = 1/4.
We also have (µX + µY )/2 ≥ 10µY , so by the second bound from Theorem 9.3 invoked
with λ = 9 we get
Pr[Y ≥ (µX + µY )/2] < (e/10)10µY ≤ (e/10)µX /2 .
Clearly, (e/10)µX /2 = e−cµX for some constant c > 0, as required.
Finally, we need the following inequality on the degree sequence d.
Claim 9.2 For all t ≥ 1 one has
!1/t
!(t−1)/t
X
X
X
2−1/t
2
du
≤
du
·
du
.
u
u
u
30
Proof:
By Hölder’s inequality with conjugates t and
X
d2−1/t
=
u
u
X
we have
2−2/t
d1/t
u · du
u
!1/t
≤
X
t
(d1/t
u )
X
·
u
≤
X
1
(2−2/t)· 1−1/t
du
!1−1/t
u
!1/t
du
!(t−1)/t
X
·
d2u
u
u
9.5
1
1−1/t
Proofs of the lower and upper bounds
We start by lower bounding E[Y (q)].
Lemma 9.5 Let G = (V, E), V = [n] be drawn from the Chung-Lu distribution with degree
sequence d. Suppose that node id’s are chosen uniformly at random. For any integer q ≥ 3 let
Y (q) be defined by (2). Then if d is n−δ -balanced for a constant δ > 0, the following holds for any
constant q
!q−2
X
1
−q+3
2
E[Y (q)] ≥ (1 − o(1)) · (2m)
du
.
q
u
Proof:
We have
E[Y (q)] =
X
(u1 ,...,uq )∈V
ui distinct
Pr[(u1 , . . . , uq ) is a path in G] · Pr [id(u1 ) > id(uj ), j ∈ [2..q]]
q:
(4)
By Claim 9.1 one has for any vector (u1 , . . . , uq ) of distinct nodes
Pr[(u1 , . . . , uq ) is a path in G] =
q−1
Y
j=1
q−1
duj duj+1
du1 duq Y d2uj
=
·
.
2m
2m
2m
j=2
Furthermore, for any fixed q-tuple of distinct nodes (u1 , . . . , uq ) we have
Pr [id(u1 ) > id(uj ), j ∈ [2..q]] =
1
q
since id’s are uniformly random by assumption of the lemma. (Note that this implies that
Y (q) is lower bounded by a constant times the total number of paths of length q, since q is
constant.)
31
Plugging this bound into (4), we get
X
E[Y (q)] =
Pr[(u1 , . . . , uq ) is a path in G] · Pr [id(u1 ) > id(uj ), j ∈ [2..q]]
(u1 ,...,uq )∈V q :
ui distinct
=
1
q
X
(u1 ,...,uq )∈V q :
ui distinct
q−1
du1 · duq Y d2uj
·
2m
2m
j=2
(5)
q−1
Y X d2uj
X duq
1X
·
=
du1 ·
q
2m
2m
u1 ∈V
1
−
q
j=2 uj ∈V
uq ∈V
X
(u1 ,...,uq )∈V q :
ui not distinct
du1 · duq
·
2m
q−1
Y
j=2
d2uj
2m
.
Clearly,
1
q
q−1
Y X d2uj
X duq
1
du1 ·
·
= (2m)3−q ·
2m
2m
q
X
u1 ∈V
j=2 uj ∈V
uq ∈V
!q−2
X
d2u
(6)
u∈V
We now show that
1
q
X
(u1 ,...,uq )∈V q :
ui not distinct
du1 · duq
·
2m
q−1
Y
j=2
d2uj
1
= o (2m)3−q ·
2m
q
!q−2
X
d2u
.
(7)
u
Together with (5) this gives the result.
We show that (7) follows from the assumption that the degree sequence d is balanced.
Let
U = {(u1 , . . . , uq ) ∈ V q : ui not distinct}
Wk,` = {(u1 , . . . , uq ) ∈ V q : uk = u` } , for 1 ≤ k < ` ≤ q.
Note that U ⊆ ∪1≤k<`≤q Wk,` . Let ~v denote a vector (u1 , . . . , uq ) ∈ V q . To simplify the
exposition denote
q−1
du · duq Y d2uj
·
.
S(~v ) = 1
2m
2m
j=2
We show that for every 1 ≤ k < ` ≤ q,
X
~v ∈Wk,`
1
S(~v ) ≤ n−δ (2m)3−q ·
q
!q−2
X
d2u
.
(8)
u
Since there are constant number of such sets this implies (7). To prove (8) we consider four
different cases.
32
Case 1: 2 ≤ k < ` ≤ q − 1. In this case
X
Y
d2uj
du1 · duq d4uk
·
·
2m
4m2
2m
(u1 ,...,uq )∈Wk,`
j∈[2..q−1]\{k,`}
X duq
X d4u
Y
X d2uj
1 X
k
·
=
du1 ·
·
q
2m
4m2
2m
u1 ∈V
uq ∈V
uk ∈V
j∈[2..q−1]\{k,`} uj ∈V
X duq
X
X d2u
X d2u
1
k
k
·
≤ n−δ
du1 ·
·
·
q
2m
2m
2m
1
q
S(~v ) =
~v ∈Wk,`
X
uq ∈V
u1 ∈V
Y
X
j∈[2..q−1]\{k,`} uj ∈V
1
= n−δ (2m)3−q ·
q
uk ∈V
uk ∈V
d2uj
2m
!q−2
X
d2u
u∈V
Note that the inequality follows since the degree sequence d is n−δ -bounded.
Case 2: k = 1 and ` = q. In this case
X
S(~v ) =
~v ∈W1,q
1
q
X
(u1 ,...,uq )∈W1,q
d2u1
·
2m
Y
j∈[2..q−1]
d2uj
2m
Y
X d2uj
1 X d2u1
=
·
q
2m
2m
u1 ∈V
j∈[2..q−1] uj ∈V
X
Y
X d2uj
X
1 −δ
du1
≤ n
·
du1 ·
q
2m
2m
u1 ∈V
u1 ∈V
j∈[2..q−1] uj ∈V
!q−2
X
1 −δ
d2u
= n (2m)3−q ·
q
u∈V
Case 3: k = 1 and ` ∈ [2..q − 1]. In this case
X
~v ∈W1,`
Y
d2uj
d3u1 duq
·
·
2m 2m
2m
(u1 ,...,uq )∈W1,`
j∈[2..q−1]\{`}
Y
X d2uj
1 X d3u1 X duq
=
·
·
q
2m
2m
2m
u1 ∈V
uq ∈V
j∈[2..q−1]\{`} uj ∈V
X duq
X d2u
Y
X d2uj
1 −δ X
1
·
≤ n
du1 ·
·
q
2m
2m
2m
u1 ∈V
uq ∈V
u1 ∈V
j∈[2..q−1]\{`} uj ∈V
!q−2
X
1 −δ
= n (2m)3−q ·
d2u
q
S(~v ) =
1
q
X
u∈V
33
Case 4: k ∈ [2..q − 1] and ` = q. This case is symmetric to Case 3.
The following lemma provides an upper bound on the expected runtime of the degreebased algorithm for enumerating cycles of length q ≥ 3:
Lemma 9.6 Let G = (V, E), V = [n] be drawn from the Chung-Lu distribution with degree
sequence d. For any integer q ≥ 3 let X(q) be defined by (3). Then there exists an absolute
constant C > 1 such that
!q−1
X
−q+2
2−1/(q−1)
E[X(q)] ≤ C(2m)
du
.
u∈V
We have by (3)
X
Pr[(u1 , . . . , uq ) is a path in G and ∀j ∈ [2..q] deg(u1 ) ≥ deg(uj )]
E[X(q)] ≤
Proof:
(u1 ,...,uq )∈V q :
ui distinct
where deg(u) stands for the actual degree of the node u ∈ V in the graph G. Note that while
E[deg(u)] = du , there may be deviations due to the sampling process. This fact introduces
some complications in the analysis.
Similarly to [5], we start by splitting the summation above into two. For a constant φ
(to be fixed later to 1/80) let
(9)
E[X(q)] = S1 + S2 ,
where
X
S1 :=
Pr[(u1 , . . . , uq ) is a path in G and ∀j ∈ [2..q] deg(u1 ) ≥ deg(uj )]
(u1 ,...,uq )∈V q :ui distinct,
∀j∈[2..q] du1 >φduj
X
S2 :=
Pr[(u1 , . . . , uq ) is a path in G and ∀j ∈ [2..q] deg(u1 ) ≥ deg(uj )].
(u1 ,...,uq )∈V q :ui distinct
∃j∈[2..q] s.t. du1 ≤φduj
We now bound the two summations separately.
Bounding S1
Claim 9.1
To bound S1 we use the fact that for any q-tuple of distinct u1 , . . . , uq by
Pr[(u1 , . . . , uq ) is a path in G and ∀j ∈ [2..q] deg(u1 ) ≥ deg(uj )
≤ Pr[(u1 , . . . , uq ) is a path in G] =
q−1
du1 · duq Y d2uj
·
.
2m
2m
j=2
We thus get
S1 ≤
X
(u1 ,...,uq )∈V q
∀j∈[2..q] du1 >φduj
34
q−1
du1 · duq Y d2uj
·
2m
2m
j=2
(10)
Since φduj < du1 for all potential paths (u1 , . . . , uq ) in the summation above, we have
1
1− q−1
du1
q−2
q−1
u1
=d
>
q−1
Y
(φduj )
1
q−1
=φ
q−2
q−1
q−1
Y
1
q−1
uj
d
≥φ
j=2
j=2
q−1
Y
1
duq−1
j .
j=2
It follows that
du1 ·
q−1
Y
d2uj
= du1 ·
j=2
q−1
1
Y 2− q−1
duj
j=2
·
q−1
Y
1
q−1
uj
d
j=2
q−1
1
1 Y 2− q−1
.
≤
duj
φ
j=1
Substituting this bound in (10), we get
X
S1 ≤
(u1 ,...,uq )∈V q
∀j∈[2..q] du1 >φduj
1
≤
φ
≤
X
S2 =
duq ·
(u1 ,...,uq )∈V q
1
(2m)−q+2
φ
Bounding S2
q−1
du1 · duq Y d2uj
·
2m
2m
1
≤
φ
j=2
q−1
Y
X
2−1/(q−1)
duj
1
≤
φ
2m
j=1
!q−1
1
X 2− q−1
du
.
u
duq ·
(u1 ,...,uq )∈V q
∀j∈[2..q] du1 >φduj
X
uq ∈V
duq ·
q−1
Y
j=1
q−1
Y
2−1/(q−1)
duj
2m
X du2−1/(q−1)
j
2m
j=1 uj ∈V
Recall that
X
Pr [(u1 , . . . , uq ) is a path in G and ∀j ∈ [2..q] deg(u1 ) ≥ deg(uj )]
(u1 ,...,uq )∈V q :
∃j∈[2..q] s.t. du1 ≤φduj
=
X
Pr[∀j ∈ [2..q] deg(u1 ) ≥ deg(uj ) | E(u1 , . . . , uq )] · Pr[E(u1 , . . . , uq )],
q:
(u1 ,...,uq )∈V
∃j∈[2..q] s.t. du1 ≤φduj
where we let E(u1 , . . . , uq ) := {(u1 , . . . , uq ) is a path in G}. We omit the argument of E
below to simplify notation. Note that we have
du1 ≤E[deg(u1 ) | E] ≤ du1 + 1
duj ≤E[deg(uj ) | E] ≤ duj + 2 for all j ∈ [2..q − 1]
duq ≤E[deg(uq ) | E] ≤ duq + 1.
Furthermore, deg(uj ) is a sum of independent 0/1 Bernoulli random variables even conditional on the event E. Now we would like to apply Corollary 9.4 to random variables
deg(u1 ) and deg(uj ) conditional on E, but there is one more issue: these random variables are dependent through the potential edge (u1 , uj ). To avoid this issue, we omit
the 0/1 Bernoulli random variable corresponding to this potential edge from both random variables deg(u1 ) and deg(uj ). Let deg(u1 ) and deg(uj ) be the modified random
variables. Namely, deg(u1 ) := deg(u1 ) − 1(u1 ,uj )∈E and deg(uj ) := deg(uq ) − 1(u1 ,uj )∈E ,
35
where 1(u1 ,uj )∈E is the random variable corresponding to the sampling of the potential
edge (u1 , uj ).
Note that conditional on E the random variables deg(u1 ) and deg(uj ) are independent,
and are both sums of independent Bernoulli 0/1 random variables. Note that du ≥ 1 for
every node u ∈ V , and since du1 ≤ φduj , then duj ≥ 1/φ. It follows that
E[deg(u1 ) | E] = E[deg(u1 ) − 1(u1 ,uj )∈E | E] ≤ du1 + 1 ≤ φ(duj +
1
) ≤ 2φduj
φ
(11)
On the other hand since du1 < m,
E[deg(uj ) | E] = E[deg(uj ) − 1(u1 ,uj )∈E
du1
1
≥ duj .
| E] ≥ duj 1 −
2m
2
(12)
Putting these bounds together with the assumption that φ < 1/80, we get that the
preconditions of Corollary 9.4 are satisfied, and hence
Pr[deg(u1 ) ≥ deg(uj ) | E] = Pr[deg(u1 ) − 1(u1 ,uj )∈E ≥ deg(uj ) − 1(u1 ,uj )∈E | E]
= Pr[deg(u1 ) ≥ deg(uj ) | E] ≤ 2e−Ω(duj ) .
(13)
This allows us to bound Pr[∀j ∈ [2..q] deg(u1 ) ≥ deg(uj ) | E] as follows. Let j ∗ be the
index of the highest expected degree node in {u2 , . . . , uq }, that is, j ∗ := argmaxj∈[2..q] duj .
Clearly,
Pr[∀j ∈ [2..q] deg(u1 ) ≥ deg(uj ) | E] ≤ Pr[deg(u1 ) ≥ deg(uj ∗ ) | E].
Note that if there exists j ∈ [2..q] such that du1 ≤ φduj , then also du1 ≤ φduj ∗ . Thus,
applying (13) with j = j ∗ we get
Pr[∀j ∈ [2..q] deg(u1 ) ≥ deg(uj ) | E] ≤ Pr[deg(u1 ) ≥ deg(uj ∗ ) | E]
≤ 2e
−Ω(duj ∗ )
.
Substituting
this
bound in the expression for S2 and using the fact that 2e
Q
−Ω 1q duj
2 j∈[1..q] e
(since duj ∗ ≥ duj for all j ∈ [1..q]), we get
S2 ≤
X
(2m)−q+1 · du1 e
−Ω( 1q du1 )
· duq e
−Ω( 1q duq )
(u1 ,...,uq )∈V q :
∃j∈[2..q] s.t. du1 ≤φduj
≤ (2m)−q+1 ·
X
·
Y
−Ω(duj ∗ )
≤
−Ω( 1q duj )
d2uj e
j∈[2..q−1]
du e
−Ω( 1q du )
u∈V
!2
·
X
−Ω( 1q du )
d2u e
!q−2
.
u∈V
Recall that the first two moments of the exponential distribution are constants and thus
since q is assumed to be constant we get that both
X
X
−Ω( 1q du )
−Ω( 1q du )
du e
and
d2u e
u∈V
u∈V
are constants and thus S2 = O (2m)−q+1 = o(S1 ).
36
9.6
Comparing the lower and upper bounds
We show that the bound on E[X(q)] is not much worse than our bound on E[Y (q)] for any
degree sequence d that is n−δ -balanced. We later show that our bound gives polynomially
smaller runtime if the degree sequence d satisfies the truncated power law.
Lemma 9.7 For any q ≥ 3 and any n−δ -balanced degree sequence d, for a constant δ > 0,
E[X(q)] = O(E[Y (q)]).
Proof:
We start with the bound from Lemma 9.6:
X
E[X(q)] ≤ C(2m)2−q
du
1
2− q−1
!q−1
.
u∈V
By Claim 9.2 with t = q − 1 we have
X
du
1
2− q−1
!
X
≤
u
! q−1−1
1
q−1
q−1
X
·
du
u
d2u
.
u
Substituting this into the bound above, we get that
E[X(q)] ≤ C(2m)
2−q
X
du
1
2− q−1
!q−1
u∈V
≤ C(2m)
!
X
2−q
! q−2 q−1
1
q−1
q−1
·
du
X
u
= C(2m)2−q
u
du
u
!q−2
!
X
d2u
·
X
d2u
!q−2
= C(2m)3−q
u
X
d2u
.
u
The lemma follows by comparing the above bound to the bound on E[Y (q)] provided by
Lemma 9.5 (note that the preconditions are satisfied, as d is n−δ -balanced for a constant
δ > 0 by assumption of the lemma).
We now compare the bounds from Lemma 9.6 and Lemma 9.5 when the graph G is
drawn from the Chung-Lu distribution with a degree sequence that satisfies the truncated
power law.
Lemma 9.8 Let G be a random graph drawn from the Chung-Lu distribution with degree sequence
d that satisfies the truncated power law with exponent α, for a constant α ∈ (1, 2). The following
bounds hold for any constant
q ≥ 3.
1
(1) E[Y (q)] = Ω nα−1+ 2 (2−α)q
1 1
(2a) E[X(q)] = O n 2 + 2 (2−α)(q−1) , for α ∈ (1, 2 −
1
q−1 ).
1
(2b) E[X(q)] = O (n log n), for α ∈ [2 − q−1
, 2).
Proof: We first prove the bound on E[Y (q)]. As shown later in Claim 10.1 the condition
that the degree sequence d satisfies the truncated power law implies that the degree sequence d is n−δ -balanced for a constant δ > 0, and hence preconditions of Lemma 9.5 are
37
satisfied. To apply the bound of Lemma 9.5 we bound
1
2
X
2
u du .
P
1
log2 n
log2 n
2X
X
n 2j
(2−α)j
1− 12 α
=
2
=
n
·
2
=
Θ
n
·
n
.
2αj
d2u
u
j=0
j=0
The number of edges in a graph with degree sequence that satisfies the power law with
exponent α ∈ (1, 2) satisfies
X
X
X
1
n
2(1−α)j · n = Θ(n).
m=
du = Θ
2j · αj = Θ
2
2
j≥0
j≥0
u∈V
Plugging both bounds in the bound of Lemma 9.5 we get
1
1
E[Y (q)] = Ω n−q+3 n(2− 2 α)(q−2) = Ω nα−1+ 2 (2−α)q .
To bound E[X(q)] using Lemma 9.6 we need first to bound the following summation.
X
1
2− q−1
du
1
1
2
log2 n
log2 n
2X
X
1
1
n
2− q−1
j
2−α− q−1
j
=
·2
2
=n·
.
2αj
j=0
u∈V
(14)
j=0
The sum (14) above is dominated either by the first term or by the last term, depending
1
on whether α is less than or greater than 2 − q−1
.
1
Case 1: α ∈ 1, 2 − q−1
. In this case the sum (14) is dominated by the last term and thus
bounded by
1
2
n·
log2 n
X
2
1
2−α− q−1
j
1
1
2− 12 α+ q−1
1+ 12 2−α− q−1
=O n
.
=O n
j=0
Substituting this into the bound provided by Lemma 9.6, using the fact that m = Θ(n)
we get
1 1
1
(q−1)
2− 12 α+ q−1
−q+2
E[X(q)] = O n
n
= O n 2 + 2 (2−α)(q−1)
h
Case 2: α ∈ 2 −
1
q−1 , 2
. In this case the sum (14) is dominated by the first term which is
constant and since we have 21 log n summands the sum (14) is O(n log n).
Substituting this into the bound provided by Lemma 9.6, using the fact that m = Θ(n)
we get
E[X(q)] = O n−q+2 (n log n)q−1 = O n (log n)q−1 .
Applying these bounds we conclude the following
38
Corollary 9.9 If G is a random graph drawn from the Chung-Lu distribution with degree sequence
d that satisfies the truncated power law with exponent α, for a constant α ∈ (1, 2), then E[X(q)] =
o (E[Y (q)]).
1
Proof: Using the bounds of Lemma 9.8 above. If α ∈ 1, 2 − q−1
the improvement of
the degree based algorithm over the id based (naive) algorithm is
1
1
nα−1+ 2 (2−α)q
E[Y (q)]
≥ 1 1
= n 2 (α−1)
E[X(q)]
n 2 + 2 (2−α)(q−1)
h
1
If α ∈ 2 − q−1
, 1 the improvement of the degree based algorithm over the id based
(naive) algorithm is
1
1
E[Y (q)]
nα−1+ 2 (2−α)q
= nα−2+ 2 (2−α)q · (log n)1−q
≥
q−1
E[X(q)]
n (log n)
Note that if If α ∈ 2 −
10
1
q−1 , 1
then the second term vanishes.
Power law and balanced sequences
In this section we show that a degree sequence that satisfies the truncated power law is
also balanced.
Claim 10.1 Any degree sequence d that satisfies the truncated power law with exponent
α, for
any
1
α ∈ (1, 2) that is bounded away from 1 and 2 by constants, is λ-balanced for λ = O n 2 α−1 .
Proof: The intuition behind the proof is simple: the bound that needs to be satisfied by a
balanced sequence holds because a degree sequence that satisfies the truncated power law
√
1
contains about n 2 (1−α) nodes of degree n (the largest), which means that the edge mass
is somewhat spread among these nodes, leading to the result of the lemma. We now give
the details.
Recall that by definition of the truncated power law, for each j = 0, . . . , 12 log2 n, the
number of nodes with degree Θ(2j ) is Θ(n/2αj ). For any integer s ≥ 2, we have
1
2
X
dsu = Θ
u
log2 n
X
2s·j
j=0
1
log2 n
2X
n
· αj = Θ
2(s−α)·j · n .
2
j=0
Since s ≥ 2 and α is bounded away from 2 by assumption, the summation is dominated
by the last term. Namely,
1
2
X
u
dsu = Θ
log2 n
X
2(s−α)·j · n = Θ ns/2 · n1−α/2
j=0
We now show that for for any integers a, b ≥ 1,
39
a+b
u du
P
P
P
≤ λ · ( u dau )( u dbu ).
We distinguish the following three cases:
Case 1: a ≥ 2 and b ≥ 2. In this case the derivation above with q = a and q = b gives
X
X
dbu = Θ(nb/2 · n1−α/2 ).
dau = Θ(na/2 · n1−α/2 )
u
u
It follows that
da+b
u
P au P
=Θ
( u du )( u dbu )
n(a+b)/2 · n1−α/2
na/2 · n1−α/2 · nb/2 · n1−α/2
= Θ nα/2−1 ,
P
!
which gives the result.
Case 2: a = 1, b ≥ 2. Since α is bounded away from 1 by a constant we have
1
2
X
du =
u
log2 n
X
2j ·
j=0
On the other hand, since b ≥ 2 and a + b ≥ 2,
X
da+b
= Θ(n(a+b)/2 · n1−α/2 )
u
n
= Θ(n)
2αj
X
u
(15)
dbu = Θ(nb/2 · n1−α/2 )
u
Putting the estimates above together, we get
dua+b
P au P
( u du )( u dbu )
P
=Θ
n(a+b)/2 · n1−α/2
n · nb/2 · n1−α/2
!
= Θ n−1/2 .
The result follows since nα/2−1 ≥ n−1/2 by the assumption that α ∈ (1, 2). Note that the
case a ≥ 2, b = 1 is symmetric.
Case 3: a = 1, b = 1. Then we have
X
X
X
dua+b = Θ(n(a+b)/2 · n1−α/2 )
dau = Θ(n)
dbu = Θ(n)
u
u
u
by the estimates above, and hence
dua+b
P au P
=Θ
( u du )( u dbu )
P
n(a+b)/2 · n1−α/2
n·n
= Θ n−α/2
!
The result follows since nα/2−1 ≥ n−α/2 by the assumption that α ∈ (1, 2).
40
References
[1] N. Alon, P. Dao, I. Hajirasouliha, F. Hormozdiari, and S. Sahinalp. Biomolecular network motif counting and discovery by color coding. In ISMB, 2008.
[2] N. Alon, R. Yuster, and U. Zwick. Color-coding. JACM, 42(4):844–856, 1995.
[3] D. Aparicio, P. Ribeiro, and F. da Silva. Parallel subgraph counting for multicore
architectures. In ISPA, 2014.
[4] K. Baskerville and M. Paczuski. Subgraph ensembles and motif discovery using an
alternative heuristic for graph isomorphism. Phys. Rev. E, 74:051903, 2006.
[5] J. Berry, L. Fostvedt, D. Nordman, C. Phillips, C. Seshadhri, and A. Wilson. Why do
simple algorithms for triangle enumeration work in the real world? In ITCS, 2014.
[6] M. Bhuiyan and M. Al Hassan. An iterative MapReduce based frequent subgraph
mining algorithm. IEEE Trans. Knowledge Data Eng., 27(3):608–620, 2015.
[7] M. Bhuiyan, M. Rahman, M. Rahman, and M. Al Hasan. Guise: Uniform sampling of
graphlets for large graph analysis. In ICDM, 2012.
[8] E. Bloedorn, N. Rothleder, D. DeBarr, and L. Rosen. Relational Graph Analysis with
Real-World Constraints: An Application in IRS Tax Fraud Detection. In AAAI, 2005.
[9] H. Bodlaender. Treewidth of graphs. Encyclopedia of Algorithms, 2015.
[10] S. Burt. Structural Holes and Good Ideas. The American Journal of Sociology, 110(2):349–
399, 2004.
[11] D. Chakrabarti, Y. Zhan, and C. Faloutsos. R-MAT: A Recursive Model for Graph
Mining. In SDM, 2004.
[12] D. Chen, N. Eisley, P. Heidelberger, R. M. Senger, Y. Sugawara, S. Kumar, V. Salapura,
D. Satterfield, B. Steinmacher-Burow, and J. Parker. The IBM Blue Gene/Q Interconnection Fabric. IEEE Micro, 32(1):32–43, March-April 2012.
[13] N. Chiba and T. Nishizeki. Arboricity and subgraph listing algorithms. SIAM J. Computing, 14(1):210–223, 1985.
[14] F. Chung and L. Lu. The average distances in random graphs with given expected
degrees. PNAS, 99(25):15879–15882, 2002.
[15] J. Cohen. Graph twiddling in a MapReduce world. Computing in Science and Engineering, 11(4), 2009.
[16] D. Eppstein. Parallel recognition of series-parallel graphs. Inf. Comput., 98(1):41–55,
1992.
[17] J. Grochow and M. Kellis. Network motif discovery using subgraph enumeration and
symmetry-breaking. In RECOMB, 2007.
41
[18] F. Hormozdiari, P. Berenbrink, N. Przulj, and S. Sahinalp. Not all scale-free networks
are born equal: The role of the seed graph in ppi network evolution. PLoS Computational Biology, 3(7), 2007.
[19] N. Iakovidou, S. Dimitriadis, N. Laskaris, K. Tsichlas, and Y. Manolopoulos. On the
discovery of group-consistent graph substructure patterns from brain networks. Journal of Neuroscience Methods, 213(2):2014–213, 2013.
[20] M. Jha, C. Seshadhri, and A. Pinar. Path sampling: A fast and provable method for
estimating 4-vertex subgraph counts. In WWW, 2015.
[21] T. Kolda, A. Pinar, T. Plantenga, C. Seshadhri, and C. Task. Counting triangles in
massive graphs with MapReduce. SIAM J. Scientific Computing, 36(5), 2014.
[22] M. Middendorf, E. Ziv, and C. Wiggins. Inferring network mechanisms: The
Drosophila melanogaster protein interaction network. PNAS, 102(9):3192–3197, 2005.
[23] R. Milo, S. Shen-Orr, S. Itzkovitz, N. Kashtan, D. Chklovskii, and U. Alon. Network
motifs: simple building blocks of complex networks. Science, 298(5594):824–827, October 2002.
[24] D. O’Callaghan, M. Harrigan, J. Carthy, and P. Cunningham. Network analysis of
recurring youtube spam campaigns. CoRR, abs/1201.3783, 2012.
[25] N. Przulj, D. Corneil, and I. Jurisica. Modeling interactome: scale-free or geometric?
Bioinformatics, 2004.
[26] M. Rahman, M. Bhuiyan, and M. Al Hasan. GRAFT: an approximate graphlet counting algorithm for large graph analysis. In CIKM, 2012.
[27] T. Schank and D. Wagner. Finding, counting and listing all triangles in large graphs,
an experimental study. In WEA, 2005.
[28] G. Slota and K. Madduri. Fast approximate subgraph counting and enumeration. In
ICPP, pages 210–219, 2013.
[29] G. Slota and K. Madduri. Characterizing biological networks using subgraph counting and enumeration. In PPSC, 2014.
[30] G. Slota and K. Madduri. Complex network analysis using parallel approximate motif
counting. In IPDPS, pages 405–414, 2014.
[31] S. Suri and S. Vassilvitskii. Counting triangles and the curse of the last reducer. In
WWW, 2011.
[32] G. Wu, M. Harrigan, and P. Cunningham. Classifying wikipedia articles using network motif counts and ratios. In WikiSym, 2012.
[33] Z. Zhao, M. Khan, V. Kumar, and M. Marathe. Subgraph enumeration in large social
contact networks using parallel color coding and streaming. In ICPP, 2010.
42
| 8cs.DS
|
Published as a conference paper at ICLR 2018
A BAYESIAN P ERSPECTIVE ON G ENERALIZATION AND
S TOCHASTIC G RADIENT D ESCENT
Samuel L. Smith∗ & Quoc V. Le
Google Brain
{slsmith, qvl}@google.com
arXiv:1710.06451v3 [cs.LG] 14 Feb 2018
A BSTRACT
We consider two questions at the heart of machine learning; how can we predict
if a minimum will generalize to the test set, and why does stochastic gradient
descent find minima that generalize well? Our work responds to Zhang et al.
(2016), who showed deep neural networks can easily memorize randomly labeled
training data, despite generalizing well on real labels of the same inputs. We show
that the same phenomenon occurs in small linear models. These observations are
explained by the Bayesian evidence, which penalizes sharp minima but is invariant
to model parameterization. We also demonstrate that, when one holds the learning
rate fixed, there is an optimum batch size which maximizes the test set accuracy.
We propose that the noise introduced by small mini-batches drives the parameters
towards minima whose evidence is large. Interpreting stochastic gradient descent
as a stochastic differential equation, we identify the “noise scale” g = ( N
B − 1) ≈
N/B, where is the learning rate, N the training set size and B the batch size.
Consequently the optimum batch size is proportional to both the learning rate and
the size of the training set, Bopt ∝ N . We verify these predictions empirically.
1
I NTRODUCTION
This paper shows Bayesian principles can explain many recent observations in the deep learning
literature, while also discovering practical new insights. Zhang et al. (2016) trained deep convolutional networks on ImageNet and CIFAR10, achieving excellent accuracy on both training and test
sets. They then took the same input images, but randomized the labels, and found that while their
networks were now unable to generalize to the test set, they still memorized the training labels. They
claimed these results contradict learning theory, although this claim is disputed (Kawaguchi et al.,
2017; Dziugaite & Roy, 2017). Nonetheless, their results beg the question; if our models can assign
arbitrary labels to the training set, why do they work so well in practice? Meanwhile Keskar et al.
(2016) observed that if we hold the learning rate fixed and increase the batch size, the test accuracy
usually falls. This striking result shows improving our estimate of the full-batch gradient can harm
performance. Goyal et al. (2017) observed a linear scaling rule between batch size and learning rate
in a deep ResNet, while Hoffer et al. (2017) proposed a square root rule on theoretical grounds.
Many authors have suggested “broad minima” whose curvature is small may generalize better than
“sharp minima” whose curvature is large (Chaudhari et al., 2016; Hochreiter & Schmidhuber, 1997).
Indeed, Dziugaite & Roy (2017) argued the results of Zhang et al. (2016) can be understood using
“nonvacuous” PAC-Bayes generalization bounds which penalize sharp minima, while Keskar et al.
(2016) showed stochastic gradient descent (SGD) finds wider minima as the batch size is reduced.
However Dinh et al. (2017) challenged this interpretation, by arguing that the curvature of a minimum can be arbitrarily increased by changing the model parameterization. In this work we show:
• The results of Zhang et al. (2016) are not unique to deep learning; we observe the same
phenomenon in a small “over-parameterized” linear model. We demonstrate that this phenomenon is straightforwardly understood by evaluating the Bayesian evidence in favor of
each model, which penalizes sharp minima but is invariant to the model parameterization.
∗
Work done as a member of the Google Brain Residency Program (g.co/brainresidency)
1
Published as a conference paper at ICLR 2018
• SGD integrates a stochastic differential equation whose “noise scale” g ≈ N/B, where
is the learning rate, N training set size and B batch size. Noise drives SGD away from
sharp minima, and therefore there is an optimal batch size which maximizes the test set
accuracy. This optimal batch size is proportional to the learning rate and training set size1 .
We describe Bayesian model comparison in section 2. In section 3 we replicate the observations of
Zhang et al. (2016) in a linear model, and show they are explained by the Bayesian evidence. In
section 4 we show there is an optimum batch size which maximizes the test set accuracy, and in
section 5 we derive scaling rules between the optimum batch size, learning rate, training set size and
momentum coefficient. Throughout this work, “generalization gap” refers to the gap in test accuracy
between small and large batch SGD training, not the gap in accuracy between training and test sets.
2
BAYESIAN MODEL COMPARISON
Bayesian model comparison was first applied to neural networks in MacKay (1992). We provide a
brief tutorial here, since the theory is central to the remainder of the paper. For simplicity we first
consider a classification model M with a single parameter ω, training inputs x and training labels y.
We can infer a posterior probability distribution over the parameter by applying Bayes theorem,
P (y|ω, x; M )P (ω; M )
P (ω|y, x; M ) =
(1)
P (y|x; M )
Q
−H(ω;M )
The
, where H(ω; M ) =
i P (yi |ω, xi ; M ) = e
P likelihood, P (y|ω, x; M ) =
− i ln (P (yi |ω, xi ; M )) denotes the cross-entropy of unique categorical labels. We typically
p
2
use a Gaussian prior, P (ω; M ) =
λ/2πe−λω /2 , and therefore
the posterior probability denp
sity of the parameter given the training data, P (ω|y, x; M ) ∝ λ/2πe−C(ω;M ) , where C(ω; M ) =
H(ω; M ) + λω 2 /2 denotes the L2 regularized cross entropy, or “cost function”, and λ is the regularization coefficient. The value ω0 which minimizes the cost function lies at the maximum of this
posterior. To predict an unknown label yt of a new input xt , we should compute the integral,
Z
P (yt |xt , y, x; M ) =
dω P (yt |ω, xt ; M )P (ω|y, x; M )
(2)
R
dω P (yt |ω, xt ; M )e−C(ω;M )
R
=
.
(3)
dω e−C(ω;M )
However these integrals are dominated by the region near ω0 , and since P (yt |ω, xt ; M ) is smooth
we usually approximate P (yt |xt , x, y; M ) ≈ P (yt |ω0 , xt ; M ). Having minimized C(ω; M ) to find
ω0 , we now wish to compare two different models and select the best one. The probability ratio,
P (M1 |y, x)
P (y|x; M1 ) P (M1 )
=
.
(4)
P (M2 |y, x)
P (y|x; M2 ) P (M2 )
The second factor on the right is the prior ratio, which describes which model is most plausible. To
avoid unnecessary subjectivity, we usually set this to 1. Meanwhile the first factor on the right is the
evidence ratio, which controls how much the training data changes our prior beliefs. Germain et al.
(2016) showed that maximizing the evidence (or “marginal likelihood”) minimizes a PAC-Bayes
generalization bound. To compute it, we evaluate the normalizing constant of equation 1,
Z
P (y|x; M ) =
dω P (y|ω, x; M )P (ω; M )
(5)
r
Z
λ
=
dω e−C(ω;M ) .
(6)
2π
Notice that the evidence is computed by integrating out the parameters; and consequently it is invariant to the model parameterization. Since this integral is dominated by the region near the minimum
ω0 , we can estimate the evidence by Taylor expanding C(ω; M ) ≈ C(ω0 ) + C 00 (ω0 )(ω − ω0 )2 /2,
r
Z
00
2
λ
−C(ω0 )
P (y|x; M ) ≈ e
dω e−C (ω0 )(ω−ω0 ) /2
(7)
2π
1
= exp − C(ω0 ) + ln (C 00 (ω0 )/λ) .
(8)
2
1
Equivalently, there is an optimal learning rate proportional to the batch size and the training set size.
2
Published as a conference paper at ICLR 2018
Within this “Laplace” approximation, the evidence is controlled by the value of the cost function at
the minimum, and by the logarithm of the ratio of the curvature about this minimum compared to
the regularization constant. Thus far we have considered models of a single parameter; in realistic
p
1/2
models with many parameters P (y|x; M ) ≈ λ 2 e−C(ω0 ) /|∇∇C(ω)|ω0 , where |∇∇C(ω)|ω0 is the
determinant of the Hessian, and p denotes the number of model parameters
& Raftery, 1995).
Q(Kass
p
The determinant of the Hessian is simply the product of its eigenvalues, ( i=1 λi ), and thus,
!)
(
p
1X
ln(λi /λ)
.
(9)
P (y|x; M ) ≈ exp − C(ω0 ) +
2 i=1
p
1/2
The contribution (λ 2 /|∇∇C(ω)|ω0 ) is often called the “Occam factor”, because it enforces Occam’s razor; when two models describe the data equally well, the simpler model is usually better
(Gull, 1988). Minima with low curvature are simple, because the parameters do not have to be finetuned to fit the data. Intuitively, the Occam factor describes the fraction of the prior parameter space
consistent with the data. Since this fraction is always less than one, we propose to approximate
equation 9 away from local minima by only performing the summation over eigenvalues λi ≥ λ.
The evidence can be reframed in the language of information theory, whereby Occam’s factor penalizes the amount of information the model must learn about the parameters to accurately model the
training data (Hinton & Van Camp, 1993; Achille & Soatto, 2017; Shwartz-Ziv & Tishby, 2017).
In this work, we will compare the evidence against a null model which assumes the labels are entirely
random, assigning equal probability to each class. This unusual model has no parameters, and so the
evidence is controlled by the likelihood alone, P (y|x; N U LL) = (1/n)N = e−N ln (n) , where n
denotes the number of model classes and N the number of training labels. Thus the evidence ratio,
P (y|x; M )
= e−E(ω0 ) ,
P (y|x; N U LL)
(10)
P
Where E(ω0 ) = C(ω0 ) + (1/2) i ln(λi /λ) − N ln(n) is the log evidence ratio in favor of the null
model. Clearly, we should only assign any confidence to the predictions of our model if E(ω0 ) < 0.
The evidence supports the intuition that broad minima generalize better than sharp minima, but
unlike the curvature it does not depend on the model parameterization. Dinh et al. (2017) showed
one can increase the Hessian eigenvalues by rescaling the parameters, but they must simultaneously
rescale the regularization coefficients, otherwise the model changes. Since Occam’s factor arises
from the log ratio, ln (λi /λ), these two effects cancel out2 . It is difficult to evaluate the evidence for
deep networks, as we cannot compute the Hessian of millions of parameters. Additionally, neural
networks exhibit many equivalent minima, since we can permute the hidden units without changing
the model. To compute the evidence we must carefully account for this “degeneracy”. We argue
these issues are not a major limitation, since the intuition we build studying the evidence in simple
cases will be sufficient to explain the results of both Zhang et al. (2016) and Keskar et al. (2016).
3
BAYES THEOREM AND GENERALIZATION
Zhang et al. (2016) showed that deep neural networks generalize well on training inputs with informative labels, and yet the same model can drastically overfit on the same input images when the
labels are randomized; perfectly memorizing the training set. To demonstrate that these observations
are not unique to deep networks, let’s consider a far simpler model; logistic regression. We form a
small balanced training set comprising 800 images from MNIST, of which half have true label “0”
and half true label “1”. Our test set is also balanced, comprising 5000 MNIST images of zeros and
5000 MNIST images of ones. There are two tasks. In the first task, the labels of both the training and
test sets are randomized. In the second task, the labels are informative, matching the true MNIST
labels. Since the images contain 784 pixels, our model has just 784 weights and 1 bias.
We show the accuracy of the model predictions on both the training and test sets in figure 1. When
trained on the informative labels, the model generalizes well to the test set, so long as it is weakly
regularized. However the model also perfectly memorizes the random labels, replicating the observations of Zhang et al. (2016) in deep networks. No significant improvement in model performance
2
Note however that while the evidence itself is invariant to model parameterization, one can find reparameterizations which change the approximate evidence after the Laplace approximation.
3
Published as a conference paper at ICLR 2018
(a)
(b)
Figure 1: Prediction accuracy and mean training set margin as a function of regularization coefficient, for a logistic regression trained on random (a) and informative (b) labels of the same inputs.
The weakly regularized model generalizes well on informative labels but memorizes random labels.
is observed as the regularization coefficient increases. For completeness, we also evaluate the mean
margin between training examples and the decision boundary. For both random and informative
labels, the margin drops significantly as we reduce the regularization coefficient. When weakly
regularized, the mean margin is roughly 50% larger for informative labels than for random labels.
Now consider figure 2, where we plot the mean cross-entropy of the model predictions, evaluated on
both training and test sets, as well as the Bayesian log evidence ratio defined in the previous section.
Looking first at the random label experiment in figure 2a, while the cross-entropy on the training
set vanishes when the model is weakly regularized, the cross-entropy on the test set explodes. Not
only does the model make random predictions, but it is extremely confident in those predictions. As
the regularization coefficient is increased the test set cross-entropy falls, settling at ln 2, the crossentropy of assigning equal probability to both classes. Now consider the Bayesian evidence, which
we evaluate on the training set. The log evidence ratio is large and positive when the model is
weakly regularized, indicating that the model is exponentially less plausible than assigning equal
probabilities to each class. As the regularization parameter is increased, the log evidence ratio falls,
but it is always positive, indicating that the model can never be expected to generalize well.
Now consider figure 2b (informative labels). Once again, the training cross-entropy falls to zero
when the model is weakly regularized, while the test cross-entropy is high. Even though the model
makes accurate predictions, those predictions are overconfident. As the regularization coefficient
increases, the test cross-entropy falls below ln 2, indicating that the model is successfully generalizing to the test set. Now consider the Bayesian evidence. The log evidence ratio is large and
positive when the model is weakly regularized, but as the regularization coefficient increases, the
(a)
(b)
Figure 2: The cross-entropy and log evidence ratio, evaluated on random (a) or informative (b)
labels. The evidence, evaluated on the training set, is strongly correlated with the test cross-entropy.
4
Published as a conference paper at ICLR 2018
log evidence ratio drops below zero, indicating that the model is exponentially more plausible than
assigning equal probabilities to each class. As we further increase the regularization, the log evidence ratio rises to zero while the test cross-entropy rises to ln 2. Test cross-entropy and Bayesian
evidence are strongly correlated, with minima at the same regularization strength.
Bayesian model comparison has explained our results in a logistic regression. Meanwhile, Krueger
et al. (2017) showed the largest Hessian eigenvalue also increased when training on random labels
in deep networks, implying the evidence is falling. We conclude that Bayesian model comparison
is quantitatively consistent with the results of Zhang et al. (2016) in linear models where we can
compute the evidence, and qualitatively consistent with their results in deep networks where we
cannot. Dziugaite & Roy (2017) recently demonstrated the results of Zhang et al. (2016) can also
be understood by minimising a PAC-Bayes generalization bound which penalizes sharp minima.
4
BAYES THEOREM AND STOCHASTIC GRADIENT DESCENT
We showed above that generalization is strongly correlated with the Bayesian evidence, a weighted
combination of the depth of a minimum (the cost function) and its breadth (the Occam factor).
Consequently Bayesians often add isotropic Gaussian noise to the gradient (Welling & Teh, 2011).
In appendix A, we show this drives the parameters towards broad minima whose evidence is large.
The noise introduced by small batch training is not isotropic, and its covariance matrix is a function
of the parameter values, but empirically Keskar et al. (2016) found it has similar effects, driving the
SGD away from sharp minima. This paper therefore proposes Bayesian principles also account for
the “generalization gap”, whereby the test set accuracy often falls as the SGD batch size is increased
(holding all other hyper-parameters constant). Since the gradient drives the SGD towards deep
minima, while noise drives the SGD towards broad minima, we expect the test set performance to
show a peak at an optimal batch size, which balances these competing contributions to the evidence.
We were unable to observe a generalization gap in linear models (since linear models are convex
there are no sharp minima to avoid). Instead we consider a shallow neural network with 800 hidden
units and RELU hidden activations, trained on MNIST without regularization. We use SGD with a
momentum parameter of 0.9. Unless otherwise stated, we use a constant learning rate of 1.0 which
does not depend on the batch size or decay during training. Furthermore, we train on just 1000
images, selected at random from the MNIST training set. This enables us to compare small batch
to full batch training. We emphasize that we are not trying to achieve optimal performance, but to
study a simple model which shows a generalization gap between small and large batch training.
In figure 3, we exhibit the evolution of the test accuracy and test cross-entropy during training. Our
small batches are composed of 30 images, randomly sampled from the training set. Looking first
at figure 3a, small batch training takes longer to converge, but after a thousand gradient updates
a clear generalization gap in model accuracy emerges between small and large training batches.
Now consider figure 3b. While the test cross-entropy for small batch training is lower at the end
(a)
(b)
Figure 3: The evolution during training of the test accuracy (a), and the test set cross-entropy (b).
Full batches are composed of 1000 images, while small batches comprise 30 images.
5
Published as a conference paper at ICLR 2018
(a)
(b)
Figure 4: The test accuracy for a range of batch sizes, during training (a) and after 10000 steps (b).
(a)
(b)
Figure 5: a) The test set accuracy as a function of batch size, for a range of learning rates . The
performance peak shifts to the right as we increase , but the overall performance falls once & 3.
b) The best observed batch size is proportional to the learning rate across two orders of magnitude.
of training; the cross-entropy of both small and large training batches is increasing, indicative of
over-fitting. Both models exhibit a minimum test cross-entropy, although after different numbers of
gradient updates. Intriguingly, we show in appendix B that the generalization gap between small
and large batch training shrinks significantly when we introduce L2 regularization.
From now on we focus on the test set accuracy (since this converges as the number of gradient
updates increases). In figure 4a, we exhibit training curves for a range of batch sizes between 1 and
1000. We find that the model cannot train when the batch size B / 10. In figure 4b we plot the mean
test set accuracy after 10000 training steps. A clear peak emerges, indicating that there is indeed
an optimum batch size which maximizes the test accuracy, consistent with Bayesian intuition. The
results of Keskar et al. (2016) focused on the decay in test accuracy above this optimum batch size.
5
S TOCHASTIC DIFFERENTIAL EQUATIONS AND THE SCALING RULES
We showed above that the test accuracy peaks at an optimal batch size, if one holds the other SGD
hyper-parameters constant. We argued that this peak arises from the tradeoff between depth and
breadth in the Bayesian evidence. However it is not the batch size itself which controls this tradeoff,
but the underlying scale of random fluctuations in the SGD dynamics. We now identify this SGD
“noise scale”, and use it to derive three scaling rules which predict how the optimal batch size
depends on the learning rate, training set size and momentum coefficient. A gradient update,
!!
dĈ
dC
dC
+
−
,
(11)
∆ω = −
N dω
dω
dω
6
Published as a conference paper at ICLR 2018
(a)
(b)
Figure 6: a) The test accuracy as a function of batch size, for a range of training set sizes. To reduce
noise, we average each curve over five experiments. The performance peak shift to the right as we
increase the size of the training set. Unsurprisingly, the overall model performance also improves.
b) The best observed batch size is proportional to the size of the training set once N & 20000.
PN
dĈ
i
where is the learning rate, N the training set size, dC
= i=1 dC
dω
dω the true gradient, and dω =
P
B
dCi
N
on a mini-batch. The expected gradient of a single
i=1 dω the estimated gradient
B
D evaluated
E
2
dCi
dCi dCj
1 dC
example, dω = N dω , while dω dω = N1 dC
+ F (ω)δij . F (ω) is a matrix describing the
dω
gradient covariances, which are a function of the current parameter values. We adopt the central limit
Ĉ
theorem and model the gradient error α = ( ddω
− dC
dω ) with Gaussian random noise (We discuss this
approximation briefly in appendix C). It is easy to show that hαi = 0, while hα2 i = N ( N
B −1)F (ω).
Typically N B, such that hα2 i ≈ N 2 F (ω)/B. To continue, we interpret equation 11 as the
discrete update of a stochastic differential equation (Li et al., 2017; Gardiner, 1985),
dC
dω
=−
+ η(t),
(12)
dt
dω
Where t is a continuous variable, η(t) represents noise, hη(t)i = 0 and hη(t)η(t0 )i = gF (ω)δ(t−t0 ).
The constant g controls the scale of random fluctuations in the dynamics. To relate this differential
R /N dω
R /N
dC
equation to the SGD, we compute a gradient update ∆ω = 0
η(t)dt.
dt dt = − N dω + 0
Finally, to measure g, we equate the variance in this gradient update to the variance in equation 11,
2
N
hα2 i = 2 ( − 1)F (ω)/N
N
B
Z /N
Z /N Z /N
= h(
dt η(t))2 i =
dt
dt0 hη(t)η(t0 )i = gF (ω)/N. (13)
0
0
0
Rearranging, the SGD noise scale g = ( N
B − 1) ≈ N/B. The noise scale falls when the batch
size increases, consistent with our earlier observation of an optimal batch size Bopt while holding
the other hyper-parameters fixed. Notice that one would equivalently observe an optimal learning
rate if one held the batch size constant. A similar analysis of the SGD was recently performed by
Mandt et al. (2017), although their treatment only holds near local minima where the covariances
F (ω) are stationary. Our analysis holds throughout training, which is necessary since Keskar et al.
(2016) found that the beneficial influence of noise was most pronounced at the start of training.
When we vary the learning rate or the training set size, we should keep the noise scale fixed, which
implies that Bopt ∝ N . In figure 5a, we plot the test accuracy as a function of batch size after
(10000/) training steps, for a range of learning rates. Exactly as predicted, the peak moves to the
right as increases. Additionally, the peak test accuracy achieved at a given learning rate does not
begin to fall until ∼ 3, indicating that there is no significant discretization error in integrating the
stochastic differential equation below this point. Above this point, the discretization error begins to
dominate and the peak test accuracy falls rapidly. In figure 5b, we plot the best observed batch size
as a function of learning rate, observing a clear linear trend, Bopt ∝ . The error bars indicate the
distance from the best observed batch size to the next batch size sampled in our experiments.
7
Published as a conference paper at ICLR 2018
(a)
(b)
Figure 7: a) The test set accuracy as a function of batch size for a range of momentum coefficients.
As expected, the peak moves to the right as the momentum coefficient increases. b) The best observed batch size for a range of momentum coefficients. The green curve exhibits the scaling rule.
This scaling rule allows us to increase the learning rate with no loss in test accuracy and no increase
in computational cost, simply by simultaneously increasing the batch size. We can then exploit
increased parallelism across multiple GPUs, reducing model training times (Goyal et al., 2017). A
similar scaling rule was independently proposed by Jastrzebski et al. (2017) and Chaudhari & Soatto
(2017), although neither work identifies the existence of an optimal noise scale. A number of authors
have proposed adjusting the batch size adaptively during training (Friedlander & Schmidt, 2012;
Byrd et al., 2012; De et al., 2017), while Balles et al. (2016) proposed linearly coupling the learning
rate and batch size within this framework. In Smith et al. (2017), we show empirically that decaying
the learning rate during training and increasing the batch size during training are equivalent.
In figure 6a we exhibit the test set accuracy as a function of batch size, for a range of training set sizes
after 10000 steps ( = 1 everywhere). Once again, the peak shifts right as the training set size rises,
although the generalization gap becomes less pronounced as the training set size increases. In figure
6b, we plot the best observed batch size as a function of training set size; observing another linear
trend, Bopt ∝ N . This scaling rule could be applied to production models, progressively growing
the batch size as new training data is collected. We expect production datasets to grow considerably
over time, and consequently large batch training is likely to become increasingly common.
Finally, in appendix D we extend our analysis to SGD with momentum, identifying the noise scale,
N
, where m denotes the momentum coefficient. Notice that this reduces to the noise
g ≈ B(1−m)
scale of conventional SGD as m → 0. When m > 0, we obtain an additional scaling rule Bopt ∝
1/(1 − m). This scaling rule predicts that the optimal batch size will increase when the momentum
coefficient is increased. In figure 7a we plot the test set performance as a function of batch size after
10000 gradient updates ( = 1 everywhere), for a range of momentum coefficients. In figure 7b, we
plot the best observed batch size as a function of the momentum coefficient, and fit our results to the
scaling rule above; obtaining remarkably good agreement. We propose a simple heuristic for tuning
the batch size, learning rate and momentum coefficient in appendix E.
6
C ONCLUSIONS
Just like deep neural networks, linear models which generalize well on informative labels can memorize random labels of the same inputs. These observations are explained by the Bayesian evidence,
which is composed of the cost function and an “Occam factor”. The Occam factor penalizes sharp
minima but it is invariant to changes in model parameterization. Mini-batch noise drives SGD away
from sharp minima, and therefore there is an optimum batch size which maximizes the test accuracy.
Interpreting SGD as the discretization of a stochastic differential equation, we predict this optimum
batch size should scale linearly with both the learning rate and the training set size, Bopt ∝ N .
We derive an additional scaling rule, Bopt ∝ 1/(1 − m), between the optimal batch size and the
momentum coefficient. We verify these scaling rules empirically and discuss their implications.
8
Published as a conference paper at ICLR 2018
ACKNOWLEDGMENTS
We thank Pieter-Jan Kindermans, Prajit Ramachandran, Jascha Sohl-Dickstein, Jon Shlens, Kevin
Murphy, Samy Bengio, Yasaman Bahri and Saeed Saremi for helpful comments on the manuscript.
R EFERENCES
Alessandro Achille and Stefano Soatto. On the emergence of invariance and disentangling in deep
representations. arXiv preprint arXiv:1706.01350, 2017.
Lukas Balles, Javier Romero, and Philipp Hennig. Coupling adaptive batch sizes with learning rates.
arXiv preprint arXiv:1612.05086, 2016.
Richard H Byrd, Gillian M Chin, Jorge Nocedal, and Yuchen Wu. Sample size selection in optimization methods for machine learning. Mathematical programming, 134(1):127–155, 2012.
Pratik Chaudhari and Stefano Soatto. Stochastic gradient descent performs variational inference,
converges to limit cycles for deep networks. arXiv preprint arXiv:1710.11029, 2017.
Pratik Chaudhari, Anna Choromanska, Stefano Soatto, and Yann LeCun. Entropy-SGD: Biasing
gradient descent into wide valleys. arXiv preprint arXiv:1611.01838, 2016.
Soham De, Abhay Yadav, David Jacobs, and Tom Goldstein. Automated inference with adaptive
batches. In Artificial Intelligence and Statistics, pp. 1504–1513, 2017.
Laurent Dinh, Razvan Pascanu, Samy Bengio, and Yoshua Bengio. Sharp minima can generalize
for deep nets. arXiv preprint arXiv:1703.04933, 2017.
Gintare Karolina Dziugaite and Daniel M Roy. Computing nonvacuous generalization bounds for
deep (stochastic) neural networks with many more parameters than training data. arXiv preprint
arXiv:1703.11008, 2017.
Michael P Friedlander and Mark Schmidt. Hybrid deterministic-stochastic methods for data fitting.
SIAM Journal on Scientific Computing, 34(3):A1380–A1405, 2012.
Crispin W Gardiner. Handbook of Stochastic Methods, volume 4. Springer Berlin, 1985.
Pascal Germain, Francis Bach, Alexandre Lacoste, and Simon Lacoste-Julien. PAC-bayesian theory
meets bayesian inference. In Advances in Neural Information Processing Systems, pp. 1884–
1892, 2016.
Priya Goyal, Piotr Dollár, Ross Girshick, Pieter Noordhuis, Lukasz Wesolowski, Aapo Kyrola, Andrew Tulloch, Yangqing Jia, and Kaiming He. Accurate, large minibatch SGD: Training imagenet
in 1 hour. arXiv preprint arXiv:1706.02677, 2017.
Stephen F Gull. Bayesian inductive inference and maximum entropy. In Maximum-entropy and
Bayesian methods in science and engineering, pp. 53–74. Springer, 1988.
Geoffrey E Hinton and Drew Van Camp. Keeping the neural networks simple by minimizing the
description length of the weights. In Proceedings of the sixth annual conference on Computational
learning theory, pp. 5–13. ACM, 1993.
Sepp Hochreiter and Jürgen Schmidhuber. Flat minima. Neural Computation, 9(1):1–42, 1997.
Elad Hoffer, Itay Hubara, and Daniel Soudry. Train longer, generalize better: closing the generalization gap in large batch training of neural networks. arXiv preprint arXiv:1705.08741, 2017.
Stanisław Jastrzebski, Zachary Kenton, Devansh Arpit, Nicolas Ballas, Asja Fischer, Yoshua
Bengio, and Amos Storkey. Three factors influencing minima in SGD. arXiv preprint
arXiv:1711.04623, 2017.
Robert E Kass and Adrian E Raftery. Bayes factors. Journal of the american statistical association,
90(430):773–795, 1995.
9
Published as a conference paper at ICLR 2018
Kenji Kawaguchi, Leslie Pack Kaelbling, and Yoshua Bengio. Generalization in deep learning.
arXiv preprint arXiv:1710.05468, 2017.
Nitish Shirish Keskar, Dheevatsa Mudigere, Jorge Nocedal, Mikhail Smelyanskiy, and Ping Tak Peter Tang. On large-batch training for deep learning: Generalization gap and sharp minima. arXiv
preprint arXiv:1609.04836, 2016.
David Krueger, Nicolas Ballas, Stanislaw Jastrzebski, Devansh Arpit, Maxinder S Kanwal, Tegan
Maharaj, Emmanuel Bengio, Asja Fischer, and Aaron Courville. Deep nets don’t learn via memorization. ICLR Workshop, 2017.
Qianxiao Li, Cheng Tai, and E Weinan. Stochastic modified equations and adaptive stochastic
gradient algorithms. In International Conference on Machine Learning, pp. 2101–2110, 2017.
David JC MacKay. A practical bayesian framework for backpropagation networks. Neural computation, 4(3):448–472, 1992.
Stephan Mandt, Matthew D Hoffman, and David M Blei. Stochastic gradient descent as approximate
bayesian inference. arXiv preprint arXiv:1704.04289, 2017.
Ravid Shwartz-Ziv and Naftali Tishby. Opening the black box of deep neural networks via information. arXiv preprint arXiv:1703.00810, 2017.
Samuel L. Smith, Pieter-Jan Kindermans, and Quoc V. Le. Don’t decay the learning rate, increase
the batch size. arXiv preprint arXiv:1711.00489, 2017.
Max Welling and Yee W Teh. Bayesian learning via stochastic gradient langevin dynamics. In
Proceedings of the 28th International Conference on Machine Learning (ICML-11), pp. 681–688,
2011.
Chiyuan Zhang, Samy Bengio, Moritz Hardt, Benjamin Recht, and Oriol Vinyals. Understanding
deep learning requires rethinking generalization. arXiv preprint arXiv:1611.03530, 2016.
A
BAYESIAN POSTERIOR SAMPLING AND L ANGEVIN DYNAMICS
Instead of minimizing the cost function, Bayesian usually prefer to sample parameter values from
the posterior (MacKay, 1992),
P (ω|y, x; M ) ∝ e−C(ω;M ) ,
(14)
where C(ω; M ) is the regularized summed cost function, as shown in section 2 of the main text. It
is well known that one can sample this posterior by simulating the overdamped Langevin equation
(Gardiner, 1985), which is described by the stochastic differential equation,
dω
dC
=−
+ η(t),
dt
dω
(15)
where t is a continuous variable, and η(t) describes Gaussian noise with mean hη(t)i = 0 and
variance hη(t)η(t0 )i = 2T Iδ(t−t0 ). The matrix I denotes the identity, while T is the “temperature”.
Notice this Langevin equation is extremely similar to the stochastic differential equation of SGD,
discussed in section 5 of the main text. Indeed, if the gradient covariances F (ω) were stationary
and proportional to the identity, then the SGD would integrate an overdamped Langevin equation
with temperature proportional to the SGD noise scale g. As t → ∞, the probability of sampling any
particular parameter vector ω from the Langevin equation, P (ω, t → ∞) ∝ e−C/T .
We obtain posterior samples if T = 1. In order to draw posterior samples in practice, we repeatedly
integrate the Langevin equation (at temperature T = 1), over a finite step t → t + /N ,
Z t+ N
dC
∆ω = −
+
η(t)dt
(16)
N dω
t
dC
= −
+ α,
(17)
N dω
10
Published as a conference paper at ICLR 2018
where α denotes a Gaussian random variable with mean hαi = 0 and variance hα2 i = 2I/N ,
which introduces isotropic noise to the gradient update as described in section 4 of the main text.
Note that, since C(ω; M ) denotes the summed cost function, we chose to scale our step size by the
training set size N . This also matches our treatment of SGD in section 5 of the main text. The
larger the step size , the greater the discretization error, but if is sufficiently small and we iterate
equation 17 sufficiently many times, we will obtain valid samples from the posterior.
Since the probability of sampling any given parameter vector ω is proportional to the posterior, the
probability of sampling a parameter vector belonging to any given local minimum is proportional to
the integral of the posterior over the bowl of attraction D which surrounds that minimum.
Z
P (ω ∈ D, t → ∞) ∝
dω e−C(ω;M ) .
(18)
D
Meanwhile we showed in section 2 of the main text that the evidence in favor of a model is proportional to the integral of the posterior over all parameter space.
Z
P (y|x; M ) ∝ dω e−C(ω;M ) .
(19)
As we discussed, this evidence is dominated by the contributions to the integral near local minima.
In a convex model, there is only one such minimum; which allows us to accurately estimate the
model evidence. Meanwhile, in non-convex models, there are many such minima, and so we can
instead define the evidence as a sum over local evidences in favor of each minimum,
X
P (y|x; M ) =
P (y|x, ω ∈ Di ; M ),
(20)
i
where we define the evidence in favor of a minimum as the integral over the local bowl of attraction,
Z
P (y|x, ω ∈ D; M ) ∝
dω e−C(ω;M ) .
(21)
D
Since the combined bowls of attraction of all the minima perfectly tile the entire parameter space,
equations 19 and 20 are equivalent. Meanwhile, equating equations 18 and 21 we find that, when one
performs Bayesian posterior sampling, the probability of sampling the parameter vector from a local
minimum is proportional to the evidence in favor of that minimum. This demonstrates that bayesian
posterior samples are biased in favor of local minima whose evidence is large, which explains why
a single posterior sample ωp often achieves lower test error than the cost function minimum ω0 .
B
T HE EFFECT OF REGULARIZATION ON THE GENERALIZATION GAP
In the experiments of section 4 of the main text, the L2 regularization coefficient λ = 0. In figure
8, we plot the evolution of the training curves when λ = 0.1, for both small batch and full batch
(a)
(b)
Figure 8: The mean test accuracy (a) and the mean test cross-entropy (b) of a regularized model
during training. While full batch training takes longer to converge, it achieves similar performance
at long times. The noise inherent in small batch training causes the performance to fluctuate.
11
Published as a conference paper at ICLR 2018
(a)
(b)
Figure 9: The gradient distribution of a randomly selected parameter in the softmax layer, when
measured over a single training example (a), and when averaged over mini-batches of 30 images (b).
training. Excluding the regularization parameter, these experiments are identical to figure 3. To our
surprise, regularized full batch training took longer to converge than small batch training. In another
surprise, regularization significantly reduced the size of the generalization gap. While large batch
regularized training achieves slightly lower test set accuracy than unregularized small batch training,
it also achieves lower test cross-entropy. The test cross-entropy of our regularized models does not
degrade after many gradient updates, removing the need for early stopping.
C
T HE G AUSSIAN APPROXIMATION TO THE MINI - BATCH ERROR
In section 5 of the main text, we approximated the difference between the full batch gradient and
Ĉ
the mini-batch gradient estimate, α = ( ddω
− dC
dω ), by a Gaussian random variable. This enabled
us to derive the scaling rules, which we verified empirically. We motivated this assumption by
reference to the central limit theorem, which states that the gradient error will tend towards Gaussian
noise as {N → ∞, B → ∞, B N }, so long as the distribution of gradients over individual
training examples does not have heavy tails. In practice neither N nor B is infinite, and the gradient
distribution may be heavy tailed, especially when gradients are sparse. Nonetheless the central limit
theorem tends to be surprisingly robust in practice, and is consequently widely used.
It is beyond the scope of this work to perform a thorough study of the gradient noise distribution
in deep networks. However as a brief proof of principle, we present the distribution of the gradient
immediately after random initialization in figure 9, for the shallow neural network discussed in
sections 4 and 5 of the main text. In figure 9a, we present the distribution over the individual
training examples, of the gradient of a single matrix element in the softmax output layer, chosen
randomly. The distribution is double peaked and clearly not Gaussian. However in figure 7b, we plot
the distribution of the gradient of the same matrix element, when averaged over randomly sampled
mini-batches of 30 images (without replacement). A single peak emerges, and while the distribution
is still slightly skewed, it is clearly already approaching the Gaussian limit. We conclude that the
Gaussian approximation is likely to be reasonable for commonly used mini-batch sizes.
D
D ERIVING THE SCALING RULES FOR SGD WITH MOMENTUM
Momentum simulates a generalized Langevin equation (with structured fluctuations),
dω dC
d2 ω
= −λ
−
+ η(t).
2
dt
dt
dω
(22)
λ is the “damping coefficient” and η(t) describes Gaussian noise, whose statistics hη(t)i = 0 and
hη(t)η(t0 )i = gλF (w)δ(t − t0 ). As before, the coefficient g describes the scale of random fluctuations in the dynamics, and F (ω) describes the gradient covariances between parameters. We include
a factor of λ in the noise variance to satisfy the fluctuation-dissipation theorem, which states that
12
Published as a conference paper at ICLR 2018
we can vary the damping coefficient without changing the probability of sampling any particular
configuration of parameters in the limit t → ∞, if we proportionally increase the noise variance.
To relate this Langevin equation to the usual momentum equations, we first re-express it as two
coupled first order differential equations,
dp
dC
= −λp −
+ η(t),
(23)
dt
dω
dω
= p.
(24)
dt
Integrating over a single step ∆t/N ,
∆t dC
∆p = −(λ∆t/N )p −
+ η,
(25)
N dω
∆ω = p∆t/N.
(26)
Where now hηi = 0 and hη 2 i = g∆tλF (w)/N . We define the accumulation A = p/∆t,
η
1 dC
+
,
(27)
∆A = −(λ∆t/N )A −
N dω
∆t
∆ω = (∆t)2 A/N.
(28)
These equations can be compared to the TensorFlow update equations for momentum,
1 dC
+α ,
(29)
∆A = (m − 1)A −
N dω
∆ω = A.
(30)
Ĉ
Where α = ddω
− dC
denotes the error in the gradient update. As discussed in the main text,
dω
we can approximate this error as Gaussian noise with statistics hαi = 0 and hα2 i ≈ N 2 F (ω)/B.
Equations 27 and 28 match equations 29 and 30 if the step size = (∆t)2 /N , and the momentum
parameter m = (1 − λ∆t/N ). Finally we equate the noise by setting hα2 i/N 2 = hη 2 i/(∆t)2 , and
solve for the noise scale g to obtain,
(∆t)2
∆tN
=
(31)
g ≈
Bλ
B(1 − m)
N
=
.
(32)
B(1 − m)
As observed in the main text, if we wish to keep the scale of random fluctuations constant, then
we should scale the batch size B ∝ N . We also predict an additional scaling relation between
the batch size and the momentum parameter, B ∝ 1/(1 − m). Note that one can also interpret
ef f = /(1 − m) as the “effective learning rate”.
E
H OW TO ACHIEVE LARGE BATCH TRAINING
Here we propose a simple heuristic for tuning the batch size, learning rate and momentum parameter; in order to maximize both test accuracy and batch size (enabling parallel training across many
machines). Note that this is only worthwhile if one expects to retrain a model many times.
1. Set the learning rate to 0.1 and the momentum coefficient to 0.9. Run experiments at a range
of batch sizes on a logarithmic scale, and identify the optimal batch size which maximizes
the validation set accuracy. If training is not stable, reduce the learning rate, and repeat.
2. Repeatedly increase the batch size by a factor of 3, while scaling the learning rate ∝ B,
until the validation set accuracy starts to fall. Then repeatedly increase the batch size by
a factor of 3, while scaling the momentum coefficient (1 − m) ∝ 1/B, until either the
validation set accuracy falls or the batch size reaches the limits of your hardware.
3. Having identified the final learning rate and momentum parameter, retune the batch size on
a linear scale in the local neighborhood of the current batch size.
We believe that this simple procedure will increase the test accuracy, reduce the cost of tuning hyperparameters, and significantly reduce the final number of gradient updates required to train a model.
13
| 2cs.AI
|
Approximation of Continuous-Time Infinite-Horizon
Optimal Control Problems Arising in Model Predictive
Control - Supplementary Notes
arXiv:1608.08823v1 [math.OC] 31 Aug 2016
Michael Muehlebach and Raffaello D’Andrea∗
Abstract
These notes present preliminary results regarding two different approximations of linear infinitehorizon optimal control problems arising in model predictive control. Input and state trajectories are
parametrized with basis functions and a finite dimensional representation of the dynamics is obtained via
a Galerkin approach. It is shown that the two approximations provide lower, respectively upper bounds
on the optimal cost of the underlying infinite dimensional optimal control problem. These bounds get
tighter as the number of basis functions is increased. In addition, conditions guaranteeing convergence
to the cost of the underlying problem are provided.
1
Introduction
Model predictive control (MPC) takes input and state constraints fully into account and is therefore a
promising control strategy with various applications. The standard MPC approach relies on discrete dynamics and a finite prediction horizon, which leads inevitably to issues related to closed-loop stability.
In [5], an approximation of the underlying infinite dimensional infinite-horizon optimal control problem
has been proposed, which is based on a parametrization of input and state trajectories with basis functions.
The infinite prediction horizon is maintained, and therefore closed-loop stability and recursive feasibility arise naturally from the problem formulation. Moreover, it is conjectured that the underlying infinite
dimensional optimization problem is well-approximated even with a low basis function complexity.
Herein, we compare the approach from [5] to a different finite dimensional approximation. We analyze
both with respect to convergence of the optimal costs as the number of basis functions is increased. In
particular, the optimal cost of the approximation given in [5] decreases monotonically and approaches the
cost of underlying infinite dimensional problem from above. It is shown that the corresponding optimal
trajectories are guaranteed to converge and that the second approximation approaches the optimal cost
of the infinite dimensional problem from below. In addition, we will establish conditions guaranteeing
convergence of both approximations to the cost of the underlying infinite dimensional problem.
This report focuses on the technical proofs and complements [4], where the underlying ideas are discussed in detail and a numerical example is provided.
∗
Michael Muehlebach and Raffaello D’Andrea are with the Institute for Dynamic Systems and Control, ETH Zurich. The
contact author is Michael Muehlebach, [email protected]. This work was supported by ETH-Grant ETH-48 15-1.
1
2
Problem Formulation
We present and analyze two approximations of the following optimal control problem,
1
1
J∞ := inf ||x||22 + ||u||22
2
2
s.t. ẋ(t) = Ax(t) + Bu(t), x(0) = x0 ,
x(t) ∈ X , u(t) ∈ U, ∀t ∈ [0, ∞),
x ∈ L2n , u ∈ L2m , ẋ ∈ L2n ,
(1)
where X and U are closed and convex subsets of Rn and Rm , respectively, containing 0; the space of square
integrable functions mapping from [0, ∞) to Rq is denoted by L2q , where q is a positive integer; and the
L2q -norm is defined as
Z
∞
L2q → R,
xT x dt,
x → ||x||22 :=
(2)
0
where dt denotes the Lebesgue measure. We assume that J∞ is finite and that the corresponding minimizers,
x and u, are unique.
Input and state trajectories will be approximated as a linear combination of basis functions τi ∈ L21 ,
i = 1, 2, . . . , that is
x̃s (t, ηx ) := (In ⊗ τ s (t))T ηx ,
ũs (t, ηu ) := (Im ⊗ τ s (t))T ηu ,
(3)
where ⊗ denotes the Kronecker product, ηx ∈ Rns and ηu ∈ Rms are the parameter vectors, and τ s (t) :=
(τ1 (t), τ2 (t), . . . , τs (t)) ∈ Rs . In order to simplify notation we will omit the superscript s in τ s , x̃s , and ũs ,
and simply write τ , x̃, and ũ whenever the number of basis functions is clear from the context. Similarly,
the dependence of x̃ and ũ on ηx and ηu is frequently omitted. Without loss of generality we assume that
the basis functions are orthonormal. Note that orthonormal basis functions can be constructed with the
Gram-Schmidt procedure, [3, p. 50].
As motivated in [5], the following additional assumptions on the basis functions are made:
A1) They are linearly independent.
A2) They fulfill τ̇ (t) = M τ (t) for all t ∈ [0, ∞), for some matrix M ∈ Rs×s . The eigenvalues of M have
strictly negative realparts.
2.1
Resulting optimization problems
In [5], the following finite dimensional approximation of the original problem (1) is introduced,
1
1
Js := inf ||x̃||22 + ||ũ||22
2
Z2 ∞
s.t.
(In ⊗ τ ) Ax̃ + B ũ − x̃˙ dt = 0,
(4)
0
x̃(0) − x0 = 0,
ηx ∈ X s , ηu ∈ U s .
Note that the subscript s refers to the number of basis functions used. More precisely, the optimization
problem with optimal cost Js corresponds to the case where input and state trajectories are spanned by the
first s basis functions. The trajectories x̃ and ũ, which satisfy the equality constraint, fulfill the equations
˙ for all t ∈ [0, ∞) and x̃(0) = x0 , see [5]. This is needed
of motion exactly, that is, Ax̃(t) + B ũ(t) = x̃(t)
to guarantee that the minimizers of (4) achieve the cost Js on the nominal system. We make the following
assumptions on the sets X s and U s1
1
The assumptions are listed for the state constraints X and are analogous for the input constraints U.
2
B0) X s is closed and convex
B1) is (X s ) ⊂ X s+1
B2) ηx ∈ X s implies (In ⊗ τ (t))T ηx ∈ X for all t ∈ [0, ∞),
where the inclusion is , mapping from Rns to Rn(s+1) is defined by
x̃s (t, ηx ) = x̃s+1 (t, is (ηx )), ∀t ∈ [0, ∞), ∀ηx ∈ Rns .
(5)
Assumption B0) implies that the optimization problem (4) is convex, and that corresponding minimizers
exist, provided the existence of feasible trajectories. Assumption B1) is used to show that the cost Js is
monotonically decreasing in s, whereas Assumption B2) implies that Js is bounded below by J∞ , see Sec. 3.
In the context of MPC, Assumption B2) guarantees recursive feasibility and closed-loop stability, [5]. In
addition, the cost Js is achieved on the nominal system, as the resulting input and state trajectories respect
input and state constraints and fulfill the dynamics exactly.
The following alternative approximation is introduced,
1
1
J˜s := inf ||x̃||22 + ||ũ||22
2
2
Z
∞
(In ⊗ τ ) Ax̃ + B ũ − x̃˙ dt − (In ⊗ τ (0)) (x̃(0) − x0 ) = 0,
s.t.
(6)
0
ηx ∈ X̃ s , ηu ∈ Ũ s ,
whose purpose is to provide a monotonically increasing sequence J˜s bounding J∞ from below. To that
extent, the following assumptions on the sets X̃ s and Ũ s are made:2
C0) X̃ s is closed and convex.
C1) πs (X̃ s+1 ) ⊂ X̃ s .
C2) For each x ∈ L2n with x(t) ∈ X for all t ∈ [0, ∞) it holds that π s (x) ∈ X̃ s ,
where the projections πs and π s are defined as
π
s
:L2n
ns
Z
→ R ,x →
∞
(In ⊗ τ s )x dt,
(7)
0
πs :Rn(s+1) → Rns , ηx → π s x̃s+1 (·, ηx ) .
(8)
Assumption C0) ensures that the optimization problem (6) is convex, and that corresponding minimizers
exist, provided the existence of feasible trajectories. Assumption C1) is used to demonstrate that J˜s is
monotonically increasing, whereas Assumption C2) implies that J˜s is bounded above by J∞ , see Sec. 3.
Examples fulfilling Assumptions B0)-B2) and C0)-C2) are provided in [4].
3
Main results
In the following we will analyze the two approximations (4) and (6) and prove the following result:
Theorem 3.1 Let N0 be such that JN0 is finite.
1) If Assumptions B0), B1), and B2) hold, then the sequence Js is monotonically decreasing for s ≥ N0 ,
converges as s → ∞, and is bounded below by J∞ . The corresponding optimizers x̃ and ũ converge
strongly in L2n , respectively L2m as s → ∞.
2) If Assumptions C0), C1), and C2) are fulfilled, then J˜s is monotonically increasing for s ≥ 1, converges
as s → ∞, and is bounded above by J∞ .
2
The assumptions are again listed for the state constraints X and are analogous for the input constraints U.
3
We start by summarizing the results from [5], stating that the optimal cost of (4) is a monotonically
decreasing sequence providing an upper bound on the optimal cost of (1).
Proposition 3.2 Let N0 be such that JN0 is finite. If Assumptions B0) and B1) are fulfilled, then the sequence Js is monotonically decreasing for s ≥ N0 and converges as s → ∞. If Assumptions B0) and B2)
are fulfilled, Js is bounded below by J∞ .
Proof See [5].
The fact that Js converges can be used to demonstrate convergence of the optimizer x̃s and ũs , as well
as the corresponding parameters ηxs and ηus . Therefore, the parameter vectors ηxs and ηus are interpreted
as square summable sequences, i.e. as elements in `2 .3
Proposition 3.3 Let N0 be such that JN0 is finite and let Assumptions B0) and B1) be fulfilled. Then,
the minimizers of (4) converge (strongly) in `2 as s → ∞, and the corresponding trajectories x̃s and ũs
converge (strongly) in L2n , respectively L2m .
Proof We fix s ≥ N0 and denote the minimizers corresponding to Js by ηxs , ηus , and the minimizers
corresponding to Js+1 by ηxs+1 , ηus+1 , which we consider to be elements of `2 . The following observation
can be made: The vectors ηx := λis (ηxs ) + (1 − λ)ηxs+1 and ηu := λis (ηus ) + (1 − λ)ηus+1 with λ ∈ [0, 1]
are feasibly candidates for the optimization problem (4) over s + 1 basis functions.4 This is because the
constraints is (ηx ) ∈ X s+1 , is (ηu ) ∈ U s+1 are satisfied by Assumptions B0) (convexity) and B1). Moreover,
as the dynamics are fulfilled exactly it follows for x̃ := (In ⊗ τ )T ηx and ũ := (Im ⊗ τ )T ηu that
x̃(t) = λx̃s (t) + (1 − λ)x̃s+1 (t)
Z t
At
eA(t−t̂) B(λũs (t̂) + (1 − λ)ũs+1 (t̂))dt̂
= e x0 +
Z0 t
eA(t−t̂) B ũ(t̂)dt̂,
= eAt x0 +
0
where x̃s := (In ⊗ τ )T ηxs , ũs := (In ⊗ τ )T ηus , x̃s+1 := (In ⊗ τ )T ηxs+1 , and ũs+1 := (In ⊗ τ )T ηus+1 ,
which concludes that the equality constraint is likewise fulfilled. Since ηxs+1 and ηus+1 are the minimizers
corresponding to Js+1 , we have that
1
1
Js+1 ≤ ||x̃||22 + ||ũ||22 ,
2
2
(9)
for all λ ∈ [0, 1]. The objective function is quadratic, and hence strongly convex with respect to the L2n and
L2m -norm, which leads to
1
1
Js+1 ≤ λJs + (1 − λ)Js+1 − λ(1 − λ)||x̃s − x̃s+1 ||22 − λ(1 − λ)||ũs − ũs+1 ||22 ,
2
2
(10)
for all λ ∈ [0, 1]. We set λ = 1/2 and obtain
||x̃s − x̃s+1 ||22 + ||ũs − ũs+1 ||22 ≤ 4(Js − Js+1 ).
(11)
According to Prop. 3.2, the sequence Js converges as s → ∞. As a result, it follows that (x̃s , ũs ) ∈ L2n ×L2m
is a Cauchy sequence. The space L21 is a Banach space, [7, p. 67] and so is L2n × L2m . Consequently,
(x̃s , ũs ) converges strongly as s → ∞, [8, p. 4]. The orthonormality of the basis functions implies by
Bessel’s inequality, [3, p. 51] that the parameters ηxs and ηus form a Cauchy sequence in `2 × `2 . The
square summable sequences form likewise a Banach space and therefore the parameter vectors ηxs and ηus
converge strongly in `2 as s → ∞.
The set of square summable sequences is denoted by `2 .
Notation is slightly abused, since is is used to denote both the inclusion Rns → Rn(s+1) and Rms → Rm(s+1) , which is
defined in analogy to (5).
3
4
4
We establish results for the optimization problem (6), which are similar to the ones given by Prop. 3.2.
More precisely, we will show that under Assumptions C0)-C2) the optimal cost of (6) bounds J∞ from
below and is monotonically increasing in s.
Proposition 3.4 Let Assumptions C0) and C2) be fulfilled. Then J˜s ≤ J∞ holds for all s ≥ 1.
Proof We will denote the minimizers of (1) by x and u, which are both square integrable and fulfill ẋ(t) =
Ax(t) + Bu(t) for all t ∈ [0, ∞). We define ηx := π s (x) ∈ Rns , ηu := π s (u) ∈ Rms (where notation is
slightly abused to denote both the projection from L2n → Rns and the projection from L2m → Rms , defined
in analogy to (7), by π s ). From Assumption C2) it follows that ηx ∈ X̃ s and ηu ∈ Ũ s . We will argue that ηx
and ηu fulfill the equality constraints in (6). Therefore we rewrite the equality constraint as
A ⊗ Is − In ⊗ M T − In ⊗ (τ (0)τ (0)T ) ηx + (B ⊗ Is )ηu + (In ⊗ τ (0))x0 = 0,
(12)
where orthonormality of the basis functions, the properties of the Kronecker product, and Assumption A2)
is used. We note further that the identity
Z ∞
Z ∞
T
T
τ τ̇ T dt
τ̇ τ dt = −τ (0)τ (0) −
M=
(13)
0
0
T
T
= −τ (0)τ (0) − M ,
which follows from integration by parts, simplifies the previous equation to
(A ⊗ Is + In ⊗ M )ηx + (B ⊗ Is )ηu + (In ⊗ τ (0))x0 = 0.
Moreover, it holds that
Z ∞
Z
(In ⊗ τ )ẋdt = −(In ⊗ τ (0))x0 − (In ⊗ M )
0
(14)
∞
(In ⊗ τ )xdt
(15)
0
= −(In ⊗ τ (0))x0 − (In ⊗ M )ηx
= (A ⊗ Is )ηx + (B ⊗ Is )ηu ,
(16)
(17)
where integration by parts (1st step), the definition of the projection π s (2nd step), and the fact that x and u
fulfill the (linear) equations of motion exactly (3rd step) has been used. Clearly, (16) and (17) are equivalent
to (14) and therefore ηx and ηu are feasible candidates for (6). Bessel’s inequality, [3, p. 51], implies that
1
1
1
1
1
1
|ηx |22 + |ηu |22 = |π s (x)|22 + |π s (u)|22 ≤ ||x||22 + ||u||22 ,
2
2
2
2
2
2
(18)
where the Euclidean norm is denoted by | · |2 . Therefore ηx and ηu are feasible candidates achieving a cost
that is smaller than J∞ , and hence, J˜s ≤ J∞ for all s ≥ s0 .
In order to establish that the sequence J˜s is monotonically increasing, we will work with the dual
problem. It turns out that the finite dimensional representation of the adjoint equations are fulfilled exactly
by (6). We will use this fact to construct feasible candidates for the optimization over s + 1 basis functions.
Proposition 3.5 Let Assumptions C0), C1), and C2) be fulfilled. Then J˜s is monotonically increasing and
bounded above by J∞ for all s ≥ 1.
Proof We first derive the dual of (6). We use Lagrange duality to rewrite (6) as
Z ∞
1 T
1
˜
Js = inf sup
x̃ x̃ + ũT ũ + p̃T Ax̃ + B ũ − x̃˙ dt − p̃(0)T (x̃(0) − x0 ),
ηx ,ηu ηp
2
2
0
s
s
ηx ∈ X̃ , ηu ∈ Ũ .
5
(19)
From Assumptions C0) and C1) we can infer that J˜s ≤ J∞ for all s ≥ 1 by Prop. 3.4. The fact that
0 ≤ J˜s ≤ J∞ implies further that the infimum in (6) is attained, and that the set of minimizers is nonempty
due to Assumption C0) (X̃ s and Ũ s are closed). According to [6, p. 503, Thm. 11.39] strong duality holds,
and the infimum and supremum can be interchanged, which yields
Z ∞
1
1 T
˜
x̃ x̃ + ũT ũ + p̃T Ax̃ + B ũ − x̃˙ dt − p̃(0)T (x̃(0) − x0 ),
Js = sup inf
2
2
ηp ηx ,ηu 0
(20)
s
s
ηx ∈ X̃ , ηu ∈ Ũ .
The minimization over ηx and ηu is a convex problem and can be rewritten in terms of convex-conjugate
˙ resulting in
functions, [6, p. 473]. To that extent, we apply first integration by parts on the term p̃T x̃,
Z ∞
1
1 T
˜
Js = sup inf
x̃ x̃ + x̃T AT p̃ + p̃˙ + ũT ũ + ũT B T p̃ dt + p̃(0)T x0 ,
2
2
ηp ηx ,ηu 0
(21)
s
s
ηx ∈ X̃ , ηu ∈ Ũ .
By defining ṽ(t, ηv ) := (In ⊗ τ (t))T ηv such that
Z ∞
˙ dt = 0,
δ λ̃T (ṽ + AT p̃ + p̃)
δ λ̃ := (In ⊗ τ )T δηλ ,
0
for all δηλ ∈ Rns , which is equivalent to −ṽ := AT p̃ + p̃˙ as shown in [5], the minimization over x̃ can be
interpreted as a (extended real-valued) function of ṽ, i.e.
Z ∞
Z ∞
1 T
1
T
inf
x̃ x̃ − x̃ ṽ dt = − sup
x̃T ṽ − x̃T x̃ dt
(22)
2
2
ηx ∈X̃ s 0
ηx ∈X̃ s 0
Z ∞
1
= − sup
x̃T ṽ − x̃T x̃ dt =: −Iϕ∗s (ṽ).
(23)
2
π s (x̃)∈X̃ s 0
Note that Iϕ∗s maps from L2n to the extended real line and is well-defined. In a similar way, we can regard
the minimization over ηu as (extended real-valued) function of p̃,
Z ∞
1 T
inf
ũ ũ + ũT B T p̃ dt =: −Iψ∗ s (−B T p̃),
(24)
s
s
2
π (ũ)∈Ũ
0
where in this case π s denotes the projection L2m → Rsm defined in analogy to (7) (with a slight abuse of
notation). Thus, (21) is reformulated as
J˜s = sup −Iϕ∗s (ṽ) − Iψ∗ s (−B T p̃) + p̃(0)T x0 ,
ηp ∈Rns
Z ∞
s.t.
(In ⊗ τ ) p̃˙ + AT p̃ + ṽ dt = 0.
(25)
0
The functions ṽ and p̃ satisfy the adjoint equations exactly and it holds that limt→∞ p̃(t) = 0 by Assumption
A1).
Let ηv ∈ Rns and ηp ∈ Rns , with corresponding trajectories ṽ s (t, ηv ) and p̃s (t, ηp ), be maximizers of
(25). The set of maximizers is non-empty due to the fact that we optimize over Rns and 0 ≤ J˜s ≤ J∞
holds. The equality constraint implies that the adjoint equation p̃˙s (t) + AT p̃s (t) + ṽ s (t) = 0 is fulfilled for
all times t ∈ [0, ∞), see [5], and thus, the adjoint equation is likewise fulfilled by the augmented trajectories
ṽ s+1 (t, is (ηv )) and p̃s+1 (t, is (ηp )). Hence, ṽ s+1 (t, is (ηv )) and p̃s+1 (t, is (ηp )) are feasible candidates to the
optimization (25) over s + 1 basis functions, and it holds that p̃s+1 (0, is (ηp )) = p̃s (0, ηp ). It remains to
6
establish the relation between Iϕ∗s and Iϕ∗s+1 , as well as Iψ∗ s+1 and Iψ∗ s , which is done via the order reversing
property of the convex-conjugation. Therefore the function Iϕ∗s is regarded as the conjugate of
(
1
||x̃s (t, π s (x))||22 π s (x) ∈ X̃ s ,
Iϕs (x) := 2
(26)
∞
otherwise.
We note that Assumption C1) implies Iϕs+1 (x) ≥ Iϕs (x) for all x ∈ L2n . This is due to the fact that
any square integrable function x with π s+1 (x) ∈ X̃ s+1 automatically fulfills π s (x) ∈ X̃ s , since πs (X̃ s+1 )
is contained in X̃ s by Assumption C1), and |π s+1 (x)|22 ≥ |π s (x)|22 holds for all x ∈ L2n . The convexconjugation reverses ordering, which implies
Iϕ∗s+1 (v) ≤ Iϕ∗s (v)
(27)
for all v ∈ L2n , see [1, Prop. 4.4.1, p. 171]. The same reasoning applies to Iψ∗ s , which is the convex-conjugate
of
(
1
||ũs (t, π s (u))||22 π s (u) ∈ Ũ s ,
(28)
Iψs (u) := 2
∞
otherwise.
This leads to the conclusion that
−Iϕ∗s (v) − Iψ∗ s (−B T p) ≤ −Iϕ∗s+1 (v) − Iψ∗ s+1 (−B T p),
(29)
for any v, p ∈ L2n . Hence, we have that ṽ s+1 (t, is (ηv )) and p̃s+1 (t, is (ηp )) are feasible candidates to the
optimization problem over s + 1 basis functions with higher corresponding cost and therefore J˜s+1 ≥ J˜s .
Next, we would like to establish that lims→∞ J˜s = lims→∞ Js . In order to do so, we need the following
assumptions:
D0) lim sups→∞ X̃s ⊂ lim inf s→∞ X s
D1) The basis functions τi , i = 1, 2, . . . , are dense in C0∞ (in the topology of uniform convergence).5
Proposition 3.6 Let N0 be such that JN0 is finite and let Assumptions B0)-D1) be fulfilled. Then, lims→∞ J˜s =
lims→∞ Js holds.
Proof By Prop. 3.2 and Prop. 3.3 it follows that Js is monotonically decreasing, lims→∞ Js is finite, and that
the corresponding optimizers converge. From Prop. 3.4 and Prop. 3.5, we can infer that J˜s is monotonically
increasing and bounded above by J∞ for all s ≥ 1. This implies further that the sequence of minimizers
of (6) is bounded in the L2 -sense. Due to the fact that L21 (and likewise L2n × L2m ) is a Hilbert space, there
exists a subsequence s(q) such that the corresponding minimizer of (6) converge weakly, i.e. x̃s(q) * x̃,
ũs(q) * ũ, [2, p. 163].
We pick anyP
δp := (δp1 , . . . , δpn )T , with δpi ∈ C0∞ , i = 1, 2, . . . , n, and δp(0) = 0, and choose a
sequence δ p̃k = ki=1 τi δηpi , δηpi ∈ Rn , converging uniformly to δp. According to Assumption D1) such a
sequence exists. Hence for any > 0 we can find an integer N0 large enough, such that
|δ p̃k (0)T (x̃s(q) (0) − x0 )| ≤ |x̃s(q) (0) − x0 |2
holds for all k ≥ N0 . We claim that |x̃s(q) (0) − x0 |2 is uniformly bounded. This can be seen by right
multiplying the equality constraint of (6) by ηxT , resulting in
Z ∞
x̃s(q)T (Ax̃s(q) + B ũs(q) − x̃˙ s(q) )dt − x̃s(q) (0)T (x̃s(q) (0) − x0 ) = 0,
(30)
0
5
The set of smooth functions with compact support mapping from [0, ∞) to R is denoted by C0∞ .
7
which can be further simplified to
Z ∞
1
1
2
x̃s(q)T (Ax̃s(q) + B ũs(q) )dt = |x̃s(q) (0) − x0 |22 ,
|x0 |2 +
2
2
0
(31)
using limt→∞ x̃s(q) (t) = 0 (by Assumption A1)) and completing the squares. From the fact that J˜s(q) ≤ J∞
for all q, it follows that x̃s(q) and ũs(q) are bounded in L2n , respectively L2m . As a consequence, |x̃s(q) (0)−x0 |22
is uniformly bounded, as can be verified with the Cauchy-Schwarz inequality, limk→∞ δ p̃k (0)T (x̃s(q) (0) −
x0 ) converges uniformly, and the limits over q and k can be interchanged,
lim lim δ p̃k (0)T (x̃s(q) (0) − x0 ) = lim lim δ p̃k (0)T (x̃s(q) (0) − x0 ) = 0.
q→∞ k→∞
k→∞ q→∞
The equality constraint of (6) reads therefore as
Z ∞
Z
T
s(q)
s(q)
s(q)
˙
δ p̃k (Ax̃ + B ũ − x̃ )dt = lim lim
lim lim
k→∞ q→∞
q→∞ k→∞
0
(32)
∞
δ p̃Tk (Ax̃s(q) + B ũs(q) − x̃˙ s(q) )dt = 0,
0
(33)
where both limits agree. We will show that the limit over
PN k commutes with the integration. To that extent,
we make the following claim: For any function ṽ := i=1 τi ηvi , where ηvi are bounded vectors in Rn and
N is a positive integer, it holds that
Z ∞
Z ∞
T
lim
δ p̃k ṽdt =
δpT ṽdt.
(34)
k→∞
0
0
We will prove the claim below, but assume for now that it holds. As a consequence of Assumption A1),
implying that x̃˙ s(q) is a linear combination of the basis functions, the claim results in
Z ∞
Z ∞
s(q)
˙
δpT x̃˙ s(q) dt,
(35)
δ p̃k x̃ dt =
lim
k→∞
0
0
for any integer s(q). Using integration by parts (twice) and the fact that x̃s(q) converges weakly leads to
Z ∞
Z ∞
T ˙ s(q)
δ ṗT x̃s(q) dt
(36)
lim
δp x̃ dt = lim −
q→∞ 0
q→∞
0
Z ∞
δ ṗT x̃dt
(37)
=−
0
Z ∞
˙
=
δpT x̃dt.
(38)
0
Note that δ ṗ has compact support, is bounded (by continuity), and is therefore square integrable in [0, ∞).
The claim implies further that (33) simplifies to
Z ∞
0 = lim
δpT (Ax̃s(q) + B ũs(q) − x̃˙ s(q) )dt
(39)
q→∞ 0
Z ∞
˙
=
δpT (Ax̃ + B ũ − x̃)dt.
(40)
0
The same argument can be repeated for any δp = (δp1 , . . . δpn ), δpi ∈ C0∞ , i = 1, 2, . . . , n, vanishing at 0,
and therefore, as s(q) → ∞, the equality constraint of (6) reads as
Z ∞
˙
0=
δpT (Ax̃ + B ũ − x̃)dt,
∀δp ∈ C0∞ .
0
8
˙
Due to the fundamental lemma of the calculus of variations, [9, p. 18], this is equivalent to x̃(t)
= Ax̃(t) +
B ũ(t) for all t ∈ [0, ∞) (almost everywhere). A similar argument based on variations that do not vanish at
time 0 ensures limq→∞ x̃s(q) (0) = x0 . As a result, the equality constraint of (6) is equivalent to the one of (4)
in the limit as s(q) → ∞. Combined with Assumption D1), it implies that x̃ and ũ are feasible candidates
for (4) and therefore limq→∞ J˜s(q) ≥ lims→∞ Js . From Prop. 3.2, Prop. 3.4, and Prop. 3.5 it follows that J˜s
is monotonically increasing and bounded by J∞ ≤ Js for all s ≥ N0 , resulting in
lim J˜s(q) = lim J˜s = lim Js .
q→∞
s→∞
(41)
s→∞
PN
n
It remains to prove the claim. Let ṽ :=
i=1 τi ηvi where the ηvi s are bounded vectors in R and N is
fixed. The matrix M in Assumption A1) is negative definite and therefore it holds that |ṽ(t)|2 ≤ C1 e−βt
for all t ≥ T0 for some constants C1 > 0, β > 0 and time T0 > 0. As a result, it follows from Hölder’s
inequality, [7, p. 76], that
Z ∞
Z ∞
T
(δ p̃k − δp) ṽdt ≤ sup |δ p̃k (t) − δp(t)|2
|ṽ|2 dt.
(42)
0
t∈[0,∞)
0
The second term can be bounded by invoking Hölder’s inequality once more,
Z ∞
Z ∞
Z T0
C1 −βT0
e
.
|ṽ|2 dt +
C1 e−βt dt ≤ T0 ||ṽ||2 +
|ṽ|2 dt ≤
β
T0
0
0
(43)
Hence, the right-hand side of (42) converges to zero due to the uniform convergence of the δ p̃k to δp as
k → ∞. This proves the claim.
4
Conclusion
We introduced two different approximations to a class of infinite-horizon optimal control problems encountered in MPC. The approximations bound the optimal cost of the underlying problem from above and below,
and their optimal costs converge as the number of basis functions tends to infinity. Under favorable circumstances, the resulting input trajectories of the first approximation are found to approximate the optimal input
of the underlying infinite dimensional problem arbitrarily accurately, and the corresponding optimal costs
converge to the optimal cost of the underlying infinite dimensional problem. The second approximation
yields a lower bound on the cost of the underlying optimal control problem, and can therefore be used to
quantify the approximation quality of both approximations.
Acknowledgment
The first author would like to thank Jonas Lührmann for a fruitful discussion regarding the proof of
Prop. 3.6.
References
[1] J. M. Borwein and J. D. Vanderwerff. Convex Functions. Cambridge University Press, 2010.
[2] J. B. Conway. A Course in Functional Analysis. Springer, second edition, 1990.
[3] R. Courant and D. Hilbert. Methods of Mathematical Physics, volume 1. Interscience Publishers, 1953.
9
[4] M. Muehlebach and R. D’Andrea. Approximation of continuous-time infinite-horizon optimal control
problems arising in model predictive control. Proceedings of the IEEE Conference on Decision and
Control, 2016.
[5] M. Muehlebach and R. D’Andrea. Parametrized infinite-horizon model predictive control for linear
time-invariant systems with input and state constraints. Proceedings of the American Control Conference, pages 2669–2674, 2016.
[6] R. T. Rockafellar and R. J.-B. Wets. Variational analysis. Springer, 2009.
[7] W. Rudin. Real and Complex Analysis. McGraw-Hill, third edition, 1987.
[8] W. Rudin. Functional Analysis. McGraw-Hill, second edition, 1991.
[9] L. C. Young. Lectures on the Calculus of Variations and Optimal Control Theory. AMS Chelsea
Publishing, second edition, 1980.
10
| 3cs.SY
|
On LDPC Code Based
Massive Random-Access Scheme
for the Gaussian Multiple Access Channel
Luiza Medova∗† , Anton Glebov‡, Pavel Rybin∗ , and Alexey Frolov‡∗
∗ Inst.
for Information Transmission Problems, Moscow, Russia, [email protected]
Inst. of Physics and Mathematics, Moscow, Russia, [email protected]
‡ Skolkovo Institute of Science and Technology, Moscow, Russia, {[email protected], [email protected]}
arXiv:1803.08377v1 [cs.IT] 22 Mar 2018
† Moscow
Abstract—This paper deals with the problem of massive
random access for Gaussian multiple access channel (MAC). We
continue to investigate the coding scheme for Gaussian MAC
proposed by A. Vem et al in 2017. The proposed scheme consists
of four parts: (i) the data transmission is partitioned into time
slots; (ii) the data, transmitted in each slot, is split into two
parts, the first one set an interleaver of the low-density paritycheck (LDPC) type code and is encoded by spreading sequence
or codewords that are designed to be decoded by compressed
sensing type decoding; (iii) the another part of transmitted
data is encoded by LDPC type code and decoded using a joint
message passing decoding algorithm designed for the T-user
binary input Gaussian MAC; (iv) users repeat their codeword
in multiple slots. In this paper we are concentrated on the third
part of considered scheme. We generalized the PEXIT charts to
optimize the protograph of LDPC code for Gaussian MAC. The
simulation results, obtained at the end of the paper, were analyzed
and compared with obtained theoretical bounds and thresholds.
Obtained simulation results shows that proposed LDPC code
constructions have better performance under joint decoding
algorithm over Gaussian MAC than LDPC codes considered by
A. Vem et al in 2017, that leads to the better performance of
overall transmission system.
I. I NTRODUCTION
Current wireless networks are designed with the goal of
servicing human users. Next generation of wireless networks
is facing a new challenge in the form of machine-type communication: billions of new devices (dozens per person) with
dramatically different traffic patterns are expected to go live in
the next decade. The main challenges are associated with: (a)
huge number of autonomous devices connected to one access
point, (b) low energy consumption, (c) short data packets. This
problem has attracted attention (3GPP and 5G-PPP) under the
name of mMTC (massive machine-type communication).
There are K ≫ 1 users, of which only T have data to
send in each time instant. A base station (BS) sends periodic
beacons, announcing frame boundaries, so that the uplink
(user-to-BS) communication proceeds in a frame-synchronized
fashion. Length of each frame is N , where a typical interesting
value is n ≈ 104 − 105 . Each active user has k bits that it
intends to transmit during a frame, where a typical value is
k ≈ 100 bit. The main goal is to minimize the energy-perbit spent by each of the users. We are interested in grant-free
access (5G terminology). That is, active users transmit their
data, without any prior communication with the BS (without
resource requests). We will focus on the Gaussian multipleaccess channel (GMAC) with equal-power users, i.e.
Y =
T
X
Xt + Z,
t=1
where Z ∼ N (0, N0 /2) and E |Xi |2 ≤ P .
This paper deals with construction of low-complexity random coding schemes for GMAC (indeed we restrict our
consideration to the case of binary input GMAC). Let us
emphasize the main difference from the classical setting.
Classical information theory provided the exact solutions for
the case of all-active users, i.e. T = K. Almost all well-known
low-complexity coding solutions for the traditional MAC
channel (e.g. [1]) implicitly assume some form of coordination
between the users. Due to the gigantic number users we
assume them to be symmetric, i.e. the users use the same codes
and equal powers. Here we continue the line of work started
in [2]–[4]. In [2] the bounds on the performance of finitelength codes for GMAC are presented. In [3] Ordentlich and
Polyanskiy describe the first low-complexity coding paradigm
for GMAC. The improvement (it terms of required Eb /N0 )
was given in [4].
We continue to investigate the coding scheme from [4]. The
proposed scheme consists of four parts:
• the data transmission is partitioned into time slots;
• the data, transmitted in each slot, is split into two parts,
the first one (preamble) allows to detect users that were
active in the slot. It also set an interleaver of the lowdensity parity-check (LDPC) type code [5], [6] and is
encoded by spreading sequence or codewords that are
designed to be decoded by compressed sensing type
decoding;
• the second part of transmitted data is encoded by LDPC
type code and decoded using a joint message passing
decoding algorithm designed for the T -user binary input
GMAC;
• users repeat their codeword in multiple slots and use
successive interference cancellation.
The overall scheme can be called T-fold irregular repetition
slotted ALOHA (IRSA, [7], [8]) scheme for GMAC. The main
difference of this scheme in comparison to IRSA is as follows:
any collisions of order up to T can be resolved with some
probability of error introduced by Gaussian noise.
In this paper we are concentrated on the third part of considered scheme. Our contribution is as follows. We generalized
the protograph extrinsic information transfer charts (EXIT)
to optimize the protograph of LDPC code for GMAC. The
simulation results, obtained at the end of the paper, were
analyzed and compared with obtained theoretical bounds and
thresholds. Obtained simulation results shows that proposed
LDPC code constructions have better performance under joint
decoding algorithm over Gaussian MAC than LDPC codes
considered in [4], that leads to the better performance of
overall system.
II. I TERATIVE JOINT
DECODING ALGORITHM
By C (t) , t ∈ [T ], we denote the codes used by users (the
codes are binary). Recall, that n and R are accordingly the
length and the rate of C (t) , k ∈ [T ].
Thus T users send codewords c(1) , c(2) , . . . , c(T ) . After
(1)
(2)
(K)
BPSK modulator
√ we√have the sequences x , x , . . . , x ,
x(i) ∈ {− P , + P }n . The channel output (y) is the
element-wise sum of the sequences affected by Gaussian noise.
The aim of joint multi-user decoder is to recover the all the
codewords based on received vector y. The decoder employs a
low-complexity iterative belief propagation (BP) decoder that
deals with a received soft information presented in LLR (log
likelihood ratio) form. The decoding system can be represented
as a graph (factor graph), which is shown in Fig. 1. User
LDPC codes are presented with use of Tanner graphs with
variable and check nodes. At the same time there is a third
kind of nodes in the figure – functional nodes (marked with
green color). These nodes correspond to the elements of the
received sequence y.
Following the by now standard methodology of factor
graphs, see [9], [10], we can write down the corresponding
message passing decoding algorithm.
1) initialize the LLR values of variable nodes for each user
code with zero values assuming equal probability for 1
and −1 values;
2) perform IO outer iterations, where each iteration consists
of the following steps:
a) perform maximum likelihood decoding of functional nodes (i.e. calculate update messages for
variable nodes);
b) perform II inner iterations of BP decoder for users’
LDPC codes and update LLR values of variable
nodes (this is done in parallel);
The BP part is standard, i.e. each user utilizes standard
BP decoding algorithm (Sum-Product or Min-Sum) to decode
an LDPC code. The most interesting part is the decoding of
functional nodes. Following the principles of message-passing
algorithms, the update rule to compute the message (µ) sent to
i-th variable node of k-th user (k = 1, . . . , K, i = 1, . . . , N )
from a functional node Fi is the following:
Let mkvc,j denotes the message sent by the variable node v
to the check node c along its jth edge of user k:
mvc,j =
dX
v −1
mkcv,i + mksv ;
i=1, i6=j
where mkcv,i is the message outgoing from the check node :
dY
c −1
m
vc,i
;
mcv,j = 2 tanh−1
tanh
2
i=1, i6=j
mksv
and
is the message outgoing from the state node.
It is also necessary to describe the rule for computing
messages outgoing from state nodes. Let xki denote the ith
transmitted code bit and yi denote the channel output. The
outgoing message from the ith variable node of user k to the
connected state node is computed as
k
p(xki = 1)
p(xki = 1)
, emvs,i =
;
k
p(xi = −1)
p(xki = −1)
mkvs,i = log
Considering standard function node message-passing rules [9],
we compute the message sent to user k ith variable node from
the state node:
mksv,i = log
P Q
p(xki = 1|y)
=
p(xki = −1|y)
(1)
(k)
p(xji = 1)p(yi |xi , ..., xi
(n)
= 1, ..., xi )
∼x(k) j6=k
i
log P Q
(1)
(k)
(n)
j
p(xi = −1)p(yi |xi , ..., xi = −1, ..., xi )
(k) j6=k
∼xi
We can simplify it in the following way:
mksv,i =
P Q
e
1xj Xj
(1)
(k)
p(yi |xi , ..., xi
(n)
= 1, ..., xi )
∼x(k) j6=k
i
log P Q 1 X
, (1)
(1)
(k)
(n)
x
j
e j p(yi |xi , ..., xi = −1, ..., xi )
(k) j6=k
∼xi
(
(j)
1, xi = 1
.
(j)
0, xi = −1.
The number of computations necessary to obtain the outgoing messages from the node Fi grows exponentially with the
number of users, nevertheless, this number of users usually
remains small, and we will therefore not be concerned with
this fact.
where
1xk =
III. PEXIT C HARTS
Extrinsic Information Transfer (EXIT) charts [11] can be
used for the accurate analysis of the behavior of LDPC
decoders. But since the usual EXIT analysis cannot be applied
to the study of protograph-based [12] LDPC codes we will use
a modified EXIT analysis for protograph-based LDPC codes
(PEXIT) [13]. This method is similar to the standard EXIT
analysis in that it tracks the mutual information between the
message edge and the bit value corresponding to the variable
node on which the edge is incident, while taking into account
LDPC code of 1st user
LDPC code of 2nd user
check nodes
LDPC code of 3rd user
. . .
. . .
variable nodes
. . .
. . .
. . .
. . .
. . .
functional nodes
Fig. 1. Joint decoder graph representation for T = 3
the structure of the protograph. In our work we use the notation
from [13] to describe EXIT charts for protograph-based LDPC
codes.
Let IEv denote the extrinsic mutual information between a
message at the output of a variable node and the codeword bit
associated to the variable node:
IEv = IEv (IAv , IEs ) ,
where IAv is the mutual information between the codeword
bits and the check-to-variable messages and IEs is the mutual information between the codeword bits and the stateto-variable messages. Since the PEXIT tracks the mutual
information on the edges of the protograph, we define IEv (i, j)
as the mutual information between the message sent by the
variable node Vj to the check node Ci and the associated
codeword bit:
sX
[J −1 (IAv (s, j))]2 + [J −1 (IEs (j))]2
IEv (i, j) = J
s6=i
where J(σ) is given by [11]:
J(σ) = 1
Z∞
1
1
√
exp −
−
2
2
2πσ
2
y − σ2 x
σ
−∞
!2
log2 (1 + e−y )dy.
Similarly, we define IEc , the extrinsic mutual information
between a message at the output of a check node and the
codeword bit associated to the variable node receiving the
message:
IEc = IEc (IAc ) ,
where IAc is the mutual information between one input
message and the associated codeword bit and IAc = IEv .
Accordingly, the mutual information between the message sent
by Ci to Vj and the associated codeword bit is described as:
sX
[J −1 (1 − Iac (i, s))]2 .
Iec (i, j) = 1 − J
s6=j
The mutual information between the variable node Vj and
the message passed to the state node is denoted as IEvs (j)
and is given by:
s
X
IEvs (j) = J
[J −1 (Iav (s, j))]2 .
s
Next we need to compute the mutual information IEs . In
order to get an idea about the probability density function of
(1) for user j, we generate samples of the outgoing LLRs
through (1) based on the samples of the received LLRs from
other users whose PDF is approximated with N (µEvs , 2µEvs ),
−1
where µEvs = J (I2 Evs ) . To numerically estimate µEs and
obtain the required mutual information as IEs = J(µEs ),
we refer to [14], where the following three approaches are
proposed:
• Mean-matched Gaussian approximation : the mean µ is
2
estimated from samples and we set µEvs = µ and σEvs
=
2µ.
• Mode-matched Gaussian approximation : given a sufficiently large number of N samples generated through
(1), the mode m is estimated from samples and we set
2
µEvs = m and σEvs
= 2m.
• Gaussian mixture approximation: mean values µ1 , ..., µk
and the weights a1 , ..., ak are estimated from samples and
IEs = a1 J(µ1 ) + ... + ak J(µk ).
The rationale for using these approximations was shown in
[14]. Furthermore, the authors compared the performance of
these approaches. The mode-matched method was found to
give the maximum output mutual information and the joint
codes designed by using this approximation also yield the
lowest decoding bit error probability compared to the other
two approaches.
Each user calculate IAP P (j), the mutual information between the posteriori probability likelihood ratio evaluated by
the variable node Vj and the associated codeword bit.
s
X
[J −1 (IAv (s, j))]2 + [J −1 (IEs (j))]2 .
IAP P (j) = J
s
The convergence is declared if each IAP P (j) reaches 1 as the
iteration number tends to infinity.
IV. N UMERICAL R ESULTS
In this section the simulation results, obtained for the cases
T=2 and T=4, are represented. Let us at first consider the
simulation results for T=2 (Fig. 2). For this case we compare
the Frame Error Rate (FER) performance of rate-1/4 LDPC
code (364, 91) from [4] obtained by repetition of each code
bit of regular (3,6) LDPC code twice, rate-1/4 LDPC code
(364, 91) optimized by PEXIT charts method described above
and Polyanskiy’s finite block length (FBL) bound for 2 user
less than 3 dB.
10 0
V. S PARSE
FER
10 -1
10 -2
10 -3
Repeated reg. (3,6) LDPC code
PEXIT optimized LDPC code
10 -4
0.5
1
1.5
2
2.5
3
3.5
4
4.5
5
5.5
E b /N0 , dB
Fig. 2. Simulation results for T=2 and LDPC code (364, 91)
case. As we can see in Fig. 2 proposed PEXIT-optimized
LDPC code construction outperforms LDPC code construction
from [4] by about 0.5 dB. In the same time the gap between
Polyanskiy’s FBL bound and PEXIT-optimized LDPC code
is about 3 dB. But we would like to point out that used
here Polyanskiy’s FBL bound is for Gaussian signal and not
for Binary Phase-Shift Keying (BPSK) modulation, used for
simulation. So, we believe that this gap will be reduced is FBL
bound for BPSK modulation is used.
Now let us consider simulation results for T=4 (Fig. 3). For
this case we obtain another PEXIT-optimized rate-1/4 LDPC
code (364, 91) and compare FER performance of same LDPC
code from [4] and Polyanskiy’s FBL bound for 4 users. As
10
0
FER
10 -1
10
LDPC
CODES
In this section we answer a very natural question: how to
increase the order of collision, that can be decoded in a slot.
E.g. consider the case from the previous section. Let the slot
length n′ = 364. We want to increase T up to 8. Here we face
with two problems:
• The performance of LDPC joint decoder rapidly becomes
bad with grows of T . We were not able to find (364, 91)
LDPC codes, that work well for T = 8.
• The number of computations necessary to obtain the
outgoing messages from the functional node grows exponentially with the number of users T .
We address both these problems in a scheme, which is
proposed below (see Fig. 4). The idea is to use sparse
spreading signatures [15] for LDPC codes, such that the degree
of functional node is reduced from T to dc . The slot length is
now n′ , n′ 6= n.
Users
C1
C2
CT
…
2
…
1
…
n
n
dc
dc
…
…
1
2
K
…
n
dc
…
…
n’
Resources
Tn
n′ =
dc
Fig. 4. Sparse spreading of LDPC codes
-2
10 -3
Repeated reg. (3,6) LDPC code
PEXIT optimized LDPC code
10 -4
SPREADING OF
2
3
4
5
6
7
8
9
10
E b /N0 , dB
Fig. 3. Simulation results for T=4 and LDPC code (364, 91)
we can see in Fig. 3 proposed PEXIT-optimized LDPC code
construction outperforms LDPC code construction from [4]
by more than 3 dB. And again the gap between Polyanskiy’s
FBL bound and PEXIT-optimized LDPC code is a little bit
In Fig. 5 we present the simulation results. As we were
not able to find (364, 91) LDPC codes, that work well for
T = 8 we consider 2 times shorter LDPC codes and compare
2 strategies:
• split the slot into 2 parts and send 4 users in each part;
• use sparse spreading;
We see, that our approach is much better and works practically the same in comparison to the case of 2 times longer
LDPC codes and 2 times smaller number of users (see the
previous section).
VI. C ONCLUSION
We generalized the protograph extrinsic information transfer
charts (EXIT) to optimize the protograph of LDPC code for
GMAC. The simulation results, obtained at the end of the
paper, were analyzed and compared with obtained theoretical
bounds and thresholds. Obtained simulation results shows that
10 0
FER
10 -1
10 -2
10 -3
T=8, 2 parts split slot, LDPC code (180, 45)
T=8, slot spreading, LDPC code (180, 45)
T=4, LDPC code (364, 91)
10 -4
3
3.5
4
4.5
5
5.5
6
6.5
7
7.5
8
E b /N0 , dB
Fig. 5. Simulation results for spreading
proposed LDPC code constructions have better performance
under joint decoding algorithm over Gaussian MAC than
LDPC codes considered by A. Vem et al in 2017, that leads
to the better performance of overall system.
ACKNOWLEDGMENT
We want to thank Y. Polyanskiy for fruitful discussions.
R EFERENCES
[1] B. Rimoldi and R. Urbanke, “A rate-splitting approach to the gaussian
multiple-access channel,” IEEE Transactions on Information Theory,
vol. 42, no. 2, pp. 364–375, 1996.
[2] Y. Polyanskiy, “A perspective on massive random-access,” in Information Theory (ISIT), 2017 IEEE International Symposium on. IEEE,
2017, pp. 2523–2527.
[3] O. Ordentlich and Y. Polyanskiy, “Low complexity schemes for the random access gaussian channel,” in 2017 IEEE International Symposium
on Information Theory (ISIT), June 2017, pp. 2528–2532.
[4] A. Vem, K. R. Narayanan, J. Cheng, and J.-F. Chamberland, “A userindependent serial interference cancellation based coding scheme for the
unsourced random access gaussian channel,” in Proc. IEEE Information
Theory Workshop (ITW), 2017, pp. 1–5.
[5] R. G. Gallager, Low-Density Parity-Check Codes, 1963.
[6] R. Tanner, “A recursive approach to low complexity codes,” IEEE
Transactions on information theory, vol. 27, no. 5, pp. 533–547, 1981.
[7] G. Liva, “Graph-based analysis and optimization of contention resolution
diversity slotted aloha,” IEEE Transactions on Communications, vol. 59,
no. 2, pp. 477–487, February 2011.
[8] K. R. Narayanan and H. D. Pfister, “Iterative collision resolution for
slotted aloha: An optimal uncoordinated transmission policy,” in 2012
7th International Symposium on Turbo Codes and Iterative Information
Processing (ISTC), Aug 2012, pp. 136–139.
[9] T. Richardson and R. Urbanke, Modern coding theory. Cambridge
university press, 2008.
[10] F. R. Kschischang, B. J. Frey, and H.-A. Loeliger, “Factor graphs and
the sum-product algorithm,” IEEE Transactions on information theory,
vol. 47, no. 2, pp. 498–519, 2001.
[11] S. ten Brink, “Convergence behavior of iteratively decoded parallel
concatenated codes,” IEEE Transactions on Communications, vol. 49,
no. 10, pp. 1727–1737, Oct 2001.
[12] J. Thorpe, “Low-Density Parity-Check (LDPC) Codes Constructed from
Protographs,” Interplanetary Network Progress Report, vol. 154, pp. 1–
7, Aug. 2003.
[13] G. Liva and M. Chiani, “Protograph ldpc codes design based on exit
analysis,” in IEEE GLOBECOM 2007 - IEEE Global Telecommunications Conference, Nov 2007, pp. 3250–3254.
[14] I. Shahid and P. Yahampath, “Distributed joint source-channel code
design for gmac using irregular ldpc codes,” EURASIP Journal on
Wireless Communications and Networking, vol. 2014, no. 1, p. 3, Jan
2014. [Online]. Available: https://doi.org/10.1186/1687-1499-2014-3
[15] R. Hoshyar, F. P. Wathan, and R. Tafazolli, “Novel low-density signature
for synchronous cdma systems over awgn channel,” IEEE Transactions
on Signal Processing, vol. 56, no. 4, pp. 1616–1626, April 2008.
| 7cs.IT
|
Proceedings of Machine Learning Research vol 75:1–22, 2018
Minimax Distribution Estimation in Wasserstein Distance
Shashank Singh
SSS 1@ CS . CMU . EDU
Barnabás Póczos
BAPOCZOS @ CS . CMU . EDU
arXiv:1802.08855v1 [math.ST] 24 Feb 2018
Machine Learning Department, Carnegie Mellon University, Pittsburgh, PA 15213, USA
Abstract
The Wasserstein metric is an important measure of distance between probability distributions, with
several applications in machine learning, statistics, probability theory, and data analysis. In this
paper, we upper and lower bound minimax rates for the problem of estimating a probability distribution under Wasserstein loss, in terms of metric properties, such as covering and packing numbers,
of the underlying sample space.
Keywords: Wasserstein Distance; Density Estimation; Minimax Theory; Covering Number; Packing Number
1. Introduction
The Wasserstein metric is an important measure of distance between probability distributions, based
on the cost of transforming either distribution into the other through mass transport, under a metric
on the sample space. Originating in the optimal transport literature,1 the Wasserstein metric has,
owing to its intuitive and general nature, been utilized in such diverse areas as probability theory
and statistics, economics, image processing, text mining, robust optimization, and physics (Villani,
2008; Fournier and Guillin, 2015; Esfahani and Kuhn, 2015; Gao and Kleywegt, 2016).
In the analysis of image data, the Wasserstein metric has been used for various tasks such as
texture classification and face recognition (Sandler and Lindenbaum, 2011), reflectance interpolation, color transfer, and geometry processing (Solomon et al., 2015), image retrieval (Rubner et al.,
2000), and image segmentation (Ni et al., 2009), and, in the analysis of text data, for tasks such as
document classification (Kusner et al., 2015) and machine translation (Zhang et al., 2016).
In contrast to a number of other popular notions of distance2 between probability distributions,
such as Lp distances or Kullback-Leibler and other f -divergences (Morimoto, 1963; Csiszár, 1964;
Ali and Silvey, 1966), which require distributions to be absolutely continuous with respect to each
other or to a base measure, Wasserstein distance is well-defined between any pair of probability
distributions over a sample space equipped with a metric.3 As a particularly important consequence,
Wasserstein distances between discrete (e.g., empirical) distributions and continuous distributions
are well-defined and informative (i.e., often finite).
Partly for this reason, Wasserstein distances have been widely used in probability theory to
prove central limit and related approximation theorems (Rüschendorf, 1985; Johnson et al., 2005;
1. The Wasserstein metric has been variously attributed to Monge, Kantorovich, Rubinstein, Gini, Mallows, and others;
see Chapter 3 of Villani (2008) for detailed history.
2. We use the term distance to refer both to proper metrics as well as to similar notions, such as information divergences,
that fail to satisfy symmetry or the triangle inequality.
3. For this reason, we use “distribution estimation” in this paper, rather than the more popular “density estimation”.
c 2018 S. Singh & B. Póczos.
D ISTRIBUTION E STIMATION
IN
WASSERSTEIN D ISTANCE
Chatterjee et al., 2008; Rio et al., 2009, 2011; Chen et al., 2010; Reitzner et al., 2013). Within machine learning and statistics, this same property has inspired a class of so-called minimum Wasserstein distance estimates (del Barrio et al., 1999, 2003; Bassetti et al., 2006; Bernton et al., 2017)
of distributions, ranging from exponential distributions (Baı́llo et al., 2016) to more exotic models
such as restricted Boltzmann machines (RBMs) (Montavon et al., 2016) and generative adversarial
networks (GANs) (Arjovsky et al., 2017). k-means clustering (and, similarly, k-medians clustering)
also fall in this class of estimators, where the hypothesis class is taken to be discrete distributions
supported on at most k points; more flexible multi-level clustering algorithms can also be expressed
in this way, using a more elaborate hypothesis class (Ho et al., 2017). PCA can also be expressed
and generalized to manifolds using Wasserstein distance minimization (Boissard et al., 2015). In
principle, these estimators are equivalent to empirical risk minimization, taking advantage of the
fact that Wasserstein distances between the empirical distribution and distributions in the relevant
hypothesis class are well-defined and finite. Moreover, these estimates often perform well in practice because they are free of both tuning parameters and strong distributional assumptions.
In this paper, we study, in a minimax sense, the problem of estimating a distribution using
Wasserstein distance as a loss function. We make minimal assumptions, and some examples of the
varied metric spaces to which our results apply are given in Section 7. Specifically, we assume that
the sample space is a totally bounded metric space, but make no assumptions at all on the distribution. We then prove bounds on the minimax convergence rates of distribution estimation, utilizing
covering numbers of the sample space for upper bounds and packing numbers for lower bounds.
It may at first be surprising that positive results can be obtained under such mild assumptions; this
highlights that the Wasserstein metric is quite a weak metric (see our Lemma 8 and the subsequent
remark for some detailed discussion of this). Moreover, our results imply that, without further assumptions on the population distribution, the empirical distribution is often minimax rate-optimal.
O RGANIZATION
The remainder of this paper is organized as follows. Section 2 provides notation required to formally state both the problem of interest and our results, while Section 3 reviews previous work
studying convergence of distributions in Wasserstein distance. Section 4 provides several lemmas
on which the proofs of our main results rely; all lemmas are proven in Appendix A. Sections 5 and
6 respectively contain our main upper and lower bound results, along with their proofs. Finally, in
Section 7, we apply our upper and lower bounds to identify minimax convergence rates in a number of concrete examples. Section 8 concludes with a summary of our contributions and suggested
avenues for future work.
2. Notation and Problem Setting
For any integer n ∈ N, [n] = {1, 2, ..., n} denotes the set of the first n positive integers. For
sequences {an }n∈N and {bn }n∈N of non-negative real numbers, an . bn and, equivalently bn & an ,
are used to indicate the existence of a constant C > 0 such that lim supn→∞ abnn ≤ C. an ≍ bn
indicates an . bn and an & bn simultaneously.
2
D ISTRIBUTION E STIMATION
IN
WASSERSTEIN D ISTANCE
2.1. Problem Setting
For the remainder of this paper, fix a metric space (Ω, ρ), over which Σ denotes the Borel σ-algebra,
and let P denote the family of all Borel probability distributions on Ω. The main object of study in
this paper is the Wasserstein distance on P, defined as follows:
Definition 1 (r-Wasserstein Distance) Given two Borel probability distributions P and Q over Ω
and r ∈ [1, ∞), the r-Wasserstein distance Wr (P, Q) ∈ [0, ∞] between P and Q is defined by
!1/r
Wr (P, Q) :=
inf
E
µ∈Π(P,Q) (X,Y )∼µ
[ρr (X, Y )]
,
where Π(P, Q) denotes all couplings between X ∼ P and Y ∼ Q; that is,
Π(P, Q) := µ : Σ2 → [0, 1] for all A ∈ Σ, µ(A × Ω) = P (A) and µ(Ω × A) = Q(A) ,
is the set of joint probability measures over Ω × Ω with marginals P and Q.
Wr is a proper metric over P; namely, it is symmetric in its arguments and satisfies the triangle inequality. Intuitively, Wr (P, Q) quantifies the r-weighted total cost of transforming mass distributed
according to P to be distributed according to Q, where the cost of moving a unit mass from x ∈ Ω
to y ∈ Ω is ρ(x, y). Note that Wr (P, Q) is sometimes defined in terms of equivalent (e.g., dual)
formulations (especially for r = 1); however, they will not be necessary in this paper.
We now give a formal statement of the problem studied in this paper:
Formal Problem Statement: Suppose (Ω, ρ) is a metric space. Suppose P is an unknown
IID
Borel probability distribution on Ω, from which we observe n IID samples X1 , ..., Xn ∼ P . We
are interested in studying the minimax rates at which P can be estimated from X1 , ..., Xn , in terms
of the (r th power of the) r-Wasserstein loss. Specifically, we are interested in upper and lower
bounding the quantity
i
h
Wrr P, Pb ,
inf sup
E
Pb P ∈P X ,...,XnIID
∼ P
1
where the infimum is taken over all estimators Pb of P (i.e., all functions Pb : Ωn → P of the data).
2.2. Definitions for Stating our Results
Here, we give notation and definitions needed to state our theoretical results in Sections 5 and 6 and
intermediate lemmas in Section 4.
Ω
Let 2Ω denote the power set of Ω. Let S ⊆ 22 denote the family of all Borel partitions of Ω:
(
)
[
S := S ⊆ Σ : Ω ⊆
S and ∀S, T ∈ S, S ∩ T = ∅ .
S∈S
We first define some metric notions that will later be useful for bounding Wasserstein distances:
Definition 2 (Diameter and Separation of a Set, Resolution of a Partition) For any set S ⊆ Ω,
the diameter Diam(S) of S is defined by Diam(S) := supx,y∈S ρ(x, y), and the separation Sep(S)
of S is defined by Sep(S) := inf x6=y∈S ρ(x, y). If S ∈ S is a partition of Ω, then the resolution
Res(S) of S is defined by Res(S) := supS∈S Diam(S).
3
D ISTRIBUTION E STIMATION
IN
WASSERSTEIN D ISTANCE
Definition 3 (Refinement of a Partition, Nested Partitions) Suppose S, T ∈ S are partitions of
Ω. T is said to be a refinement of S if, for every T ∈ T , there exists S ∈ S with T ⊆ S. A sequence
{Sk }k∈[K] of partitions is called nested if, for each k ∈ [K]\{K}, Sk+1 is a refinement of Sk ,
We now define the covering and packing number of a metric space, which are classic and widely
used measures of the size or complexity of a metric space (Dudley, 1967; Haussler, 1995; Zhou,
2002; Zhang, 2002). Our main convergence results will be stated in terms of these quantities, as
well as the packing radius, which acts, approximately, as the inverse of the packing number.
Definition 4 (Covering Number, Packing Number, and Packing Radius of a Metric Space) The
covering number N : (0, ∞) → N of (Ω, ρ) is defined for all ε by
N (ε) := min {|S| : S ∈ S
and
Res(S) ≤ ε} .
The packing number M : (0, ∞) → N of (Ω, ρ) is defined for all ε by
M (ε) := max {|S| : S ⊆ Ω
and
Sep(S) ≥ ε} .
Finally, the packing radius R : N → [0, ∞] is defined for all n ∈ N by
R(n) := sup{Sep(S) : S ⊆ Ω
and
|S| ≥ n}.
Sometimes, we use the covering or packing number of a metric space, say (Θ, τ ), other than (Ω, ρ);
in such cases, we write N (Θ; τ ; ε) or M (Θ; τ ; ε) rather than N (ε) or M (ε), respectively. For
specific ε > 0, we will also refer to N (Θ; τ ; ε) as the ε-covering number of (Θ, τ ).
Remark 5 The covering and packing numbers of a metric space are closely related. In particular,
for any ε > 0, we always have
M (ε) ≤ N (ε) ≤ M (ε/2).
(1)
The packing number and packing radius also have a close approximate inverse relationship. In
particular, for any ε > 0, we always have
R(M (ε)) ≥ ε.
(2)
However, we may have R(M (ε)) > ε, and it need not, in general, be the case that, for all n ∈ N,
M (R(n)) ≥ n or, for all n ∈ N, M (R(n)) ≤ n.
Remark 6 We defined the covering number slightly differently from usual (using partitions rather
than covers). However, the given definition is equivalentSto the usual definition, since (a) any partition is itself a cover (i.e., a set C ⊆ 2Ω such that Ω ⊆ C∈C C), and (b), for any countable cover
C := {C1 , C2 , ...} ⊆ 2Ω ,Sthere exists a partition S ∈ S with |S| ≤ |C| and each Si ⊆ Ci , defined
recursively by Si := Ci \ i−1
j=1 Si . S is often called the disjointification of C.
4
D ISTRIBUTION E STIMATION
IN
WASSERSTEIN D ISTANCE
3. Related Work
A long line of work (Dudley, 1969; Ajtai et al., 1984; Dereich et al., 2013; Boissard et al., 2014;
Fournier and Guillin, 2015; Weed and Bach, 2017) has studied distribution estimation in Wasserstein distance, with all the above papers focusing on analyzing convergence of the empirical distribution to the population distribution in Wasserstein distance. In terms of upper bounds, the most
general and tight upper bounds are those of Weed and Bach (2017), who provide results in general
metric spaces, in terms of the covering numbers of those spaces. Their main results are expressed
in terms of a particular notion of dimension, which they call the Wasserstein dimension s because it
allows them to derive convergence rates of order n−1/s (matching the n−1/D rate achieved on the
unit cube [0, 1]D ). Our general upper bounds as a function of covering number are somewhat easier
to apply than those of Weed and Bach (2017), and we use these in Section 7 to derive convergence
rates even in high-dimensional cases (where D → ∞ as n → ∞) or infinite dimensional cases.
To the best of our knowledge, there has been no work on deriving minimax lower bounds for
distribution estimation under Wasserstein loss. As noted above, most previous works have focused
on studying convergence rate of the empirical distribution to the true distribution in Wasserstein distance. For this rate, several lower bounds have been established, matching known upper bounds in
many quite general cases. However, many distribution estimators besides the empirical distribution
can be considered. For example, it is tempting (especially given the infinite dimensionality of the
distribution to be estimated) to try to reduce variance by techniques such as smoothing or importance sampling (Bucklew, 2013). However, our lower bound results, given in Section 6, imply that
the empirical distribution is already minimax optimal, up to constant factors, in many cases.
4. Preliminary Lemmas
We begin by providing a few basic lemmas; these lemmas are not fundamentally novel, but they
will be used in the subsequent proofs of our main upper and lower bounds, and also help provide
intuition for the behavior of the Wasserstein metric and its connections to other metrics between
probability distributions. Our first lemma relates Wasserstein distance to the notion of resolution of
a partition.
Lemma 7 Suppose S ∈ S is a countable Borel partition of Ω. Let P and Q be Borel probability
measures such that, for every S ∈ S, P (S) = Q(S). Then, for any r ≥ 1, Wr (P, Q) ≤ Res(S).
Our next lemma gives simple lower and upper bounds on the Wasserstein distance between distributions supported on a countable subset X ⊆ Ω, in terms of Diam(X ) and Sep(X ). Since our
main results will utilize coverings and packings to approximate Ω by finite sets, this lemma will
provide a first step towards approximating (in Wasserstein distance) distributions on Ω by distributions on these finite sets. Indeed, the lower bound in Inequality (3) will suffice to prove our lower
bounds, although a tighter upper bound, based on the upper bound in (3), will be necessary to obtain
tight upper bounds.
Lemma 8 Suppose (Ω, ρ) is a metric space, and suppose P and Q are Borel probability distributions on Ω with countable support; i.e., there exists a countable set X ⊆ Ω with P (X ) = Q(X ) = 1.
Then, for any r ≥ 1,
X
X
|P ({x}) − Q({x})| . (3)
|P ({x}) − Q({x})| ≤ Wrr (P, Q) ≤ (Diam(X ))r
(Sep(X ))r
x∈X
x∈X
5
D ISTRIBUTION E STIMATION
Remark 9 Recall that the term
P
IN
WASSERSTEIN D ISTANCE
|P ({x}) − Q({x})| in Inequality (3) is the L1 distance
X
kp − qk1 :=
|p(x) − q(x)|
x∈X
x∈X
between the densities p and q of P and Q with respect to the counting measure on X , and that this
same quantity is twice the total variation distance
T V (P, Q) := sup |P (A) − Q(A)| .
A⊆Ω
Hence, Lemma 8 can be equivalently written as
Sep(Ω) (kp − qk1 )1/r ≤ Wr (P, Q) ≤ Diam(Ω) (kp − qk1 )1/r
and as
Sep(Ω) (2T V (P, Q))1/r ≤ Wr (P, Q) ≤ Diam(Ω) (2T V (P, Q))1/r ,
bounding the r-Wasserstein distance in terms of the L1 and total variation distance. As noted in
Example 1, equality holds in (3) precisely when ρ is the unit discrete metric given by ρ(x, y) =
1{x6=y} for all x, y ∈ Ω.
On metric spaces that are discrete (i.e., when Sep(Ω) > 0), the Wasserstein metric is (topologically) at least as strong as the total variation metric (and the L1 metric, when it is well-defined),
in that convergence in Wasserstein metric implies convergence in total variation (and L1 , respectively). On the other hand, on bounded metric spaces, the converse is true. In either of these cases,
rates of convergence may differ between metrics, although, in metric spaces that are both discrete
and bounded (e.g., any finite space), we have Wr ≍ T V 1/r .
While Lemma 8 gave a simple upper bound on the Wasserstein distance, the factor of Diam(Ω)
turns out to be too large to obtain tight rates for a number of cases of interest (such as the Ddimensional unit cube Ω = [0, 1]D , discussed in Example 2). The following lemma gives a tighter
upper bound, based on a hierarchy of nested partitions of Ω; this allows us to obtain tighter bounds
(than Diam(Ω)) on the distance that mass must be transported between P and Q. Note that, when
K = 1, Lemma 10 reduces to a trivial combination of Lemmas 7 and 8; indeed, these lemmas are
the starting point for proving Lemma 10 by induction on K.
Note that the idea of such a “multi-resolution” upper bound has been utilized extensively before,
and numerous versions have been proven before (see, e.g., Fact 6 of Do Ba et al. (2011), Lemma 6
of Fournier and Guillin (2015), or Proposition 1 of Weed and Bach (2017)). However, most of these
versions have been specific to Euclidean space; to the best of our knowledge, only Proposition 1 of
Weed and Bach (2017) applies to general metric spaces.
Lemma 10 Let K be a positive integer. Suppose {Sk }k∈[K] is a nested sequence of countable Borel
partitions of (Ω, ρ), with S0 = Ω. Then, for any r ≥ 1 and Borel probability measures P and Q on
Ω,
K
X
X
Wrr (P, Q) ≤ (Res(SK ))r +
|P (S) − Q(S)| .
(Res(Sk−1 ))r
(4)
S∈Sk
k=1
6
D ISTRIBUTION E STIMATION
WASSERSTEIN D ISTANCE
IN
Lemma 10 requires a sequence of partitions of Ω that is not only multi-resolution but also
nested. While the ε-covering number implies the existence of small partitions with small resolution,
these partitions need not be nested as ε becomes small. For this reason, we now give a technical
lemma that, given any sequence of partitions, constructs a nested sequence of partitions of the same
cardinality, with only a small increase in resolution.
Lemma 11 Suppose S and T are partitions of (Ω, ρ), and suppose S is countable. Then, there
exists a partition S ′ of (Ω, ρ) such that:
a) |S ′ | ≤ |S|.
b) Res(S ′ ) ≤ Res(S) + 2 Res(T ).
c) T is a refinement of S ′
5. Upper Bounds
In this section, we utilize the covering number bounds from the previous section to develop error
bounds for density estimation using the empirical distribution under Wasserstein loss.
5.1. Expectation Bounds
Here, we use Lemmas 10 and 11 to prove bounds on the expected Wasserstein distance between the
empirical and true distributions. We begin with a simple technical lemma bounding the expected L1
deviation of a multinomial random variable from its mean.
Lemma 12 Suppose (X1 , ..., XK ) ∼ Multinomial(n, p1 , ..., pK ). Let
Z := kX − pk1 =
Then, E [Z/n] ≤
p
K
X
k=1
|Xk − pk | .
21K/n.
Theorem 13 Let (Ω, ρ) be a metric space on which P is a Borel probability measure. Let Pb denote
IID
the empirical distribution of n IID samples X1 , ..., Xn ∼ P , give by
n
1X
Pb(S) :=
1{Xi ∈S} ,
n
i=1
∀S ∈ Σ.
Then, for any sequence {εk }k∈[K] ∈ (0, ∞)K with ε0 = Diam(Ω),
r
r
K
K
i
X
X
p
21
r
r
2j−k εj
N (εk ).
E Wr (P, Pb) ≤ εK +
n
h
k=1
7
j=k−1
D ISTRIBUTION E STIMATION
IN
WASSERSTEIN D ISTANCE
Proof By recursively applying Lemma 11, there exists a sequence {Sk }k∈[K] of partitions of (Ω, ρ)
satisfying the following conditions:
1. for each k ∈ [K], |Sk | = N (εk ).
2. for each k ∈ [K], Res(Sk ) ≤
K
X
2j−k εj .
j=k
3. {Sk }k∈[K] is nested.
Note that, for any k ∈ [K], the vector nPb(S) (indexed by S ∈ Sk ) follows an n-multinomial
distribution over |Sk | categories, with means given by P (S); i.e.,
(nPb(S1 ), ..., nPb(Sk )) ∼ Multinomial(n, P (S1 ), ..., P (Sk )).
Thus, by Lemma 12, for each k ∈ [K],
r
r
X
21|S
|
21N (εk )
k
P (S) − Pb(S) ≤
=
.
E
n
n
S∈Sk
Thus, by Lemma 10,
r
r
E [Wr (P, Q)] ≤ E εK +
≤ εrK +
K
X
k=1
K
X
k=1
K
X
j=k
r
2j−k εj
X
S∈Sk
|P (S) − Q(S)|
r
K
X
X
2j−k εj E
|P (S) − Q(S)|
S∈Sk
j=k
r
r
K
K
X
X
p
21
N (εk )
2j−k εj
≤ εrK +
n
k=1
j=k
6. Lower Bounds
In this section, we prove minimax lower bounds (over the family P of all densities on the metric
space (Ω, ρ)) for density estimation in Wasserstein distance (that is, the quantity
i
h
(5)
Wrr P, Pb ,
inf sup
E
Pb P ∈P X1 ,...,Xn IID
∼ P
where the infimum is over all estimators Pb of P (i.e., all functions Pb : Ωn → P)). Our bounds are
in terms of the packing radius R of (Ω, ρ).
8
D ISTRIBUTION E STIMATION
IN
WASSERSTEIN D ISTANCE
Theorem 14 Suppose P ∈ P is an unknown Borel probability measure on (Ω, ρ), and suppose we
IID
observe n IID samples X1 , ..., Xn ∼ P . Then, there exists a universal constant CHJW > 0 such
that
r
h
i
k−1
r
r
Wr (P, Pb) ≥ CHJW sup R (k)
inf sup
.
E
n
Pb P ∈P X ,...,X IID
k∈[n]
∼ P
1
n
Proof Our proof is based on a reduction to a minimax lower bound for estimating the mean parameter vector of a multinomial distribution in L1 norm. Specifically, we will use Corollary 4 of
Han et al. (2015), which implies the existence of a universal constant CHJW > 0 (independent of n)
such that, for any k ≤ n,
r
k−1
,
p(X1 , ..., Xn ) − pk1 ] ≥ CHJW
inf sup
E [kb
n
pb p∈∆k X1 ,...,Xn
where
∆k :=
p ∈ [0, 1]k :
k
X
j=1
pj = 1
denotes the standard (k − 1)-dimensional probability simplex, X1 , ..., Xn ∈ [k] are distributed IID
according to a multinomial distribution with parameter vector p ∈ ∆k , and the infimum is taken
over all estimators pb of p (i.e., all functions pb : [k]n → ∆k ).
For some k ≤ n, let PD denote the class of discrete distributions over an R(k)-packing D of
(Ω, ρ) with |D| = k. Then, Corollary 4 of Han et al. (2015) implies that
r
i
h
|D| − 1
kPb (X1 , ..., Xn ) − P k1 ≥ CHJW
,
inf sup
E
n
Pb p∈PD X1 ,...,Xn
where the infimum is taken over all estimators Pb of P (i.e., all functions Pb : Ωn → P).
Recalling Lemma 8 (which implies (Sep(D))r kPb − P k1 ≤ Wrr (P, Pb)), we have
i
h
h
i
inf sup E Wrr (Pb, P ) ≥ (Sep(D))r inf sup E kPb − P k1
Pb P ∈P
Pb P ∈P
i
h
≥ (Sep(D))r inf sup E kPb − P k1
Pb P ∈PD
r
|D| − 1
r
≥ CHJW (Sep(D))
n
r
k−1
≥ CHJW Rr (k)
.
n
The theorem follows by taking the supremum over k ≤ n on both sides.
9
(6)
D ISTRIBUTION E STIMATION
IN
WASSERSTEIN D ISTANCE
7. Example Applications
We conclude by exploring applications of our results for several metric spaces (Ω, ρ). In each of the
following examples, P is an unknown Borel probability measure over the specified Ω, from which
we observe n IID samples. For upper bounds, Pb denotes the empirical distribution of these samples.
Example 1 (Finite Space) Consider the case where Ω is a finite set, over which ρ is the discrete
metric given, for some δ > 0, by
ρ(x, y) = δ1{x=y} ,
x, y ∈ Ω.
for all
Then, for any ε ∈ (0, δ), the covering number is N (ε) = |Ω|. Thus, setting K = 1 and sending
ε1 → 0 in Theorem 13 gives
r
h
i
21|Ω|
r
r
b
.
E Wr (P, P ) ≤ δ
n
On the other hand, for any k ≤ |Ω|, R(k) = δ, while, for k > |Ω|, R(k) = 0. Thus, setting
k = |Ω| in Theorem 14 gives
r
h
i
|Ω| − 1
r
r
b
.
Wr (P, P ) ≥ CHJW δ
inf sup
E
n
Pb P ∈P X ,...,X IID
∼ P
1
n
Example 2 (Unit Cube, Euclidean Metric) Consider the case where Ω = [0, 1]D is the unit cube
and ρ is the Euclidean metric. Suppose P is a Borel probability measure on (Ω, ρ), and let Pb
denote the empirical distribution of n IID samples from P . Setting εk := 2−2k and recalling that
N (ε) ≤ (2D)D ε−D , one can show that Theorem 13 gives
r
K
h
i
21(2D)D X k(D−2r)
r
−2Kr
r+1
b
2
W
(P,
P
)
≤
2
+
2
E
r
n
k=1
If D = 2r, then this gives
E
and setting K ≍
1
4r
h
i
Wrr (P, Pb )
−2Kr
≤2
r+1
+2
r
21(2D)D
K,
n
h
i
log n gives the rate E Wrr (P, Pb) . n−1/2 log n. Otherwise, if D 6= 2r, then
h
i
r
21(2D)D 2K(D−2r) − 1
n
2D−2r − 1
i
h
For 2r > D, sending K → ∞ gives a rate of E Wrr (P, Pb) . n−1/2 . For 2r < D, setting
i
h
r
1
K ≍ 2D
log n gives a rate of E Wrr (P, Pb) . n− D . To summarize
E
Wrr (P, Pb)
−2Kr
≤2
D+1
+2
−1/2
i n
r
n−1/2 log n
E Wr (P, Pb) .
−r/D
n
h
10
if 2r > D
if 2r = D
if 2r < D
D ISTRIBUTION E STIMATION
IN
WASSERSTEIN D ISTANCE
(reproducing results of Fournier and Guillin (2015)). On the
√ other hand, it is easy to check that
the packing radius R satisfies R(n) ≥ n−1/D and R(2) ≥ D. Thus, applying Theorem 14 with
k = n and k = 2, we have
o
h
i
n
inf sup E Wrr (Pb, P ) ≥ CHJW max (n + 1)−r/D , D r/2 n−1/2 .
Pb P ∈P
Combining these upper and lower bounds gives that the minimax rate for density estimation over P
in the loss Wrr is
h
i n−1/2 if 2r > D
r b
inf sup E Wr (P , P ) ≍
n−r/D if 2r < D
Pb P ∈P
(except for the case 2r = D, when our upper and lower bounds are separated by a factor of log n).
Moreover, this rate is achieved by using the empirical distribution to estimate P .
We now consider a high dimensional problem, where the size of Ω can grow with n.
Example 3 (Binary Hypercube, Hamming Metric) Suppose Ω = {0, 1}D is the P
D-dimensional
binary hypercube, and suppose ρ is the Hamming metric given by ρδ (x, y) = δ D
j=1 1{xi 6=yi } .
Then, the covering number can easily be upper bounded for ε ∈ [D] by N (ε) ≤ 2D−ε . Thus,
setting K = 1 and sending ε1 → 0 in Theorem 13 gives
r
h
i
21
r
r D
.
E Wr (P, Pb ) ≤ D 2
n
The famous Varshamov-Gilbert bound (see, e.g., Lemma
25.12 of Rigollet (2015)) can be interpreted as the statement that, for any γ ∈ (0, 1/2), R ⌊eγ D ⌋ ≥ 21 − γ Dδ. Thus, setting
p
2
k = ⌊eγ D ⌋ (so that k ≤ n as long as γ ≤
(1/D) log n) in Theorem 14 and applying the
Varshamov-Gilbert bound gives
s
r
h
i
1
⌊eγ 2 D ⌋ − 1
−
γ
Dδ
.
Wrr (P, Pb) ≥ CHJW
sup
inf sup
E
q
2
n
Pb P ∈P X ,...,X IID
log n
1
∼ P
1
n
γ∈ 0,min
2
,
D
To summarize, in the high-dimensional case where D, n → ∞, consistent estimation is possible if
D log D ∈ o(log n) and only if D ∈ o(log n).
Finally, we consider distributions over an infinite dimensional space of smooth functions.
Example 4 (Hölder Ball, L∞ Metric) Suppose that, for some α ∈ (0, 1],
Ω := f [0, 1]D → [−1, 1]
∀x, y ∈ [0, 1]D , |f (x) − f (y)| ≤ kx − ykα2
is the class of unit α-Hölder functions on the unit cube and ρ is the L∞ -metric given by
ρ(f, g) =
sup |f (x) − g(x)|,
x∈[0,1]D
11
for all f, g ∈ Ω.
D ISTRIBUTION E STIMATION
IN
WASSERSTEIN D ISTANCE
The covering and packing numbers of (Ω, ρ) are well-known to be of order exp ε−D/α (DeVore and Lorentz,
1993); specifically, there exist positive constants 0 < c1 < c2 such that, for all ε ∈ (0, 1),
c1 exp ε−D/α ≤ N (ε) ≤ M (ε) ≤ c2 exp ε−D/α .
It also follows from Inequality (2) that, for all k ∈ N,
R(k) ≥
1
log(k/c1 )
α
D
.
Since Diam(Ω) = 2, setting K = 1 and
ε1 =
in Theorem 13 gives
E
h
D
2D log n − αr log log n
i
Wrr (P, Pb )
.
1
log n
αr
αr
D
D
.
On the other hand, setting k = n in Theorem 14 gives
inf sup
Pb P ∈P X
1
αr
h
i 1
D
1
r
Wr (P, Pb ) ≥
,
E
2 log(n/c1 )
IID
,...,X ∼ P
n
−αr
showing that distribution estimation over (P, Wrr ) has the extremely slow minimax rate (log n) D .
Although we considered only α ∈ (0, 1] (due to the notational complexity of defining higher-order
Hölder spaces), analogous rates hold for all α > 0. Also, since our rates depend only on covering
and packing numbers of Ω, identical rates can be derived for related Sobolev and Besov classes.
One might wonder why we are interested studying Wasserstein convergence of distributions
over spaces of smooth functions, as in Example 4. Our main motivation comes from the fact that,
historically, smooth function spaces have been widely used for modeling images and other complex
naturalistic signals (Mallat, 1999; Peyré, 2011). Far more recently, empirical breakthroughs have
been made in generative modeling, particularly of images, based on the principle of minimizing
Wasserstein distance between the empirical distribution and a large class of models encoded by a
deep neural network (Montavon et al., 2016; Arjovsky et al., 2017; Gulrajani et al., 2017).
However, little is known about the theoretical properties of these methods; while there has
been some work studying the optimization landscape of such models (Nagarajan and Kolter, 2017),
we know of no work exploring their statistical properties. Given the extremely slow minimax
convergence rate we derived above, it must be the case that the class of distributions encoded by
such models is far smaller or sparser than P. An important avenue for further work is thus to
explicitly identify stronger assumptions that can be made on distributions over interesting classes
of signals, such as images, to bridge the gap between empirical performance and our theoretical
understanding.
12
D ISTRIBUTION E STIMATION
IN
WASSERSTEIN D ISTANCE
8. Conclusion
In this paper, we derived upper and lower bounds for distribution estimation under Wasserstein loss.
Our upper bounds generalize prior results and are tighter in certain cases, while our lower bounds
are, to the best of our knowledge, the first minimax lower bounds for this problem. We also provided
several simple examples for which our upper and lower bounds agree.
8.1. Future Work
We studied only minimax rates over the entire class P of all distributions on the metric space (Ω, ρ).
It would be useful to understand how minimax rates improve when additional assumptions, such as
smoothness or finite-moment assumptions are made (see Fournier and Guillin (2015) for somewhat
improved upper bounds under finite-moment assumptions when (Ω, ρ) is Euclidean space). Given
the rather slow convergence rates we found over P in many cases, studying minimax rates under
assumptions may help to explain the relatively favorable empirical performance of popular distribution estimators based on empirical risk minimization in Wasserstein loss. Moreover, while rates over
all of P are of interest only for very weak metrics such as the Wasserstein distance (as stronger metrics may be infinite or undefined), studying minimax rates under additional assumptions will allow
for a better understanding of the Wasserstein metric in relation to other commonly used metrics.
Acknowledgments
This work was partly supported by a NSF Graduate Research Fellowship DGE-1252522 to S.S. and
by NSF IIS-1563887, Darpa D3M program, and AFRL grants to B.P..
References
Miklós Ajtai, János Komlós, and Gábor Tusnády. On optimal matchings. Combinatorica, 4(4):
259–264, 1984.
Syed Mumtaz Ali and Samuel D Silvey. A general class of coefficients of divergence of one distribution from another. Journal of the Royal Statistical Society. Series B (Methodological), pages
131–142, 1966.
Martin Arjovsky, Soumith Chintala, and Léon Bottou.
arXiv:1701.07875, 2017.
Wasserstein GAN.
arXiv preprint
Amparo Baı́llo, Javier Cárcamo, and Konstantin V Getman. The estimation of Wasserstein and
Zolotarev distances to the class of exponential variables. arXiv preprint arXiv:1603.06806, 2016.
Federico Bassetti, Antonella Bodini, and Eugenio Regazzini. On minimum Kantorovich distance
estimators. Statistics & probability letters, 76(12):1298–1302, 2006.
Espen Bernton, Pierre E Jacob, Mathieu Gerber, and Christian P Robert. Inference in generative
models using the Wasserstein distance. arXiv preprint arXiv:1701.05146, 2017.
Emmanuel Boissard, Thibaut Le Gouic, et al. On the mean speed of convergence of empirical and
occupation measures in Wasserstein distance. In Annales de l’Institut Henri Poincaré, Probabilités et Statistiques, volume 50, pages 539–563. Institut Henri Poincaré, 2014.
13
D ISTRIBUTION E STIMATION
IN
WASSERSTEIN D ISTANCE
Emmanuel Boissard, Thibaut Le Gouic, Jean-Michel Loubes, et al. Distributions template estimate
with wasserstein metrics. Bernoulli, 21(2):740–759, 2015.
James Bucklew. Introduction to rare event simulation. Springer Science & Business Media, 2013.
Sourav Chatterjee et al. A new method of normal approximation. The Annals of Probability, 36(4):
1584–1610, 2008.
Louis HY Chen, Larry Goldstein, and Qi-Man Shao. Normal approximation by Steins method.
Springer Science & Business Media, 2010.
Imre Csiszár. Eine informationstheoretische ungleichung und ihre anwendung auf beweis der ergodizitaet von markoffschen ketten. Magyer Tud. Akad. Mat. Kutato Int. Koezl., 8:85–108, 1964.
Eustasio del Barrio, Evarist Giné, and Carlos Matrán. Central limit theorems for the Wasserstein
distance between the empirical and the true distributions. Annals of Probability, pages 1009–
1071, 1999.
Eustasio del Barrio, Evarist Giné, Carlos Matrán, et al. Correction: Central limit theorems for the
Wasserstein distance between the empirical and the true distributions. The Annals of Probability,
31(2):1142–1143, 2003.
Steffen Dereich, Michael Scheutzow, Reik Schottstedt, et al. Constructive quantization: Approximation by empirical measures. In Annales de l’Institut Henri Poincaré, Probabilités et Statistiques, volume 49, pages 1183–1203. Institut Henri Poincaré, 2013.
Ronald A DeVore and George G Lorentz. Constructive approximation, volume 303. Springer
Science & Business Media, 1993.
Luc Devroye. The equivalence of weak, strong and complete convergence in L1 for kernel density
estimates. The Annals of Statistics, pages 896–904, 1983.
Khanh Do Ba, Huy L Nguyen, Huy N Nguyen, and Ronitt Rubinfeld. Sublinear time algorithms for
earth movers distance. Theory of Computing Systems, 48(2):428–442, 2011.
Joseph L Doob. Measure theory, volume 143. Springer Science & Business Media, 2012.
Richard M Dudley. The sizes of compact subsets of hilbert space and continuity of gaussian processes. Journal of Functional Analysis, 1(3):290–330, 1967.
RM Dudley. The speed of mean Glivenko-Cantelli convergence. The Annals of Mathematical
Statistics, 40(1):40–50, 1969.
Peyman Mohajerin Esfahani and Daniel Kuhn. Data-driven distributionally robust optimization using the wasserstein metric: Performance guarantees and tractable reformulations. arXiv preprint
arXiv:1505.05116, 2015.
Nicolas Fournier and Arnaud Guillin. On the rate of convergence in Wasserstein distance of the
empirical measure. Probability Theory and Related Fields, 162(3-4):707–738, 2015.
14
D ISTRIBUTION E STIMATION
IN
WASSERSTEIN D ISTANCE
Rui Gao and Anton J Kleywegt. Distributionally robust stochastic optimization with wasserstein
distance. arXiv preprint arXiv:1604.02199, 2016.
Ishaan Gulrajani, Faruk Ahmed, Martin Arjovsky, Vincent Dumoulin, and Aaron C Courville. Improved training of wasserstein gans. In Advances in Neural Information Processing Systems,
pages 5769–5779, 2017.
Yanjun Han, Jiantao Jiao, and Tsachy Weissman. Minimax estimation of discrete distributions under
ℓ1 loss. IEEE Transactions on Information Theory, 61(11):6343–6354, 2015.
David Haussler. Sphere packing numbers for subsets of the boolean n-cube with bounded vapnikchervonenkis dimension. Journal of Combinatorial Theory, Series A, 69(2):217–232, 1995.
Nhat Ho, XuanLong Nguyen, Mikhail Yurochkin, Hung Hai Bui, Viet Huynh, and Dinh Phung.
Multilevel clustering via wasserstein means. arXiv preprint arXiv:1706.03883, 2017.
Oliver Johnson, Richard Samworth, et al. Central limit theorem and convergence to stable laws in
mallows distance. Bernoulli, 11(5):829–845, 2005.
Matt Kusner, Yu Sun, Nicholas Kolkin, and Kilian Weinberger. From word embeddings to document
distances. In International Conference on Machine Learning, pages 957–966, 2015.
Stéphane Mallat. A wavelet tour of signal processing. Academic press, 1999.
Grégoire Montavon, Klaus-Robert Müller, and Marco Cuturi. Wasserstein training of restricted
Boltzmann machines. In Advances in Neural Information Processing Systems, pages 3718–3726,
2016.
Tetsuzo Morimoto. Markov processes and the H-theorem. Journal of the Physical Society of Japan,
18(3):328–331, 1963.
Vaishnavh Nagarajan and J Zico Kolter. Gradient descent GAN optimization is locally stable. arXiv
preprint arXiv:1706.04156, 2017.
Kangyu Ni, Xavier Bresson, Tony Chan, and Selim Esedoglu. Local histogram based segmentation
using the wasserstein distance. International journal of computer vision, 84(1):97–111, 2009.
Gabriel Peyré. The numerical tours of signal processing. Computing in Science & Engineering, 13
(4):94–97, 2011.
Matthias Reitzner, Matthias Schulte, et al. Central limit theorems for u-statistics of poisson point
processes. The Annals of Probability, 41(6):3879–3909, 2013.
Phillippe Rigollet. High-dimensional statistics. Lecture notes for course 18S997, 2015.
Emmanuel Rio et al. Upper bounds for minimal distances in the central limit theorem. In Annales de
l’Institut Henri Poincaré, Probabilités et Statistiques, volume 45, pages 802–817. Institut Henri
Poincaré, 2009.
Emmanuel Rio et al. Asymptotic constants for minimal distance in the central limit theorem. Electronic Communications in Probability, 16:96–103, 2011.
15
D ISTRIBUTION E STIMATION
IN
WASSERSTEIN D ISTANCE
Yossi Rubner, Carlo Tomasi, and Leonidas J Guibas. The earth mover’s distance as a metric for
image retrieval. International journal of computer vision, 40(2):99–121, 2000.
Ludger Rüschendorf. The wasserstein distance and approximation theorems. Probability Theory
and Related Fields, 70(1):117–129, 1985.
Roman Sandler and Michael Lindenbaum. Nonnegative matrix factorization with earth mover’s
distance metric for image analysis. IEEE Transactions on Pattern Analysis and Machine Intelligence, 33(8):1590–1602, 2011.
Justin Solomon, Fernando De Goes, Gabriel Peyré, Marco Cuturi, Adrian Butscher, Andy Nguyen,
Tao Du, and Leonidas Guibas. Convolutional wasserstein distances: Efficient optimal transportation on geometric domains. ACM Transactions on Graphics (TOG), 34(4):66, 2015.
Cédric Villani. Optimal transport: old and new, volume 338. Springer Science & Business Media,
2008.
Jonathan Weed and Francis Bach. Sharp asymptotic and finite-sample rates of convergence of
empirical measures in Wasserstein distance. arXiv preprint arXiv:1707.00087, 2017.
Meng Zhang, Yang Liu, Huan-Bo Luan, Maosong Sun, Tatsuya Izuha, and Jie Hao. Building earth
mover’s distance on bilingual word embeddings for machine translation. In AAAI, pages 2870–
2876, 2016.
Tong Zhang. Covering number bounds of certain regularized linear function classes. Journal of
Machine Learning Research, 2(Mar):527–550, 2002.
Ding-Xuan Zhou. The covering number in learning theory. Journal of Complexity, 18(3):739–767,
2002.
16
D ISTRIBUTION E STIMATION
IN
WASSERSTEIN D ISTANCE
Appendix A. Proofs of Lemmas
Lemma 7 Suppose S ∈ S is a countable Borel partition of Ω. Let P and Q be Borel probability
measures such that, for every S ∈ S, P (S) = Q(S). Then, for any r ≥ 1, Wr (P, Q) ≤ Res(S).
Proof This fact is intuitively obvious; clearly, there exists a transportation map µ from P to Q that
moves mass only within each S ∈ S and therefore without moving any mass further than δ. For
completeness, we give a formal construction.
Let µ : Σ2 → [0, 1] denote the coupling that is conditionally independent given any set S ∈ S
with P (S) = Q(S) > 0 (that is, for any A, B ∈ Σ, µ(A×B ∩S ×S)P (S) = P (A∩S)Q(B ∩S)).4
It
S is easy to verify that µ ∈ C(P, Q). Since S is a countable partition and µ is only supported on
S∈S S × S,
Wr (P, Q) ≤
Z
Ω×Ω
XZ
=
S∈S
XZ
≤
S∈S
=δ
1/r
ρ (x, y) dµ(x, y)
r
X
S∈S
!1/r
r
ρ (x, y) dµ(x, y)
S×S
!1/r
r
δ dµ(x, y)
S×S
!1/r
µ(S × S)
=δ
X P (S)Q(S)
P (S)
S∈S
!1/r
=δ
X
!1/r
Q(S)
S∈S
= δ.
Lemma 8 Suppose (Ω, ρ) is a metric space, and suppose P and Q are Borel probability distributions on Ω with countable support; i.e., there exists a countable set X ⊆ Ω with P (X ) = Q(X ) = 1.
Then, for any r ≥ 1,
X
X
|P ({x}) − Q({x})| .
|P ({x}) − Q({x})| ≤ Wrr (P, Q) ≤ (Diam(X ))r
(Sep(X ))r
x∈X
x∈X
P
Proof The term x∈X |P ({x}) − Q({x})| = T V (P, Q) is precisely the (unweighted) amount
of mass that must be transported to transform between P and Q. Hence, the result is intuitively
fairly obvious; all mass moved has a cost of at least Sep(Ω) and at most Diam(Ω). However, for
completeness, we give a more formal proof below.
To prove the lower bound, suppose µ ∈ Π(P, Q) is any coupling between P and Q. For x ∈ X ,
µ({x} × {x}) + µ({x} × (Ω\{x})) = µ({x} × Ω) = P ({x})
and, similarly,
µ({x} × {x}) + µ((Ω\{x}) × {x}) = µ(Ω × {x}) = Q({x}).
4. The existence of such a measure can be verified by the Hahn-Kolmogorov theorem, similarly to that of the usual
product measure (see, e.g., Section IV.4 of Doob (2012)).
17
D ISTRIBUTION E STIMATION
IN
WASSERSTEIN D ISTANCE
Since P ({x}), Q({x}) ∈ [0, 1], it follows that
µ({x} × (Ω\{x})) + µ(µ((Ω\{x}) × {x})) ≥ |P ({x} − Q({x})| .
Therefore, since ρ(x, y) = 0 whenever x = y and ρ(x, y) ≥ Sep(Ω) whenever x 6= y,
Z
Z
ρr (x, y) dµ(x, y)
ρr (x, y) dµ(x, y) =
X ×X
Ω×Ω
Z
XZ
r
ρr (x, y) dµ(x, y)
ρ (x, y) dµ(x, y) +
=
x∈X
{x}×(Ω\{x})
≥ (Sep(Ω))r
r
≥ (Sep(Ω))
X
x∈X
X
x∈X
(Ω\{x})×{x}
µ({x} × (Ω\{x})) + µ((Ω\{x}) × {x})
|P ({x}) − Q({x})| .
Taking the infimum over µ on both sides gives
X
|P ({x}) − Q({x})| ≤ Wrr (P, Q).
(Sep(Ω))r
x∈X
To prove the upper bound, since ρ is upper bounded by Diam(Ω), it suffices to construct a coupling
µ that only moves mass into or out of each given point, but not both; that is, for each x ∈ X ,
min{µ({x} × (Ω\{x})), µ((Ω\{x}) × {x})} = 0.
One way of doing this is as follows. Fix an ordering x1 , x2 , ... of the elements of X . For each i ∈ N,
define
i
i
X
X
Xi :=
(P (xℓ ) − Q(xℓ ))+ and Yi :=
(Q(xℓ ) − P (xℓ ))+ ,
ℓ=1
ℓ=1
and further define
ji := min{j ∈ N : Xi ≤ Yj }
and
ki := min{k ∈ N : Xj ≥ Yi }.
Then, for each i ∈ N, move Xi mass from {x1 , ..., xi } to {y1 , ..., yji } and move Yi mass from
{y1 , ..., yi } to {x1 , ..., xki }. As i → ∞, by construction of Xi and Yi , the total mass moved in this
way is
X
|P (x) − Q(x)| .
µ((X × X )\{(x, x) : x ∈ X }) = lim Xi + Yi =
i→∞
x∈X
Lemma 10 Let K be a positive integer. Suppose {Sk }k∈[K] is a sequence of nested countable Borel
partitions of (Ω, ρ), with S0 = Ω. Then, for any r ≥ 1 and any Borel probability distributions P
and Q on Ω,
K
X
X
(Res(Sk−1 ))r
Wrr (P, Q) ≤ (Res(SK ))r +
|P (S) − Q(S)| .
S∈Sk
k=1
18
D ISTRIBUTION E STIMATION
IN
WASSERSTEIN D ISTANCE
Proof Our proof follows the same ideas as and slightly generalizes of the proof of Proposition 1 in
Weed and Bach (2017). Intuitively, to prove Lemma 10 it suffices to find a transportation map such
that For each k ∈ [K], recursively define
Pk := P −
k−1
X
µk
and
j=0
Qk := Q −
k−1
X
νk ,
j=0
where, for each k ∈ [K], µk and νk are Borel measures on Ω defined for any E ∈ Σ by
µk (E) :=
X
(Pk (S) − Qk (S))+
Pk (E ∩ S)
Pk (S)
X
(Qk (S) − Pk (S))+
Qk (E ∩ S)
.
Qk (S)
S∈Sk :Pk (S)>0
and
νk (E) :=
S∈Sk :Qk (S)>0
P
By construction of µk and νk , each µk and νk is a non-negative measure and K
k=1 µk ≤ P and
PK
k=1 νk ≤ Q. Furthermore, for each k ∈ [K − 1], for each S ∈ Sk , µk+1 (S) = νk+1 (S), and
X
|P (S) − Q(S)| .
µk (Ω) = νk (Ω) ≤
S∈Sk
Consequently, although µ and ν are not probability measures, we can slightly generalize the definition of Wasserstein distance by writing
Wrr (µk , νk ) := µ(Ω)
τ ∈Π
inf
ν
µk
, k
µk (Ω) νk (Ω)
[ρr (X, Y )]
E
(X,Y )∼µ
(or Wrr (µk , νk ) = 0 if µk = νk = 0). In particular, this is convenient because we one can easily
show that, by construction of the sequences {Pk }k∈[K] and {Qk }k∈[K] ,
Wrr (P, Q) ≤ Wrr (PK , QK ) +
For each k ∈ [K], Lemma 8 implies that
X
Wrr (µk , νk ) ≤
(Diam(S))r
S∈Sk−1
≤ (Res(Sk−1 ))
r
= (Res(Sk−1 ))
r
X
K
X
Wrr (µk , νk ) .
k=1
X
T ∈Sk :T ⊆S
|P (T ) − Q(T )|
X
S∈Sk−1 T ∈Sk :T ⊆S
X
T ∈Sk
|P (T ) − Q(T )| .
Furthermore, for each S ∈ SK , PK = QK , Lemma 7 gives that
Wrr (PK , QK ) ≤ (Res(SK ))r
19
|P (T ) − Q(T )|
(7)
D ISTRIBUTION E STIMATION
IN
WASSERSTEIN D ISTANCE
Plugging these last two inequalities into Inequality (7) gives the desired result:
Wrr (P, Q)
r
≤ (Res(SK )) +
K
X
(Res(Sk−1 ))r
X
S∈Sk
k=1
|P (S) − Q(S)| .
Lemma 11 Suppose S and T are partitions of (Ω, ρ), and suppose S is countable. Then, there
exists a partition S ′ of (Ω, ρ) such that:
a) |S ′ | ≤ |S|.
b) Res(S ′ ) ≤ Res(S) + 2 Res(T ).
c) T is a refinement of S ′ .
Proof Enumerate the elements of S as S1 , S2 , .... Define S0′ := ∅, and then, for each i ∈ {1, 2, ...},
recursively define
-
i−1
[
[
T
Sj′ ,
Si′ :=
j=1
T ∈T :T ∩Si 6=∅
and set S ′ = {S1′ , S2′ , ...}. Clearly, |S ′ | ≤ |S| (equality need not hold, as we may have some
Si′ = ∅). By the triangle inequality, each
[
Diam(Si′ ) ≤ Diam
T ≤ δS + 2δT .
T ∈T :T ∩Si 6=∅
Finally, since T is a partition and we can write
-
i−1
[
[
T
Si′ =
j=1 T ∈T
T ∈T :T ∩Si 6=∅
T is a refinement of S ′ .
[
:T ∩Sj′ 6=∅
T ,
Before proving Lemma 12, we recall a simple concentration inequality bounding the L1 deviation of a multinomial random variable from its mean:
Lemma 15 (Lemma 3 P
of Devroye (1983)) Suppose (X1 p
, ..., XK ) ∼ Multinomial(n, p1 , ..., pK ).
20K/n, for all ε ∈ [ε0 , 1),
Let Z := kX − pk1 = K
|X
−
p
|.
Then,
for
ε
:=
0
k
k
k=1
P [Z > nε] ≤ 3 exp −nε2 /25 .
Lemma 12 now follows as an easy corollary.
20
D ISTRIBUTION E STIMATION
IN
WASSERSTEIN D ISTANCE
Lemma 12 Consider the setting of Lemma 15. Then,
p
Z
≤ 21K/n.
E
n
Proof Since, almost surely, Z < n, by Lemma 15,
Z ∞
P [Z > nε] dε
[Z/n]
=
E
0
Z 1
≤ ε0 + 3
exp −nε2 /25 dε
Z
ε0
1−ε0
exp −n (ε + ε0 )2 /25 dε
0
Z
1
p
3
2
≤ ε0 + 3 exp −nε0
exp −nε2 /25 dε ≤ ε0 + exp −nε20 ≤ 21K/n.
2
0
= ε0 + 3
21
D ISTRIBUTION E STIMATION
IN
22
WASSERSTEIN D ISTANCE
| 7cs.IT
|
Fitting Jump Models
arXiv:1711.09220v1 [cs.LG] 25 Nov 2017
A. Bemporad† , V. Breschi† , D. Piga‡ , and S. Boyd§
†
IMT School for Advanced Studies Lucca, Piazza San Francesco 19, 55100 Lucca,
Italy. Email: {alberto.bemporad,valentina.breschi}@imtlucca.it
‡
Dalle Molle Institute for Artificial Intelligence Research - USI/SUPSI, Galleria 2,
Via Cantonale 2c, CH-6928 Manno, Switzerland. Email: [email protected]
§
Department of Electrical Engineering, Stanford University, Stanford CA 94305,
USA. Email: [email protected]
November 28, 2017
Abstract
We describe a new framework for fitting jump models to a sequence of
data. The key idea is to alternate between minimizing a loss function to
fit multiple model parameters, and minimizing a discrete loss function to
determine which set of model parameters is active at each data point. The
framework is quite general and encompasses popular classes of models,
such as hidden Markov models and piecewise affine models. The shape of
the chosen loss functions to minimize determine the shape of the resulting
jump model.
1
Introduction
In many regression and classification problems the training dataset is formed
by input and output observations with time stamps. However, when fitting
the function that maps input data to output data, most algorithms used in
supervised learning do not take the temporal order of the data into account. For
example, in linear regression problems solved by least squares minθ kAθ − bk22
each row of A and b is associated with a data-point, but clearly the solution
θ? is the same no matter how the rows of A and b are ordered. In systems
identification of dynamical models, temporal information is often used only to
construct the input samples (or regressors) and outputs, but then it is neglected.
For example, in estimating autoregressive models with exogenous inputs (ARX),
the regressor is a finite collection of current and past signal observations, but
the order of the regressor/output pairs is irrelevant when least squares are used.
Similarly, in logistic regression and support vector machines the order of the
data points does not change the result. In training forward neural networks
using stochastic gradient descent, the samples may be picked up randomly (and
1
more than once) by the solution algorithm, and again their original temporal
ordering is neglected.
On the other hand, there are many applications in which relevant information
is contained not only in the data but also in their temporal order. In particular,
if the time each data-point was collected is taken into account, one can detect
changes in the type of regime the data were produced. Examples range from
video segmentation [19, 9] to speech recognition [23, 24], asset-price models
in finance [25, 14], human action classification [22, 21], and many others. All
these examples are characterized by the need of fitting multiple models and
understanding when switches from one model to another occur.
Piecewise affine (PWA) models attempt at fitting multiple affine models to
a data set, where each model is active based on the location of the input sample in a polyhedral partition of the input space [12, 8]. However, as for ARX
models, the order of the data is not relevant in computing the model parameters and the polyhedral partition. In some cases, mode transitions are captured
by finite state machines, for example in hybrid dynamical models with logical
states, where the current mode and the next logical state are generated deterministically by Boolean functions [4, 7]. Hidden Markov models (HMMs) treat
instead the mode as a stochastic discrete variable, whose temporal dynamics is
described by a Markov chain [23]. Natural extensions of hidden Markov models
consider the cases in which each mode is associated with a linear function of
the input [13, 10, 20]. Hidden Markov models are usually trained using the
Baum-Welch algorithm [1], a forward-backward version of the more general Expectation Maximization (EM) algorithm [11].
In this paper we consider rather general jump models to fit a temporal sequence of data that takes the ordering of the data into account. The proposed
fitting algorithm alternates two steps: estimate the parameters of multiple models and estimate the temporal sequence of model activation, until convergence.
The model fitting step can be carried out exactly when it reduces to a convex optimization problem, which is often the case. The mode-sequence step is
always carried out optimally using dynamic programming.
Our jump modeling framework is quite general. The structure of the model
depends on the shape of the function that is minimized to obtain the model parameters, the way the model jumps depends on the function that is minimized
to get the sequence of model activation. When we impose no constraints or
penalty on the model sequence, our method reduces to automatically splitting
the dataset in K clusters and fitting one model per cluster, which is a generalization of K-means [16, Algorithm 14.1]. Hidden Markov models (HMMs) are a
special case of jump models, as we will show in the paper. Indeed, jump models
have broader descriptive capabilities than HMMs, for example the sequence of
discrete states may not be necessarily generated by a Markov chain and could
be a deterministic function. Moreover, as stated above, jump models can have
rather arbitrary model shapes.
After introducing jump models in Section 2 and giving a statistical interpretation of the loss function in Section 3, we provide algorithms for fitting jump
models to data and to estimate output values and hidden modes from available
2
input samples in Section 4, emphasizing differences and analogies with HMMs.
Finally, in Section 5 we show four examples of application of our approach for
regression and classification, using both synthetic and experimental data sets.
The code implementing the algorithms described in the paper is available at
http://cse.lab.imtlucca.it/~bemporad/jump_models/.
1.1
Setting and goal
We are given a training sequence of data pairs (xt , yt ), t = 1, . . . , T , with xt ∈ X ,
yt ∈ Y. We refer to t as the time or period, xt as the regressor or input, and
yt as the outcome or output at time t. The training sequence is used to build
a regression model that provides a prediction ŷt of yt given the available inputs
x1 , . . . , xt , and possibly past outputs y1 , . . . , yt−1 . We are specifically interested
in models where ŷt is not simply a static function of xt , but rather we want
to exploit the additional information embedded in the temporal ordering of
the data. As we will detail later, our regression model is implicitly defined by
the minimization of a fitting loss J that depends on x1 , . . . , xt , y1 , . . . , yt−1 , yt
and other variables and parameters. The chosen shape for J determines the
structure of the corresponding regression model.
Given a production data sequence (x̃1 , ỹ1 ), . . ., thought to be generated by
a similar process that produced the training data, the quality of the regression
model over a time period t = 1, . . . , T̃ will be judged by the average true loss
true
L
T̃
1 X true
` (ŷt , ỹt )
=
T̃ t=1
(1)
where `true : Y × Y → R penalizes the mismatch between ŷt and ỹt , with
`(y, y) = 0 for all y ∈ Y.
2
2.1
Regression models
Single model
A simple form of deriving a regression model is to introduce a model parameter
θ ∈ Rd , a loss function ` : X × Y × Rd → R ∪ {+∞}, and a regularizer r : Rd →
R ∪ {+∞} defining the fitting objective
J(X, Y, θ) =
T
X
`(xt , yt , θ) + r(θ)
(2a)
t=1
where X = (x1 , . . . , xT ), Y = (y1 , . . . , yT ). For a given training data set (X, Y ),
let
θ? = arg minθ J(X, Y, θ)
(2b)
be the optimal model parameter. By fixing θ = θ? we get the following regression
model
ŷt = arg minyt J(X, Y, θ? ) = arg minyt `(xt , yt , θ? ) =: ϕ(xt )
3
(2c)
where we have exploited the separability of the loss J and set ϕ : X → Y as the
regression model, with ties in the arg min broken arbitrarily.
Model (2) can be enriched by adding output information sets Yt ⊆ Y that
augment the information that is available about yt ,
ŷt = arg miny∈Yt `(x, y, θ? )
(3)
where Yt = Y if no extra information on yt is given. For example, if we know a
priori that yt ≥ 0 we can set Yt equal to the nonnegative orthant.
2.2
K-models
Let us add more flexibility and introduce multiple model parameters θs ∈ Rd ,
s = 1, . . . , K, and a latent mode variable st that determines the model parameter
θst that is active at step t. Fitting a K-model on the training data set (X, Y ),
entails choosing the K models by minimizing
J(X, Y, Θ, S) =
T
X
`(xt , yt , θst ) +
t=1
K
X
r(θi )
(4)
i=1
with respect to Θ = (θ1 , . . . , θK ) and S = (s1 , . . . , sT ). The optimal parameters
?
define the K-model
θ1? , . . . , θK
(ŷt , ŝt ) = arg minyt ,st `(xt , yt , θs?t ).
(5)
For example, the loss function
`(xt , yt , θst ) = kyt − θs0 t xt k22
(6)
corresponds to K possible linear models ŷt = θs0 t xt . By setting
r(θ) = 0,
`(xt , yt , θst ) = kyt − θst k22 + kxt − θst k22
(7)
and yt = xt , model (5)–(7) is the standard K-means model [16], where θs are
the centroids of the clusters. In this case, ŝt denotes the index of the centroid
θŝ?t which is closest to xt , and ŷt = θŝ?t is the best estimate of xt based on the
regression model defined by (7). More generally,
0
`(xt , yt , θst ) = kyt − θy,s
[ x1t ] k22 + ρkxt − θx,st k22
t
(8)
with θst = (θy,st , θx,st ) and ρ > 0, defines a piecewise affine (PWA) model over
?
?
the piecewise linear partition generated by the Voronoi diagram of (θx,1
, . . . , θx,K
),
i.e., the regression model is
?
ŝt = arg mins kxt − θx,s
k22 ,
?
ŷt = (θy,ŝ
)0 [ x1t ] .
t
The hyper-parameter ρ in (8) trades off between fitting the output yt and clustering the inputs (x1 , . . . , xt ) based on Euclidean distance.
4
A more general PWA model can be defined by considering a piecewise linear
separation function ϑ : X → R [5, 8]
0
ϑ(x) = max{θx,s
[ x1 ]}
(9)
s
and setting
`(xt , yt , θst ) =
0
kyt − θy,s
t
[ x1t
] k22 + ρ
K
X
2
max {0, (θx,j − θx,st )0 [ x1t ] + 1} . (10)
j=1
j6=st
In this case it is immediate to verify that the regression model is
ŝt = arg maxs {(θx?t ,s )0 [ x1 ]},
2.3
?
ŷt = (θy,ŝ
)0 [ x1t ] .
t
(11)
Jump model
The models introduced above do not take into account the temporal order in
which the samples (xt , yt ) are generated. To this end, we add a mode sequence
loss L in the fitting objective (4)
J(X, Y, Θ, S) =
T
X
`(xt , yt , θst ) +
t=1
K
X
r(θk ) + L(S),
(12)
k=1
where S = (s0 , s1 , . . . , sT ) is the mode sequence. We define L : KT +1 → R ∪
{+∞} in (12) as
L(S) = Linit (s0 ) +
T
X
Lmode (st ) +
t=1
T
X
Ltrans (st , st−1 )
(13a)
t=1
where K = {1, . . . , K}, Linit : K → R ∪ {+∞} is the initial mode cost, Lmode :
K → R ∪ {+∞} is the mode cost, and Ltrans : K2 → R ∪ {+∞} is the mode
transition cost. We discuss possible choices for L in Sections 2.3.1 and 3.
With a little abuse of notation, we write
J(X, Y, Θ, S) = `(X, Y, Θ, S) + r(Θ) + L(S)
where
`(X, Y, Θ, S) =
T
X
`(xt , yt , θst ),
t=1
r(Θ) =
K
X
r(θk ).
(13b)
(13c)
k=1
As with any model, the choice of the fitting objective (13) should trade off
between fitting the given data and prior assumptions we have about the models
and the mode sequence. In particular, the mode sequence loss L in (13a) takes
into account the temporal structure of the mode sequence, for example that the
mode might change (i.e., st 6= st−1 ) rarely.
5
A jump model can be used for several tasks beyond inferring the values ŷt .
In anomaly identification, we are interested in determining times t for which
the jump model does not fit the data point yt well. In model change detection
we interested in identifying times t for which ŝt 6= ŝt−1 . In control systems
jump models can be used to approximate nonlinear/discontinuous dynamics
and design model-based control policies, state estimators, and fault-detection
algorithms.
2.3.1
Mode loss functions
We discuss a few options for choosing the mode loss functions Linit , Lmode ,
Ltrans defining the mode sequence loss L in (13a).
As mentioned above, the case L(S) = 0 leads to a K-model. By choosing
Ltrans (i, j) = λ for all i 6= j, Lmode (i) = Ltrans (i, i) = 0, one penalizes mode
transitions equally by λ ≥ 0, where λ → ∞ leads to regression of a single model
on the data (that is, st ≡ s0 ), while λ → 0 leads again to a K-model. Note
that choosing the same constant λ for all transitions makes the fitting problem
exhibit multiple solutions, as indexes i, j can be arbitrarily permuted. The
mode loss Lmode can be used to break such symmetries. For example, smaller
values for st will be preferred by making Lmode (i) < Lmode (j) for i < j. The
shape of the increasing finite sequence {Lmode (i)}K
i=1 can be used to reduce the
number of possible modes: larger increasing values of Lmode (i) will discourage
the use of an increasing number of modes.
The initial mode cost Linit summarizes prior knowledge about the initial
mode s0 . For example, Linit (s0 ) ≡ 0 if no prior information on s0 is available. On the contrary, if the initial mode s0 is known and say equal to j, then
Linit (s0 ) = 0 for s0 = j and +∞ otherwise.
Next Section 3 suggests criteria for choosing L in case statistical assumptions
about the underlying process that generates st are available. Alternative criteria
are discussed in Section 4.4 for choosing L directly from the training data.
3
Statistical interpretations
Let Y = (y1 , . . . , yT ), X = (x1 , . . . , xT ), S = (s0 , . . . , sT ), Θ = (θ1 , . . . , θK ). We
provide a statistical interpretation of the loss functions for the special case in
which the following modeling assumptions are satisfied:
A1. The mode sequence S, the model parameters Θ and the input data X are
statistically independent, i.e.,
p(S|X, Θ) = p(S),
p(Θ|S, X) = p(Θ)
A2. The conditional likelihood of Y is given by
p(Y |X, S, Θ) =
T
Y
p(yt |X, S, Θ) =
t=1
T
Y
t=1
6
p(yt |xt , θst )
where p(yt |xt , θst ) is the likelihood of the outcome yt given xt and θst ;
A3. The priors on the model parameters θ1 , . . . , θK are all equal to p(θ), i.e.,
p(θ1 ) = · · · = p(θK ) = p(θ)
and the model parameters are statistically independent, i.e.,
p(Θ) =
K
Y
p(θk )
k=1
A4. The probability of being in mode st given s0 , . . . , st−1 is p(st |st − 1) =
πst ,st−1 (Markov property);
A5. The initial mode s0 has probability p(s0 ) = πs0 .
Proposition 1 Let Assumptions A1-A5 be satisfied and define
`(xt , yt , θst )
r(θk )
Ltrans (st , st−1 )
init
L
(s0 )
mode
L
(st )
= − log p(yt |xt , θst )
(14a)
= − log p(θk )
(14b)
= − log πst ,st−1
(14c)
= − log πs0
(14d)
=
(14e)
0.
Then minimizing J(X, Y, Θ, S) as defined in (12)–(14) with respect to Θ and S
is equivalent to maximizing the joint likelihood p(Y, S, Θ|X).
Proof. Because of the Markov property (Assumption A4), the likelihood of the
mode sequence S is
T
Y
p(S) = p(s0 )
p(st |st−1 ).
(15)
t=1
From (15) and Assumptions A1-A3, we have:
p(Y, S, Θ|X)
=
p(Θ|X)p(Y, S|X, Θ) = p(Θ|X)p(S|X, Θ)p(Y |S, X, Θ)
=
p(Θ)p(S)p(Y |S, X, Θ)
=
K
Y
k=1
p(θk )p(s0 )
T
Y
p(st |st−1 )p(yt |xt , θst )
t=1
whose logarithm is
log p(Y, S, Θ|X) =
K
X
T
T
X
X
log p(θk )+log p(s0 )+
log p(st |st−1 )+
log p(yt |xt , θst ).
t=1
k=1
t=1
(16)
By defining the loss functions `, r, Ltrans , Linit , and Lmode as in (14), the
minimization of the fitting objective J(X, Y, Θ, S) as in (12)–(13) with respect
7
to Θ and S is equivalent to maximizing the logarithm of the joint likelihood
p(Y, S, Θ|X), and therefore p(Y, S, Θ|X).
It is also possible to give an inverse result, namely a statistical interpretation of minimizing a given generic J(X, Y, Θ, S) defined as in (13). Define the
following probabilities
p(yt |xt , θst )
=
e−`(xt ,yt ,θst )
ν(θst , xt )
p(Θ|X) = p(Θ)
=
Z
(17a)
e−r(Θ)
(17b)
e−r(Θ) dΘ
Rd×K
−L̄(S,Θ,X)
p(S|Θ, X)
e
X
=
e−L̄(S̄,Θ,X)
(17c)
S̄∈K T +1
where
Z
ν(θst , xt )
=
e−`(xt ,y,θst ) dy
(18a)
Y
L̄(S, Θ, X)
= L(S) −
T
X
log ν(θst , xt ).
(18b)
t=1
Under the assumption that the outputs yt are independent given (S, X, Θ), we
QT
have that p(Y |S, X, Θ) = t=1 p(yt |xt , θst ). Furthermore, since
p(Y, S, Θ|X) = p(Y |S, X, Θ)p(S|X, Θ)p(Θ|X)
we get
arg max log p(Y, S, Θ|X) = arg min J(X, Y, Θ, S).
(19)
PK
Note that the standard quadratic regularization r(Θ) = k=1 r(θk ), with r(θk ) =
kθ k2
q
− k 22
1
2
, which amounts to assuming
ρkθk k2 , leads to p(θk ) = ce 2σθ , where σθ = 2ρ
θk ∼ N (0, σθ2 I). Similarly, the quadratic penalty on the prediction error
`(xt , yt , θst ) = ckyt − θs0 t xt k22
(20)
q
1
.
leads to assuming that yt ∼ N (θs0 t xt , σy2 I), where σy = 2c
The special cases highlighted in Section 2.3.1 also have a statistical interpretation. Let us assume for simplicity that ν(θst , xt ) = T1 in (18a), so that
we can replace L̄(S, Θ, X) = L(S) in (18b) without changing the optimizer of
J(X, Y, Θ, S) in (19), and also assume that Linit (s) = Lmode (s).
When Ltrans = 0 we can set
p(S|Θ, X) = p(S) =
T
Y
t=0
8
mode
e−L
PK
k=1
(st )
e−Lmode (k)
Algorithm 1 Jump model fitting
Input: Training data set X = (x1 , . . . , xT ), Y = (y1 , . . . , yT ), number K of
models, initial mode sequence S 0 = {s00 , . . . , s0T }.
1. iterate for k = 1, . . .
1.1. Θk ← arg minΘ `(X, Y, Θ, S k−1 ) + r(Θ);
k
k
1.2. S ← arg minS `(X, Y, Θ , S) + L(S);
(model fitting)
(mode sequence fitting)
2. until S k = S k−1 .
Output: Estimated model parameters Θ? = Θk and mode sequence S ? = S k .
in (17c).
The case L(S) = 0 corresponds to assuming that the modes st are i.i.d.
1
(independently and identically distributed) with uniform probability K
for all
t = 0, . . . , T .
The case Ltrans = 0, Linit (s) = Lmode (s) = s corresponds to assuming the
(e−1) −s
modes st i.i.d. with probability distribution p(s) = 1−e
.
−K e
mode
trans
Regarding the case L
= 0 and L
(i, j) = λ for i 6= j and 0 for i = j,
QT
by assuming p(S|Θ, X) = p(S) = p(s0 ) t=1 p(st |st−1 ) we get
(
e−λ
if st 6= st−1
1
−λ
1+(K−1)e
p(s0 ) = , p(st |st−1 ) =
1
K
if st = st−1 .
1+(K−1)e−λ
4
Algorithms
We provide now algorithms for fitting a jump model to a given data set and to
infer predictions ŷt , ŝt from it.
4.1
Model fitting
Given a training sequence X = (x1 , . . . , xT ) of inputs and Y = (y1 , . . . , yT ) of
outputs, for fitting a jump K-model we need to attempt minimizing the cost
J(X, Y, Θ, S) with respect to Θ and S. A simple algorithm to solve this problem
is Algorithm 1, a coordinate descent algorithm that alternates minimization
with respect to Θ and S. If ` and r are convex functions, Step 1.1 can be solved
globally (up to the desired precision) by standard convex programming [6].
Step 1.2 can be solved by standard discrete dynamic programming (DP) [2].
The latter is achieved by computing the following matrices M ∈ RK×(T +1) of
9
costs and U ∈ K × RT of indexes
M (s, T ) = Lmode (s) + `(xT , yT , θs )
(21a)
trans
Us,t = arg minj {M (j, t + 1) + L
mode
M (s, t) = L
init
M (s, 0) = L
(j, s)}, t = 1, . . . , T − 1
trans
(s) + `(xt , yt , θs ) + M (Us,t , t + 1) + L
(s) + min{M (j, 1) + L
trans
j
(Us,t , s)
(j, s)}
(21b)
(21c)
(21d)
backwards in time, and then reconstructing the minimum cost sequence S forward in time by setting
s0 = arg minj M (j, 0)
(21e)
st = Ust−1 ,t , t = 1, . . . , T.
(21f)
Note that if the time order of operations in (21) is reversed, the DP iterations (21) become Viterbi algorithm [23, p. 264]:
M (s, 0) = Linit (s)
(22a)
trans
Us,t = arg minj {M (j, t − 1) + L
mode
M (s, t) = L
(j, s)}, t = 1, . . . , T
trans
(s) + `(xt , yt , θs ) + M (Us,t , t − 1) + L
(Us,t , s)
(22b)
(22c)
followed by the backwards iterations
sT = arg minj M (j, T )
(22d)
st = Ust+1 ,t , t = 0, . . . , T − 1.
(22e)
Since at each iteration the cost J(X, Y, Θ, S) is non-increasing and the number
of sequences S is finite, Algorithm 1 always terminates in a finite number of
steps, assuming that in case of multiple optima one selects the optimizers in
Steps 1.1 and 1.2 according to some predefined criterion. However, there is no
guarantee that the solution found is the global one, as it depends on the initial
guess S 0 . To improve the quality of the solution, we may run Algorithm 1 N
times from different random initial sequences S 0 and select the best result. Our
experience is that a small N , say N = 5, is usually enough.
During the execution of Algorithm 1 it may happen that a mode s does not
appear in the sequence S k−1 . In this case, the fitting loss `(X, Y, Θ, S k−1 ) does
not depend on θs , and the latter will be determined in Step 1.1 based only on
the regularizer r(Θ).
In case L(S) = 0, the ordering of the training data becomes irrelevant and
Algorithm 1 reduces to fitting K models to the data set. If in addition ` and
r are specified as in (7) and Y = X, Algorithm 1 is the standard K-means
algorithm, where the starting sequence S 0 is the initial clustering of the data
points (x1 , . . . , xT ), Step 1.1 computes the collection Θk of cluster centroids at
iteration k, and Step 1.2 reassigns data points to clusters by updating their
labels skt .
10
When again L(S) = 0 and the mode loss in (10) is used for getting a PWA
model, the cost function minimized in Step 1.1 of Algorithm 1 is separable with
respect to θy,s , θx,s . Then the minimization with respect to θx,s produces the
PWL separation function ϑ in (9) [8], while Step 1.2 looks for the optimal latent
variables st that best trade off between assigning the corresponding data point
0
0
xt to the polyhedron {x ∈ X : θx,s
[ x ] ≥ θx,j
[ x1 ] , ∀j 6= st , j ∈ K} and matching
t 1
xt
0
the predicted output yt ≈ θy,st [ 1 ].
Finally, we remark that Algorithm 1 is also applicable to the more general
case in which the mode loss L also depends on Θ, by simply replacing Steps 1.1
and 1.2 with
Θk ← arg minΘ `(X, Y, Θ, S k−1 ) + r(Θ) + L(S k−1 , Θ)
k
k
k
S ← arg minS `(X, Y, Θ , S) + L(S, Θ ).
(23a)
(23b)
This would cover the case in which L contains parameters to be estimated.
4.2
Inference
4.2.1
One-step ahead prediction
Assume that the model parameters Θ? have been estimated and that new production data X̃t = (x̃1 , . . . , x̃t ) and outputs Ỹt−1 = (ỹ1 , . . . , ỹt−1 ) are given.
Because of the structure of the mode loss function L defined in (13a), the estimates ŷt and ŝ0 , . . . , ŝt do not depend on future inputs x̃j and modes ŝj for
j > t.
The same fitting objective (12) can be used to estimate ŷt and Ŝt = (ŝ0 , . . . , ŝt ),
(ŷt , Ŝt ) = arg minyt ,St Jt (X̃t , Ỹt−1 , yt , Θ? , St )
s.t. yt ∈ Yt
(24)
where Yt ⊆ Y is a possible additional output information set and
Jt (X̃t , Ỹt−1 , yt , Θ? , St )
=
`(x̃t , yt , θs?t ) +
t−1
X
`(x̃j , ỹj , θs?j ) + Linit (s0 )
j=1
+
t
X
Lmode (sj ) +
j=1
t
X
Ltrans (sj , sj−1 ).
(25)
j=1
Algorithm 2 attempts at solving problem (24). Step 1 is solved again by the
DP iterations (21) over the time span [0, t], with the only difference that in (21a)
we set the terminal penalty equal to M (s, t) = Lmode (s) + miny {`(xt , y, θs )},
since the last output yt is determined later at Step 2.
Note that open-loop prediction, that is the task of predicting ŷt without
acquiring Ỹt−1 , can be simply obtained by replacing Ỹt−1 = (ỹ1 , . . . , ỹt−1 ) with
Ŷt−1 = (ŷ1 , . . . , ŷt−1 ). Arbitrary combinations of one-step ahead and open-loop
predictions are possible to handle the more general case of intermittent output
data availability.
11
Algorithm 2 Inference
Input: Model set Θ? , production data set X̃t = (x̃1 , . . . , x̃t ), past outputs
Ỹt−1 = (ỹ1 , . . . , ỹt−1 ).
1. Ŝt ← arg minSt
L(St ) +
t−1
X
j=1
`(x̃j , ỹj , θs?j ) + min `(x̃t , y, θs?t ) ;
y∈Yt
2. ytk ← arg miny∈Yt `(x̃t , y, θŝ?t );
Output: Estimated output ŷt and mode sequence Ŝt .
4.2.2
Recursive inference
When Ltrans = 0 problem (24) becomes completely separable and simplifies to
(ŷt , ŝt ) = arg miny,s `(x̃t , y, θs? ) + Lmode (s)
s.t. y ∈ Yt .
(26)
For example, in the case of K-means (7) (L(s) = 0), the estimate obtained
by (26) is
ŝt = arg mins kx̃t − θs? k22 , ŷt = θŝt
that is ŷt is the centroid of the cluster which x̃t is closest to.
When the mode transition loss function Ltrans 6= 0, the simplification in (26)
does not hold anymore. Nonetheless, an incremental version of (24), producing
exactly the same results, can be still derived in the following form
(ŷt , ŝt ) = arg miny,s `(x̃t , y, θs ) + Lt (s)
s.t. y ∈ Yt
(27a)
where Lt : K → R is the arrival cost recursively defined as
Lt (st ) = Lmode (st ) + min `(x̃t−1 , ỹt−1 , θst−1 ) + Lt−1 (st−1 )
st−1
+L
L0 (s0 )
trans
init
= L
(st , st−1 ) ,
(s0 ).
t = 1, 2, . . .
(27b)
(27c)
Clearly, while producing exactly the same results, the formulation in (27) is
much more efficient than Algorithm 2 as the number of computations does not
increase with t.
4.2.3
Smoothing
The same approach described in Section 4.2.1 can be generalized to other inference tasks than one-step ahead or open-loop prediction, such as smoothing.
Assume ỹk is only known at steps k ∈ Tt ⊆ {1, . . . , t}. Steps 1–2 are replaced
12
by
Ŝt
yj
n
P
L(St ) + j∈Tt `(x̃j , ỹj , θs?j )
o
P
+ j∈T¯t minyj ∈Yj `(x̃j , yj , θs?j )
← arg miny∈Yj `(x̃j , y, θŝ?j ), ∀j ∈ T̄t
← arg minSt
where T̄t = {1, . . . , t} \ Tt .
4.2.4
Pure mode estimation
In case we are interested in estimating only the latent mode ŝt given x̃1 , . . . , x̃t ,
ỹ1 , . . . , ỹt−1 and also ỹt , we can keep using (27) by simply changing (27a) to
ŝt = arg mins `(x̃t , ỹt , θs ) + Lt (s).
(28)
This allows reconstructing the mode sequence ŝ1 , . . . , ŝT̃ recursively from the
available data set, which may be useful for example to detect changes in the
relation between the input x̃t and the output ỹt .
4.3
Relation with Hidden Markov Models
Jump models have several common features with hidden Markov models (HMMs)
[23]. First, both models consider the presence of discrete latent states st . While
HMMs assume that the sequence S of such states satisfy the Markov property
p(st |st−1 , . . . , s0 ) = p(st |st−1 )
in jump models the particular form chosen in (13a) for the mode sequence loss
L makes estimating ŝt incrementally as in (27) possible.
Second, in HMMs the observed outputs are such that
p(yt |xt , . . . , x1 , yt−1 , . . . , y1 , st , . . . , s0 ) = p(yt |xt , st ).
Similarly, in jump models ŷt is a unique function of a given pair (xt , st ), as (27a)
becomes
ŷt = arg miny∈Yt `(xt , y, θst ).
Indeed, an HMM is a special case of a jump model. Consider the case in which
the output observation yt is discrete, that is Y = {1, . . . , L}. An HMM is
characterized by the set of discrete probabilities
p(st+1 = i|st = j) = πi,j ,
i, j ∈ K
p(s0 ) = πs0
p(yt = v|st = j) = βj,v , v ∈ Y.
(29a)
(29b)
(29c)
Let us set xt = 1, θs = s, and define the loss function ` as
`(x, y, θs ) = − log(βθs ,y ).
13
(30)
Similarly to (14), by also setting r(θ) = 0, Ltrans (st , st−1 ) = − log πst ,st−1 ,
Linit (s0 ) = − log πs0 , and Lmode (s) = 0, the jump model defined by the inference
rule (24)–(25) returns the mode sequence Ŝt that best matches the observed
sequence Ỹt of outputs and that sets the output ŷt equal to the value v ∈ Y
that maximizes the probability βŝt ,v . An extension to HMMs with continuous
observation densities can be obtained by properly redefining the loss function `
in (30).
In case the probabilities βs,y are not given, but rather must be estimated
from a training data set, we can set instead θs = [βs,1 . . . βs,L ]0 along with the
loss function `
`(y, θs ) = − log(e0y θs )
(31)
where ey is the yth column of the identity matrix of size L. If the initial probability distribution πs0 and the state transition probabilities πi,j are unknown,
they can be estimated by minimizing J(X, Y, Θ, S) in (12) with Ltrans (st , st−1 )
and Linit (s0 ) as in (14c) and (14d), respectively. This implies that the unknown
model parameter Θ should also include πs0 and πi,j , leading to the general case
of having the mode sequence loss L also dependent on Θ as in (23).
The well-known Expectation-Maximization (EM) algorithm [11] determines
the parameters of an HMM by maximizing the log-likelihood
X
LHMM (Θ|X, Y ) = log p(Y |X, Θ) = log
p(Y, S|X, Θ)
(32)
S∈KT +1
with respect to Θ. Instead, as shown by Proposition 1, our approach maximizes
log p(Y, S, Θ|X) with respect to Θ and S.
The case of HMMs in which the observations y are a mode-dependent linear
function of x rather than discrete has been dealt with for example in [13],
under the assumption that such a linear relation between input and output
samples is perturbed by Gaussian noise. This is a special case of our jump
model framework, obtained by setting ` as in (20), Ltrans as in (14c), Linit
as in (14d), Lmode (s) = 0, and r(θ) = 0. The training algorithm described
in [13], however, completely relies on the probabilistic assumptions made about
the normal distribution of output noise and the Markovian nature of mode
transitions.
In conclusion, jump models are more descriptive than HMMs. The sequence
of modes may not be generated by a Markov chain, such as in the case of PWA
models (10) and (11), where the mode st is a deterministic function of xt . In
addition, the loss and mode loss functions can have rather arbitrary shapes. For
example, we may choose `(x, y, θs ) as the Huber function of y − θs0 x for robust
regression, which is still a convex loss.
4.4
Selecting the mode sequence loss from data
Selecting the right mode sequence loss L may not be obvious and require several
attempts that involve fitting and cross-validation. A simple approach to choose
L directly from the training data is to update the mode loss function L after
14
executing Algorithm 1 based on the best sequence S ? found so far, and run
Algorithm 1 again, executing the algorithm N times in total.
Assuming Lmode = 0, we update Ltrans , Linit from one run of Algorithm 1
to another by setting
#{t ∈ {1, . . . , T } : s?t−1 = j}
T
#{t ∈ {1, . . . , T } : s?t = i, s?t−1 = j}
µij ←
T
µij
trans
L
(i, j) ← −τi log
, i, j = 1, . . . , K
µj
#{t ∈ {0, . . . , T } : s?t = j}
µ0j ←
T
Linit (j) ← −τ0 log µ0j
µj
←
(33a)
(33b)
(33c)
(33d)
(33e)
where # denotes the cardinality (number of elements) of a set, S ? = (s?0 , . . . , s?T ),
and we set
PK
trans
(i, j)
j=1 L
, i = 1, . . . , K
τi ← P
µij
K
j=1 log µj
PK
init
j=1 L0 (j)
τ0 ← PK
0
j=1 log µj
to preserve the relative weight of L with respect to the loss ` and regularization r.
Choosing L as in (33) is motivated by the statistical interpretation (14c)–(14d)
and used routinely for estimating state probabilities in HMMs [23]. Clearly, (33)
are well defined only if µij , µj , µ0j > 0 forall i, j = 1, . . . , K. If the latter condition is not satisfied, one may consider adding the following Laplace smoothing
µj
←
µij
←
µ0j
←
1 + #{t ∈ {1, . . . , T } : s?t−1 = j}
T +K
1 + #{t ∈ {1, . . . , T } : s?t = i, s?t−1 = j}
T + K2
1 + #{t ∈ {0, . . . , T } : s?t = j}
T +K
(34a)
(34b)
(34c)
when estimating µj , µij and µ0j .
Estimating L as in (33) is especially useful for improving the quality of
inference, both when using (25) or (27).
5
Examples
We test the algorithms proposed in the previous sections on various problems
of regression and classification using jump models. As the DP computation in
15
Step 1.2 of Algorithm 1 also provides the optimal cost V k , J(X, Y, Θk , S k ),
when running the tests we replace the termination criterion in Step 2 with
V k−1 − V k ≤ V
(35)
where V is a small tolerance.
All tests were run on a MacBook Pro 3 GHz-Intel i7 in MATLAB R2016b.
The test code is available for download at http://cse.lab.imtlucca.it/
~bemporad/jump_models/.
5.1
Jump linear model regression
We consider a dataset of T = 10000 training data and T̃ = 10000 production
data generated by the following jump linear model with K = 3 modes
yt = θst xt + ζt
with yt ∈ R, xt ∈ R20 , ζt ∼ N (0, σy2 ). The coefficients of the parameter vectors
θi are randomly selected from the normal distribution N (0, 1). The true mode
st has probability π = 5% of being different from st−1 , starting from s0 = 1.
As loss functions we consider
`(xt , yt , θst )
=
r(θk )
=
Ltrans (st , st−1 )
kyt − θs0 t xt k22
10−5 kθk k22
−τ log(1 − (K − 1)π) if st = st−1
=
−τ log π
if st =
6 st−1
Linit (s0 )
=
0
Lmode (st )
=
0
where τ is treated as a hyper-parameter to be tuned. The tolerance V = 10−8
is used in (35) as a stopping criterion for Algorithm 1, which is executed N = 5
times from different random initial guesses. Each execution is limited to kmax =
1000 iterations.
We run Algorithm 1 on the training data for different magnitudes σy of output noise and values of the hyper-parameter τ . The resulting model coefficients
Θ? are then used in Algorithm 2 for recursive inference on the production data.
For assessing the quality of inference we use the true loss
L̃true =
T̃
1X
kŷ − ỹk22 .
T̃ i=1
In addition, assuming the latent modes s̃t are available only for validation purposes, we consider the following mode-mismatch figure
T̃
`true
=
s
100 X
δŝt ,s̃t
T̃ t=1
16
(36)
8.5
20
8
19
18
7.5
17
% mode mismatch
7
fit
6.5
6
5.5
16
15
14
13
5
12
4.5
4
10 -6
11
10 -4
10 -2
10
10 -6
10 0
10 -4
10 -2
10 0
Figure 1: Jump linear model fit and validation: true loss L̃true (left) and mode
(right), optimal theoretical value τ ? = 2σy2 (dashed line)
mismatch `true
s
where δi,j is the Kronecker delta function. The results are summarized in Figure 5.1.
By recalling (20) and (14c), in order to minimize − log p(yt |xt , θ)−log πst ,st−1
one should set `(xt , yt , θst ) = 2σ1 2 and Ltrans (st , st−1 ) = − log πst ,st−1 , or equivy
alently `(xt , yt , θst ) = 1, Ltrans (st , st−1 ) = −τ ? log πst ,st−1 with τ ? = 2σy2 . Figure 5.1 also reports the value of τ ? (dashed line) corresponding to different
values of σy . As expected, the best value for τ obtained by cross validation,
corresponding to the minimum of the plotted curves, corresponds to the theoretical one τ ∗ that would be obtained if σy were known. For large values of τ
the percentage of mode mismatch becomes close to K−1
K ≈ 66% (not shown in
the figure), that is the value one gets when the mode ŝt is assigned randomly.
The average CPU time for executing Algorithm 1 is 342 ms, for Algorithm 2
is 0.89 µs per data point.
5.2
Jump binary classification
We consider T = 10000 training data and T̃ = 10000 production data generated
by the following jump linear model with K = 3 modes
yt = sign(θst xt + ζt )
17
with
θ1
θ2
−1
1.1812
−0.7585
−1.1096
θ3 =
−0.8456
−0.5727
−0.5587
0.1784
−1
−0.5587
0.1784
−0.1969
0.5864
0.8759
−0.2428
0.1668
−1
0.8003
−1.5094
0.8759
−0.2428
0.6037
1.7813
1.7737
and yt ∈ {−1, 1}, xt ∈ R8 , ζt ∼ N (0, σy2 ), σy = 0.1. The true mode st changes
every 500 samples during the generation of the data, covering all modes.
We want to train a binary classifier defined by the following losses
`(xt , yt , θst )
=
max(1 − yt θs0 t xt , 0)
r(θk )
=
10−5 kθk k22
(st , st−1 )
=
τ (1 − δi,j )
=
0
=
0.
trans
L
L
init
(s0 )
mode
L
(st )
Figure 5.2 shows the results obtained for different values of the hyper-parameter
τ . We consider the mismatch between the true labels yt and the estimated
labels yt? = sign((θs∗?t )0 xt ) returned by Algorithm 1 on the training data, and
also between the true labels ỹt and the labels ŷt = sign((θŝ∗t )0 x̃t ) returned by
Algorithm 2 on the production data. In addition, we consider the detection
of model changes, comparing the true modes st , s̃t and their corresponding
estimates s∗t , ŝt . Good values for τ are in the range 1 ÷ 10, for which model
changes are correctly detected on both training and production data.
The CPU time for executing Algorithm 1 ranges between 4.24 and 80.76 s,
with Step 1 computed using the QP solver of GUROBI 7.02 [15]. Algorithm 2
takes an average of 0.57 µs per data point for inference.
5.3
Markov jump linear dynamical system
We consider the Markov jump linear dynamical system with K = 4 modes
xt+1 = Ast xt + Bst ut + ζt
where xt , ζt ∈ R8 , ut ∈ R2 , ζtj ∼ N (0, σy ) for all j = 1, . . . , 8, the matrix
pairs (Ai , Bi ) are random stable systems for all i = 1, . . . , K. The modes st are
randomly generated according to an (unknown) transition probability matrix
Π ∈ R4×4 . The goal is to estimate the system matrices (Ai , Bi ), i = 1, . . . , K,
and the transition probability Π from T = 50000 data pairs (xt , ut ) available
for training, and validate the results on T̃ = 50000 new samples.
Algorithm 1 is executed N = 5 times on the training data with loss function
kxt+1 − Ast xt − Bst ut k22 , uniform mode transition loss Ltrans (i, j) = τ , and zero
18
90
50
45
training
inference
training
inference
80
40
70
% mode mismatch
% label mismatch
35
30
25
20
60
50
40
30
15
20
10
10
5
0
10 -2
10 -1
10 0
10 1
0
10 -2
10 2
10 -1
10 0
10 1
10 2
Figure 2: Jump binary classifier: percentage of misclassified labels (left) and
mode mismatch (right) on training and production data
losses Linit , Lmode . Note that, since the output sample yt = xt+1 is multidimensional, we cannot train a model for each component of y independently, as
they are linked by the common mode st .
After training and before performing inference via (27), the transition probability matrix Π̂ is reconstructed using (33) on the estimated mode sequence S ?
returned by Algorithm 1.
The results are reported in Figure 3. The coefficients of the models (Ai , Bi )
are estimated with an error of 10−8 (σy = 0), 10−3 (σy = 0.01), and 10−2 (σy =
.05), respectively, while the transition probability matrix with error kΠ − Π̂k2 of
0.01 for all values of σy = 0. The average CPU time for executing Algorithm 1 is
68 ms (corresponding to about 15 iterations), while Algorithm 2 takes of 0.57 µs
per data point on average for inference.
5.4
Experimental example: PWA dynamical model
We consider the problem of modeling the dynamics of a placement process of
electronic components in a pick-and-place machine described in [17]. The process consists of a mounting head carrying the electronic component which is
placed on a printed circuit board, and then released. This process is characterized by two main operating modes, the free and the impact mode. In free
mode the machine carries the electronic component in an unconstrained environment, i.e., without being in contact with the circuit board. In impact mode the
19
mounting head moves in contact with the circuit board. Because of its switching
behaviour, this process has been used as a benchmark to assess the performance
of several identification algorithms for hybrid dynamical systems [3, 18, 20].
A data record over an interval of 15 s is gathered from an experimental bench
(see [17] for details), with a sampling frequency of 2 kHz. We denote by u the
voltage applied to the motor driving the mounting head and by y the vertical
position of the mounting head. The data record is split in two disjoint subsets: a
training set with T = 24000 samples, which consist of the observations gathered
in the first 12 s of the experiments, and a test set with T̃ = 6000 samples, which
consist of the observations gathered in the last 3 s.
We want to fit a PWA model (9)–(10) with K = 2 discrete modes. Each
0
0
regression model is given by yt = θy,s
[ x1t ], where xt = [yt−1 yt−2 ut−1 ut−2 ] .
t
A piecewise linear separation function ϑ defined in (9) is used to describe the
partition of X = R4 .
Algorithm 1 is executed N = 5 times on the first 22000 samples of the
training set with loss function `(xt , yt , θst ) as in (10) and mode sequence loss
L = 0. The remaining 2000 samples are used to tune the hyper-parameter
ρ, whose best value is ρ = 1.23 · 10−5 . The average CPU time for executing
Algorithm 1 for a fixed value of ρ is 200 ms.
Figure 4 shows the outputs ỹt collected from the production dataset, the
open-loop prediction ŷt of the output reconstructed by feeding the same inputs ũt to the estimated PWA model, and
modes
thersequence of estimated
ŝt . The resulting best fit rate BFR = 100 1 −
PT̃
Pt=1
T̃
t=1
kỹt −ŷt k2
kỹt −ȳk2
% is equal to
81%, where ȳ denotes the average of the outputs ỹ1 , . . . , ỹT̃ . It is interesting to
note that the reconstructed mode correctly captures the two different operating
conditions of the process (mode 1 = free mode, mode 2 = impact mode).
6
Conclusions
We have presented a new framework for fitting a jump model to a temporal
sequence of data. A main strength of the approach is its versatility in describing
a large class of parametric models, as the shape of the model and the way it
jumps depends on the shape of the loss functions used for fitting the model
parameters and for inference. Another strength of the approach is its numerical
efficiency, due to using a simple coordinate-descent optimization algorithm for
fitting model parameters and a recursive formulation for inferring outputs and
latent modes.
Future research will address an incremental version of the fitting algorithm,
so to update models and infer output/mode pairs when data are streaming
on-line.
20
50
9
45
8
40
35
% mode mismatch
fit
7
6
5
30
25
20
15
4
10
3
10 -2
10 -1
10 0
10 1
5
10 -2
10 2
10 -1
10 0
10 1
10 2
Figure 3: Markov jump linear dynamical system: true loss L̃true (left) and mode
(right).
mismatch `true
s
References
[1] L.E. Baum, T. Petrie, G. Soules, and N. Weiss. A maximization technique
occurring in the statistical analysis of probabilistic functions of Markov
chains. The Annals of Mathematical Statistics, 41(1):164–171, 1970.
[2] R. Bellman. Dynamic Programming. Princeton University Press, Princeton,
NJ, USA, 1957.
[3] A. Bemporad, A. Garulli, S. Paoletti, and A. Vicino. A bounded-error
approach to piecewise affine system identification. IEEE Transactions on
Automatic Control, 50(10):1567–1580, 2005.
[4] A. Bemporad and N. Giorgetti. Logic-based methods for optimal control of
hybrid systems. IEEE Transaction on Automatic Control, 51(6):963–976,
2006.
[5] K.P. Bennett and O.L. Mangasarian. Multicategory discrimination via
linear programming. Optimization Methods and Software, 3:27–39, 1994.
[6] S. Boyd and L. Vandenberghe. Convex Optimization. Cambridge University
Press, New York, NY, USA, 2004. http://www.stanford.edu/~boyd/
cvxbook.html.
21
25
output
20
15
10
5
0
0
0.5
1
0
0.5
1
1.5
2
2.5
3
1.5
2
2.5
3
mode
2
1
time (seconds)
Figure 4: Pick-and-place machine: simulated and actual output (top) and estimated mode sequence (bottom).
22
[7] V. Breschi, A. Bemporad, and D. Piga. Identification of hybrid and linear
parameter varying models via recursive piecewise affine regression and discrimination. In European Control Conference, pages 2632–2637, Aalborg,
Denmark, 2016.
[8] V. Breschi, D. Piga, and A. Bemporad. Piecewise affine regression via recursive multiple least squares and multicategory discrimination. Automatica,
73:155–162, November 2016.
[9] A.B. Chan and N. Vasconcelos. Modeling, clustering, and segmenting video
with mixtures of dynamic textures. IEEE transactions on pattern analysis
and machine intelligence, 30(5):909–926, 2008.
[10] O. L. V. Costa, M. D. Fragoso, and R. P. Marques. Discrete-time Markov
jump linear systems. Springer Science & Business Media, 2006.
[11] A.P. Dempster, N.M. Laird, and D.B. Rubin. Maximum likelihood from
incomplete data via the EM algorithm. Journal of the Royal Statistical
Society, Series B., 39(1):1–38, 1977.
[12] G. Ferrari-Trecate, M. Muselli, D. Liberati, and M. Morari. A clustering
technique for the identification of piecewise affine systems. Automatica,
39(2):205–217, 2003.
[13] M. Fridman. Hidden Markov model regression. Technical report, Institute
of Mathematics, University of Minnesota, Minneapolis, MN, 1994.
[14] M. Guidolin. Markov switching models in empirical finance. In Missing
Data Methods: Time-Series Methods and Applications, pages 1–86. Emerald Group Publishing Limited, 2011.
[15] Gurobi Optimization, Inc. Gurobi Optimizer Reference Manual, 2017.
[16] T. Hastie, R. Tibshirani, and J. Friedman. The elements of statistical
learning. Springer, New York, 2nd edition, 2009.
[17] A.L. Juloski, W.P.M.H. Heemels, and G. Ferrari-Trecate. Data-based hybrid modelling of the component placement process in pick-and-place machines. Control Engineering Practice, 12(10):1241–1252, 2004.
[18] A.L. Juloski, W.P.M.H. Heemels, G. Ferrari-Trecate, R.Vidal, S. Paoletti,
and J.H.G. Niessen. Comparison of four procedures for the identification of
hybrid systems. Lecture Notes in Computer Science, 3414:354–369, 2005.
[19] S.M. Oh, J.M Rehg, T. Balch, and F. Dellaert. Learning and inferring motion patterns using parametric segmental switching linear dynamic systems.
International Journal of Computer Vision, 77(1):103–124, 2008.
[20] H. Ohlsson and L. Ljung. Identification of switched linear regression models
using sum-of-norms regularization. Automatica, 49(4):1045–1050, 2013.
23
[21] N. Ozay, Mario M. Sznaier, and C. Lagoa. Model (in)validation of switched
arx systems with unknown switches and its application to activity monitoring. In 49th IEEE Conference on Decision and Control, pages 7624–7630,
Atlanta, GA, 2010.
[22] V. Pavlovic, J.M. Rehg, and J. MacCormick. Learning switching linear
models of human motion. In Advances in neural information processing
systems, pages 981–987, 2001.
[23] L.R. Rabiner. A tutorial on hidden Markov models and selected applications in speech recognition. Proceedings of the IEEE, 77(2):257–286, 1989.
[24] B. Schuller, M. Wöllmer, T. Moosmayr, G. Ruske, and G. Rigoll. Switching
linear dynamic models for noise robust in-car speech recognition. Pattern
Recognition, pages 244–253, 2008.
[25] A. Timmermann. Markov Switching Models in Finance, volume 4. John
Wiley & Sons, Ltd, 2015.
24
| 3cs.SY
|
Global Pose Estimation with an Attention-based Recurrent Network
arXiv:1802.06857v1 [cs.CV] 19 Feb 2018
Emilio Parisotto∗,1,2 , Devendra Singh Chaplot∗,1,2 , Jian Zhang1 , Ruslan Salakhutdinov1,2
∗
Equal Contribution.
1
Apple Inc., 1 Infinite Loop, Cupertino, CA 95014, USA.
2
Carnegie Mellon University, 5000 Forbes Ave, Pittsburgh, PA 15213, USA.
{eparisot,dchaplot}@cs.cmu.edu
{jianz,rsalakhutdinov}@apple.com
Abstract
The ability for an agent to localize itself within an
environment is crucial for many real-world applications.
For unknown environments, Simultaneous Localization and
Mapping (SLAM) enables incremental and concurrent
building of and localizing within a map. We present a new,
differentiable architecture, Neural Graph Optimizer, progressing towards a complete neural network solution for
SLAM by designing a system composed of a local pose
estimation model, a novel pose selection module, and a
novel graph optimization process. The entire architecture
is trained in an end-to-end fashion, enabling the network
to automatically learn domain-specific features relevant to
the visual odometry and avoid the involved process of feature engineering. We demonstrate the effectiveness of our
system on a simulated 2D maze and the 3D ViZ-Doom environment.
Figure 1. Components of the proposed model along with sample
input, output and ground truth. The Local Pose Estimation model
predicts the relative pose change between consecutive observations and Neural Graph Optimization model jointly optimizes the
predictions of the Local Pose Estimation model to predict global
pose changes. The local pose estimates, global pose estimates,
and ground truth trajectory are shown in green, orange and blue,
respectively.
1. Introduction
The ability for an agent to localize itself within an environment is a crucial prerequisite for many real-world applications, such as household robots [42], autonomous drones
[14], augmented and virtual reality applications, and video
game AI [36]. In most cases, the main challenge for an
agent localizing itself is that, the agent is not provided
with a map of the environment and therefore the agent
must simultaneously map the environment and localize itself within the incomplete map it has produced. A wide variety of algorithms to solve this Simultaneous Localization
and Mapping (SLAM) task have been developed over a long
history [42, 4], with modern methods achieving impressive
accuracy and real-time performance [33, 25, 34, 12]. These
methods still have several shortcomings, owing mainly to
the hand-engineered features, dense matching, and heuristics used in the design of these algorithms. For example,
most methods are brittle in certain scenarios, such as varying lighting conditions (e.g. changing time of day), differ-
ent weather conditions or seasons [40], repetitive structures,
textureless objects, extremely large viewpoint changes, dynamic elements within the environment, and faulty sensor
calibration [4]. Because these situations are common in
real-world scenarios, robust applications of those systems
are difficult.
In this paper, we develop a method which can be made
more robust to the common situations where previous
SLAM algorithms typically degrade. To do this, we formulate a novel neural network architecture called “Neural
Graph Optimizer”. Neural Graph Optimizer consists of differentiable analogues of the common types of subsystems
used in modern SLAM algorithms, such as a local pose
estimation model, a pose selection module (key frame selection, essential graph), and a graph optimization process.
Because each component in the system is differentiable, the
entire architecture can be trained in an end-to-end fashion,
1
enabling the network to learn invariances to the types of
scenarios observed during training.
To demonstrate the ability of our method to learn pose
estimation, we use trajectories sampled from several simulated environments. The first environment is a 2D maze
where the agent has a single-pixel row-scan as input. We
then scale the model up to 3D mazes based on the ViZDoom
environment [21], where the agent receives an image of the
first-person view of the world as input.
ods focus on dense recontruction of the scene, for instance
[47] builds dense globally consistent surfel-based maps of
room scale environments explored using an RGB-D camera, without pose graph optimisation, while KinectFusion
[35] obtains depth measurements directly using active sensors and fuse them over time to recover high-quality surface maps. These approaches still suffer from strict calibration and synchronization requirements, and the data association modules require extensive parameter tuning in order
to work correctly for a given scenario.
In light of the limitations of feature-based and direct approaches, deep networks are proposed to learn suitable feature representations that are robust against motion blur, occlusions, dynamic scenes, illumination, texture, and viewpoint changes. They have been successfully applied to several related multiview vision problems, including learning
optical flow [11], depth [28], homography between frame
pairs [9], and localization [5] and re-localization problems.
Recent work includes re-formulating the localization
problem as a classification task [46], a regression task
[22, 44], end-to-end trainable filtering [18], and differentiable RANSAC [3]. More specifically, PlaNet [46] formulates localization as a classification problem, predicting the
corresponding tile from a set of tiles subdividing Earth surface for a given image, thus providing the approximate position from which a photo was taken. PoseNet [22] formulates 6-DoF pose estimation as a regression problem. One
drawback of the PoseNet approach is its relative inaccuracy,
compared to state-of-the-art SIFT methods. Similarly, [30]
fine-tunes a pretrained classification network to estimate the
relative pose between two cameras. To improve its performance, [44] added Long-Short Term Memory (LSTM) units
to the fully-connected layers output, to perform structured
dimensionality reduction, choosing the most useful feature
correlations for the task of pose estimation. From a different angle, DSAC [3] proposes a differentiable RANSAC
so that a matching function that optimizes pose quality can
be learned. These approaches are not robust to repeated
structure or similar looking scenes, as they ignore the sequential and graphical nature of the problem. Addressing
this limitation, work in [6] fused additional sequential inertial measurement with visual odometry. SemanticFusion
[29] combines convolutional neural networks (CNNs) and
a dense ElasticFusion [47]. However, classic feature-based
methods still outperform CNN-based methods published to
date in terms of accuracies.
Recently, there has been an increasing interest in combining navigation and plannning in an end-to-end deep reinforcement learning (DRL) framework. The efforts to date
can be divided into two categories depending on the presence of external memory in the architecture or not. Targetdriven visual navigation takes a visual observation and an
image of the target [50] or range findings [41] as input, and
2. Related Work
SLAM is a process in which an agent needs to localize itself in an unknown environment and build a map of
this environment at the same time, with uncertainties in
both its motions and observations. SLAM has evolved
from filter-based to graph-based (optimization-based) approaches. Some EKF-based systems have demonstrated
state-of-the-art performance, such as the Multi-State Constraint Kalman Filter [33], the VIN [23], and the system of
Hesch et al. [19]. Those methods, even though efficient,
heavily depend on linearization and Gaussian assumptions,
and thus under-perform their optimization-based counterparts, such as OK-VIS [27], ORB-SLAM [34], and LSDSLAM [12].
Graph-based SLAM typically includes two main components: the front-end and the back-end. The front-end extracts relevant information (e.g. salient features) from the
sensor data and associates each measurement to a specific
map feature, while the back-end performs graph optimization on a graph of abstracted data produced by the front-end.
Graph-based SLAM can be categorized either as featurebased or direct methods depending on the type of frontend. Feature-based methods rely on local features (e.g.
SIFT, SURF, FAST, ORB, etc.) for pose estimation. For
example, ORB-SLAM [34] performs data association and
camera relocalization with ORB features and DBoW2 [16].
RANSAC [13] is commonly used for geometric verification
and outlier rejection, and there are also prioritized feature
matching approaches [39]. However, hand-engineered feature detector and descriptors are not robust to motion blur,
illumination changes, or strong viewpoint changes, any of
which can cause localization to fail.
To avoid some of the aforementioned drawbacks of
feature-based approaches, direct methods, such as LSDSLAM [12], utilize extensive photometric information from
the images to determine the pose, by minimizing the photometric error between corresponding pixels. This approach
is in contrast to feature-based methods, which minimize the
reprojection error. However, such methods are usually not
applicable to wide baseline settings [4] during large viewpoint changes. Recent work in [14] [15] combines feature
and direct methods by minimizing the photometric error of
features lying on intensity corners and edges. Some meth2
plans goal seeking actions in a 3D indoor simulated environment as the output.
In simulated environments, [31] uses stacked LSTM in a
goal-driven RL problem with auxilary tasks of depth prediction and loop-closure classification, while [48] added successor features to ease transfer from previously mastered
navigation tasks to new ones. Work in [1] augmented DRL
with Faster-RCNN for object detection and SLAM (ORBSLAM2) for pose estimation; observing images and depth
from VizDoom, they built semantic maps with 3D reconstruction and bounding boxes as input to a RL policy.
To deal with the limited memory of standard recurrent
architures (such as LSTM) more structured external memories have been developed to take the spatial relations of
memories into account. [17] assumes known ego-motion
and constructs a metric egocentric multi-scale belief map
(top-down-view latent representation of free space) of the
world with a 2D spatial memory, upon which RL plans a
sequence of actions towards goals in the environment with
a value iteration network. Neural Map in [36] is a writable
structured 2D external memory map for an agent to learn
to navigate within 2D and 3D maze environments. These
works all assume precise egomotion and thus perfect localization, a prerequisite that can rarely be met in real-world
scenarios. Relaxing this assumption and resembling traditional occupancy grid SLAM, Neural SLAM [49] uses an
occupancy-grid-like memory map, assuming only an initial
pose is provided, and updates the pose beliefs and grid map
using end-to-end DRL.
One of key ingredient for the success of graph-based
SLAM is the back-end optimization. The back-end builds
the pose graph, in which two pose nodes share an edge
if an odometry measurement is available between them,
while a landmark and a robot-pose node share an edge if
the landmark was observed from the corresponding robot
pose. In pose graph optimization, the variables to be estimated are poses sampled along the trajectory of the robot,
and each factor imposes a constraint on a pair of poses.
Modern SLAM solvers exploit the sparse nature of the underlying factor graph and apply iterative linearization and
optimization methods (e.g. nonlinear least squares via the
Gauss-Newton or Levenberg-Marquardt algorithm). Several such solvers achieve excellent performance, for example, g2o [25], TSAM [8], Ceres, iSAM [20], SLAM++
[38], and recently [2] for optimization with semantic data
association. The SLAM back-end offers a natural defense
against data association and perceptual aliasing errors from
the front-end, where similarly looking scenes, corresponding to distinct locations in the environment, would deceive
place recognition. However, they depend heavily on linearization of the sensing and motion models, and require
good initial guesses. Current systems can be easily induced
to fail when either the motion of the robot or the environ-
ment are too challenging (e.g. fast robot dynamics or highly
dynamic environments) [4].
In this work we formulate a complete end-to-end trainable solution to the graph-based SLAM problem. We
present a novel architecture that combines a CNN-based
local front-end and an attention-based differentiable backend. We learn effective features automatically and perform
implicit loop closure by designing an additional differentiable Neural Graph Optimizer to perform global optimization over entire pose trajectories and correct errors accumulated by the local estimation model.
3. Method
The Neural Graph Optimizer architecture is split into
distinct differentiable components. Similar to many of the
previous methods, we split the process into local adjustments between temporally adjacent frames combined with a
global optimization procedure which distributes error over
the entire observed trajectory. As will be shown in the experiments, the global graph optimization procedure is critical to removing drift (the accumulation of small errors over
long trajectories). The graph optimization procedure does
this by learning to do loop closures, recognizing when the
agent has revisted the same location, and enforcing a constraint that those poses should be nearly equal. The local model is crucial for providing a good starting point for
the global optimization. It does this by estimating relative
transformations between two temporally adjacent frames.
By accumulating transformations from the start of the trajectory to the end, we can use this model to get the initial
pose estimate within the global frame.
The complete model architecture is shown in Fig. 2. We
will describe relative poses as ∆P = (∆p1 , . . . , ∆pT ) with
the first pose set as the origin, i.e. ∆p1 is the transformation
from origin to pose 1, ∆p2 is the transformation from pose 1
to pose 2, and so on. These relative poses can be transformed into a global frame of reference by accumulating the
relative pose changes along the trajectory, i.e. p1 = ∆p1 I,
p2 = ∆p2 ∆p1 I, and so on. These global poses will be
refered to as P = (p1 , . . . , pT ). There exists a differential function r2g = g2r−1 such that P = r2g(∆P ) and
∆P = g2r(P ). Each component is described in more detail in the next sections.
3.1. Local Pose Estimation Network
The Local Pose Estimation network learns to predict the
relative pose change between two consecutive frames. From
two consecutive observations, where each observation is,
for example, an RGB frame, this component predicts the xcoordinate, y-coordinate, and orientation (∆x, ∆y and ∆θ)
of the second frame with respect to the first frame. It can
also optionally take in side information, such as the action
taken by the agent between the two frames. The architecture
3
Figure 2. The architecture of the proposed model, showing the Local Pose Estimation, the Pose Aggregation, and the Neural Graph
Optimization modules.
In addition to the pose features being aggregated into
meta-pose features by the temporal convolution, we also
compose all the local pose transformations that were predicted by the Local Pose Estimation model. This composition gives us an initial global pose estimate for each of the
meta-poses. The combined meta-features and meta-poses
are then passed onto the Neural Graph Optimization layer
for the final global pose adjustments, as shown in Fig. 2.
of the Local Pose Estimation network is shown in Fig. 3.
Both frames are stacked and passed through a series of convolutional layers. The output of the convolutional layers
is flattened and passed to two fully-connected layers that
predict the translational and rotational pose change respectively.
Some of the recent work showed that optical flow is useful in predicting frame-to-frame ego-motion [7]. The architecture of the Local Pose Estimation network is inspired by
the architecture of Flownet [11] which predicts the optical
flow between two frames. The convolutional layers in the
Local Pose Estimation network are identical to the convolutional layers in Flownet. Prior work on visual odometry
and visual inertial odometry has also used the convolutional
layer architecture of Flownet [6, 45].
3.3. Neural Graph Optimization
The final component of our system is the “Neural Graph
Optimizer”. This submodule aggregates information over
the entire pose trajectory with the goal of redistributing error to minimize drift. The Neural Graph Optimizer model
is a neural analogue of the global optimization procedures
commonly used in traditional state-of-the-art SLAM packages, such the g2o framework [25]. We define the Neural
Graph Optimizer as a recurrent network submodule which
takes as input sequential pose features and outputs a refined
estimate of these poses.
In more detail, the Neural Graph Optimizer takes as input some initial T relative pose estimates (i.e. the aggregated output of the local
pose estimation network) ∆P(0) =
(0)
(0)
∆p1 , . . . , ∆pT
and produces two outputs for each
pose:
(1)
(1)
∇P(1) = ∇p1 , . . . , ∇pT , and
(1)
(1)
β (1) = β1 , . . . , βT .
3.2. Pose Aggregation
The next step of the architecture is a Pose Aggregation
network which takes in a large number of low-level poses
and pose features (up to 2000 for 2D, 1000 for 3D VizDoom
environment) and reduces them into a smaller number of
more temporally distant “meta-poses” and “meta-pose features” (around 250 for 2D, 125 for 3D VizDoom). These resulting meta-poses and meta-pose features are then passed
to the Neural Graph Optimization procedure.
For pose feature aggregation, we utilize a deep temporal convolutional network with several alternating layers of
(kernel size 3, stride 1, padding 1) dimension-preserving
convolutions and (kernel size 2, stride 2, padding 0)
dimension-reducing max pooling (where each max pooling
operation halves the sequence size). The number of times
we halve the sequence length is a hyperparameter. Instead
of temporal convolutions, we could have utilized recurrent
networks, but we decided to focus on convolutions for computational and memory-efficiency reasons.
New pose estimates are then constructed by performing an
iterative update:
(1)
(0)
(1)
(1)
∆pi = ∆pi + βi ∇pi .
The Neural Graph Optimizer procedure can then be rerun
(1)
(1)
on the new pose estimates ∆P(1) = (∆p1 , . . . , ∆pT )
4
Figure 3. The architecture of the Local Pose Estimation network. The architecture of the convolutional layers is adapted from the architecture of the Flownet [11].
(2)
(2)
to produce ∆P(2) = (∆p1 , . . . , ∆pT ), and so on. The
process is repeated until some pre-specified number of iterations M has taken place. We then transform the refined relative pose estimates into the final global output:
P(M ) = r2g(∆P(M ) ).
The specific architecture of the Neural Graph Optimizer
is based on two priors that are intuitively useful for pose
optimization. The first prior is the notion that poses that are
temporally adjacent should have similar outputs, while the
second prior is that visually similar but temporally disparate
poses should also have similar outputs since this provides a
hint that a place has been revisited, thereby potentially enabling a loop closure-like correction of drift. We express
these priors by using two architectural systems in the Neural
Graph Optimizer. The first is a Transformer-like [43] attention phase where information is propagated over the entire
sequence, and the second is a convolutional phase where
local temporal information is aggregated.
3.3.1
where the superscripts (i − 1) were omitted for clarity of notation. This produces a sequence of attention
(i−1)
(i−1)
vectors (a1
, . . . , aT
), which are passed along with
(i−1)
(i−1)
(f1
, . . . , fT
) to the next “Optimization” phase.
3.3.2
The optimization phase aggregates local temporal information by passing the pose features through several temporal
convolutions and is responsible for producing the iterative
(i)
(i)
(i)
(i)
adjustments: {∇p1 , . . . , ∇pT } and {β1 , . . . , βT }. The
optimization phase proceeds as follows: First, the attention
and feature vectors are concatenated into a new sequence of
features:
"
#
"
#!
(i−1)
to produce the current iteration’s adjustments (∇P(i) and
β (i) ) as well as the feature layer for the next iteration of the
process (F(i) ).
For our experiments, we use 9 layers of convolutions
with filter size 3 and ReLU activations. While temporal
convolutions have a limited receptive field which provides
a hard upper limit on how far they can transmit information
across time, we found that in practice they worked better
than using a bidirectional LSTMs.
3.3.3
Ctu = hqt , fu i,
Ctu
αtu = PT
,
v=1 Ctv
αtu
.
These features are then passed through several layers of 1D
convolutions hl and activations σl :
(i)
#! !!
"
# "
F
(i−1)
(i−1)
f
f
1
T
∇P(i) = σL hL ... h1
...
(i−1)
(i−1) ...
a
a
(i)
1
T
β
Suppose there is a meta-pose sequence of T steps, processed by the pose aggregation network into an initial set
(0)
(0)
of features at each time step: F(0) = (f1 , . . . , fT ). The
attention phase computes, for each pose, a soft-attention
operation over the entire trajectory. This attention operation allows each pose to query information over long time
spans. The attention phase takes as input the pose fea(i−1)
(i−1)
ture sequence (f1
, . . . , fT
) and produces for each
(i−1)
(i−1)
time step a query vector: (q1
, . . . , qT
) using a fully(i−1)
connected layer. Then, for each query vector qt
, a softattention operation is carried out to produce an attention
(i−1)
vector at
as follows:
T
X
(i−1)
f1
fT
(i−1) , . . . ,
(i−1)
a1
aT
Attention Phase
at =
Optimization
Induced Attention Graph
We now provide some intuition on why the attention phase
enables higher performance than only using the optimization phase, or running all pose features through bidirectional LSTMs. We can see that during the attention phase,
some similarity graph C is constructed such that each element Ctu is the inner product between the query vector qt
fu ,
v=1
5
300
300
250
250
200
200
150
groundtruth
predicted (opt)
predicted (unopt)
100
150
100
50
50
0
−100
0
100
200
300
300
250
250
100
150
100
200
300
100
200
300
groundtruth
predicted (opt)
predicted (unopt)
50
0
0
0
RMSE
17.80
10.21
3.16
100
200
groundtruth
predicted (opt)
predicted (unopt)
150
50
Results on the 2D Environment
Model
Only Local Estimation
Global Estimation - 1 Attend-Opt iteration
Global Estimation - 5 Attend-Opt iterations
0
300
200
Figure 4. Left: A screenshot of the 2D environment based on
Box2D. Right: A bird’s eye view of the 3D environment based
on the Doom game engine.
groundtruth
predicted (opt)
predicted (unopt)
0
100
200
300
0
Figure 5. Images visually demonstrating the effect on pose estimates of adding the Neural Graph Optimizer module on top of
the local pose estimation model in the 2D environment. We can
see that the global optimization procedure greatly reduces drift.
These figures were generated with the 5 iteration Neural Graph
Optimization model.
Table 1. Results for different Neural Graph Optimizer architectures and hyperparameters, in terms of test set Global RMSE. We
can see that the addition of the global optimization procedure reduces the loss by more than 80% as compared to solely using the
local pose model.
rays uniformly in front of itself with an effective field of
view of 300◦ . The observation of the agent includes the
RGB values as well as the depth of the points where these
rays hit a wall. An example of the 2D environment is shown
in Fig. 4. Each cell in the maze has a random color. The
agent can take one of three discrete actions at every time
step: move-forward, turn-left, or turn-right. These actions
result in translational acceleration if the action is moveforward or angular acceleration if the action is turn-left or
turn-right. Data is collected by visiting four different corners on the maze using Dijkstra’s algorithm [10].
and the pose feature vector fu . Therefore C represents a
similarity matrix between the queries and pose features, and
those with very similar features will thus have high information bandwidth through the attention operator because
the attention weight αtu will be near 1 for highly similar
query and pose features, and near 0 otherwise. The attention operation is thus inducing a connectivity graph between
poses with highly similar features. This therefore resembles
a soft, differentiable analogue of the pose graph constructed
in SLAM algorithms such as ORB-SLAM [34].
For this environment, the training data is generated by
worker threads in parallel with the model training and each
training datapoint is used only once. A test set is fixed and
common for all experiments. Each epoch of training consists of 200, 000 datapoints. The error metric is Root Mean
Squared Error (RMSE) in pose estimation.
4. Experiments
We use two simulation environments for our experiments, a 2D environment based on Box2D and a 3D environment based on the Doom game engine. To train the system, we pretrained the local pose estimation model and then
trained the global optimizer with the local pose model held
fixed. This was mainly due to the large sequence lengths we
were required to process (on the order of 1000 time steps).
This limited the amount of sequences we could process due
to the large memory requirements. Training the system in
stages enabled us to preprocess the sequence images into a
far more memory-efficient compressed representation.
To improve upon the results produced by the local pose
estimation model, we train a Neural Graph Optimizer on the
pose outputs of a pretrained Local Pose Estimation model.
For the 2D environment, as shown in Table 1, we observed
over 80% improvement in the correction of drift compared
to using only the local pose estimation model, as measured
by the root mean squared error loss. We can see that increasing the number of iterations (applying the attention operator and then the temporal aggregation operator) improved
results from 1 to 5 iterations. We show some sample trajectories in Fig. 5 before and after the Neural Graph Optimizer
procedure.
4.1. 2D Environment
For the 2D Environment, random maze designs are generated using Prim’s algorithm [37], and the environment is
created using Box2D (box2d.org). The agent projects 241
6
Results on the 3D Doom Environment
Seen
% Err. Trans.
% Err. Rot.
Only Local Estimation
1.65
0.117
Global Estimation - 1 Attend-Opt iteration
1.42
0.071
Global Estimation - 5 Attend-Opt iterations
1.25
0.057
DeepVO [45]
1.78
0.079
Model
Unseen
% Err. Trans.
% Err. Rot.
1.62
0.122
1.16
0.071
1.04
0.056
2.39
0.091
Table 2. Results for different Neural Graph Optimizer architectures and hyperparameters, in terms of % translation and rotation error
on maps either seen or unseen during training time. We can see that the addition of the global optimization procedure reduces error
significantly compared to using only the local pose model. In addition, increasing the number of attention iterations provides an increase
in performance.
Translational RMSE Ratio during training
4.2. 3D Environment
Rot RMSE Ratio ( ∘ /m∘
Trans RMSE Ratio (%)
1.5
For the 3D Environment, random maze designs are generated using the Kruskal’s algorithm [24], and the environment is created using the ViZDoom API [21]. The agent
observes the environment in a first-person view with a fieldof-view of 108◦ . An example of the 3D environment design
is shown in Fig. 4. Similr to the 2D environment, the pose
predictions are 3-dimensional tuples (x, y, angle) and the
agent can take one of three discrete actions at every time
step: move-forward, turn-left, or turn-right, which results
in translational or angular acceleration. For collecting data
in this environment, a navigation network [26] is trained
to maximize the distance travelled by the agent using the
Asynchronous Advantage Actor-Critic algorithm [32]. The
data is collected by using the policy learned by the navigation network.
Like the 2D environment, the training data is generated
by worker threads in parallel with the model training, and
each training datapoint is used only once. We additionally
sample two test sets, one containing 39 trajectories sampled
from maze geometries that were seen during training and
one containing 39 trajectories sampled from novel maze geometries that the agent had not encountered during training.
4.2.1
Rotational RMSE Ratio during training
0.12
1.6
1.4
1.3
1.2
1.1
1.0
0
2000
4000
6000
Epoch
8000
10000 12000
0.11
0.10
local een
global een
local un een
global un een
0.09
0.08
0.07
0.06
0.05
0
2000
4000
6000
Epoch
8000
10000 12000
Figure 6. Training curves for Doom over 13, 000 updates for
the 5 iteration Attend-Opt model. We show the performance on
both seen and unseen test sets as training progress. The dotted line
represents the estimate provided by using only the local model.
We can see there is a large reduction in error when making use of
the global optimizer.
The baseline DeepVO [45] is one of the state-of-the-art
methods using deep neural nets for monocular visual odometry. It stacks 2 consecutive frames and passes them through
9 convolutional layers followed by 2 LSTM layers to estimate the pose changes. As compared to the proposed Local Pose Estimation model which observes only the last 2
frames at the time, the DeepVO model can potentially utilize information from all the prior frames using the LSTM
layer. However, the DeepVO model does not correct its previous predictions as it observes new information. The Neural Graph Optimizer has the ability to correct its predictions
using the Attention operation and consequently leads to improved performance.
Results
Results are shown in Table 2. Here we report % Error
in Translation and Rotation for seen/unseen mazes, where
the accumulated drift error is divided by the entire distance
traveled in each trajectory. Observe that the local model
is significantly improved by using global optimization and
performance of the global model improves as we increase
the number of Attend-Opt iterations from 1 to 5. The global
model outperforms the DeepVO [45] baseline on both the
test sets. Additionally, we can clearly see that the model
itself does not overfit to the training environments it experienced, and gets similar or even lower error on unseen
test mazes. Learning curves are shown in Fig 6. We can
see that performance plateaus decrease significantly early
on and then progress is much slower after around 2000 updates.
4.2.2
Analysis
We next plot the total rotational and translational errors as
a function of number of steps in the trajectory in Figures 7
(for unseen mazes) and 8 (for seen mazes). The global
model reduces the slope of the rate of increase of both translational and rotation errors as compared to the local estimates. Figures 9 and 10 display the ratio of the translational
(left) and rotational (right) drift error over distance traveled.
We can see from these plots that the trend is negative, meaning that drift accumulates much slower than the distance being traveled. This indicates that the model is likely to generalize well to arbitrarily long trajectories. Additionally, in
all plots, we can see a clear ordering of the performance of
7
Figure 9. Ratio of the Translational (Left) and Rotational (Right)
RMSE to the distance travelled as a function of number of images
in the trajectory in unseeen mazes.
Figure 7. Translational (Left) and Rotational (Right) RMSE as a
function of number of images in the trajectory in unseen mazes.
Figure 8. Translational (Left) and Rotational (Right) RMSE as a
function of number of images in the trajectory in seen mazes.
Figure 10. Ratio of the Translational (Left) and Rotational (Right)
RMSE to the distance travelled as a function of number of images
in the trajectory in seen mazes.
the models, where the local model performs worst, one iteration of Attend-Opt increases model performance significantly, and increasing the number of Attend-Opt iterations
to 5 further increases model performance.
The plots in Figures 7 and 8 as well as the numbers in
Table 2 show that the improvement in rotational errors due
to the neural optimization is higher than the improvement
in translation errors. Fig 11 shows sample trajectories with
estimates of both global and local pose estimates. As seen
in the figure, the neural graph optimizer considerably improves the rotation estimates, consequently leading to significant improvements in the drift reduction.
14
groundtruth
predicted (opt)
predicted (unopt)
12
groundtruth
predicted (opt)
predicted (unopt)
8
10
6
8
4
6
2
4
0
2
−2
0
−2
−17.5 −15.0 −12.5 −10.0 −7.5
10
−2.5
0.0
2.5
−4
−2.5
0.0
2.5
5.0
7.5
10.0
20.0
groundtruth
predicted (opt)
predicted (unopt)
8
5. Conclusion
−5.0
12.5
15.0
groundtruth
predicted (opt)
predicted (unopt)
17.5
15.0
12.5
6
10.0
4
In this paper, we designed a novel attention-based architecture to perform an end-to-end trainable global pose estimation. Compared to the previous work on using deep
networks to do pose estimation, our method uses an attention operation to re-estimate its trajectory at each time step
and therefore enables iterative refinement of the quality of
its estimates as more data is available. We demonstrate the
benefit of the model on two simulators, the first is a topdown 2D maze world and the second is a 3D random maze
environment running the Doom engine. Our results show
that our method has an increased performance compared to
models which used only temporally local information.
The proposed method can be further extended to a complete end-to-end graph-based SLAM system by adding a relocalization module which uses pose features to relocalize
in a known map [5]. It can also be extended to an Active
SLAM system where the agent also decides the actions, in
order to map the environment as fast as possible.
7.5
5.0
2
2.5
0.0
0
−4
−2
0
2
4
6
8
10
−5
0
5
10
15
Figure 11. Images visually demonstrating the effect on pose estimates of adding the Neural Graph Optimizer module on top of
the local pose estimation model in the 3D environment. We can
see that the global optimization procedure greatly reduces drift.
These figures were generated with the 5 iteration Neural Graph
Optimization model. The agent always starts at the origin (0, 0).
Acknowledgments
We thank Tim Barfoot and Russ Webb for helpful comments and discussions. We would also like to thank Barry
Theobald and Megan Maher for helpful feedback on the
manuscript.
8
References
[16] D. Gálvez-López and J. D. Tardos. Bags of binary words for
fast place recognition in image sequences. IEEE Transactions on Robotics, 28(5):1188–1197, 2012.
[17] S. Gupta, J. Davidson, S. Levine, R. Sukthankar, and J. Malik. Cognitive mapping and planning for visual navigation.
In Proceedings of the IEEE Conference on Computer Vision
and Pattern Recognition, pages 2616–2625, 2017.
[18] T. Haarnoja, A. Ajay, S. Levine, and P. Abbeel. Backprop
kf: Learning discriminative deterministic state estimators. In
Advances in Neural Information Processing Systems, pages
4376–4384, 2016.
[19] J. A. Hesch, D. G. Kottas, S. L. Bowman, and S. I. Roumeliotis. Camera-imu-based localization: Observability analysis
and consistency improvement. The International Journal of
Robotics Research, 33(1):182–201, 2014.
[20] M. Kaess, H. Johannsson, R. Roberts, V. Ila, J. J. Leonard,
and F. Dellaert. isam2: Incremental smoothing and mapping
using the bayes tree. The International Journal of Robotics
Research, 31(2):216–235, 2012.
[21] M. Kempka, M. Wydmuch, G. Runc, J. Toczek, and
W. Jaśkowski. Vizdoom: A doom-based ai research platform
for visual reinforcement learning. In Computational Intelligence and Games (CIG), 2016 IEEE Conference on, pages
1–8. IEEE, 2016.
[22] A. Kendall, M. Grimes, and R. Cipolla. Posenet: A convolutional network for real-time 6-dof camera relocalization. In
Proceedings of the IEEE international conference on computer vision, pages 2938–2946, 2015.
[23] D. G. Kottas, J. A. Hesch, S. L. Bowman, and S. I. Roumeliotis. On the consistency of vision-aided inertial navigation.
In Experimental Robotics, pages 303–317. Springer, 2013.
[24] J. B. Kruskal. On the shortest spanning subtree of a graph
and the traveling salesman problem. Proceedings of the
American Mathematical society, 7(1):48–50, 1956.
[25] R. Kümmerle, G. Grisetti, H. Strasdat, K. Konolige, and
W. Burgard. g 2 o: A general framework for graph optimization. In Robotics and Automation (ICRA), 2011 IEEE
International Conference on, pages 3607–3613. IEEE, 2011.
[26] G. Lample and D. S. Chaplot. Playing fps games with deep
reinforcement learning. In AAAI, pages 2140–2146, 2017.
[27] S. Leutenegger, S. Lynen, M. Bosse, R. Siegwart, and P. Furgale. Keyframe-based visual–inertial odometry using nonlinear optimization. The International Journal of Robotics
Research, 34(3):314–334, 2015.
[28] F. Liu, C. Shen, and G. Lin. Deep convolutional neural fields
for depth estimation from a single image. In Proceedings
of the IEEE Conference on Computer Vision and Pattern
Recognition, pages 5162–5170, 2015.
[29] J. McCormac, A. Handa, A. Davison, and S. Leutenegger.
Semanticfusion: Dense 3d semantic mapping with convolutional neural networks. In Robotics and Automation (ICRA),
2017 IEEE International Conference on, pages 4628–4635.
IEEE, 2017.
[30] I. Melekhov, J. Ylioinas, J. Kannala, and E. Rahtu. Relative
camera pose estimation using convolutional neural networks.
In International Conference on Advanced Concepts for Intelligent Vision Systems, pages 675–687. Springer, 2017.
[1] S. Bhatti, A. Desmaison, O. Miksik, N. Nardelli, N. Siddharth, and P. H. Torr.
Playing doom with slamaugmented deep reinforcement learning. arXiv preprint
arXiv:1612.00380, 2016.
[2] S. L. Bowman, N. Atanasov, K. Daniilidis, and G. J. Pappas.
Probabilistic data association for semantic slam. In Robotics
and Automation (ICRA), 2017 IEEE International Conference on, pages 1722–1729. IEEE, 2017.
[3] E. Brachmann, A. Krull, S. Nowozin, J. Shotton, F. Michel,
S. Gumhold, and C. Rother. Dsac-differentiable ransac for
camera localization. In IEEE Conference on Computer Vision and Pattern Recognition (CVPR), volume 3, 2017.
[4] C. Cadena, L. Carlone, H. Carrillo, Y. Latif, D. Scaramuzza,
J. Neira, I. Reid, and J. J. Leonard. Past, present, and
future of simultaneous localization and mapping: Toward
the robust-perception age. IEEE Transactions on Robotics,
32(6):1309–1332, 2016.
[5] D. S. Chaplot, E. Parisotto, and R. Salakhutdinov. Active
neural localization. In International Conference on Learning
Representations (ICLR), 2018.
[6] R. Clark, S. Wang, H. Wen, A. Markham, and N. Trigoni.
Vinet: Visual-inertial odometry as a sequence-to-sequence
learning problem. In AAAI, pages 3995–4001, 2017.
[7] G. Costante, M. Mancini, P. Valigi, and T. A. Ciarfuglia. Exploring representation learning with cnns for frame-to-frame
ego-motion estimation. IEEE robotics and automation letters, 1(1):18–25, 2016.
[8] F. Dellaert. Factor graphs and gtsam: A hands-on introduction. Technical report, Georgia Institute of Technology,
2012.
[9] D. DeTone, T. Malisiewicz, and A. Rabinovich. Deep image
homography estimation. arXiv preprint arXiv:1606.03798,
2016.
[10] E. W. Dijkstra. A note on two problems in connexion with
graphs. Numerische mathematik, 1(1):269–271, 1959.
[11] A. Dosovitskiy, P. Fischer, E. Ilg, P. Hausser, C. Hazirbas,
V. Golkov, P. van der Smagt, D. Cremers, and T. Brox.
Flownet: Learning optical flow with convolutional networks.
In Proceedings of the IEEE International Conference on
Computer Vision, pages 2758–2766, 2015.
[12] J. Engel, T. Schöps, and D. Cremers. Lsd-slam: Large-scale
direct monocular slam. In European Conference on Computer Vision, pages 834–849. Springer, 2014.
[13] M. A. Fischler and R. C. Bolles. Random sample consensus:
a paradigm for model fitting with applications to image analysis and automated cartography. In Readings in computer
vision, pages 726–740. Elsevier, 1987.
[14] C. Forster, M. Pizzoli, and D. Scaramuzza. Svo: Fast semidirect monocular visual odometry. In Robotics and Automation (ICRA), 2014 IEEE International Conference on, pages
15–22. IEEE, 2014.
[15] C. Forster, Z. Zhang, M. Gassner, M. Werlberger, and
D. Scaramuzza. Svo: Semidirect visual odometry for
monocular and multicamera systems. IEEE Transactions on
Robotics, 33(2):249–265, 2017.
9
[31] P. Mirowski, R. Pascanu, F. Viola, H. Soyer, A. Ballard,
A. Banino, M. Denil, R. Goroshin, L. Sifre, K. Kavukcuoglu,
et al. Learning to navigate in complex environments. arXiv
preprint arXiv:1611.03673, 2016.
[32] V. Mnih, A. P. Badia, M. Mirza, A. Graves, T. Lillicrap,
T. Harley, D. Silver, and K. Kavukcuoglu. Asynchronous
methods for deep reinforcement learning. In International
Conference on Machine Learning, pages 1928–1937, 2016.
[33] A. I. Mourikis and S. I. Roumeliotis. A multi-state constraint kalman filter for vision-aided inertial navigation. In
Robotics and automation, 2007 IEEE international conference on, pages 3565–3572. IEEE, 2007.
[34] R. Mur-Artal, J. M. M. Montiel, and J. D. Tardos. Orb-slam:
a versatile and accurate monocular slam system. IEEE Transactions on Robotics, 31(5):1147–1163, 2015.
[35] R. A. Newcombe, S. Izadi, O. Hilliges, D. Molyneaux,
D. Kim, A. J. Davison, P. Kohi, J. Shotton, S. Hodges, and
A. Fitzgibbon. Kinectfusion: Real-time dense surface mapping and tracking. In Mixed and augmented reality (ISMAR),
2011 10th IEEE international symposium on, pages 127–
136. IEEE, 2011.
[36] E. Parisotto and R. Salakhutdinov. Neural map: Structured
memory for deep reinforcement learning. arXiv preprint
arXiv:1702.08360, 2017.
[37] R. C. Prim. Shortest connection networks and some generalizations. Bell Labs Technical Journal, 36(6):1389–1401,
1957.
[38] R. F. Salas-Moreno, R. A. Newcombe, H. Strasdat, P. H.
Kelly, and A. J. Davison. Slam++: Simultaneous localisation and mapping at the level of objects. In Proceedings of
the IEEE conference on computer vision and pattern recognition, pages 1352–1359, 2013.
[39] T. Sattler, B. Leibe, and L. Kobbelt. Efficient & effective prioritized matching for large-scale image-based localization.
IEEE transactions on pattern analysis and machine intelligence, 2016.
[40] T. Sattler, W. Maddern, A. Torii, J. Sivic, T. Pajdla, M. Pollefeys, and M. Okutomi. Benchmarking 6dof urban visual localization in changing conditions. arXiv preprint
arXiv:1707.09092, 2017.
[41] L. Tai, G. Paolo, and M. Liu. Virtual-to-real deep reinforcement learning: Continuous control of mobile robots for mapless navigation. In Intelligent Robots and Systems (IROS),
2017 IEEE/RSJ International Conference on, pages 31–36.
IEEE, 2017.
[42] S. Thrun, W. Burgard, and D. Fox. Probabilistic robotics.
MIT press, 2005.
[43] A. Vaswani, N. Shazeer, N. Parmar, J. Uszkoreit, L. Jones,
A. N. Gomez, Ł. Kaiser, and I. Polosukhin. Attention is all
you need. In Advances in Neural Information Processing
Systems, pages 6000–6010, 2017.
[44] F. Walch, C. Hazirbas, L. Leal-Taixé, T. Sattler, S. Hilsenbeck, and D. Cremers. Image-based localization with spatial
lstms. CoRR, abs/1611.07890, 2016.
[45] S. Wang, R. Clark, H. Wen, and N. Trigoni. Deepvo:
Towards end-to-end visual odometry with deep recurrent
[46]
[47]
[48]
[49]
[50]
10
convolutional neural networks. In Robotics and Automation (ICRA), 2017 IEEE International Conference on, pages
2043–2050. IEEE, 2017.
T. Weyand, I. Kostrikov, and J. Philbin. Planet-photo geolocation with convolutional neural networks. In European
Conference on Computer Vision, pages 37–55. Springer,
2016.
T. Whelan, R. F. Salas-Moreno, B. Glocker, A. J. Davison,
and S. Leutenegger. Elasticfusion: Real-time dense slam
and light source estimation. The International Journal of
Robotics Research, 35(14):1697–1716, 2016.
J. Zhang, J. T. Springenberg, J. Boedecker, and W. Burgard. Deep reinforcement learning with successor features
for navigation across similar environments. arXiv preprint
arXiv:1612.05533, 2016.
J. Zhang, L. Tai, J. Boedecker, W. Burgard, and M. Liu. Neural SLAM. CoRR, abs/1706.09520, 2017.
Y. Zhu, R. Mottaghi, E. Kolve, J. J. Lim, A. Gupta, L. FeiFei, and A. Farhadi. Target-driven visual navigation in indoor scenes using deep reinforcement learning. In Robotics
and Automation (ICRA), 2017 IEEE International Conference on, pages 3357–3364. IEEE, 2017.
| 1cs.CV
|
ANGLE CONSTRAINED PATH TO CLUSTER
MULTIPLE MANIFOLDS
Amir Babaeian
University of California, San Diego
Abstract. In this paper, we propose a method to cluster multiple intersected
manifolds. The algorithm chooses several landmark nodes randomly and then
checks whether there is an angle-constrained path between each landmark node
and every other node in the neighborhood graph. When the points lie on different
manifolds with intersection they should not be connected using a smooth path,
thus the angle constraint is used to prevent connecting points from one cluster to
another one. The resulting algorithm is implemented as a simple variation of
Dijkstra’s algorithm used in Isomap. However, Isomap was specifically designed
for dimensionality reduction in the single-manifold setting, and in particular, cannot handle intersections. Our method is simpler than the previous proposals in
the literature and performs comparably to the best methods, both on simulated
and some real datasets.
1 Introduction
Since last decade, the amount of collected data is increasing exponentially. Data mining techniques represent each data sample with an extracted very high-dimensional
fea-ture vector. However, the intrinsic dimension of these features is much smaller that
the dimension of ambient space. Precisely, the features are sampled from the surface
of a low-dimensional manifold embedded in the ambient space [7, 19]. The most wellknown manifold learning algorithms such as Locally Linear Embedding (LLE) [16],
Laplacian Eigenmap (LE) [2], and Isomap [18]. All these methods assume that the
data is coming from a single manifold. However, this assumption is not correct in many
application such as video segmentation [5], where the data is sampled from multiple
intersected manifolds. However, clustering the data points into several intersected
man-ifolds is itself a challenging problem.
In this work, we consider the problem of clustering data points sampled
from the vicinity of multiple intersected manifolds embedded in ambient
space. Here, we focus on a nonparametric approach based on the
assumption that the manifolds are smooth. See Fig. 1.
Related works in the area are based on different assumptions. Most methods are
designed for the cases where the surfaces do not intersect [15, 14, 6], while others work
only when the surfaces that intersect have different intrinsic dimension or density [8, 12].
The method of [1] is only able to separate intersecting curves. Methods that pur-posefully
aim at resolving intersections are fewer. [17] implement some variant of K-means where the
centers are surfaces. [11] propose to minimize a (combinatorial) en-ergy that includes local
orientation information, using a tabu search. The state-of-the-art
(a)
(b)
Fig. 1. Simulated data illustrating the problem of multi-manifold clustering; (a) A
sample 3D data containing three intersected manifolds; (b) three clustered manifolds
as the output of the proposed method.
methods work based on local principal component analysis (PCA). An early proposal
was the elaborate multi-scale spectral method of [13], while the clustering routine of
[9], developed in the context of semi-supervised learning, inspired by the works of
[20] and [10]. Here, we propose a markedly different approach based on
connecting points to landmarks via angle-constrained paths. It can be seen as a
constrained variant of Isomap [18]. Isomap was specifically designed for
dimensionality reduction in the single-manifold setting, and in particular, cannot
handle intersections. The angle con-straint on paths is there to prevent connecting
points from one cluster to points from a different, intersecting cluster. The resulting
algorithm is implemented as a simple varia-tion of Dijkstra’s algorithm. The
experimental results, conducted on both synthetic and real real data sets, confirm
that our method performs comparably to the state-of-the-art methods.
The rest of the paper is organized as follows. In Section 2, we describe in
detail the problem of multi-manifold clustering followed by discussing the methods
of [17] and [20], as well as the subspace clustering method of [4], which will serve
as bench-marks in our experiments. In Section 3 we motivate and describe our
proposed method. In Section 4, we present some numerical experiments on
synthetic and real data sets. Finally, in Section 5, we draw our conclusion.
2 Multi-Manifold Clustering
D
We observe points x1, . . . , xN in a Euclidean space R , with D ≥ 2, that are assumed
D
to be sampled from the vicinity of embedded submanifolds S1, . . . , SK ⇢ R . The
surfaces are assumed to be compact and smooth — defined here as having bounded
maximum pointwise curvature and a boundary (if any) satisfying the same property
— and may be of possibly different intrinsic dimensions. The points are independent
and identically drawn from a mixture distribution supported in the neighborhood of the
surfaces. A simple model assumes
KX
k
x i = s i + z i, s i ⇠
=1
⇡kµSk kzik ⌧,
(1)
where µS denotes the uniform distribution over S. (s ⇠ µ means
Pthat the random point
s has distribution µ.) The mixture weights satisfy ⇡i ≥ 0 and k ⇡k = 1, and zi is a
uniform noise bounded by ⌧ . Let Ik index the points si that were sampled from
surface Sk, meaning Ik = {i : si ⇠ µSk }. Note that I1, . . . , IK forms a partition of the
data. (This implicitly assumes there are no outliers, which we take to be the case in
this pa-per.) The goal of multi-manifold clustering is to recover this partition up to
permutation of {1, . . . , K}. We note that the term ’submanifold’ is unduly restrictive for
us since we allow the underlying surfaces to self-intersect. Our method works
essentially as well on such surfaces as on sub-manifolds. See 2.
(a)
(b)
Fig. 2. Simulated data illustrating our method working well in a case where the
underlying sur-faces self-intersect. Left: data. Right: output from our method.
We also note that, when the underlying surfaces are allowed to intersect, some
sort of smoothness assumption seems necessary in a nonparametric setting like ours,
for otherwise the clustering objective is ambiguous — think of a T-intersection.
Authors in [17] suggest an algorithm that mimics K-means algorithm by replacing
centroid points with centroid sub-manifolds. The method starts like Isomap by build-ing
a neighborhood graph and computes the shortest path distances within the graph. In
[20], dissimilarity is used to factor the Euclidean distance and the discrepancy between the local orientation of the data. The surfaces are assumed to be of same
dimen-sion d known to the user. We chose the subspace clustering method of [4]
among a few other methods that perform well in this context.In [4], people proposed a
spectral method for subspace clustering based on the affinity of underlying manifolds.
3 Approach
Our algorithm is quite distinct from all the other methods for multi-manifold clustering
we are aware of, although it starts by building a q-nearest neighbor graph like many
others. The idea is very simple and amounts to clustering together points that are connected by an angle-constrained path in the neighborhood graph. For an angle ✓ 2 [0, ⇡],
1
m
−−−−− !
−−−−−tt!
we say that a path (xi , . . . , xi ) is ✓-constrained if (x i − xi , x i xi )
✓ for all
t = 2, . . . , m − 1.
The rationale is the following. Take two surfaces S1 and S2 intersecting at
a strictly positive angle. Then for ‘most’ pairs of data points xi1 2 S1 and xi2 2
S2, a path in the graph going from xi1 to xi2 has at least one large angle
between two successive edges, on the order of the incidence angle between
the surfaces; while for ‘most’ pairs of data points xi1 , xi2 2 S1, there is a path
with all angles between successive edges relatively small.
To speedup the implementation, we select M landmarks (with M slightly
larger than K) at random among the data points and only identify what data
points are con-nected to what landmark via a ✓-constrained path in the graph.
M and ✓ are parameters of the algorithm.
Let ⇠`i = 1 if point i and landmark ` are connected that way, and ⇠`i = 0 if
not. We use ⇠ i := (⇠`i : ` = 1, . . . , M) as feature vectors that we group
together and cluster using hierarchical clustering with complete linkage.
\
t1 t
+1
3.1 Computational Complexity
Building a symmetric q-nearest neighbor graph using cover trees [3] takes order O(qN
log N), where the implicit constant depends exponentially on the intrinsic dimensions of
the surfaces and linearly on the ambient dimension D. The angle-constrained pathfinder
routine is a simple variant of Dijkstra’s algorithm, whose implementation by Fibonacci
heaps runs in O(qN log N) . Hence, calling this routine once for each landmark costs
O(qM N log N). Grouping the feature vectors O(F M N) and then clustering them by
2
complete linkage costs O(F log F ), where F is the (data-dependent) number of distinct feature vectors ⇠i, often of the order of K in our experiments.
Algorithm 1 Path-Based Clustering (PBC)
Input: data (xi); parameters q, K, M,✓
Build q-nearest neighbor graph
Choose M landmarks are random
for i = 1 to n do
For each landmark xˆ`, identify which points xi it is connected to via a ✓constrained path in the graph, and set ⇠`i = 1 if so, and ⇠`i = 0 otherwise.
end for
Group and then apply hierarchical clustering to the feature vectors ⇠1, . . . , ⇠n to
find K clus-ters, where ⇠i := (⇠`i : ` = 1, . . . , M).
(a)
(b)
(c)
(d)
(e)
(f)
(g)
(h)
Fig. 3. Result of our method on 8 synthetic datasets.
4 Experiments
In this section, we compare the performance of our methods with those
described in 2 on various synthetic and real data sets.
4.1 Synthetic Data
The synthetic datasets we generated are similar to those appearing in the literature.
They are Three Planes (TP), Two Spirals (TSI), Five Segments (FS), Dollar-Sign,
Plane and Roll (DSPR), Roll and Plane (RP), Cone and Plane (CP), Two Spheres
(TSH), Rose Curve and Circle (RCC). Figure 3 shows the result of our method on eight
synthetic datasets, with good performance in all cases. The misclustering rates for our
method, and the other three methods, are presented in Table 1, where we see that our
method achieves a performance at least comparable to the best of the other three
methods on each dataset. Spectral Curvature Clustering (SCC) works well on linear
manifolds (as expected) while it fails when there is curvature. See Fig. 4(a). KManifolds fails in the more complicated examples. We found that this algorithm is very
slow since it has to compute the shortest path between all the points, so that we could
not apply it to some of the largest datasets. We mention that it assumes that clusters
intersect, and otherwise does not work properly. Fig. 4(b) and Fig. 4(c) shows two
cases where K-Manifolds fails. Our method and Spectral Multi-Manifolds Clustering
(SMMC) perform compa-rably on most datasets, but SMMC fails in the Rose Curve
and Circle example shown in 4(d). We note that K-Manifold, SCC and SMMC all
require that all surfaces are of same dimension, which is a parameter of these
methods, why our method does not need knowledge of the intrinsic dimensions of the
surfaces and can operate even when these are different.
Table 1. Clustering accuracy on synthetic data.
DATA SET K-MANIFOLDS
TP
95.1%
TSI
95.2%
FS
59.1%
DSPR
46.1%
RP
53.5%
CP
TS
RCC
62.9%
SCC SMMC PBC
98.8% 99.3% 93.0%
54.8% 99.6% 98.2%
94.9% 99.6% 98.1%
- 99.4% 97.5%
- 95.6% 94.6%
- 99.3% 94.9%
- 95.5% 97.9%
- 64.7% 99.4%
4.2 Real Data
We applied our method on the COIL-20 dataset which includes 1440 gray-scale images of
the size 128 ⇥ 128 pixels from 20 objects. Each object contains 72 images taken from
different view points. We first projected the data set onto the top 10 principal
(a)
(b)
(c)
(d)
Fig. 4. (a) An example where SCC fails; (b-c) examples where K-Manifolds fails; (d) An
example where SMMC fails.
components, then applied our path-based clustering algorithm. We tested our method on
the three very similar objects 3, 6 and 19. The algorithm is 99% accurate (misclusters are
only 2 images out of 216) bringing a significant improvement over the state-of-the-art result
of 70% reported in [20]. Lastly, we evaluated our method on the whole dataset obtaining an
83.6% accuracy, improving on the 70.7% accuracy reported in [20]. Since in this case we
have 20 different classes, we increased the number of landmarks to 100 to make sure we
have sampled at least a few landmarks from each class. In 5, we show the results of our
algorithm and those of SCC and SMMC on first 5 objects in a 3D plot where the points are
represented by their first 3 principal components. As it can be seen other 3 methods fail to
cluster multiple manifold with intersection whereas Path-Based Clustering seems much
closer to the ground truth.
4.3 Clustering of Human Motion Sequences
In this section, we test our algorithm on a sequence of video frames including differ-ent
activities performed by a subject. [21, 22, 24, 23] studied unsupervised tracking of
human bodies in the presence of temporal changes. We use a different model to track
4 mixed actions from subject 86, trial number 9 of the CMU MoCap dataset. The data
Real Clusters
(a)
SMMC
(c)
SCC
(b)
PBC
(d)
Fig. 5. Output of three different methods on the first 5 objects of COIL-20. Points are
represented by their 3 top principal components.
consists in a temporal sequence of 62-dimensional representation of the
human body via markers in R3. One motion sequence of 4794 frames and
corresponding result of path-based multi-manifold clustering are given in
Figure 6. Four activities are labeled from 1 to 4. When the subject switches
from one action to another one the labels are ambiguous.
5 Conclusion
In this paper, we proposed a novel method to cluster multiple intersected
manifolds that perform well in practice and is consistent in the large-sample
limit. The variant we introduced is the best on numerical experiments. For
now, the variant we introduced here is the best on numerical experiments.
4
3
2
1
1000
200 0
3 000
4000
1000
200 0
3 000
4000
4
3
2
1
Fig. 6. Result of human activity segmentation using Path-Based Clustering. There are
4 activities: walking (1), looking (2), sitting (3) and standing (4). Top: a sample of the
sequence. Middle: ground truth. Bottom: output of our algorithm.
Acknowledgments
This work was partially supported by a grant from the National Science
Foundation DMS- 09-15160.
References
1. E. Arias-Castro, G. Chen, and G. Lerman. Spectral clustering based on local linear
approxi-mations. Electron. J. Statist., 5:1537–1587, 2011.
2. Mikhail Belkin and Partha Niyogi. Laplacian eigenmaps for dimensionality
reduction and data representation. Neural computation, 15(6):1373–1396, 2003.
3. Alina Beygelzimer, Sham Kakade, and John Langford. Cover trees for nearest
neighbor. In ICML ’06: Proceedings of the 23rd international conference on
Machine learning, pages 97–104, New York, NY, USA, 2006. ACM.
4. G. Chen and G. Lerman. Spectral curvature clustering (SCC). IJCV, 81(3):317–330, 2009.
5. Ehsan Elhamifar and Rene´ Vidal. Sparse subspace clustering. In Computer Vision and Pattern Recognition, 2009. CVPR 2009. IEEE Conference on, pages 2790–2797. IEEE, 2009.
6. Ehsan Elhamifar and Rene Vidal. Sparse manifold clustering and embedding. In J.
Shawe-Taylor, R.S. Zemel, P. Bartlett, F.C.N. Pereira, and K.Q. Weinberger, editors,
Advances in Neural Information Processing Systems 24, pages 55–63. 2011.
7. Ehsan Elhamifar and Rene Vidal. Sparse subspace clustering: Algorithm, theory,
and ap-plications. Pattern Analysis and Machine Intelligence, IEEE Transactions
on, 35(11):2765– 2781, 2013.
8. Aristides Gionis, Alexander Hinneburg, Spiros Papadimitriou, and Panayiotis
Tsaparas. Di-mension induced clustering. In KDD ’05: Proceedings of the eleventh
ACM SIGKDD inter-national conference on Knowledge discovery in data mining,
pages 51–60, New York, NY, USA, 2005. ACM.
9. A.B. Goldberg, X. Zhu, A. Singh, Z. Xu, and R. Nowak. Multi-manifold semisupervised learning. In AISTATS, pages 169–176, 2009.
10. Dian Gong, Xuemei Zhao, and Gerard´ Medioni. Robust multiple manifolds
structure learn-ing. In Proc. 29th Intl. Conf. on Machine Learning (ICML), 2012.
11. Qiyong Guo, Hongyu Li, Wenbin Chen, I-Fan Shen, and Jussi Parkkinen. Manifold
cluster-ing via energy minimization. In ICMLA ’07: Proceedings of the Sixth
International Confer-ence on Machine Learning and Applications, pages 375–380,
Washington, DC, USA, 2007. IEEE Computer Society.
12. G. Haro, G. Randall, and G. Sapiro. Stratification learning: Detecting mixed density
and di-mensionality in high dimensional point clouds. Advances in Neural
Information Processing Systems, 19:553, 2007.
13. D. Kushnir, M. Galun, and A. Brandt. Fast multiscale clustering and manifold
identification. Pattern Recogn., 39(10):1876–1891, 2006.
14. A.Y. Ng, M.I. Jordan, and Y. Weiss. On spectral clustering: Analysis and an
algorithm. Advances in neural information processing systems, 2:849–856, 2002.
15. M. Polito and P. Perona. Grouping and dimensionality reduction by locally linear
embedding. Advances in Neural Information Processing Systems, 14:1255–1262, 2001.
16. S. Roweis and L. Saul. Nonlinear dimensionality reduction by locally linear
embedding. Science, 290(5500):2323–2326, 2000.
17. R. Souvenir and R. Pless. Manifold clustering. In Computer Vision, 2005. ICCV 2005.
Tenth IEEE International Conference on, volume 1, pages 648–653 Vol. 1, 2005.
18. J. B. Tenenbaum, V. de Silva, and J. C. Langford. A global geometric framework
for nonlin-ear dimensionality reduction. Science, 290(5500):2319–2323, 2000.
19. Rene´ Vidal and Paolo Favaro. Low rank subspace clustering (lrsc). Pattern
Recognition Letters, 43:47–61, 2014.
20. Y. Wang, Y. Jiang, Y. Wu, and Z.H. Zhou. Spectral clustering on multiple
manifolds. Neural Networks, IEEE Transactions on, 22(7):1149–1161, 2011.
21. Bardia Yousefi, Julien Fleuret, Seyed Alireza Ghaffari, Simon Frechet,´ Felix´
Labrie Larrivee,´ Marcelo Sung Ma Jo, Matthieu Klein, Xavier Maldague, and
Raymon Watts. Unsupervised automatic tracking of thermal changes in human
body. Advanced Infrared Technology & Applications, page 152, 2015.
22. Bardia Yousefi and Chu Kiong Loo. Comparative study on interaction of form and
motion processing streams by applying two different classifiers in mechanism for
recognition of biological movement. The Scientific World Journal, 2014, 2014.
23. Bardia Yousefi and Chu Kiong Loo. Slow feature action prototypes effect
assessment in mechanism for recognition of biological movement ventral stream.
International Journal of Bio-Inspired Computation, 8(6):410–424, 2016.
24. Bardia Yousefi, Chu Kiong Loo, and Ali Memariani. Biological inspired human
action recognition. In Robotic Intelligence In Informationally Structured Space
(RiiSS), 2013 IEEE Workshop on, pages 58–65. IEEE, 2013.
| 1cs.CV
|
arXiv:1604.01168v2 [math.CO] 12 Nov 2016
An Estimation of the Size of Non-Compact Suffix
Trees
Bálint Vásárhelyi∗
Abstract
A suffix tree is a data structure used mainly for pattern matching. It is known
that the space complexity of simple suffix trees is quadratic in the length of the
string. By a slight modification of the simple suffix trees one gets the compact
suffix trees, which have linear space complexity. The motivation of this paper is
the question whether the space complexity of simple suffix trees is quadratic not
only in the worst case, but also in expectation.
1
Introduction
A suffix tree is a powerful data structure which is used for a large number of combinatorial
problems involving strings. Suffix tree is a structure for compact storage of the suffixes
of a given string. The compact suffix tree is a modified version of the suffix tree, and
it can be stored in linear space of the length of the string, while the non-compact suffix
tree is quadratic (see [11, 14, 18, 19]).
The notion of suffix trees was first introduced by Weiner [19], though he used the name
compacted bi-tree. Grossi and Italiano mention that in the scientific literature, suffix
trees have been rediscovered many times, sometimes under different names, like compacted bi-tree, prefix tree, PAT tree, position tree, repetition finder, subword tree etc.
[10] .
Linear time and space algorithms for creating the compact suffix tree were given soon
by Weiner [19], McCreight [14], Ukkonen [18], Chen and Sciferas [4] and others.
The statistical behaviour of suffix trees has been also studied. Most of the studies
consider improved versions.
∗
Szegedi Tudományegyetem, TTIK, Szeged, 6720, Hungary. E-mail: [email protected]
1
The average size of compact suffix trees was examined by Blumer, Ehrenfeucht and
Haussler [3]. They proved that the average number of nodes in the compact suffix tree
is asymptotically the sum of an oscillating function and a small linear function.
An important question is the height of suffix trees, which was answered by Devroye,
Szpankowski and Rais [6], who proved that the expected height is logarithmic in the
length of the string.
The application of suffix trees is very wide. We mention but only a few examples.
Apostolico et al. [2] mention that these structures are used in text searching, indexing,
statistics, compression. In computational biology, several algorithms are based on suffix
trees. Just to refer a few of them, we mention the works of Höhl et al. [12], Adebiyi et
al. [1] and Kaderali et al. [13]
Suffix trees are also used for detecting plagiarism [2], in cryptography [15, 16], in data
compression [7, 8, 16] or in pattern recognition [17].
For the interested readers further details on suffix trees, their history and their applications can be found in [2], in [10] and in [11], which sources we also used for the overview
of the history of suffix trees.
It is well-known that the non-compact suffix tree can be quadratic in space as we referred
before. In our paper we are setting a lower bound on the average size, which is also
quadratic.
2
Preliminaries
Before we turn to our results, let us define a few necessary notions.
Definition 1. An alphabet Σ is a set of different characters. The size of an alphabet is
the size of this set, which we denote by σ(Σ), or more simply σ. A string S is over the
alphabet Σ if each character of S is in Σ.
⋄
Definition 2. Let S be a string. S[i] is its ith character, while S[i, j] is a substring of
S, from S[i] to S[j], if j ≥ i, else S[i, j] is the empty string. Usually n(S) (or n if there
is no danger of confusion) denotes the length of the string.
⋄
Definition 3. The suffix tree of S is a rooted directed tree with n leaves, where n is the
length of S.
Its structure is the following:
Each edge e has a label ℓ(e), and the edges from a node v have different labels (thus,
the suffix tree of a string is unique). If we concatenate the edge labels along a path P,
we get the path label L(P).
2
We denote the path from the root to the leaf j by P(j). The edge labels are such that
L(j) = L(P(j)) is S[j, n] and a $ sign at the end. The definition becomes more clear
if we check the example on Figure 1 and Algorithm 4.
⋄
b
c
b
c
c
a
$
a
b
6
$
b
c
c
b
b
c
c
$
b
c
$
b
5
$
4
Growth of the string
3
2
$
1
Figure 1: Suffix tree of string aabccb
A naive algorithm for constructing the suffix tree is the following:
Algorithm 4. Let S be a string of length n. Let j = 1 and T be a tree of one vertex r
(the root of the suffix tree).
Step 1: Consider X = S[j, n] + $. Set i = 0, and v = r.
Step 2: If there is an edge vu labelled X[i + 1], then set v = u and i = i + 1.
Step 3: Repeat Step 2 while it is possible.
Step 4: If there is no such an edge, add a path of n − j − i + 2 edges from v, with
labels corresponding to S[j + i, n] + $, consecutively on the edges. At the end of
the path, number the leaf with j.
Step 5: Set j = j + 1, and if j ≤ n, go to Step 1.
⋄
Notice that in Algorithm 4 a leaf always remain a leaf, as $ (the last edge label before a
leaf) is not a character in S.
3
Definition 5. The compact suffix tree is a modified version of the suffix tree. We get it
from the suffix tree by compressing its long branches.
⋄
The structure of the compact suffix tree is basically similar to that of the suffix tree, but
an edge label can be longer than one character, and each internal node (i.e. not leaf)
must have at least two children. For an example see Figure 2.
b
c
b$
5
ccb$
cb$
4
a
bccb$
$
3
6
abccb$
2
1
Figure 2: Compact tree of string aabccb
With a regard to suffix trees, we can define further notions for strings.
Definition 6. Let S be a string, and T be its (non-compact) suffix tree.
A natural direction of T is that all edges are directed from the root towards the leaves.
If there is a directed path from u to v, then v is a descendant of u and u is an ancestor
of v.
We say that the growth of S (denoted by γ(S)) is one less than the shortest distance of
leaf 1 from an internal node v which has at least two children (including leaf 1), that is,
we count the internal nodes on the path different from v. If leaf j is a descendant of v,
then the common prefix of S[j, n] and S[1, n] is the longest among all j’s.
⋄
If we consider the string S = aabccb, the growth of S is 5, as it can be seen on Figure 1.
An important notion is the following one.
Definition 7. Let Ω(n, k, σ) be the number of strings of length n with growth k over an
alphabet of size σ.
⋄
Observe that the connection between the growth and the number of nodes in a suffix
tree is the following:
Observation 8. If we construct the suffix tree of S by using Algorithm 4, we get that the
sum of the growths of S[n − 1, n], S[n − 2, n], . . . , S[1, n] is a lower bound to the number
4
of nodes in the final suffix tree. In fact, there are only two more internal nodes, the root
vertex, the only node on the path to leaf n, and we have the leaves.
In the proofs we will need the notion of period and of aperiodic strings.
Definition 9. Let S be a string of length n. We say that S is periodic with period d, if
there is a d|n for which S[i] = S[i + d] for all i ≤ n − d. Otherwise, S is aperiodic.
The minimal period of S is the smallest d with the property above.
⋄
Definition 10. µ(j, σ) is the number of j-length aperiodic strings over an alphabet of
size σ.
⋄
A few examples for the number of aperiodic strings are given in Table 1.
σ
2
3
4
5
µ(1, σ) µ(2, σ) µ(3, σ) µ(4, σ) µ(5, σ) µ(6, σ) µ(7, σ) µ(8, σ)
2
6
12
30
54
126
240
504
3
6
24
72
240
696
2184
648
4
12
60
240
1020
4020
16380 65280
5
20
120
600
3120
15480 78120 390000
Table 1: Number of aperiodic strings for small alphabets. σ is the size of the alphabet,
and µ(j, σ) is the number of aperiodic strings of length j
3
Main results
Our main results are formulated in the following theorems.
Theorem 11. On an alphabet of size σ for all n ≥ 2k, Ω(n, k, σ) ≤ φ(k, σ) for some
function φ.
Theorem 12. There is a c > 0 and an n0 such that for any n > n0 the following is
true. Let S ′ be a string of length n − 1, and S be a string obtained from S ′ by adding a
character to its beginning chosen uniformly random from the alphabet. Then the expected
growth of S is at least c · n.
Theorem 13. There is a d > 0 that for any n > n0 (where n0 is the same as in
Theorem 12) the following holds. On an alphabet of size σ the simple suffix tree of a
random string S of length n has at least d · n2 nodes in expectation.
5
4
Proofs
Proof. (Theorem 13)
Considering Observation 8 we have that the expected size of the simple suffix tree of a
random string S is at least
E
n
X
(γ(S[n − m, n])) ≥
m=1
n
X
(1)
E(γ(S[n − m, n])).
m=1
If m ≤ n0 , Theorem 12 is obvious. If m > n0 , we can divide the sum into two parts:
n
X
E(γ(S[n − m, n])) =
m=1
n0
X
E(γ(S[n − m, n])) +
n
X
E(γ(S[n − m, n])).
(2)
m=n0 +1
m=1
The first part of the sum is a constant, while the second part can be estimated with
Theorem 12:
n
n
X
X
E(γ(S[n − m, n])) ≥
cn = d · n2 .
(3)
m=n0 +1
m=n0 +1
This proves Theorem 13.
First, we show a few lemmas about the number of aperiodic strings. Lemma 14 can be
found in [9] or in [5], but we give a short proof also here.
Lemma 14. For all j > 0 integer and for all alphabet of size σ the number of aperiodic
strings is
X
µ(d, σ).
(4)
µ(j, σ) = σ j −
d|j
d6=j
Proof. µ(1, σ) = σ is trivial.
There are σ j strings of length j. Suppose that a string is periodic with minimal period
d. This implies that its first d characters form an aperiodic string of length d, and there
are µ(d, σ) such strings. This finishes the proof.
Specially, if p is prime, then µ(p, σ) = σ p − σ.
t
Corollary 15. If p is prime and t ∈ N, then µ (pt , σ) = σ p − σ p
size σ.
6
t−1
for all alphabet of
t
Proof. We count the aperiodic strings of length pt . There are σ p strings. Consider the
minimal period of the string, i.e. the period which is aperiodic. If we exclude all minimal
periods of length k, we exclude µ(k, σ) strings. This yields the following equality:
X
t
µ pt , σ = σ p −
µ (ps , σ) .
(5)
1≤s<t
With a few transformations and using Lemma 14, we have that (5) is equal to
X
X
X
t
t
t−1
µ (ps , σ) −
µ (ps , σ) , (6)
σ p −µ pt−1 , σ −
µ (ps , σ) = σ p −σ p +
1≤s<t−1
1≤s<t−1
1≤s<t−1
which is
t
σp − σp
t−1
(7)
.
Lemma 16. For all j > 1 and for all alphabet of size σ , µ(j, σ) ≤ σ j − σ.
Proof. From Lemma 14 we have µ(j, σ) = σ j −
P
µ(d, σ). Considering µ(d, σ) ≥ 0 and
d|j
d6=j
µ(1, σ) = σ, we get the claim of the lemma.
Lemma 17. For all j ≥ 1, and for all alphabet of size σ
µ(j, σ) ≥ σ(σ − 1)j−1 .
(8)
Proof. We prove by induction. For j = 1 the claim is obvious, as µ(1, σ) = σ.
Suppose we know the claim for j − 1. Consider σ(σ − 1)j−2 aperiodic strings of length
j − 1. Now, for any of these strings there is at most one character by appending that to
the end of the string we receive a periodic string of length j. Therefore we can append
at least σ − 1 characters to get an aperiodic string, which gives the desired result.
Observation 18. Observe that if the growth of S is k, then there is a j such that
S[1, n − k] = S[j + 1, j + n − k]. For example, if the string is abcdef abcdab (n = 12),
one can check that the growth is 8 (the new branch in the suffix tree which ends in leaf
1 starts after abcd), and with j = 6 we have S[1, 4] = S[7, 10] = abcd.
The reverse of this observation is that if there is a j < n such that S[1, n − k] =
S[j + 1, j + n − k], then the growth is at most k, as S[j + 1, n] and S[1, n] shares a
common prefix of length n − k, thus, the paths to the leaves j + 1 and n share n − k
internal nodes, and at most k new internal nodes are created.
7
Proof. (Theorem 11) For proving the theorem we count the number of strings with
growth k for n ≥ 2k.
First, we fix j, and then count the number of possible strings where the growth occurs
such that S[1, n − k] = S[j + 1, j + n − k] for that fixed j. Note that by this way, we only
have an upper bound for this number, as we might found an ℓ such that S[1, n − k + 1] =
S[ℓ + 1, ℓ + n − k + 1].
We know that j ≤ k, otherwise S[j + 1, j + n − k] does not exist.
If j = k, then we know S[1, n − k] = S[k + 1, n].
S[1, k] must be aperiodic. Suppose the opposite and let S[1, k] = p . . . p, where p is the
minimal period, and its length is d. Then S[k + 1, n] = p . . . p. Obviously, in this case
S[1, n − d] = S[d + 1, n], which by Observation 18 means that the growth would be at
most d. See also Figure 3.
Therefore this case gives us at most µ(k) strings of growth k.
p
p
1
p
p
p
p
n
k
Figure 3: Proof of Theorem 11, case j = k
If j < k, then we have S[1, n − k] = S[j + 1, j + n − k].
First, we note that S[1, j] must be aperiodic. Suppose the opposite and let S[1, j] =
p . . . p, where p is the minimal period, and its length is d. Then
S[j + 1, 2j] = S[2j + 1, 3j] = . . . = p . . . p,
(9)
k
k
· j = S j + 1, j +
· j = p . . . p.
S 1,
j
j
(10)
which means that
This implies that S[1, j + n − k] = p . . . pp′ , where p′ is a prefix of p. However, S[1, j +
n − k − d] = S[d, j + n − k] is true, and using Observation 18, we have that γ(S) ≤
n − (j + n − k) + d = k − j + d < k, which is a contradiction.
Further, S[j + n − k + 1] must not be the same as S[k + 1], which means that this
character can be chosen σ − 1 ways.
Therefore this case gives us at most µ(j)(σ − 1)σ k−j−1 strings of growth k for each j.
By summing up for each j, we have
8
p
p
p
p
p
p
j
1
p
p
p′
j+n−k
k
n
Figure 4: Proof of Theorem 11, case j < k
φ(k, σ) =
k−1
X
µ(j, σ)(σ − 1)σ k−j−1 + µ(k, σ)
(11)
j=1
This completes the proof.
Proof. (Theorem 12)
According to Lemma 16, µ(j, σ) ≤ σ j − σ (if j > 1).
In the proof of Theorem 11 at (11) we saw for k ≥ 1 and n ≥ 2k − 1 that
φ(k, σ) = µ(k, σ) +
k−1
X
µ(j, σ)(σ − 1)σ k−j−1.
(12)
j=1
We can bound the right hand side of (12) from above as it follows:
µ(k, σ)+
k−1
X
µ(j, σ)(σ−1)σ
k−j−1
= µ(k, σ)+µ(1, σ)(σ−1)σ
k−2
+
j=1
k−1
X
µ(j, σ)(σ−1)σ k−j−1,
j=2
(13)
which is by Lemma 16 at most
k
σ −σ + σ(σ −1)σ
k−2
+
k−1
X
j
(σ −σ)(σ −1)σ
k−j−1
j=2
k
k
≤ σ +σ +
k−1
X
σ j σσ k−j−1 ≤ kσ k . (14)
j=2
Thus, φ(k, σ) ≤ kσ k , which means
m
X
k=1
φ(k, σ) ≤
m
X
kσ k ≤ (m + 1)σ m+1 .
(15)
k=1
The left hand side of 15 is an upper bound for the strings of growth at most m.
Let m = n2 .
n
As σ n ≫ n2 σ 2 , this implies that in most cases the suffix tree of S has at least
nodes than the suffix tree of S[1, n − 1].
Thus, a lower bound on the expectation of the growth of S is
9
n
2
more
1 n n n n n n
2 +
2
σ
σ
+
1
,
σ
−
σn 2
2
2
n
1 n+2 n
n n(n + 2)
2
σ
σ
+
−
= cn,
σn
2
2
4
E (γ(S)) ≥
which is
(16)
(17)
with some c, if n is large enough.
With this, we have finished the proof and gave a quadratic lower bound on the average
size of suffix trees.
References
[1] E.F. Adebiyi, T. Jiang, and M. Kaufmann. An efficient algorithm for finding short
approximate non-tandem repeats. Bioinformatics, 17:5S–12S, 2001.
[2] A. Apostolico, M. Crochemore, M. Farach-Colton, Z. Galil, and S. Muthukrishnan.
40 years of suffix trees. Communications of the ACM, 59:66–73, 2016.
[3] A. Blumer, A. Ehrenfeucht, and D. Haussler. Average sizes of suffix trees and
DAWGs. Discrete Applied Mathematics, 24:37–45, 1989.
[4] M.T. Chen and J. Sciferas. Efficient and elegant subword tree construction. In
Combinatorial algorithms on words, pages 97–107. Springer-Verlag, 1985.
[5] J.D.
Cook.
Counting
primitve
bit
strings.
http://www.johndcook.com/blog/2014/12/23/counting-primitive-bit-strings/,
2014. [Online; accessed 02-May-2016].
[6] L. Devroye, W. Szpankowski, and B. Rais. A note on the height of suffix trees.
SIAM Journal on Computing, 21:48–53, 1993.
[7] E. R. Fiala and D. H. Greene. Data compression with finite windows. Communications of the ACM, 32:490–505, 1989.
[8] C. Fraser, A. Wendt, and E.W. Myers. Analyzing and compressing assembly code.
In Proceedings SIGPLAN Symposium on Compiler Construction, pages 117–121,
1984.
[9] E.N. Gilbert and J. Riordan. Symmetry types of periodic sequences. Illinois Journal
of Mathematics, 5:657–665, 1961.
10
[10] R. Grossi and G.F. Italiano. Suffix trees and their applications in string algorithms.
In Proceedings of the 1st South American Workshop on String Processing, pages
57–76, 1993.
[11] D. Gusfield. Algorithms on Strings, Trees and Sequences. Cambridge University
Press, 1997.
[12] M. Höhl, S. Kurtz, and E. Ohlebusch. Efficient multiple genome alignment. Bioinformatics, 18:312S–320S, 2002.
[13] L. Kaderali and A. Schliep. Selecting signature oligonucleotides to identify organisms
using DNA arrays. Bioinformatics, 18:1340–1348, 2002.
[14] E. M. McCreight. A space-economical suffix tree construction algorithm. Journal
of the ACM, 23:262–272, 1976.
[15] L. O’Connor and T. Snider. Suffix trees and string complexity. In Advances in
Cryptology: Proceedings of EUROCRYPT, LNCS 658, pages 138–152. SpringerVerlag, 1992.
[16] M. Rodeh. A fast test for unique decipherability based on suffix trees,. IEEE
Transactions on Information Theory, 28(4):648–651, 1982.
[17] S.L. Tanimoto. A method for detecting structure in polygons. Pattern Recognition,
13:389–494, 1981.
[18] E. Ukkonen. On-line construction of suffix trees. Algorithmica, 14:249–260, 1995.
[19] P. Weiner. Linear pattern matching algorithms. In Proceedings of the 14th IEEE
Symposium on Switching and Automata Theory, pages 1–11, 1973.
11
| 8cs.DS
|
Decision-level multi-method fusion of spatially scattered data
from nondestructive inspection of ferromagnetic parts
René Heideklanga, Parisa Shokouhib
a BAM
Federal Institute for Materials Research and Testing, Unter den Eichen 87, 12205 Berlin, Germany
[email protected]
phone: ≈49 30 8104-1838 (secretariat of Division 8.5“
b Department
of Civil and Environmental Engineering, The Pennsylvania State University, 215 Sackett
Bldg., University Park, PA 16802
[email protected]
phone: ≈1 (814“ 863-0678
Corresponding author: René Heideklang
Sensor 1
Sensor 2
Spatial
Density
Agreement
Detection
Crack
This work is licensed under a Creative Commons Attribution-NonCommercial-NoDerivatives 4.0
International License: http://creativecommons.org/licenses/by-nc-nd/4.0/
Abstract
This article deals with the fusion of flaw detections from multi-sensor nondestructive materials testing.
Because each testing method makes use of different physical effects for defect localization, a multimethod approach is promising to effectively distinguish the many false alarms from actual material
defects. To this end, we propose a new fusion technique for scattered two- or three-dimensional location
data. Using a density-based approach, the proposed method is able to explicitly address the localization
uncertainties such as registration errors. We provide guidelines on how to set all key parameters and
demonstrate the technique s robustness. Finally, we apply our fusion approach to experimental data and
demonstrate its ability to find small defects by substantially reducing false alarms under conditions where
no single-sensor method is adequate.
Keywords: multi-sensor data fusion; density estimation; scattered data; defect detection; nondestructive
testing; registration error;
1 Introduction
Industrial nondestructive testing (NDT“ refers to the inspection of materials, parts or structures concerning
their condition without compromising their integrity. NDT experts employ different sensors depending on
the material and the expected types of defects. Near-surface cracks represent one class of flaws that are
commonly encountered in critical machine parts under dynamic loading such as turbine blades and
bearings. Repeating cycles of dynamic stress promote the growth of invisible micro cracks into larger
defects that compromise the safety and shorten the service life of the component. As such, early detection
of cracking is critical to the safe operation and serviceability of flawed machinery. Among the NDT
methods that qualify for the task of near-surface crack detection, special attention is paid to those that
allow automatic data acquisition and provide accurate, objective and reproducible results. Common
candidate methods are ultrasonic testing (UT“, eddy current testing (ET“, magnetic flux leakage (MFL“
testing, and thermal testing (TT“. However, single-method inspection is often ambiguous. This is because
each method reacts to changes in specific physical properties of the tested material in the presence of a
defect. But, harmless material or geometrical characteristics may also produce similar and often
indistinguishable changes in the recorded signals. For instance, MFL is affected by surface roughness and
inhomogeneous magnetic properties, and TT s high sensitivity to negligible manufacturing and handling
marks and imperfections results in a multitude of false positive indications. These non-defect indications
are referred to as structural noise, because unlike measurement noise, they are deterministic across
repeated measurements. However, structural noise does vary across different inspection techniques,
whereas actual flaws are characterized by multi-sensor agreement. This concept lies at the heart of NDT
data fusion. On other words, multi-sensor inspection differentiates flaw indications from structural noise,
thus providing more reliable assessment.
Sensor fusion can be performed at various levels of signal abstraction [1], each with specific drawbacks
and advantages. In particular, decision-level fusion deals with the higher level aggregation of data after
individual detection. That is, each signal is first processed individually, and then fed into the fusion
algorithm. We argue that decision-level fusion has several advantages over signal-level fusion for NDT
inspection. First, unlike signal-level fusion, it does not require obtaining individual signal values at the very
same locations by interpolation. Although a spatial association step is still necessary to prepare individual
detections for fusion, decision-level fusion can work on the un-interpolated original data. As we will
demonstrate here, this approach allows for explicit handling of localization uncertainty, for instance due to
registration errors. This is in contrast to signal-level fusion, where accurate registration is crucial [2, 3]
because errors can hardly be compensated. Another advantage of high-level fusion for defect detection is
the two stage process: the measurements are sifted once for flaws as in traditional single-sensor
inspection, and a second time after fusion. This process promises to effectively eliminate many of the false
alarms that result from structural noise. A further major benefit is the strategy s modularity. The individual
data collection and processing can be carried out independently by the respective experts to tailor the
detection process specifically to each inspection method. One practical benefit of the modularity offered by
decision-level fusion is that it allows combining individual results, even if fusion was not envisioned in the
original testing plan. This also facilitates independence from the type of data source, making it possible to
aggregate heterogeneous modalities ranging from manual inspection data to the output of fully automated
scanning systems. Consequently, different sources of information can be effortlessly exchanged and the
fusion strategy is readily adapted to unknown future changes of input sources.
Although there are recent studies in NDT proposing decision-level fusion, e.g. by weighted averaging [4],
hypothesis testing [5], and Bayesian or Dempster-Shafer theory [6, 7], all of these works still rely on image
registration and interpolation to perform fusion at common grid points. In fact, we are not aware of any
data fusion publication in the field of NDT dealing with scattered decision-level fusion, i.e. using the
original measurement locations, despite the aforementioned advantages.
We propose a new fusion strategy that combines spatially scattered detection locations to improve the
detection performance compared to single-source methods. In this paper, we will address the problem of
nondestructive surface inspection, although the methodology is quite generic and can be easily extended
to the three-dimensional case of volume inspection. Our approach is detailed in section 2 where the
general idea is introduced using a simple example. Subsequently, we formalize the problem, describe our
method and provide practical techniques for automatic parameter estimation. The developed approach is
then applied to experimental multi-sensor NDT measurements by three different inspection methods.
Finally, we quantitatively demonstrate the enhanced performance of multi-sensor crack detection using
our technique over single-sensor inspection.
2 Methodology
Per-sensor detection is done after individual pre-processing, using simple rules such as thresholding. The
result of individual defect localization is a set of detections = , , also termed hits, scattered across
the specimen's surface, each exhibiting a certain signal to noise ratio (SNR“. For example, consider
Figure 1 for an outcome of individual surface inspection using two NDT methods. In cases a to d, each dot
marks a hit according to some detection rule per sensor. Among the hits, there are also false alarms, for
instance indicating changes in material properties unrelated to a defect (structural noise“. Such false
alarms are illustrated by cases b and c in Figure 1. Using single-sensor inspection, these false alarms
cannot be distinguished from indications produced by actual defects such as those shown in a. A multisensor data set, on the other hand, is able to reveal a as a real defect by assessing the agreement among
different detection methods. Here, agreement is expressed in terms of joint spatial density of hits, taking
into account all sensors. Our rationale is that the joint hit density is higher over real defects than in other
areas, provided sufficient SNR for at least two sensors. On the contrary, a clear conflict occurs where only
one sensor generates hits, and thus the joint density is not significantly increased relative to the individual
sensor density. This concept is depicted at the bottom of Figure 1, where for each sensor the spatial
density of hits across the specimen surface is symbolized. Only in case a both sensors agree in increased
spatial density, whereas in cases b and c the sensors do not agree. Although there is also agreement in
case d as well, the joint density is not significant enough, indicating the low likelihood of defect presence.
This example demonstrates the potential of the joint spatial density as a measure for multi-sensor
detection.
a)
b)
c)
d)
Sensor 1
Sensor 2
spatial density
likelihood of
true defect
true defect
sensor agreement
false alarm
sensor conflict
false alarm
sensor conflict
true non-defect
sensor agreement
Figure 1: Schematic representation of the principle of the approach. The detection outcome of two different NDT
methods are represented by circles and crosses, for four cases a“-d“. For each case, the corresponding spatial
densitiy per sensor is plotted below. The likelihood of observing a true defect (gray area“ depends on both sensors
yielding significant hit densities.
We propose evaluating the local hit density as a measure for multi-sensor data fusion at decision level.
The first step is detection per individual sensors. The initial detection rules must focus on sensitivity rather
than specificity, thus ensuring that all (unknown“ real defect indications are retained for the final detection
by fusion. At the same time, we would still want to discard as many false alarms as possible. For example,
most of the typical signal intensities are near zero. Therefore, a simple thresholding operation significantly
reduces the number of potential defect indications per sensor, and thus simplifies the subsequent data
fusion efforts. Since the detection method is independent from the proposed fusion approach, the most
suitable detection criterion for each sensor can be used.
Moreover, the locations returned by the NDT sensors have to be related, if they are not defined with
respect to the same coordinate system. Rather than measuring at the very same locations on the
specimen, it is often more practical to mathematically register the individual local coordinate systems after
the measurements were carried out. Therefore, for each sensor pair, we find corresponding locations in
the sensor data and fit a coordinate transformation model that minimizes the distance between each pair
of corresponding locations. Note that, in contrast to fusion at the signal level, this transformation is not
used here to interpolate the sensor data values. Rather, for decision-level fusion, the transformations
allow to express the locations of the individual detections in any of the sensors coordinate systems. We
note that although each NDT method usually uses gridded measurement positions, the mapped hit
locations after registration are not jointly gridded in the common coordinate system, but appear scattered
instead.
One central challenge in the decision-level fusion of non-gridded locations is the uncertainty in
localization. Two main factors contribute to this uncertainty. First, each sensor s localization ability is
limited by the physical resolution as well as the spatial sampling rate. Second, the coordinate
transformations computed during spatial registration are inaccurate to some degree, because the
manually set point pairs do not perfectly match or the data do not fit the transformation model well enough.
To be robust, a fusion approach must adequately cope with the inherent uncertainties about hit positions
and must associate nearby points for the purpose of density quantification. This loose concept of
proximity must therefore be mathematically formalized.
To this end, various non-parametric techniques have been developed such as Mean shift [8], DBSCAN
[9], OPTICS [10], Spectral clustering [11], and Kernel density estimation (KDE“ [12, 13]. We have selected
the framework of kernel density estimation for this study. This choice was motivated by considering that
our data space typically has only two or three spatial dimensions, independent from the number of
sensors. Therefore, the density can be directly modeled without being affected by the curse of
dimensionality. Furthermore, it allows us to evaluate the density at arbitrary positions, not only at the hits.
Next, we will formally introduce our method using ideas from KDE.
2.1 Kernel density estimation (KDE)
KDE is a nonparametric statistical method to estimate a probability density function
samples
from a set of
. The result is a continuous function, computed from a weighted sum of kernel functions
with
an associated bandwidth ℎ, each centered over one of the samples: = ∑ ∑
=
. Some functions qualifying as a kernel
are the uniform, triangle, Gaussian or
−
, with
Epanechnikov kernel functions. The bandwidth ℎ controls the size of the neighborhood in which each
sample influences the density. The choice of the bandwidth is critical for the overall performance of the
algorithm. If the bandwidth is chosen too wide, KDE results in an overly smoothed density, thus losing
important details of the distribution. On the other hand, if it is chosen too narrow, the estimate adapts too
much to the specific realization of the sample set, thus missing the global features of the density. This
problem has been well-studied, and several solutions have been proposed [14]. We will describe how to
automatically compute a suitable bandwidth for our problem in section 2.3.
The general formulation given above for KDE includes the normalization constant ∑ , ensuring that
the density integrates to one. We proceed with a simpler unnormalized version of KDE by dropping the
normalization constant. Furthermore, since the density estimate is a weighted sum with one term per data
point, we can partition the data set and aggregate the total density function
which include only the samples
from each partition :
= =
∈ !
−
from the sub-densities ,
" denotes a subset of all points such that partitions do not overlap and the union of all partitions covers
the complete data set. This re-arrangement is taken up in the following section to group detections by
each sensor.
KDE can be extended to vector-valued samples. To this end, let #$ denote the ith vector-valued (bold face“
sample, and let % denote the vector-valued location where to evaluate the density. Multivariate KDE is
computed from multivariate kernel functions and an associated bandwidth matrix &, which describes the
scale and the orientation of the kernels. A special kind of multivariate kernel is the product kernel, defined
by
' #
= ∏
!
) +, where one univariate kernel for each dimension is evaluated. The ℎ are the
#*
!
entries of the diagonal bandwidth matrix, i.e. product kernels are not arbitrarily oriented in the data space.
We use product kernels in our approach, which is described next.
2.2 Scattered decision-level fusion
In this section, we develop a new fusion method for our NDT problem based on concepts from KDE. Here,
the role of the vector-valued data sample #$ is taken by the two-dimensional hit location as detected by
a single sensor , during surface inspection. The fused density for defect detection
is computed as:
- = .
/ -
/∈0/1 ,…/3 4
= :
∈>/
/ -5
−
;< =/ -
max
/∈0/1 ,…/3 4
−
/ -
(1“
(2“
C
C D
A1 − ) + E F1 − A E G ,
ℎ
ℎ
; ? = @
D
0,
|C | ≤ ℎ JK LC L ≤ ℎ
NOPN
(3“
In Equation (1“, - is the estimated density evaluated at arbitrary location - = Q , Q on the specimen
surface. This density is computed from the sum of all per-sensor densities
regards detections = , ∈ R, that were found by sensor ,.
/ -,
each of which only
A key element of our method is to explicitly penalize situations in which only a single sensor makes the
major contribution to the sum of densities: for each evaluation point -, we subtract the maximum single-
sensor KDE at this location. This step realizes the quantification of agreement among sensors, because a
large fused score now requires at least two sensors with high individual densities. Thus,
is expected to
behave similarly to the function indicated by the shaded area in Figure 1. Note that in the case of only two
sensors, subtracting the maximum is equal to the minimum fusion rule, which is a fuzzy AND-operator,
and is in fact the operation used to generate the shaded area in the figure. However, as more than two
sensors are involved, requiring that all sensors indicate a defect seems too strict. Therefore, ignoring the
maximum contribution can be viewed as a much milder version of the AND fusion rule.
To
compute the per-sensor densities
/ -,
the point of interest - is mapped to the local coordinate system
through the coordinate transform =/ , where the distance between - and each detection is computed as
shown in Equation (2“. This measure of proximity is converted to a measure of density by the per-sensor
kernel function
;< . Although in principle, the use of any kernel type is valid, we suggest using a
compactly supported kernel function like the Epanechnikov product kernel as defined in Equation (3“. The
compact support has the advantage of limiting the spatial influence area of each detection, which is
expressed by the kernel bandwidth parameters ; = ℎ , ℎ , and thus facilitates faster computation than
e.g. the non-vanishing Gaussian kernel. We further adapt the kernel function
;
so that instead of
integrating to unity, its maximum is normalized to one. The rationale behind this choice is discussed in
view of automatic bandwidth selection in the following section. In addition, note that instead of
transforming the individual detection locations to the coordinate system of -, we follow the converse
approach to map the density evaluation locations - to each sensor s system. This simplifies density
estimation, because in each local system, the respective kernels are axis-aligned, as will be explained in
the next section. This facilitates using a multivariate product kernel, in contrast to general kernels with
non-diagonal bandwidth matrix. Figure 2 visualizes the role of the coordinate transformation =S in our
method.
To further adjust the quantification of density, we scale the kernel functions according to the weight : per
detection. These weights control the influence of each detection on the final KDE. We set each weight
proportional to the detection s signal to noise ratio, so that clear indications have more impact on the final
density than insignificant ones. Also, the weights offer additional flexibility to regulate the fusion result with
regards to specific sensors or different inspection areas.
In total, our fusion approach includes three mechanisms to ensure robustness against false alarms:
quantification of density, subtraction of the maximum single-sensor contribution, and decision weighting
according to significance.
VW
a)
VWX
U;
b)
U;
p
global coordinate system
sensor‘s local system
sensor‘s local system
global coordinate system
Figure 2: Coordinate transformation during the computation of the fused density. The two black coordinate systems in
a“ and b“ are related through the coordinate transform =S .
a“ Coordinate system in which the fused density will be evaluated at gridded points -. The coordinate systems of the
individual sensors, where the detections are defined, are not axis-aligned with the global system. In particular, the
kernels
would require non-diagonal bandwidth matrices.
b“ Coordinate system of one of the sensors, given by its measurement grid. For single-sensor density estimation,
kernels
are axis-aligned to the sensor s system, thus facilitating product kernels. The single-sensor density is then
evaluated at the transformed points =/ -.
2.3 Automatic bandwidth estimation
We can use our background knowledge about the nature of our data to deduce the minimum required
values for the bandwidth. Intuitively, the density estimator should always be able to smoothly interpolate
between neighboring data points. For NDT data, the smallest possible distance between any two
detections of the same sensor is given by the known spatial sampling intervals. For a measurement grid
per sensor ,, the two spatial sampling distances are denoted by Δ/ , Δ/ in the sensor s coordinate system.
To ensure that neighboring line scans crossing the same defect do not form disconnected density peaks,
the kernel functions should at least stretch across one pixel in the sensor image. However, to avoid
merging two unrelated indications, the kernels should not be made much larger1. Therefore, we propose
to use product kernels with minimum bandwidth parameters of ;S = ℎ , ℎ = Δ/ , Δ/ for each sensor. It
is natural to use product kernels for gridded individual measurements, because the bandwidths directly
correspond to the physical pixel dimensions. Note that, in conventional KDE, broader kernels are
necessarily flatter to maintain unit integral. We, in contrast, fix the kernel s peak value to one,
independently from its spatial extent. By following this procedure, each sensor contributes equally to the
joint density, regardless of its spatial localization characteristics. Figure 3 illustrates this concept.
The aforementioned kernel size is a minimal setting. In practice, the most significant factor contributing to
the localization uncertainty may not be the spatial sampling, but inevitable registration errors. The kernel
sizes should be set large enough to smooth out these unwanted variations and to associate poorly
registered indications. As a general approach, we propose the following kernel size ;S = ℎ , ℎ =
max0Δ/ , CY4, maxZΔ/ , CY[, where CY denotes an estimate of the localization uncertainty, for instance the
mean registration error. However, note that in this approach, the advantage of having spatially accurate
sensors may be lost. Also, with increasing kernel size, closely situated defects become harder to
separate. Therefore, our density-based approach will substantially benefit from high-quality registration by
facilitating small kernel sizes.
1
ℎ/
1
ℎ/D
1
ℎ/\
Figure 3: Kernel functions with automatically computed bandwidths, for three sensors with different spatial sampling
distances, e.g. in mm. For clarity, only one spatial dimension is shown per sensor. Registration errors also influence
the kernel widths, but this is ignored in the depiction for clarity. Moreover, if data weights were regarded, the kernel
amplitudes would be scaled.
1
In fact, some NDT methods might have coarse spatial localization ability due to physical limitations,
regardless of the spatial sampling rate. Those sensors are characterized by spatially slowly varying
signals. Because in such cases, the smallest possible distance between any two detections is larger than
one pixel, it is sensible to increase the kernel size.
3 Application to experimental data
To demonstrate the fusion technique s performance under realistic conditions, a test specimen containing
15 surface flaws was inspected using three different NDT methods. In this section, we describe the
specimen, the individual data collection and processing as well as the application of our fusion algorithm.
Finally, we quantitatively evaluate the effect of various conditions on the fusion result, and compare fused
detection against single-sensor detection.
3.1 Specimen
The ring-shaped test specimen is a bearing shell [15, pp. 173-175] made of surface-hardened steel. As
illustrated in Figure 4, it has an outer diameter of 215 mm and is 73 mm long in its axial direction. Cracks
are simulated by 15 machined grooves that are regularly spaced across the surface. These grooves were
created by electrical discharge machining and vary in depth from 11 to 354 µm, as detailed in Table 1.
Grooves have constant lengths of 1 mm and their openings vary between 25 µm and 51 µm. The
specimen s surface is uncoated and its roughness is very low, thus enabling high-quality near-contact
measurement.
Figure 4: Schematic view of the ring specimen, not to scale. Top: outer proportions of the ring. Bottom: unrolled outer
surface. Short vertical lines indicate the positions of the 15 grooves.
Groove
1
2
3
4
5
6
7
8
9
10
11
12
13
14
15
Depth / µm
354
224
170
105
82
61
57
53
43
40
39
27
29
20
11
Table 1: Groove depths. Labels correspond to those shown in Figure 4.
3.2 Individual measurements and processing
For crack detection, thee different inspection techniques were used that are well-suited for the automated
nondestructive evaluation of cracks in ferromagnetic metals. The first method is called eddy current testing
(ET“. An excitation coil is run over the specimen surface. Using this coil, an alternating current induces
circular eddy currents in the specimen s near-surface region. These currents are blocked by the presence
of defects, thus affecting the impedance of the probe coil, which is the measured signal. The second
method employed here is the laser-induced active thermography testing (TT“. A high-power laser is run
across the specimen to locally heat up the surface. In defect-free regions, the heat flow is able to
dissipate, whereas defects cause localized heat accumulation. An infrared camera monitors the heat flow
decay and generates a digital image sequence for processing. The third method is magnetic flux leakage
testing (MFL“. The specimen is exposed locally to a static magnetic field, which spreads inside the
ferromagnetic material. At near-surface defects, the field is forced to leak out of the specimen into the
air, although air has lower magnetic permeability than the specimen. This leakage can be detected by
magnetic field sensors, such as giant magneto resistance (GMR“ sensors. The following three inspections
were performed sequentially during the course of about one year.
ET was carried out at an excitation frequency of 500 kHz. An automated scanning device rotates the
specimen under the fixed probe. Signal processing is based only on the imaginary part of the measured
impedance. The obtained one-dimensional signals are preprocessed by high-pass filtering for trend
correction, and by low-pass filtering to improve SNR. An image is formed by stacking the line scans in
axial direction of the ring.
The MFL data were collected using the same scanner as for ET, and a GMR sensor array developed at
BAM [16]. Using these gradiometers, the normal component of the magnetic stray field was measured
while the specimen was locally magnetized. Preprocessing comprised trend correction by high-pass
filtering per line scan, and an adaptive 2D wiener filter2 for noise suppression. The image was then Sobelfiltered to highlight the steep flanks that are generated by the gradiometers over the grooves.
Thermography testing was performed by rotating the specimen under a 40 W powered laser while
recording with an infrared camera. The movie frames were then composed to form an image of the
specimen surface. This image is processed by 2D background subtraction using median filtering, and
noise was suppressed by an adaptive 2D wiener filter.
We note that the presented signal acquisition is tailored to the known groove orientation. In a realistic
setting, we would have to perform a second scan for ET and MFL testing to detect any circumferentially
oriented defects as well. We further emphasize that for our specific ring specimen, the GMR sensors yield
far superior results compared to ET and TT, and would suffice by themselves for surface crack detection.
However, such performance is not guaranteed for other materials or in case of suboptimal surface
conditions, so that a multi-method approach is still in demand. To take advantage of as many independent
sources of information as possible during fusion, we intentionally lowered the quality of the MFL image
before preprocessing and detection. This was done by separating the true defect indications from the
2
We used MATLAB s function wiener2. See http://de.mathworks.com/help/images/ref/wiener2.html
background signal variations using the shift-invariant wavelet transform [17], and by reconstructing the
signal with a factor of 0.1 for the noise-free component. Although this does not simulate a lower-quality
MFL measurement in a physically realistic way, the approach allows making use of the acquired GMR
signals as a third source of information. We justify this alteration in favor of demonstrating the capabilities
of our fusion technique in other settings where individual inspection is in fact not reliable enough.
Therefore, for the rest of the article, we consider only this modified version of the MFL data.
To convey an impression of the signals, we show an exemplary portion of each preprocessed inspection
image in Figure 5. The displayed part of the specimen surface is a 10 mm by 6.5 mm region around
groove nr. 13 which is quite shallow, and thus generates relatively weak indications. The figure
demonstrates the different signal patterns among sensors, concerning both the groove and the
background variations. Also, the different pixel sizes are evident. A related plot is shown in Figure 6,
where one-dimensional line scans crossing the groove reveal more clearly the individual sensor
responses. The different spatial sampling positions are demonstrated by the line markers. Table 2 offers a
quantitative comparison of the individual data sets.
After individual preprocessing, the same detection routine was used for all three images to extract hit
locations and confidences, which will later be fused at the decision level. To this end, we convert the
signal intensities to confidence values and apply a threshold to extract only significant indications.
Confidence values are computed by estimating the distribution function of background signal intensities
from a defect-free area. This estimate serves as the null distribution "]^__ in the significance test. For
each image pixel s intensity , the probability "]^__ ≥ is computed as the confidence. Pixels are
considered significant here, if their confidence exceeds 99 %. Additionally, only those hits that are local
maxima with regard to their neighboring pixels along the horizontal axis (thus crossing the grooves“ are
retained. This constraint further filters many false alarms while making the detection results invariant to
different peak widths. After registration to a common coordinate system based on manually defined
location correspondences in the data, the final set of hits from all sensors is plotted in Figure 7. Obviously,
the false alarms considerably outnumber the actual groove detections. This is due to our sensitive
detection rules, intending that no actual defect is missed during individual processing. Of course, in a
single-sensor inspection task, a much more stringent detection criterion is appropriate to limit the number
of false alarms. However, this possibly leads to worse sensitivity to small material flaws. In contrast, our
data fusion approach is supposed to discard most false hits while maintaining high sensitivity to small
defects.
We will now briefly compare the individual sensor results. In contrast to ET and TT, the MFL hits cluster
spatially. This is because the background variations in this data set are not homogeneous, possibly due to
inhomogeneities in the internal magnetic field. MFL data are missing in the strip between the two groove
rows. Interestingly, in Figure 7, spatial defect-like patterns are formed, although the specimen is not
expected to contain any flaws other than the known grooves. For example, see the vertical lines from TT
(e.g. indicated by the right arrow“, or the diagonally oriented lines from ET indicated by the left arrow. As
previously discussed, using individual inspection, it is not easy to classify these obvious indications as
structural indications or flaws. In spite of their regular structure, we label these patterns as non-defect
indications during the following evaluation, if our multi-sensor data set is not able to give a reliable
confirmation. On the other hand, there are a few off-groove locations where different sensors behave
consistently. These regions could in fact represent unknown but real defects, and will therefore be
excluded from the following evaluation. Note that detections within disregarded areas are not shown in this
figure. Moreover, the confidence associated with each detection is not shown, because all hits exceed the
chosen threshold of 99 % as explained before.
Figure 5: Preprocessed sensor intensity images, zoomed to a region around the groove nr. 13. Higher intensities
correspond to indications. Top: ET, Middle: MFL, Bottom: TT. The vertical black line marks the location of the groove.
Each image is shown in the respective sensor s coordinate system, thus explaining the different axis labels.
Figure 6: Preprocessed line scan per inspection method around groove nr. 13. The signals are shifted so that each
peak value is located at
= 0. Note the different y scales.
Table 2: Quantitative properties of the individual data sets
ET
MFL
TT
0.029
0.029
0.469
0.200
0.200
0.126
indication, in mm
2 69Δ
0.6 20Δ
Avg. nr. of hits per pixel
0.0023
0.0031
Δ in mm
Δ in mm
Width of a typical
0.5 Δa
0.0068
Figure 7: Hit locations per sensor in a common coordinate system similar to that in Figure 4. Axes are not to scale.
The tips of the triangle markers indicate the groove positions. The arrows indicate crack-shaped false positives from
ET (left arrow“ and TT (right arrow“. A colored version of this figure appears in the article online.
3.3 Fusion and final detection
To compute the kernel density per sensor, we used Alexander Ihler s KDE Toolbox for MATLAB [18]. The
fused density, according to Equations (1“-(3“, is a continuous function that must be evaluated at discrete
locations. In fact, to circumvent the discrete sampling, a multivariate mode-seeking algorithm could be
used for detection. However, for simplicity, we set up a discrete evaluation grid that is designed fine
enough to not miss any mode of the density. Modes are then traced similarly to per-sensor detection by
finding local density maxima along parallel lines on the specimen surface. Figure 8 displays the principle
of the detection. Finding local maxima along one-dimensional lines is straightforward due to the density s
smoothness, and also makes the detection results more stable across different kernel sizes.
Figure 8: Final detection by evaluating the fused continuous density function at gridded points (circles“ and finding
local maxima (filled circles“ among them. These detections approximate the true density modes (dashed lines“.
The final detections after fusion are presented in Figure 9. Most of the single-sensor false alarms from
Figure 7 were discarded by our fusion method by recognizing the sensor conflicts. Yet, there are a
considerable number of remaining false hits. These spurious detections originate from single-sensor hits
that overlap purely by chance. Nevertheless, all grooves but the shallowest one, nr. 15, clearly stand out
against the false alarms considering the fused density measure, which is represented by the marker colors
in Figure 9. Because higher values of the fused measure correspond to increased defect likelihood, a
threshold can be applied to produce a binary decision. In the following, the detection performance will be
quantitatively assessed under various conditions.
Figure 9: Result of decision level fusion. Darker markers correspond to increased detection confidence. The colors
range from zero (white“ to the fused intensity at the shallow defect nr. 14 or higher (black“. Axes are not to scale.
3.4 Evaluation
In the following sub-sections, our fusion method is quantitatively evaluated with regard to the presented
specimen. This evaluation focuses on detectability, meaning the ability to distinguish between grooves
and background in the fusion result. Consequently, the ability to accurately localize a defect after fusion is
not a part of this evaluation. The evaluation is realized by manually labeling detections near the known
groove locations as on-groove , and others as off-groove . Specifically, for each method to be evaluated
and for each groove, we define a polygon, inside which hit locations are labeled as on-groove . In the
same way, several regions on the specimen surface are marked to be excluded from the evaluation.
These are areas near the border of the specimen, indications that result from experimental modification of
the specimen surface, off-groove areas where real unplanned defects exist (which would otherwise be
counted as false alarms“, and a margin around each groove polygon. Excluding this margin is necessary,
because the defect indications might exceed the actual defect region on the specimen, as is seen in
Figure 5. Hits in this margin usually have reduced intensity compared to the center of the flaw. Therefore,
if they were included in the on-groove class, these intensities might not exceed the detection threshold
and thus the groove would often seem to be detected incompletely. On the other hand, if they were
counted as off-groove, the number of false alarms would be spuriously increased, because the intensities
are often still larger than the background signal. Not only are all of these disregarded regions removed
from evaluation after fusion, but already the hits in these regions are excluded from the density estimation,
so that they don t affect the density in the surrounding regions. Furthermore, to evaluate detection
performance per flaw depth, after fusion each groove is assessed individually while ignoring all others.
We note that there are several issues concerning the comparability of different detection methods using
this evaluation approach. The fundamental issue is that detection, and therefore evaluation, is not carried
out per (hit“ region, but per point. This approach complicates evaluation, because different detection
methods (e.g. single sensor detection and fused detection“ yield different numbers of hits at different
locations. This heterogeneity is compensated here by tailoring the ground truth to each investigated
detection method, using the described polygons. Yet, results are never fully comparable, because each is
evaluated on a slightly different basis. Nevertheless, the evaluation was designed to be as fair and
expressive as possible.
Detection performance is measured by area under the precision-recall curve. This method is similar to the
well-known area under the Receiver Operating Characteristic (ROC“ curve, but we compute precision
prec = #true hits/#all hits instead of false alarm rate FAR = #false hits/max possible number of false hits
and plot it against the true positive rate TPR = #true hits/#max possible number of true hits. Precision is
suited for class-imbalanced problems. This is clearly the case here, because the groove regions are much
smaller than the background region. Consequently, bcd is always small, even if there are many times
more false alarms than true hits. In contrast, QeNf relates hits to hits, rather than hits to non-hits.
Furthermore, we do not compute the area under the whole curve, but only for the curve region where
recall = true positive rate = #heCN ℎihP⁄max possible number of true hits > 0.5. We denote this measure by
AUC-PR-0.5. This focuses our evaluation on thresholds that are low enough to ensure that at least half of
a groove is detected. Furthermore, a single false alarm hit with higher intensity than the groove suffices to
force the curve down to zero precision for small true positive rates, i.e. high thresholds, and therefore
dominates the whole AUC measure. This is another reason for ignoring the left half of the diagram in the
computation of AUC-PR-0.5.
Using this evaluation framework, we investigate the performance of our fusion approach in the following.
3.4.1 Influence of kernel size
In section 2.3, we presented an automatic method to set up the kernel bandwidth parameters. In this
section, we quantitatively compare the influence of different kernel sizes on detection performance in our
data set.
For our inspections, the localization uncertainty due to registration errors3 is up to CY = 0.2 mm. We take
this value as the basis for the range of suitable kernel sizes. Specifically, the following kernel sizes are
investigated: ;S = ℎ , ℎ = max0Δ/ , JCY4, maxZΔ/ , JCY[ with J ∈ 00, 0.5, 1, 2, 5,104. See Table 2 for the
specific values of Δ/ . For J = 1, we obtain the setting proposed in section 2.3. For J = 0, the kernel sizes
are computed only from the sampling distances of each sensor, disregarding the registration errors. For
high values of J, the effect of over-smoothing should be observable. For all investigated settings, the
3
As measured by the following: One location on the specimen surface corresponds to point A in the
coordinate system C1, and to point B in the coordinate system C2. The Euclidean distance between point
A mapped to C2 and point B is a registration error. From the distribution of these errors a suitable
summary statistic is generated.
density was evaluated at the same locations. These locations are defined by a super-sampled
measurement grid of one of the individual sensors.
Figure 10 compares the results. The detection performance appears to be stable across a broad range of
kernel sizes. Perfect detection (AUC-PR-0.5=0.5“ is possible for almost all defects. However, at the
shallowest groove nr. 15, SNR in the individual sensors is not sufficient to yield multi-sensor hits. In
contrast, the second-shallowest groove nr. 14 is detectable, although performance depends on the kernel
size. The SNR at this region is relatively low, so that larger kernels increasingly produce false indications
in the fused image. However, the smallest investigated kernel size is also not suited. For this groove, the
optimal kernel size indeed matches the registration error (J = 1“. The deeper grooves, indicated by lower
indices, are readily distinguished from the large number of per-sensor false hits. Only for J = 0, the
kernels apparently are too narrow to associate the individual hits that indicate a groove. This is due to our
relatively large registration error, which this setting does not account for. Interestingly, no over-smoothing
effect due to large values of J can be observed here. This could be attributed to the other implemented
mechanisms for false alarm rejection, such as the SNR-based weights : and the subtraction of the
maximum individual sub-KDE. At very large sizes J = 10, there are two outlier flaws in the evaluation. The
seemingly poor performance is due to a spatial shift of the fused indications, thus exceeding groove region
that was pre-defined for the evaluation.
We suggest using J = 1, that is setting the kernel sizes for sensor S to
;S = ℎ , ℎ = max0Δ/ , CY4, maxZΔ/ , CY[, if information about localization uncertainty CY is available.
Figure 10: Evaluation of different kernel sizes for each groove, measured by AUC-PR-0.5. Larger is better, maximum
possible score is 0.5. Grooves with higher indices are shallower and thus harder to detect.
3.4.2 Influence of weights
In the previous experiment, the individual sensors detections were weighted by factors : to take into
account the local SNR. To this end, SNR is computed from a null distribution of background signal
intensities, similarly to what was described in section 3.2. But in instead of forming probabilities, we
compute a z-score : = − |̂ ]^__ ⁄~Y]^__ from the respective signal intensity
under the null distribution
using robust estimates of |] S and ~] S . This linear transform normalizes the intensities among different
sensors while preserving the intensity information.
In this section, we repeat the experiment of section 3.4.1, but without weighting the individual hits by
setting all weights : = 1. Therefore, the fused density is only influenced by the proximity of the individual
hits. This setting represents inspection results for which no measure of confidence is available.
See Figure 11 for the resulting detection performance, based on unweighted hits. In comparison to the
weighted results shown in Figure 10, detection performance is generally worse. In particular, groove nr 14
is barely detectable with this approach. However, for larger kernel sizes at J = 5, optimal detection results
are still possible for most defects. We explain this by the observation that 5CY equals the groove length for
our specimen. Therefore, all hits from a groove contribute to the fused intensity, whereas for smaller
kernel sizes, different parts of the grooves are unrelated. This is clearly seen at J = 0, where performance
is the poorest, in agreement with the previous experiment. Note that the general trend of improved
performance for larger kernels can partly be explained by a bias in the evaluation regime: Larger kernels
lead to smoother densities that form fewer local maxima ( peaks “. Therefore, the number of potential
false alarms after fusion decreases, which benefits precision.
We conclude that, if possible, the weighted approach should be favored over the unweighted method. If
weighting is not possible, kernel sizes should not be made too small to fully exploit the density information.
We further suggest setting the kernel size smaller than or equal to the expected flaw length.
Figure 11: Evaluation of different kernel sizes, for unweighted hits.
3.4.3 Influence of individual sensors
The choice which sensors to fuse may have an impact on the detection capability. In particular, we are
interested in the contribution that each individual source has on the fusion result. Therefore, we compute
fused densities while excluding one of the three inspections, and compare these results to fusion using all
three data sets. Based on the results of the previous two experiments, we choose a weighted KDE
technique with kernel sizes set according to J = 0.5 which proved to perform well. The same evaluation
locations were used as in section 3.4.1. We would like to remind the reader that our fusion approach
involves suppressing single-sensor detections, i.e. areas in which the joint density is generated by only a
single sensor. Therefore, if one sensor misses a defect due to poor SNR, both other sensors must yield
hits to indicate this defect in the fusion result. In our experiment, the exclusion of one of these sensors will
lead to strongly decreased performance.
As Figure 12 shows, the full set of inspections is necessary to achieve the best detection performance
here. Each two-sensor subset of inspection methods shows slightly different effects. Apparently, the
thermographic data mainly help in detecting the shallow grooves 10-14. However, the same inspection
seems to have missed the flaws 2, 3, 8 and 9, because the information from both MFL and ET is crucial
for detection here. Among the deeper grooves, nrs. 1, 4, 5, 6 and 7 are perfectly found using any twosensor configuration, thus indicating that they are clearly represented in all three measurements. For the
shallower set of grooves, the inclusion of the ET inspection seems to bring slightly smaller gains than
including the MFL or TT inspection. In particular, the second shallowest groove, nr. 14, is readily detected
using a combination of only MFL and TT. This is especially interesting in light of the following experiment,
where single-sensor performance is contrasted with fusion performance.
Figure 12: AUC-PR-0.5 values for fusion using different sub-sets of inspections.
3.4.4 Comparison against single-sensor detection
Here we quantify the gain of fusion over single-sensor inspection. Again, we choose a weighted KDE
technique with kernel sizes set to half the registration error (J = 0.5“. Unlike the previous experiments,
which were all conducted in the same spatial coordinate system, here each single sensor has its own
system. Since we compare each individual sensor against the fusion result, it is natural to evaluate the
density from each of these individual comparisons in the respective single sensor s coordinate system. We
super-sample each measurement grid to avoid missing density peaks. Note that in all three comparisons,
the same density function is evaluated, albeit at different locations. To simplify the evaluation, we report
only a single fusion result by selecting the worst outcome per groove among the three results.
For fair comparison of single-sensor inspection against fusion, only the pixels that represent a local
intensity maximum ( peak “ define the set of evaluated single-sensor locations. In particular, these points
are not chosen according to their intensity. This set of hits is further filtered by the regions that should be
ignored throughout the evaluation (see section 3.3“. A performance curve is formed by varying a threshold
over this distribution of intensity peak values and counting precision and recall for each threshold value. In
contrast, for the assessment decision-level fusion, the per-sensor detections are first judged according to
their confidence, as described in section 3.2. From these hits, our fused density is then computed, which
is in turn subjected to local maximum detection. The density values at these peak positions are then used
to form the ROC curve for the fusion method.
At this point, we emphasize that the results presented in this article are not representative for the general
performance of each individual inspection method. It is possible that better individual results than shown
here may be obtained by optimizing e.g. the specimen preparation, the sensors or the processing
routines. Rather, in the following our focus is to demonstrate that our technique is able to perform well in
the face of imperfect sources of information. Moreover, the single-sensor results are not fully comparable,
because the inspected area of the specimen differs between sensors. Instead, each single-sensor
inspection should only be compared against the fusion result.
We report AUC-PR-0.5 values in Figure 13. According to our evaluation, the presented fusion technique
achieves perfect detection results of AUC-PR-0.5 = 0.5 for the first 12 flaws and outperforms singlesource detection for grooves not deeper than 61 µm (groove nr. 6“. Interestingly, although the
performance index drops below 0.05 for each individual testing technique from groove nr. 10 onwards,
high performance above 0.4 is still possible using our fusion approach. This means that most actual defect
indications along a flaw generate indications in the fused image that dominate the many off-groove false
alarms, which can therefore be suppressed. In contrast, although the individual groove indications may
surmount the measurement noise, they are not significantly different from the many off-groove indications,
thus impairing precision. The last and shallowest groove (11 µm“ is not reflected in the fused density
image, because only the MFL sensor is able to generate hits here. However, at the two shallow defects nr.
13 and 14, fusion performance is still above 0.28. Because we report the worst out of three density
evaluations here and also the ground truths differ between experiments, this value is lower than the result
presented in the previous section, where already a two-sensor combination provided perfect detection. In
contrast, no individual sensor is able to indicate these shallow grooves at acceptable levels of precision.
Figure 13: Per-groove performance comparison (bars“ of single sensor inspection versus the fusion approach. The
maximum possible score is 0.5 (left vertical axis“. Here, fused actually denotes the worst per groove out of the three
density evaluations (see section 3.4.3“. The set of grooves is divided into two sub-figures for clarity. Groove depth is
indicated by the plotted line corresponding to the right vertical axis. Note the different axis scales for groove depth in
the two subplots.
4 Discussion
Our experiments demonstrate that our density-based approach is well-suited to incorporate indications
from heterogeneous sensors. The number of false alarms can be strongly reduced relative to singlesensor inspection while retaining most of the defects. Regarding our evaluation index, for all grooves but
the shallowest one, the fusion method performs as well as or better than single-sensor detection. The
performance gain is most pronounced for the shallower defects, which usually generate less significant
indications.
We note that the principle to quantify agreement among sensors requires that all sensors yield redundant
information about the object of interest, e.g. near-surface cracks in our case. That is, in NDT, all sensors
must respond to the same flaw type in the same size range. If, in contrast, one of the sensors reports a
defect that is not detectable by the other methods, it will be discarded as a false alarm by our technique,
since it is not designed to fuse complementary information.
Another point concerns the relationship between fusion performance and spatial uncertainty. Specifically,
the fusion performance is expected to improve with registration accuracy. This is because kernels can be
made narrower for smaller registration errors, and therefore the likelihood of a non-defect-related
indication due to spurious multi-sensor agreement is reduced. In any case, the actual registration error
must be quantified to set the kernel size accordingly. Moreover, the fusion technique strongly benefits
from realistic estimates of the local signal to noise ratios, which enter the fused density through weights.
For the final detection after fusion, a threshold could be chosen to retain only the significant density peaks.
Thus, only few parameters (localization uncertainty, density threshold“ fully describe the methodology and
are usually readily determined. Furthermore, if one is unsure about a fused indication, the original
individual detections can always be reconsidered to collect additional evidence for or against the presence
of a defect. After all, the density-based approach spatially associates neighboring hits and thus may serve
as the basis for multi-sensor detection after feature extraction. For example, our density measure
identifies narrow regions of increased defect likelihood. These regions can further be assessed by
extracting features from each individual sensor in this region, which could be combined by some
classification algorithm to reach a final conclusion. Concerning the number of sensors, we suggest using
at least three different sources of information, as presented here. However, further investigations show
that improved performance over single-sensor inspection is possible already for two sensors. Note that the
more fusion inputs are provided, the higher the likelihood of the purely coincidental agreement between at
least two sensors. Therefore, our proposal to subtract the largest contribution from the sum of individual
densities will have to be extended if even more sensors are included.
We would also like to point out the limitations of the current study. Whereas the eroded grooves facilitate
detection assessment for well-defined defect depths, their linear shapes do not resemble natural defects.
Also, whereas the orientation of our flaws is well-defined, natural defects often vary in orientation.
Therefore, directionally sensitive measurements, such as ET using a differential probe and MFL using
gradiometers, must be carried out multiple times in different directions. However, because this issue is
only relevant for per-sensor detection prior to fusion, it is not further elaborated here.
5 Conclusion / Outlook
We developed a density-based method for the fusion of spatially scattered data and applied it to sensor
signals from the nondestructive testing of a bearing shell. This high-level fusion approach has the
advantage of being independent from the processes that generate the scattered points. Three different
mechanisms are implemented to increase robustness against false alarms. Practical suggestions on how
to determine the parameters, especially the kernel size, are given. We quantitatively evaluated our
technique using a defect detection experiment. The results demonstrate that single-sensor inspections of
the specimen are outperformed by the proposed technique, especially for defects that are too shallow to
be reliably indicated otherwise. Moreover, the proposed method is quite generic, as it receives spatial
locations from single-source detection routines and returns areas of multi-sensor agreement. Therefore, it
may be applied for detection tasks in other domains, such as medical image fusion.
Acknowledgements
We acknowledge the help of our colleagues in divisions 8.4 and 8.7 for kindly providing the measurements
and for their support: M. Pelkner, R. Stegemann, T. Erthner, R. Pohl, M. Kreutzbruck, M. Ziegler, D.
Mikolai and C. Maierhofer.
[1] R.C. Luo, M.G. Kay, A tutorial on multisensor integration and fusion, in: Proceedings of the 16th
Annual Conference of IEEE Industrial Electronics Society, 1990. IECON '90, 1990, pp. 707-722.
[2] B. Zitová, J. Flusser, Image registration methods: a survey, Image and Vision Computing, 21 (2003“
977-1000.
[3] L.G. Brown, A Survey of Image Registration Techniques, ACM Comput. Surv., 24 (1992“ 325 376.
[4] S. De, K. Gupta, R. Joe Stanley, R. Zoughi, K. Doering, D. Van Aken, G. Steffes, M. O'Keefe, D.
Palmer, A Comprehensive Structural Analysis Process for Failure Assessment in Aircraft Lap-Joint Mimics
Using Intramodal Fusion of Eddy Current Data, Research in Nondestructive Evaluation, 23 (2012“ 146170.
[5] N. Brierley, T. Tippetts, P. Cawley, Data fusion for automated non-destructive inspection, Proceedings
of the Royal Society A: Mathematical, Physical and Engineering Science, 470 (2014“.
[6] M. Friedrich, G. Dobie, C.C. Chan, S.G. Pierce, W. Galbraith, S. Marshall, G. Hayward, Miniature
Mobile Sensor Platforms for Condition Monitoring of Structures, IEEE Sensors Journal, 9 (2009“ 14391448.
[7] J. Moysan, A. Durocher, C. Gueudré, G. Corneloup, Improvement of the non-destructive evaluation of
plasma facing components by data combination of infrared thermal images, NDT & E International, 40
(2007“ 478-485.
[8] K. Fukunaga, L. Hostetler, The estimation of the gradient of a density function, with applications in
pattern recognition, IEEE Transactions on Information Theory, 21 (1975“ 32-40.
[9] M. Ester, H.-p. Kriegel, J. S, X. Xu, A density-based algorithm for discovering clusters in large spatial
databases with noise, in: Proc. 2nd int. Conf. on Knowledge Discovery and Data Mining (KDD '96“, AAAI
Press, Portland, Oregon, 1996, pp. 226 231.
[10] M. Ankerst, M.M. Breunig, H.-P. Kriegel, J. Sander, OPTICS: Ordering Points to Identify the
Clustering Structure, in: Proceedings of the 1999 ACM SIGMOD International Conference on
Management of Data, ACM, New York, NY, USA, 1999, pp. 49 60.
[11] Y. Weiss, Segmentation using eigenvectors: a unifying view, in: The Proceedings of the Seventh
IEEE International Conference on Computer Vision, 1999, 1999, pp. 975-982.
[12] E. Parzen, On Estimation of a Probability Density Function and Mode, The Annals of Mathematical
Statistics, 33 (1962“ 1065-1076.
[13] M. Rosenblatt, Remarks on Some Nonparametric Estimates of a Density Function, The Annals of
Mathematical Statistics, 27 (1956“ 832-837.
[14] N.-B. Heidenreich, A. Schindler, S. Sperlich, Bandwidth selection for kernel density estimation: a
review of fully automatic selectors, AStA Advances in Statistical Analysis, 97 (2013“ 403-433.
[15] M. Pelkner, Entwicklung, Untersuchung und Anwendung von GMR-Sensorarrays für die
Zerstörungsfreie Prüfung von ferromagnetischen Bauteilen, Universität des Saarlandes, 2014, pp. 215.
[16] M. Pelkner, V. Reimund, T. Erthner, N. Panke, M. Kreutzbruck, Automated inspection of surface
breaking cracks using GMR sensor arrays, in: 40th Annual Review of Progress in Quantitative
Nondestructive Evaluation, AIP Publishing, 2014, pp. 1393-1399.
[17] R.R. Coifman, D.L. Donoho, Translation-invariant de-noising, Springer, 1995.
[18] A. Ihler, M. Mandel, Kernel Density Estimation Toolbox for MATLAB (R13“, 2003,
http://www.ics.uci.edu/~ihler/code/kde.html.
| 5cs.CE
|
1
Performance Analysis of NOMA in Training Based
Multiuser MIMO Systems
arXiv:1711.01813v1 [cs.IT] 6 Nov 2017
Hei Victor Cheng, Emil Björnson, and Erik G. Larsson
Department of Electrical Engineering (ISY), Linköping University, Sweden
Email: {hei.cheng, emil.bjornson, erik.g.larsson}@liu.se
Abstract—This paper considers the use of NOMA in multiuser
MIMO systems in practical scenarios where CSI is acquired
through pilot signaling. A new NOMA scheme that uses shared
pilots is proposed. Achievable rate analysis is carried out for
different pilot signaling schemes including both uplink and
downlink pilots. The achievable rate performance of the proposed
NOMA scheme with shared pilot within each group is compared
with the traditional orthogonal access scheme with orthogonal
pilots. Our proposed scheme is a generalization of the orthogonal
scheme, and can be reduced to the orthogonal scheme when
appropriate power allocation parameters are chosen. Numerical
results show that when downlink CSI is available at the users,
our proposed NOMA scheme outperforms orthogonal schemes.
However with more groups of users present in the cell, it is
preferable to use multi-user beamforming in stead of NOMA.
I. I NTRODUCTION
Non-orthogonal-multiple-access (NOMA) is a new
multiple-access concept proposed for next generation wireless
networks [2]. The key idea behind NOMA is the use
of superposition coding [3], and associated interference
cancellation techniques, to serve multiple terminals in the
same time-frequency slot. This is classified as NOMA in
the power domain. NOMA provides the ability to increase
capacity, especially when the number of users exceeds the
dimension of the channel coherence interval, or the number
of spatial dimensions (antennas) available for multiplexing is
limited. The technology is currently attracting much attention
[4]–[7]. In the standardization of 3GPP-LTE-Advanced
networks, a NOMA technique for the downlink (DL), called
multiuser superposition transmission (MUST), was recently
proposed [8].
Concurrently, multiuser MIMO is becoming a cornerstone
technology in emerging standards for wireless access. The
idea is to use multiple, phase-coherently operating antennas
at the base station to simultaneously serve many terminals
and separate them in the spatial domain. The basic multiuser
MIMO concepts and the associated information theory go
back a long time [9]–[11]. The most useful form of multiuser
MIMO is massive MIMO, which emerged more recently [12],
[13]. In massive MIMO, the base stations use hundreds of
antennas to serve tens of terminals – harnessing a large spatial
This work was supported by the Swedish Research Council (VR), the
Linköping University Center for Industrial Information Technology (CENIIT),
and the ELLIIT. Part of this work has been presented at IEEE International Workshop on Signal Processing Advances in Wireless Communications
(SPAWC) 2017 [1], however there is an error in the Fig. 2 which is corrected
in this paper.
multiplexing gain for high area throughputs, as well as a large
array gain for improved coverage.
The question addressed in this paper is under what circumstances the use of NOMA can provide gains in multiuser
MIMO systems. While this question per se is not new, no existing study to the authors’ knowledge addressed it under realistic
assumptions on the availability of channel state information
(CSI). Specifically, previous work either assumed perfect CSI
[5], [14] or only statistical CSI [15]. In contrast, we consider
the use of training (pilot transmission) to acquire estimated
CSI, and we derive rigorous capacity bounds for NOMA-based
access under these practical conditions. Training-based NOMA
schemes have been considered in [16], but only for singleantenna systems and hence only downlink pilots are sent to the
users for estimating their effective channel gains. Moreoever,
in multi-user MIMO the effective channels depend on the
beamforming, which complicates the analysis. Beamforming
with imperfect CSI also creates extra interference to the
users, which has not been investigated in the literature. In
contrast, in this work pilots are transmitted on the uplink (UL),
facilitating the base station to estimate all channels. By virtue
of reciprocity and time-division-duplex (TDD) operation, the
so-obtained estimates constitute legitimate estimates of the
downlink channel as well and can be used for coherent
beamforming. However, since the terminals do not know their
effective channels, we consider also the possibility of sending
(beamformed) pilots in the DL.
The assumptions made on availability of CSI are critical in
the analysis of wireless access performance: Perfect CSI (or
even high-quality CSI) is unobtainable in environments with
mobility, and performance analyses conducted under perfectCSI assumptions often yield significantly overoptimistic results. Conversely, the reliance on only statistical CSI precludes
the full exploitation of spatial multiplexing gains, rendering
any performance results overpessimistic. The quality of the
channel estimates that can ultimately be obtained is dictated
by the length of the channel coherence interval (CI) (product of
the coherence time and the coherence bandwidth): the higher
mobility, the less room for pilots, the lower-quality CSI – and
vice versa. Since the coherence time is proportional to the
wavelength, the use of higher carrier-frequencies accentuates
this problem. In high mobility and at high frequencies, the
channel coherence may become very short and eventually one
is forced to use non-coherent communication techniques [17].
The specific technical contributions of this paper are:
‚ We propose a training scheme to obtain CSI and utilize
2
‚
‚
the NOMA concept in a DL multiuser MIMO system.
The derivation of new, rigorous, semi-closed form lower
bounds on the DL capacity in multiuser MIMO with
NOMA, with and without DL pilots.
A numerical demonstration that NOMA can give gains
in multiuser MIMO with estimated CSI under appropriate conditions, and a discussion of relevant application
scenarios, most importantly that of rate-splitting and
multicasting.
II. S YSTEM M ODEL
We consider a single-cell massive MIMO system with M
antennas at the base station (BS) and K (even number) singleantenna users. Among these users, K{2 of them are located
in the cell center, while the other K{2 users are at the cell
edge. TDD operation is assumed and therefore the BS acquires
downlink channel estimates through uplink pilot signaling, by
exploiting channel reciprocity. These estimates are used to
perform downlink multiuser beamforming. These operations
have to be done within the same CI, where the channels are
approximately constant. Therefore the more symbols spent
on uplink training, the fewer symbols are available for data.
We consider non-line-of-sight communication and model the
small-scale fading for each user as independent Rayleigh
fading. We denote the small-scale fading realizations for the
users at the cell center as
gk „ CN p0, IM q,
k “ 1, . . . , K{2.
(1)
βgk
The corresponding large-scale fading parameters are
ą
0, k “ 1, . . . , K{2 for the b
users in the cell center; the actual
channel realization is then βgk gk .
Similarly, the small-scale fading realizations for the users
at the cell edge are denoted as
hk „ CN p0, IM q,
k “ 1, . . . , K{2.
(2)
The corresponding large-scale fading parameters are βhk ą
0, k “ 1, . . . , K{2
b for the users at the cell edge. The actual
channel is then
βgk gk . The large-scale fading is widely
different between the two sets of users: βgk " βhk . Note that
this is the scenario of interest to us, but the formulas will
actually be valid for any values of βgk and βhk . The names
”cell edge” and ”cell center” are just descriptive, but should
not be interpreted literally.
The BS is assumed to know the deterministic parameters
βgk and βhk . However the small-scale fading realizations are
unknown a priori and changing independently from one CI to
another CI. To estimate the small-scale fading realizations at
the BS, in traditional TDD multiuser MIMO, orthogonal uplink pilots are transmitted from the users in the cell. However,
the number of available orthogonal pilot sequences is limited
by the size of the CI and this effectively limits the number of
users that can be scheduled simultaneously. In this study, we
are interested in the case when K is greater than the number of
available pilot sequences. To facilitate discussion and analysis,
we assume that there are only K{2 orthogonal pilot sequences
available. With this assumption, we compare two schemes that
make use of the K{2 pilot sequences differently.
A. Orthogonal Access Scheme
The first scheme is the traditional orthogonal access scheme
[18] that schedules K{2 users in a fraction η of time-frequency
resources, and then serve the others in the remaining fraction
1 ´ η of the resources. To minimize near-far effects, we
schedule the K{2 users at the cell center in the first fraction
η, followed by the other K{2 users at the cell edge in the
remaining 1 ´ η of the resources. From now on we call this
Scheme-O.
B. Proposed NOMA Scheme
The second scheme is a generalization of an existing scheme
in the NOMA literature [19], which creates K{2 groups, each
with one user at the cell edge and one at the cell center.
In [19], the beamformers are selected based on the channel
of the cell center user, but NOMA with superposition coding
is applied within each group so that the cell edge user can
get a separate data signal. The beamformers can be selected
to mitigate the inter-group interference. For example, in [19],
zero-forcing beamforming is applied to cancel inter-group
interference. However, this existing scheme can only provide
the user at the cell edge with a small data rate. This is so
because the beams are directed to the stronger user in the
group, thus the weaker user will not have any beamforming
gain and this results in low received power and no interference
suppression. Moreover, the existing work is based on the
impractical assumption of perfect CSI. In [20] a two-stage
beamformer is proposed where the outer stage aims to cancel
the inter-group interference and the inner stage beamformer is
optimized to enhance the rate performance for the users within
the group. However this approach needs perfect CSI at the BS
which is hard to obtain in practice and therefore we do not
consider it here.
We propose a generalization of the NOMA scheme from
[19] and devise a way to estimate the channels in practice. To
resolve the pilot-shortage problem, we propose to reuse the
same pilot for multiple terminals in the same cell. In particular,
the BS allocates the same pilot to the two terminals in a group,
where one is in the cell center and one is at the cell edge.1
Since the two users are using the same pilot and have the
same small-scale fading statistics, we will later see that the
BS cannot distinguish their channel responses. However, the
BS can estimate a linear combination of the channels to both
terminals from the pilot transmission. This estimate provides
a useful description of the combined channel, particularly, if
power control is used to even out the pilot signal strengths of
the two terminals. In our proposed scheme, the BS beamforms
a combination of the data symbols intended for the two
terminals using the estimated channels. We make use of the
NOMA concept for which the symbols intended for different
users are super-imposed using super-position coding. The
cell edge user performs the decoding by treating inter-user
interference as noise, while the cell center user decodes the
other user’s data first and performs interference cancellation
before decoding its own data. Since the beamformers are based
1 This scheme can be extended to more than two users, and we will briefly
discuss about this in Section V.
3
eplacements
UL
Pilots
UL
Pilots
DL Data:
DL
Pilots
All Users
UL
Pilots
DL
DL Data:
Cell Center Users
UL
Pilots
Pilots
DL
Pilots
DL
Pilots
One CI
DL Data:
All Users
DL Data:
Cell Edge Users
Another CI
Fig. 1. Frame structure in the considered training based multiuser MIMO
systems. Upper figure: frame structure for the proposed Scheme-N, where all
users are scheduled by sharing pilots. Bottom figure: common frame structure
for Scheme-O, where users are scheduled in different CIs.
user (2,g)
eplacements
user (1,g)
user (2,g)
user (1,g)
user (2,h)
user (1,h)
(a)
user (1,h)
user (2,h)
(b)
Fig. 2. The training and the beamforming stages for Scheme-N. (a) the
transmissions during the uplink training stage where two users with the same
index share the same orthogonal pilot. (b) the beamforming transmission for
the data where the same beam is formed for every two users.
on the channels of all users, the proposed scheme can deliver
good data rates to everyone.
From now on we call this generalized NOMA scheme
Scheme-N.
Fig. 1 shows the frame structure for the two schemes,
Scheme-O and Scheme-N, and Fig. 2 shows the training and
beamforming operations for Scheme-N.
III. U PLINK C HANNEL E STIMATION
In this section we consider the uplink channel estimation for
the two schemes that we are comparing. The channel estimation is different from in conventional systems since the number
of users scheduled in one slot and the pilot orthogonality are
different. We denote the pilot matrix by Φ P CK{2ˆK{2 that
contains the K{2 orthogonal pilot sequences in its rows, i.e.
ΦΦH “ IK{2 .
For Scheme-O, the K{2 users at the cell center are schedO
P CMˆK{2
uled first, and the received uplink pilot signal Yuc
is
?
O
Yuc
“ pu GDg Φ ` Nuc ,
(3)
b
b
K{2
where Dg is a diagonal matrix with βg1 , . . . , βg on its
diagonal. Then the K{2 users at the cell edge are scheduled
O
in a subsequent CI and the received uplink pilot signal Yue
P
MˆK{2
C
is
?
O
(4)
Yue
“ pu HDh Φ ` Nue ,
b
a
K{2
where Dh is a diagonal matrix with βh1 , . . . , βh
on
its diagonal. For Scheme-N, the received uplink pilot signal
YuN P CMˆK{2 is
?
?
YuN “ pu GDg Ag Φ ` pu HDh Ah Φ ` Nu , (5)
where Ab
and Ah are diagonal
matrices with
g
b
b
a
K{2
K{2
and
α1g , . . . , αg
α1h , . . . , αh
on the diagonal
respectively. Nuc , Nue and Nu represent the additive noise
during pilot transmission with independent and identically
distributed (i.i.d.) CN p0, 1q entries. αkh ď 1 and αkg ď 1 are
the positive power control parameters applied to the pilot to
(potentially) even out the channel estimation quality between
the users in the same group.
Without loss of generality, the k th user at the cell center is
paired with the k th user at the cell edge to form the k th group
in Scheme-N, and they are using the same pilot sequence.
From now on we call the cell edge user in the k th group
“user pk, hq” and the cell center user in the k th group “user
pk, gq”.
A. MMSE Channel Estimation for Scheme-O
In this subsection, we consider the channel estimation for
Scheme-O. The estimates will be used in the next section
for performance analysis. The BS first processes the received
pilots signals by multiplying with ΦH from the right. The
processed pilot signal in (3) becomes
b
O H
O
ȳuc,k
“ rYuc
Φ sk “ pu βgk gk ` n̄uc,k , k “ 1, . . . , K{2,
(6)
where n̄uc,k “ rNuc ΦH sk „ CN p0, IM q, for the users at the
cell center, and where r¨sk denotes the kth column of a matrix
. The processed pilot signal in (4) becomes
b
O
O H
ȳue,k
“ rYue
Φ sk “ pu βhk hk ` n̄ue,k , k “ 1, . . . , K{2,
(7)
where n̄ue,k “ rNue ΦH sk „ CN p0, IM q, for the users at
the cell edge.
Based on the processed received pilots, the BS then performs channel estimation. We consider MMSE channel estimation here. Using classical results from [21], we obtain the
MMSE channel estimate of gk is
b
pu βgk
ȳuc,k , k “ 1, . . . , K{2
(8)
ĝk “
pu βgk ` 1
for users at the cell center and the MMSE estimate of hk is
b
pu βhk
ĥk “
ȳue,k , k “ 1, . . . , K{2
(9)
pu βhk ` 1
for users at the cell edge.
B. MMSE Channel Estimation for Scheme-N
Similar to the case of Scheme-O, the BS first processes the
received pilot signal by multiplying with ΦH from the right
in (5) and obtains the processed received signals
b
N
“ rYuN ΦH sk “ pu αkg βgk gk
ȳu,k
b
(10)
` pu αkh βhk hk ` n̄u,k , k “ 1, . . . , K{2,
4
where n̄u,k “ rNu ΦH sk „ CN p0, IM q. Then the MMSE
channel estimate of gk for a user in the cell center is
b
pu αkg βgk
ĝk “
ȳ N , k “ 1, . . . , K{2.
pu αkg βgk ` pu αkh βhk ` 1 u,k
(11)
The MMSE channel estimate of hk for a user at the cell edge
is
b
pu αkh βhk
ĥk “
ȳ N , k “ 1, . . . , K{2. (12)
pu αkg βgk ` pu αkh βhk ` 1 u,k
We observe that ĝk and ĥk are parallel, thus the BS
cannot distinguish between the channel “direction” of users
that share the same pilot. This effect is a consequence of
pilot contamination. Pilot contamination is a major issue in
massive MIMO system, since it makes it hard for the BS
from performing coherent beamforming only towards one of
the users that share a pilot [12]. In contrast, if the same
data is multicasted to multiple users, it is desirable to jointly
beamform towards all of them. Pilot contamination is then
useful to reduce the pilot overhead [22]. In this paper, we will
show how to exploit NOMA to send different data to the users
that share a pilot.
One alternative way to utilize the uplink pilots is to estimate
the linear combination
b
b
wk “ αkg βgk gk ` αkh βhk hk
of the channels. The MMSE estimate of wk for group k is
?
?
pu αkg βgk ` pu αkh βhk N
ȳ ,
ŵk “
pu αkg βgk ` pu αkh βhk ` 1 u,k
k “ 1, . . . , K{2. (13)
Note that ŵk is also parallel with ĝk and ĥk . The choice
of channel estimate does not matter because in either case the
channel estimates are linearly scaled versions of the processed
pilot signal. Hence the beamforming directions suggested by
the estimates are the same by using any one of the estimators.
Since we need to normalize the beamformer to satisfy the
power constraint, the scaling disappears after normalization
and therefore does not affect the rate.
C. Interference-Limited Scenarios
We can obtain a special case by assuming there is no noise
during the uplink training, or equivalently that the uplink
power pu goes to infinity. This yields as an upper bound
on the performance of all the schemes. It is also a good
approximation of the interference-limited scenario with high
SNR, but large inter-user interference.
For Scheme-O, noise-free channel estimation implies that
the channels are perfectly known at the BS, due to the fact
that all users use orthogonal pilots in the uplink training, i.e.,
ĝk “ gk ,
k “ 1, . . . , K{2,
(14)
ĥk “ hk ,
k “ 1, . . . , K{2.
(15)
and
In contrast, for Scheme-N, the channel estimate at the BS will
still be a linear combination of the channels because of the
use of the same pilot in each group. The noise-free estimate
of wk becomes
ŵk “
b
b
αkh βhk hk ` αkg βgk gk “ wk ,
k “ 1, . . . , K{2.
(16)
IV. P ERFORMANCE A NALYSIS
In this section, we analyze the ergodic achievable rates of
Scheme-O and Scheme-N under imperfect channel estimation.
In wireless systems with fast fading channels, channel codes
span many realizations of the fading process. Therefore the
ergodic achievable rate is an appropriate metric to characterize
the performance of coded systems in fast fading environment.
It is commonly adopted in the multiuser MIMO literature,
especially when the number of antennas is large. We make
use of the UL channel estimates from Section III for downlink
beamforming, by assuming perfect reciprocity between UL
and DL. The channel estimation errors are taken into account
in the ergodic achievable rate expressions. We separate the
analysis into three parts, namely the cases with and without
instantaneous DL CSI, and the case with estimated DL channel
gains. The case with instantaneous downlink CSI is unobtainable in practice, and used only as a benchmark.
that
˘ the effective ergodic rate have a prelog penalty
` Note
K
for the case without DL pilots, where T is the size
1 ´ 2T
of the CI. This penalty accounts for the loss from spending
K
2T of every CI to estimate the
˘ For the case with DL
` channels.
pilots, the pre-log penalty is 1 ´ K
T .
A. Downlink Signal Model
Denote by pd the DL transmission power normalized by the
noise variance. For Scheme-O, the received signal for user k
in the cell center is
b
(17)
yc,k “ pd βgk gkT xg ` nc,k , k “ 1, . . . , K{2,
and the received signal for user k at the cell edge is
ye,k “
b
pd βhk hTk xh ` ne,k ,
k “ 1, . . . , K{2,
(18)
where xg (xh ) is the signal vector containing data for the cell
center users (cell edge users), and nc,k (ne,k ) is the normalized
i.i.d. zero mean unit variance complex Gaussian noise at the
k th user at the cell center (edge). Before transmission, each
data symbol is multiplied with a beamforming vector as
xg “
K{2
ÿ
bk
K{2
ÿ
ak
k“1
b
O s
γk,g
k,g
(19)
for the users in the cell center and
xh “
k“1
b
O s
γk,h
k,h
(20)
for the users at cell edge. In the above equations γk,h (γk,g )
represents the non-negative power control coefficients for user
5
k at the cell edge (cell center), and sk,h (sk,g ) is the data
symbol intended for user k at the cell edge (cell center) which
is zero mean and unit variance. The combined signal vectors
xh and xg need to satisfy the power constraint ErxH
h xh s ď 1
and ErxH
g xg s ď 1.
In this work we focus on maximum ratio transmission
(MRT) which is simple to implement and performs close to
optimality in low SNR scenarios,
ĝ ˚
bk “ a k
Er||ĝk ||2 s
for the cell center users and
ĥ˚k
ak “ b
Er||ĥk ||2 s
for the cell edge users. With the normalized beamforming
řK{2 O
vectors, the power constraint becomes k“1 γk,g
ď 1 and
řK{2 O
γ
ď
1.
k“1 k,h
For Scheme-N, the received downlink signal for users in the
cell center is
yk,g “
K{2
b
ÿ
?
gkT ai γi,h si,h
pd βgk
i“1
K{2
b
ÿ
?
` pd βgk
gkT bi γi,g si,g ` nk , k “ 1, . . . , K{2.
i“1
(21)
Similarly, the received downlink signal for users at the cell
edge can be written as
yk,h “
K{2
b
ÿ
?
pd βhk
hTk ai γi,h si,h
i“1
K{2
b
ÿ
?
` pd βhk
hTk bi γi,g si,g ` nk , k “ 1, . . . , K{2.
i“1
(22)
In Scheme-N, where the BS knows only the linear combination of the channels for the users in the same NOMA
group, it regards the estimate as the true channel for both
users pk, gq and pk, hq since that is the best estimate available.
The combined symbols from both terminals in the same group
?
are weighted with the power control coefficients γk,h and
?
γk,g . The transmitted symbol in the k th NOMA group is
?
?
hence
ř ` γk,g sk,g . Therefore the power constraint
ř γk,h sk,h
is k γk,h ` k γk,g ď 1. In this case we have the MRT
beamforming vector with normalization
ŵk˚
ak “ bk “ a
.
Er}ŵk }2 s
(23)
B. Performance With Perfect CSI at the Users
In this subsection, we compute the ergodic achievable rate
for the two schemes under the assumption that the DL pilots
make perfect DL CSI available at the users. This assumes
that DL pilots are sent in each CI and users perform channel
estimation to obtain their own channel gain coefficients and the
cross-channel gains between different users. The achievable
rate is obtained by averaging over all sources of randomness
in the channel and noise.
For Scheme-O, every user decodes its own data symbol by
treating interference as noise. Since perfect CSI is available,
an ergodic achievable rate of user k with beamforming vector a1 , . . . , aK and b1 , . . . , bK can be computed using [13,
Section 2.3.5]
¸ff
«
˜
˙
ˆ
k O
T
2
p
β
γ
|g
b
|
K
d
k
g
k,g
k
O
ř O T 2
ηE log2 1 `
Rc,k
“ 1´
T
|gk bj | ` 1
pd βgk j γj,g
(24)
for the users in the cell center and
«
˜
¸ff
ˆ
˙
k O
T
2
p
β
γ
|h
a
|
K
d
k
h
k,h
k
O
ř O T 2
Re,k
“ 1´
p1´ηqE log2 1 `
T
|hk aj | ` 1
pd βhk j γj,h
(25)
for the users at the cell edge.
The ergodic achievable rates are measured in b/s/Hz, and
they can be achieved by using Gaussian signaling and codewords that span over all channel realizations. The pre-log
factors account for the loss in achievable rate due to the fact
that each user is only scheduled for a fraction of the CIs, in
time or frequency.
For Scheme-N, recall that we name the k th user at the cell
edge as pk, hq and the k th user at the cell center as pk, gq. The
instantaneous SINR of sk,h of user pk, gq is
pd βgk γk,h |gkT ak |2
ř
pd βg j‰k γj,h |gkT aj |2 ` pd βgk j γj,g |gkT bj |2 ` 1
(26)
and similarly the instantaneous SINR of sk,h at user pk, hq
can be written as
SINRk,g “
ř
k
pd βhk γk,h |hTk ak |2
ř
.
T
k
T
2
2
j‰k γj,h |hk aj | ` pd βh
j γj,g |hk bj | ` 1
(27)
The condition that user pk, gq can decode the data intended
for user pk, hq is that the ergodic achievable rate of sk,h at
user pk, gq is no less than the ergodic achievable rate of sk,h
at user pk, hq, which is explicitly
SINRk,h “
pd βhk
ř
Erlog2 p1 ` SINRk,g qs ě Erlog2 p1 ` SINRk,h qs.
(28)
When this condition does not hold, we need to lower the
data rate to user pk, hq such that it can be decoded at user
pk, gq. This can be done by choosing
NP
Rk,h
“ min pErlog2 p1 ` SINRk,g qs, Erlog2 p1 ` SINRk,h qsq .
(29)
Since Erlog2 p1 ` SINRk,h qs is an achievable rate for user
pk, hq, from an information-theoretic perspective any rate that
is lower than that is also achievable. Therefore by transmitting
NP
with the chosen Rk,h
both users are able to decode the data.
In practice, for (28) to hold we just need to properly control
the pilot powers such that (28) holds. Then user pk, gq gathers
all received signals over all channel realizations (coherence
intervals) and decodes the data for user pk, hq. Notice that the
6
SIC is done after the whole codeword is decoded, and not
performed in every CI. Therefore it is not a problem if the
instantaneous SINR is lower at user pk, gq, as long as (28) is
satisfied in the long term.
In the typical scenarios of βhk ! βgk , there is a wide
range of possible choices of power control parameters on the
pilots available to satisfy (28). With any choice of power
control satisfying (28) we transmit with the super-position
coding scheme such that user pk, hq decodes the signal sk,h
from yk,h by treating the signal from user pk, gq as noise.
Then user pk, gq performs successive interference cancellation
such that it first decodes sk,h from yk,g and then subtracts
b
?
pd βgk gkT ak γk,h sk,h from yk,g and decodes sk,g afterwards.
With the superposition coding scheme, the achievable rate
of user pk, gq is given in (30) and the achievable rate of user
pk, hq is given in (31) on top of next page.
It is worth noticing that when αkg “ γk,g “ 0 @k and
k
N
O
αh “ 1, one can obtain Rk,h
“ Re,k
with η “ 1. Similarly
k
when αh “ γk,h “ 0 , @k and αkg “ 1, one can obtain
N
O
Rk,g
“ Rc,k
with η “ 0. By using time-sharing between
these two extremes, we obtain all the ergodic achievable rates
that Scheme-O can attain. This shows that Scheme-N is more
general than the traditional scheme with orthogonal access.
C. Performance Without Downlink CSI
In this section we investigate the case when instantaneous
DL CSI is not available, however we assume the channel
statistics are known by all parties. This corresponds to the case
when no DL pilots are sent and serves as a lower bound on the
performance of all the schemes with estimated DL channels.
In this case users utilize the long term statistics as the channel
gain and decode the signals, that is, they take the statistical
average of the effective gain as an estimate of that gain. Then
the achievable rate is obtained by gathering all the symbols
over different channel realizations and decoding the signal.
Assume the BS uses the estimated CSI for beamforming
to all terminals. Since we are considering MRT beamforming,
ak and bk are scaled versions of the channel estimate ŵk
N
which is a scaled version of the processed pilots ȳu,k
. Then
N˚
the beamforming vector is ak “ bk “ ck ȳu,k where the
normalizing constant ck that meets the power constraint can
be calculated as
1
1
. (32)
“b
ck “ b
N }2 s
Er}ȳu,k
ppu αkh βhk ` pu αkg βgk ` 1qM
Therefore the received signal at user pk, gq is
b
N ˚?
yk,g “ ck βgk gkT ȳu,k
pd γk,h sk,h
b
N ˚?
pd γk,g sk,g ` Ik,g ` nk,g ,
` ck βgk gkT ȳu,k
where
Ik,g “
b ÿ
N ˚?
βgk
pd γj,h sj,h
cj gkT ȳu,j
j‰k
b ÿ
N ˚?
pd γj,g sj,g
cj gkT ȳu,j
` βhk
j‰k
(33)
(34)
is the interference from other groups of users. Similarly, the
received signal at user pk, gq is
b
N ˚?
pd γk,h sk,h
yk,h “ ck βhk hTk ȳu,k
b
(35)
N ˚?
` ck βhk hTk ȳu,k
pd γk,g sk,g ` Ik,h ` nk,h ,
where
Ik,h “
b ÿ
N ˚?
βhk
pd γj,h sj,h
cj hTk ȳu,j
j‰k
b ÿ
N ˚?
` βhk
cj hTk ȳu,j
pd γj,g sj,g
(36)
j‰k
is the interference from other groups of users.
Now we make use of the channel statistics to write the
received signal at terminal pk, hq as
„ b
N ˚?
(37)
pd γk,h sk,h ` zk,h
yk,h “ E ck βhk hTk ȳu,k
where we have introduced the following effective noise term
„ b
˙
ˆ b
N ˚?
N ˚?
zk,h “ ck βhk hTk ȳu,k
pd γk,h ´ E ck βhk hTk ȳu,k
pd γk,h
b
N ˚?
sk,h ` ck βhk hTk ȳu,k
pd γk,g sk,g ` Ik,h ` nk,h .
(38)
It can be easily verified that zk,h is uncorrelated with the signal
term in (37). Therefore (37) can be regarded as an equivalent
scalar channel with deterministic known gain and additive
uncorrelated noise. Using the fact that additive Gaussian noise
is the worst case uncorrelated noise [13, Section 2.3.2], the
following rate is achievable for user pk, hq:
Proposition 1. The following ergodic rate is achievable for
user pk, hq with Scheme-N:
ˆ
˙
ˆ
˙
K
pd λk,h βhk γk,h M
N ip
Rk,h
“ 1´
,
log2 1 `
2T
pd λk,h βhk γk,g M ` pd βhk ` 1
(39)
where λk,h is defined as
λk,h “
pu αkh βhk
.
pu αkh βhk ` pu αkg βgk ` 1
(40)
Proof. The proof is given in Appendix A.
We define
λk,g “
pu αkg βgk
pu αkh βhk ` pu αkg βgk ` 1
(41)
to quantify the channel estimation quality for the following
discussion. Under the condition αkh βhk ď αkg βgk , the effective
SINR of the signal sk,h at user pk, gq is greater than the
effective SINR of the signal sk,h at user pk, hq, i.e.,
pd λk,g βgk γk,h M
pd λk,h βhk γk,h M
.
ď
pd λk,g βgk γk,g M ` pd βgk ` 1
pd λk,h βhk γk,g M ` pd βhk ` 1
(42)
Therefore we can use NOMA where user pk, gq decodes data
from user pk, hq and then subtracts it from the received signal
yk,g . From the sufficient condition αkh βhk ď αkg βgk we see that it
is better to let the user with larger large-scale fading coefficient
7
NP
Rk,g
˜
¸ff
˙ «
ˆ
pd βgk γk,h |gkT ak |2
K
ř
ř
E log2 1 `
“ 1´
T
pd βgk j‰k γj,h |gkT aj |2 ` pd βgk j‰k γj,g |gkT bj |2 ` 1
NP
Rk,h
¸ff
˜
ˆ
˙ «
K
pd βhk γk,h |hTk ak |2
ř
ř
“ 1´
E log2 1 `
T
pd βhk j‰k γj,h |hTk aj |2 ` pd βhk j γj,g |hTk bj |2 ` 1
perform successive interference cancellation as the condition
is easier to satisfy. We have the new received signal
ı
” b
N˚ ?
pd γk,h sk,h
ȳk,g “ yk,g ´ E ck βgk gkT ȳu,k
ı?
” b
N˚
pd γk,g sk,g
“ E ck βgk gkT ȳu,k
´ b
” b
ı¯ ?
N˚
N˚
` ck βgk gkT ȳu,k
pd γk,h sk,h
´ E ck βgk gkT ȳu,k
” b
ı¯ ?
´ b
N˚
N˚
pd γk,g sk,g
´ E ck βgk gkT ȳu,k
` c βgk gkT ȳu,k
` Ik,g ` nk,g .
(43)
We can similarly write the effective noise as
” b
ı¯ ?
´ b
N˚
N˚
pd γk,h sk,h
´ E ck βgk gkT ȳu,k
zk,g “ ck βgk gkT ȳu,k
” b
ı¯ ?
´ b
N˚
N˚
pd γk,g sk,g
´ E ck βgk gkT ȳu,k
` c βgk gkT ȳu,k
` Ik,g ` nk,g .
(44)
Proposition 2. The following ergodic rate is achievable for
user pk, gq with Scheme-N:
˜
¸
´
λk,g βgk γk,g M
τ ¯
N ip
Rk,g “ 1 ´
log2 1 `
.
(45)
2T
pd βgk ` 1
Proof. The proof is given in Appendix B.
From the ergodic rate expressions, we can observe that the
signal terms are proportional to M , which is the array gain
from coherent beamforming. Moreover, we observe that the
total interference from other groups of users is a constant
that only depends the user’s own large-scale fading, but not
on the number of antennas or channel estimation quality.
Therefore the only parameters that affect the rate are the
power control parameters and the uplink channel estimation
quality. Adding more groups of users in Scheme-N will only
change the amount of power that is allocated to each group,
but not the total interference. Each user at the cell edge is
affected by coherent interference from the signal intended
for the cell center user in its group. However, for the users
in the cell center, coherent interference disappears in the
successive interference cancellation and the only effect of the
pilot contamination is the degraded channel estimation quality.
Using similar calculations, we obtain the ergodic achievable
rate expressions for Scheme-O. For users in the cell center, we
have
˜
¸
˙
ˆ
k O
λO
K
k,g βg γk,g M
Oip
η log2 1 `
,
Rc,k “ 1 ´
2T
pd βgk ` 1
(46)
k “ 1, . . . , K{2
where
λO
k,g “
pu βgk
.
pu βgk ` 1
(30)
(31)
(47)
For users at the cell edge, we have
¸
˜
ˆ
˙
O
λO
K
k,h γk,h M
Oip
,
Re,k “ 1 ´
p1 ´ ηq log2 1 `
2T
pd βhk ` 1
(48)
k “ 1, . . . , K{2
where
λO
k,h “
pu βhk
.
pu βhk ` 1
(49)
As in the case with perfect CSI at the users, when we set
αkh “ γk,h “ 0, @k and αkg “ 1 in Scheme-N we get the
achievable rate of the users in the cell center in Scheme-O
with η “ 1. Setting αkh “ γk,h “ 0, @k and αkh “ 1 we get the
achievable rate of the users at the cell edge in Scheme-O with
η “ 0. By using time-sharing between these two extremes,
we obtain all the ergodic achievable rates that Scheme-O can
attain.
D. Performance With Estimated Downlink CSI
DL CSI does not come for free. In practice some form of
estimation of the beamformed channel gain is usually needed.
In this subsection, we investigate the performance of the two
schemes when we send DL (beamformed) pilots [23] for the
channel estimation. For Scheme-O every user receives its own
orthogonal pilot. For Scheme-N, since we are using the same
beamformer for the pair of users in every group k, only one
downlink pilot is needed for every pair of users. In this case
the users estimate their effective channel gain and perform a
form of “equalization” using the estimated channel gain (see
below for the details).
We denote the channel gain at user pk, hq as fk,h fi hTk ak .
Then the received pilot at each of these users is
b
(50)
ydpk,h “ fk,h pd βhk ` ndpk,h , k “ 1, . . . , K{2.
Assuming LMMSE estimation [21] at the user, we obtain the
estimate
fˆk,h “ Erfk,h s
b
˙
b
βhk pd Varrfk,h s ˆ
ydpk,h ´ βhk pd Erfk,h s ,
` k
βh pd Varrfk,h s ` 1
(51)
of the channel gain where
a
Erfk,h s “ M λk,h ,
(52)
Varrfk,h s “ 1.
8
The estimation quality will improve with M as the mean of
the channel gain is increasing with M while the variance is
constant.
Similarly, denote the channel gain at user pk, gq as fk,g fi
gkT bk . The received pilot at each of these user is
b
(53)
ydpk,g “ fk,g pd βgk ` ndpk,g , k “ 1, . . . , K{2.
Applying LMMSE estimation yields the estimate
fˆk,g “ Erfk,g s
b
b
¯ (54)
βgk pd Varrfk,g s ´
ydpk,g ´ βgk pd Erfk,g s ,
` k
βg pd Varrfk,g s ` 1
where
Erfk,g s “
a
M λk,g ,
Varrfk,g s “ 1.
A. User Pairing
In this paper we are investigating the effects of imperfect
CSI obtained through uplink training. The channels are not
known a priori; the only information available at the BS
regarding the channel strength is the large scale fading coefficients of the users. As a result the user pairing has to be
done based on the large-scale fading coefficients tβk u. This
can also be observed from the achievable rate expressions.
This is the same condition that has been discussed in [24].
However the differences are that first, in our case there is
a beamforming gain of order M which effectively increases
the SNR and second, the existence of self-interference caused
by channel estimation errors. A detailed analysis would be
interesting, but has to be left for future work.
(55)
With these estimates of the channel gains, we first divide the
received signal at user pk, hq by the channel estimate. This
can be seen as a form of equalization, and ideally the ratio
fk,g
is one. Then we use the same method as above to obtain
fˆk,g
the achievable rate of user pk, hq in (56) on top of next page.
Similarly for user pk, gq, an achievable rate is given in (57)
on top of next page.
For Scheme-O, similar techniques can be applied to obtain
the achievable rate for the users in the cell center given in (58)
on top of next page. The corresponding achievable rate for the
users at the cell edge is given in (59) on top of next page.
As in the case with perfect CSI at the users, when we set
αkh “ γk,h “ 0, @k and αkg “ 1 we get the achievable rate of
the users at the cell center in Scheme-O with η “ 1. Setting
αkh “ γk,h “ 0, @k and αkh “ 1 we get the achievable rate of
the users at the cell edge in Scheme-O with η “ 0. By using
time sharing between these two extremes, we obtain all the
ergodic achievable rates that Scheme-O can attain.
Table I summarizes all the ergodic rate expressions we have
obtained, they are all listed in Table I with reference to the
equation numbers.
Comparing the achievable rates of the different schemes
under different CSI assumptions, we observe that the main
difference among them is that imperfect CSI at the users is
causing self-interference. Without any downlink pilots, this
self-interference is proportional to the received power (pd β),
which fundamentally limits the achievable rate of the user.
Therefore we can conclude that neither increasing the DL
power nor putting the user closer to the BS would help much.
This would not create a large SINR difference at the user, and
thus we expect that Scheme-N would not provide much gain.
However with DL pilots, the self-interference can be reduced
substantially if we increase the DL SNR. This creates a larger
SINR difference at the users and thus we expect that SchemeN would provide higher gains.
V. P RACTICAL I SSUES AND E XTENSIONS
In this section we discuss various issues when implementing
the proposed Scheme-N in practical systems and some possible
extensions. Due to space limitations, these issues are discussed
briefly and in-depth investigations are left for future work.
B. More than Two Users Per Group
The proposed Scheme-N can be extended to include more
than two users per group. Suppose there are L users in each
group k and each user is labeled as user pk, 1q to user pk, Lq.
In the channel estimation phase they are assigned the same
pilot. The BS estimates a linear combination of the channels
from all L users in the group. Then the BS uses this for MRT
beamforming. Without loss of generality, assume they have
large-scale fading coefficients ordered as β1k ď β2k ď . . . ď
βLk . The required condition such that NOMA can be applied
is that user pk, iq can decode all messages intended for user
pk, jq for all j ď i. The condition can be written as
Erlog2 p1 ` SINRk,i qs ě Erlog2 p1 ` SINRk,j qs
@ i ě j,
(60)
where SINRk,i is the effective SINR of user pk, iq which
has different forms according to the availability of CSI. This
condition can be met by controlling the pilot power of the
users. Detailed analysis of this extension is out of scope and
has to be left for future work due to the limit of space.
C. Users with Multiple Antennas
In the case when users are equipped with more than one
antenna, adding more antennas can be viewed as adding users
at the same distance. Thus the same analysis and results can
be applied by putting the different antennas of the same user
in different groups in Scheme-N. This argument does not
consider the possibility of receive beamforming at the users as
it requires accurate channel estimation at the users. Since the
scenario we considered is when the pilot resources are scarce,
the consideration of receive beamforming at the user side is
out of scope.
D. Power Control
Power control in any communication systems is crucial. We
have considered both power control in the UL for the pilots
and in the DL for the data. They are optimized according to the
requirement of the users. In Section VII we will look at the rate
region and a particular operating point on the Pareto boundary
of the rate region which is obtained by performing power
control on both UL pilots and DL data. However these are
9
N dp
Rk,h
¨
˙
ˆ
˚
K
log2 ˚
“ 1´
˝1 `
T
N dp
Rk,g
Odp
Rc,k
O
Re,k
˛
ˇ ”
ıˇ2
ˇ
ˇ f
pd βhk γk,h ˇE fˆk,h ˇ
‹
k,h
„ˇ
„ˇ
ˇ2
ˇ2 ‹
ˇ ”
ı
ıˇ2
”
ˇI ˇ
ˇ
ˇ ‚
ˇ
ˇ f
f
pd βhk γk,h Var fˆk,h ` pd βhk γk,g ˇE fˆk,h ˇ ` E ˇ fˆk,h ˇ ` E ˇ fˆ1 ˇ
k,h
¨
˙
ˆ
˚
K
log2 ˚
“ 1´
˝1 `
T
k,h
k,h
k,h
˛
ˇ ”
ıˇ
ˇ fk,g ˇ2
ˇE fˆ ˇ
‹
k,g
„ˇ
„ˇ
ˇ2
ˇ2 ‹
”
ı
ˇ
ˇ ‚
ˇI ˇ
f
pd βgk Var fˆk,g ` E ˇ fˆk,g ˇ ` E ˇ fˆ1 ˇ
pd βgk γk,g
k,g
k,g
(57)
k,g
˛
ˇ „
ˇ2
O
ˇ
ˇ
f
k,g
ˇE O ˇ
‹
˚
ˆ
˙
ˇ
fˆk,g ˇ
‹
˚
K
˚
ff
«
ff
«
“ 1´
η log2 ˚1 `
ˇ
ˇ2
ˇ2 ‹
ˇ
„
T
ˇ
ˇ ‹
ˇ IO ˇ
fO
˝
ˇ ` E ˇ O1 ˇ ‚
pd βgk Var fˆk,g
` E ˇˇ fˆk,g
O
O ˇ
ˆ
ˇ fk,g ˇ
k,g
k,g
¨
(56)
O
pd βgk γk,g
˛
ˇ „
ˇ2
O
ˇ
ˇ
f
k,h ˇ
O ˇ
‹
˚
ˆ
˙
pd βhk γk,h
O
ˇE fˆk,h
ˇ
‹
˚
K
‹
«
ff
«
ff
“ 1´
p1 ´ ηq log2 ˚
1
`
ˇ
ˇ
ˇ
ˇ
„
˚
2
2
T
O
O ˇ
ˇ Ik,h
ˇ 1 ˇ ‹
‚
˝
fk,h
k
pd βh Var fˆO ` E ˇˇ fˆO ˇˇ ` E ˇˇ fˆO ˇˇ
k,h
k,h
k,h
¨
done by a grid search over different power control coefficients.
More efficient algorithms for this purpose would be useful but
have to be left for future work.
VI. OTHER A PPLICATIONS
A. Application in Multicasting
A specific application of the techniques in Scheme-N is to
multiresolution multicasting [20]. In multiresoultion multicasting, signals of different resolutions are multicasted to multiple
users requesting the same data. Users with low SINR decode
only the low resolution signal treating the high resolution
signal as noise, while users with high received SINR decode
both the low and high resolution signals. It is natural to apply
NOMA here since the low resolution signal is wanted by all
users in the cell. In this setup we only need to use one uplink
pilot for channel training and the same beamforming vector is
applied to all users in the cell. This can be viewed as a special
case of Scheme-N where the data intended for all users pk, hq
are the same and data intended for all users pk, gq are the
same.
B. Rate-Splitting for Improving Sum Degree of Freedom
Recently a rate-splitting approach was proposed to improve
the sum degree of freedom in broadcast channels [25] which
is an approach that was first used for interference channels
and then called the the Han-Kobayashi scheme [26]. In the
rate-splitting scheme, one selected user’s message is split into
a common part and a private part where the common part can
be decoded by all users. The common part is super-imposed
on the private part and sent with a different beamformer.
All NOMA schemes can be viewed as a special case of the
(58)
(59)
rate-splitting approach where there is no private part for the
user pk, hq and all message to user pk, hq is contained in the
common part. Our proposed Scheme-N can be adapted for the
rate-splitting scheme to handle the problem of pilot shortage
by decomposing the message of user pk, hq into two parts and
the analysis can be carried out using similar techniques.
VII. N UMERICAL R ESULTS
In this section we compare the performance of the two
schemes in different settings. The comparison is done by
comparing the complete achievable rate regions. The achievable rate region is obtained by considering a grid of pilot
power control and data power control coefficients to obtain
the rate pairs for each set of power control parameters, and
then take the convex hull of all the rate pairs. This assumes
the use of time-sharing between different sets of power control
parameters. This gives an approximate rate region which is a
lower bound on the actual rate region.
A. Small-Scale Antenna Systems
The first setup that we are looking into is the case with a
small number of antennas at the BS. In the simulations we
choose M “ 10, K “ 2, βh “ 1, βg “ 100, pu “ pd “ 1.
Since we compare schemes that use the same number of pilots,
we omit the pre-log penalty caused by the use of pilots for
acquiring CSI. For the case without downlink CSI it has fewer
pilots than the other cases.
Fig. 3 shows the performance with noise free uplink estimation and perfect CSI at the users. This case represents an upper
bound on the performance for practically realizable schemes.
From this figure we observe that with perfect CSI available,
10
TABLE I
S UMMARY OF A CHIEVABLE R ATE R ESULTS
Schemes (users)
Scheme-O (cell center)
Scheme-O (cell edge)
Scheme-N (cell center)
Scheme-N (cell edge)
Estimated CSIT, Perfect CSIR
(24)
(25)
(30)
(31)
Estimated CSIT, no CSIR
(46)
(48)
(45)
(39)
12
3.5
Scheme−N
Scheme−O
Scheme−N
Scheme−O
Orthogonal UL pilots
3
Rate of User 2 (b/s/Hz)
Rate of User (k,g) (b/s/Hz)
10
8
6
4
2.5
2
1.5
1
2
0
0.5
0
0.5
1
1.5
2
2.5
Rate of User (k,h) (b/s/Hz)
3
3.5
4
Fig. 3. Achievable rate region with noise free uplink channel estimation and
perfect CSI at the users. M “ 10, K “ 2 βh “ 1, βg “ 100, pu “ pd “ 1.
10
Scheme−N
Scheme−O
9
8
Rate of User (k,g) (b/s/Hz)
Estimated CSIT, CSIR
(58)
(59)
(57)
(56)
7
6
5
4
3
2
1
0
0
0.5
1
1.5
2
Rate of User (k,h) (b/s/Hz)
2.5
3
Fig. 4. Achievable rate region with noisy uplink channel estimation and
perfect CSI at the users. M “ 10, K “ 2, βh “ 1, βg “ 100, pu “ pd “ 1.
the performance gained by using NOMA is quite significant.
For example, when the rate of user pk, hq is 2.5 b/s/Hz, the
rate of user pk, gq can be increased by almost 2 b/s/Hz.
Fig. 4 shows the performance with noisy uplink estimation
and perfect CSI at the users. Comparing with Fig. 3 we observe
that the uplink channel estimation errors do not lower the
performance much for the user in the cell center. However
the rate of the user at the cell edge loses more than 20%, due
to the poor quality of the uplink channel estimate. Never the
less, the gain from using NOMA is still large.
0
0
0.5
1
Rate of User 1 (b/s/Hz)
1.5
2
Fig. 5. Achievable rate region with noisy uplink channel estimation and no
CSI at the users. M “ 10, K “ 2, βh “ 1, βg “ 100, pu “ pd “ 1.
Fig. 5 shows the achievable rate region with noisy uplink
estimation and no CSI at the users. Comparing to Fig. 4 we see
that CSI at the users is critical as Scheme-N and Scheme-O are
overlapping. Without CSI, Scheme-N is performing the same
as Scheme-O which means there is no gain from using NOMA.
We also plot the performance with orthogonal UL pilots for
all K users as reference. In Scheme-N we send K{2 uplink
pilots, while with the ‘Orthogonal UL Pilots’ scheme we send
K uplink pilots. In this comparison all schemes do not require
downlink pilots. This shows that without taking the penalty of
using more pilots, it is better to use orthogonal pilots when
DL CSI is not available. When the number of pilot symbols
is limited and sending K orthogonal pilots is not possible, we
can only compare Scheme-O and Scheme-N. There are still
some gains from using NOMA with other sets of parameters
(when M is of the order of thousands) than the one considered
in this figure, but they are marginal and applying NOMA may
not be worth it since it increases the complexity and delays at
the user.
Fig. 6 shows the achievable rate region with noisy uplink
estimation and estimated channel gains at the users, which is
the most practical scenario. Comparing to Fig. 5 we see that
with the estimated channel gains, we see some gains from
using NOMA. We also plot the performance with orthogonal
UL pilots for all K users as reference. In Scheme-N we send
K{2 uplink pilots and K{2 downlink pilots, while with the
‘Orthogonal UL Pilots’ scheme we send K uplink pilots and
no downlink pilots. Comparing the rate regions we see that our
proposed Scheme-N outperforms both traditional schemes.
11
9
18
Scheme−N
Scheme−O
Orthogonal UL Pilots
Sum Ergodic Achievable Rate (b/s/Hz)
8
Rate of User (k,g) (b/s/Hz)
7
6
5
4
3
2
1
0
14
12
10
8
6
0
0.2
0.4
0.6
0.8
1
1.2
1.4
Rate of User (k,h) (b/s/Hz)
1.6
1.8
11
Scheme−N
Scheme−O
Orthogonal UL Pilots
Sum Ergodic Achievable Rate (b/s/Hz)
9
8
7
6
5
4
3
10
20
30
40
50
60
70
80
Number of Antennas at the BS (M)
90
5
10
15
20
25
30
Path Loss Difference (dB)
35
40
2
Fig. 6. Achievable rate region with noisy uplink channel estimation and
estimated CSI at the users. M “ 10, K “ 2, βh “ 1, βg “ 100, pu “
pd “ 1.
10
Scheme−N
Scheme−O
Orthogonal UL Pilots
16
100
Fig. 7. Sum rate with noisy uplink channel estimation and estimated channel
gains at the users for different number of antennas at the BS (M). K “ 2,
βh1 “ 1, βg1 “ 100, pu “ pd “ 1. The rate of the user pk, hq is constrained
to be the rate it would get when using Scheme-O with η “ 0.5.
B. Constrained Sum Rate Comparison
In this subsection we compare a specific operating point
on the achievable rate region. We choose the point such that
users at the cell edge get the same rate as in Scheme-O with
η “ 0.5. This means that users at the cell edge do not lose any
rate by using NOMA. We compare the sum rate of the whole
cell under this constraint and vary the number of antennas,
large-scale fading parameters. In all plots we choose K “ 2
with 1 user at the cell edge and 1 user in the cell center. For
Scheme-O, 1 user is scheduled in one slot, thus full power is
O
O
“ 1. For Scheme-N, we vary the
used with γ1,g
“ 1 and γ1,h
power between the two users to find the optimal constrained
sum rate.
In Fig. 7 we compare the constrained sum rate with different
Fig. 8. Sum rate with noisy uplink channel estimation and estimated channel
gains at the users with path loss differences (large-scale fading of user pk, hq
is fixed while large-scale fading of user pk, gq is varying). M “ 100, K “ 2,
βh1 “ 1, pu “ pd “ 1. The rate of the user pk, hq is constrained to be the
rate it would get when using Scheme-O with η “ 0.5.
numbers of antennas M at the BS with βh1 “ 1, βg1 “ 100,
and pu “ pd “ 1. From the plot we see that the sum rate
difference between Scheme-O and Scheme-N is increasing
when M increases. This contradicts the common notion that
NOMA is only useful when the number of antennas at the BS
is less than the total number of antennas at the users [5]. The
reason for this is that CSI at the users is very important in
NOMA, and when M is small, the estimation quality is not
good enough, resulting in a lower rate. When M increases,
the estimation quality at the users increases (due to the array
gain that increases the SNR with M in the DL estimation)
and hence the gain from NOMA is more significant. We
also observe that the performance gap between Scheme-O and
the ‘Orthogonal UL Pilots’ decreases with M and eventually
Scheme-O performs worse than the latter. This is due to the
channel hardening effect. The more antennas at the BS, the
less fluctuation in the norm of the channel vector (normalized
by the number of antennas): the norm of the realization of the
channel vector is almost equal to its statistical mean.
In Fig. 8 we compare the constrained sum rate with different
large-scale fading coefficients between the paired users, with
M “ 100, pu “ pd “ 1, βh1 is fixed to be 1 while βg1 varies.
From the plot we see that the sum rate difference between
Scheme-O and Scheme-N is increasing with the large-scale
fading difference. This is expected and matches the results
for single antenna NOMA systems [2]. When the large-scale
fading difference is small, the orthogonal UL pilots scheme
gives the best performance because both users have low SNR
and therefore DL estimates are of poor quality. This verifies
the importance of user pairing in NOMA.
C. Effect of Number of Users or Number of Antennas at the
User
In this subsection, we look into the effect of increasing the
number of users in the cell, or equivalently, the number of
12
VIII. C ONCLUSION
26
Scheme-N
Scheme-O
Orthogonal UL Pilots
Sum Ergodic Achievable Rate (b/s/Hz)
24
In this work we analyzed the performance of NOMA in
multiuser MIMO under practical scenarios where the CSI
was obtained through pilot signaling. The performance analysis was done for a conventional orthogonal scheme and a
NOMA scheme under this setup. Extensive simulations were
done using the derived achievable rate expressions. From the
simulation results we draw the following conclusions:
22
20
18
16
14
12
10
8
2
3
4
5
6
7
8
9
10
Number of Users in the Cell (K)
Fig. 9. Sum rate with noisy uplink channel estimation and estimated channel
gains at the users with different number of users K. M “ 100, βh “ 1,
βg “ 100 and pu “ pd “ 1. The rate of the user pk, hq is constrained to be
the rate it would get when using Scheme-O with η “ 0.5.
antennas at the users. We compare the same operating point
as in the previous subsection. In the simulation we have the
same number of users at the cell edge and in the cell center.
The users at the cell edge have the same large-scale fading
βhk “ βh , k “ 1, . . . , K{2 and the users at the cell center
have the same large-scale fading βgk “ βg , k “ 1, . . . , K{2.
For Scheme-O, all users that are scheduled in one slot have
the same large-scale fading, thus equal power allocation with
O
O
γk,g
“ 2{K, k “ 1, . . . , K{2 and γk,h
“ 2{K, k “
1, . . . , K{2 is optimal in terms of achievable sum rate. For
Scheme-N, we allocate equal power to each group which is
also optimal for the sum rate due to the symmetry in the K{2
groups. The length of the coherence interval T is chosen to be
200 which is corresponding to a typical fast fading scenario.
In Fig. 9, we compare the constrained sum rates with
different numbers of users, with M “ 100, pu “ pd “ 1,
βh “ 1, and βg “ 100. From the figure we see that SchemeN outperforms the other schemes only when there is one group
of users. As soon as there are more than one group, Scheme-N
and Scheme-O are the same (Scheme-O is a special case of
Scheme-N) and they are both worse than the ‘Orthogonal UL
Pilots’ scheme. This is because the inter-group interference
lowers the SINR difference between the cell center user and
the cell edge user, and thus NOMA does not provide any
gain. This shows that the SINR difference is the key factor
for NOMA to outperform the orthogonal scheme, but not the
SNR difference. Moreover, from Fig. 9 we also observe that
when we have more users in the cell, it is better to user
multiuser beamforming instead of NOMA. That is because the
inter-group interference levels are the same for all schemes
and that is the major factor that lowers the SINR. In this
case, increasing the beamforming gain is more effective than
removing the intra-group interferences.
1) NOMA works well only when high quality CSI is available at the user and there is no inter-group interference;
2) When there is more than one group, it is preferable to
use multiuser beamforming instead of NOMA. In this
case, we need a higher beamforming gain to enhance
the SINR;
3) The gain of NOMA increases with the path loss difference between the users in the same NOMA group.
When the difference is small, multiuser beamforming is
preferable.
The above conclusions hold when the BS precoding is restricted to MR. For other more advanced precoding methods,
most importantly zero-forcing, the observations may change
because accurate channel estimates are required by these
methods. Some initial simulations have shown that the proposed shared-pilot scheme only provides little gain with zeroforcing precoding. More exploration is needed to find out the
strategy of applying NOMA in training based systems with
zero-forcing type precoding and it is left for future work.
Moreover, from our simulation results we see that CSI at the
user is critical for the NOMA scheme. Instead of sending DL
pilots, blind channel estimation methods designed for NOMA
can help to reduce pilot overhead and therefore is worthy of
exploration.
A PPENDIX A
P ROOF OF P ROPOSITION 1
Using results from [13, Section 2.3.2], we have the capacity
lower bound:
N ip
Rk,h
ˇ2 ˛
ˇ ” b
ı
ˇ
ˇ
k hT ȳ N ˚ ?p γ
β
E
c
ˇ
d k,h ˇ ‹
k
h k u,k
˚
“ log2 ˝1 `
‚. (61)
Varpzk,h q
¨
The numerator can be calculated as
ˇ „ b
ˇ2
ˇ
ˇ
ˇE ck β k hTk ȳ N ˚ ?pd γk,h ˇ “ pd λk,h βhk γk,h M,
u,k
h
ˇ
ˇ
(62)
13
and the denominator can be calculated as
„ˆ b
N ˚?
pd γk,h
Varrzk,h s “ Var ck βhk hTk ȳu,k
„ b
˙
T N ˚?
k
´ E ck βh hk ȳu,k pd γk,h sk,h
b
N ˚?
` ck βhk hTk ȳu,k
pd γk,g sk,g ` Ik,h ` nk,h
„ b
N˚
` VarrIk,h s ` Varrnk,h s
“ Var ck βhk hTk ȳu,k
ˇ „ b
ˇ2
ˇ
ˇ
N˚ ?
` ˇˇE ck βhk hTk ȳu,k
pd γk,g ˇˇ
“ pd βhk ` 1 ` pd λh,k βhk γk,g M.
P ROOF
(63)
A PPENDIX B
OF P ROPOSITION 2
Using the results from [13, Section 2.3.2], we have the
capacity lower bound:
¨
ˇ2 ˛
ˇ ” b
ı
ˇ
ˇ
N˚ ?
ck βgk gkT ȳu,k
pd γk,g ˇ
ˇE
‹
˚
N ip
Rk,g “ log2 ˝1 `
‚. (64)
Varpzk,g q
The numerator can be calculated as
ˇ2
ˇ ” b
ı
ˇ
ˇ
N˚ ?
pd γk,h ˇ “ pd λk,g βgk γk,g M,
ˇE ck βgk gkT ȳu,k
(65)
and the denominator can be calculated as
„ˆ b
N ˚?
Varrzk,g s “ Var ck βgk gkT ȳu,k
pd γk,h
” b
ı˙
N ˚?
´ E ck βgk gkT ȳu,k
pd γk,h sk,h
ˆ b
N ˚?
pd γk,g
` ck βgk gkT ȳu,k
” b
ı˙
T N ˚?
k
´ E ck βg gk ȳu,k pd γk,g sk,g ` Ik,g ` nk,g
ı
” b
N˚
` VarrIk,g s ` Varrnk,g s
“ Var ck βgk gkT ȳu,k
“ pd βgk ` 1.
(66)
R EFERENCES
[1] H. V. Cheng, E. Björnson, and E. G. Larsson, “NOMA in multiuser
MIMO systems with imperfect CSI,” in 2017 IEEE International
Workshop on Signal Processing Advances in Wireless Communications
(SPAWC).
[2] Y. Saito, Y. Kishiyama, A. Benjebbour, T. Nakamura, A. Li, and
K. Higuchi, “Non-orthogonal multiple access (NOMA) for cellular future radio access,” in 2013 IEEE 77th Vehicular Technology Conference
(VTC Spring), June 2013, pp. 1–5.
[3] T. Cover, “Broadcast channels,” IEEE Transactions on Information
Theory, vol. 18, no. 1, pp. 2–14, Jan 1972.
[4] L. Dai, B. Wang, Y. Yuan, S. Han, C. l. I, and Z. Wang, “Non-orthogonal
multiple access for 5G: solutions, challenges, opportunities, and future
research trends,” IEEE Communications Magazine, vol. 53, no. 9, pp.
74–81, September 2015.
[5] Z. Ding, F. Adachi, and H. V. Poor, “The application of MIMO
to non-orthogonal multiple access,” IEEE Transactions on Wireless
Communications, vol. 15, no. 1, pp. 537–552, Jan 2016.
[6] Z. Ding, Y. Liu, J. Choi, Q. Sun, M. Elkashlan, C. L. I, and H. V.
Poor, “Application of non-orthogonal multiple access in LTE and 5G
networks,” IEEE Communications Magazine, vol. 55, no. 2, pp. 185–
191, February 2017.
[7] V. W. Wong, R. Schober, D. W. K. Ng, and L. Wang, Key Technologies
for 5G Wireless Systems. Cambridge, U. K.: Cambridge University
Press, 2017.
[8] 3GPP TD RP-150496, “Study on downlink multiuser superposition
transmission.”
[9] S. C. Swales, M. A. Beach, D. J. Edwards, and J. P. McGeehan, “The
performance enhancement of multibeam adaptive base-station antennas
for cellular land mobile radio systems,” vol. 39, no. 1, pp. 56–67, 1990.
[10] S. Anderson, M. Millnert, M. Viberg, and B. Wahlberg, “An adaptive
array for mobile communication systems,” vol. 40, no. 1, pp. 230–236,
1991.
[11] G. Caire and S. Shamai, “On the achievable throughput of a multiantenna
Gaussian broadcast channel,” IEEE Transactions on Information Theory,
vol. 49, no. 7, pp. 1691–1706, July 2003.
[12] T. L. Marzetta, “Noncooperative cellular wireless with unlimited numbers of base station antennas,” IEEE Transactions on Wireless Communications, vol. 9, no. 11, pp. 3590–3600, November 2010.
[13] T. Marzetta, E. G. Larsson, H. Yang, and H. Ngo, Fundamentals of
Massive MIMO. Cambridge, U.K.: Cambridge University Press, 2016.
[14] Z. Ding, Z. Yang, P. Fan, and H. V. Poor, “On the performance of
non-orthogonal multiple access in 5G systems with randomly deployed
users,” IEEE Signal Processing Letters, vol. 21, no. 12, pp. 1501–1505,
Dec 2014.
[15] Q. Sun, S. Han, C. L. I, and Z. Pan, “On the ergodic capacity of MIMO
NOMA systems,” IEEE Wireless Communications Letters, vol. 4, no. 4,
pp. 405–408, Aug 2015.
[16] Z. Yang, Z. Ding, P. Fan, and G. K. Karagiannidis, “On the performance of non-orthogonal multiple access systems with partial channel
information,” IEEE Transactions on Communications, vol. 64, no. 2, pp.
654–667, Feb 2016.
[17] L. Jing, E. D. Carvalho, P. Popovski, and A. O. Martinez, “Design and
performance analysis of noncoherent detection systems with massive
receiver arrays,” IEEE Transactions on Signal Processing, vol. 64,
no. 19, pp. 5000–5010, Oct 2016.
[18] D. Tse and P. Viswanath, Fundamentals of Wireless Communication.
Cambridge, U. K.: Cambridge University Press, 2005.
[19] B. Kim, S. Lim, H. Kim, S. Suh, J. Kwun, S. Choi, C. Lee, S. Lee,
and D. Hong, “Non-orthogonal multiple access in a downlink multiuser
beamforming system,” in IEEE Military Communications Conference,
Nov 2013, pp. 1278–1283.
[20] J. Choi, “Minimum power multicast beamforming with superposition
coding for multiresolution broadcast and application to NOMA systems,”
IEEE Transactions on Communications, vol. 63, no. 3, pp. 791–800,
March 2015.
[21] S. M. Kay, Fundamentals of Statistical Signal Processing: Estimation
Theory. Upper Saddle River, NJ, USA: Prentice-Hall, Inc., 1993.
[22] H. Yang, T. L. Marzetta, and A. Ashikhmin, “Multicast performance of
large-scale antenna systems,” in 2013 IEEE 14th Workshop on Signal
Processing Advances in Wireless Communications (SPAWC), June 2013,
pp. 604–608.
[23] H. Q. Ngo and E. G. Larsson, “No downlink pilots are needed in
TDD massive MIMO,” IEEE Transactions on Wireless Communications,
vol. 16, no. 5, pp. 2921–2935, 2017.
[24] Z. Ding, P. Fan, and H. V. Poor, “Impact of user pairing on 5G
nonorthogonal multiple-access downlink transmissions,” IEEE Transactions on Vehicular Technology, vol. 65, no. 8, pp. 6010–6023, Aug 2016.
[25] S. Yang, M. Kobayashi, D. Gesbert, and X. Yi, “Degrees of freedom
of time correlated MISO broadcast channel with delayed CSIT,” IEEE
Transactions on Information Theory, vol. 59, no. 1, pp. 315–328, Jan
2013.
[26] A. E. Gamal and Y. H. Kim, Network Information Theory. Cambridge,
U. K.: Cambridge University Press, 2012.
| 7cs.IT
|
AMINO ACID INTERACTION NETWORK
PREDICTION USING MULTI-OBJECTIVE
OPTIMIZATION
Md. Shiplu Hawlader1 and Saifuddin Md. Tareeq2
1
Department of Computer Science & Engineering, University of Asia Pacific, Dhaka,
Bangladesh
[email protected]
2
Department of Computer Science & Engineering, University of Dhaka, Dhaka,
Bangladesh
[email protected]
ABSTRACT
Protein can be represented by amino acid interaction network. This network is a graph whose vertices
are the proteins amino acids and whose edges are the interactions between them. This interaction
network is the first step of proteins three-dimensional structure prediction. In this paper we present a
multi-objective evolutionary algorithm for interaction prediction and ant colony probabilistic
optimization algorithm is used to confirm the interaction.
KEYWORDS
Protein Structure, Interaction Network, Multi-objective Optimization, Genetic Algorithm, Ant Colony
Optimization
1. INTRODUCTION
Proteins are biological macromolecules performing a vast array of cellular functions within
living organisms. The roles played by proteins are complex and varied from cell to cell and
protein to protein. The best known role of proteins in a cell is performed as enzymes, which
catalyze chemical reaction and increase speed several orders of magnitude, with a remarkable
specificity. And the speed of multiple chemical reactions is essential to the organism survival
like DNA replication, DNA repair and transcription. Proteins are storage house of a cell and
transports small molecules or ions, control the passages of molecules through the cell
membranes, and so forth. Hormone, another kind of protein, transmits information and allow the
regulation of complex cellular processes.
Genome sequencing projects generate an ever increasing number of protein sequences. For
example, the Human Genome Project has identified over 30,000 genes [1] which may encode
about 100,000 proteins. One of the first tasks when annotating a new genome is to assign
functions to the proteins produced by the genes. To fully understand the biological functions of
proteins, the knowledge of their structure is essential.
Proteins are amino acids chain bonded together in peptide bonds, and naturally adopt a native
compact three-dimensional form. The process of forming three-dimensional structure of a
protein is called protein folding and this is not fully understood yet in System Biologoy. The
process is a result of interaction between amino acids which form chemical bond to make
protein structure.
In this paper, we proposed a new algorithm to predict a interaction network of amino acids
using two new emerging optimization techniques, multi-objective optimization based on
evolutionary clustering and ant colony optimization.
2. AMINO ACID INTERACTION NETWORK
Amino acids are the building blocks of proteins. Protein a sequences of amino acids linked by
peptide bond. Each amino acid has the same fundamental structure, differing only in the sidechain, designated the R-group. The carbon atom to which the amino group, carboxyl group, and
side chain (R-group) are attached is the alpha carbon (Cα). The alpha carbon is the common
reference point for coordinates of an amino acid structure. Among the 20 amino acids some are
acidic, some are basic, some are polar, some non-polar. To make a protein, these amino acids
are joined together in a polypeptide chain through the formation of a peptide bond. The
structure, function and general properties of a protein are all determined by the sequence of
amino acids that makes up the primary sequence. The primary structure of a protein is the linear
sequence of its amino acid structural units and it is a part of whole protein structure. The two
torsion angles of the polypeptide chain, also called Ramachandran angles, describe the rotations
of the polypeptide backbone around the bonds between N – Cα (called Phi angle, φ) and Cα – C
(called Psi angle, ψ). Torsion angle is one of the most important parameter of protein structure
and controls the protein folding. For each type of the secondary structure elements there is a
characteristic range of torsion angle values, which can clearly be seen on the Ramachnadran
plot [2].
Figure 1: SSE-IN of 1DTP protein. Green edges are to be
predicted by ant colony algorithm
Another important property of protein is hydrophobicity. Proteins tertiary structure’s core are
hydrophobic and the amino acids inside core part do not interact much as like their counterpart
hydrophilic, those made the outer side of the protein structure. Many systems, both natural and
artificial, can be represented by networks, that is by site or vertices connected by link or edges.
Protein also be represented as a network of amino acid whose edges are the interactions or the
protein functions between amino acids. The 3D structure of a protein is determined by the
coordinates of its atoms. This information is available in Protein Data Bank (PDB) [3], which
regroups all experimentally solved protein structures. Using the coordinates of two atoms, one
can compute the distance between them. We define the distance between two amino acids as the
distance between their Cα atoms. Considering the Cα atom as a center of the amino acid is an
approximation, but it works well enough for our purposes. Let us denote by N the number of
amino acids in the protein. A contact map matrix is an N X N, 0 - 1 matrix, whose element (i, j)
is 1 if there is a contact between amino acids i and j and 0 otherwise. It provides useful
information about the protein. For example, the secondary structure elements can be identified
using this matrix. Indeed, α - helices spread along the main diagonal, while β - sheets appear as
bands parallel or perpendicular to the main diagonal [4]. There are different ways to define the
contact between two amino acids. In [5], the notion is based on spacial proximity, so that the
contact map can consider non–covalent interactions. Gaci et al. in [5] says that two amino acids
are in contact iff the distance between them is below a given threshold. A commonly used
threshold is 7 Å. Consider a contact map graph with N vertices (each vertex corresponds to an
amino acid) and the contact map matrix as incidence matrix. It is called contact map graph. The
contact map graph is an abstract description of the protein structure taking into account only the
interactions between the amino acids. Now let us consider the sub-graph induced by the set of
amino acids participating in SSE, where SSE is secondary structure element of protein like
alpha helix, beta sheet etc. We call this graph SSE interaction network (SSE - IN). The reason of
ignoring the amino acids not participating in SSE is simple. Evolution tends to preserve the
structural core of proteins composed from SSE. In the other hand, the loops (regions between
SSE) are not so important to the structure and hence, are subject to more mutations. That is why
homologous proteins tend to have relatively preserved structural cores and variable loop
regions. Thus, the structure determining interactions are those between amino acids belonging
to the same SSE on local level and between different SSEs on global level. In [6] and [7] the
authors rely on similar models of amino acid interaction networks to study some of their
properties, in particular concerning the role played by certain nodes or comparing the graph to
general interaction networks models. Thanks to this point of view the protein folding problem
can be tackled by graph theory approaches.
Gaci et al. in [8], has described the topological properties of a network and compared them with
some All alpha and beta to prove that a protein can be treat as a network of amino acids.
According to the diameter value, average mean degree and clustering coefficient shown in the
experiment in [8], we can say a protein is a network of amino acids.
3. PREDICT AMINO ACID INTERACTION NETWORK
We can define the problem as prediction of a graph G consist of N vertices V and E edges. If
two amino acids interact with each other in protein we mention it as an edge (u, v) ∈ E, u ∈ V,
v ∈ V of the graph. A SSE-IN is a highly dense sub-graph GSEE-IN with edge set ESEE-IN.
Probability of the edge (u, v) ∈ ESEE-INA, u ∈ VSSE-INA, v ∈ VSSE-INA is very high and
probability of the edge (u, v) ∉ ESEE-INA, u ∈ VSSE-INA, v ɛ VSSE-INA is very low where VSSE-INA
and VSSE-INB are respectively the vertex set of SSE-IN A and SSE-IN B. SCOP and CATH are
the two databases generally accepted as the two main authorities in the world of fold
classification. According to SCOP there are 1393 different folds. To predict the network we
have to solve three problems, as i) find a associate SCOP protein family from the given protein
sequence ii) predict a network of amino acid secondary structure element (SSE) from the known
SCOP protein family and iii) Predict interactions between amino acids in the network, including
internal edges of SSE-IN and external edges.
We are going to avoid the description the first problem, because it can be solve using a good
sequence alignment algorithm like BLAST as discussed in [8]. In this paper we are going to
solve the second and third problem individually with multi-objective optimization using genetic
algorithm and ant colony optimization respectively.
Gaci et al. in [9], described a solution to the prediction of amino acid interaction network. He
used a genetic algorithm with single objective as the distance between to amino acid in protein
atom. But it is very difficult to define real world problems like amino acid interaction problem
in terms of a single objective. A multi-objective optimization problem deals with more than one
objective functions that are to be minimized or maximized. These objectives can be conflicting,
subject to certain constraints and often lead to choosing the best trade-off among them. As we
have described before, the interaction between amino acids in protein depends not only distance
between two amino acids but also the torsion angles and hydrophobic property of the amino
acid. So to get more accurate interaction network of amino acid we have to consider it is as a
multi-objective problem rather than single objective.
4. ALGORITHM
As we have mentioned before, we can solve the amino acid interaction network prediction
problem as well as the protein folding problem using two new and emerging algorithms. The
multi-objective optimization algorithm will predict structural motifs of a protein and will give a
network or graph of secondary structural element (SSE) of the protein. On the other hand, the
ant colony optimization (ACO) algorithm will find the interactions between amino acids
including the intra-SSE-IN and inter-SSE-IN interactions. In our algorithm we have considered
a folded protein in the PDB as an unknown sequence if it has no SCOP v1.73 family
classification. According to [8], we can associate the most compatible and best fit structural
family based on topological criteria like average diameter, average mean distance etc.
4.1. Prediction of SSE interaction network using Multi-objective Optimization
There are several ways to solve multi-objective optimization problem. In this research we have
decided to use Genetic Algorithm (GA) as multi-objective optimization. The GA has to predict
the adjacency matrix of unknown sequence when it is represented by chromosome.
In this paper we proposed a evolutionary clustering algorithm to predict the SSE-IN, which is a
modified algorithm of the second version of strength pareto evolutionary algorithm (SPEA2) in
[10]. SPEA2 preserve better solutions than NSGA-II [11] and its diversity mechanism is better
than the others, this is the reason to choose SPEA2 to implement the evolutionary clustering
algorithm.
Figure 2: Network of 7 nodes clustered into 1,2,3,4 and 5,6,7
and their genetic representation
As proposed in [12], we are using a local-based adjacency representation. In this representation
an individual of the population consist of N genes g1,...,gN, where N is the number of nodes.
Each gene can hold allele value in the range 1,...,N. Genes and alleles represents nodes in the
graph G = (V,E) modelling a network N. A value j assigned in i-th gene interpreted as a link
between node i and j and in clustering node i and j will be in the same cluster as in Figure 2. In
decoding step all the components are identified and nodes participating in the same component
are assigned to the same cluster.
Alogorith1Multiobjective genetic algorithm to predict SSE interaction algorithm
1:
Input: A protein sequence , T = total time steps, NE = Archive size, NP =Population size
2:
Output : A predicted incident matrix M and clustering for each network N i of N
3:
Use BLAST to find a associate protein family of the given sequence from PDB
4:
Generate initial cluster CR 1 = {C 11 , . . . , C k1}of the network N 1 with number of vertex
equal to number of SSE of the associate protein family
for t = 2 to T do
5:
6:
Create initial population of random individual P0 and set E0 = 0 , i = 0
7:
Loop
8:
Decode each individual of Pi ⋃E i
9:
11:
Evaluate each individual of Pi ⋃E i to find rank and density value using equation 1
and 2
Assign fitness value to each individual, as the sum of rank and inverse of density
value
Copy all no dominating solution to E i+1
12:
if | E i+1 | > NE then
10:
13:
truncate | E i+1 | - NE solutions according to topological property
14:
Else
15:
16:
copy best NE - | E i+1 | dominated solution according to their fitness value and
topological property
end if
17:
if stopping criteria does not satisfies then
18:
19:
return non-dominated solution in | E i+1 |
else
20:
Select some individual form— E i+1 for mating pool as parents using binary
tournament with replacement
21:
Apply crossover and mutation operators to the mating pool to the mating pool to
create Np offspring solution and copy to Pi+1
i := i+1
22:
23:
end if
24:
end loop
25:
26:
From the returned solution in E take the best cluster according to the highest
modularity value
end for
It takes a dynamic network N = N1, N2, ... NT, the sequence of graphs G = G1,G2, ... ,GT and the
number of timestamps T as input and gives a clustering of each network Ni of N as output.
In the amino acid interaction network, total number of gene is the number SSE in the associate
protein family found from the first step and each SSE represents one gene or allele notably
considering its size that is the number of amino acids which compose it, of the population. We
represent a protein as an array of alleles. The position of an allele corresponds to the SSE
position it represents in the sequence. At the same time, an incident matrix is associates for each
genome.
For the first time-stamp of first input network there is no temporal relation with the previous
network. The only objective function is snapshot quality or snapshot score. Thus we can apply
any static clustering algorithm or trivial genetic algorithm to find the initial cluster. In this
algorithm we used genetic algorithm to find the best cluster by maximizing the only objective
function. As it is single objective algorithm we can find the single best cluster from this step.
As a first step in each time-stamp from 2nd time-stamp to T, it creates a population of random
individuals. Each individual is a vector of length equal to number of nodes in the graph Gt.
Genetic variant operators will be applied on this population for a fixed number of pass.
Each individual of the population and archive is decoded into component graph. As each
individual gene is working as an adjacency list, if a node in x of graph is reachable from y by
maintaining the edges in the individual, then x and y is in same cluster of component.
Give each individual chromosome of the population and chromosome in archive a rank value.
Smaller the rank value better it is as fitness value. Each non-dominated individual gets the rank
0. After removing the 0 ranked individuals, give the rank 1 to the next non-dominated
individuals and so on. After giving each individuals a rank value, sort the individuals according
to the ascending rank.
r(x) = ∑xp y s( y)
(1)
There could be many individuals of in same area of solution space or objective space. If we take
all these solution into account, we could loss diversity in the population. To remain the
population diverse, we are using distance of k-th nearest neighbour. The fitness value of each
individual is the sum of its non-dominated rank and the inverse of the distance of k-th nearest
neighbours distance. More the distance between solutions, better the fitness functions value.
m(x) = (σ x +1) −1
k
(2)
where σxk is the distance between individual x and its k-th nearest neighbour. To calculate the
distance between chromosome, we have to take account the three objectives, atomic distance of
amino acids, torsion angles and hydrophobicity.
After evaluating fitness values of each of the population and archive, the best individuals are
selected as a new population. From the total individuals of population and archive population
size individuals are selected as new population. From the rank 0 to the highest rank, all the
individuals are added if number of population of this rank is not exceeding the current
population size. If it is exceeding, then some individuals are truncated according to the value of
each individuals.
Table 1: Example of uniform crossover
Parent 1
4
3
2
2
6
5
6
Parent 2
3
3
1
5
4
7
6
Mask
0
1
1
0
0
1
1
Offspring
4
3
1
2
6
7
6
After selecting the new population, a mating pool is created of pool size from the new
population to apply the genetic variation operators. To choose the mating pool, binary
tournament with replacement has been used in this algorithm. According to binary tournament,
two individuals are randomly selected from the new population and the better fitness valued
individual is chosen for the mating pool.
4.1.1. Genetic Variation Operators
Genetic operators are used to create offspring from parent or mating pool. As other genetic
algorithms, in this algorithm two widely used genetic variation operators have been used. These
are crossover and mutation.
Crossover is the operator which is used to create offspring from two parents. The offspring bear
the genes of each parent. As a genetic variation operator there is very high probability to
crossover occurs other than mutation. In this algorithm we are using uniform crossover. A
random bit vector of length of number of the node in the current graph is created. If i-th bit is 0
then the value of the i-th gene comes from the first parent otherwise it comes from the i-th gene
of second parent. As each of the parents holding true adjacency information, the offspring will
also hold it.
One of the most widely used variation operator in genetic algorithm, which perform the
operation in a single individual is mutation. Though the probability of mutation is normally very
low, but it is the best way to make small variation in the individual. To mutate and create a
offspring, some position of the of the individuals are chosen randomly and changed to other
values. But the value should be one of its neighbours in the current graph.
A topological operator is used to exclude incompatible population generated by the algorithm.
We compute the diameter, the characteristic path length and the mean degree to evaluate the
average topological properties of the family for the particular SSE number.
4.1. Ant Colony Optimization (ACO) to Predict Interactions
After predicting the SSE-IN network we have to identify the interactions involve between the
amino acids in the folded protein. We have used an ant colony optimization (ACO) approach to
select and predict the edges which link different SSE's, considering about the correction of the
matrix of motifs previously predicted.
We have built a two steps algorithm as the hierarchical structure of the SSE-IN.
•
In interaction, consider each pair of SSE's separately. This is the local step. We use an
ant colony algorithm to identify the suitable interactions between amino acids belonging
to these SSE's.
•
A global ant colony algorithm is run to predict the interaction between amino acids
from different SSE-IN.
4.2.1. Parameters for Interaction Network Prediction
To predict the interactions, firstly we have to know how many edges to be add in the network
and which nodes we should consider in interactions. To find and evaluate these parameters, we
incorporated the template proteins from the associate family.
We select some template proteins from the associate family whose SSE number is same as the
sequence to predict the edge rate of the sequence and represent them as chromosome or array of
alleles as in the multi-objective genetic algorithm. Thus, we build a comparative model to
compute the edge ratio, which is used to fold the sequence SSE-IN.
We calculate the average chromosome from all the template proteins in associate protein family.
Here we used the distance between two chromosomes as discussed in the previous section to
compare the sequence with the average chromosome. We add up the distance allele by allele to
obtain a distance between the sequence and the average family chromosome. After that, we
calculate the cumulated size by adding up the chromosome cell values. If the distance is less
than 20% of the sequence cumulated size and the average family chromosome then the sequence
is closer to the template protein. Then we compute the average edge rate in the closer protein to
add the initial edges in the disconnected network of the sequence. If we can't find a sequence
closer to the template one, we add the sequence with the average family chromosome and start
again the same procedure.
We do the same procedure to find the designation of the vertices, which vertices should interact
with each other as they also use comparative model.
To define, which edges link two SSE's, we consider the following problem.
Let X = x1, x2, ... xn and Y = y1, y2, ... ym be two SSEs in interaction. We want to add e edges
among the n x m possible combinations. For i ∈ [1, n] and j ∈ [1, m] the probability to interact
the amino acid xi with yj, is correlated with the occurrence matrix of the predicted edges ratios,
represented by Q(xi, yj) and we can assume sij ~ Q(xi, yj). To add approximately e edges, we
need
n
m
∑∑S
i =1 j =1
ij
=e
(3)
and
S ij =
eQ( xi , y j )
n
m
∑p=1 ∑p=1 Q( x p , yq )
4.2.2. Ant Colony Algorithm
The prediction of interaction network consists of two approaches, local and global algorithm.
(4)
4.2.2.1. Local Algorithm
The local algorithm is used to predict the suitable shortcut edges between pair of SSEs in the
network. Thus, we differentiate each pair of SSEs which have connection and build a graph
where each vertex of the first SSE is connected to each vertex of the other SSE. The connection
or the edges are weighted (Sij). Then we used an ant colony approach consists of an ant number
equals to the number of vertices in two SSE. The ant system has to reinforce the suitable edges
between the SSEs. We use these edges in the global algorithm which is described in the next
section.
The local ant colony algorithm first creates n ants which is total number of vertices in the two
SSEs related in the search. For an ant to be positioned we choose a random vertex of to SSE
involved and place it. All the n ants are positioned this way and two ants can share same vertex.
An ant in vertex i will choose the vertex j with probability pij, defined as follows:
p ij =
τ ij α . s ij β
∑
α
β
τ
.
s
ij
ij
k ∈V ( i )
(5)
The weight sij also called heuristic vector, calculated before. If the vertices i and j are in the
same SSE, then the edge between these two vertices has weight equal to the average weight of
the shortcut edges:
1
S =
nm
n
m
∑∑
i =1
j =1
s ij
(6)
After each move of an ant we update the pheromone value on the inter-SSE edges using the
formula,
τ ij = (1 − ρ )τ ij + n ij ∆ τ
(7)
where sij is the number of ants that moved on the edge (i, j) and ∆τ is the quantity of pheromone
dropped by each ant. As far as the edges belonging to the same SSE are concerned, we keep the
pheromone rate on them equals to the average pheromone rate on the inter-SSE edges
1
τ =
nm
n
m
∑ ∑τ
i =1 j =1
ij
(8)
Ants are move inside an SSE randomly, described above, on the other hand if they decide to
change the SSE they are guided by the edge weight and the weight is guided by the pheromone
value. The algorithm stops after a predefined number of iteration or the maximum pheromone
rate is e time bigger than the average pheromone rate on the edge. After the execution of the
algorithm we keep the edges whose pheromone quantity exceeds a threshold λmin.
Algorithm 2 Local algorithm to find Inter-SSE edges
1:
Input: The predicted network from the multiobjective genetic algorithm
2:
Output: Predicted inter-SSE edges
3:
Create n ants, where n is the total number of nodes in process
4:
while stopping criteria does not meet do
5:
for all ant α do
6:
moveAnt(α)
7:
end for
8:
updatePheromone()
9:
end while
10:
selectEdges(λmin)
Algorithm 3 Global algorithm to predict edges into SSEs
1:
2:
Input: The network with predicted edges Es from local algorithm and Ep, number edges
to predict
Output: The network with total Ep edges
3:
buildSSEIN(Es)
4:
create n ants
5:
while stopping criteria does not meet do
6:
for all ant a do
7:
moveAnt(α)
8:
end for
9:
updatePheromone()
10:
end while
11:
selectEdges(Ep)
4.2.2.2. Global Algorithm
After the local algorithm execution, we get the SSE-IN composed of these specific inter-SSE
edges. The global algorithm will keep the number of edges exactly Ep, which was predicted
before. As the local one, the global algorithm uses the ant colony approach with the number of
vertices equal to the SSE-IN vertex number. The ants movements contribute to emerge the
specific shortcut who’s only a number Ep is kept. We rank the shortcut edges as a function of
the pheromone quantity to extract the Ep final shortcuts. Finally, we measure the resulting SSEINs by topological metrics to accept it or not.
We compute the diameter, the characteristic path length and the mean degree to evaluate the
average topological properties of the family for a particular SSE number. Then, after we have
built the sequence SSE-IN, we compare its topological properties with the template ones. We
allow an error up to 20% to accept the built sequence SSE-IN. If the built SSE-IN is not
compatible, it is rejected. We compare the predicted value, denoted Ep, with the real value,
denoted ER
AC = 1 −
| ER − E p |
Ep
(9)
where AC is the accuracy of the prediction.
5. PERFORMANCE ANALYSIS
In this paper we have discussed two algorithms to predict the interaction network of amino acid.
We are going to analysis each algorithm independently.
5.1. Analysis of Genetic Algorithm as Multi-objective Optimization
In order to test the performance of proposed multi-objective genetic algorithm, we randomly
pick three chromosomes from the final population and we compare their associated matrices to
the sequence SSE-IN adjacency matrix. To evaluate the difference between two matrices, we
use an error rate defined as the number of wrong elements divided by the size of the matrix. The
dataset we use is composed of 698 proteins belonging to the All alpha class and 413 proteins
belonging to the All beta class. A structural family has been associated to this dataset as in [8].
All alpha class has an average error rate of 14.6% and for the All beta class it is 13.1% and the
maximum error rate shown in the experiment is 22.9%. Though, the error rate depends on other
criteria like the three objectives described before but according to the result we can firmly assert
that the error rate is depends on the number of initial population, more the number of initial
population less the error rate. With sufficient number of individuals in the initial population we
can ensure the genetic diversity as well as the improved SSE-IN prediction. When the number
of initial population is at least 15, the error rate is always less than 10%.
As compared to the work in [8] we can claim better and improved error rate in this part of SSEIN prediction algorithm.
5.2. Analysis of Ant Colony Optimization
We have experimented and tested this part of our proposed method according to the associated
family protein because the probability of adding edge is determined by the family occurrence
matrix. We have used the same dataset of sequences whose family has been deduced.
For each protein, we have done 150 simulations and when the topological properties are become
compatible to the template properties of the protein we accepted the built SSE-IN. The results
are shown in Table 2. The score is the percentage of correctly predicted shortcut edges between
the sequence SSE-IN and the SSE-IN we have reconstructed [8]. In most cases, the number of
edges to add were accurate according to the Figure 3. From this we can percept that, global
interaction scores depends on the local algorithm lead for each pair of SSEs in contact. The plot,
in Figure 3 confirms this tendency, if the local algorithm select at least 80% of the correct
shortcut edges, the global intersection score stays better than the 80% and evolve around 85%
for the All alpha class and 73% for the All beta class.
Figure 3: Precision in number of edges to be added in All Alpha (left) and All Beta (right).
After the discussion we can say that, though for the big protein of size more than 250 amino
acids the average score decreases, but in an average the score remains for the global algorithm
around 80%.
5.3. Algorithm Complexity
Our proposed algorithm is independent of specific time bound. Both the optimization algorithm
used as multi-objective genetic algorithm and ant colony algorithm, is iteration based. We can
stop the algorithm at any time. Though the result of the algorithm depends on the number of
iteration but if we give sufficient amount of iteration it provides good result. In compare to other
state of art algorithms, those uses exponential complexity algorithm, our is linear in terms of
time and memory.
6. CONCLUSIONS
We have proposed an computational solution to an biological problem. We have described how
we can formulate a biological problem like folding protein into optimization and graph theory
problem. The formulation consists of finding the interactions between secondary structure
element (SSE) network and interaction between amino acids of the protein. The first problem
was solving by an multi-objective genetic algorithm and the second one solve by ant colony
optimization approach.
As discussed before, we have given theoretical and statically proof that our proposed algorithm
gives more accurate result in terms of accuracy and score to predict the amino acid interaction
Table 2: Folding a SSE-In by ant colony approach. The algorithm parameter values are : α = 25,
β = 12, ρ = 0.7, ∆τ = 4000, e = 2, λmin = 0.8.
SCOP Family Number of Proteins Protein Size
Class
All Alpha
All Beta
Score
Average Deviation
46688
17
27-46
83.973
3.277
47472
10
98-125
73.973
12.635
46457
25
129-135
76.125
7.849
48112
11
194-200
69.234
14.008
48507
18
203-214
66.826
5.504
46457
16
241-281
63.281
17.025
48507
20
387-422
62.072
9.304
50629
6
54-66
79.635
2.892
50813
11
90-111
74.006
4.428
48725
24
120-124
80.881
7.775
50629
13
124-128
76.379
9.361
50875
14
133-224
77.959
10.67
network. Though it can be furnished further with improved data structure and parallel
algorithms.
REFERENCES
[1]
E. Pennisi. A low number wins the genesweep pool. In Science, volume 300, page 1484, 2003.
[2]
K Madhusudan Reddy, Sunkara V Manorama, and A Ramachandra Reddy. Bandgap studies on
anatase titanium dioxide nanoparticles. Materials Chemistry and Physics, 78(1):239–245, 2003.
[3]
Helen M Berman, John Westbrook, Zukang Feng, Gary Gilliland, TN Bhat, Helge Weissig, Ilya
N Shindyalov, and Philip E Bourne. The protein data bank. Nucleic acids research, 28(1):235–
242, 2000.
[4]
Amit Ghosh, KV Brinda, and Saraswathi Vishveshwara. Dynamics of lysozyme structure
network: probing the process of unfolding. Biophysical journal, 92(7):2523–2535, 2007.
[5]
Omar Gaci and Stefan Balev. The small-world model for amino acid interaction networks. In
Advanced Information Networking and Applications Workshops, 2009. WAINA’09. International
Conference on, pages 902–907. IEEE, 2009.
[6]
Usha K Muppirala and Zhijun Li. A simple approach for protein structure discrimination based
on the network pattern of conserved hydrophobic residues. Protein Engineering Design and
Selection, 19(6):265–275, 2006.
[7]
KV Brinda and Saraswathi Vishveshwara. A network representation of protein structures:
implications for protein stability. Biophysical journal, 89(6):4159–4170, 2005.
[8]
Gaci, Omar. Building a topological inference exploiting qualitative criteria. Evolutionary
Bioinformatics, 2010.
[9]
Gaci, Omar. How to fold amino acid interaction networks by computational intelligence
methods. In BioInformatics and BioEngineering (BIBE), 2010 IEEE International Conference
on, pages 150 – 155, 2010.
[10]
Marco Laumanns Eckart Zitzler and Lothar Thiele. SPEA2: Improving the Strength Pareto
Evolutionary Algorithm. Zurich, Switzerland, 2001.
[11]
Deb K, Agrawal S, Pratap A, and Meyarivan T. A fast elitist non-dominated sorting genetic
algorithm for multi-objective optimization: Nsga-ii. In Proc. of 6th international conference on
parallel problem solving on nature, Paris,France, September 2000. Springer.
[12]
Y.J. Park and M.S. Song. A genetic algorithm for clustering problems. In Proc. of 3rd Annual
Conference on Genetic Algorithm, pages 2–9, 1989.
Authors
Md. Shiplu Hawlader, born 1987 in
Dhaka, Bangladesh, graduated from
the Faculty of Engineering of
University of Dhaka in 2011 at the
Computer Science and Engineering
department. After the graduation he
completed his M.Sc.degree at the same
department. He is a lecturer of
Department of Computer Science and
Engineering, University of Asia
Pacific.
Saifuddin Md. Tareeq, Dr., graduated
from Department of Computer Science
and Engineering, University of Dhaka
and working as Associate Professor at
the same Department.
| 9cs.NE
|
Tagger: Deep Unsupervised Perceptual Grouping
arXiv:1606.06724v2 [cs.CV] 28 Nov 2016
Klaus Greff* , Antti Rasmus, Mathias Berglund, Tele Hotloo Hao,
Jürgen Schmidhuber* , Harri Valpola
The Curious AI Company {antti,mathias,hotloo,harri}@cai.fi
*
IDSIA {klaus,juergen}@idsia.ch
Abstract
We present a framework for efficient perceptual inference that explicitly reasons
about the segmentation of its inputs and features. Rather than being trained for
any specific segmentation, our framework learns the grouping process in an unsupervised manner or alongside any supervised task. We enable a neural network to
group the representations of different objects in an iterative manner through a differentiable mechanism. We achieve very fast convergence by allowing the system
to amortize the joint iterative inference of the groupings and their representations.
In contrast to many other recently proposed methods for addressing multi-object
scenes, our system does not assume the inputs to be images and can therefore directly handle other modalities. We evaluate our method on multi-digit classification
of very cluttered images that require texture segmentation. Remarkably our method
achieves improved classification performance over convolutional networks despite
being fully connected, by making use of the grouping mechanism. Furthermore,
we observe that our system greatly improves upon the semi-supervised result of a
baseline Ladder network on our dataset. These results are evidence that grouping
is a powerful tool that can help to improve sample efficiency.
1
Introduction
Humans naturally perceive the world as being structured into different
objects, their properties and relation to each other. This phenomenon
which we refer to as perceptual grouping is also known as amodal
perception in psychology. It occurs effortlessly and includes a segmentation of the visual input, such as that shown in in Figure 1. This
grouping also applies analogously to other modalities, for example
in solving the cocktail party problem (audio) or when separating the
sensation of a grasped object from the sensation of fingers touching
each other (tactile). Even more abstract features such as object class,
color, position, and velocity are naturally grouped together with the
inputs to form coherent objects. This rich structure is crucial for many
real-world tasks such manipulating objects or driving a car, where
awareness of different objects and their features is required.
In this paper, we introduce a framework for learning efficient iterative inference of such perceptual grouping which we call iTerative
Amortized Grouping (TAG). This framework entails a mechanism for
iteratively splitting the inputs and internal representations into several
different groups. We make no assumptions about the structure of this
segmentation and rather train the model end-to-end to discover which
Figure 1: An example of perare the relevant features and how to perform the splitting.
ceptual grouping for vision.
30th Conference on Neural Information Processing Systems (NIPS 2016), Barcelona, Spain.
By using an auxiliary denoising task we focus directly on amortizing the posterior inference of the
object features and their grouping. Because our framework does not make any assumptions about the
structure of the data, it is completely domain agnostic and applicable to any type of data. The TAG
framework works completely unsupervised, but can also be combined with supervised learning for
classification or segmentation.
Another class of recently proposed mechanisms for addressing complex structured inputs is attention [31, 2, 9]. These methods simplify the problem of perception by learning to restrict processing
to a part of the input. In contrast, TAG simply structures the input without directing the focus or discarding irrelevant information. These two systems are not mutually exclusive and could complement
each other: the group structure can help in deciding what exactly to focus on, which in turn may help
simplify the task at hand.
We apply our framework to two artificial datasets: a simple binary one with multiple shapes and
one with overlapping textured MNIST digits on a textured background. We find that our method
learns intuitively appealing groupings that support denoising and classification. Our results for the
2-digit classification are significantly better than a strong ConvNet baseline despite the use of a
fully connected network. The improvements for semi-supervised learning with 1,000 labels are even
greater, suggesting that grouping can help learning by increasing the sample efficiency.
2
Iterative Amortized Grouping (TAG)
Grouping. Our goal is to enable neural networks to split inputs and internal representations into
coherent groups that can be processed separately. We hypothesize that processing the whole input
in one clump is often difficult due to unwanted interference. However, if we allow the network
to separately process groups, it can make use of invariant distributed features without the risk of
ambiguities. We thus define a group to be a collection of inputs and internal representations that are
processed together (largely) independently of the other groups.
The “correct” grouping is often dynamic, ambiguous and task dependent. For example, when driving
along a road, it is useful to group all the buildings together. To find a specific house, however, it is
important to separate the buildings, and to enter one, they need to be subdivided even further. Rather
than treating segmentation as a separate task, we provided a mechanism for grouping as a tool for
our system. We make no assumptions about the correspondence between objects and groups. If the
network can process several objects in one group without unwanted interference, then the network is
free to do so.
Processing of the input is split into K different groups, but it is left up to the network to learn
how to best use this ability in a given problem, such as classification. To make the task of instance
segmentation easy, we keep the groups symmetric in the sense that each group is processed by the
same underlying model. We introduce latent binary variables gk,j to encode if input element xj is
assigned to group k.1
Amortized Iterative Inference. We want our model to reason not only about the group assignments
but also about the representation of each group. This amounts to inference over two sets of variables:
the latent group assignments and the individual group representations; A formulation very similar to
mixture models for which exact inference is typically intractable. For these models it is a common
approach to approximate the inference in an iterative manner by alternating between (re-)estimation
of these two sets (e.g., EM-like methods [8]). The intuition is that given the grouping, inferring the
object features becomes easy, and vice versa. We employ a similar strategy by allowing our network
to iteratively refine its estimates of the group assignments as well as the object representations. If the
model can improve the estimates in each step, then it will converge to a final solution.
Rather than deriving and then running an inference algorithm, we train a parametric mapping to arrive
at the end result of inference as efficiently as possible [13]. This is known as amortized inference [34],
and it is used, for instance, in variational autoencoders where the encoder learns to amortize the
posterior inference required by the generative model represented by the decoder. Here we instead
apply the framework of denoising autoencoders [10, 18, 37] which are trained to reconstruct original
1
More formally, we introduce discrete random variables Gj for each input element indexed by j. As a
shorthand for p(Gj = k) we write p(gk,j ) and denote the discrete-valued vector of elements gk,j by gk .
2
δz0
z
δz1
z
L(m0)
m1
L(m1)
m2
1
q1(x)
2
δz2
L(m2)
q2 (x)
z
m3
δz i-1
zi
L(mi-1)
3
m i-1
z
m0
0
mi
z i-1
PARAM
ETRIC
MAPPIN
G
PARAMETRIC MAPPING
PARAM
ETRIC
MAPPIN
G
PARAM
ETRIC
MAPPIN
G
q3 (x)
x̃
q1 (x|g)
x
x
iteration 1
iteration 2
q1(x)
x̃
iteration 3
Figure 2: Illustration of the TAG framework used for training. Left: The system learns by denoising
its input over iterations using several groups to distribute the representation. Each group, represented
by several panels of the same color, maintains its own estimate of reconstructions z i of the input,
and corresponding masks mi , which encode the parts of the input that this group is responsible for
representing. These estimates are updated over iterations by the same network, that is, each group
and iteration share the weights of the network and only the inputs to the network differ. In the case of
images, z contains pixel-values. Right: In each iteration z i−1 and mi−1 from the previous iteration,
are used to compute a likelihood term L(mi−1 ) and modeling error δz i−1 . These four quantities are
fed to the parametric mapping to produce z i and mi for the next iteration. During learning, all inputs
to the network are derived from the corrupted input as shown here. The unsupervised task for the
network is to learn to denoise, i.e. output an estimate q(x) of the original clean input. See Section 2.1
for more details.
inputs x from corrupted versions x̃. This encourages the network to implement useful amortized
posterior inference without ever having to specify or even know the underlying generative model
whose inference is implicitly learned.
This situation is analogous to normal supervised deep learning, which can also be viewed as amortized
inference [6]. Rather than specifying all the hidden variables that are related to the inputs and labels
and then deriving and running an inference algorithm, a supervised deep model is trained to arrive at
an approximation q(class | input) of the true posterior p(class | input) without the user specifying
or typically even knowing the underlying generative model. This works as long as the network is
provided with the input information and mechanisms required for an an efficient approximation of
posterior inference.
2.1
Definition of the TAG mechanism
A high-level illustration of the TAG framework is presented in Figure 2: We train a network with a
learnable grouping mechanism to iteratively denoise corrupted inputs x̃. The output at each iteration
is an approximation qi (x) of the true probability p(x | x̃), which is refined over iterations indexed
by i. As the cost function for training the network, we use the negative log likelihood
X
C(x) = −
log qi (x),
(1)
i
where the summation is over iterations i. From here on we mostly omit i from the equations for
readability. Since this cost function does not require any class labels or intended grouping information,
training can be completely unsupervised, though additional terms for supervised tasks can be added
too.
Group representation. Internally, the network maintains K versions of its representations indexed
by k. This can also be thought of as running K separate copies of the same network, where
each network only sees a subset of the inputs and outputs zk = q(x|gk ) (the expected value
of the input for that group), and mk = q(gk ) (the group assignment probabilities). Each zk
and mk has the same dimensionality as the input, and they are updated over iterations. Each
group k makes its own prediction about the original input based on zk . In the binary case we use
3
q(xj | gk,j ) = sigmoid(zk,j ), and in the continuous case we take zk,j to represent the mean of a
Gaussian distribution with variance v. We assumed the variance of the Gaussian distribution to be
constant over iterations and groups but learned it from the data. It would be easy to add a more
accurate estimate of the variance.
The final prediction of the network is defined as:
X
X
q(xj ) =
q(gk,j )q(xj | gk,j ) =
mk,j q(xj | gk,j ).
k
(2)
k
The group assignment probabilities q(gk,j ) = mk are forced to be non-negative and sum up to one
over k:
X
mk,j ≥ 0,
mk,j = 1.
(3)
k
Inputs. In contrast to a normal denoising autoencoder that receives the corrupted x̃, we feed in
estimates mik and zki from the previous iteration and two additional quantities: L(mik ) and δzki .
They are functions of the estimates and the corrupted x̃ and carry information about how the estimates
could be improved. A parametric mapping (here a neural network) then produces the new estimates
mi+1
and zki+1 . The initial values for m0k are randomized, and zk0 is set to the data mean for all k.
k
Because zk are continuous variables, their likelihood is a function over all possible values of zk , and
not all of this information can be easily represented. Typically, the relevant information is found close
to the current estimate zk ; therefore we use δzk , which is proportional to the gradient of the negative
log likelihood. Essentially, it represents the remaining modeling error:
"
#
X
∂
δzk,j = mk,j (x̃j − zk,j ) ∝
− log(
q(x̃j | zh,j , gh,j )) .
(4)
∂zk,j
h
The derivation of the analogous term in the binary case is presented in Appendix A.5.
Since we are using denoising as a training objective, the network can only be allowed to take inputs
through the corrupted x̃ during learning. Therefore, we need to look at the likelihood p(x̃ | z: , g: ) of
the corrupted input when trying to determine how the estimates could be improved. Since gk,j are
discrete variables unlike zk,j , we treat them slightly differently: For gk,j it is feasible to express the
complete likelihood table assuming other values constant. We denote this function by
L(mk,j ) ∝ q(x̃j | z:,j , g:,j ) .
(5)
Note that we normalize L(mk,j ) over k such that it sums up to one for each value of j. This amounts
to providing each group information about how likely each input element belongs to them rather than
some other group. In other words, this is equivalent to likelihood ratio rather than the raw likelihood.
Intuitively, the term L(mk ) describes how well each group reconstructs the individual input elements
relative to the other groups.
Parametric mapping. The final component needed in the TAG framework is the parametric model,
which does all the heavy lifting of inference. This model has a dual task: first, to denoise the estimate
zk of what each group says about the input, and second, to update the group assignment probabilities
mk of each input element. The information about the remaining modeling error is based on the
corrupted input x̃; thus, the parametric network has to denoise this and in effect implement posterior
inference for the estimated quantities. The mapping function is the same for each group k and for each
iteration. In other words, we share weights and in effect have only a single function approximator
that we reuse.
The denoising task encourages the network to iteratively group its inputs into coherent groups that can
be modeled efficiently. The trained network can be useful for a real-world denoising application, but
typically, the idea is to encourage the network to learn interesting internal representations. Therefore,
it is not q(x) but rather mk , zk and the internal representations of the parametric mapping that we
are typically concerned with.
4
Data: x, K, T, σ, v, Wh , Wu , Θ
Result: z T , mT , C
begin Initialization:
x̃ ← x + N (0, σ 2 I);
m0 ← softmax(N (0, I));
z 0 ← E[x];
end
for i = 0 . . . T − 1 do
for k = 1 . . . K do
z̃k ← N (x̃; zki , (v + σ 2 )I);
δzki ← (x̃ − zki )mik z̃k ;
L(mik ) ← Pz̃kz̃h ;
h
hik ← f (Wh zki , mik , δzki , L(mik ) );
uik ← Ladder(hik , Θ);
i
[zki+1 , mi+1
k ] ← Wu uk ;
end
mi+1 ← softmax(mi+1 );
PK
qi+1 (x) ← k=1 N (x; zki+1 , vI)mi+1 ;
end
PT
C ← − i=1 log qi (x);
Figure 3: An example of how Tagger
would use a 3-layer-deep Ladder Network
as its parametric mapping to perform its
iteration i + 1. Note the optional class prediction output ygi for classification tasks.
Algorithm 1: Pseudocode for running Tagger on a sin- See supplementary material for details.
gle real-valued example x. For details and a binaryinput version please refer to supplementary material.
P
Summary. By using the negative log likelihood C(x) = − i log qi (x) as a cost function, we
train our system to compute an approximation qi (x) of the true denoising posterior p(x|x̃) at each
iteration i. An overview of the whole system is given in Figure 2. For each input element xj we
introduce K latent binary variables gk,j that take a value of 1 if this element is generated by group
k. This way inference is split into K groups, and we can write the approximate posterior in vector
notation as follows:
X
X
qi (x) =
qi (x|gk )qi (gk ) =
N (x; zki , vI)mik ,
(6)
k
k
where we model the group reconstruction qi (x|gk ) as a Gaussian with mean zki and variance v, and
the group assignment posterior qi (gk ) as a categorical distribution mk .
The trainable part of the TAG framework is given by a parametric mapping that operates independently
on each group k and is used to compute both zki and mik (which is afterwards normalized using an
elementwise softmax over the groups). This parametric mapping is usually implemented by a neural
network and the whole system is trained end-to-end using standard backpropagation through time.
The input to the network for the next iteration consists of the vectors zki and mik along with two
additional quantities: The remaining modelling error δzki and the group assignment likelihood ratio
L(mik ) which carry information about how the estimates can be improved:
δzki ∝
∂C(x̃)
∂zki
qi (x̃|gk )
L(mik ) ∝ P
h qi (x̃|gh )
and
Note that they are derived from the corrupted input x̃, to make sure we don’t leak information about
the clean input x into the system.
2.2
The Tagger: Combining TAG and Ladder Network
We chose the Ladder network [22] as the parametric mapping because its structure reflects the
computations required for posterior inference in hierarchical latent variable models. This means
that the network should be well equipped to handle the hierarchical structure one might expect to
find in many domains. We call this Ladder network wrapped in the TAG framework Tagger. This is
illustrated in Figure 3 and the corresponding pseudocode can be found in Algorithm 1.
5
We mostly used the specifications of the Ladder network as described by Rasmus et al. [22], but there
are some minor modifications we made to fit it to the TAG framework. We found that the model
becomes more stable during iterations when we added a sigmoid function to the gating variable v [22,
Equation 2] used in all the decoder layers with continuous outputs. None of the noise sources or
denoising costs were in use (i.e., λl = 0 for all l in Eq. 3 of Ref. [22]), but Ladder’s classification
cost (Cc in Ref. [22]) was added to the Tagger’s cost (Equation 1) for the semi-supervised tasks.
All four inputs (zki , mik , δzki , and L(mik )) were concatenated and projected to a hidden representation
that served as the input layer of the Ladder Network. Subsequently, the values for the next iteration
were simply read from the reconstruction (x̂ in Ref. [22]) and projected linearly into zki+1 and via
softmax to mi+1
to enforce the conditions in Equation 3. For the binary case, we used a logistic
k
sigmoid activation for zki+1 .
3
Experiments and results
We explore the properties and evaluate the performance of Tagger both in fully unsupervised settings
and in semi-supervised tasks in two datasets2 . Although both datasets consist of images and grouping
is intuitively similar to image segmentation, there is no prior in the Tagger model for images: our
results (unlike the ConvNet baseline) generalize even if we permute all the pixels .
Shapes. We use the simple Shapes dataset [24] to examine the basic properties of our system. It
consists of 60,000 (train) + 10,000 (test) binary images of size 20x20. Each image contains three
randomly chosen shapes (4, 5, or ) composed together at random positions with possible overlap.
Textured MNIST. We generated a two-object supervised dataset (TextureMNIST2) by sequentially
stacking two textured 28x28 MNIST-digits, shifted two pixels left and up, and right and down,
respectively, on top of a background texture. The textures for the digits and background are different
randomly shifted samples from a bank of 20 sinusoidal textures with different frequencies and
orientations. Some examples from this dataset are presented in the column of Figure 5. We use a 50k
training set, 10k validation set, and 10k test set to report the results. The dataset is assumed to be
difficult due to the heavy overlap of the objects in addition to the clutter due to the textures. We also
use a textured single-digit version (TextureMNIST1) without a shift to isolate the impact of texturing
from multiple objects.
3.1
Training and evaluation
We train Tagger in an unsupervised manner by only showing the network the raw input example
x, not ground truth masks or any class labels, using 4 groups and 3 iterations. We average the cost
over iterations and use ADAM [17] for optimization. On the Shapes dataset we trained for 100
epochs with a bit-flip probability of 0.2, and on the TextureMNIST dataset for 200 epochs with a
corruption-noise standard deviation of 0.2. The models reported in this paper took approximately 3
and 11 hours in wall clock time on a single Nvidia Titan X GPU for Shapes and TextureMNIST2
datasets respectively.
To understand how model size, length of the iterative inference, and the number of groups affect
the modeling performance, we evaluate the trained models using two metrics: First, the denoising
cost on the validation set, and second we evaluate the segmentation into objects using the adjusted
mutual information (AMI) score [38] and ignore the background and overlap regions in the Shapes
dataset (consistent with Greff et al. [12]). Evaluations of the AMI score and classification results
in semi-supervised tasks were performed using uncorrupted input. The system has no restrictions
regarding the number of groups and iterations used for training and evaluation. The results improved
in terms of both denoising cost and AMI score when iterating further, so we used 5 iterations for
testing. Even if the system was trained with 4 groups and 3 shapes per training example, we could
test the evaluation with, for example, 2 groups and 3 shapes, or 4 groups and 4 shapes.
6
Denoising cost
AMI
Iter 1
0.094
0.58
Iter 2
0.068
0.73
Iter 3
0.063
0.77
Iter 4
0.063
0.79
Iter 5
0.063
0.79
Denoising cost*
AMI*
0.100
0.70
0.069
0.90
0.057
0.95
0.054
0.96
0.054
0.97
RC [12]
Tagger
Tagger*
AMI
0.61 ± 0.005
0.79 ± 0.034
0.97 ± 0.009
(b) Method comparison
(a) Convergence of Tagger over iterative inference
Table 1: Table (a) shows how quickly the algorithm evaluation converges over inference iterations
with the Shapes dataset. Table (b) compares segmentation quality to previous work on the Shapes
dataset. The AMI score is defined in the range from 0 (guessing) to 1 (perfect match). The results
with a star (*) are using LayerNorm [1] instead of BatchNorm.
3.2
Unsupervised Perceptual Grouping
Table 1 shows the median performance of Tagger on the Shapes dataset over 20 seeds. Tagger is able
to achieve very fast convergences, as shown in Table 1a. Through iterations, the network improves its
denoising performances by grouping different objects into different groups. Comparing to Greff et al.
[12], Tagger performs significantly better in terms of AMI score (see Table 1b).
Figure 4 and Figure 5 qualitatively show the learned unsupervised groupings for the Shapes and
textured MNIST datasets. Tagger uses its TAG mechanism slightly differently for the two datasets.
For Shapes, zg represents filled-in objects and masks mg show which part of the object is actually
visible. For textured MNIST, zg represents the textures and masks mg texture segments. In the case
of the same digit or two identical shapes, Tagger can segment them into separate groups, and hence, it
performs instance segmentation. We used 4 groups for training even though there are only 3 objects
in the Shapes dataset and 3 segments in the TexturedMNIST2 dataset. The excess group is left empty
by the trained system but its presence seems to speed up the learning process.
The hand-picked examples A-C in Figure 4 illustrate the robustness of the system when the number of
objects changes in the evaluation dataset or when evaluation is performed using fewer groups.
Example E is particularly interesting; E1 shows how the normal evaluation looks like but E2
demonstrates how we can remove the topmost digit from the scene and let the system fill in digit
below and the background. We do this by setting the corresponding group assignment probabilities
mg to a large negative number just before the final softmax over groups in the last iteration.
To solve the textured two-digit MNIST task, the system has to combine texture cues with high-level
shape information. The system first infers the background texture and mask which are finalized
on the first iteration. Then the second iteration typically fixes the texture used for topmost digit,
while subsequent iterations clarify the occluded digit and its texture. This demonstrates the need for
iterative inference of the grouping.
3.3
Classification
We investigate the role of grouping for the task of classification. We evaluate the Tagger against
four baseline models on the textured MNIST task. As our first baseline we use a fully connected
network (FC) with ReLU activations and batch normalization after each layer. Our second baseline is
a ConvNet (Conv) based on Model C from Springenberg et al. [33], which has close to state-of-the-art
results on CIFAR-10. We removed dropout, added batch normalization after each layer and replaced
the final pooling by a fully connected layer to improve its performance for the task. Furthermore, we
compare with a fully connected Ladder [22] (FC Ladder) network.
All models use a softmax output and are trained with 50,000 samples to minimize the categorical cross
entropy error. In case there are two different digits in the image (most examples in the TextureMNIST2
dataset), the target is p = 0.5 for both classes. We evaluate the models based on classification errors.
For the two-digit case, we score the network based on the two highest predicted classes (top 2).
2
The datasets and a Theano [36] reference implementation of Tagger are available at http://github.com/
CuriousAI/tagger
7
i=1
i=2
i=3
i=4
i=5
C
m3
B
z3
A
m2
0: 60
z2
0: 85
m1
1: 00
z1
1: 00
m0
1: 00
z0
1: 00
1: 00
reconst:
i=0
original
reconst:
Figure 4: Results for Shapes dataset. Left column: 7 examples from the test set along with their
resulting groupings in descending AMI score order and 3 hand-picked examples (A, B, and C) to
demonstrate generalization. A: Testing 2-group model on 3 object data. B: Testing a 4-group model
trained with 3-object data on 4 objects. C: Testing 4-group model trained with 3-object data on
2 objects. Right column: Illustration of the inference process over iterations for four color-coded
groups; mg and zg .
8
i=1
i=2
i=3
i=4
i=5
0: 95
reconst:
i=0
E2
m3
E1
Pred: : no class
z3
D
m2
Pred: : 2
0: 85
z2
0: 86
m1
0: 87
Pred: : no class
z1
0: 89
m0
0: 90
Pred: : 0
z0
0: 92
Class
original
reconst:
Figure 5: Results for the TextureMNIST2 dataset. Left column: 7 examples from the test set along
with their resulting groupings in descending AMI score order and 3 hand-picked examples (D,
E1, E2). D: An example from the TextureMNIST1 dataset. E1-2: A hand-picked example from
TextureMNIST2. E1 demonstrates typical inference, and E2 demonstrates how the system is able to
estimate the input when a certain group (topmost digit 4) is removed. Right column: Illustration of
the inference process over iterations for four color-coded groups; mg and zg .
9
For Tagger, we first train the system in an unsupervised phase for 150 epochs and then add two
fresh randomly initialized layers on top and continue training the entire system end to end using the
sum of unsupervised and supervised cost terms for 50 epochs. Furthermore, the topmost layer has a
per-group softmax activation that includes an added ’no class’ neuron for groups that do not contain
any digit. The final classification is then performed by summing the softmax output over all groups
for the true 10 classes and renormalizing this sum to add up to one.
The final results are summarized in Table 2. As shown in this table, Tagger performs significantly
better than all the fully connected baseline models on both variants, but the improvement is more
pronounced for the two-digit case. This result is expected because for cases with multi-object overlap,
grouping becomes more important. It, moreover, confirms the hypothesis that grouping can help
classification and is particularly beneficial for complex inputs. Remarkably, Tagger, despite being
fully connected, is on par with the convolutional baseline for the TexturedMNIST1 dataset and even
outperforms it in the two-digit case. We hypothesize that one reason for this result is that grouping
allows for the construction of efficient invariant features already in the low layers without losing
information about the assignment of features to objects. Convolutional networks solve this problem
to some degree by grouping features locally through the use of receptive fields, but that strategy is
expensive and can break down in cases of heavy overlap.
3.4
Semi-Supervised Learning
Training TAG does not rely on labels and is therefore directly usable in a semi-supervised context.
For semi-supervised learning, the Ladder [22] is arguably one of the strongest baselines with SOTA
results on 1,000 MNIST and 60,000 permutation invariant MNIST classification. We follow the
common practice of using 1,000 labeled samples and 49,000 unlabeled samples for training Tagger
and the Ladder baselines. For completeness, we also report results of the convolutional (ConvNet)
and fully-connected (FC) baselines trained fully supervised on only 1,000 samples.
From the results in Table 2, it is obvious that all the fully supervised methods fail on this task with
1,000 labels. The best result of approximately 52 % error for the single-digit case is achieved by
ConvNet, which still performs only at chance level for two-digit classification. The best baseline
result is achieved by the FC Ladder, which reaches 30.5 % error for one digit but 68.5 % for
TextureMNIST2.
For both datasets, Tagger achieves by far the lowest error rates: 10.5 % and 24.9 %, respectively.
Again, this difference is amplified for the two-digit case, where the Tagger with 1,000 labels even
outperforms the Ladder baseline with all 50k labels. This result matches our intuition that grouping
can often segment out objects even of an unknown class and thus help select the relevant features for
learning. This is particularly important in semi-supervised learning where the inability to self-classify
unlabeled samples can easily mean that the network fails to learn from them at all.
To put these results in context, we performed informal tests with five human subjects. The task
turned out to be quite difficult and the subjects needed to have regular breaks to be able to maintain
focus. The subjects improved significantly over training for a few days but there were also significant
individual differences. The best performing subjects scored around 10 % error for TextureMNIST1
and 30 % error for TextureMNIST2. For the latter task, the test subject took over 30 seconds per
sample.
4
Related work
Attention models have recently become very popular, and similar to perceptual grouping they help
in dealing with complex structured inputs. These approaches are not, however, mutually exclusive
and can benefit from each other. Overt attention models [31, 9] control a window (fovea) to focus on
relevant parts of the inputs. Two of their limitations are that they are mostly tailored to the visual
domain and are usually only suited to objects that are roughly the same shape as the window. But
their ability to limit the field of view can help to reduce the complexity of the target problem and thus
also help segmentation. Soft attention mechanisms [29, 7, 43] on the other hand use some form of
top-down feedback to suppress inputs that are irrelevant for a given task. These mechanisms have
recently gained popularity, first in machine translation [2] and then for many other problems such as
10
Dataset
TextureMNIST1
Method
FC MLP
FC Ladder
FC Tagger (ours)
ConvNet
Error 50k
Error 1k
31.1 ± 2.2
7.2 ± 0.1
4.0 ± 0.3
3.9 ± 0.3
89.0 ± 0.2
30.5 ± 0.5
10.5 ± 0.9
52.4 ± 5.3
Model details
2000-2000-2000 / 1000-1000
3000-2000-1000-500-250
3000-2000-1000-500-250
based on Model C [33]
FC MLP
55.2 ± 1.0 79.4 ± 0.3 2000-2000-2000 / 1000-1000
FC Ladder
41.1 ± 0.2 68.5 ± 0.2 3000-2000-1000-500-250
FC Tagger (ours)
7.9 ± 0.3 24.9 ± 1.8 3000-2000-1000-500-250
ConvNet
12.6 ± 0.4 79.1 ± 0.8 based on Model C [33]
Table 2: Test-set classification errors for textured one-digit MNIST (chance level: 90 %) and top-2
error on the textured two-digit MNIST dataset (chance level: 80 %). We report mean and sample
standard deviation over 5 runs. FC = Fully Connected
TextureMNIST2
image caption generation [42]. Because they re-weigh all the inputs based on their relevance, they
could benefit from a perceptual grouping process that can refine the precise boundaries of attention.
Our work is primarily built upon a line of research based on the concept that the brain uses synchronization of neuronal firing to bind object representations together. This view was introduced by von
der Malsburg [40] and has inspired many early works on oscillations in neural networks (see the
survey [39] for a summary). Simulating the oscillations explicitly is costly and does not mesh well
with modern neural network architectures (but see [20]). Rather, complex values have been used to
model oscillating activations using the phase as soft tags for synchronization [21, 23]. In our model,
we further abstract them by using discretized synchronization slots (our groups). It is most similar to
the models of Wersing et al. [41], Hyvärinen & Perkiö [16] and Greff et al. [12]. However, our work
is the first to combine this with denoising autoencoders in an end-to-end trainable fashion.
Another closely related line of research [26, 25] has focused on multi-causal modeling of the inputs.
Many of the works in that area [19, 35, 32, 15] build upon Restricted Boltzmann Machines. Each
input is modeled as a mixture model with a separate latent variable for each object. Because exact
inference is intractable, these models approximate the posterior with some form of expectation
maximization [8] or sampling procedure. Our assumptions are very similar to these approaches, but
we allow the model to learn the amortized inference directly (more in line with Goodfellow et al.
[11]).
Since recurrent neural networks (RNNs) are general purpose computers, they can in principle
implement arbitrary computable types of temporary variable binding [28, 29], unsupervised segmentation [27], and internal [29] and external attention [31]. For example, an RNN with fast weights [29]
can rapidly associate or bind the patterns to which the RNN currently attends. Similar approaches
even allow for metalearning [30], that is, learning a learning algorithm. Hochreiter et al. [14], for example, learned fast online learning algorithms for the class of all quadratic functions of two variables.
Unsupervised segmentation could therefore in principle be learned by any RNN as a by-product of
data compression or any other given task.
The recurrent architecture most similar to the Tagger is the Neural Abstraction Pyramid (NAP; [3]) – a
convolutional neural network augmented with lateral connections which help resolve local ambiguities
and feedback connections that allow incorporation of high-level information. In early pioneering
work the NAP was trained for iterative image binarization [5] and iterative image denoising [4], much
akin to the setup we use. Being recurrent, the NAP layers too, could in principle learn a perceptual
grouping as a byproduct. That does not, however, imply that every RNN will, through learning, easily
discover and implement this tool. The main improvement that our framework adds is an explicit
mechanism for the network to split the input into multiple representations and thus quickly and
efficiently learn a grouping mechanism. We believe this special case of computation to be important
enough for many real-world tasks to justify this added complexity.
11
5
Future Work
So far we’ve assumed the groups to represent independent objects or events. However, this assumption
is unrealistic in many cases. Assuming only conditional independence would be considerably more
reasonable, and could be implemented by allowing all groups to share the same top-layer of their
Ladder network.
The TAG framework assumes just one level of (global) groups, which does not reflect the hierarchical
structure of the world. Therefore, another important future extension is to rather use a hierarchy of
local groupings, by using our model as a component of a bigger system. This could be achieved by
collapsing the groups of a Tagger network by summing them together at some hidden layer. That
way this abstract representation could serve as input for another tagger with new groupings at this
higher level. We hypothesize that a hierarchical Tagger could also represent relations between objects,
because they are simply the couplings that remain from the assumption of independent objects.
Movement is a strong segmentation cue and a simple temporal extensions of the TAG framework
could be to allow information to flow forward in time between higher layers, not just via the inputs.
Iteration would then occur in time alongside the changing inputs. We believe that these extensions
will make it possible to scale the approach to video.
6
Conclusion
In this paper, we have argued that the ability to group input elements and internal representations
is a powerful tool that can improve a system’s ability to handle complex multi-object inputs. We
have introduced the TAG framework, which enables a network to directly learn the grouping and
the corresponding amortized iterative inference in a unsupervised manner. The resulting iterative
inference is very efficient and converges within five iterations. We have demonstrated the benefits
of this mechanism for a heavily cluttered classification task, in which our fully connected Tagger
even significantly outperformed a state-of-the-art convolutional network. More impressively, we have
shown that our mechanism can greatly improve semi-supervised learning, exceeding conventional
Ladder networks by a large margin. Our method makes minimal assumptions about the data and can
be applied to any modality. With TAG, we have barely scratched the surface of a comprehensive
integrated grouping mechanism, but we already see significant advantages. We believe grouping to
be crucial to human perception and are convinced that it will help to scale neural networks to even
more complex tasks in the future.
Acknowledgments
The authors wish to acknowledge useful discussions with Theofanis Karaletsos, Jaakko Särelä, Tapani
Raiko, and Søren Kaae Sønderby. And further acknowledge Rinu Boney, Timo Haanpää and the rest
of the Curious AI Company team for their support, computational infrastructure, and human testing.
This research was supported by the EU project “INPUT” (H2020-ICT-2015 grant no. 687795).
References
[1] Ba, J. L., Kiros, J. R., and Hinton, G. E. Layer normalization. arXiv:1607.06450 [cs, stat], July 2016.
[2] Bahdanau, D., Cho, K., and Bengio, Y. Neural machine translation by jointly learning to align and translate.
arXiv preprint arXiv:1409.0473, 2014.
[3] Behnke, Sven. Hebbian learning and competition in the neural abstraction pyramid. In Neural Networks,
1999. IJCNN’99. International Joint Conference on, volume 2, pp. 1356–1361. IEEE, 1999.
[4] Behnke, Sven. Learning iterative image reconstruction in the neural abstraction pyramid. International
Journal of Computational Intelligence and Applications, 1(04):427–438, 2001.
[5] Behnke, Sven. Learning Iterative Binarization using Hierarchical Recurrent Networks. cell, 2:2i, 2003.
[6] Bengio, Yoshua, Laufer, Eric, Alain, Guillaume, and Yosinski, Jason. Deep Generative Stochastic Networks
Trainable by Backprop. pp. 226–234, 2014.
[7] Deco, G. Biased competition mechanisms for visual attention in a multimodular neurodynamical system.
In Emergent Neural Computational Architectures Based on Neuroscience, pp. 114–126. Springer, 2001.
[8] Dempster, A. P., Laird, N. M., and Rubin, D. B. Maximum likelihood from incomplete data via the EM
algorithm. Journal of the royal statistical society., pp. 1–38, 1977.
12
[9] Eslami, S. M., Heess, N., Weber, T., Tassa, Y., Kavukcuoglu, Y., and Hinton, G. E. Attend, infer, repeat:
Fast scene understanding with generative models. preprint arXiv:1603.08575, 2016.
[10] Gallinari, P., LeCun, Y., Thiria, S., and Fogelman-Soulie, F. Mémoires associatives distribuées: Une
comparaison (distributed associative memories: A comparison). In Cesta-Afcet, 1987.
[11] Goodfellow, I. J., Bulatov, Y., Ibarz, J., Arnoud, S., and Shet, V. Multi-digit number recognition from street
view imagery using deep convolutional neural networks. arXiv preprint arXiv:1312.6082, 2013.
[12] Greff, K., Srivastava, R. K., and Schmidhuber, J. Binding via reconstruction clustering. arXiv:1511.06418
[cs], November 2015.
[13] Gregor, K. and LeCun, Y. Learning fast approximations of sparse coding. In Proceedings of the 27th
International Conference on Machine Learning (ICML-10), pp. 399–406, 2010.
[14] Hochreiter, S., Younger, A. S., and Conwell, P. R. Learning to learn using gradient descent. In Proc.
International Conference on Artificial Neural Networks, pp. 87–94. Springer, 2001.
[15] Huang, J. and Murphy, K. Efficient inference in occlusion-aware generative models of images. arXiv
preprint arXiv:1511.06362, 2015.
[16] Hyvärinen, A. and Perkiö, J. Learning to segment any random vector. In The 2006 IEEE International
Joint Conference on Neural Network Proceedings, pp. 4167–4172. IEEE, 2006.
[17] Kingma, D. and Ba, J. Adam: A method for stochastic optimization. CBLS, 2015.
[18] Le Cun, Y. Modèles Connexionnistes de L’apprentissage. PhD thesis, Paris 6, 1987.
[19] Le Roux, N., Heess, N., Shotton, J., and Winn, J. Learning a generative model of images by factoring
appearance and shape. Neural Computation, 23(3):593–650, 2011.
[20] Meier, M., Haschke, R., and Ritter, H. J. Perceptual grouping through competition in coupled oscillator
networks. Neurocomputing, 141:76–83, 2014.
[21] Rao, R. A., Cecchi, G., Peck, C. C., and Kozloski, J. R. Unsupervised segmentation with dynamical units.
Neural Networks, IEEE Transactions on, 19(1):168–182, 2008.
[22] Rasmus, A., Berglund, M., Honkala, M., Valpola, H., and Raiko, T. Semi-supervised learning with ladder
networks. In NIPS, pp. 3532–3540, 2015.
[23] Reichert, D. P. and Serre, T. Neuronal synchrony in complex-valued deep networks. arXiv:1312.6115 [cs,
q-bio, stat], December 2013.
[24] Reichert, D. P., Series, P, and Storkey, A. J. A hierarchical generative model of recurrent object-based
attention in the visual cortex. In ICANN, pp. 18–25. Springer, 2011.
[25] Ross, D. A. and Zemel, R. S. Learning parts-based representations of data. The Journal of Machine
Learning Research, 7:2369–2397, 2006.
[26] Saund, E. A multiple cause mixture model for unsupervised learning. Neural Computation, 7(1):51–71,
1995.
[27] Schmidhuber, J. Learning complex, extended sequences using the principle of history compression. Neural
Computation, 4(2):234–242, 1992.
[28] Schmidhuber, J. Learning to control fast-weight memories: An alternative to dynamic recurrent networks.
Neural Computation, 4(1):131–139, 1992.
[29] Schmidhuber, J. Reducing the ratio between learning complexity and number of time varying variables in
fully recurrent nets. In ICANN’93, pp. 460–463. Springer, 1993.
[30] Schmidhuber, J. A ‘self-referential’ weight matrix. In ICANN’93, pp. 446–450. Springer, 1993.
[31] Schmidhuber, J. and Huber, R. Learning to generate artificial fovea trajectories for target detection.
International Journal of Neural Systems, 2(01n02):125–134, 1991.
[32] Sohn, K., Zhou, G., Lee, C., and Lee, H. Learning and selecting features jointly with point-wise gated
Boltzmann machines. In Proceedings of The 30th International Conference on Machine Learning, pp.
217–225, 2013.
[33] Springenberg, J. T., Dosovitskiy, A., Brox, T., and Riedmiller, M. Striving for simplicity: The all
convolutional net. arXiv preprint arXiv:1412.6806, 2014.
[34] Srikumar, V., Kundu, G., and Roth, D. On amortizing inference cost for structured prediction. In EMNLPCoNLL ’12, pp. 1114–1124, Stroudsburg, PA, USA, 2012. Association for Computational Linguistics.
[35] Tang, Y., Salakhutdinov, R., and Hinton, G. Robust boltzmann machines for recognition and denoising. In
Computer Vision and Pattern Recognition (CVPR), 2012 IEEE Conference on, pp. 2264–2271. IEEE, 2012.
[36] Team, The Theano Development. Theano: A Python framework for fast computation of mathematical
expressions. arXiv:1605.02688 [cs], May 2016.
[37] Vincent, P., Larochelle, H., Bengio, Y., and Manzagol, P. A. Extracting and composing robust features
with denoising autoencoders. In ICML, pp. 1096–1103. ACM, 2008.
[38] Vinh, N. X., Epps, J., and Bailey, J. Information theoretic measures for clusterings comparison: Variants,
properties, normalization and correction for chance. JMLR, 11:2837–2854, 2010.
[39] von der Malsburg, C. Binding in models of perception and brain function. Current opinion in neurobiology,
5(4):520–526, 1995.
13
[40] von der Malsburg, Christoph. The Correlation Theory of Brain Function. Departmental technical report,
MPI, 1981.
[41] Wersing, H., Steil, J. J., and Ritter, H. A competitive-layer model for feature binding and sensory
segmentation. Neural Computation, 13(2):357–387, 2001.
[42] Xu, K., Ba, J., Kiros, R., Courville, A., Salakhutdinov, R., Zemel, R., and Bengio, Y. Show, attend and tell:
Neural image caption generation with visual attention. arXiv preprint arXiv:1502.03044, 2015.
[43] Yli-Krekola, A., Särelä, J., and Valpola, H. Selective attention improves learning. In Artificial Neural
Networks–ICANN 2009, pp. 285–294. Springer, 2009.
14
A
Supplementary Material
A.1
Notation
Symbol
Space
N
K
H
i
j
k
N
N
N
N
{1, . . . , N }
{1, . . . , K}
x
x̃
zk
mk
δzk
L(mk )
C(x)
RN
RN
RN
RN
RN
RN
R
v
Wh
Wu
Θ
R
RH×4N
R2N ×H
f ()
g()
softmax()
Gj
gk,j
g
p(x | x̃)
q(x)
q(xj | gk,j )
A.2
Description
input dimensionality
total number of groups
input and output dimension of the parametric mapping
iteration index
input element index
group index
input vector with elements xj
corrupted input
the predicted mean of input for group k
probabilities for each input to be assigned to group k
modeling error for group k
group assignment likelihood ratio
the training loss for input x
variance of the input estimate. Only used in the continuous case
Projection weights from tagger inputs to ladder inputs h
Projection weights from ladder output to z and m
Contains all parameters of the ladder
rectified linear activation function
logistic sigmoid activation function
elementwise softmax over the groups
Latent random variable that encodes which group xj belongs to.
Shorthand for Gj = k. Mostly used for p(gk,j ) = p(Gj = k).
a vector of all gj .
posterior of the data given the corrupted data
learnt approximation of p(x | x̃)
Shorthand for q(xj | Gj = k)
Input
In its basic form (without supervision) Tagger receives as input only a datapoint x. It corresponds
to either a binary vector or a real-valued vector and is then corrupted with either bitflip or Gaussian
noise. The training objective is the removal of this noise.
Bitflip Noise In the case of binary inputs we use bitflip noise for corruption:
x̃ = x ⊕ B(β),
where ⊕ denotes componentwise XOR, and B(β) is Bernoulli distributed noise with probability β.
In our experiments on the Shapes dataset we use β = 0.2.
Gaussian Noise If the inputs are real-valued, we corrupt it using Gaussian noise:
x̃ = x + N (0, σ 2 ),
where σ is the standard deviation of the input noise. We used σinput = 0.2.
15
δz0
z
δz1
z
L(m0)
m1
L(m1)
m2
1
q1(x)
2
δz2
L(m2)
q2 (x)
z
m3
δz i-1
zi
L(mi-1)
3
m i-1
z
m0
0
mi
z i-1
PARAM
ETRIC
MAPPIN
G
PARAMETRIC MAPPING
PARAM
ETRIC
MAPPIN
G
PARAM
ETRIC
MAPPIN
G
q3 (x)
x̃
q1 (x|g)
x
x
iteration 1
iteration 2
q1(x)
x̃
iteration 3
Figure 6: Illustration of the TAG framework used for training. Left: The system learns by denoising
its input over iterations using several groups to distribute the representation. Each group, represented
by several panels of the same color, maintains its own estimate of reconstructions z i of the input,
and corresponding masks mi , which encode the parts of the input that this group is responsible for
representing. These estimates are updated over iterations by the same network, that is, each group
and iteration share the weights of the network and only the inputs to the network differ. In the case of
images, z contains pixel-values. Right: In each iteration z i−1 and mi−1 from the previous iteration,
are used to compute a likelihood term L(mi−1 ) and modeling error δz i−1 . These four quantities are
fed to the parametric mapping to produce z i and mi for the next iteration. During learning, all inputs
to the network are derived from the corrupted input as shown here. The unsupervised task for the
network is to learn to denoise, i.e. output an estimate q(x) of the original clean input.
A.3
Group Assignments
Within the TAG framework the group assignment is represented by the K vectors mk which contain
one entry for each input element or pixel. These entries mk,j = q(gk,j ) of mk represent the discreet
probability distribution over K groups for each input xj . They therefore sum up to one:
K
X
mk,j = 1
for all j = 1 . . . N
(7)
k=1
Initialization Similar to expectation maximization, the group assignment is initialized randomly,
but such that Equation 7 holds. So we first sample an auxiliary m0k,j from a standard Gaussian
distribution and then normalize it using a softmax:
m0k,j ∼ N (0, 1)
e
mk,j = PK
(8)
m0k,j
h=1
0
emh,j
(9)
(10)
A.4
Predicted Inputs
Tagger maintains an input reconstruction zk for each group k.
Binary Case In the binary case we use a sigmoid activation function on zk and interpret it directly
as the probability
sigmoid(zk ) = q(x = 1|gk ).
16
(11)
We can use it to compute z̃k = q(x̃|gk ) which will be used for the modeling error (Section A.5) and
the group likelihood:
X
q(x̃ = 1|gk ) =
q(x̃|x, gk )q(x|gk )
(12)
x
=
X
q(x̃|x)q(x|gk )
(13)
(x(1 − β) + (1 − x)β) q(x|gk )
(14)
(x(1 − 2β) + β) zk
(15)
= β(1 − zk ) + (1 − β)zk
= zk (1 − 2β) + β
(16)
(17)
z̃k = x̃(zk (1 − 2β) + β) + (1 − x̃)(1 − zk (1 − 2β) − β)
(18)
x
=
X
x
=
X
x
Therefore we have:
Continuous Case For the continuous case we interpret zk as the means of an isotropic Gaussian
with learned variance v:
(x−zk )2
1
e 2v
q(x|gk ) = N (x; zk , vI) = √
(19)
2πv
Using the additivity of Gaussian distributions we directly get:
z̃k = q(x̃|gk ) = N (x̃; zk , (v + σ 2 )I)
(20)
Initialization For simplicity we initialize all zk to the expectation of the data for all k. In our
experiments these values are 0.5 for the TextureMNIST datasets and 0.26 for the Shapes dataset.
A.5
Modeling Error
As explained in Section 2.1, δz carries information about the remaining modeling error. During
training as a denoiser, we can only allow information about the corrupted x̃ as inputs but not about
the original clean x. Therefore, we use the derivative of the cost on the corrupted input as helpful
information for the parametric mapping. Since we work with the input elements individually we skip
the index j in the following:
δz k ∝ −∂C(x̃)/∂zk .
(21)
More precisely for a single iteration (omitting the index i) we have::
∂C(x̃)
δzk = −
(22)
∂zk
!
X
∂
=
log
q(x̃|gh )q(gh )
(23)
∂zk
h
P
∂ h q(x̃|zh )q(gh )
1
=P
(24)
∂zk
h q(x̃|zh )q(gh )
1
∂ z̃k
=P
mk
(25)
z̃
m
h h h ∂zk
(26)
Continuous Case For the continuous case this gives us:
1
∂ z̃k
δzk = P
mk
(27)
z̃
m
h h h ∂zk
1
x̃ − zk
=P
z̃ m
(28)
2+v k k
z̃
m
σ
h h h
∝ (x̃ − zk )mk z̃k
(29)
Note that since the network will multiply its inputs with weights, we can always omit any constant
multipliers.
17
Binary Case Let us denote the corruption bit-flip probability by β and define
ξk := q(x̃ = 1|gk ) = (1 − 2β)zg + β .
Then we get:
z̃k = x̃ξk + (1 − x̃)(1 − ξk )
and thus:
∂ z̃k
1
mk
h z̃h mh ∂zk
(x̃(1 − 2β) − (1 − x̃)(1 − 2β))mk
= P
h (x̃ξh + (1 − x̃)(1 − ξh ))mh
δzk = P
(30)
(31)
which simplifies for x̃ = 1 as
mk
(1 − 2β)mk
≈ −P
= P
h ξh mh
h ξh mh
and for x̃ = 0 as
=
(1 − 2β)mk
m
mk
P
Pk
≈
=P
1 − h ξh mh
1 − h ξh mh
h ξh m h − 1
Putting it back together:
mk
h ξh mh − 1 + x̃
δzk = P
A.6
Ladder Modifications
We mostly used the specifications of the Ladder network as described by Rasmus et al. [22], but there
are some minor modifications we made to fit it to the TAG framework. We found that the model
becomes more stable during iterations when we added a sigmoid function to the gating variable v [22,
Equation 2] used in all the decoder layers with continuous outputs. None of the noise sources or
denoising costs were in use (i.e., λl = 0 for all l in Eq. 3 of Ref. [22]), but Ladder’s classification
cost (Cc in Ref. [22]) was added to the Tagger’s cost for the semi-supervised tasks.
All four inputs (zki , mik , δzki , and L(mik )) were concatenated and projected to a hidden representation
that served as the input layer of the Ladder Network. Subsequently, the values for the next iteration
were simply read from the reconstruction (x̂ in Ref. [22]) and projected linearly into zki+1 and via
softmax to mi+1
to enforce the conditions in Equation 7. For the binary case, we used a logistic
k
sigmoid activation for zki+1 .
18
A.7
Pseudocode
In this section we put it all together and provide the pseudocode for running Tagger both on binary
(Algorithm 3) and real-valued inputs (Algorithm 2). The provided code shows the steps needed to run
for T iterations on a single example x using G groups. Here we use three activation functions: f (x) =
max(x, 0) is the rectified linear function, g(x) = 1+e1−x is the logistic sigmoid, and softmax(x)g =
xg
PGe
xh is a softmax operation over the groups. All three include a batch-normalization operation,
h=1 e
which we omitted for clarity. Only the forward pass for a single example is shown, but derivatives of
the cost C wrt. parameters v, Wh , Wu and Θ are computed using regular backpropagation through
time. For training we use ADAM with a batch-size of 100.
Data: x, K, T, σ, v, Wh , Wu Θ
Result: z T , mT , C
begin Initialization:
x̃ ← x + N (0, σ 2 I);
m0 ← softmax(N (0, I));
z 0 ← E[x];
end
for i = 0 . . . T − 1 do
for k = 1 . . . K do
z̃k ← N (x̃; zki , (v + σ 2 )I);
δzki ← (x̃ − zki )mik z̃k ;
L(mik ) ← Pz̃kz̃h ;
h
hk ← f (Wh zki , mik , δzki , L(mik ) );
[zki+1 , mi+1
k ] ← Wu Ladder(hk , Θ);
end
mi+1 ← softmax(mi+1 );
PK
qi+1 (x) ← k=1 N (x; zki+1 , vI)mi+1 ;
end
PT
C ← − i=1 log qi (x);
Algorithm 2: Pseudocode for running Tagger on a single real-valued example x.
Data: x, K, T, β, Wh , Wu , Θ
Result: z T , mT , C
begin Initialization:
x̃ ← x ⊕ B(β);
m0 ← softmax(N (0, I));
z 0 ← E[x];
end
for i = 0 . . . T − 1 do
for k = 1 . . . K do
ξk ← z i (1 − 2β) + β;
mik
;
ξh mih −1+x̃
x̃ξ
+(1−x̃)(1−ξ
k
k)
i
L(m ) ← P x̃ξh +(1−x̃)(1−ξ
;
h)
h i
i
i
hk ← f (Wh zk , mk , δzk , L(mik ) );
[zki+1 , mi+1
k ] ← Wu Ladder(hk , Θ);
δzki ←
P
h
end
mi+1 ← softmax(mi+1 );
PK
qi+1 (x) ← k=1 N (x; zki+1 , vI)mi+1 ;
end
PT
C ← − i=1 log qi (x);
Algorithm 3: Pseudocode for running Tagger on a single binary example x.
19
| 9cs.NE
|
JOURNAL OF COMPUTER SCIENCE AND ENGINEERING, VOLUME 1, ISSUE 1, MAY 2010
72
An approach to find dynamic slice for C++
Program
Santosh Kumar Pani and Priya Arundhati
Abstract—Object-oriented programming has been considered a most promising method in program development and maintenance. An
important feature of object-oriented programs (OOPs) is their reusability which can be achieved through the inheritance of classes or reusable components.Dynamic program slicing is an effective technique for narrowing the errors to the relevant parts of a program when debugging. Given a slicing criterion, the dynamic slice contains only those statements that actually affect the variables in the slicing criterion. This
paper proposes a method to dynamically slice object-oriented (00) programs based on dependence analysis. It uses the Control Dependency
Graph for object program and other static information to reduce the information to be traced during program execution. In this paper we
present a method to find the dynamic Slice of object oriented programs where we are finding the slices for object and in case of function overloading.
Index Terms— Program Slicing, Dynamic Slicing, Control dependency Graph.
—————————— ——————————
1 INTRODUCTION
Program slicing is an effective technique for narrowing
the focus of attention to the relevant parts of a program.
The slice of a program consists of those statements and
predicates that may directly or indirectly affect the variables computed at a given program point [1]. The program point s and the variable set V, denoted by <s, V>, is
called a slicing criterion. The basic idea of program slicing
is to remove irrelevant statements from source codes
while preserving the semantics of the program such that
at the specified point the variable produces the same value as its original program. Program slicing has been
widely used in many software activities, such as software
analyzing, understanding, debugging, testing, maintenance, and so on [2, 3, 4]. Slicing algorithm can be classed
according to whether they only use static information
(static slicing) or dynamic execution information for a
specific program input (dynamic slicing). This paper focuses on the dynamic slicing methods for finding the slices for class,object and in case of function overloading.
2 REVIEW OF RELATED WORKS
Agrawal proposes a dynamic slicing method by marking
nodes or edges on a static program dependence graph
during execution [5]. The result is not precise, because
some dependencies might not hold in dynamic execution.
Agrawal also proposes a precise method based on the
dynamic dependence graph (DDG) [5], and Zhao applies
it to slice object-oriented programs [6]. The shortcoming
is that the size of the DDG is unbound.
Korel [8], Song [7] and Tibor [9] propose forward dynamic slicing methods and Song also proposes method to slice
OOPs using dynamic object relationship diagram
(DORD). In these methods, they compute the dynamic
slices for each statement immediately after this statement
is executed. After the last statement is executed, the dynamic slices of all statements executed have been obtained. However, only some special statements in loops
need to compute dynamic slices. In this paper we
represent the object oriented program in control dependence Graph (CDG).
3 BASIC CONCEPT & DEFINITION
Slicing object oriented program is quite difficult as it
comprises of many features like Class, Object, Polymorphism, inheritance etc. In this paper we are trying to finding the slices for object and in case of function overloading.
In object oriented programming language it is possible to
define more then one function that share a common name
this feature is called function overloading.Hence our aim
is to find the Slice where the function is overloaded.In
that case the Slicer should know that which version of the
function to call.For this,the Slicer uses the following information about the function Call.
1. The type of the argument.
2. The no.of argument.
3. The name of the function.
After getting these information from the calling node the
unquiue function will be called and the slice of the function will be found out in normal method.
The following Data Structures are used for this purpose.
————————————————
Santosh Kumar Pani is with the School of Computer Engineering, KIIT
University, Bhubaneswar.
Priya Arundhati is with the Department of CSE, Nalanda Institute of
Technology, Bhubaneswar.
1. Control Dependence Graph: The control depen-
dence graph (CDG) G of an object oriented program P is a graph G=(N,E),where each node nεN
represents a statement of the program P. For any
© 2010 JCSE
http://sites.google.com/site/jcseuk/
73
2.
3.
4.
5.
6.
7.
pair of nodes x and y,(x,y)εE if node x is control
dependent on node y.
Def(dvar) : If dvar is a variable in a program P.A
node u of CDG Gp is said to be a Def(dvar) node
if u is definition statement that defines the variable dvar.
Use(dvar) : If dvar is a variable in a program P.A
node u of CDG Gp is said to be a Use(dvar) node
if he statement u uses the value of the variable
dvar.
Defvarset(u):Let u be a node in the CFG Gp of a
program P.The set DefVarSet(u)={dvar:dvar is a
data variable in the program P,and u is a
Def(dvar) node
Usevarset(u):Let u be a node in the CFG Gp of
program P. The set UseVarSet(u)={dvar:dvar is a
data variable in the program P,and u is a
Use(dvar) node}.
ActiveControlSlice(s): Let s be a test
node(predicate statement) in, the CDG Gp of a
program
P
and
UseVarSet(s)={var1,var2,…..vark}. Before execution of
the program P, ActiveControlSlice(s) = Ø.After
each execution of the node s in the actual run of
the
program,
ActiveControlSlice(s)={s}UActiveDataSlice(var1)…UActiveDat
aSlice(vark) U ActiveControlSlice(t), where t is
most recently executed successor node of s in
Gp.If s is a loop control node and the present execution of the node s corresponds to exit from
loop,then ActiveControlSlice(s)= Ø
ActiveDataSlice(var):Let dvar be a data variable
in the object oriented program P. Before execution
of
the
program
P
,ActiveDataSlice(obj.var)=Ø.Let
u
be
a
Def(obj.var)
node,
and
UseVarSet(u)={var1,var2,..vark).After each execution of
the node u in the actual run of the program, ActiveDataSlice(obj.var)={s}U ActiveDataSlice(var1)
U………UActiveDataSlice(dvark) U ActiveControlSlice(t), where t is most recently executed
successor node of s in Gp.
/*For each data member of an object a separate
slice will be mainted in a shared memory which
is only accessable to the member function called
on that object.*/
8. DyanSlice(obj):-Let obj be an object of the class
in the proram P. Before execution of program P,
DyanSlice(obj)=ф. Let the object contains the data
member mem1.mem2,…memn then DyanSlice(obj)=DyanSlice(obj.mem1)
U
DyanSlice(obj.mem2)U…..U DyanSlice(obj.memn)
9. DyanSlice(s,var):-Let s be a node of the CDG of a
program P,& var be a variable in the set i.e. var(DefVarSet(s) U UseVarSet(s) Before executionof
the program P DyanSlice(s,var)=фAfter each ex-
ecution of the node s in an actual run DyanSlice(s,var)=Active Data Slice(var) U Active Control Slice(t)
10. Active Call Slice:-Let Gp represents the CDG of
the multiprocedure program P. Before execution
of the program P Active Call Slice=ф. Let Uactive
be an active Call node then Active Call
Slice={Uactive} U Active Call Slice U Active Control Slice(t).
11. Call Slice Stack:-This stack is maintained to keep
track of the Active Call slice during the actual
run of the program.
12. Formal(x,var),Actual(x,var)): Let p1 be a procedure of a program p having multiple procedures,
and x be a calling node to the procedures, and x
be a calling node to the procedure p1.Let f be a
formal parameter of procedure P1 and its corresponding actual parameter at the calling node to
the procedure P1 and its corresponding actual
parameter at the calling node x be a. We define
Formal(x,a)=f and Actual(x,f)=a.
4
AN ALGORITHM FOR FINDING DYNAMIC SLICE
FOR OBJECT ORIENTED PROGRAMS IN CASE OF
FUNCTION OVERLOADING.
Step 1: Start
Step2: At first Control Dependence Graph G of object
oriented program P is statically constructed .
Step 3: Before each execution of the program do step 4 to
step 7.
Step 4: For each node u of G do step 5 to step 6.
Step 5 : If u is a test node, then ActiveControlSlice(u) = Ø.
Step 6: For each variable var εDefVarSet(u) U UseVarSet(u) do DyanSlice(u,var) = Ø.
Step 7: CallSliceStack=NULL.
ActiveCallSlice= Ø.
Step 7a. For each data variable dvar of the program
do
ActiveDataSlice(var)= Ø.
Step 7b. If var is a data member of an corresponding class
of an object to then
do
ActiveDataSlice(obj.var) = Ø
for each member declared with in the class of an
object
Step 8: Run the program P with the given set of input
values and repeat steps 9 to 19 until the program terminates.
Step 9: Before execution of each call node u do step 10
and step 11
Step 10a. Let u be a call node to a procedure Q, update
CallSliceStack and ActiveCallSlice.
//In case of function having same name the signature of
the function is foundout as each function have unquiue
signature ,hence the slicer will know that which version
of the function to call.//
Step 10b. For each actual parameter var in the procedure
74
call Q do ActiveDataSlice (Formal(u,var)) = ActiveDataSlice(var) U ActiveCallSlice.
//Here when a member finction of an class is called with
an object then the actual parameter of the function is individually copied to the formal parameter of the member
function.//
Step 11:Update ActiveReturnSlice before execution of the
return node.
Step 12: After execution of each node u of the program P
do step 13 to step 18.
Step 13:If u is a Def(var) node and not a call node then
Update ActiveDataSlice(var).
//var can be a data variable or data member of an object
then Update ActiveDataSlice(var) will be done accordingly //
Step 14: If u is a call node to a procedure Q then for every
formal reference parameter var in the procedure Q do
ActiveDataSlice(Actual(u,var))= ActiveDataSlice(var).
Step 15: If u is a Def(var) node then ActiveDataSlice(var)=ActiveReturnSlice.
Step16a. For every variable var if declared as automatic
local in the procedure Q do ActiveDataSlice= Ø.
//For local static variable and global variable the ActiveDataSlice remains unchanged
Step16b. Update CallSliceStack and ActiveCallSlice and
Set ActiveReturnSlice= Ø .
Step 17: For every variable var ε DefVarSet(u) U UseVarSet(u) Update DyanSlice(u,var).
Step 18: If u is a test node, then update ActivecontrolSlice(u).
Step 19:Exit
22.b=tp1.b+tp2.b;
}
test add(test tp3,int s)
{
23.a=tp3.a +s;
24.b=tp3.b +s;
}
};
void main()
{
test T1,T2,T3,T4;
int p,q;
1.Cout<<”Enter the value of p”;
2.Cin>>p;
3.Cout<<”Enter the value of q”;
4.Cin>>q;
5.T1.get(p,q);
6.T1.display();
7.Cout<<”Enter the value of p”;
8.Cin>>p;
9.Cout<<”Enter the value of q”;
10.Cin>>q;
11.T2.get(p,q)
12.T2.display()
13.T3.add(t1,t2);
14.T3.display();
15.T4.add(T3,5);
16.T4.display();
}
Control Dependency Graph
5 SAMPLE PROGRAM AND WORKING OF THE
ALGORITHM
Sample Program
The following object oriented Program where only the
concept of class and object are used is discussed to iilustrate our proposed Algorithm.
Class test
{
int a;
int b;
public:
void get(int x,int y)
{
17.a=x;
18.b=y;
}
Void display()
{
19.Cout<<a;
20.Cout<<b;
}
Test add (test tp1,test tp2)
{
21.a=tp1.a+tp2.a;
Working of the Algorithm
After Execution of node-2
ADS(p)={2}
DyanSlice(2,p)={2}
After Execution of node-4
ADS(q)={4}
75
DyanSlice(4,q)={4}
After Execution of node-5
Here one of the object of test class is T1.get() is a member
function of the test class.The formal parameter of get() is x
& y as well as these are the local variable to get().The actual parameter is p & q.
So,ADS(x)=ADS(p) U Active Call Slice
={2} U {5}={2,5}
ADS(y)=ADS(q) U Active Call Slice
={4} U{5}={4,5}
Again T1 is calling the get() so here a & b are T1’s variable
ADS(T1.a)=ADS(x)U{15}={2,5} U{15}={2,5,17}
DyanSlice(15,T1.a)={2,5,17}
ADS(T1.b)=ADS(y)U{16}={4,5,18}
DyanSlice(16,T1.b)={4,5,18}
After Execution of node-6
Now T1 is calling display()
So,DyanSlice(17,T1.a)=ADS(T1.a)={2,5,17}
DyanSlice(18,T1.b)=ADS(T1.b)={4,5,18}
DyanSlice(T1)=DyanSlice(T1.a) U DyanSlice(T1.b)
={2,5,17} U {4,5,18}
={2,4,5,17,18}
After Execution of node-8
ADS(p)={8}
DyanSlice(8,p)={8}
After Execution of node-10
ADS(q)={10}
DyanSlice(10,q)={10}
After Execution of node-11
ADS(x)=ADS(p) U Active Call Slice
={8} U{11}={8,11}
ADS(y)=ADS(q) U Active Call Slice
={10}U{11}={10,11}
ADS(T2.a)=ADS(x) U 17
={8,11,17}
DyanSlice(17,T2.a)={8,11,17}
ADS(T2.b)=ADS(y) U 18
={10,11} U 18={10,11,18}
After Execution of node-12
DyanSlice(19,T2.a)=ADS(T2.a)={8,11,17}
DyanSlice(20,T2.b)=ADS(T2.b)={10,11,18}
DyanSlice(T2)=DyanSlice(T2.a) U DyanSlice(T2.b)
={8,11,17} U {10,11,118}
={8,10,11,17,18}
After Execution of node-13
Here T3 is calling the add() where the formal parameter
are tp1 & tp2 but here the object T1 & T2 are passed as the
actual parameter.and there are two version of
add()function.Now the Slicer will decide which function
to call by checking the signature of the function.when this
node will be executed the control will go to the first
add()function where the argument are the objects.
So,ADS(tp1.a)=ADS(T1.a) U Active Call Slice
={2,,5,17} U {13}
={2,5,17,13}
ADS(tp1.b)= ADS(T1.b) U Active Call Slice
={4,5,18} U 13
={4,5,18,13}
ADS(tp1)=ADS(tp1.a) U ADS(tp1.b)={ 2,5,17,13}
{4,5,18,13}={ 2,5,17,13,4,5,18,13}
ADS(tp2.a)=ADS(T2.a) U Active Call Slice
={8,11,17} U {13}
={8,11,17,13}
ADS(tp2.b)= ADS(T2.b) U Active Call Slice
={10,11,18} U 13
={10,11,18,13}
ADS(tp2)=ADS(tp2.a) U ADS(tp2.b)={ 8,11,17,13}
{10,11,18,13}={ 8,10,11,13,17,18}
ADS(T3.a)=ADS(tp1.a) U ADS(tp2.a) U {21}
={2,,5,13,17} U {8,11,17,13} U{21}
={2,5,8,11,13,17,21}
ADS(T3.b)= ADS(tp1.b) U ADS(tp2.b) U{20}
={4,5,13,18} U {10,11,13,18} U{22}
={4,5,10,11,13,18,22,}
DyanSlice(13,T3.a)=ADS(T3.a)={2,5,8,11,13,17,21,}
DyanSlice(13,T3.b)=ADS(T3.b)={4,5,10,11,13,18,22}
After Execution of node-14
DyanSlice(19,T3.a)=ADS(T3.a)={ 2,5,8,11,13,17,21}
DyanSlice(18,T3.b)=ADS(T3.b)={ 4,5,10,11,13,18,22}
DyanSlice(T3)=DS(T3.a) U DS(T3.b)
=
{2,5,8,11,13,17,21}
{4,5,10,11,13,18,22}={2,4,5,8,10,11,13,17,18,21,22}
After Execution of node-15
ADS(tp3.a)=ADS(T3.a) U Active Call Slice
={2,5,8,11,13,17,21} U 15
={2,5,8,11,13,15,17,21}
ADS(tp3.b)=ADS(T3.b) U Active Call Slice
={4,5,10,11,13,18,22,} U15
={4,5,10,11,13,15,18,22,}
ADS(T4.a)=ADS(tp3.a) U Active Call Slice
={2,5,8,11,13,15,17,21} U{23}
={2,5,8,11,13,15,17,21,23}
ADS(T4.b)=ADS(tp3.b) U Active Call Slice
={4,5,10,11,13,15,18,22,} U {24}
={4,5,10,11,13,15,18,22,24}
After Execution of node-16
DyanSlice(16,T4.a)=ADS(T4.a)
={2,5,8,11,13,15,17,21,23}
DyanSlice(16,T4.b)=ADS(T4.b)
={4,5,10,11,13,15,18,22,24}
DyanSlice(T4)= DyanSlice(T4.a) U DyanSlice(T4.b)
={2,5,8,11,13,15,17,21,23}
{4,5,10,11,13,15,18,22,24}
={2,4,5,8,10,11,13,17,18,21,22,15,23,24}
6
U
U
U
U
ANALYSIS OF THE ALGORITHM
The improved InterSlice Algorithm also uses a collection of
control dependence graphs as the intermediate program
representation & computes precise dynamic slice. The space
complexity of the improved Interslice algorithm is also
O(n2)The time complexity of this algorithm is linear as no
extra data structures are required for this enhancement .This
algorithm does not require to store the past execution histo-
76
ry.
7
CONCLUSION
In this paper of the dynamic Slicing of Object Oriented Programs is discussed. The Object Oriented Programs has various features like Class, Object, Dynamic Binding, Polymorphism, Inheritance etc. In this proposed Algorithm only the
Class, Object and function overloading are discussed. Finding the Slices for Inheritance, Dynamic binding and other
features are the future work of this Research
REFERENCES
[1]
M. Weiser, Program Slicing, IEEE Trans. Software Engineering,
1984, 16(5): 498-509.
[2] H. Agrawal, et al. Debugging with Dynamic Slicing and Backtracking. Software – Practice and Experience, 1993, 23(6): 589616
[3] B. Korel. Application of Dynamic Slicing in Program Debugging. Third International Workshop on Automated Debugging,
1997: 59-74
[4] F. Tip. A Survey of Program Slicing Techniques, J. Programming Languages, 1995, 3(3): 121-189
[5] H. Agrawal, J. Horgan. Dynamic Program Slicing. Proceedings
of the ACM SIGPLAN’90 Conference on Programming Language Design and Implementation, 1990: 246-256
[6] J. Zhao, Dynamic Slicing of Object-Oriented Programs. Technical-Report SE-98-119, Information Processing Society of Japan, May 1998: 17-23
[7] Y. Song, D. Huynh, Forward Dynamic Object-Oriented Program Slicing, Application-Specific Systems and Software Engineering and Technology (ASSET '99), IEEE CS Press, 1999: 230 –
237
[8] B. Korel, S. Yalamanchili. Forward Derivation of Dynamic Slices. Proceedings of the International Symposium on Software
Testing and Analysis, 1994, pp. 66-79.
[9] G, Tibor, et al. An Efficient Relevant Slicing Method for Debugging, Software Engineering Notes, Software Engineering
―
―
ESEC/FSE’99 Springer ACM SIGSFT, 1999: 303-321
[10] G.B. Mund, R. Mall, S. Sarkar,2002a. An efficient dynamic program slicing technique. Information and Software Technology
44(2),123-132.
| 6cs.PL
|
Noname manuscript No.
(will be inserted by the editor)
Estimating Distances via Received Signal Strength and Connectivity in
Wireless Sensor Networks
arXiv:1801.09350v1 [eess.SP] 29 Jan 2018
Qing Miao, Baoqi Huang∗ and Bing Jia
Received: date / Accepted: date
Abstract Distance estimation is vital for localization and
many other applications in wireless sensor networks (WSNs).
Particularly, it is desirable to implement distance estimation
as well as localization without using specific hardware in
low-cost WSNs. As such, both the received signal strength
(RSS) based approach and the connectivity based approach
have gained much attention. The RSS based approach is
suitable for estimating short distances, whereas the connectivity based approach obtains relatively good performance
for estimating long distances. Considering the complementary features of these two approaches, we propose a fusion
method based on the maximum-likelihood estimator (MLE)
to estimate the distance between any pair of neighboring
nodes in a WSN through efficiently fusing the information
from the RSS and local connectivity. Additionally, the method
is reported under the practical log-normal shadowing model,
and the associated Cramer-Rao lower bound (CRLB) is also
derived for performance analysis. Both simulations and experiments based on practical measurements are carried out,
and demonstrate that the proposed method outperforms any
single approach and approaches to the CRLB as well.
Keywords Distance Estimation, Maximum-Likelihood
Estimator, Error Distributions, Cramer-Rao lower bound
1 INTRODUCTION
Wireless sensor networks (WSNs), composed of hundreds
or thousands of small and inexpensive nodes with constrained
computing power, limited memories, and short battery lifetime, can be used to monitor and collect data in a region
of interests. Accurate and low-cost node localization is imInner Mongolia University, Hohhot, 010021, China,
E-mail: [email protected], [email protected]
[email protected]
and
ji-
portant for various applications in WSNs, and thus, great
efforts have been devoted to developing various localization
algorithms, categorized into distance based algorithms and
connectivity-based algorithms [1]. The distance-based localization algorithms rely on distance estimates and are able to
achieve relatively good localization accuracy, whereas the
connectivity-based localization algorithms generally achieve
coarse-grained localization accuracy since only local connectivity information (the numbers of the common and noncommon one-hop neighbors) is employed for distance estimation. Besides, distance estimation is also useful for sensor network management, such as topology control [2,3] and
boundary detection [4, 5].
In reality, distance estimation can be realized by using
information such as RSS, time of arrival (TOA) and time
difference of arrival (TDOA) [1]. The RSS approach (using
RSS measurements) does not require any dedicated hardware, but is able to provide coarse-grained distance estimates; in contrast, the TOA and TDOA methods can provide distance estimates with high accuracy at the cost of
extra hardware, but it is unaffordable to equip each sensor
with a dedicated measurement device in a large-scale WSN
due to the costs in both hardware and energy. Therefore, it
is of great importance to enhance the accuracy of low-cost
distance measurement approaches, and many efforts have
been imposed in the literature. Due to its intrinsic simplicity and independence of dedicated hardware, the RSS-based
distance methods have gained much attraction [6–9]. But,
both simulations and theoretical analysis indicate that the
performance of the RSS-based methods is relatively poor,
and degrades with the increasing actual distances [6, 7, 10].
Hence, it is necessary to advance the RSS-based methods by
adopting various techniques [11, 12].
Apart from RSS measurements, local connectivity information of each node is also independent of extra device, and
can be employed to estimate distances from itself to other
2
neighboring nodes [10, 13, 14]. However, Unlike the RSSbased methods that return ideal estimates for short distances,
the connectivity-based methods obtain relatively good performance for estimating long distances. Therefore, the complementary features of these two types of distance estimation methods motivate us to design a fusion method which
is able to sufficiently exploit the advantages of both types of
methods.
In this paper, the maximum-likelihood estimator (MLE)
is utilized to efficiently fuse the information from the RSS
measurement and local connectivity, so as to provide good
performance regardless the actual sizes of distances to be
estimated. The strengths of the proposed fusion method lie
in the following aspects. Firstly, it is well known that the
MLE is asymptotically optimal, indicating that the proposed
method is able to obtain superior performance. Secondly, the
practical log-normal (shadowing) model is adopted to characterize the probabilistic distributions of the errors respectively induced by the RSS-based method and the connectivitybased method, so as to ensure the applicability of the proposed method. Thirdly, the Cramer-Rao lower bound (CRLB)
associated with the proposed method is also formulated and
can be used to evaluate the optimality of the proposed method.
Finally, both experiments based on measurements in a real
environment and simulations are carried out to thoroughly
validate the effectiveness of the proposed method. However,
even if the proposed method outperforms the other two available methods and approaches to the CRLB in most time, it
still suffers from a few limitations: it is computationally expensive due to the complicated cost function involved; as
the connectivity-based method, the performance also relies
on the sensor density. This study not only contributes to
improving sensor localization in low-cost WSNs, but also
paves the way for advancing many other researches and applications relying on inter-node distances.
The remainder of the paper is organized as follows. Section 2 briefly reviews related works in the literature. Section
3 introduces the WSN model and both the RSS-based and
connectivity-based distance estimation methods. Section 4
presents the models of both the RSS-based and connectivitybased methods, and proposes the fusion method based on the
MLE and formulates the corresponding CRLB. Section 5 reports the performance of the proposed method through both
simulations and experiments. Finally, section 6 concludes
the paper and sheds lights on future works.
Qing Miao, Baoqi Huang∗ and Bing Jia
ceiver as long as after the model depicting the relationship
between power losses and distances is available. In [6], an
estimator was designed based on the MLE and the log-normal
model, but the theoretical analysis indicated that the estimator is inefficient in the sense that the error variance increases exponentially with powers. However, the practice
in [7, 9] reveal that the RSS-based distance estimation is unreliable. Additionally, the distance estimation is even complicated in indoor environments since the factors, like furniture, hand grip and human bodies, affect the distance estimation [8]. Hence, in [11] a dynamic calibration method was
proposed to update the log-normal model parameters which
fluctuate with environmental changes, and in [12], an averaging method based on multichannel RSS measurements
was presented to mitigate the variability of RSS measurements. Therefore, it can be concluded that it is still challenging to apply the RSS-based methods.
The connectivity-based methods infer distances from local connectivity information among different nodes in WSNs
[10, 15, 16]. The neighborhood intersection distance estimation scheme (NIDES) presented in [15] heuristically relates
the distance, e.g. from node A to node B, to an easily observed ratio, i.e. the number of their common immediate
neighbors to the number of immediate neighbors of A, and
then performs distance estimation at node A according to
this ratio and other a priori known information. NIDES assumes a unit disk model, namely that the communication
coverage of each node is a perfect disk, and all nodes are
uniformly and randomly deployed in the WSN. Its enhanced
version presented in [16] adapted the ratio by taking into account the number of immediate neighbors of node B, and
heuristically stated that NIDES could be applied under arbitrary communication models. In [10], a novel method is presented based on the MLE under a generic channel model,
including the unit disk model and the more realistic lognormal model, and its error characteristics were analyzed in
light of the CRLB. However, the performance of the connectivitybased method obtains obvious errors when estimating short
distances.
In summary, both of the above methods are restricted in
practical applications, but fortunately, are complementary to
each other, which motivates us to combine them to obtain
better performance. As such, this paper presents a fusion
method to estimate distances by making use of RSS measurements and local connectivity under the practical lognormal model.
2 RELATED WORKS
3 PRELIMINARIES
In this section, we shall briefly review the studies on distance
estimation in low-cost WSNs, which can be categorized into
RSS-based methods and connectivity-based methods.
The RSS-based methods infer distances from power losses
incurred by signals travelling between transmitter and re-
This section first briefly introduces the static WSN model
which is considered in this paper, and then elaborates the
RSS-based and connectivity-based methods, respectively. Throughout this paper, we shall use the following mathematical no-
Estimating Distances via Received Signal Strength and Connectivity in Wireless Sensor Networks
3
tation: p(·) denotes the probability density function of an
event, and E(·) denotes the statistical expectation.
3.1 The WSN Model
In a static WSN, nodes are often assumed to be randomly
and uniformly distributed on account of the random nature
of network deployment, e.g. nodes being dropped from a flying plane. Since a homogeneous Poisson process provides
an accurate model for the uniform distribution of nodes as
the network size approaches infinity, we define the static
WSN to be deployed over an infinite plane according to the
homogeneous Poisson process of intensity λ.
Fig. 1 The communication coverage of two nodes under the unit disk
model.
3.2 The RSS-based Method
The RSS-based method estimates the distance between any
pair of nodes using the received signal power, i.e. RSS. When
a signal is propagated between transmitter and receiver, the
power loss or attenuation is unavoidable, and generally rises
with increasing the separation between transmitter and receiver. Moreover, as is commonly made in both theoretical studies (e.g. [17–19]) and experimental studies (e.g. [20,
21]), the power loss can be formulated by using the lognormal model, namely
PR (d)(dBm) = PR (d0 )(dBm) − 10α log10
d
+ Z,
d0
(1)
where PR (d)(dBm) is the received signal power at d in dBm,
PR (d0 )(dBm) is the mean received signal power at a reference distance d0 in dBm, α is the path loss exponent, and
Z is a random variable representing the shadowing effect,
2
normally distributed with mean zero and variance σdB
.
Based on the log-normal model in (1), it is straightforward to infer the distance d from the received signal power
PR (d) by using any parameter estimator. For instance, dˆR is
defined to be the distance estimate between two nodes via
an associated RSS measurement, and can be formulated as
follows (see [22] for more details)
dˆR = 10
PR (d0 )(dBm)−PR (d)(dBm)
10α
d0 .
(2)
3.3 The Connectivity-based Method
The connectivity-based method estimates the distance between any pair of neighboring nodes on the basis of their local connectivity information. In this subsection, we present
this method under both the simple unit disk model and the
generic channel model. Specifically, the unit disk model assumes an ideal communication coverage for each node, i.e.
a perfect disk with the radius of r, whereas the generic channel model, including the log-normal model, takes into consideration the random noises (e.g. the shadowing effect) in
the communication channels, so as to characterize the communication coverage in a more practical way.
3.3.1 The Unit Disk Model Case
Given a static WSN, suppose two nodes A and B with coordinates (xA , yA ), (xB , yB ) and separation d (d ≤ r) and
the disks with the common radius r represent their individual communication coverage under the unit disk model, as
shown in Fig. 1. Because of d ≤ r, the two disks intersect
and create three disjoint regions. Regarding r as constant,
define S = πr2 and f (d) to be the area of the middle region
in Figure 1, where
r
d
d2
2S
− d r2 − .
arccos
(3)
f (d) =
π
2r
4
It is obvious that the nodes residing in the middle region
are common immediate neighbors of A and B, the nodes residing in the left (or right) one are non-common immediate
neighbors of A (or B). Define three random variables M , P ,
and Q to be the numbers of the three categories of neighbors; according to the assumption of the Poisson point process, they are mutually independent and Poisson with means
λf (d), λ(S − f (d)) and λ(S − f (d)), as pointed out in [23].
However, the actual values of M , P , and Q can be easily
obtained after A and B exchange their neighborhood information. On the basis of the observations of M , P and Q and
the method of MLE, the distance estimate of d, denoted dˆc ,
can be summarized as follows (see [10] for details)
(
f −1 (S),
if M = P = Q = 0;
ˆ
dc =
(4)
−1
f (ρ̂S), otherwise
4
Qing Miao, Baoqi Huang∗ and Bing Jia
2M
2M+P +Q .
where ρ̂ =
3.3.2 The Generic Channel Model Case
In the generic channel model [10], the randomness on the
RSS can be characterized by a function g(d), denoting the
probability that a directional communication link exists from
transmitter to receiver with distance d. In particular, in the
log-normal model, we can have
g(d) =
2
exp − 2σz2
dB
∞
Z
k log
√
dz
2πσdB
d
r
(5)
where k = 10α/ log 10; r denotes a pseudo transmission
range which depends on the antenna gains, the wavelength
of the propagating signal, the transmission power and the
communication threshold for RSS.
Let M , P , and Q continuously denote the numbers of
common and non-common immediate neighbors associated
with two nodes. we can compute their expectations as follows
E(M + P ) = E(M + Q) =
Z ∞Z ∞
p
λ
×g( (x − xB )2 + (y − yB )2 )dxdy,
−∞
(6)
4 The Proposed Fusion Method
In this section, we introduce the new distance estimation
method based on the aforementioned RSS-based method and
connectivity-based method. To do so, it is necessary to understand the statistical distributions of the RSS-based and
connectivity-based distance estimates, respectively. As such,
a thorough theoretical analysis is carried out to investigate
both of the distance estimation methods. After that, the MLE
method can be instantly applied to estimate the distance,
and particularly, the Newton-Raphson method is adopted
to solve the corresponding likelihood function. Besides, the
CRLB associated with the proposed fusion distance estimation method is formulated to further observe its performance.
4.1 Statistical Analysis of Distance Estimates
−∞
E(M ) =λ
Z
∞
−∞
Z
∞
−∞
p
g( (x − xA )2 + (y − yA )2 )
(7)
p
× g( (x − xB )2 + (y − yB )2 )dxdy.
Then, by generalizing S and f (d) to specify the expectations of M , P and Q under the generic channel model instead of the areas defined under the unit disk model, we can
have the following formulas
Z ∞Z ∞ p
S=
g( (x − xB )2 + (y − yB )2 )dxdy,
(8)
−∞
f (d) =
However, since the closed-form formulae for f (d) and
its inverse are hard or even impossible to obtain, we thus
substitute its inverse by using an approximate piecewise linear function, namely that a linear regression model is established to predict d for each affine segment.
Z
−∞
∞
−∞
Z
∞
−∞
p
g( (x − xA )2 + (y − yA )2 )
(9)
p
× g( (x − xB )2 + (y − yB )2 )dxdy.
Moreover, by using (5) and (9), we can derive the formula for f (d) under the log-normal model. Similar to 3.3.1,
the distance estimate can be calculated based on the inverse
of f (d); that is,
if M = P = Q = 0;
0,
dˆc = f −1 (ρ̂S), if f (dth ) ≤ ρ̂ ≤ f (0);
(10)
dth ,
if ρ̂S < f (dth )
where ρ̂ = 2M/(2M + P + Q), and dth denotes the longest
distance between two neighboring nodes.
In what follows, the RSS-based and connectivity-based distance estimation methods shall be analyzed by formulating
their statistical distributions, which paves the way for clarifying the proposed distance estimation method.
4.1.1 The RSS-based Case
In light of the RSS-based distance estimation method presented in Subsection 3.2, it follows from Equation (1) that d
can be formulated as follows
d = 10
PR (d0 )(dBm)−PR (d)(dBm)
10α
Z
× 10 10α .
(11)
By replacing the RSS-based distance estimate dˆR in (2), we
can have
Z
d = dˆR × 10 10α .
(12)
Then, define ǫR to be the multiplicative error of the RSSbased distance estimate, namely
dˆR = d × ǫR ,
(13)
and
Z
ǫR = 10− 10α .
(14)
Since ǫR is dependent on the normal variable Z, their
probability density functions, denoted by pǫR (·) and pZ (·)
respectively, satisfy the following equation
pǫR (x) = pZ (zǫR (x)) × |zǫ′ R (x)|,
(15)
Estimating Distances via Received Signal Strength and Connectivity in Wireless Sensor Networks
where
zǫR (x) = −10α × log10 x.
(16)
Then, we can obtain the probability density function of ǫR
as follows
1
log210 x
1
exp −
,
(17)
pǫR (x) = √
2
2σR
x ln 10
2πσR
where
2
σR
=
2
σdB
(10α)2 .
Since the relationship of ǫR and dˆR is expressed as (13)
and the above method, the probability density function of the
RSS-based distance estimation, denoted by pdˆR (·), satisfies
the following equation
!
log210 xd
1
1
exp −
.
(18)
pdˆR (x) = √
2
2σ
x
ln
10
2πσR
R
4.1.2 The Connectivity-based Case
In Subsection 3.3, we have introduced the connectivity-based
distance estimation method and a piecewise linear function
to approximate the function f (d). Similarly, we can approximate f (d) by
f (d) ≈ kd + b,
(19)
where k and b are constant. Thus according to (10), the
connectivity-based distance estimate dˆc satisfies
ρ̂S = k dˆc + b,
(20)
where ρ̂ = 2M/(2M + P + Q) is a random variable.
Then, define ǫc to be the error of the connectivity-based
distance estimate, namely
ǫc = dˆc − d,
(21)
and
ǫc =
1
(ρ̂S − f (d)).
k
(22)
Then, we are interested in the distribution of the error
base on the formula of ǫc . Because of the variable M , P and
Q are mutually independent Poisson random variables with
means λf (d), λ(S − f (d)), λ(S − f (d)), respectively, and
the additivity property of the independent Poisson random
variables. The distribution of 2M +P +Q can be formulated
as
2M + P + Q ∼ P (2λS).
(23)
In [24], it has proofed that the Poisson random variable
with the mean λ larger than five can approximately equal to a
normal distribution with the mean and variance are equal to
λ. The Poisson random variables 2M and 2M + P + Q with
means 2λf (d) and 2λS satisfy above condition, thence the
5
Poisson random variables can approximately be presented
as follows
2M ∼ N (2λf (d), 2λf (d)),
(24)
2M + P + Q ∼ N (2λS, 2λS).
(25)
Next, in order to analyse the distribution of the ρ̂ =
2M/(2M + P + Q), we consider the ratio of two independent normal random variables. In [25], it has proofed that the
two independent normal variables X and Y with means and
variances (µx , σx2 ) and (µy , σy2 ), respectively. The random
variable Z = X/Y could be approximated to the normal
distribution with the mean and variance
µx
,
µz =
µy
2 2 2 !
(26)
σx
σy
µx
2
.
+
σz =
µy
µx
µy
By (24), (25) and (26), we can have the statistical distribution of ρ̂ satisfies that
f (d) f 2 (d)
1
1
ρ̂ ∼ N
.
(27)
,
+
S
S2
2λf (d) 2λS
According to the formal in (22), we can obtain the error distribution of the connectivity-based distance estimation when S and f (d) are constant
ǫc ∼ N (0, σc2 ),
2
(28)
1
1
where σc2 = f k(d)
2
2λf (d) + 2λS .
Therefore, the probability density function of the error
ǫc , denoted by pǫc (·), satisfies the following equation
x2
1
exp − 2 ,
(29)
pǫc (x) = √
2σc
2πσc
and then, the probability density function, denoted by pdˆc (·),
satisfies the following equation
(x − d)2
1
exp −
pdˆc (x) = √
.
(30)
2σc2
2πσc
4.2 The Proposed Distance Estimation Method
Given the error distributions of the RSS-based and connectivitybased distance estimates, the MLE can be applied to fuse
the distance estimates by the above two methods. Since the
distance estimates dˆR and dˆc rely on different sources of information, we assume that dˆR and dˆc are independent from
each other and express the likelihood function as follows
L(d) = pdˆR (x1 ) × pdˆc (x2 )
log2 x1
(x2 − d)2
1
exp − 102 d −
=
2πσR σc
2σR
2σc2
!
1
.
x1 ln 10
6
Qing Miao, Baoqi Huang∗ and Bing Jia
(31)
Then, the natural logarithm of the likelihood function is
ln L = − ln(2πσR σc x1 ln 10) −
log210 xd1
2
2σR
2
−
(x2 − d)
.
2σc2
(32)
In order to obtain the maximum value of ln L, the NewtonRaphson method is adopted to derive the root of the first
derivative of ln L, denoted by
ˆ =
F (d)
x1
d̂
ˆ 2 − d)
ˆ
d(x
+
.
2
2
σR ln 10
σc
log10
(33)
ˆ be the derivative function of F (d),
ˆ and
To do so, let F ′ (d)
the specific steps of successively finding better approximaˆ include
tions to the root of the function F (d)
1. Select an initial guess d0 , where dˆ0 = (x1 + x2 )/2;
2. Calculate the values of the function F (dk ) and derivative
F ′ (dk );
3. Update dˆk+1 by using the iterative equation dˆk+1 =
ˆ
dˆk − FF′((ddˆk )) ;
k
4. Repeat Step 2 and 3 until |dˆk+1 − dˆk | < ξ, where ξ is a
sufficiently small positive number.
4.3 CRLB
The CRLB expresses a lower bound on the variance of any
unbiased estimator [26, 27] and is equal to the inverse of
the corresponding Fisher Information Matrix (FIM). In this
subsection, the CRLB regarding the proposed distance estimation problem, namely estimating the distance d from M ,
P , Q and the noisy RSS measurement, is formulated under
the log-normal model. Specifically, with the unknown parameters d and λ, the associated FIM, denoted FIM(d, λ), is
formulated as
!
1
+ S−f2 (d) ) + dκ2 −f ′ (d)
λf ′ (d)2 ( f (d)
,
FIM(d, λ) =
2S−f (d)
−f ′ (d)
λ
(34)
2
. The detailed derivation of (34) can
where κ = σdB10α
ln 10
be found in Appendix A.
As was proved in [10], f (d) is a first order differentiable
function, such that the CRLB for d by using any unbiased
estimator, denoted CRLB(d), can be formulated as follows
−1
κ
2λS 2 (f ′ (d))2
+
.
CRLB(d) =
f (d)(2S − f (d))(S − f (d)) d2
(35)
With the CRLB, the comparison will be made in the experimental analyses section to verify the effectiveness of the
proposed method.
5 EXPERIMENTAL ANALYSES
In this section, we aim to investigate the accuracy of the
proposed distance estimation method, and further analyse
the influences of different factors through both simulative
and practical measurements.
5.1 Simulative Analyses
In the numerical simulations, the root-mean-square error (RMSE),
which equals to the square root of the squared biases plus
variances of the errors in distance estimates, is evaluated
in Matlab to measure the accuracy of the proposed method.
Moreover, the CRLB and the RMSE produced by the RSSbased method and the connectivity-based method are also
calculated in the simulations.
The simulative parameters in relation to the WSNs and
wireless channels are described below.
1. The mean value of RSS measurements at the reference
distance, i.e. PR (d0 ), is −37.47dBm;
2. The minimum acceptable RSS value is −100dBm;
3. Each WSN is deployed in a square region, the side length
of which varies with different configurations of α and
σdB ;
4. The sensors are deployed under the random and uniform
distribution of mean λ;
5. The RMSE in each case is evaluated after simulating
10000 distance estimates;
6. µ is the expected number of immediate neighbors of a
sensor, namely µ = E(M + P ) = E(M + Q), and will
take different values given various configurations of σdB
and α.
For better presentation, the connectivity index µ will be used
in the following discussions instead of the sensor density λ.
To analyze the error characteristics of the proposed method,
the influences of different factors, including the expected
number of immediate neighbors of a sensor, the variance
of shadowing effect and the path loss exponent, are investigated in what follows.
Firstly, the effect of the expected number of immediate
neighbors of a sensor (i.e. µ) on the RMSE is considered.
Given σdB = 4 and α = 4, Fig. 2 depicts the RMSE with µ
varying from 10 to 40. As can be seen, we can observe that
– given two nearby sensors, the performance of the proposed method approaches to that of the RSS-based distance estimate, and is superior to that of the connectivitybased method; on the contrary, given two sensors far
away from each other, the performance of the proposed
method is close to that of the connectivity-based method
and is much better than the RSS-based method; that is to
say, the proposed method always outperforms the other
two methods;
Estimating Distances via Received Signal Strength and Connectivity in Wireless Sensor Networks
Secondly, the influence of the noise level on the RMSE
is investigated. As shown in Fig. 3, the RMSE and CRLB
are plotted given µ = 20, α = 4 and σdB is number varying
from 4 to 8, and it can be concluded that
14
12
Finally, the effect of the path loss exponent is studied.
As illustrated in Fig. 4, the RMSE and CRLB are plotted
given µ = 20, σdB = 4 and α varying from 3 to 6. It can be
seen that
3
20 17
2
19
5
7
8
22
4
4
2
36
37
18
16
35
39
21
6
23
6
0
– with σdB increasing, the performance of the RSS-based
methods deteriorates, which is on account of the increasing noises in RSS measurements, whereas the proposed
method and the connectivity-based method incur slight
changes, which is also consistent with the CRLB;
– the overall performance of the proposed method is also
better than the other two methods.
1
10
Y Coordinate (m)
– the RMSE of the proposed method is slightly higher than
the CRLB when two sensors are not far way from each
other, and can even be smaller than the CRLB due to the
fact that the boundary information (i.e. the upper bound
on the distance estimate) is introduced.
7
34
38
40
33
15
27
8
10
28
9
11
−2
−5
25
26
12
0
41
24
14
43
29
13
30
5
X Coordinate (m)
31
32
42
44
10
Fig. 5 The layout of the sensors in [28]
To sum up, the proposed method is able to achieve more
accurate distance estimates than both the RSS-based method
and the connectivity-based method under various simulative
and practical environments, which confirms the effectiveness of the proposed method.
6 CONCLUSIONS
– with α increasing, both the RMSE of these three methods and the CRLB decrease, which is because the communication range decreases;
– the proposed method always outperforms the other two
methods and approaches to the CRLB.
5.2 Implementing the Method in Practice
To better demonstrate its superiority, the proposed method
is implemented using practical RSS measurements and deployment information provided in [28]. Specifically, a WSN
consisting of 44 sensors was deployed in a real environment
as shown in Fig. 5, and the RSS measurements between any
two sensors were reported. On the basis of their RSS measurements, the proposed method can be run to produce practical distance estimates.
According to [28], define α = 2.3, σdB = 3.92, PR (d0 ) =
−37.47 dBm, and the minimum value of the RSS measurements is −55 dBm. To avoid boundary effects as much as
possible, we consider the four sensors near the center of the
deployment region, i.e. sensors 15, 23, 24 and 25. The distance estimates among these four sensors are calculated by
RSS-based, connectivity-based and proposed methods, respectively, and are listed in Tab. 1. As can be seen, the errors
of the proposed method are always less than the corresponding errors obtained by the other two methods.
In this paper, we fused the RSS measurements and local connectivity between two neighboring nodes and implemented
the low-cost and accurate distance estimation method. The
advantages of the proposed method lie in the following aspects. Firstly, the practical log-normal model was applied
to deduce the error characteristics of the RSS-based method
and the connectivity-based method, which enables us to fuse
two sources of information based the MLE. Secondly, both
simulations and experiments were conducted, and it was shown
that the proposed method outperforms its counterparts and
approaches to the CRLB in most cases.
Regarding future works, we would like to apply the proposed method with the existing low-cost localization algorithms (e.g. DV-Hop) so as to improve the localization performance of WSNs without using extra devices.
References
1. Mao, G., Fidan, B., & Anderson, B. D. (2007). Wireless sensor network localization techniques. Computer networks, 51(10), 25292553.
2. Li, L., Halpern, J. Y., Bahl, P., Wang, Y. M., & Wattenhofer,
R. (2005). A cone-based distributed topology-control algorithm
for wireless multi-hop networks. IEEE/ACM Transactions on Networking (TON), 13(1), 147-159.
3. Siripongwutikorn, P., & Thipakorn, B. (2008). Mobility-aware
topology control in mobile ad hoc networks. Computer Communications, 31(14), 3521-3532.
4. Sitanayah, L., Datta, A., & Cardell-Oliver, R. (2010). Heuristic algorithm for finding boundary cycles in location-free low density
wireless sensor networks. Computer Networks, 54(10), 1630-1645.
8
Qing Miao, Baoqi Huang∗ and Bing Jia
20
20
Connectivity
RSS
Proposed
CRLB
18
16
16
14
RMSE(m)
RMSE(m)
14
12
10
8
12
10
8
6
6
4
4
2
2
0
Connectivity
RSS
Proposed
CRLB
18
0
20
40
60
0
80
0
20
40
d(m)
(a) RMSE (µ = 10)
20
Connectivity
RSS
Proposed
CRLB
18
16
Connectivity
RSS
Proposed
CRLB
18
16
14
RMSE(m)
14
RMSE(m)
80
(b) RMSE (µ = 20)
20
12
10
8
12
10
8
6
6
4
4
2
2
0
60
d(m)
0
20
40
60
0
80
0
20
d(m)
40
60
80
d(m)
(c) RMSE (µ = 30)
(d) RMSE (µ = 40)
Fig. 2 The RMSE and CRLB given α = 4, σdB = 4 and µ = 10, 20, 30, 40.
Table 1 The errors by different distance estimation methods
Sensor pair
The proposed method
The RSS-based method
The connectivity-based method
True distance
(15, 23)
0.603
0.974
0.615
2.286
(15, 24)
0.520
1.168
0.530
2.656
5. Huang, B., Wu, W., & Zhang, T. (2013, October). An improved
connectivity-based boundary detection algorithm in wireless sensor
networks. In Local Computer Networks (LCN), 2013 IEEE 38th
Conference on (pp. 332-335). IEEE.
6. Chitte, S. D., & Dasgupta, S. (2008, October). Distance estimation
from received signal strength under log-normal shadowing: Bias
and variance. In Signal Processing, 2008. ICSP 2008. 9th International Conference on (pp. 256-259). IEEE.
7. Benkic, K., Malajner, M., Planinsic, P., & Cucej, Z. (2008, June).
Using RSSI value for distance estimation in wireless sensor networks based on ZigBee. In Systems, signals and image processing, 2008. IWSSIP 2008. 15th international conference on (pp. 303306). IEEE.
8. Della Rosa, F., Pelosi, M., & Nurmi, J. (2012). Human-induced effects on rss ranging measurements for cooperative positioning. International Journal of Navigation and Observation, 2012.
9. Heurtefeux, K., & Valois, F. (2012, March). Is RSSI a good choice
for localization in wireless sensor network?. In Advanced Information Networking and Applications (AINA), 2012 IEEE 26th International Conference on (pp. 732-739). IEEE.
10. Huang, B., Yu, C., Anderson, B., & Mao, G. (2014). Estimating
distances via connectivity in wireless sensor networks. Wireless
Communications and Mobile Computing, 14(5), 541-556.
(15, 25)
0.202
0.442
0.778
1.649
(23, 24)
0.776
1.419
0.828
2.781
(23, 25)
0.963
1.654
1.257
2.891
(24, 25)
0.313
0.745
0.905
1.371
11. Botta, M., & Simek, M. (2013). Adaptive distance estimation
based on RSSI in 802.15. 4 network. Radioengineering, 22(4),
1162-1168.
12. Zanella, A., & Bardella, A. (2014). RSS-based ranging by multichannel RSS averaging. IEEE Wireless Communications Letters,
3(1), 10-13.
13. Nagpal, R., Shrobe, H., & Bachrach, J. (2003). Organizing a
global coordinate system from local information on an ad hoc sensor network. In Information processing in sensor networks (pp. 553553). Springer Berlin/Heidelberg.
14. Shang, Y., Ruml, W., Zhang, Y., & Fromherz, M. P. (2003, June).
Localization from mere connectivity. In Proceedings of the 4th
ACM international symposium on Mobile ad hoc networking &
computing (pp. 201-212). ACM.
15. Buschmann, C., Pfisterer, D., & Fischer, S. (2006, September).
Estimating distances using neighborhood intersection. In Emerging
Technologies and Factory Automation, 2006. ETFA’06. IEEE Conference on (pp. 314-321). IEEE.
16. Buschmann, C., Hellbrck, H., Fischer, S., Kr?ller, A., & Fekete,
S. P. (2007, January). Radio propagation-aware distance estimation
based on neighborhood comparison. In EWSN (pp. 325-340).
17. Li, X. (2007). Collaborative localization with received-signal
strength in wireless sensor networks. IEEE Transactions on Vehic-
Estimating Distances via Received Signal Strength and Connectivity in Wireless Sensor Networks
25
9
35
Connectivity
RSS
Proposed
CRLB
20
Connectivity
RSS
Proposed
CRLB
30
RMSE(m)
RMSE(m)
25
15
10
20
15
10
5
5
0
0
20
40
60
80
0
0
20
d(m)
(a) RMSE (σdB = 5)
60
80
(b) RMSE (σdB = 6)
40
45
Connectivity
RSS
Proposed
CRLB
35
30
RMSE(m)
40
d(m)
Connectivity
RSS
Proposed
CRLB
40
35
30
25
25
20
20
15
15
10
10
5
0
5
0
20
40
60
80
0
0
20
40
60
80
d(m)
(c) RMSE (σdB = 7)
(d) RMSE (σdB = 8)
Fig. 3 The RMSE and CRLB given α = 4, µ = 20 and σdB = 5, 6, 7, 8.
ular Technology, 56(6), 3807-3817.
18. Ouyang, R. W., Wong, A. K. S., & Lea, C. T. (2010). Received signal strength-based wireless localization via semidefinite programming: Noncooperative and cooperative schemes. IEEE Transactions
on Vehicular Technology, 59(3), 1307-1318.
19. So, H. C., & Lin, L. (2011). Linear least squares approach for
accurate received signal strength based source localization. IEEE
Transactions on Signal Processing, 59(8), 4035-4040.
20. Patwari, N., Hero, A. O., Perkins, M., Correal, N. S., & O’dea, R.
J. (2003). Relative location estimation in wireless sensor networks.
IEEE Transactions on signal processing, 51(8), 2137-2148.
21. Cheng, Y. Y., & Lin, Y. Y. (2009). A new received signal strength
based location estimation scheme for wireless sensor network.
IEEE Transactions on Consumer Electronics, 55(3).
22. Botta, M., & Simek, M. (2013). Adaptive distance estimation
based on RSSI in 802.15. 4 network. Radioengineering, 22(4),
1162-1168.
23. Franceschetti, M., & Meester, R. (2008). Random networks for
communication: from statistical physics to information systems
(Vol. 24). Cambridge University Press.
24. Griffin, T. F. (1992). Distribution of the ratio of two poisson random variables (Doctoral dissertation).
25. Dłaz-Francs, E., & Rubio, F. J. (2013). On the existence of a normal approximation to the distribution of the ratio of two independent normal random variables. Statistical Papers, 1-15.
26. Salman, N., Maheshwari, H. K., Kemp, A. H., & Ghogho, M.
(2011, June). Effects of anchor placement on mean-CRB for localization. In Ad hoc networking workshop (Med-Hoc-Net), 2011
The 10th IFIP annual mediterranean (pp. 115-118). IEEE.
27. Ling, Y., Alexander, S., & Lau, R. (2012, March). On quantification of anchor placement. In INFOCOM, 2012 Proceedings IEEE
(pp. 2192-2200). IEEE.
28. Patwari, N. Wireless sensor network localization measurement
repository, 2006.
A The Derivation of the CRLB
The probability density functions of M , P , Q and the RSS
can be formulated as follows
(λf (d))x1
exp(−λf (d)),
x1 !
(λ(S − f (d)))x2
pP (x2 ) =
exp(−λ(S − f (d))),
x2 !
(λ(S − f (d)))x3
exp(−λ(S − f (d))),
pQ (x3 ) =
x3 !
1
(x4 − PR (d0 ) + 10α log10 d)2
pPR (x4 ) = √
exp −
,
2
2σdB
2πσdB
pM (x1 ) =
where PR is a random variable representing RSS in dBm,
and is normally distributed with the mean PR (d0 )−10α log10 d
2
and variance σdB
.
10
Qing Miao, Baoqi Huang∗ and Bing Jia
80
20
Connectivity
RSS
Proposed
CRLB
70
60
Connectivity
RSS
Proposed
CRLB
18
16
RMSE(m)
RMSE(m)
14
50
40
30
12
10
8
6
20
4
10
0
2
0
50
100
150
200
0
250
0
20
40
d(m)
(a) RMSE (α = 3)
4
Connectivity
RSS
Proposed
CRLB
7
3
5
4
3
2.5
2
1.5
2
1
1
0.5
0
10
Connectivity
RSS
Proposed
CRLB
3.5
RMSE(m)
6
RMSE(m)
80
(b) RMSE (α = 4)
8
0
60
d(m)
20
30
40
d(m)
(c) RMSE (α = 5)
0
0
5
10
15
20
25
d(m)
(d) RMSE (α = 6)
Fig. 4 The RMSE and CRLB given σdB = 4, µ = 20 and α = 3, 4, 5, 6.
According to the probability density functions, it is straightforward to formulate the likelihood function as
L(d, λ) = pM (x1 ) × pP (x2 ) × pQ (x3 ) × pPR (x4 )
∂ ln L x1 f ′ (d) (x2 + x3 )f ′ (d)
=
−
+ λf ′ (d)
∂d
f (d)
S − f (d)
(f (d))x1 (S − f (d))x2 +x3
1
= λx1 +x2 +x3 √
x4 − PR (d0 ) + 10α log10 d) 10α
x1 !x2 !x3 !
2πσdB
−
,
2
σdB
d ln 10
(x4 − PR (d0 ) + 10 log10 d)2
. ∂ ln L x1 + x2 + x3
× exp −λ(2S − f (d)) −
2
2σdB
=
− (2S − f (d)),
∂λ
λ
And the natural logarithm of the likelihood function can be
expressed as follows
(f (d))x1 (S − f (d))x2 +x23
√
x1 (f ′ (d))2
κ
(x2 + x3 )(f ′ (d))2
∂ ln L
=
−
− 2
−
x1 !x2 !x3 ! 2πσdB
2
2
2
∂d
f (d)
(S − f (d))
d
(x4 − PR (d0 ) + 10α log10 d)2
.
− λ(2S − f (d)) −
(x
−
10α
P
(d
)
+
10α
log
d)
4
R 0
10
2
2σdB
+
,
2
σdB
d2 ln 10
ln L(d, λ) = (x1 + x2 + x3 ) ln λ + ln
Therefore, the FIM is defined as
2
2
ln(L)
E ∂ ∂dln2 L
E ∂ ∂λ∂d
.
2
FIM(d, λ) = − ∂ 2 ln(L)
ln(L)
E ∂λ∂d
E ∂ ∂λ
2
x1 + x2 + x3
∂ 2 ln L
=−
,
∂λ2
λ2
∂ 2 ln L
=f ′ (d).
∂λ∂d
Since f (d) approximates a linear function, the partial
derivative can be formulated as
Because M , P , Q and PR are independent random variables with means λf (d), λ(S − f (d)), λ(S − f (d)) and
Estimating Distances via Received Signal Strength and Connectivity in Wireless Sensor Networks
PR (d0 ) − 10α log10 d , respectively, we can have
2
κ
1
∂ ln L
2
′
2
E
− 2,
=
−λf
(d)
+
2
∂d
f (d) S − f (d)
d
2
∂ ln L
2S − f (d)
E
=−
,
∂λ2
λ
2
∂ ln L
E
= f ′ (d),
∂λ∂d
2
where κ = σdB10α
.
ln 10
Then, the FIM can be expressed as follows
!
1
+ S−f2 (d) + dκ2 −f ′ (d)
λf ′ (d)2 f (d)
,
FIM(d, λ) =
2S−f (d)
−f ′ (d)
λ
and thus, we can have
−1
κ
2λS 2 (f ′ (d))2
+
.
CRLB(d) =
f (d)(2S − f (d))(S − f (d)) d2
11
| 3cs.SY
|
Analysis of the Pólya-Gamma block Gibbs sampler for
Bayesian logistic linear mixed models
arXiv:1708.00100v2 [math.ST] 17 Nov 2017
Xin Wang∗ and Vivekananda Roy†
Department of Statistics, Iowa State University, Ames, IA
November, 2017
Abstract
In this article, we construct a two-block Gibbs sampler using Polson et al.’s
(2013) data augmentation technique with Pólya-Gamma latent variables for Bayesian
logistic linear mixed models under proper priors. Furthermore, we prove the uniform
ergodicity of this Gibbs sampler, which guarantees the existence of the central limit
theorems for MCMC based estimators.
key words: Data augmentation, Markov Chain, Logit link, Pólya-Gamma distribution,
Uniform ergodicity
1
Introduction
Consider the logistic linear mixed model set-up (Charles E. McCulloch and Neuhaus,
2008; McCulloch, 2003). Let (Y1 , Y2 , . . . , YN ) denote the vector of Binomial(ni , pi ) random variables, xi and zi be the p × 1 and q × 1 known covariates and random effect
design vectors respectively associated with the ith observation for i = 1, . . . , N . Let
β ∈ Rp be the unknown vector of regression coefficients and u ∈ Rq be the random
effects vector. Assume that pi = F (xTi β + ziT u), where F is the standard logistic distribution function, that is F (t) ≡ et /(1 + et ) for t ∈ R. Suppose we have r random effects
with u = (uT1 , . . . , uTr )T , where uj is a qj × 1 vector with qj > 0, q1 + · · · + qr = q,
ind
and uj ∼ N (0, Iqj 1/τj ), where τj ∈ R+ = (0, ∞) is the precision parameter associated with uj for j = 1, . . . , r. The joint distribution of u is N 0, D(τ )−1 , where
D(τ ) = ⊕rj=1 τj Iqj , τ = (τ1 , . . . , τr ) and ⊕ denotes the direct sum. The data model for
the logistic linear mixed model is
∗
†
Yi |β, u
ind
Binomial(ni , pi ) for i = 1, . . . , N with
pi
=
uj |τj
ind
F (xTi β + ziT u) for i = 1, . . . , N,
1
N 0, Iqj , j = 1, . . . , r.
τj
∼
∼
Email: [email protected]
Email: [email protected]
1
(1)
Let y = (y1 , y2 , . . . , yN )T be the observed Binomial response variables. The likelihood
function for (β, τ ) is
L (β, τ |y) =
Z
N
Y
Rq i=1
! h
iyi
exp xTi β + ziT u
ni
−1
u;
0,
D(τ
)
φ
du,
q
yi 1 + exp xTi β + ziT u ni
(2)
where φq (s; a, B) is the probability density function of the q-dimensional normal distribution with mean vector a and covariance matrix B evaluated at s. In Bayesian
framework, let π (β) and π (τ ) be the prior densities for β and τ respectively. Assume
that β and τ are apriori independent. The joint posterior density of (β, τ ) is
π (β, τ |y) =
1
L (β, τ |y) π (β) π (τ ) ,
c0 (y)
(3)
where c0 (y) is the marginal density of y with c0 (y) = Rr Rp L (β, τ |y) π (β) π (τ ) dβdτ .
+
The posterior density (3) is intractable for any choice of the prior distributions of β
and τ . Generally, Markov chain Monte Carlo (MCMC) algorithms are used for exploring
these posterior densities. Even in the absence of random effects, for generalized linear
models, MCMC algorithms are needed to summarize the associated posterior densities.
For probit regression models, Albert and Chib (1993) proposed a widely used data augmentation (DA) algorithm to sample from the corresponding posterior distributions.
Roy and Hobert (2007) and Chakraborty and Khare (2017) proved the geometric ergodicity of this DA algorithm for Bayesian probit regression model under improper and
proper priors respectively. Wang and Roy (2017) recently extended the convergence rate
analysis of the block Gibbs samplers based on this DA technique for Bayesian probit
linear mixed models under both proper and improper priors.
For logistic regression models, there have been several attempts for producing a DA
algorithm similar to Albert and Chib’s (1993) algorithm for the probit regression model
(see e.g. Holmes and Held (2006) and Frühwirth-Schnatter and Frühwirth (2010)). Unfortunately, these algorithms are far more complex than Albert and Chib’s (1993) algorithm. Only recently, Polson et al. (2013) produced such a DA algorithm for logistic
regression models using Pólya-Gamma latent variables. Choi and Hobert (2013) proved
uniform ergodicity of the Pólya-Gamma DA Markov chain under normal priors on the
regression parameters. Choi and Román (2017) showed that the Markov operator based
on Polson et al.’s (2013) DA algorithm for one-way logistic ANOVA model is traceclass, which implies that the associated Markov Chain is geometrically ergodic. Both
Choi and Hobert (2013) and Choi and Román (2017) considered the special case when
the data are binary, that is ni = 1 for all i. However, there is no result in the literature about convergence analysis of any Gibbs samplers for Bayesian logistic linear mixed
models. In this article, we construct a two-block Gibbs sampler for Bayesian logistic
linear mixed models with normal priors on regression parameters and truncated Gamma
priors on precision parameters. We further establish uniform ergodicity of this Gibbs
sampler.
The article is organized as follows. In section 2, we construct the two-block Gibbs
sampler for the Bayesian logistic linear mixed model under proper priors. In section 3,
R
2
R
we prove the uniform ergodicity of the underlying Markov chain. Finally, we have some
discussions in section 4.
2
Two-block Gibbs sampler
In Polson et al. (2013), a logistic linear mixed model example is introduced. In their
example, normal distribution is used as the prior for regression coefficients and Gamma
distribution is used as the prior for precision parameters. We assume the following
priors: β ∼ Np (Q−1 µ0 , Q−1 ) for some p × p positive definite matrix Q and µ0 ∈ Rp ,
ind
τj ∼ truncated Gamma(aj , bj , τ0 ), j = 1, . . . , r, where bj > 0 for j = 1, . . . , r. The
density function of truncated Gamma(aj , bj , τ0 ) is
a −1
f (τj |aj , bj , τ0 ) = [c (τ0 , aj , bj )]−1 τj j
exp (−bj τj ) I(τj ≥ τ0 ),
(4)
where c (τ0 , aj , bj ) = τ∞
τ aj −1 exp (−bj τ ) dτ and τ0 > 0 is a known constant.
0
By Theorem 1 in Polson et al. (2013),
R
h
exp xTi β + ziT u
iyi
1 + exp xTi β + ziT u
i
h
ni = 2−ni exp κi xTi β + ziT u
×
Z
∞
0
exp −ωi
xTi β
+
ziT u
2
/2 p (ωi ) dωi ,
(5)
where p (ωi ) is the probability density function of the random variable ωi ∼ PG (ni , 0)
and κi = yi − ni /2 for i = 1, . . . , N . Here, PG(ni , 0) denotes the Pólya-Gamma distribution with parameters ni and 0 with density
f (x|ni , 0) =
∞
Γ(n + ni ) (2n + ni ) − (2n+ni )2
2ni −1 X
8x
√
e
(−1)n
, x > 0.
Γ(ni ) n=0
Γ(n + 1)
2πx3
Let ω = (ω1 , . . . , ωN ) and the joint (posterior) density of β, u, ω and τ be
π (β, u, ω, τ |y) ∝
N
Y
i=1
exp κi xTi β + ziT u − ωi xTi β + ziT u
× φq u; 0, D(τ )−1 φp β; Q−1 µ0 , Q−1
r
Y
j=1
a −1 −bj τj
τj j
e
2
/2 p (ωi )
I(τj ≥ τ0 ).
(6)
From (2), (3) and (5), it follows that Rq RN π(β, u, ω, τ |y)dωdu = π(β, τ |y), which
+
is our target posterior density. Using draws from all full conditional distribution distributions of (6), we can run a Gibbs sampler with stationary density (6). It is known
that by combining and simultaneously drawing multiple parameters, the convergence of
the Gibbs sampler can be improved (Liu et al., 1994), although the “blocking” to be
computationally efficient, the corresponding joint conditional distributions need to be
R
R
tractable. Here we construct a two-block Gibbs sample for (6). Let η = β T , uT
3
T
,
κ = (κ1 , . . . , κn ) and M = (X, Z) with ith row mTi and Ω be the n × n diagonal matrix
with ith diagonal element ωi . Standard calculations show that the conditional density
of η is
1 T
1 T T
T
T
T
π (η|ω, τ , y) ∝ exp − η M ΩM η + η M κ exp − η A (τ ) η + η l ,
2
2
where l = µT0 , 01×q
where Σ =
T
, and A (τ ) = Q ⊕ D(τ ). That is,
η|ω, τ , y ∼ N Σ−1 µ, Σ−1
X T ΩX + Q
X T ΩZ
Z T ΩX
Z T ΩZ + D(τ )
!
(7)
= M T ΩM + A (τ ) ,
µ = M T κ + l.
Similarly, the conditional density of (ω, τ ) is
π (ω, τ |β, u, y) ∝
×
N
Y
i=1
r
Y
exp −ωi xTi β + ziT u
a +qj /2−1
τj j
j=1
h
2
/2 p (ωi )
i
exp − uTj uj /2 + bj τj I(τj ≥ τ0 ).
So, given β, u, y, we have that ω and τ are conditionally independent with
ind
ωi |η, y ∼ P G ni , |mTi η| , i = 1, . . . , N,
uTj uj
qj
τj |η, y ∼ truncated Gamma aj + , bj +
, τ0 , j = 1, . . . , r.
2
2
ind
!
Remark 1. As in Wang and Roy (2017), we assume that the prior distribution for τj is
a truncated Gamma distribution. However, while implementing the block Gibbs sampler
in practice, a number slightly larger than the machine precision zero can be treated as
τ0 , practically avoiding the need to use any rejection sampling algorithms to draw from
the truncated conditional distribution of τ .
Thus, one single iteration of the block Gibbs sampler {η (m) , ω (m) , τ (m) }∞
m=0 has the
following two steps:
Algorithm: The (m + 1)st iteration for the two-block Gibbs sampler
1:
(m+1)
Draw τj
(m+1) ind
∼ P G ni , |mTi η (m) | for i = 1, . . . , N .
j = 1, . . . , r, and independently draw ωi
2:
from truncated Gamma aj + qj /2, bj + uTj uj /2, τ0 with u = u(m) for
Draw η (m+1) from (7), η (m+1) ∼ Np+q Σ(m)−1 M T κ + l , Σ(m)−1 , where
(m+1)
Σ(m) = M T Ω(m+1) M +A(τ (m+1) ) and the diagonal elements of Ω(m+1) are ωi
i = 1, . . . , N .
,
Polson et al. (2013) developed an efficient method for sampling from PG distribution,
which is the only nonstandard distribution involved in the above Gibbs sampler.
4
3
Uniform ergodicity of the two-block Gibbs sampler
In this section, we prove the uniform ergodicity of the two-block Gibbs sampler
{η (m) , ω (m) , τ (m) }∞
m=0 , which has the same rate of convergence as the η-marginal Markov
(Roberts
chain {η (m) }∞
and Rosenthal, 2001). Below we analyze the Ψ ≡ {η (m) }∞
m=0
m=0
chain.
Let η ′ be the current state and η be the next state, then the Markov transition
density (Mtd) of Ψ is
′
k(η|η ) =
Z
Rr+
Z
RN
+
π(η|ω, τ , y)π(ω, τ |η ′ , y)dωdτ ,
(8)
where π(·|·, y)’s are the conditional densities from section 2. Routine calculations show
that k(η|η ′ ) is reversible
and thus invariant with respect to the marginal density of η
R R
denoted as π(η|y) ≡ Rr RN π(η, ω, τ |y)dωdτ , where π(η, ω, τ |y) is defined in (6).
+
+
Since k(η|η ′ ) is strictly positive, the Markov chain Ψ is Harris ergodic (Hobert, 2011).
Let B denote the Borel σ-algebra of Rp+q and K(·, ·) be the Markov transition
function corresponding to the Mtd k(·, ·) in (8), that is, for any set A ∈ B, η ′ ∈ Rp+q
and any j = 0, 1, . . . ,
′
K(η , A) = Pr(η
(j+1)
∈ A|η
(j)
′
=η)=
Z
k(η|η ′ )dη.
(9)
A
Then the m-step Markov transition function is K m (η ′ , A) = Pr(η (m+j) ∈ A|η (j) = η ′ ).
Let Π(·|y) be the probability measure with density π(η|y). The Markov chain Ψ is
geometrically ergodic if there exists a constant 0 < t < 1 and a function G : Rp+q 7→ R+
such that for any η ∈ Rp+q ,
||K m (η, ·) − Π(·|y)|| := sup |K m (η, A) − Π(A|y)| ≤ G(η)tm .
(10)
A∈B
If G(η) is bounded above, then the corresponding Markov chain is uniformly ergodic.
The following theorem establishes uniform ergodicity of the Markov chain Ψ by constructing a minorization condition.
Theorem 1. Assume that aj + qj /2 ≥ 1 and bj > 0 for all j, then the Markov chain Ψ
is uniformly ergodic.
Proof. We show that there exists a δ > 0 and a density function h : Rp+q → [0, ∞) such
that, for all η ′ , η ∈ Rp+q ,
k(η|η ′ ) ≥ δh(η).
(11)
By Roberts and Rosenthal (2004)[Theorem 8], (11) implies that the Markov chain
Ψ is uniformly ergodic. Furthermore, under (11), (10) holds with G = 1 and t = 1 − δ.
For τj ≥ τ0 , j = 1, . . . , q, Σ ≥ M T ΩM + A (τ0 ), that is Σ − (M T ΩM + A (τ0 ))
is positive semidefinite. So |Σ| ≥ M T ΩM + A (τ0 ) = |A (τ0 )| M̃ T ΩM̃ + I ≥
5
−1
|A (τ0 )| , where M̃ = M A (τ0 )−1/2 . And µT Σ−1 µ ≤ µT M T ΩM + A (τ0 )
µT A (τ0 )
−1
µ. Then,
π(η|ω, τ , y) = (2π)
− p+q
2
≥ (2π)−
p+q
2
= (2π)−
p+q
2
µ ≤
T
1
|Σ| exp − η − Σ−1 µ Σ η − Σ−1 µ
2
1
1 T
−1
1/2
T
T
|A (τ0 )| exp − η Ση − 2η µ − µ A (τ0 ) µ
2
2
1
1
1 T
1/2
−1
T
T
T
|A (τ0 )| exp − β Qβ − u D(τ )u + η µ − µ A (τ0 ) µ
2
2
2
1
2
N
2
1X
.
ωi mTi η
× exp −
2 i=1
#
"
Therefore,
π(η|ω, τ , y)π(ω, τ |η ′ , y)
≥ (2π)
×
×
r
Y
i=1
1/2
|A (τ0 )|
1
1
1
exp − β T Qβ − uT D(τ )u + η T µ − µT A (τ0 )−1 µ
2
2
2
1
a +qj /2−1
τj j
′
c τ0 , aj + qj /2, bj + u′T
j uj /2
j=1
N
Y
− p+q
2
mTi η ′
coshni
exp
−
2
mTi η ′
2
+ mTi η
2
h
2
i
′
exp − bj + u′T
j uj /2 τj I(τj ≥ τ0 )
ωi p (ωi ) .
According to Polson et al. (2013) and Choi and Hobert (2013),
Z
R+
mTi η ′
exp −
≥ cosh
mTi η ′
2
+
2
mTi η
−ni
2
+ mTi η
2
2
implying
ni
cosh
−ni
≥2
mTi η ′
2
exp
mTi η
2
Z
R+
2
≥ 2 cosh
exp −
−ni
q
2
2 −ni
mTi η ′ + mTi η
ωi p (ωi ) dωi = cosh
mTi η ′
2
mTi η ′
2
+ mTi η
2
2
T
mi η
−ni
≥2
exp
4
6
2
cosh
mTi η
−ni
2
,
mTi η
−ni
−ni
cosh
ωi p (ωi ) dωi ≥ 2
−ni
+ 1
2
−ni −ni /4
=2
e
ni T 2
exp −
mi η
.
4
So we have,
Z
RN
+
π(η|ω, τ , y)π(ω, τ |η ′ , y)dω
≥ (2π)−
1
1
|A (τ0 )|1/2 exp − β T Qβ + η T µ − µT A (τ0 )−1 µ
2
2
1 T T
exp − η M ΛM η
4
n
× 2−n e− 4
×
r
Y
p+q
2
a +qj /2−1
′
−1 j
[c τ0 , aj + qj /2, bj + u′T
j uj /2 ] τj
j=1
PN
where n =
Thus,
k(η|η ′ ) =
i=1 ni
Z
Rr+
Z
RN
+
h
i
′
T
exp − bj + u′T
j uj /2 + uj uj /2 τj I(τj ≥ τ0 ),
and Λ is the N × N diagonal matrix with ith diagonal element ni .
π(η|ω, τ , y)π(ω, τ |η ′ , y)dωdτ
n
1
1
1
|A (τ0 )|1/2 exp − β T Qβ + η T µ − µT A (τ0 )−1 µ × 2−n e− 4 exp − η T M T ΛM η
2
2
4
Z ∞
r
h
i
Y
1
a +q /2−1
×
τj j j
exp − bj + u′T
u′j /2 + uTj uj /2 τj dτj .
j
′T ′
τ0
j=1 c τ0 , aj + qj /2, bj + uj uj /2
≥ (2π)−
p+q
2
Now consider
1
′
c τ0 , aj + qj /2, bj + u′T
j uj /2
aj +qj /2
′
bj + u′T
j uj /2
Z
∞
τ0
=
a +q /2
′ /2 + uT u /2 j j
bj + u′T
u
j
j
j
j
a +qj /2−1
τj j
h
i
′
T
exp − bj + u′T
j uj /2 + uj uj /2 τj dτj
R∞
xaj +qj /2−1 exp (−x) dx,
T
′
(bj +u′T
j uj /2+uj uj /2)τ0
R∞
·
aj +qj /2−1 exp (−x) dx
(bj +u′T u′ /2)τ0 x
j
j
aj +qj /2−1 exp (−t) dt, f (x) = ∞
aj +qj /2−1 exp (−t) dt
For x ≥ 0, define, f1 (x) = (∞
2
(bj +x)τ0 t
b +x+uT
/2)τ0 t
j uj
j
and g (x) = f1 (x) − exp −τ0 uTj uj /2 f2 (x). Since aj + qj /2 − 1 ≥ 0 by assumption,
′
it can
h be shown ithat g (x) ≤ 0. And g (x) ≥ limx→∞ g (x) = 0. Thus f1 (x) /f2 (x) ≥
T
T
′T ′
′
exp −τ0 uTj uj /2 . Also (bj + u′T
j uj /2)/(bj + uj uj /2 + uj uj /2) ≥ bj /(bj + uj uj /2),
So
R
κ(η|η ′ ) ≥ (2π)−
p+q
2
−n − n
4
×2
e
R
1
1
|A (τ0 )|1/2 exp − β T Qβ + η T µ − µT A (τ0 )−1 µ
2
2
r
Y
bj
1 T T
exp − η M ΛM η ×
4
bj + uTj uj /2
j=1
7
!aj +qj /2
exp −τ0 uTj uj /2 .
Let
c1 (M , y) =
×
1
1
exp − β T Qβ + η T µ − η T M T ΛM η
2
4
Rp+q
Z
r
Y
bj
bj + uTj uj /2
j=1
p+q
2
≤ (2π)
!aj +qj /2
−1/2
|A (τ0 )|
exp −τ0 uTj uj /2 dη
n
1 T
exp
µ A (τ0 ) µ 2n e 4 < ∞.
2
So there exists a density function h (η) and δ > 0 such that,
k(η|η ′ ) =
Z
Rr+
Z
Rn
+
π(η|ω, τ , y)π(ω, τ |η ′ , y)dωdτ ≥ δh (η) ,
where
1
1
1
h (η) =
exp − β T Qβ + η T µ − η T M T ΛM η
c1 (M , y)
2
4
×
r h
Y
iaj +qj /2
bj /(bj + uTj uj /2)
j=1
and δ = (2π)−
p+q
2
exp −τ0 uTj uj /2 ,
n
1
|A (τ0 )|1/2 2−n e− 4 · exp − µT A (τ0 )−1 µ c1 (M , y) .
2
Hence the Markov chain is uniformly ergodic.
Remark 2. Since Theorem 1 does not put any conditions on y, X, Z, N , p and q, it
is applicable in high dimensional situations where p (or q) can be much larger than N .
Remark 3. Following the proof of Theorem 1, the uniform ergodicity result in Choi and Hobert
(2013) can be extended to binomial data.
i
h
Remark 4. Since 1/(x+1) ≥ exp (−x), we have bj /(bj +uTj uj /2) ≥ exp −uTj uj /(2bj ) .
Using this inequality, we have
−n − n
4
δ≥2
e
1/2
|A (τ0 )|
|Σ1 |
−1/2
1
1
exp − µT A (τ0 )−1 µ + µT Σ−1
1 µ ,
2
2
i
h
where Σ1 = 12 M T ΛM + Q ⊕ ⊕rj=1 {τ0 + (aj + qj )/(2bj )}Iqj . This in turn, gives a
computable upper bound to the total variation distance to the stationary in (10).
4
Discussion
We prove uniform ergodicity of the two-block Gibbs sampler for Bayesian logistic
linear mixed models, which guarantees the existence of central limit theorem for MCMC
8
estimators under a finite second moment condition (Jones, 2004). Thus, our result has
important practical implications as it allows for obtaining valid asymptotic standard
errors for the posterior estimates (Flegal and Jones, 2010). Convergence rates analysis
of Gibbs samplers for Bayesian logistic linear mixed models with improper priors is a
potential future project.
References
Albert, J. H. and Chib, S. (1993). Bayesian analysis of binary and polychotomous
response data. Journal of the American statistical Association, 88(422):669–679.
Chakraborty, S. and Khare, K. (2017). Convergence properties of Gibbs samplers for
Bayesian probit regression with proper priors. Electronic Journal of Statistics, 11:177–
210.
Charles E. McCulloch, S. R. S. and Neuhaus, J. M. (2008). Generalized, linear, and
mixed models. Wiley, 2nd ed edition.
Choi, H. M. and Hobert, J. P. (2013). The Pólya-Gamma Gibbs sampler for Bayesian
logistic regression is uniformly ergodic. Electronic Journal of Statistics, 7:2054–2064.
Choi, H. M. and Román, J. C. (2017). Analysis of Pólya-Gamma Gibbs sampler for
Bayesian logistic analysis of variance. Electronic Journal of Statistics, 11(1):326–337.
Flegal, J. M. and Jones, G. L. (2010). Batch means and spectral variance estimators in
Markov chain Monte Carlo. The Annals of Statistics, 38(2):1034–1070.
Frühwirth-Schnatter, S. and Frühwirth, R. (2010). Data augmentation and MCMC for
binary and multinomial logit models. In Statistical Modelling and Regression Structures, pages 111–132. Springer.
Hobert, J. P. (2011). The data augmentation algorithm: Theory and methodology. In
Brooks, S., Gelman, A., Jones, G., and Meng, X.-L., editors, Handbook of Markov
Chain Monte Carlo. Chapman & Hall/CRC Press.
Holmes, C. C. and Held, L. (2006). Bayesian auxiliary variable models for binary and
multinomial regression. Bayesian analysis, 1(1):145–168.
Jones, G. L. (2004). On the markov chain central limit theorem. Probability surveys,
1:299âĂŞ–320.
Liu, J. S., Wong, W. H., and Kong, A. (1994). Covariance structure of the Gibbs
sampler with applications to the comparisons of estimators and augmentation schemes.
Biometrika, pages 27–40.
McCulloch, C. E. (2003). Generalized linear mixed models. NSF-CBMS regional conference series in probability and statistics, 7:i–84.
9
Polson, N. G., Scott, J. G., and Windle, J. (2013). Bayesian inference for logistic models
using Pólya-Gamma latent variables. Journal of the American statistical Association,
108(504):1339–1349.
Roberts, G. O. and Rosenthal, J. S. (2001). Markov chains and de-initializing processes.
Scandinavian Journal of Statistics, 28:489–504.
Roberts, G. O. and Rosenthal, J. S. (2004). General state space markov chains and
MCMC algorithms. Probability Surveys, 1:20–71.
Roy, V. and Hobert, J. P. (2007). Convergence rates and asymptotic standard errors for
Markov chain Monte Carlo algorithms for Bayesian probit regression. Journal of the
Royal Statistical Society: Series B, 69(4):607–623.
Wang, X. and Roy, V. (2017). Convergence analysis of block Gibbs samplers for Bayesian
probit linear mixed models. arXiv preprint arXiv:1706.01846.
10
| 10math.ST
|
Zero-Shot Learning to Manage a Large Number of Place-Specific
Compressive Change Classifiers
arXiv:1709.05397v1 [cs.CV] 15 Sep 2017
Tanaka Kanji
Abstract— With recent progress in large-scale map maintenance and long-term map learning, the task of change detection
on a large-scale map from a visual image captured by a
mobile robot has become a problem of increasing criticality.
Previous approaches for change detection are typically based
on image differencing and require the memorization of a
prohibitively large number of mapped images in the above
context. In contrast, this study follows the recent, efficient
paradigm of change-classifier-learning and specifically employs
a collection of place-specific change classifiers. Our changeclassifier-learning algorithm is based on zero-shot learning
(ZSL) and represents a place-specific change classifier by its
training examples mined from an external knowledge base
(EKB). The proposed algorithm exhibits several advantages.
First, we are required to memorize only training examples
(rather than the classifier itself), which can be further compressed in the form of bag-of-words (BoW). Secondly, we
can incorporate the most recent map into the classifiers by
straightforwardly adding or deleting a few training examples
that correspond to these classifiers. Thirdly, we can share the
BoW vocabulary with other related task scenarios (e.g., BoWbased self-localization), wherein the vocabulary is generally
designed as a rich, continuously growing, and domain-adaptive
knowledge base. In our contribution, the proposed algorithm
is applied and evaluated on a practical long-term cross-season
change detection system that consists of a large number of
place-specific object-level change classifiers.
I. I NTRODUCTION
This paper considers the problem of scalable change detection using a novel, compact representation of an environment
map (Fig. 1). With recent progress in large-scale map maintenance [2] and long-term map learning [3], the task of change
detection on a large-scale map, from a visual image captured
by a mobile robot, has become a problem of increasing
criticality [4]. We aim to enhance the scalability from the
novel perspective of map compactness while maintaining the
effectiveness of the change detection system.
In general, change detection has the goal of detecting
changes between a robot’s view image and a previously
constructed map. A major challenge in change detection is
to effectively address appearance variations of the change
objects. As the variations are inherently place-specific [5] and
attributed to various factors (e.g., objects, viewpoint, background, illumination conditions, occlusions), it is challenging
to obtain a general-purpose change model. Till the present,
the most fundamental scheme reported for addressing this
challenge is to directly compare each view image against
Our work has been supported in part by JSPS KAKENHI Grant-in-Aid
for Young Scientists (B) 23700229, for Scientific Research (C) 26330297,
and for Scientific Research (C) 17K00361.
The authors are with Graduate School of Engineering, University of
Fukui, Japan. [email protected]
Fig. 1. Scalable change detection for long-term map maintenance. (a) Experimental environment. The trajectories of the four datasets “2012/01/22,”
“2012/03/31,” “2012/08/04,” and “2012/11/17” used in our experiments are
visualized in green, purple, blue, and light-blue curves and overlaid on the
bird’s-eye-view imagery obtained from NCLT dataset [1]. (b) Examples of
changes. For each panel, the left and right figures are a query image and
its corresponding reference (i.e., mapped) image, respectively.
the corresponding reference (i.e., mapped) image using handcrafted features [6] or deep-learning techniques [7]. However,
this fundamental scheme requires explicit memorization of
every possible mapped image and exhaustive many-to-many
image comparisons, which severely limits the scalability in
time and space.
Our approach is inspired by the recent success in the
community of change-classifier-learning [8]–[11]. Instead
of memorizing the large collection of mapped images, the
change-classifier-learning approach simply learns an essential change classifier that is often significantly compact
and nonetheless exhibits generalization capability. In [8],
Gueguen et al. presented a change detection method for
overhead imagery wherein they train a support vector machine (SVM) with linear kernel with L1-regularization and
L2-loss as a semi-supervised scene-specific change classifier
from manually labeled positive examples (i.e., changes) and
plenty of available negative examples (i.e., no-changes).
Moreover, it is preferable to use only a single change
classifier; however, it is challenging for a single classifier
to capture the place-specific variations of changes [5] or
to flexibly incorporate the latest local changes in the map
[12]. Therefore, we opt to use multiple place-specific change
classifiers, each of which is responsible for each specific
place region and the place-specific change classifiers can be
flexibly and locally updated, added, or deleted at a marginal
fixed cost. Each classifier receives local feature-descriptors
(e.g., ORB descriptors) as input and evaluates their likelihood
of change. However, it is not straightforward to apply the
typical framework of change-classifier-learning to our application domain of autonomous robotics, wherein we are
not provided examples with labels (i.e., change/no-change).
Instead, the mapper-robot itself should collect examples in
an unsupervised manner.
In this study, we address the above issue from the viewpoint of zero-shot learning (ZSL) [13]. ZSL is a domainadaptation technique in the field of machine learning and
was originally proposed as a strategy to obtain classifiers
for arbitrary, novel categories when no labeled example is
available. In particular, we are inspired by mining-based ZSL
[14], wherein training examples for novel unseen categories
(in the target domain) are collected by mining an external
knowledge base (EKB) such as a search engine. To realize
our ZSL-based change-classification framework, it is necessary to address the following questions: (1) How does one
train classifiers? (2) How does one represent examples? (3)
How does one prepare an EKB?
The key concepts required for answering these questions
are as follows:
1) We train classifiers with the latest examples when
necessary to incorporate the latest changes in the map;
2) We use bag-of-words (BoW) as a compact and discriminative representation of examples;
3) We employ a visual word vocabulary as the EKB,
which are inspired by the use of BoW as a compressed
form of place-specific classifier in our previous study [15].
The above concepts exhibit several advantages: (1) We only
have to memorize training examples rather than the classifier themselves because typical classifier-learning algorithms
such as SVM [16] and nearest neighbor (NN) [17] allow us
to reproduce a classifier, given the same training examples
[15]. (2) We can incorporate the latest changes into the classifiers by simply adding or deleting a few training examples
that correspond to these classifiers, while the training-time
overhead per place can be reasonably low by using efficient
training algorithms such as NN and SVM. (3) We can
represent each training example in a significantly compact
form of a visual word [15], implying that the training
example is approximated by its nearest neighbor exemplar
feature from a visual word vocabulary. (4) We can share
the vocabulary with other related task-scenarios including
BoW-based self-localization, SLAM, lifelong learning, and
open-set recognition, wherein the vocabulary is generally designed as a rich, continuously growing, and domain-adaptive
knowledge base. In experiments, the proposed framework is
evaluated in a cross-season change detection setting using
the publicly available NCLT dataset [1].
A. Relation to Other Works
A majority of the present state-of-the-art change detection
systems are based on image differencing or pairwise image
comparison [5]. In [18], a scene alignment method for image
differencing is proposed based on ground surface reconstruction, texture projection, image rendering, and registration
refinement. In [4], a deep deconvolutional network for pixelwise change detection was trained and used for comparing
query and reference image patches.
Some recent studies use change-classifier-learning to realize more efficient and compact change detection [8]–
[11]. Our algorithm is inspired by these classifier-learning
approaches and advances a step further from the perspective
of unsupervised ZSL. The work in [8] can be considered as
one of the most relevant works to our study. In their work,
change detection in overhead imagery is addressed. A BoW
model with tree-of-shape features is employed to achieve
more effective accuracy-efficiency tradeoff. Based on these
features, linear canonical correlation analysis is employed to
learn a subspace to encode the notion of change between
images. To reduce the cost of label acquisition by human
photo-interpreters, a semi-supervised SVM framework is
introduced. In contrast, we address unsupervised settings,
wherein examples have to be labeled automatically by the
robot itself rather than by human labelers.
Our study focuses on a monocular camera as the sole input
device. This is a significantly challenging setting compared
with other change detection settings, which assume a more
richer information source, including 3D model [19], stereo
or image sequence [6], 3D data [20], multi-spectral overhead
imagery [21], and object model [22].
In the area of field robotics, there is substantial work
on change detection systems for patrolling [23], agriculture
[24], tunnel inspection [25], and damage detection [8]. We
consider that our extension would also contribute to these
applications from the novel perspective of map compactness.
This work is a part of our study on compact map model for
scalable change detection and long-term map maintenance
[15], [26]–[29]. In previous studies, two scenarios have been
considered. One is the “global localization” scenario [28],
wherein the change detection system is required to work
under global viewpoint uncertainty. As this scenario requires
access to the entire large-size maps in memory, individual
images were used in their compressed form of bag-of-words
[26]. In contrast, the current study focuses on the alternative
“pose-tracking” scenario, wherein the system can assume
that the robot’s viewpoint is tracked over time. A similar
setting is addressed in our recent study in [27] with a key
variation being that classifiers are not compressed. As the
pose-tracking scenario requires access to only a marginal
portion of the submaps that correspond to the robot’s surroundings, we are allowed to use mapped images in a lesscompact form of the decompressed classifiers. Nevertheless,
it is necessary to maintain the other classifiers compressed.
Hence, the efficiency of compression/decompression is another critical topic, as also discussed in this paper.
II. A PPROACH
Our goal is to enhance the compactness of map representation for scalable change detection. As mentioned,
this rules out typical image-differencing approaches that
require memorization of a number of high-resolution mapped
images proportional to the map size. Therefore, we decide
to use the change-classifier approach. More specifically, we
employ multiple place-specific change classifiers to capture
the place-specific nature of changes and to enable flexible
and local update of the change model. The key concept is
to represent each classifier by their training examples and
compress each training example to a visual word; this enables
a significantly compact and domain-adaptive change model.
Based on the above consideration, we use a set of placespecific change classifiers and represent each by a BoW. Following literature [30], a BoW is represented as an unordered
collection of visual words, or visual features vector quantized
by a codebook of exemplar features called vocabulary. In our
approach, an i-th classifier is represented by positive (i.e.,
change) and negative (i.e., no-change) sets of visual words,
{+,−}
Si+ and Si− , respectively. Each visual word w ∈ Si
is
represented in the form:
a
r
w = hw , w i.
(1)
wa is a B-bit code called appearance word and is an
identifier for its nearest neighbor local-feature-descriptor in
the vocabulary. We employ ORB feature descriptor [31] as
local-feature-descriptor because it is a rapid binary keypoint descriptor that has been applied to numerous realtime computer-vision and robotic-mapping problems [32].
The number of ORB descriptors per image is set to 2000.
We denote the appearance-word vocabulary as V (i.e., B =
log2 |V |). wr is called pose word and represents the spatial
location of the feature keypoint descriptor with respect to the
object region; it incurs a small constant space cost B′ =log2 |A|
(bits) where |A| is the area [pixels] of the object region.
Consequently, the total space-cost (bits) of our BoW-based
model approximately sums up to
##
"
"
C=
∑ ∑ ∑
x∈X
r∈Rx
log2 |V | + B′
,
(2)
w∈Wr
where X is the set of place regions, Rx is the object regions
(described in II-A) in the x-th place, and Wr are the visual
words belonging to the object r.
Apparently, it is necessary to minimize the sizes of X,
Rx , Wr and V without compromising change-classification
accuracy. Joint minimization of these is generally intractable,
and in this study, we straightforwardly consider each minimization problem separately. Minimizing Rx , Wr and V are
the tasks of change-aware object proposal, feature extraction,
and vocabulary design. We have also addressed the issue
of vocabulary design and the use of BoW as a compressed
form of classifiers in an alternative context of visual robot
localization in [15]. Minimizing P is called unsupervised
place-definition and workspace-partitioning discovery. We
have addressed this problem in our previous studies in the
Fig. 2.
Pipeline of classifier-compression/-decompression task. Red:
positive examples. Blue: negative examples.
1e+06
100000
10000
1000
0
50
100
150
200
250
300
Fig. 3. Effect of map compression. Vertical axis: Total space cost (bits) for
all the place-specific change classifiers. Horizontal axis: Frame ID. Black
and red curves exhibit space costs for compressed and non-compressed (or
decompressed) maps. Each k-th curve corresponds to the k-th object cluster,
and the intervals between the (k − 1)-th and k-th curves equals the space
cost of the k-th object cluster.
context of place recognition [33] and change recognition
[27]. In this study, we straightforwardly partition the entire
sequence of mapped images in the workspace into equallength subsequences, each of which corresponds to each
place class.
Our algorithm consists of two distinct phases: training
and classification. The training procedure runs through the
following steps: (t1) Train classifiers. (t2) Represent classifiers by their positive/negative examples. (t3) Approximate
the examples by BoWs. (t4) Memorize the examples in the
form of BoW. The classification procedure runs through the
following steps: (c1) Lookup the vocabulary to determine
examples that correspond to the specified reference image.
(c2) Reproduce classifiers from the examples. (c3) Classify
query features using the reproduced classifiers. (c4) Delete
classifiers. Note that steps (c2) and (c4) function as decompression and compression of classifiers, respectively (Fig.
2). The time overhead of the compression task is more
or less negligible, while that of the decompression task
is marginal albeit noticeable. Therefore, it is necessary to
perform task scheduling of decompression task adequately.
In this study, we follow a straightforward strategy: When
the robot approaches or enters a new submap’s region,
decompress the classifiers of its surrounding submaps.
Figure 3 demonstrates our strategy. In the figure, the frame
IDs ranging from 137 to 237 correspond to the classifiers for
the robot’s surroundings. As illustrated, the space cost of the
Fig. 4.
Overview of scene recognition.
compressed classifiers is significantly lower than that of noncompressed or decompressed classifiers, which demonstrates
the effectiveness of our approach.
A. Classifier-learning
The place-specific change model aims to detect changes
at the level of object. It consists of a set of object-level
classifiers, each of which aims to learn the appearance of
a known object (i.e., no-change object) in the reference
image of a specific place and then, to classify an unseen
query feature as either change or no-change with respect to
the learned object. To this end, object segmentation both at
training and at classification tasks significantly influences the
performance of the object-level change detection.
The training phase begins with the extraction of a collection of object proposals from the specified reference image
(Fig. 4b left). We use BING object-proposal algorithm [34]
to achieve this because it is highly efficient to provide object
proposals with category-independent image windows. Then,
the proposed object-regions are further grouped into clusters
of spatially-near regions, termed object clusters (Fig. 4b
right). We use a criterion in this clustering wherein any two
object regions belong to the same cluster if their bounding
boxes overlap each other. Then, an additional “background”
object cluster, which treats the entire image region as the
object region, is added. Then, a change classifier is trained
for each object cluster. We will implement and evaluate two
types of classifiers: nearest neighbor, or SVM, and four types
of SVM kernel: linear, sigmoid, polynomial, or RBF. The
above methods of object proposal and classifier-learning are
detailed below.
The BING algorithm generally produces a large number of object proposals (e.g., 2×103 proposals per image).
The proposals generally contain numerous false positives.
Evaluating all the proposals is computationally intractable.
Therefore, we decide to select a marginal portion of the
object proposals. From the viewpoint of effectiveness of our
BoW model, we evaluate spatial density of visual words
N/A, where A and N are the area of the object proposal’s
bounding box and the number of visual words inside the
bounding box, respectively. Then, we select 400 object
proposals with highest density. Subsequently, near-duplicate
object regions that overlap significantly with object regions
with higher density are eliminated. We consider two object
regions i and j to be in the near-duplicate condition if their
overlap ratio Ai j / min(Ai , A j ) exceeds a threshold of 0.5;
here, Ai , A j and Ai j are the areas [pixels] of the objects i
and j and their overlap region.
The change classifier addresses one-sided classification
problem, wherein only negative examples (i.e., no-change)
are provided during the training stage. In order to obtain a
sufficiently large number of positive examples, we consider
the task that we call change mining [27], the aim of which
is to mine features of potential change objects from a large
feature collection of the EKB. In this study, we consider
three strategies for change mining: “uniform,” “farthest,” and
“nearest.” The “uniform” strategy uniformly samples positive
examples from the EKB, while the “farthest” and “nearest”
strategies sample positive examples that are respectively the
farthest and nearest in Hamming distance from the specified
negative examples. For all the strategies, each example in
the EKB that is nearer than 10 bit from its nearest neighbor
negative is considered inappropriate as a positive example
and eliminated from the candidates of positive examples prior
to the change mining. To enable efficient training, the number
of training examples is truncated to a maximum number 400.
B. Change Classification
The change classification process aims to rank local
feature-descriptors in query images in the order of likelihood
of changes. It consists of two distinct steps: (1) image
registration and (2) change ranking. Both the steps are
detailed below.
The image registration step, which aims to align query and
reference images into the same coordinate frame (Fig. 4c), is
a necessary pre-processing step for a substantial majority of
change detection tasks [5]. Note that image alignment from
monocular image-pairs is significantly ill-posed, particularly
when we are provided only BoW representations of the images. We have tested three strategies for feature matching—
ORB keypoint matching with and without post-verifications
using RANSAC [35] and using vector field consensus (VFC)
[36]—and observed that these two post-verification strategies
(i.e., RANSAC and VFC) are not adequately stable in our
scenario of highly-complex scenes. Therefore, we use ORB
keypoint matching as a method for image alignment.
In this study, we follow a fundamental procedure for image
registration [18], which assumes a linear transformation from
reference- to query-image coordinate (Fig. 4c), although a
key difference is that we are given the BoW representation
instead of raw feature descriptors. Note that the transformation algorithm requires pairs of feature keypoints matched
between query and reference images. In order to filter out
outlier matches to the maximum, an image region commonly
visible in both images is estimated in the form of bounding
box (Fig. 4a). More formally, keypoints of matched visual
words M are sorted in ascending order of x- or y-coordinates,
and we define ⌊δ |M|⌋-th and ⌈(1 − δ )|M|⌉-th elements (δ =
0.1) in the sorted lists as the x- and y-locations of the
boundary of the visible regions.
Two types of additional spatial cues are obtained as a
byproduct of the image registration. The first one is global
spatial-information for visibility analysis (Fig. 4a), wherein
local features outside the commonly visible region are not
considered as change-object candidates. The second is local spatial-information by which each object region (i.e.,
bounding box) in the reference image is first transformed
to the query image’s coordinate (Fig. 4c); following this,
the query image’s local features outside the transformed
bounding box (Fig. 4d) are not considered as matching
candidates of the specific object in the reference image.
Taking into account the transformation errors, we expand
the transformed bounding box by ∆L [pixel], wherein the
margin ∆L is currently set to 10% of the image width.
The change-ranking step aims to rank query features by the
likelihood of change, given the learned place-specific objectlevel change classifiers. First, each query feature is assigned
to the spatially-near object cluster in the corresponding
reference image (Fig. 4d). We assign each query feature to
an object cluster if its keypoint is located within one of the
object cluster’s bounding box. Then, each query feature is
input to the assigned object cluster’s change classifier, and
the classifier outputs the probability pr of the query feature
not originating from the learned object cluster r. We employ
Platt scaling [37] with five-fold cross-validation to convert an
output value of the SVM classifier to the probability estimate
or use a Gaussian [1 − exp(−d 2 /σd2 )] to convert an output
distance of the nearest neighbor classifier d into a probability
value. Given the output probability value pc from each object
cluster c, the probability of the query feature of interest being
changed is computed by
p = min pc .
c
(3)
In order to render the set of top-ranked features more
diverse, we introduce the idea of non-maximal suppression
and penalize features that are considered similar to a higherranked feature. In our view, two features are similar if they
belong to the same object cluster. Based on the concept,
we extract object clusters from a specified query image and
group query features into clusters in a similar manner as
in the reference images; then, features in each cluster are
ranked in descending order of the probability p. Then, the
entire set of query features is re-ranked in the ascending
order of augmented score r + (1 − p), where r(∈ [1,C]) is the
abovementioned intra-cluster rank, C is the size of the largest
cluster, and p(∈ [0, 1]) is the original probability-estimate.
Note that all query features do not always belong to object
clusters, and there may be isolated query features that do not
belong to any object cluster. We empirically determined that
such isolated features are generally less-reliable. Therefore,
we assign them a worst intra-cluster rank r = C + 1.
Fig. 5. Examples of scene recognition. Left panel: ORB keypoints (red
dots) and commonly visible regions estimated between query and reference
images (purple boxes). Middle and right panels: Object regions and object
clusters extracted from both images, using different colors for different
regions and clusters.
III. E XPERIMENTS
We evaluated various change detection strategies using
NCLT dataset [1]. NCLT dataset is a long-term autonomy
dataset for robotics research collected on the University of
Michigan’s North Campus. The dataset consists of omnidirectional imagery, 3D lidar, planar lidar, GPS, and odometry,
and we use the monocular images from the front-directed
camera (“camera #5”) for our change detection tasks. Fig. 5
shows some examples of scene recognition.
During the vehicle’s travel through both indoor and outdoor environments (Fig. 1a), it encounters various types
of changes, which originate from the movement of individuals, parking of cars, furniture, building construction,
opening/closing of doors, and placement/removal of posters
(Fig. 1b). There are also nuisance changes that originate
0.8
success ratio
0.7
0.6
0.5
0.4
0.3
object
non-object
w/o suppression
0.2
0.1
0.125
Fig. 6.
0.25
0.5
1
top-X (%)
2
4
Change detection performance.
from illumination alterations, viewpoint-dependent changes
of objects’ appearances and occlusions, weather variations,
and falling leaves and snow. A critical and significant challenge in a substantial majority of change detection tasks is to
discriminate changes of interest from nuisances. This renders
our change detection task significantly more demanding.
We use four datasets “2012/1/22,” “2012/3/31,”
“2012/8/4,” and “2012/11/7” that correspond to four
sessions of vehicle navigation. These datasets consist of
5095, 3994, 4877, and 5118 images of size 1232 × 1616.
We manually created 7571 pairs of corresponding query
and reference images, which correspond to “revisiting” or
“loop-closing” situations [28], using the global viewpoint
information. While 548 of the pairs are change image-pairs,
6714 pairs are no-change image-pairs; the remaining 309
pairs are not independent of the 548 change pairs and are
not used in the experiments. We categorized small changes
(e.g., 10×10 [pixels]) that typically originated from distant
objects, into no-change because it is challenging to detect
such small changes by visual change detection algorithm.
As a result, such small objects are likely to cause pseudo
false-positive detection by change detection algorithms. We
use specific combinations of query and reference image sets:
(query, reference) = (2012/1/22, 2012/3/31), (2012/3/31,
2012/8/4), (2012/8/4, 2012/11/7), or (2012/11/7, 2012/1/22).
The vocabulary is sized 1 M (i.e., 20-bit visual words) by
default, and its exemplar features are randomly sampled
from the reference set’s visual features. Change objects in
the 548 change image pairs are manually annotated in the
form of bounding boxes.
The average time-overhead for compression and decompression per classifier was 0.3 s and time-overhead per place
was 3.0 s using a non-optimized implementation of SVM
from SVM light [38] on a laptop PC (Intel Core i5-4200 2.5
GHz). This implies that the compression/decompression can
function in real-time when the robot moves at the speed of
1 m/s and the workspace is divided into 3-m-length places.
The total travel distance was 22,181 m.
We consider a straightforward change detection task scenario: Given a collection of 100 query images acquired by
a robot, identify the image that is the changed one and
the locations of the change features within that image. The
size of an image collection is set to 100, and it consists
of a change image and 99 random no-change images. We
created 548 collections based on the above 548 change image
pairs, for each of which 99 no-change images are randomly
sampled from the 6714 no-change pairs, and the collections
are commonly used as dataset by each change detection
algorithm.
We consider various strategies of change mining, classifiers, and image registration. For the change mining, we
implemented three methods: “uniform,” “farthest,” and “nearest,” which are described in II-A. For the classifier, we
implemented the SVM classifier with four kernels—linear,
sigmoid, polynomial, and RBF—and the nearest neighbor
classifier with Hamming distance metric. For the image
registration, we aim to verify the efficacy of the visibility analysis strategy presented in II-B and implemented
the common visible region in the form of bounding box
(“horizontal-vertical”) as well as two additional methods for
visibility analysis. One of these two is a method that omits
y-direction boundaries, which is motivated by an observation
that y-direction spatial information is generally unreliable
(“horizontal”). The other is a method that completely omits
the bounding-box information, which corresponds to not
using the visibility analysis strategy (“none”). By default, we
use SVM with RBF kernel as the classifier, uniform sampling
with 10-bit distance threshold as the change mining strategy,
and the 20-bit vocabulary.
Figures 6, 7, 8, 9 and 10 exhibit the performance results.
We evaluated the various methods over the independent
548 image collections. Each image collection consists of
100 pairing of query and reference images, and each query
image contains 2000 features to score. We assess whether the
change detection task on an image collection is successful
or not and compute the success ratio over all the image
collections. For the assessment, all the 100×2, 000=200, 000
features are sorted in the ascending order of the augmented
score (described in II-B), and if a ground-truth change feature
is ranked within the top-X (%) with respect to the sorted list,
the change detection is considered as success; otherwise, it
is considered as failure. We performed the evaluation for
various X values: 0.1%, 0.25%, 0.5%, 1%, 2.5%, and 5%,
which respectively correspond to the top 200, 500, 1000,
2000, 5000, and 10000 ranked query features.
Figure 6 illustrates results of evaluating the fundamental effects of the proposed strategies—object-level change
detection (II-A) and non-maximal suppression (II-B). We
developed a comparison method termed “non-object,” which
employs a single image-level classifier rather than the objectlevel classifier. The image-level classifier is implemented as
a single object-level classifier that treats the entire image
as the object region. We also developed another comparison
method termed “non-suppression,” which does not employ
the non-maximal suppression technique. In the figure, “object” indicates the proposed method (classifier: SVM w/ RBF
kernel, vocabulary size 20 bit), and “non-object” and “nonsuppression” indicate the two comparison methods. It can
be observed that the proposed “object” method evidently
outperforms the other two, and it is verified the object-level
change classifiers are substantially more powerful than the
comparison feature-level methods.
Figure 7 illustrates the result of comparing classifiers
0.8
0.7
0.7
0.6
0.6
success ratio
success ratio
0.8
0.5
0.4
SVM w/ RBF
SVM w/ linear
SVM w/ polynomial
SVM w/ sigmoid
nearest neighbor
0.3
0.2
0.4
0.3
horizontal-vertical
horizontal
none
0.2
0.1
0.1
0.125
Fig. 7.
0.25
0.5
1
top-X (%)
2
4
0.125
Influence of classifiers and kernels.
Fig. 9.
0.25
0.5
1
top-X (%)
2
4
Influence of visibility-analysis strategies.
0.8
0.8
0.7
success ratio
0.7
success ratio
0.5
0.6
0.5
0.4
0.2
Fig. 8.
0.25
0.5
1
top-X (%)
2
0.4
20-bit
16-bit
14-bit
12-bit
10-bit
0.2
0.1
0.1
0.125
0.5
0.3
uniform (10-bit)
uniform (20-bit)
uniform (30-bit)
farthest (10-bit)
nearest (10-bit)
0.3
0.6
4
Influence of change-mining strategies.
and kernels. We evaluated different classifiers and kernels
(nearest neighbor, SVM with linear, sigmoid, polynomial,
and RBF kernels). It can be seen that SVM with RBF kernel
outperforms the other classifiers and kernels.
Figure 8 shows a result of comparing different change
mining strategies. We evaluated different change mining
strategies (uniform with 10-bit, 20-bit, and 30-bit distance
threshold, farthest, and nearest). It can be seen that uniform sampling with 10-bit distance threshold outperforms
the other methods. The reason might be that the uniform
sampling strategy tends to produce a more diverse set of
positive examples than other strategies, and thus the resulted
classifier exhibits adequate generalization capability.
Figure 9 illustrates the result of evaluating the visibility
analysis strategies. We evaluated different visibility analysis
strategies (horizontal-vertical, horizontal, and none). It can
be observed that the horizontal-vertical strategy outperforms
the other two strategies when the parameter X is sufficiently
marginal.
Figure 11 illustrates examples of successful and failed
change detection. Here, we used the abovementioned default
settings. As illustrated, changes originating from moving
objects such as cars and pedestrians are generally easy to
detect owing to the fact that their visual appearances are
significantly discriminative from other background objects
such as roads and trees. However, the detection of moving
objects is rendered a substantially challenging task when a
similar moving object appears in the corresponding reference
image also (Fig. 11 failure examples). Although spatial
information (e.g., location and size) of such similar objects
are generally dissimilar to that of the query object, they
are ordinarily accepted as matching candidates of the query
object because the search region is expanded to overcome
coordinate transformation error (as explained in II-B). Other
challenging cases include change objects such as boxes and
tables (Fig. 11 failure examples) whose visual appearance
0.125
0.25
Fig. 10.
0.5
1
top-X (%)
2
4
Influence of vocabulary size.
are significantly similar to background objects such as floors
and walls.
IV. C ONCLUSIONS
We presented a change detection framework that realizes
map compactness while maintaining detection efficiency.
Rather than memorizing pre-trained classifiers, our ZSLbased approach has to memorize only the compact indexes to their training examples that are mined from EKB.
Experimental results on place-specific object-level change
classifiers have demonstrated high potential. The proposed
algorithm is efficient and very simple to implement. It
should therefore be convenient to integrate it into existing
frameworks of change detection (e.g., change detection in
3D, stereo images) to enhance their compactness.
R EFERENCES
[1] N. Carlevaris-Bianco, A. K. Ushani, and R. M. Eustice, “University of
michigan north campus long-term vision and lidar dataset,” The Int.
Journal of Robotics Research, vol. 35, no. 9, pp. 1023–1035, 2016.
[2] F. Pomerleau, P. Krüsi, F. Colas, P. Furgale, and R. Siegwart, “Longterm 3d map maintenance in dynamic environments,” in Proc. IEEE
ICRA, 2014.
[3] M. Fehr, M. Dymczyk, S. Lynen, and R. Siegwart, “Reshaping our
model of the world over time,” in Proc. IEEE ICRA, 2016.
[4] P. F. Alcantarilla, S. Stent, G. Ros, R. Arroyo, and R. Gherardi, “Streetview change detection with deconvolutional networks.” in Robotics:
Science and Systems, 2016.
[5] R. J. Radke, S. Andra, O. Al-Kofahi, and B. Roysam, “Image
change detection algorithms: a systematic survey,” IEEE Trans. image
processing, vol. 14, no. 3, pp. 294–307, 2005.
[6] J. Košecka, “Detecting changes in images of street scenes,” in Proc.
ACCV, 2012, pp. 590–601.
[7] S. Stent, R. Gherardi, B. Stenger, and R. Cipolla, “Detecting change
for multi-view, long-term surface inspection.” in BMVC, 2015.
[8] L. Gueguen and R. Hamid, “Large-scale damage detection using
satellite imagery,” in Proc. IEEE Conf. CVPR, 2015, pp. 1321–1328.
[9] A. A. Nielsen, “The regularized iteratively reweighted mad method for
change detection in multi-and hyperspectral data,” IEEE Trans. Image
processing, vol. 16, no. 2, pp. 463–478, 2007.
[10] F. Bovolo, L. Bruzzone, and M. Marconcini, “A novel approach to
unsupervised change detection based on a semisupervised svm and
a similarity measure,” IEEE Trans. Geoscience and Remote Sensing,
vol. 46, no. 7, pp. 2070–2082, 2008.
Fig. 11.
Change detection examples.
[11] L. Bruzzone and D. F. Prieto, “Automatic analysis of the difference
image for unsupervised change detection,” IEEE Trans. Geoscience
and Remote sensing, vol. 38, no. 3, pp. 1171–1182, 2000.
[12] B. Yamauchi and P. Langley, “Place learning in dynamic real-world
environments,” Proc. RoboLearn, vol. 96, pp. 123–129, 1996.
[13] E. Gavves, T. Mensink, T. Tommasi, C. G. Snoek, and T. Tuytelaars,
“Active transfer learning with zero-shot priors: Reusing past datasets
for future tasks,” in Proc. IEEE ICCV, 2015, pp. 2731–2739.
[14] C. H. Lampert, H. Nickisch, and S. Harmeling, “Attribute-based
classification for zero-shot visual object categorization,” IEEE Trans.
PAMI, vol. 36, no. 3, pp. 453–465, 2014.
[15] K. Tanaka, “Cross-season place recognition using NBNN scene descriptor,” in Proc. IEEE/RSJ Int. Conf. IROS.
[16] C. Cortes and V. Vapnik, “Support vector machine,” Machine learning,
vol. 20, no. 3, pp. 273–297, 1995.
[17] T. Cover and P. Hart, “Nearest neighbor pattern classification,” IEEE
Trans. information theory, vol. 13, no. 1, pp. 21–27, 1967.
[18] D. W. van de Wouw, G. Dubbelman, and P. H. de With, “Hierarchical
2.5-d scene alignment for change detection with large viewpoint
differences,” IEEE Robotics and Automation Letters, vol. 1, no. 1,
pp. 361–368, 2016.
[19] A. Taneja, L. Ballan, and M. Pollefeys, “Geometric change detection
in urban environments using images,” IEEE Trans. PAMI, vol. 37,
no. 11, pp. 2193–2206, 2015.
[20] J. P. Underwood, D. Gillsjö, T. Bailey, and V. Vlaskine, “Explicit 3d
change detection using ray-tracing in spherical coordinates,” in Proc.
IEEE ICRA, 2013, pp. 4735–4741.
[21] W. Li, X. Li, Y. Wu, and Z. Hu, “A novel framework for urban change
detection using vhr satellite images,” in Proc. ICPR, 2006.
[22] S. Ren, K. He, R. Girshick, and J. Sun, “Faster r-cnn: Towards realtime object detection with region proposal networks,” IEEE Trans.
PAMI, vol. 39, no. 6, pp. 1137–1149, 2017.
[23] H. Andreasson, M. Magnusson, and A. Lilienthal, “Has somethong
changed here? autonomous difference detection for security patrol
robots,” in Proc. IEEE/RSJ Int. Conf. IROS, 2007, pp. 3429–3435.
[24] P. Ross, A. English, D. Ball, B. Upcroft, G. Wyeth, and P. Corke,
“Novelty-based visual obstacle detection in agriculture,” in Proc. IEEE
ICRA, 2014, pp. 1699–1705.
[25] S. Stent, R. Gherardi, B. Stenger, K. Soga, and R. Cipolla, “An imagebased system for change detection on tunnel linings.” in MVA, 2013.
[26] T. Murase, K. Tanaka, and A. Takayama, “Change detection with
global viewpoint localization,” in IAPR Asian Conf. Pattern Recognition, 2017.
[27] X. Fei and K. Tanaka, “Unsupervised place discovery for placespecific change classifier,” CoRR, vol. abs/1706.02054, 2017. [Online].
Available: http://arxiv.org/abs/1706.02054
[28] K. Tanaka, Y. Kimuro, N. Okada, and E. Kondo, “Global localization
with detection of changes in non-stationary environments,” in Proc.
IEEE ICRA, vol. 2, 2004, pp. 1487–1492.
[29] M. Ando, Y. Chokushi, K. Tanaka, and K. Yanagihara, “Leveraging
image-based prior in cross-season place recognition,” in Proc. IEEE
ICRA, 2015, pp. 5455–5461.
[30] J. Yang, Y.-G. Jiang, A. G. Hauptmann, and C.-W. Ngo, “Evaluating
[31]
[32]
[33]
[34]
[35]
[36]
[37]
[38]
bag-of-visual-words representations in scene classification,” in Proc.
Int. W. multimedia information retrieval. ACM, 2007, pp. 197–206.
E. Rublee, V. Rabaud, K. Konolige, and G. Bradski, “Orb: An efficient
alternative to sift or surf,” in IEEE ICCV, 2011, pp. 2564–2571.
R. Mur-Artal, J. M. M. Montiel, and J. D. Tardos, “Orb-slam: a
versatile and accurate monocular slam system,” IEEE Tran. Robotics,
vol. 31, no. 5, pp. 1147–1163, 2015.
X. Fei, K. Tanaka, K. Inamoto, and G. Hao, “Unsupervised place
discovery for visual place classification,” in IAPR Int. Conf. Machine
Vision Applications, 2017, pp. 109–112.
M.-M. Cheng, Z. Zhang, W.-Y. Lin, and P. Torr, “Bing: Binarized
normed gradients for objectness estimation at 300fps,” in Proc. IEEE
Conf. CVPR, 2014, pp. 3286–3293.
R. Raguram, O. Chum, M. Pollefeys, J. Matas, and J.-M. Frahm,
“Usac: a universal framework for random sample consensus,” IEEE
Trans. PAMI, vol. 35, no. 8, pp. 2022–2038, 2013.
J. Ma, J. Zhao, J. Tian, A. L. Yuille, and Z. Tu, “Robust point matching
via vector field consensus,” IEEE Trans. Image Processing, vol. 23,
no. 4, pp. 1706–1721, 2014.
C. Elkan and K. Noto, “Learning classifiers from only positive
and unlabeled data,” in Proc. ACM SIGKDD Int. Conf. Knowledge
discovery and data mining, 2008, pp. 213–220.
T. Joachims, “Svmlight: Support vector machine,” SVM-Light Support
Vector Machine http://svmlight. joachims. org/, University of Dortmund, vol. 19, no. 4, 1999.
| 1cs.CV
|
arXiv:1603.06002v1 [cs.DS] 18 Mar 2016
A Message Passing Algorithm for the Problem of
Path Packing in Graphs
Patrick Eschenfeldt and David Gamarnik
March 22, 2016
Abstract
We consider the problem of packing node-disjoint directed paths in
a directed graph. We consider a variant of this problem where each
path starts within a fixed subset of root nodes, subject to a given
bound on the length of paths. This problem is motivated by the socalled kidney exchange problem, but has potential other applications
and is interesting in its own right.
We propose a new algorithm for this problem based on the message
passing/belief propagation technique. A priori this problem does not
have an associated graphical model, so in order to apply a belief propagation algorithm we provide a novel representation of the problem
as a graphical model. Standard belief propagation on this model has
poor scaling behavior, so we provide an efficient implementation that
significantly decreases the complexity. We provide numerical results
comparing the performance of our algorithm on both artificially created graphs and real world networks to several alternative algorithms,
including algorithms based on integer programming (IP) techniques.
These comparisons show that our algorithm scales better to large instances than IP-based algorithms and often finds better solutions than
a simple algorithm that greedily selects the longest path from each
root node. In some cases it also finds better solutions than the ones
found by IP-based algorithms even when the latter are allowed to run
significantly longer than our algorithm.
1
Introduction.
In this paper we consider the problem of packing node-disjoint directed paths into a directed graph, with each path starting within a
designated subset of “root” nodes. For a given maximum path length,
our objective is to include as many nodes as possible in paths. This
1
problem is motivated by the kidney exchange problem (KEP) studied in [2], in which donors and recipients must be matched together to
maximize the number of donations. In some situations these donations
are performed in chains starting from a lone donor, and logistical concerns encourage limiting the length of chains. This problem also has
other potential applications such as coordinating chained transactions,
and is an interesting and hard combinatorial optimization problem in
general.
We present a new algorithm for this problem based on the message
passing/belief propagation technique (BP for short). This algorithm
depends on a new representation of the problem that allows us to create a graphical model which embeds the optimization problem. After
describing this representation we will show that a direct application of
standard BP results in poor scaling, so we provide an efficient implementation that provides significant scaling advantages. Letting n be
the number of nodes in the graph, K be the maximum path length,
and ∆ be the maximum degree, the efficient implementation can perform a full iteration of the algorithm in O (n∆K) computations, while
a direct implementation of belief propagation requires O n2 ∆6 K 2 .
We compare our algorithm to several alternatives and show that it
consistently scales well and often finds better solutions than other fast
algorithms. In particular, we compare our efficiently implemented BP
to a simple greedy algorithm, called Greedy from now on for simplicity, and an existing Integer Programming (IP) algorithm used in [2]
which we call KEP. The Greedy algorithm which we consider makes
greedy selections only between constructed paths and not during the
construction of individual paths, as at every root node it searches exhaustively for the longest path which can be added to the solution
starting from that node. We also use our new representation of the
problem to derive a secondary novel algorithm based on IP, which we
call Parent-Child-Depth (PCD). Comparisons are made on a variety
of randomly created graphs as well as on several real world networks.
Our random graphs are generated with each edge existing independently with probability c/n for various values of c. We find that for
n = 1000 and n = 10000 with a path length bound of 5 nodes, with
10% to 25% of nodes designated as root nodes, and c between 2 and
5, our message passing algorithm consistently finds more nodes than
Greedy while running in comparable time. In many cases BP and
Greedy run significantly faster than the IP algorithms, which are cut
off after a time limit if they have not yet found the optimal solution,
and in several of these cases BP also finds solutions with more nodes
than any other algorithm. We find that BP also performs well relative
to alternative algorithms when the path length bound is increased to
10 or 15 nodes. In the regime where paths as long as 15 nodes are
allowed, PCD often finds the optimal solution in reasonable time while
2
KEP almost never does. We also provide numerical results for real
world networks, with sizes ranging from 5518 nodes to 260982 nodes.
In all but the smallest network the IP algorithms produce no useful
results, and in most graphs BP finds better solutions on average with
respect to a random choice of root nodes than Greedy, while running
in comparable time.
Our approach to this problem is motivated by previous work applying message passing techniques to the problem of prize-collecting
Steiner trees (PCST) done by Bayati, et. al. [4] (also used by Bailly,
et. al. [3]). In particular, the PCST problem does not have a natural
representation as a graphical model so the authors in [4] design a new
representation of the problem that leads to a graphical model and then
show how to implement BP on that model. This paper uses a similar
approach for the path packing problem.
In another relevant paper, Altarelli, et. al. [1] apply message passing techniques to edge-disjoint paths. They directly derive a message
passing algorithm for general graphs from an iterative cost calculation
scheme valid for trees and apply it to locally tree-like graphs and find
that their algorithm performs better in terms of paths found than various alternatives. Unlike this work, in our model paths may end at any
non-root node (as we do not consider a lone root node to constitute a
path in itself, every path must include at least one edge and thus one
non-root node).
Another example of applying a message passing approach while
making major modifications to standard methods is [5]. The authors
introduce a new type of graphical model, called a memory factor network, along with a message passing style algorithm to perform inference on such a model. Designing the message passing algorithm to
match the structure of the problem provides efficiency advantages over
variants of belief propagation and also provides convergence guarantees.
The remainder of this paper is organized as follows: In section 2
we describe the problem formulation, in section 3 we detail the core
of our message passing algorithm, in section 4 we detail our method
for efficiently implementing this algorithm, in section 5 we discuss how
to translate the result of the message passing into a solution to the
original problem, in section 6 we discuss alternative algorithms, and
in section 7 we provide numerical results comparing our algorithm to
these alternatives.
2
Problem setup.
The problem description consists of a directed graph G with node set
[n] = {1, 2, . . . , n}, an edge set E, and a positive integer K. Let E ′ be
3
the undirected version of the edge set of G. We assume that G contains
no isolated nodes, so for every i ∈ [n] there exists some j ∈ [n] such
that (i, j) ∈ E ′ . The nodes [n] are partitioned into subsets U and V ,
and the edges are elements of U × V and V × V only. In particular,
no edges start in V and end in U . We define a directed simple path
in a directed graph as a sequence of distinct nodes {i1 , i2 , . . . , ik } such
that (il , il+1 ) ∈ E for all l = 1, . . . , k − 1. All paths considered in this
paper will be directed simple paths. For a path P = {i1 , . . . , ik } we
define the length of P , denoted |P |, to be the number of nodes in the
path, namely k. Our goal is to find a collection P = {P1 , . . . , Pk } of
node-disjoint paths with each path Pj starting with a node in U and
each having length at most K. Specifically, we define
X
v(P) =
|P |
P ∈P
to be the count of the nodes in a collection of paths. Let Π be the set
of feasible collections P. Then our problem is
max v(P).
P∈Π
(1)
We now introduce a formalization for the problem (1) that is conducive
to message passing.
We will use special symbols ∗ and • to represent certain relationships to be described below. We will also introduce the notation ∂i
to represent the set of nodes adjacent to the node i in the undirected
version of the graph G. That is, ∂i is the set of nodes j ∈ [n] such that
(i, j) ∈ E or (j, i) ∈ E. We also let ∂ X i = ∂ i ∩ X and let ∂i \ j stand
for the more formally correct ∂i \ {j}.
For each node i ∈ [n], we introduce variables (di , pi , ci ) where di ∈
[K] ∪ {∗}, pi ∈ ∂i ∪ {∗, •}, ci ∈ ∂ V i ∪ {∗, •}. These variables intend to
represent the depth of the node (di ), its parent (pi ), and its child (ci )
in a given solution. The special value ∗ indicates that the node does
not participate in a path, in which case we will have di = pi = ci = ∗.
The special value • indicates that the node is at one end of a path.
Thus if pi = • this means node i does not have a parent and thus starts
a path, whereas if ci = • node i does not have a child and thus ends a
path. Note that if ci is a node it must be a node in V because nodes
in U must either act as the root of a path or not participate in a path.
To formulate the optimization problem, we need to define a feasible
set and an objective function. We begin with the feasible set. We first
define a function that acts as an indicator for the consistency of the
variables (di , pi , ci ) at a single node i. For i ∈ [n], di ∈ [K] ∪ {∗},
4
pi ∈ ∂i ∪ {∗, •}, ci ∈ ∂ V i ∪ {∗, •}, let
fi (di , pi , ci ) = 1{di = ci = pi = ∗}
△
+ 1{pi = •, ci 6∈ {∗, •}, di = 1, i ∈ U }
+ 1{pi 6∈ {∗, •}, ci 6∈ {∗, pi }, di 6∈ {∗, 1}, i ∈ V }.
(2)
(3)
(4)
Term (2) is the case in which node i does not participate, term (3) is
the case in which node i is the root of a path, and term (4) is all cases
where node i participates in a path but is not the root. Note that
these three cases are mutually exclusive, so fi is an indicator function.
It is equal to zero if the variables at node i are not consistent with
any global configuration (d, p, c) = (dj , pj , cj )j∈[n] representing a set of
valid paths for the graph G, and equal to one otherwise.
For each pair of nodes i and j that are adjacent in the undirected
version of G, we define a function gij (di , pi , ci , dj , pj , cj ) of the variables
(di , pi , ci , dj , pj , cj ) for the nodes i and j which is an indicator that
these variables are consistent with each other and the existence or
nonexistence of the edges (i, j) and (j, i). Namely, they correspond to
a locally valid configuration, so if, e.g., node i reports that its parent is
node j, then the edge (j, i) exists, node j reports that its child is node
i, and we have di = dj + 1. Formally, for i ∈ [n], j ∈ ∂i, di ∈ [K] ∪ {∗},
pi ∈ ∂i ∪ {∗, •}, ci ∈ ∂ V i ∪ {∗, •}, dj ∈ [K] ∪ {∗}, pj ∈ ∂j ∪ {∗, •},
cj ∈ ∂ V j ∪ {∗, •}, let
△
gij (di , pi , ci , dj , pj , cj ) =
1{pi = j, cj = i, pj 6= i, ci 6= j, (j, i) ∈ E, di = dj + 1}
+ 1{pj = i, ci = j, pi 6= j, cj 6= i, (i, j) ∈ E, dj = di + 1}
+ 1{pi 6= j, cj 6= i, pj 6= i, ci 6= j}.
(5)
(6)
(7)
Lines (5) and (6) capture the cases where j is the parent of i and i is
the parent of j, respectively, and enforce the relationships that must
occur in those cases. Line (7) captures all cases where there is no
direct connection between nodes i and j. All three cases are mutually
disjoint, so gij is an indicator function. In summary, gij is equal to
one if parent/child relationships agree, the edges (i, j) and (j, i) are
used only in the appropriate directions, and depth relationships are
consistent.
Finally we define a indicator function for variable consistency at
both the node and edge level, which we define for each i ∈ [n], j ∈ ∂i
as
△
hij (di , pi , ci , dj , pj , cj ) = gij (di , pi , ci , dj , pj , cj )fi (di , pi , ci )fj (dj , pj , cj ).
(8)
With this definition of hij , we can define the set
△
M = (d, p, c) : hij (di , pi , ci , dj , pj , cj ) = 1 ∀ i ∈ [n], j ∈ ∂i
(9)
5
which will serve as the feasible set for our optimization.
Note that each collection P ∈ Π of node disjoint paths is associated
with a unique assignment of variables {(di , pi , ci )}i∈[n] , which we will
denote by (d(P), p(P), c(P)). In fact the converse is also true:
Proposition 1. For every (d, p, c) ∈ M there exists a unique P ∈ Π
such that (d, p, c) = (d(P), p(P), c(P)).
Proof. Given a set of variables (d, p, c) ∈ M , recall that for every i ∈ [n]
we have fi (di , pi , ci ) = 1 and for every j ∈ ∂i we have gij (di , pi , ci , dj , pj , cj ) =
1.
We will construct a feasible set of node-disjoint paths by the following procedure. Find the set of nodes i such that di = 1, which we
denote R.
If R is empty then for every i ∈ U we must be in case (2) because
case (3) is excluded. Thus di = pi = ci = ∗. For i ∈ V , by (2) and
(4) we either have di = ∗ or di > 1. But by (5) di = 2 only if dj = 1
for some j ∈ ∂i. This implies that di 6= 2 for any i ∈ V . Repeating
this argument implies di 6= l for any l ∈ [K], so we conclude di = ∗.
Then (2) implies pi = ci = ∗. Thus for every node i ∈ [n] we have
di = pi = ci = ∗, so we let P = ∅ because di (∅) = pi (∅) = ci (∅) = ∗ for
all i ∈ [n].
Otherwise, if R is nonempty, we will construct one path for each
i ∈ R. Consider some i1 ∈ [n] such that di1 = 1. By (3) we must have
i1 ∈ U , so it will be valid to start a path at i1 . Again by (3) we know
ci1 6= ∗, •, so ci1 = i2 for some i2 ∈ ∂ V i. By (6) with i = ii and j = i2
we have di2 = di1 + 1 = 2, (i1 , i2 ) ∈ E, pi2 = i1 and ci2 6= i1 . Thus it is
valid to start our path with the node sequence {i1 , i2 }. If ci2 = • then
we end this path and add P = {i1 , i2 } to P. If ci2 6= • then by di2 = 2
and (4) we have ci2 = i3 for some i3 ∈ ∂ V i2 . By (6) with i = i2 and
j = i3 we have di3 = 3, (i2 , i3 ) ∈ E, pi3 = i2 , and ci3 6= i2 . Thus it is
valid to start our path with the node sequence {i1 , i2 , i3 }. If ci3 = • we
end this path and add it to P. Otherwise we iterate, because (4) implies
ci3 = i4 for some i4 ∈ ∂ V i3 . In the generic step, we are considering
cil−1 = il for some il ∈ ∂ V il−1 and we have dil−1 = l − 1. Then (6)
for i = il−1 and j = il implies dil = l, (il−1 , il ) ∈ E, pil = il−1 and
cil 6= il−1 . If cil = • we terminate the path and add P = {i1 , i2 , . . . , il }
to P. Otherwise there exists some il+1 ∈ ∂ V il such that cil = il+1 so
we iterate. This process is guaranteed to terminate because dil = l and
dil ≤ K. Once we have added the path starting at node i1 we repeat
the process for each node i with di = 1.
To see that (d(P), p(P), c(P)) = (d, p, c) after this process, first
consider a node j which does not participate in any path in P. We
have dj (P) = pj (P) = cj (P) = ∗ so we want to show dj = pj = cj = ∗.
Because every node i with di = 1 starts a path in the above procedure,
we must have dj 6= 1. Suppose dj = l for some l ∈ {2, . . . , K}. Then
6
(4) implies pj 6= {∗, •} so there exists some i ∈ ∂j such that pj = i
and (6) implies di = l − 1. If i participates in a path in P then our
above procedure would also include j. If i does not participate then
we have found a non-participating node with depth l − 1. Thus if there
are were any non-participating j with dj ∈ [K] there would be a nonparticipating j with dj = 1, but we have already shown there is no
such j. We conclude that dj = ∗ for all non-participating dj . Then (2)
implies pj = cj = ∗, as desired.
Next consider a node il which participates in a path in P at depth l.
If l = 1 then il starts a path and has no parent and (by the above procedure) has child cil . Thus we have dil (P) = 1, pil (P) = •, and cil (P) =
cil . The only nodes il that start paths are those with dil = 1 and (3)
implies pil = •, so we have (dil (P), pil (P), cil (P)) = (dil , pil , cil ), as desired. If l > 1 then il has dl = l, parent il−1 = pil (as noted during the
procedure above) and either ends the path or has a child il+1 = cil . By
the procedure the path ends only if cil = •, so we have cil (P) = • = cil .
Thus we have (dil (P), pil (P), cil (P)) = (dil , pil , cil ), as desired. If the
path continues it continues to il+1 = cil so cil (P) = cil , and again we
have (dil (P), pil (P), cil (P)) = (dil , pil , cil ), as desired.
Because we have a bijection between feasible collections of nodedisjoint paths in G and the set M , we can treat M as the feasible region
for our optimization problem.
To define our objective, we define the function
(
0 di = pi = ci = ∗
△
η(di , pi , ci ) =
(10)
1 otw
which assigns value 0 to the case where the node i does not participate
in a path and value 1 to all other cases. We further define the function
△ X
η(di , pi , ci )
(11)
H(d, p, c) =
i∈[n]
that sums the value of each node across the entire graph. Note that
when applied to a valid (d, p, c), i.e. an element of the feasible set M ,
H simply counts the number of nodes of the graph which participate
in some path.
Thus we can reformulate the optimization problem (1) as
max
(d,p,c)∈M
3
H(d, p, c).
(12)
Algorithm.
Belief propagation algorithms are used to estimate either the marginal
distributions of variables or the solution to the maximum a posteriori
7
(MAP) problem in graphical models. We will use the framework corresponding to the latter option. The MAP solution for a probability
distribution is the most likely joint assignment of all the variables in
the distribution. To utilize this approach, we transform the optimization problem (12) into a problem of finding a MAP assignment for a
probability distribution. To achieve this transformation we introduce a
corresponding undirected graphical model. A maximum likelihood assignment of all variables in this distribution will be an optimal solution
to (12).
To create our undirected graphical model, we introduce node potentials
(
e−β di = pi = ci = ∗
φi (di , pi , ci ) =
(13)
1
otw
where β > 0 is a system parameter to be chosen. We use hij as defined
in (8) as our edge potentials. Together these define the probability
distribution P
Y
Y
hij (di , pi , ci , dj , pj , cj ).
φi (di , pi , ci )
P(d, p, c) ∝
(i,j)∈E ′
i∈[n]
This probability distribution assigns zero probability to any global configuration (d, p, c) which does not correspond to an element of M , and
among positive probability configurations it assigns higher probability
to those which include more nodes.
We can now solve our original optimization problem by solving the
MAP problem on this undirected graphical model. Note that we may
have multiple solutions.
The core component of the algorithm is the set of messages, which
we define for each ordered pair (j, i) with j ∈ [n] and i ∈ ∂j, representing the message from node j to node i. Thus each undirected
edge (i, j) ∈ E ′ is associated with two sets of messages, representing
the information sent from i to j and from j to i, respectively. The
message from j to i is denoted bj→i (di , pi , ci ) for all values of di , pi ,
and ci in the previously specified ranges di ∈ [K] ∪ {∗}, pi ∈ ∂i ∪ {∗, •},
ci ∈ ∂ V i ∪ {∗, •}. Our approach to this problem will be to implement
parallel belief propagation in the standard Min-Sum form (which can
be derived from the Max-Product form by taking the negative log of
all messages; see, e.g., [6] or [8]), with messages sent according to the
update rule
− log φj (dj , pj , cj )
bj→i (di , pi , ci ) = min
(dj ,pj ,cj )
− log hij (di , pi , ci , dj , pj , cj )
X
bk→j (dj , pj , cj ) .
+
k∈∂j\i
8
(14)
We use the convention − log(0) = ∞. In other words, if hij = 0 for
some configuration then that configuration will not be included in the
minimization. Otherwise the log(hij ) term contributes log(1) = 0.
Because our graphical model potentially has cycles, we must compute these messages iteratively from some starting values, which we set
to be bj→i (di , pi , ci ) = 1 for all j ∈ [n], i ∈ ∂j and all values of di , pi , ci ,
with the goal of finding a fixed point solution. From the messages we
can compute the max-marginals at each node as
!
X
p̄i (di , pi , ci ) = exp log φi (di , pi , ci ) −
bk→i (di , pi , ci ).
(15)
k∈∂i
These max-marginals are, according to the Belief Propagation approach, intended to approximate the probability of a maximum-probability
configuration of the whole system given the arguments di , pi , ci are
fixed. As such, they can be used in reconstructing the maximum global
configuration. The non-uniqueness of solutions to (12) complicates
this, however, which we discuss below in section 5.
The update rule (14) and max-marginal equation (15) characterize
the entirety of our algorithm. For some a priori maximum number of
iterations T and initial messages b0j→i (di , pi , ci ) = 1 for all j ∈ [n],
i ∈ ∂j and all values of di , pi , ci we compute bt from bt−1 according
to (14) for t ∈ [T ]. If the messages reach a fixed point, i.e. bt = bt+1
for some t < T , we halt iterations at that point. Furthermore, for
any 0 < t ≤ T we can use bt and (15) to compute estimates of the
max-marginals for any node in any configuration at that step of the
algorithm.
P
In this form, the number of messages is O K i∈[n] ∆3i where
∆i = |∂i| is the degree of node i in the undirected graph. Indeed,
for each node i ∈ [n] there are ∆i incoming messages, and for each
message there are K choices of di and O ∆2i choices of (pi , ci ) that
define messages. For ∆ = maxi∈[n] ∆i the number of messages is
O n∆3 K .
(16)
O n2 ∆6 K 2 .
(17)
At each iteration of the algorithm a message from j to i must be
updated by computing a minimization over all O K∆2j choices of
the variables (dj , pj , cj ). The argument of the minimization includes
a sum over the neighbors of j, which can be performed once for each
j ∈ [n] for a total computational costO (n) per iteration.
a
P Thus
P
full iteration of all messages requires O nK 2 i∈[n] ∆3i j∈∂i ∆2j
computations. For ∆ = maxi∈[n] ∆i , the number of computations is
9
Rather than implementing this scheme directly we will instead introduce a more efficient representation of the messages, which is described
in the next section.
4
Efficient representation.
In this section we introduce a more efficient representation of the message passing algorithm described above. This representation
will re
quire only O (nK∆) messages compared to O nK∆3 for the original
description (14), and will require O (nK∆) computations
to be per
formed per iteration compared to O n2 K 2 ∆6 computations for the
original representation.
Before defining the messages for the new representation, we now
introduce an intermediate set of messages and rewrite (14) as two steps,
with
bj→i (di , pi , ci ) =
min
dj ,pj ,cj s.t.
hij (di ,pi ,ci ,dj ,pj ,cj )=1
ψj→i (dj , pj , cj ),
ψj→i (dj , pj , cj ) = − log φj (dj , pj , cj ) +
X
bk→j (dj , pj , cj ).
(18)
(19)
k∈∂j\i
If the set over which we are minimizing in (18) is empty we define the
minimum to be ∞. The messages bj→i (di , pi , ci ) evolve exactly as they
do in (14).
We will define a collection of messages in terms of ψ and b, intended
as an efficient representation of the message passing algorithm, which
we call the A-H form. After defining the messages we will provide
the rules by which these messages are updated. We will also obtain
algorithmic complexity bounds for our representation, which are summarized in Theorem 1. Then we will demonstrate in Theorem 2 the
equivalence of the A-H form algorithm to the original description (14)
by showing that iterating the A-H form allows us to compute the maxmarginals (15) as if we had done the iterations in the original form
(14). This theorem will be proved in two parts, with Lemma 1 establishing a way to write all original messages bj→i (di , pi , ci ) in terms of
A-H form messages and Lemma 2 establishing how to update the A-H
form messages.
4.1
Message definitions.
The messages to be defined fall into three categories based on the type
of the node which sends the message, and the type of the node which
receives it. We first define messages sent from a non-root node j ∈ V
to another non-root node i ∈ V . Next we will describe messages sent
10
from a non-root node j ∈ V to a root node u ∈ U , and finally we
describe messages sent from a root node u ∈ U to a non-root node
i∈V.
For (i, j) ∈ E ′ , we define
(
∞ (i, j) 6∈ E
△
δi→j =
0 (i, j) ∈ E.
Fix (i, j) ∈ E ′ such that i, j ∈ V .
• For 3 ≤ d ≤ K, let
△
Adj→i = δi→j + min ψj→i (d, i, cj )
cj 6=∗,i
X
bk→j (d, i, cj ).
= δi→j + min
cj 6=∗,i
(20)
(21)
k∈∂j\i
Note that − log φj (d, i, cj ) = 0 for cj 6= ∗, so ψj→i (d, i, cj ) =
P
k∈∂j\i bk→j (d, i, cj ).
• For 2 ≤ d ≤ K − 1, let
△
d
Bj→i
= δj→i + min ψj→i (d, pj , i)
pj 6=∗,i,•
X
bk→j (d, pj , i).
= δj→i + min
pj 6=∗,i,•
k∈∂j\i
• For 2 ≤ d ≤ K, let
△
d
Fj→i
= min ψj→i (d, pj , cj )
pj 6=∗,i,•
cj 6=∗,i
pj 6=cj
X
= min
pj 6=∗,i,•
cj 6=∗,i k∈∂j\i
pj 6=cj
(22)
bk→j (d, pj , cj ).
• Let
△
Gj→i = ψj→i (∗, ∗, ∗)
X
bk→j (∗, ∗, ∗).
=β+
(23)
k∈∂j\i
• Let
△
d
.
Hj→i = min Gj→i , min Fj→i
2≤d≤K
Fix (j, u) ∈ E ′ such that j ∈ V and u ∈ U .
11
(24)
• Let
△
Aj→u = min ψj→u (2, u, cj )
cj 6=u,∗
X
bk→j (2, u, cj ).
= min
cj 6=u,∗
k∈∂j\u
• For 2 ≤ d ≤ K, let
△
d
=
Fj→u
=
min ψj→u (d, pj , cj )
pj 6=∗,u,•
cj 6=∗,u
pj 6=cj
X
min
pj 6=∗,u,•
cj 6=∗,u k∈∂j\u
pj 6=cj
bk→j (d, pj , cj ).
• Let
△
Gj→u = ψj→u (∗, ∗, ∗)
X
bk→j (∗, ∗, ∗).
=β+
k∈∂j\u
• Let
△
d
.
Hj→u = min Gj→u , min Fj→u
2≤d≤K
Fix (u, i) ∈ E ′ such that u ∈ U and i ∈ V .
• Let
△
Bu→i = ψu→i (1, •, i)
X
bk→u (1, •, i).
=
k∈∂u\i
• Let
△
min ψu→i (1, •, cu )
X
bk→u (1, •, cu ).
= min
Fu→i =
cu 6=i,∗,•
cu 6=i,∗,•
k∈∂u\i
• Let
△
Gu→i = ψu→i (∗, ∗, ∗)
X
bk→u (∗, ∗, ∗).
=β+
k∈∂u\i
12
• Let
△
Hu→i = min {Gu→i , Fu→i }
Note that when i ∈ U or j ∈ U there are different numbers of
messages in each direction. To refer to the set of messages from one
node to another we define for (i, j) ∈ E ′
K−1
3
K
2
2
K
i, j ∈ V
j→i , . . . , Fj→i , Gj→i , Hj→i
Aj→i , . . . , Aj→i , Bj→i , . . . , Bj→i , F
2
K
Xj→i =
j ∈ V, i ∈ U
Aj→i , Fj→i , . . . , Fj→i , Gj→i , Hj→i
(Bj→i , Fj→i , Gj→i , Hj→i )
j ∈ U, i ∈ V.
(25)
Note that in all cases the number of messages from i to j is O (K).
4.2 Update rules and the complexity of the A-H
form.
The iteration update rules for these new messages are listed in Tables
1-3. The derivation of these update rules from the definitions of the
messages is given in the proof of Theorem 2. These update rules allow
us to provide complexity bounds for the system:
Theorem 1. The A-H form has a total of O (n∆K) messages per
iteration. Each iteration requires a total of O (n∆K) computations to
update the entire system.
Remark. Recall
from (16) that the original form (14) has a total of
O n∆3 K messages, while by Theorem 1 the A-H form has O (n∆K)
total messages.
To perform aniteration of the system, the original form (14) requires O n2 ∆6 K 2 computations by (17), while the A-H form requires
only O (n∆K).
Proof of Theorem 1. For each undirected edge, the A-H form has O (K)
messages, so the total number of messages is O (|E ′ |K) = O (n∆K), as
desired.
In performing an iteration of the A-H form we proceed node-bynode. As we will show below, at a node j we will perform
j)
P O (K∆
total computations. This implies that we require O K j∈[n] ∆j =
O (n∆K) total computations for a full iteration of the A-H form, as
claimed.
We now provide more detail on the computations required per node
j. We consider the case j ∈ V ; if j ∈ U the analysis is similar. We
will also focus our attention on the most complex message to compute,
13
Adj→i
= δi→j +
X
Hk→j
k∈∂j\i
AK
j→i = δi→j +
X
d+1
+ min 0, min Ak→j − Hk→j
k∈∂ V j\i
Hk→j
k∈∂j\i
2
Bj→i
= δj→i +
X
Hk→j + min (Bw→j − Hw→j )
X
Hk→j + min
w∈∂ U j
k∈∂j\i
d
Bj→i
= δj→i +
k∈∂ V j\i
k∈∂j\i
2
Fj→i
=
X
=
w∈∂ U j
X
Hk→j + min
X
Hk→j + min
l∈∂ V j\i
k∈∂j\i
K
Fj→i
=
k∈∂j\i
Gj→i = β +
Hk→j + min (Bw→j − Hw→j ) + min 0, min
k∈∂j\i
d
Fj→i
d−1
− Hk→j
Bk→j
X
l∈∂ V j\i
d−1
Bl→j
− Hl→j + min 0,
K−1
− Hl→j
Bl→j
k∈∂ V j\i
min
A3k→j
k∈∂ V j\{i,l}
− Hk→j
Ad+1
k→j
Hk→j
Hj→i = min Gj→i , min
2≤d≤K
d
Fj→i
Table 1: Messages from j ∈ V to i ∈ V , for (i, j) ∈ E ′ and 3 ≤ d ≤ K − 1
14
− Hk→j
k∈∂j\i
Aj→u =
X
Hk→j + min 0, min
k∈∂j\u
2
Fj→u
=
X
Hk→j +
X
Hk→j
X
Hk→j
k∈∂j\u
d
Fj→u
=
k∈∂j\u
K
Fj→u
=
k∈∂j\u
Gj→u = β +
X
k∈∂ V j
A3k→j
− Hk→j
min (Bw→j − Hw→j ) + min 0, min
w∈∂ U j\u
k∈∂ V j
A3k→j
− Hk→j
d+1
d−1
+ min Bl→j − Hl→j + min 0, min Ak→j − Hk→j
l∈∂ V j
k∈∂ V j\l
K−1
− Hl→j
+ min Bl→j
l∈∂ V j
Hk→j
k∈∂j\u
Hj→u = min Gj→u , min
2≤d≤K
d
Fj→u
Table 2: Messages from j ∈ V to u ∈ U , for (u, j) ∈ E ′ and 3 ≤ d ≤ K − 1
Bu→i =
X
Hk→u
k∈∂u\i
Fu→i =
X
Hk→u + min (Ak→u − Hk→u )
k∈∂u\i
Gu→i = β +
X
k∈∂u\i
Hk→u
k∈∂u\i
Hu→i = min (Gu→i , Fu→i )
Table 3: Messages from u ∈ U to i ∈ V , for (u, i) ∈ E ′
15
d
which is Fj→i
for 3 ≤ d ≤ K − 1. We will require O (∆j ) computations
to compute this for all i ∈ ∂j and thus O (K∆j ) computations for all
messages sent from j because there are O (K) messages from j → i for
each i ∈ ∂j.
d
To compute Fj→i
it is necessary to compute two values:
X
Hk→j ,
k∈∂j\i
and
d−1
Bl→j − Hl→j + min 0,
min
l∈∂ V j\i
min
k∈∂ V j\{i,l}
Ad+1
k→j
− Hk→j
.
(26)
Rather than compute these directly for each i, we first compute
X
Hk→j ,
(27)
k∈∂j
select the three neighbors l ∈ ∂ V j with the smallest values of
d−1
Bl→j
− Hl→j ,
(28)
and select the three neighbors l ∈ ∂ V j with the smallest values of
Ad+1
l→j − Hl→j .
(29)
We also record the values of (28) and (29) for the nodes we record.
Once we have these quantities, for eachP
i ∈ ∂j we can compute
d
Fj→i
in O (1) computations. To compute
k∈∂j\i Hk→j we simply
subtract Hi→j from (27). To compute (26) we pick l and k according
to the minimum values we have recorded, subject to the constraints
that l 6= i, k 6= i, and l 6= k. These constraints are the reason we
record three minimizing values, as the first two choices for either l or
k may be eliminated by matching i or the other choice. For a given
d
i ∈ ∂j we compute Fj→i
in O (1) computations so computing all the
d
F messages sent out from j requires O (∆j ) computations.
All of the other messages are computed using similar patterns and
there are O (K) of them so a total of O (K∆j ) total computations are
required to compute all the messages sent out from j.
4.3
Initialization and termination.
The equations in Tables 1-3 collectively define update rules for our
0
algorithm. To initialize the system we set Xj→i
= 1 for all (i, j) ∈ E ′ ,
0
by which we mean we set each element of Xj→i to 1.
Given some a priori bound T on the number of iterations, we then
compute X t from X t−1 for t ∈ [T ] according to the update rules in
Tables 1-3. If the messages reach a fixed point with X t = X t+1 for
some t < T we halt the algorithm.
16
4.4
Validity of the A-H Form.
Identities relating original form messages bj→i (di , pi , ci ) to A-H form
messages for various choices of di , pi , and ci are given in Tables 4-6. It
can be checked that for any choice of (di , pi , ci ) that does not appear
in a table we have bj→i (di , pi , ci ) = ∞ by (18) because no choice of
(dj , pj , cj ) can satisfy hij (di , pi , ci , dj , pj , cj ) = 1. The validity of these
identities given the definitions of the A-H form variables is proved in
Lemma 1 below.
To use the original form (14) to find solutions to (12) we need to
compute the max-marginals (15) for different values of (di , pi , ci ) from
the messages bk→j (dj , pj , cj ). Specifically, we need to find the choices
of (di , pi , ci ) that maximize p̄i (di , pi , ci ). Thus to show that we can
iterate the A-H form to find the same approximate solutions we would
find by iterating the original form (14) we prove that the max-marginal
can be computed from the A-H form. In particular, we define
Definition 1. For node i ∈ [n], let ρi be any real function of AH variables {Xj→i }j∈∂i defined by substituting for bk→j (dj , pj , cj ) in
(15) according to the identities represented by Tables 4-6.
We prove the following:
Theorem 2. For all (j, i) ∈ E ′ , let the original form be initialized with
b0j→i (di , pi , ci ) = 1 for all (di , pi , ci ) and let the A-H form be initialized
0
with Xj→i
= 1. Then for each i ∈ [n], di ∈ [K] ∪ {∗}, pi ∈ ∂i ∪ {∗, •},
V
ci ∈ ∂ i ∪ {∗, •} and for 0 < t ≤ T ,
t
}j∈∂i
p̄i (di , pi , ci ) = ρi {Xj→i
when both sides are computed with appropriate messages btj→i and
t
Xj→i
updated t times according to (14) and Tables 1-3, respectively.
We will use two lemmas to prove this theorem:
Lemma 1. The identities described in Tables 4-6 hold.
Proof. Fix i, j ∈ V .
We focus on two representative messages, namely Adj→i for 3 ≤
d ≤ K − 1 and Hj→i . In the first example we will prove the fifth
identity of Table 4 and in the second will prove the sixth. The other
identities described in Table 4 and all identities described in Tables 5-6
are proved similarly so we omit the details.
We will first show that the fifth identity of Table 4 holds. Namely,
for 3 ≤ d ≤ K − 1 and p ∈ ∂ V i \ j we will show
bj→i (d, p, j) = Ad+1
j→i .
17
(30)
di
∗
2
2
3≤d≤K −1
3≤d≤K −1
3≤d≤K −1
K
K
pi
∗
p ∈ ∂U i
p ∈ ∂U i
j
p ∈ ∂V i \ j
p ∈ ∂V i \ j
j
p ∈ ∂V i \ j
ci
∗
j
V
c ∈ ∂ i \ j ∪ {•}
c ∈ ∂ V i \ j ∪ {•}
j
V
c ∈ ∂ i \ j ∪ {•}
•
•
bj→i (di , pi , ci )
Hj→i
A3j→i
Hj→i
d−1
Bj→i
Ad+1
j→i
Hj→i
K−1
Bj→i
Hj→i
Table 4: Classes of bj→i messages for i, j ∈ V .
du
∗
1
1
pu
∗
•
•
cu
∗
j
c ∈ ∂V u \ j
bj→u (du , pu , cu )
Hj→u
Aj→u
Hj→u
Table 5: Classes of bj→u messages for j ∈ V , u ∈ U .
di
∗
2
2
3≤d≤K −1
K
pi
∗
u
p ∈ ∂U i \ u
p ∈ ∂V i
p ∈ ∂V i
ci
∗
c ∈ ∂ V i∪ {•}
c ∈ ∂V i ∪
{•}
c ∈ ∂ V i \ j ∪ {•}
•
bu→i (di , pi , ci )
Hj→i
Bu→i
Hu→i
Hu→i
Hu→i
Table 6: Classes of bu→i messages for u ∈ U , i ∈ V .
18
Plugging di = d, pi = p, and ci = j into (18) gives us
bj→i (d, p, j) =
min
dj ,pj ,cj s.t.
hij (d,p,j,dj ,pj ,cj )=1
ψj→i (dj , pj , cj )
(31)
Suppose (i, j) 6∈ E. Then hij (d, p, j, dj , pj , cj ) = 0 for any choice of
dj , pj , cj and thus bj→i (d, p, j) = ∞. By (20) we have Ad+1
j→i = ∞, so
d+1
bj→i (d, p, c) = Aj→i , as desired.
Now suppose (i, j) ∈ E. We now use hij (d, p, j, dj , pj , cj ) = 1 to
infer information about dj , pj , cj . From (6) we must have dj = d + 1,
pj = i, and cj 6= i. Then (4) further implies cj 6= ∗. Substituting these
facts into (31) shows
bj→i (d, p, j) = min ψj→i (d + 1, i, cj ).
cj 6=∗,i
Observe that because δi→j = 0 the right hand is the definition of Ad+1
j→i
as given in (20), so we conclude
bj→i (d, p, j) = Ad+1
j→i ,
as desired.
Now we will show the sixth identity of Table 4 holds. Namely, for
3 ≤ d ≤ K − 1, p ∈ ∂ V i \ j, and c ∈ (∂ V i \ j) ∪ {•}, we will show
bj→i (d, p, c) = Hj→i .
Plugging di = d, pi = p, and ci = c into (18) gives us
bj→i (d, p, c) =
min
dj ,pj ,cj s.t.
hij (d,p,c,dj ,pj ,cj )=1
ψj→i (dj , pj , cj ).
We now want to break the minimization into cases based on how we
satisfy hij (d, p, c, dj , pj , cj ) = 1. First observe that p 6= j and c 6= j
implies we must be in case (7), which implies pj 6= i and cj 6= i. Then
there are two possible cases for how to satisfy fj (dj , pj , cj ) = 1.
First we can choose (2) and let dj = pj = cj = ∗. Otherwise we
are in (4) so we have dj = d′ for some 2 ≤ d ≤ K, pj 6∈ {∗, •} and
cj 6∈ {∗, pj }. Combining these cases implies
′
ψ
(d
,
p
,
c
)
.
bj→i (d, p, c) = min ψj→i (∗, ∗, ∗), min
min
j→i
j
j
2≤d′ ≤K pj 6=∗,i,•
c
=
6
∗,i
j
pj 6=cj
Observe that the right hand side contains the definitions of Gj→i from
d′
(23) and Fj→i
from (22), so we have
d′
= Hj→i ,
(32)
bj→i (d, p, c) = min Gj→i , min
F
j→i
′
2≤d ≤K
19
as desired, where the last step is applying the definition of Hj→i from
(24).
Lemma 2. The relations in Tables 1-3 for updating A-H form messages are valid.
Proof. We will show how to derive one representative update equation,
namely
X
, (33)
Ad+1
−
H
Hk→j +min 0, min
Adj→i = δi→j +
k→j
k→j
k∈∂ V j\i
k∈∂j\i
for 3 ≤ d ≤ K −1, which is the first equation of Table 1. The remainder
of Table 1 is derived via similar methods, as are the relations in Tables
2 and 3.
Recall (21):
X
bk→j (d, i, cj ).
Adj→i = δi→j + min
cj 6=∗,i
k∈∂j\i
We can split this minimization into two cases, namely cj = • and
cj ∈ ∂j \ i, which allows us to write
X
bk→j (d, i, •),
Adj→i = δi→j + min
k∈∂j\i
X
bl→j (d, i, k)
min bk→j (d, i, k) +
k∈∂j\i
l∈∂j\{i,k}
We now replace each instance of bk→j with the appropriate value
according to the identities in Tables 4 and 6. First note that for
k ∈ ∂ U j \ i we have bk→j (d, i, k) = ∞ because d ≥ 3 and k ∈ U
implies hjk (d, i, k, dk , pk , ck ) = 0 since gjk (d, i, k, dk , pk , ck ) = 1 if and
only if dk = d + 1 and if dk = d + 1 > 1 then fk (dk , pk , ck ) = 0. Thus
minimizing over k ∈ ∂j \ i is equivalent to minimizing over k ∈ ∂ V j \ i
so we can write
X
bk→j (d, i, •),
Adj→i = δi→j + min
k∈∂j\i
X
bl→j (d, i, k) .
min bk→j (d, i, k) +
k∈∂ V j\i
l∈∂j\{i,k}
We now replace these particular messages via bk→j (d, i, k) = Ad+1
k→j
from (30) (which is the fifth line of Table 4) and bl→j (d, i, k) = Hl→j
20
and bk→j (d, i, •) = Hk→j from (32) (which is the sixth line of Table
4), so we have
X
X
H
Hk→j , min Ad+1
+
Adj→i = δi→j + min
l→j
k→j
k∈∂j\i
l∈∂j\{i,k}
k∈∂j\i
X
,
Hk→j + min 0, min Ad+1
= δi→j +
k→j − Hk→j
k∈∂j\i
k∈∂j\i
where the second line simply extracts the sum from the minimization.
This equation now matches (33).
We are now prepared to prove Theorem 2.
Proof of Theorem 2. Tables 4-6 directly provide the functions ρi required to compute max-marginals. This substitution is valid after any
number of iterations because it is valid from the original definitions
of the A-H form variables by Lemma 1 and Lemma 2 proves the A-H
variables are iterated by rules that follow directly from the iteration
rules for the original form variables.
5
Constructing solutions.
Because we have no guarantee that messages will converge, or that
if they converge they will converge to an optimal solution, we elect
to compute a feasible solution at every iteration of the algorithm and
return the best solution found over all iterations. In particular we run
our algorithm for a fixed number of iterations unless we converge to a
fixed point before this limit. Finding all these solutions comes at an
additional computational cost, but it gives consistently better solutions
than simply using the result at the end of the iterations.
To initialize our solution construction algorithm, let A = V be
the set of non-root nodes which we have not yet assigned to a path.
Nodes will be removed from this set as we determine their role in
our constructed solution. For each node u ∈ U , we will construct a
(possibly empty) path Pu , and our full solution will be the collection
of these paths. The first step is to compute the max-marginals for
the possible choices at node u, which are either “do not participate in
a path” (in which case Pu = ∅) or “start a path by choosing a child
from among available neighbors of u”. We want to maximize the maxmarginal among these choices. Since our variables are in Min-Sum
form we actually compute the negative log of the max-marginals and
choose the minimizing configuration. To choose a configuration at the
21
node u, we use the functions ρi from Definition 1 to compute
X
− log p̄u (∗, ∗, ∗) =
Hk→u + β
(34)
k∈∂u
− log p̄u (1, •, j) =
X
Hk→u + Aj→u − Hj→u ,
j ∈ ∂ V u ∩ A. (35)
k∈∂u
and find the minimum value. If the minimum is (34), we let Pu = ∅
and proceed to the next root node. If the minimum is (35) for some
neighbor j, we initialize our new path as Pu = {u, j} and remove j
from the set of available nodes by setting A = A \ {j}. Note that (35)
considers only neighbors which are also elements of A insuring that
Pu does not include any nodes used in a previous path. Further note
that we must have (u, j) ∈ E so this is a valid start to a path because
j ∈ ∂ V u and E contains no edges in V × U . We next proceed to the
node j to determine whether Pu should be extended or added to the
solution as is.
This process is essentially the same as that for u, involving choosing
the optimal value of the max-marginal at j, but we are now constrained
to configurations with parent u. This means that we are choosing
whether to end the path at j or continue to one of j’s neighbors in V ,
so we compute
X
− log p̄j (2, u, •) =
Hk→j + Bu→j − Hu→j
(36)
k∈∂j
− log p̄j (2, u, i) =
X
Hk→j + Bu→j − Hu→j + A3i→j − Hi→j ,
i ∈ ∂V j ∩ A
k∈∂j
(37)
and choose the minimum. If the minimum is (36) the path is terminated so we add Pu to our solution as is and proceed to the next root
node. Otherwise the minimum is achieved by a particular choice i,
so we set Pu = {u, j, i}, remove i from the set of available nodes by
setting A = A \ {i}, and proceed to determine whether Pu should be
extended or added to the solution as is. Note again that i ∈ A guarantees we have not previously selected i to participate in any other path.
To confirm that (j, i) ∈ E, note that the first line of Table 1 implies
A3i→j = ∞ if (j, i) 6∈ E, in which case (37) is infinite for that choice
of i. Such an i cannot be the chosen minimum value because (36) is
always finite as can be seen from the last two lines of Table 1 and the
first line of Table 6. Thus if we choose i to continue the path we must
have (j, i) ∈ E which guarantees this is a valid choice.
After the first two nodes of a path the process becomes more
generic, so let us suppose we have some Pu such that |Pu | = d ≥ 3
and the last two nodes of the path Pu are are p and j. We still need
22
to determine whether to extend the path Pu or add it to the solution
as is. If d = K then we know that we cannot extend the path, so we
add it to the solution and move on to the next root node. Otherwise,
we compute
X
d−1
− log p̄j (d, p, •) =
Hk→j + Bp→j
− Hp→j
(38)
k∈∂j
− log p̄j (d, p, i) =
X
d−1
Hk→j + Bp→j
− Hp→j + Ad+1
i→j − Hi→j ,
i ∈ ∂V j ∩ A
k∈∂j
(39)
and choose the minimum. If the minimum is (38) we leave Pu as is, add
it to the solution and move on to the next root node. If the minimum
is (39) for some i, then we append i to the end of Pu and remove
i from A. We now have a path of length d + 1 and can repeat the
above procedure until we terminate the path. As above, we restrict
our attention to nodes in A to insure no node is assigned to multiple
paths. Furthermore, for i ∈ ∂ V j ∩ A if (j, i) 6∈ E then Ad+1
i→j = ∞ but
(38) is always finite because Hk→j is finite for any pair of nodes k and
d−1
j and Bp→j
is finite because we are guaranteed to have (p, j) ∈ E by
the fact that these are the last two nodes of the partially constructed
path Pu . Thus if we choose to continue the path we do so along an
edge with the correct direction to a node that is not participating in
another already constructed path, and so the end result of this process
is a feasible collection of node-disjoint directed paths.
Notice, however, that the result depends on the order in which
we process roots, so in practice we often repeat the procedure several
times with different orders and choose the best solution among those
we find.
6
Alternative algorithms.
Our numerical results in section 7 below will compare our efficient
BP algorithm to three alternatives: a Greedy algorithm, an IP based
algorithm for KEP from [2], and a novel IP based algorithm. We now
describe these algorithms.
We propose the Greedy algorithm as a simple to implement approximation algorithm with good scaling behavior. It is greedy in the
sense of adding paths to the solution one at a time and adding the
longest possible path at each step, but when adding a path the search
for the longest path is exhaustive. Given an arbitrary ordering of root
notes, the Greedy algorithm proceeds as follows. Let A = V be the
set of non-root nodes that have not yet been added to a path. From
a root node u ∈ U , all possible directed paths with length at most K
23
consisting of u and a collection of nodes in A are explored and the
longest such path is chosen, with ties being broken arbitrarily. The
longest possible path is found by recursively looking for the longest
path from each neighbor, keeping track of the depth bound to insure
a path longer than K is not chosen. The chosen path is then added
to the solution and its nodes are removed from A. If the path consists
only of the node u it is discarded. The algorithm now proceeds to the
next root node in the given order.
This algorithm is guaranteed to generate a feasible collection of
directed paths, but the size of that collection is highly dependent on
the order in which we consider root nodes, so we repeat the process
for a large number of different orderings of the root nodes, choosing
the best solution. For the numerical results to follow we often used
about 200 randomly generated orders. In most cases, the number of
orderings is chosen so the running time of the Greedy algorithm is
roughly equal to the message passing algorithm (which also includes a
similar randomization step, as noted in section 5).
The KEP algorithm from [2] is based on a Traveling Salesman IP
formulation approach to the kidney exchange problem and can find
bounded length cycles as well as paths, which may be bounded or
unbounded in length. For our comparisons we enforce our path length
bound K and disallow all cycles.
Finally we have implemented a simple IP formulation of the problem based on the same description we used in the development of
the message passing algorithm. We call this the Parent-Child-Depth
(PCD) formulation. We let
△
∂ + i = {j ∈ ∂i : (i, j) ∈ E}
and
△
∂ − i = {j ∈ ∂i : (j, i) ∈ E}.
24
The formulation is as follows:
X
xi
maximize
(40)
i∈[n]
s.t. xi =
X
pji + p•i
∀i ∈ [n]
(41)
∀i ∈ [n]
(42)
∀i ∈ [n]
(43)
∀i ∈ [n]
(44)
∀i ∈ [n]
(45)
∀i ∈ [n]
∀i ∈ V
(46)
(47)
∀(j, i) ∈ E
∀(j, i) ∈ E
(48)
(49)
∀i ∈ [n], j ∈ ∂ − i ∩ ∂ + i
∀i ∈ [n]
(50)
(51)
∀i ∈ [n], j ∈ ∂ − i
(52)
j∈∂ − i
X
cij ≤ xi
j∈∂ + i
p•i ≤
X
cij
j∈∂ + i
di = d•i +
X
dji
j∈∂ − i
0 ≤ di ≤ K
d•i = p•i
p•i = 0
pji = cji
dji = pji (dj + 1)
pji + cij ≤ 1
xi , p•i ∈ {0, 1},
pji ∈ {0, 1},
di , d•i ∈ Z
dji ∈ Z
+
cij ∈ {0, 1}
∀i ∈ [n], j ∈ ∂ i
(53)
Proposition 2. There is a bijection between feasible solutions to the
IP problem (40)-(53) and the set M defined in (9) and an optimal
solution to the IP problem maximizes the objective function H(d, p, c)
defined in (11).
Remark. This proposition states that the IP problem (40)-(53) is in
fact the same problem we deal with elsewhere in this paper. Before
providing a formal proof, we provide some intuition for the formulation.
The variable xi acts as an indicator for whether the node i participates
in a path, with the various pji and p•i indicators representing each
possible choice of parent for node i (including the possibility of starting
a path). Similarly the cij variables represent the possible choices of
child for the node i. The integer variable di represents the depth of i
in the path. This depth variable is decomposed into variables dji and
d•i which will be equal to di if the parent of i is j or i starts a path,
respectively, and zero otherwise. Thus at most one of d•i and the dji
will be nonzero, representing the actual choice of parent implied by the
values of p•i and the pji . The various constraints serve to guarantee
that this interpretation of these variables is accurate and all feasibility
constraints for the feasible set M .
25
More specifically, constraint (41) insures that a node participates if
and only if it has a parent or starts a path. The remaining constraints
insure that each node has at most one parent and one child, and that
all variables agree locally with each other. Note (43), which requires
a root to have a child (so each path contains at least one edge). The
depth bound is enforced by requiring di ≤ K and insuring that di
accurately represents the depth of i. Note in particular (49), which
insures that depth increases from parent to child. This constraint is
quadratic.
Proof of Proposition 2. Given (d′ , p′ , c′ ) ∈ M , we can construct IP
variables (x, d, p, c) node-by-node. For each i ∈ [n] there are three
possible cases, represented by (2)-(4) because fi (d′i , p′i , c′i ) = 1 is guaranteed by (d′ , p′ , c′ ) ∈ M :
• If d′i = p′i = c′i = ∗, then the node i does not participate in a path,
so we let xi = p•i = di• = 0, for j ∈ ∂ − i we let pji = dji = 0,
and for j ∈ ∂ + i we let cij = 0.
• If p′i = •, then the node i starts a path, so we let xi = p•i = 1,
di = d•i = 1, and for each j ∈ ∂ − i we let pij = dij = 0. By (3)
we know i has a child c′i ∈ ∂ + i, so we can let cic′i = 1 and for
j ∈ ∂ + i \ c′i we let cij = 0. Note that we must also have d′i = 1,
so we have di = d′i .
• If p′i 6∈ {∗, •} then the node i has a parent p′i ∈ ∂ − i, so we let
xi = pp′i i = 1 and let p•i = 0 and for j ∈ ∂ − i \ p′i let pji = 0. By
(4) we have 2 ≤ d′i ≤ K, so we let di = dp′i i = d′i and let d•i = 0
and for j ∈ ∂ − i \ p′i let dji = 0. If c′i = • we let cij = 0 for all
j ∈ ∂ + i because i has no child. Otherwise we let cic′i = 1 and for
all j ∈ ∂ + i \ c′i let cij = 0.
With variables constructed in this way it is straightforward to check
that all of the constraints are satisfied, so we omit the details.
Now suppose we have a feasible set of IP variables (x, d, p, c). We
will construct (d′ , p′ , c′ ) as described in Section 2 and show that it is
an element of the set M . We proceed node-by-node. For i ∈ [n]:
• If xi = 0, let d′i = p′i = c′i = ∗.
• If xi = 1 and p•i = 1, let d′i = 1, p′i = •. By (43) there is some
j ∈ ∂ + i such that cij = 1. Let c′i = j.
• If xi = 1 and p•i = 0, by (41) there exists some j ∈ ∂ − i such that
pji = 1. Let p′i = j, and let d′i = di . If cik = 0 for all k ∈ ∂ + i, let
c′i = •. Otherwise (42) implies there is exactly one k ∈ ∂ + i such
that cik = 1, so we let c′i = k.
We will now show (d′ , p′ , c′ ) ∈ M .
For i ∈ [n] the three cases above correspond exactly to the cases
(2)-(4) to guarantee fi (d′i , p′i , c′i ) = 1. Note in particular that (47)
26
guarantees i ∈ U for any i with p′i = •. If p′i 6∈ {∗, •} then we have
i ∈ V because p′i = j if and only if pji = 1 for some j ∈ ∂ − i but by
assumption ∂ − i = ∅ for i ∈ U .
For i ∈ [n], j ∈ ∂i, we have three cases. If p′i = j then we know
pji = 1 and j ∈ ∂ − i, so (48) implies cji = 1 and thus c′j = 1. (50)
guarantees p′j 6= i and c′i 6= j. j ∈ ∂ − i guarantees (j, i) ∈ E and (49)
guarantees d′i = d′j + 1. The case p′j = i is similar. If p′i 6= j and p′j 6= i
(48) guarantees c′i 6= j and c′j 6= i. Thus we are always in one of the
cases (5)-(7), so gij (d′i , p′i , c′i , d′j , p′j , c′j ) = 1.
Thus we conclude (d′ , p′ , c′ ) ∈ M , so we have established a bijection
between the feasible set of the IP and the set M .
To see that the optimization is the same, we need only note that
′ ′
′
xi = 1 if and only if p′i 6= ∗. Thus byP(10) we have
P xi = ηi′(pi′, ci′, di )
for all i ∈ [n] and so by (11) we have i∈[n] xi = i∈[n] η(di , pi , ci ) =
H(d, p, c). In other words, both are simply measuring the number of
nodes that participate in a path.
Note that this formulation is polynomial in n, both in the number
of variables and in the number of constraints, and is therefore simple
to implement. As we will see below, for a number of instances, with
n = 1000 and K = 15, PCD produces optimal solutions in more than
50% of tested graphs (see Table 9), while KEP does not find optimal
solutions in the same allowed time.
7
Numerical results.
We compare the performance of our algorithm, which we call BP, to the
three alternative algorithms described in section 6. Note that the two
IP-based algorithms (KEP and PCD) allow us to solve the problem to
optimality given sufficient time, whereas the Greedy and BP algorithms
have no optimality guarantees.
When running the BP algorithm we use the value β = 0.01 for
the parameter defined in (13). Informal testing of different values of
β indicated that this value had, in some cases, a small advantage of
about 1% in the size of solutions over larger β values while smaller β
values resulted in significantly worse solutions. When this advantage
was not present the value of β had no noticeable effect on the solution
size.
For all cases except the three largest real networks we use 200 orders
of root nodes for the Greedy algorithm and 5 orders of root nodes for
each iteration of BP. For the largest graphs these are decreased to
50 and 2, respectively. In all instances BP is run for a maximum of
T = 50 iterations, terminating earlier only if it has converged to a
fixed point or exceeded a preset time limit. In most cases BP did not
27
Root%
10
10
10
20
20
20
25
25
25
c
2
3
4
2
3
4
2
3
4
Nodes in solution
KEP Greedy
BP
384.9
364.4
376.6
460.8
446.4
457.8
426.4
480.4
469.2
641.6
554.6
603.4
813.6
685.9
746.3
721.8
763.1
769.0
700.7
600.3
654.5
893.7
726.6
787.0
818.7
801.3
834.9
Running Time
KEP Greedy
3.0
0.6
16.2
0.9
63.5
0.9
3.3
0.9
63.5
1.3
63.3
1.8
2.9
1.0
56.4
1.7
63.9
2.3
(s)
BP
2.2
4.5
4.0
2.6
3.8
2.1
2.4
3.9
2.5
Table 7: Random graphs on n = 1000 nodes with maximum path
length K = 5 and edge probability c/n.
100 samples were run
for each row, with β = 0.01 for BP. Dark shaded KEP results
are cases where KEP was optimal in at least 99% of samples. BP results are labeled by finding more nodes than Greedy and
finding the most nodes of any method .
consistently converge, though it did converge somewhat more often for
smaller graphs with n = 1000 and for longer allowed paths with K =
15. We have not seen a significant correlation between instances where
BP converges and those where it finds significantly better solutions,
but convergence does improve running time simply by allowing the
algorithm to terminate earlier.
7.1
Random graph comparisons.
We report our results on random graphs constructed as follows. We
let each ordered pair of nodes in V × V and U × V have an edge
independently with probability p. We set p = c/n and vary c. Thus
for every pair (i, j) ∈ U ∪ V × V the directed edge (i, j) is present with
probability c/n and absent with probability 1 − c/n.
Table 1 shows results for some varieties of random graphs. The column “Root%” indicates the percentage of nodes which are in the set
U , namely |U |/n. The column “c” indicates the value of the parameter
c for the graphs tested. The remaining columns summarize the results
of testing the algorithms on 100 random graphs, first showing the average number of nodes in the returned solution and then showing the
average running time in seconds. For the IP based algorithms a time
limit of 60 seconds was implemented by the IP solver and if the optimal solution was not found within that time the best feasible solution
28
identified by the solver was used. While in many cases KEP returned
a nontrivial feasible solution, if PCD did not find the optimal solution
it almost always returned the trivial feasible solution zero. For the
cases represented in Table 7 PCD never returned a nontrivial feasible
solution so we omit PCD from the table. It does, however, appear in
other cases below and performs well in some ranges outside those that
appear in tables, as will be discussed later.
Testing over a wide range of parameter values we find that in many
instances one or both of the IP algorithms runs to termination in time
comparable to BP or Greedy. In these cases Greedy and BP both
generally return suboptimal solutions, with both algorithms having
ranges of parameters where they enjoy a small advantage over the
other. This advantage ranges from Greedy finding about 5% more
nodes than BP to BP finding around 10% more nodes than Greedy.
Table 7 shows results for random graphs with 1000 nodes and a variety
of percentages of root nodes. The algorithms were run with path length
bound K = 5 and edge probability c/n for c = 2, 3, 4. For c = 2 KEP
is optimal in every sample and has a running time comparable to BP.
BP finds feasible solutions with 93%-98% of the nodes of the optimal
solution, whereas the Greedy solutions have 85%-95%. As we increase
c, KEP finds fewer optimal solutions, though it is still optimal in all
but 1 of our 100 samples when 10% of the nodes of the graph are roots
for c = 3. In most cases BP finds more nodes than Greedy does while
falling short of even the suboptimal KEP solutions, but for c = 4 with
20% and 25% root nodes, BP performs the best among the compared
algorithms, finding the most nodes in time comparable to Greedy and
significantly shorter than KEP.
While we do not show full results here, we can provide some sense
of the behavior of the algorithms for graphs just outside the regimes
in Table 7. If the root percentage is decreased below 10% KEP is
generally optimal and both BP and Greedy often find this optimal
solution, likely because the scarcity of roots leads to less interaction
between paths in feasible solutions. If the root percentage is increased
above 25% BP still finds more nodes than Greedy but significantly
fewer than KEP, which generally finds optimal solutions. Though PCD
did not find nontrivial solutions for the cases in Table 7, it is often
optimal for these cases with many roots. If the root percentage is kept
in the 10%-25% range but c is decreased to 1, both KEP and PCD are
always optimal, with Greedy missing a small percentage of nodes and
BP slightly behind Greedy. If c is increased above 4, both Greedy and
BP find significantly (10%-20%) more nodes than KEP, which is never
optimal, and Greedy usually has a small (1%-3%) advantage over BP.
In this regime PCD finds no nontrivial solutions.
Table 8 shows tests over some of the same parameters as Table 7
on larger graphs, with 10,000 nodes. For these larger graphs the time
29
c
2
3
4
Nodes in solution
KEP
Greedy
BP
6426.0 5371.5
5973.7
6792.8 6640.2
7332.5
7199.9 7413.1
7492.0
Running Time
KEP Greedy
150.3
39.3
1229.4
60.0
1250.4
87.3
(s)
BP
42.6
65.8
48.6
Table 8: Random graphs on n = 10000 nodes with 20% root nodes, maximum path length K = 5, and edge probability c/n. 100 samples were run
for each row, with β = 0.01 for BP. Dark shaded KEP results were optimal
in all samples. BP results are labeled by finding more nodes than Greedy
and finding the most nodes of any method .
limit for KEP and PCD is increased to 20 minutes, and again PCD
is omitted because it never finds nontrivial solutions within the time
limit. Most general trends remain the same, with KEP always finding
optimal solutions within the time limit for c = 2 but BP performing
better for c = 3 and c = 4. Specifically, BP now has an advantage over
KEP for c = 3. Another factor to note is that BP has a significant
advantage over Greedy in running time in the c = 4 case, which we
can credit to the fact that BP converged in all samples for that case,
running an average of 19 iterations as compared to 50 for the other
cases.
In Table 9 we provide summary results for problems with longer
allowed paths; that is with K = 10 and K = 15. Here KEP is almost
never optimal, though it does often provide a feasible solution with
more nodes than BP or Greedy, albeit at the cost of running to the
cutoff time rather than terminating in a few seconds. In all cases
for this regime BP finds more solutions than Greedy while running in
comparable time, and in two of these cases it also on average finds more
nodes than any other tested algorithm. In some cases the solution size
advantage over Greedy is significant, with BP solutions more than 10%
larger than Greedy solutions. Note that PCD performs rather well for
K = 15, often finding the optimal solution, though the average size of
the solution is reduced by the fact that PCD rarely returns a nontrivial
feasible solution, generally either returning the optimal solution or an
empty solution. PCD found the optimal solution in at least 50% of
samples for all the K = 15 cases, while KEP only found any for c = 2,
for which it was optimal in only 6% of cases. Thus there is an advantage
to PCD over KEP for all K = 15 cases as it much more regularly finds
the optimal solution. When it does find the optimal solution it is also
relatively fast, with an average runtime of 23 seconds for cases across
all 4 values of c in which PCD found the optimal solution. BP and
30
K
10
10
10
10
15
15
15
15
c
2
3
4
5
2
3
4
5
KEP
656.1
784.3
830.2
841.2
686.3
868.0
921.0
938.0
Nodes in solution
PCD Greedy
418.3
571.3
9.0
715.9
0
796.5
0
845.5
709.9
584.9
779.7
728.6
760.2
806.6
693.8
855.0
BP
671.7
758.2
812.6
852.6
684.7
789.7
835.6
869.6
Running Time (s)
KEP PCD Greedy
63.0 59.9
1.4
67.8 60.2
1.8
69.1 60.2
2.3
72.6 60.2
2.8
63.1 10.0
1.3
68.4 34.1
2.0
69.3 40.1
2.7
72.2 47.4
3.2
BP
3.7
3.3
2.6
2.7
3.6
3.7
3.0
3.1
Table 9: Random graphs on n = 1000 nodes with 15% root nodes, maximum
path length K, and edge probability c/n. 100 samples were run for each
row, with β = 0.01 for BP. Shaded PCD results were optimal in 96% of
samples, and in the other K = 15 cases PCD was optimal in at least 50%
of samples. BP results are labeled by finding more nodes than Greedy and
finding the most nodes of any method .
Greedy both remain significantly faster, and BP does on average find
better solutions than either PCD or Greedy.
For similar graphs just outside the regime presented in Table 9
we have a couple situations. As in the case of a shorter path bound,
decreasing c to 1 makes KEP and PCD optimal in all cases, with
Greedy missing a small percentage of nodes and BP slightly behind
Greedy. Increasing c above 5 leads to Greedy and BP having essentially
identical performance, both generally better than KEP and PCD. PCD
still finds optimal solutions in a reasonable percentage of K = 15 cases,
however. If the path length bound is kept as long as 15 or increased up
to 25, PCD performs very well over a wide variety of c values and root
percentages from 15% to 30%, generally finding the optimal solution.
In many of these cases KEP also finds an optimal solution but does
not terminate and confirm that the solution is optimal so PCD has a
significant advantage in running time in those cases. BP and Greedy
are suboptimal in these high-K cases, both finding 80%-90% of the
nodes in the optimal solution.
7.2
Computational results for real world networks.
We compared the performance of KEP, PCD, Greedy, and BP algorithms on several real world networks taken from the Stanford Large
Network Dataset collection [7]. We chose directed graphs and randomly selected a subset of nodes to act as roots (discarding directed
31
edges pointing in to roots and discarding isolated nodes). We present
performance comparisons for five networks, which are as follows:
• Epinions: this network is drawn from the website Epinions.com
and represents “trust” relationships between users, in which each
user can indicate whether they trust the opinions of each other
user. Thus an edge from i to j indicates that user i trusts user j.
• Gnutella: a snapshot of the Gnutella peer-to-peer file sharing
network, with nodes representing hosts and edges representing
the connections between those hosts.
• Slashdot: users of the technology news site Slashdot were allowed
to tag other users as friends or foes; our network has users as
nodes and tags as directed edges, drawn from a snapshot in 2008.
• Wiki-vote: A record of votes on administration positions for
Wikipedia, in which administrators vote on the promotion of
other potential administrators. Nodes are wikipedia users and a
directed edge from node i to node j represents that user i voted
on user j.
• Amazon: product co-purchasing on Amazon.com, based on the
“customers who bought this item also bought feature”. A directed edge from product i to product j indicates that j appears
on the page for i as a frequently co-purchased item. The data is
a snapshot from March 2003.
The results of our tests on the full graphs are summarized in Table
10. The “Nodes” column indicates the size of the given graph, which
is given as an average over the five tests because the selection of root
nodes results in a different number of nodes becoming isolated and
being dropped in each test. The column “∆” is the average degree of
the graph, which is also averaged across the choices of root nodes for
each trial. Remaining columns show the number of nodes in the average
solution and the running time in minutes for each algorithm. With the
exception of the graph Gnutella, KEP in all cases crashed rather than
running to the stopping time and so provided no solution, which we
indicate by “-”. We believe these crashes resulted from KEP exceeding
the available memory. They also caused the running times for KEP to
be erratic, with the algorithm crashing before the time limit in most
cases. The extreme running time for Amazon, the largest graph, can
be explained by the external software we used failing to adhere to
the requested time limits before crashing. In all cases PCD failed to
return a nontrivial solution, so it is omitted from the table. For all
graphs except Slashdot, BP enjoys an advantage in solution size over
Greedy, finding 4%-15% more nodes while running only slightly longer
on average. On Slashdot Greedy has an advantage in both solution
size and running time. We also note that even on the graph Gnutella,
32
Graph
Epinions
Gnutella
Slashdot
Wiki-vote
Amazon
Graph
Epinions
Gnutella
Slashdot
Wiki-vote
Amazon
Nodes
68507
5518
74893
6439
260982
Nodes
68507
5518
74893
6439
260982
∆
6.0
3.0
9.7
12.9
3.8
Nodes in solution
KEP
Greedy
BP
21,219.4
24,347.4
1,889.6
1,610.8
1,708.2
36,770.8
31,266.8
2,045.0
2,173.0
174,035.8 180,902.6
∆
6.0
3.0
9.7
12.9
3.8
Running Time (min)
KEP
Greedy
BP
56.5
9.8
17.9
150.3
0.2
0.4
58.9
5.1
20.5
75.5
0.8
1.5
1,042.8
182.0
259.9
Table 10: Real world networks with 20% root nodes (randomly selected) and
maximum path length K = 5. We use β = 0.01 for BP. ∆ is the average
degree of the sampled graph. 5 samples were run for each graph, with −
representing graphs for which KEP crashed without returning any solution.
Shaded cells represent the best solutions, with lighter shading for the case
when BP provides a better solution than Greedy.
33
the one case where KEP returned feasible solutions, BP has a huge
advantage in running time for a solution with 90% of the nodes of that
returned by KEP.
References
[1] F. Altarelli, A. Braunstein, L. Dall’Asta, C. De Bacco, and
S. Franz. The edge-disjoint path problem on random graphs by
message-passing. ArXiv e-prints, March 2015.
[2] Ross Anderson, Itai Ashlagi, David Gamarnik, and Alvin E. Roth.
Finding long chains in kidney exchange using the traveling salesman problem. Proceedings of the National Academy of Sciences,
112(3):663–668, 2015.
[3] M. Bailly-Bechet, C. Borgs, A. Braunstein, J. Chayes,
A. Dagkessamanskaia, J.-M. Franois, and R. Zecchina. Finding undetected protein associations in cell signaling by belief propagation.
Proceedings of the National Academy of Sciences, 108(2):882–887,
2011.
[4] M. Bayati, C. Borgs, A. Braunstein, J. Chayes, A. Ramezanpour,
and R. Zecchina. Statistical mechanics of steiner trees. Phys. Rev.
Lett., 101:037208, Jul 2008.
[5] P. Eschenfeldt, D. Schmidt, S. Draper, and J. Yedidia. Proactive
Message Passing on Memory Factor Networks. ArXiv e-prints, January 2016. Submitted to Journal of Machine Learning Research.
[6] D Koller and N Friedman. Probabilistic Graphical Models: Principles and Techniques. MIT Press, 2009.
[7] Jure Leskovec and Andrej Krevl. SNAP Datasets: Stanford large
network dataset collection. http://snap.stanford.edu/data,
June 2014.
[8] Jonathan S. Yedidia, William T. Freeman, and Yair Weiss. Exploring artificial intelligence in the new millennium. chapter Understanding Belief Propagation and Its Generalizations, pages 239–
269. Morgan Kaufmann Publishers Inc., San Francisco, CA, USA,
2003.
34
| 8cs.DS
|
Vehicle Color Recognition using Convolutional
Neural Network
Reza Fuad Rachmadi∗ , I Ketut Eddy Purnama† , and Mauridhi Hery Purnomo‡
arXiv:1510.07391v2 [cs.CV] 29 Oct 2017
Department of Multimedia and Networking Engineering
Institut Teknologi Sepuluh Nopember, Surabaya, Indonesia 60111
Email: ∗ [email protected], † [email protected], ‡ [email protected]
Abstract—Vehicle color information is one of the important
elements in ITS (Intelligent Traffic System). In this paper, we
present a vehicle color recognition method using convolutional
neural network (CNN). Naturally, CNN is designed to learn
classification method based on shape information, but we proved
that CNN can also learn classification based on color distribution.
In our method, we convert the input image to two different
color spaces, HSV and CIE Lab, and run it to some CNN
architecture. The training process follow procedure introduce by
Krizhevsky, that learning rate is decreasing by factor of 10 after
some iterations. To test our method, we use publicly vehicle color
recognition dataset provided by Chen. The results, our model
outperform the original system provide by Chen with 2% higher
overall accuracy.
I. I NTRODUCTION
Intelligent Transport System (ITS) is a system that manages
transportation from traffic management to law enforcement.
One important object that widely explored by ITS is a vehicle and their properties, including type, color, and license
plate. Vehicle color is an important property for vehicle
identification and provide visual cues for fast action law
enforcement. Recognize vehicle color is very challenging task
because several factors including weather condition, quality of
video/image acquisition, and strip combination of the vehicle.
The first factor, weather condition, may dramatically change
the color illumination of the acquisition image. For example,
if the image/video taken at haze condition then there a lot
of ”soft” white noise added to the image. Soft white noise
means that the noise is not random but continues and blended
with the foreground and background objects. The quality of
video/image acquisition is affected the final decision of the
vehicle color recognition system and its depends of the optical
sensor in the camera. Camera that can capture object at high
speed is recommended for ITS, but not all installed camera in
the road can do that. A lot of cameras installed in the road only
used to monitor the traffic, pedestrians, and street conditions.
The last factor is strip combination of the vehicle, which
is very affected to the vehicle recognition system. Region
selection is very important to tackle the problem.
There are some research paper published to tackle vehicle
color recognition problem, like in [1], [2], [6], [8], [9]. Chen et
al. [2] use feature context and linear SVM classifier to tackle
the problem. Feature context is a collection of histogram that
build with several areas, like spatial pyramid structure but with
different region configuration. In other paper [6], they try to
tackle vehicle color recognition problem using 2D histogram
with some ROI configuration as features and neural network
as classifier. Baek et al. [8] also use 2D histogram but without
ROI configuration and SVM as classifier. Another approach
is described by Son et al. [9] which using convolution kernel
to extract similarity between positive and negative images and
then feed up those similarity score to SVM classifier.
Color spaces are very important to color recognition applications, like vehicle color recognition. The selection of
color space will impact the recognition performance. The most
usable color space in digital photography is RGB color space,
but RGB color space has problem to color recognition because
channel of RGB color space contribute equal for each channel
so to distinct color is more difficult. Usually, researcher will
not use RGB as their primary color space and convert it to
other color spaces that separate illumination and color, like
CIE Lab or HSV [6], [8], [9]. Another approach is to make
2D histogram of two channels, like H and S channel in HSV
color space, and do classification using those 2D histogram.
In this paper, we present vehicle color recognition method
using convolutional neural network (CNN). CNN is type of
neural network but instead of using fully connected layer,
CNN use layer called convolution layer to extract features
from data. The training mechanism is very similar to normal
neural network and use stochastic gradient descent as training
algorithm. CNN is become very popular after winning the
ILSVRC (ImageNet Large Scale Visual Recognition Challenge) 2012 [5]. In those paper, they use more than 600,000
neuron and 7 hidden layer to provide good model of the
data. To avoid overfitting Krizhevsky et al. [5] employed
regularization method called dropout to the fully connected
layer [3]. The Krizhevsky model is huge and as reported in
the paper, the model trained in six day for 450,000 iteration in
GPU hardware. Before going into details, in section two we
describe detils related works in color recognition. Section two
describe details architecture of our CNN model. Section three
reports the experiments we have done and discuss the results.
II. R ELATED W ORKS
There are several research that try to tackle vehicle color
recognition problem including in [1], [2], [6], [8], [9]. The
newest research is describe by Chen et al. [2] in 2014 and
Fig. 1. The CNN architecture used in our system consist 8 layers with 2 base networks with total 16 layers. First two layers and fifth layer does normalization
and pooling after convolution process. The third and fourth layer does only convolution process. Before feed up to fully-connected layers, the networks do
channel concatenation process. Sample of input image is taken from Chen [2] dataset.
Hsieh et al. [1] in 2015. Chen et al. use feature context
(FC) with selected configuration to divide the images into
subregions, create histogram for each subregion, and learned
it using linear SVM. Not all value in histogram is used to
classify the vehicle color but the values clustered to form
codebook for the problem and then choose the codebook as
feature for the classifier. This mechanism know as Bag-ofWord (BoW) method. Chen et al. done preprocessing using
haze removal method [7] and color contrast normalization
method. The accuracy of system proposed by Chen et al. is
very high, over 92%.
Another paper by Hsieh et al. [1] proposed color correction
using background image and two frame image of car. Not only
color correction method, Hsieh et al. also proposed window
removal method that remove the window part of the car images
and classify vehicle color using lower part, like bumper and
doors, of the car. The window removal done by taking the
orientation of the car, fit the detail segmented car image by
ellipse shape and cut a half of the ellipse. Hsieh et al. done
the experiments using three different classifier, G-Classifier,
DC-Classifier, and DG-Classifier. G-Classifier responsible for
classify gray and non-gray color. The method is very simple
threshold method with assumption that for gray color the
avarage of three channel, RGB, is very close with color value
of each channel. The DC-Classifier and DG-Classifier trained
using SVM with features extracted from RGB and CIE Lab
color space. Red, green, blue, and yellow color class classified
using DC-Classifier and the rest of the color class classified
using DG-Classifier. From the experiments, Hsieh et al. report
that the average accuracy for the system is 93,59% with 7
color class including black, silver, white, yellow, red, green,
and blue.
FC also used by Dule et al. [6] to tackle vehicle color
recognition problem. The different between FC used by Chen
et al. and Dule et al. is that Dule et al. only used two ROI
(smooth hood peace and semi front vehicle). ROI is selected
automatically using plate detection method and otsu thresholding to search smooth hood peace and heuristic approach for
semi front vehicle. The classifier used by Dule et al. are KNN, ANN, and SVM. The best accuracy that reported in Dule
et al. paper is 83,5% with configuration of 8 bin histogram,
several combination of color spaces, and ANN classifier.
Other approach for vehicle color recognition problem is
classify vehicle color using 2D histogram features. Baek et
al. [8] proposed the vehicle color recognition system using
2D histogram features and SVM classifier. Hue and saturation
in HSV color space is used for creating the 2D histogram.
From the experiments, the average accuracy of the system is
94,92%. The dataset used in the experiment has 500 outdoor
vehicle images with five color class including black, white,
red, yellow, and blue color class.
Son et al. [9] proposed other possible approach for color
recognition using similirity method. The system using grid
kernel that run on hue and saturation channel of HSV color
space. The same dataset as in [8] is used in the experiments.
Son et al. reported only precission and recall for each color
class. The percentage of precission and recall from the experiments is very high and close to 100%. High precission and
high recall indicate that the model has good accuracy.
III. T HE CNN A RCHITECTURE
The architecture of our CNN can viewed in figure 1. Our
CNN architecture consists 2 base networks and 8 layers for
each base network with total 16 layers. The first two layers
of our CNN architecture is a convlutional layer and it does
convolution process following by normalization and pooling.
Convolutional layer is a layer that do convolution process that
same as convolution process in image processing algorithm.
For Ii is an input image and h is a some convolution kernel,
output image for convolution process Io can be written as
Io [m, n] =
∞
X
∞
X
Ii [i, j].h[m, n]
(1)
j=−∞ i=−∞
with [m, n] is pixel value at coordinate (m, n). Training process of CNN will learn h, may called as kernel, as parameters
of convolutional layer. The choice of activation function in
convolutional layer have huge impact for the networks. There
a several choice of activation function including tanh and
ReLU (Rectified Linear Unit). In our CNN networks we
use ReLU activation function for all layers including the
fully-connected layers. The normalization process done by
following equation 2 with α = 10−4 , β = 0.75, and n = 5.
β
i+n/2
α X
i
2
i
i
(kx,y )
(2)
lx,y = kx,y / 1 +
n
j=i−n/2
i
lx,y
i
with
is normalization result and kx,y
is output of layer
activation function for convolution at coordinate (x, y). Using
those normalization, the accuracy of CNN increase about 2%
according to [5]. The last process in two first layers is pooling
process. There are two type of pooling, max pooling and mean
pooling. Each type has different approach, max pooling will
take maximum respon from the convolutional process which is
shape with sharp edges and mean pooling will take the average
of the convolutional process respon which is summarize the
shape in neighborhood. In our CNN architecture, we use max
pooling with size 3x3 and stride 2 for overlapping pooling.
The second, fourth and fifth layer are grouping into two group
which each group is independent each others. The third and
fourth layer is also a convolutional layer but without pooling
and normalization process. Output of third and fourth layer is
same as input because we use 3x3 kernel with pad 1 added for
each border. The fifth layer is convolutional layer with only
pooling process without normalization.
Before going into a fully-connected layers, the pooling
output of the fifth layer from two base networks is concatenate
and flattened into one long vector. The sixth and seventh layer
is a fully-connected layer employed dropout regularization
method to reduce overfitting. The last layer is the softmax
regression layer which can describe in the following equation
T
p(y
(i)
(i)
eθ j x
= j|x ; θ) = Pk
l=1
(i)
T x(i)
eθl
(3)
with p(y (i) = j|x(i) ; θ) is probability of y (i) being class j
given input x(i) with weight parameter θ.
Overall, our CNN architecture consists 2 base networks,
8 layers each with total 16 layers. First layer use 11x11@3
kernel with total 48 kernels, second layer use 3x3@48 kernel
with total 128 kernels, third use 3x3@128 kernel with total
192 kernels, fourth layer use 3x3@192 kernel with total 192
kernels, and fifth layer use 3x3@192 with total 128 kernels.
Pooling process is employed in first, second, and fifth layer
with same parameter, pooling size of 3x3 with 2 pixel stride.
Fig. 2. Sample images from Chen dataset [2]. Some images are suffering
from noise and brightness constancy.
Sixth, seventh, and eight layers is fully-connected layers with
each 4096-4096-8 neuron with dropout regularization method
employed in sixth and seventh layer. The network’s input is
a 3 channel image with 150,228 dimensional or 227x227@3
resolution. Total neuron involved in the networks is 658,280
neurons.
IV. T HE E XPERIMENTS
A. Training Process
Our models trained using stochastic gradient descent with
115 examples per batch, momentum of 0.9 and weight decay
of 0.0005. For the experiments, we use Chen dataset [2] and
some sample images of the dataset can be viewed in figure
2. The dataset contains 15601 vehicle images with 8 classes
of vehicle color, which are black, blue, cyan, gray, green,
red, white, and yellow. In the training process, half of class
examples are used. Each example is resized into 256x256@3
resolution with certain color spaces. We use four different
color spaces, RGB, CIE Lab, CIE XYZ, and HSV. Before the
data processed for training, it cropped to 227x227@3 and subtracted by mean image of the training data. In training process
the data randomly mirrored to increase the classifier accuracy.
We use learning rate of 0.01 and reduced continuously by
a factor of 10 at multiple iteration of 50,000 with maximum
iteration of 200,000. We use caffe framework [4] to implement
our models. The weights of the networks are initialized using
a gaussian function with δ = 0.01 for connecting weights and
fixed value of 0.1 for bias value.
The stochastic gradient descent method, SGD for short, is an
optimization method that want to find minimum or maximum
value of some function. SGD will work for all function that
have gradient or first derivative. Usually the system use SGD
for minimizing the error or loss function and update the weight
parameters based on following function
wi+1 = wi − α∇L(z, wi )
(4)
TABLE I
ACCURACY OF OUR MODELS WITH 4 DIFFERENT COLOR SPACES AND
ACCURACY FROM C HEN ET AL . [2] FOR COMPARATION .
Color Space
Color Class
Chen et
al. [2]
RGB
HSV
CIE Lab
CIE XYZ
yellow
0.9794
0.9450
0.9656
0.9828
0.9553
white
0.9666
0.9624
0.9561
0.9649
0.9423
blue
0.9410
0.9576
0.9410
0.9484
0.9535
cyan
0.9645
0.9716
0.9645
0.9716
0.9787
red
0.9897
0.9866
0.9897
0.9886
0.9878
gray
0.8608
0.8503
0.8668
0.8647
0.8466
black
0.9738
0.9703
0.9703
0.9709
0.9730
green
0.8257
0.8215
0.8215
0.7676
0.7884
average
0.9447
0.9372
0.9414
0.9432
0.9282
Fig. 3. Confusion matrix from our model with RGB color space. Each cell
describe accuracy for each class measure in percentage.
TABLE II
E XECUTION TIME FOR OUR MODEL USING CPU AND GPU HARDWARE .
CPU
(1 core)
GPU
(448 cores)
Initialization time
4.849 s
4.849 s
Execution time
3.248 s
0.156 s
with wi is current weight parameters, α is learning rate, and
∇L(z, wi ) is the gradient of loss function L with respect to
input examples z. For faster model convergence, the weight
decay and momentum are added to the update equation. The
final equation of update function in SGD method is describe
following
∆wi+1 = γ∆wi + (1 − γ)(−α∇L(z, wi )
wi+1 = wi − α∇L(z, wi ) − αζwi
(5)
(6)
with γ is momentum variable and ζ is weight decay. Changing
momentum and weight decay may accelerate the training
process.
The training process done in GPU hardware to reduce the
training time. Our GPU hardware consists 14 multiprocessor
within 3 GB memory. There are two limitations of our GPU
hardware for training process, the memory limiting the size of
the networks or the batch size used in the training process and
the maximum dimension of the grid block execution (parallel
execution configuration) also limiting the batch size used in
the training process. The training process taken over 2 GB
GPU memory for the data and the networks with 4 days of
execution time.
B. Results and Discussion
For testing purpose, we use 50% examples of dataset that
not used in the training process. Table I summarize our testing
results with four different color spaces and compare the results
with the system provide by Chen et al. [2]. Each class consists
different number of examples from 141 to 2371. From table I,
it can see that RGB color space achieve the highest accuracy
of the testing process with average accuracy of 94,47%. Four
color spaces used in the models have high accuracy, more than
90%, with narrow deviation. The results show that our CNN
model outperform the original system of dataset provide by
Chen et al. [2]. Our models outperform Chen et al. system in
yellow, white, blue, red, gray, black, and green color class.
Only on cyan color class our system had lower accuracy
comparing to the Chen et al. system but with different only
0.7%. Figure 3 is a confusion matrix for our model using RGB
color space. The confusion matrix shows that the most worst
accuracy of our model is in green and gray color class. Some
examples of green class is misclassified as gray class and its
above 10%. As seen in the dataset, some green color class
examples has color that very close to gray, more like greengray color than green, so the classifier may have it wrong
classified as a gray color class. The same case appears in gray
color class which some gray color class examples misclassified
as white color class. Thus may appears because of very bright
sunlight reflection on metallic paint or the color is too light
so it’s very close to another color as well.
Another issue to tackle is execution time used to classify
vehicle color. We implement the models using two different
hardware, the first one the model running on 1 core CPU and
the second one the model running on 448 cores GPU with
NVIDIA Tesla C2050. Table II summarize average execution
time for all testing examples. As shown in table II, the models
that run on GPU have more than 20x faster than the models
that run on CPU, so the issue of execution time is solved
if the models running on appropriate hardware configuration.
The initialization time is a time for the system to prepare
the model, load it to the memory, and load mean image. For
(a) Kernel from network 1
(b) Kernel from network 2
(c) Output example from pooling process
Fig. 4. 96 kernel with 11x11@3 resolution learned by our first convolutional layer with input resolution 224x224@3 and output example from pooling
process. (a) 48 kernel from network 1, (b) 48 kernel from network 2, (c) output from pooling process in layer conv1 and conv2.
the practical implementation, we recommend using the client
server mechanism, send the vehicle detection result to the
server, do the vehicle color classification in server backend
using GPU hardware, and send back the result to the Intelligent
Transportation System for further processing.
To see how our models capturing color information in the
data, we visualize a several layer of our CNN models. The
first convolutional layer is an important part of the network
to extract low-level features. Figure 4 is a visualization of all
kernels in the first convolutional layer and an example output
of the pooling process in layer conv1 and conv2 of our CNN
architecture. As seen in figure 4, the first convolutional layer
capture rich color features in the input image. All vehicle color
variations in dataset are present in the kernels. The kernels
from network 1, figure 4a, capture a lot of cyan-like color.
Cyan-like color that appears in the kernel may contribute to
the red color class or cyan color class. Another color that
appears repeatedly in the kernel are red-blue color, greengray color, and orange-like color. For further investigation,
we capture respond from convolutional layer continuing with
normalization and pooling process and it can see in figure 4c.
We test our models using one of the test images and try to
analyze the behaviour of our models. Figure 4c show that for
yellow color class a lot of the green-like color kernel neuron is
active and its looks like our models learned that color can be
recognize from the hood color or the top color of the car. This
behaviour occurs because most all of the images in dataset take
image of the front of the car from some height and a little
deviation of angle, so the side of the car is not cover very
much. The camera configuration of taken images in dataset
simulate the CCTV or other street camera that relatively used
such that configuration.
V. C ONCLUSION
In the paper, we present the vehicle color recognition system
using convolutional neural network. Our model succesfully
capturing vehicle color in very high accuracy, 94,47%, and
outperform the original system provide by Chen [2]. From
the experiment, the best accuracy is achieve using RGB color
space and this is contradictive with several papers that not
recomend RGB color space for color recognition and using
another color space like HSV or YUV. Execution time for our
models is about 3 s for CPU (1 core) and 0.156 s for GPU
(448 cores), although the execution time is slower than system
provide by Chen [2] but its still can be used for practical
implementation with several adjustment.
R EFERENCES
[1] J.-W. Hsieh, L.-C. Chen, S.-Y. Chen, D.-Y. Chen, S. Alghyaline, and
H.-F. Chiang, Vehicle Color Classification Under Different Lighting
Conditions Through Color Correction, IEEE Sensors Journal, 2015, Vol:
15, Issue: 2, pp: 971–983.
[2] P. Chen, X. Bai and W. Liu, Vehicle Color Recognition on an Urban
Road by Feature Context, IEEE Transactions on Intelligent Transportation
Systems (TITS), 2014, Issue: 99, pp: 1-7.
[3] N. Srivastava, G. E. Hinton, A. Krizhevsky, I. Sutskever, and R.
Salakhutdinov, Dropout: A Simple Way to Prevent Neural Networks from
Overfitting, Journal of Machine Learning Research 15 (2014), pp: 1929–
1958.
[4] Y. Jia, E. Shelhamer, J. Donahue, S. Karayev, J. Long, R. Girshick, S.
Guadarrama, and T. Darrell, Caffe: Convolutional Architecture for Fast
Feature Embedding, ACM International Conference on Multimedia 2014,
pp: 675–678.
[5] A. Krizhevsky, I. Sutskever, and G. E. Hinton, ImageNet Classification
with Deep Convolutional Neural Networks, NIPS 2012, pp: 1097–1105.
[6] E. Dule, M. Gokmen, M. S. Beratoglu, A convenient feature vector construction for vehicle color recognition, Proc. 11th WSEAS International
Conference on Neural Networks, Evolutionary Computing and Fuzzy
systems, pp: 250255, (2010)
[7] K. He, J. Sun, and X. Tang, Single image haze removal using dark channel
prior, in Proc. IEEE Conf. CVPR, 2009, pp: 19561963.
[8] N. Baek, S.-M. Park, K.-J. Kim, and S.-B. Park, Vehicle Color Classification Based on the Support Vector Machine Method, International
Conference on Intelligent Computing, ICIC 2007, pp: 1133–1139.
[9] J.-W. Son, S.-B. Park, and K.-J. Kim, A convolutional kernel method
for color recognition, International Conference on Advanced Language
Processing and Web Information Technology (ALPIT 2007), pp: 242-247.
| 1cs.CV
|
A
Towards a Compiler for Reals
arXiv:1410.0198v3 [cs.PL] 11 Mar 2016
EVA DARULOVA and VIKTOR KUNCAK1 , Ecole Polytechnique Federale de Lausanne
Numerical software, common in scientific computing or embedded systems, inevitably uses an approximation of the real arithmetic in which most algorithms are designed. In many domains, roundoff errors are
not the only source of inaccuracy and measurement and other input errors further increase the uncertainty
of the computed results. Adequate tools are needed to help users select suitable approximations, especially
for safety-critical applications.
We present the source-to-source compiler Rosa which takes as input a real-valued program with error
specifications and synthesizes code over an appropriate floating-point or fixed-point data type. The main
challenge of such a compiler is a fully automated, sound and yet accurate enough numerical error estimation.
We present a unified technique for floating-point and fixed-point arithmetic of various precisions which can
handle nonlinear arithmetic, determine closed- form symbolic invariants for unbounded loops and quantify
the effects of discontinuities on numerical errors. We evaluate Rosa on a number of benchmarks from scientific
computing and embedded systems and, comparing it to state-of-the-art in automated error estimation, show
it presents an interesting trade-off between accuracy and performance.
1. INTRODUCTION
Numerical software, common in scientific computing and embedded systems, inevitably
approximates the real arithmetic in which its algorithms are typically designed. In addition
to roundoff errors from finite-precision arithmetics such as floating-point or fixed-point, many
problem domains come with additional sources of imprecision, such as measurement and
truncation errors, increasing the uncertainty on the computed results. We need adequate tools
to help developers understand whether the computed values meet the accuracy requirements
and remain meaningful in the presence of the errors. This is particularly important for
safety-critical systems.
Today, however, accuracy in numerical computations is often an afterthought. We write
programs in a user-selected finite-precision data type and then, perhaps, try to verify that
it meets our expectations. This is problematic for several reasons. Firstly, it introduces a
mismatch between the real-valued algorithms we write on paper and the low-level implementation details of floating-point or fixed-point arithmetic. Secondly, finite-precision source
code semantics prevents the compiler from applying many optimizations (soundly), as, for
instance, associativity no longer holds. And lastly, numerical errors remain implicit if they
are only the result of some separate analysis tool.
We propose a different strategy. The programmer writes the program in a real-valued
specification language and makes numerical errors explicit in pre- and postconditions. It is
then up to our compiler to determine an appropriate data type which fulfills the specification
but is as efficient as possible and to generate the corresponding code. This is in particular
attractive for fixed-point arithmetic, often used in embedded systems, where the code
generation can be quite challenging as the programmer has to ensure (implicit) decimal
points are aligned correctly.
Clearly, one of the key challenges of such a compiler is to determine how close a finiteprecision representation is to its ideal implementation in real numbers. While techniques
exist which can handle linear operations successfully [Goubault and Putot 2011; Darulova
and Kuncak 2011], precise and sound error estimation remains difficult in the presence
of nonlinear arithmetic. Roundoff errors and error propagation depend on the ranges of
variables in complex and non-obvious ways; even determining these ranges precisely for
nonlinear code poses a challenge. Furthermore, due to numerical errors, the control flow
in the finite-precision implementation may diverge from the ideal real-valued one, taking
1 This
work is supported in part by the European Research Council (ERC) project “Implicit Programming”.
ACM Journal Name, Vol. V, No. N, Article A, Publication date: January YYYY.
A:2
E. Darulova and V. Kuncak
a different branch and producing a result that is far off the expected one. Quantifying
discontinuity errors is hard due to many correlations and nonlinearity but also due to lack
of smoothness or continuity of the underlying functions that arise in practice [Chaudhuri
et al. 2011]. In loops, roundoff errors grow, in general, unboundedly. Even if an iteration
bound is known, loop unrolling approaches scale poorly when applied to nonlinear code.
We have addressed these challenges and present here our results towards the goal of a
verifying compiler for real arithmetic. In particular, we present
— a real-valued implementation and specification language for numerical programs with
uncertainties; we define its semantics in terms of verification constraints that they induce.
— an approximation procedure for computing precise range bounds for nonlinear expressions
which combines SMT solving with interval arithmetic.
— an approach for sound and fully automatic error estimation for nonlinear expressions for
floating-point as well as fixed-point arithmetic of various precisions. We handle roundoff and
propagation errors separately with affine arithmetic and a first-order Taylor approximation,
respectively. While providing accuracy, this separation also allows us to provide the
programmer with useful information about the numerical stability of the computation.
— an extension of the error estimation to programs with simple loops, where we developed a
technique to express the total error as a function of the number of loop iterations.
— a sound and scalable technique to estimate discontinuity errors which crucially relies on
the use of a nonlinear SMT solver.
— an open-source implementation in a tool called Rosa which we evaluate on a number of
benchmarks from the scientific computing and embedded systems domain and compare to
state-of-the-art tools.
2. A COMPILER FOR REALS
We first introduce Rosa’s specification language and give a high-level overview of the technical
challenges and our solutions on a number of examples.
A Real-Valued Specification Language Rosa is a ‘verifying’ source-to-source compiler which
takes as input a program written in a real-valued non-executable specification language and
outputs appropriate finite-precision code. A program is written in a functional subset of the
Scala programming language and consists of a number of methods over the Real data type.
Figures 1,4 and 5 show three such example methods. Pre- and postconditions allow the user
to explicitly state possible errors on method inputs and maximum tolerable errors on the
output(s), respectively. Taking into account all uncertainties and their propagation, Rosa
chooses a data type from a range of floating-point and fixed-point precisions and emits the
corresponding implementation in the Scala programming language.
By writing programs in a real-valued source language, programmers can reason about the
correctness of the real-valued algorithm, and leave the low-level implementation details to
the automated sound analysis in Rosa. Besides this separation of concerns, the real-valued
semantics also serves as an unambiguous ideal baseline against which to compute errors. We
believe that making numerical errors explicit in pre- and postcondition attached directly to
the source code makes it less likely that these will be forgotten or overlooked. Finally, such
a language opens the door to sound compiler optimizations exploiting properties which are
valid over reals, but not necessarily over finite-precision - as long as the accuracy specification
is satisfied.
Compilation Algorithm If a full specification (pre- and postcondition) is present on a method,
Rosa analyzes the numerical computation and selects a suitable finite-precision data type
which fulfills this specification and synthesizes the corresponding code. The user can specify
which data types are acceptable from a range of floating-point and fixed-point precisions.
The order in which these possible data types are given is important. Rosa searches through
the data types, applying a static analysis for each, and tries to find the first in the list which
ACM Journal Name, Vol. V, No. N, Article A, Publication date: January YYYY.
Towards a Compiler for Reals
A:3
def sineWithError(x: Real): Real = {
require(x > -1.57079632679 && x < 1.57079632679 && x +/- 1e-11)
x - (x*x*x)/6.0 + (x*x*x*x*x)/120.0 - (x*x*x*x*x*x*x)/5040.0
} ensuring(res => res +/- 1.001e-11)
Fig. 1. Approximation of sine with a Taylor expansion
101
10-3
10-8
10-13
10-18
10-23
10-28
10-33
10-38
10-43
10-48
10-53
10-58
10-63
Accuracy
Running time for 1000 evaluations (ms)
102
100
10-1
16bit
32bit
Float
Double oubleDoubQlueadDouble
D
Fig. 2. Running times vs accuracy for different finite-precision data types
satisfies the specification. While this analysis is currently repeated for each data type, parts
of the computation can be shared and we plan to optimize the compilation process in the
future.
The selection and the order of data types may depend on resource limits. If, for example,
a floating-point unit is not available or would be very costly, the programmer can prioritize
fixed-point data types. In general, the larger the data type (in terms of bits) the more time
and memory the execution will take. We illustrate the performance vs. accuracy trade-off
for the sine function from Figure 1 in the graph in Figure 2, which shows the runtime and
accuracy and in particular the trade-off between the two for different data types. Rosa
helps programmers navigate this trade-off space by providing automated sound accuracy
estimates.
Rosa can also be used as an analysis tool. By providing one data type (without necessarily
a postcondition), Rosa will perform the analysis and report the results, consisting of a
real-valued range and maximum error of the result, to the user. These analysis result are
also reported during the regular compilation process as they may yield useful insights.
2.1. Example 1: Straight-line Nonlinear Computations
We illustrate this compilation process on the example in Figure 1, which shows the code of
a method which computes the sine function with a 7th order Taylor approximation. The
Real data type denotes an ideal real-valued variable without uncertainty or roundoff. The
require clause specifies the range of the input parameter x as well as an initial uncertainty of
1e-11, which may stem from previous computations or measurement imprecisions. We would
like to make sure that error on the result does not grow too large, so we constrain the error
to 1.001e-11. We leave the data type unconstrained, so that Rosa considers 8, 16 and 32bit
ACM Journal Name, Vol. V, No. N, Article A, Publication date: January YYYY.
A:4
E. Darulova and V. Kuncak
/*
@param x (x +/- 1.e-11)
@return ((-1.0003675439019308 <= result && result <= 1.0003675439019308 &&
(result +/- 3.8907801077969955e-09)))
*/
def sineWithError(x : Long): Long = {
require(-1.5707963268 <= x && x <= 1.5707963268)
val
val
val
val
val
val
val
val
val
val
val
val
val
val
val
val
val
val
val
_tmp1 = ((x * x) >> 31)
_tmp2 = ((_tmp1 * x) >> 30)
_tmp3 = ((_tmp2 << 30) / 1610612736l)
_tmp4 = ((x << 1) - _tmp3)
_tmp5 = ((x * x) >> 31)
_tmp6 = ((_tmp5 * x) >> 30)
_tmp7 = ((_tmp6 * x) >> 31)
_tmp8 = ((_tmp7 * x) >> 31)
_tmp9 = ((_tmp8 << 28) / 2013265920l)
_tmp10 = ((_tmp4 + _tmp9) >> 1)
_tmp11 = ((x * x) >> 31)
_tmp12 = ((_tmp11 * x) >> 30)
_tmp13 = ((_tmp12 * x) >> 31)
_tmp14 = ((_tmp13 * x) >> 31)
_tmp15 = ((_tmp14 * x) >> 30)
_tmp16 = ((_tmp15 * x) >> 31)
_tmp17 = ((_tmp16 << 23) / 1321205760l)
_tmp18 = (((_tmp10 << 1) - _tmp17) >> 1)
result = _tmp18
}
Fig. 3. Sine function implemented in fixed-point arithmetic
fixed-point arithmetic as well as single, double, double double and quad double floatingpoint precision by default (in this order). Rosa determines that 8 bit fixed-point precision
potentially overflows, 16 or 32 bit fixed-point arithmetic is not accurate enough (total error
of 2.54e-4 and 3.90e-9 respectively) and that single-precision floating-point arithmetic is
neither (error of 2.49e-7). For double floating-point precision, Rosa determines the error
of 1.000047e-11, so that it generates code over the Double data type. If we could accept a
somewhat larger error, say 5e-9, then Rosa determines that 32bit fixed-point arithmetic
is sufficient. We show the generated code in Figure 3. (All intermediate results fit into 32
bits, but as we need up to 64 bits to perform the arithmetic operations, we simulate the
arithmetic here with the 64 bit Long data type.) Note that the generated code retains part
of the precondition. In Scala, require expressions are actually runtime checks that throw
an exception if the given condition is violated. Our error analysis, and code generation as
well, is only valid if the range and error of the input is satisfied. We can only check the
ranges at runtime, and Rosa generates the constraints such that they check the actual range
(i.e. the ideal real-valued ranges ± all possible errors). Checking the result range with a
postcondition is not necessary, since Rosa has already proven those bounds correct.
In order to determine which data type is appropriate, Rosa performs a static analysis
which computes a sound estimate of the worst-case absolute error. Since roundoff errors and
error propagation depend on the ranges of (all intermediate) computations, Rosa needs to
compute these as accurately as possible. We developed a technique which combines interval
arithmetic [Moore 1966] with a nonlinear SMT solver, which provides accuracy as well as
automation (section 4). In addition, using an SMT solver allows Rosa to take into account
arbitrary additional constraints on the inputs, which, for example, neither interval nor
ACM Journal Name, Vol. V, No. N, Article A, Publication date: January YYYY.
Towards a Compiler for Reals
A:5
def sine(x: Real): Real = {
require(-20 <= x && x <= 20)
x - x*x*x/6 + x*x*x*x*x/120
}
def pendulum(t: Real, w: Real, n: LoopCounter): (Real, Real) = {
require(-2 <= t && t <= 2 && -5 <= w && w <= 5 &&
-2.01 <= ~t && ~t <= 2.01 && -5.01 <= ~w && ~w <= 5.01)
if (n <
val h:
val L:
val m:
val g:
val
val
val
val
val
val
100)
Real
Real
Real
Real
{
=
=
=
=
0.01
2.0
1.5
9.80665
k1t = w
k1w = -g/L * sine(t)
k2t = w + h/2*k1w
k2w = -g/L * sine(t + h/2*k1t)
tNew = t + h*k2t
wNew = w + h*k2w
pendulum5(tNew, wNew, n++)
} else {
(t, w)
}
}
Fig. 4. Simulation of a pendulum
affine arithmetic [de Figueiredo and Stolfi 2004] can. Rosa decomposes the total error into
roundoff errors and propagated initial errors and computes each differently (section 5). The
accumulation of roundoff errors is kept track off with affine arithmetic, while propagation
errors are soundly estimated with a first-order Taylor approximation. The latter is fully
automatically and accurately computed with again the help of a nonlinear SMT solver.
2.2. Example 2: Loops with Constant Ranges
In general, numerical errors in loops grow unboundedly and the state-of-the-art to compute
sound error bounds in complex code is by unrolling. It turns out, however, that our separation
of errors into roundoff and propagation errors allows us to express the error as a function of
the number of loop iterations. We have identified a class of loops for which we can derive a
closed-form expression of the loop error bounds. This expression, on one hand, constitutes
an inductive invariant, and, on the other hand, can be used to compute concrete error
bounds. While this approach is limited to loops where the variable ranges are bounded, our
experiments show that this approach can already analyze interesting loops that are out of
reach for current tools.
Figure 4 shows such an example: a Runge Kutta order 2 simulation of a pendulum. t and w
are the angle the pendulum forms with the vertical and the angular velocity respectively. We
approximate the sine function with its order 5 Taylor series polynomial. We focus on roundoff
errors between the system following the real-valued dynamics and the system following the
same dynamics but implemented in finite precision (we do not attempt to capture truncation
errors due to the numerical integration, nor due to the Taylor approximation of sine). After
100 iterations, for instance, Rosa determines that the error on the results is at most 8.82e-14
ACM Journal Name, Vol. V, No. N, Article A, Publication date: January YYYY.
A:6
E. Darulova and V. Kuncak
def jetApproxGoodFitWithError(x: Real, y: Real): Real = {
require(-5<=x && x<=5 && -5<=y && y<=5 && x +/- 0.001 && y +/- 0.001)
if (y < x) {
-0.317581 + 0.0563331*x + 0.0966019*x*x + 0.0132828*y +
0.0372319*x*y + 0.00204579*y*y
} else {
-0.330458 + 0.0478931*x + 0.154893*x*x + 0.0185116*y 0.0153842*x*y - 0.00204579*y*y
}
}
Fig. 5. Approximation of a complex embedded controller
and 1.97e-13 (for t and w respectively) when implemented in double-precision floating-point
arithmetic and 7.38e-7 and 1.65e-6 in 32 bit fixed-point arithmetic.
2.3. Example 3: Discontinuities
Embedded systems often use piece-wise approximations of more complex functions. In Figure 5
we show a possible piece-wise polynomial approximation of a fairly complex jet engine
controller. We obtained this approximation by fitting a polynomial to a sample of values of
the original function. Note that the resulting function is not continuous.
A precise constraint encoding the difference between the real-valued and finite-precision
computation, if they take different paths, features variables that are tightly correlated. This
makes it hard for SMT-solvers to cope with and makes linear approaches imprecise. We
explore the separation of errors idea in this scenario as well, to soundly estimate errors
due to conditional branches. We separate the real-valued difference from finite-precision
artifacts. The individual error components are easier to handle individually, yet preserve
enough accuracy.
In our example, the real-valued difference between the two branches is bounded by 0.0428
(making it arguably a reasonable approximation given the large possible range of the result).
However, this is not a sound estimate for the discontinuity error in the presence of roundoff
and initial errors (in our example 0.001). With Rosa, we can confirm that the discontinuity
error is bounded by 0.0450 with double floating-point precision, with all errors taken into
account.
3. PROBLEM DEFINITION
Clearly, error computation is the main technical challenge of our compiler for Reals. Before
we describe our solution in detail, we first precisely define the error computation problem
that Rosa is solving internally. Formally, an input program consists of one or more methods
given by the grammar in Figure 6. args denotes possibly multiple arguments and res can be
√
a tuple. The method body may consist of the standard arithmetic operators +, −, ∗, /, ,
as well as immutable variable declarations (val t1 = ...), functional calls and conditionals.
Note that square root is only supported for floating-point arithmetic. The specification
language is functional, so we represent loops as recursive functions (denoted L), where n
denotes the integer loop iteration counter. For loop-free code D, note that more complex
conditions on branches can be expressed with nesting.
Let us denote by P a real-valued function representing our program and by x its input.
Denote by P̃ the corresponding finite-precision implementation of the program, which has
the same syntax tree but with operations interpreted in finite-precision arithmetic. Let
x̃ denote the input to this finite-precision program. The technical challenge of Rosa is to
ACM Journal Name, Vol. V, No. N, Article A, Publication date: January YYYY.
Towards a Compiler for Reals
A:7
estimate the difference:
max |P (x) − P̃ (x̃)|
x,x̃
(1)
which denotes the absolute error of the result of the program. This error is crucial for
selecting an implementation data type.
The domains of x and x̃, over which this expression is to be evaluated, are given by the userprovided precondition in the require clause. It defines range bounds xi ∈ [ai , bi ], x̃i ∈ [ci , di ]
for each component of the possibly multivariate input, as well as absolute error bounds
on the inputs of the form xi +/- λi that define the relationship |x − x̃| ≤ λ, understood
component-wise. If no errors are given explicitly, we assume roundoff as the initial error. The
ability to specify initial errors in addition to roundoff is important for modular verification
where the errors of one method may feed to following ones. Additionally, the require clause
may specify further constraints on the inputs, such as x*x + y*y <= 20.0. Method calls are
handled either by inlining the postcondition or the whole method body.
Corresponding to the syntactic program is a real-valued mathematical expression which is
the input to our core error computation procedure. Concretely, the input consists of one
or several real-valued functions f : Rm → Rn over some inputs xi ∈ R, representing the
arithmetic expressions F. We denote by f and x the exact ideal real-valued function and
variables and by f˜ : Rm → Rn , x̃i ∈ R their actual finite-precision counter-parts. Note that
for our analysis all variables are real-valued; the finite-precision variable x̃ is considered as a
noisy versions of x. We perform the error computation with respect to some fixed target
precision in floating-point or fixed-point arithmetic; this choice gives error bounds for each
individual arithmetic operation.
When P consists of a nonlinear arithmetic expression alone (F), then Equation 1
reduces to bounding the absolute error on the result of evaluating f (x) in finite precision
arithmetic: maxx,x̃ |f (x) − f˜(x̃)| (section 5). When the body of P is a loop (L), then the
constraint reduces to computing the overall error after k-fold iteration f k of f , where
f corresponds to the loop body. We define for any function H: H 0 (x) = x, H k+1 (x) =
H(H k (x)). We are then interested in bounding (section 6):
max|f k (x) − f˜k (x̃)|
x,x̃
For code containing branches (grammar rule D), Equation 1 accounts also for the discontinuity error. For example, if we let f1 and f2 be the real-valued functions corresponding
to the if and the else branch respectively with the if condition c, then, if c(x) ∧ ¬c(x̃),
the discontinuity error is given by |f1 (x) − f˜2 (x̃)|, i.e., it accounts for the case where the
P
::=
A
S
L
D
B
::=
::=
::=
::=
::=
√
::= F + F | F - F | F * F | F / F | F
::= F ≤ F | F < F | F ≥ F | F > F
::= x | const
F
C
X
def mName(args): res = {
require(A1 ∧ . . . ∧ An )
( L | D | B )
}
C | x +/- const | S
S ∧ S | S ∨ S | ¬ S | C
if (n < const) mName(B, n + 1) else args
if (C) D else D | B
val x = F; B | F
| X
Fig. 6. Rosa’s input language
ACM Journal Name, Vol. V, No. N, Article A, Publication date: January YYYY.
A:8
E. Darulova and V. Kuncak
real computation takes the if-branch, and the finite-precision one takes the else branch.
The overall error on P from Equation 1 in this case must account for the maximum of
discontinuity errors between all pairs of paths, as well as propagation and roundoff errors
for each path (section 7).
A note on relative error Our technique soundly overestimates the absolute error of the
computation. A sound estimate of the relative error can be computed from this and from
the range of the result provided that the range does not include zero. Whenever this is the
case, Rosa also reports the relative error in addition to the absolute one.
3.1. Finite-precision Arithmetic
Rosa supports analysis and code generation for floating-point and fixed-point arithmetic
of different precisions. We assume IEEE754 floating-point semantics. We regard overflow
and underflow as errors and Rosa will report the possibility of these occurring as such.
We assume rounding to nearest, however as long as the roundoff error can be determined
from the range of possible values, our analysis can be straight-forwardly adapted. Code
generation currently supports standard single and double floating-point arithmetic, as well
as double-double and quad-double precisions implemented in software [Bailey et al. 2013].
Fixed-point arithmetic also represents a subset of the rationals, but unlike floating-point
arithmetic does not require specialized hardware. Instead, it is implemented with integer
operations only, which makes it attractive especially for resource-bound systems. The
consequence of this is, however, that the alignment of (implicit) decimal points has to be
performed manually with bit shift operations at compile time. For this, the global ranges at
each intermediate computation step have to be known, respectively have to be computed.
For more details please see Anta et al. [2010], whose fixed-point semantics we follow.
While the input and output language is a subset of Scala, the analysis is programming
language agnostic, providing the IEEE 754 standard is supported, and adapting Rosa to a
different back-end would be straight-forward.
4. COMPUTING RANGES ACCURATELY
The first step to accurately estimating roundoff and propagation errors is to have a procedure
to estimate ranges as tightly as possible. This is important as these errors directly depend
on the ranges of all, including intermediate, values and coarse range estimates may result
in inaccurate errors computed or make the analysis impossible due to spurious potential
runtime errors (division by zero, etc.).
4.1. Interval and Affine Arithmetic
Traditionally, guaranteed computations have been performed with interval arithmetic [Moore
1966]. Interval arithmetic computes a bounding interval for each basic operation as
x ◦ y = [min(x ◦ y), max(x ◦ y)]
◦ ∈ {+, −, ∗, /}
and analogously for square root. For longer computations, interval arithmetic introduces
over-approximations, as it cannot track correlations between variables (e.g. x − x 6= 0).
Affine arithmetic [de Figueiredo and Stolfi 2004] partially addresses this loss of correlation
by representing possible values of variables as affine forms
x̂ = x0 +
n
X
xi i
i=1
where x0 denotes the central value (of the represented interval) and each noise symbol i is
a formal variable denoting a deviation from this central value, intended to range over [−1, 1].
The maximum magnitude of each noise term is given by the corresponding xi . The range
ACM Journal Name, Vol. V, No. N, Article A, Publication date: January YYYY.
Towards a Compiler for Reals
A:9
def getRange(expr, precondition, precision, maxIterations):
z3.assertConstraint(precondition)
[aInit, bInit] = evalInterval(expr, precondition.ranges);
//lower bound
if z3.checkSat(expr < a + precision) == UNSAT
a = aInit
b = bInit
numIterations = 0
while (b-a) < precision ∧ numIterations < maxIterations
mid = a + (b - a) / 2
numIterations++
z3.checkSat(expr < mid) match
case SAT ⇒ b = mid
case UNSAT ⇒ a = mid
case Unknown ⇒ break
aNew = a
else
aNew = aInit
//upper bound symmetrically
bNew = ...
return: [aNew, bNew]
Fig. 7. Algorithm for computing the range of an expression
represented by an affine form is computed as
[x̂] = [x0 − rad(x̂), x0 + rad(x̂)],
rad(x̂) =
n
X
|xi |
i=1
Note that the sign of the xi s does not matter in isolation, it does, however, reflect the relative
dependence between values. E.g., take x = x0 + x1 1 , then
x − x = x0 + x1 1 − (x0 + x1 1 ) = x0 − x0 + x1 1 − x1 1 = 0
If we subtracted x0 = x0 − x1 1 instead, the resulting interval would have width 2 ∗ x1 and not
zero. Linear operations are performed term wise and are computed exactly, whereas nonlinear
ones need to be approximated. Affine arithmetic can thus track linear correlations, it is,
however, not generally better than interval arithmetic: e.g. x∗y, where x = [−5, 3], y = [−3, 1]
gives [−13, 15] in affine arithmetic and [−9, 15] in interval arithmetic.
4.2. Range Estimation using Satisfiability Modulo Theories (SMT) Solvers
While interval and affine arithmetic are reasonably fast for range estimation, they tend to
introduce over-approximations, especially if the input intervals are not sufficiently small. We
improve over them by combining interval arithmetic with a nonlinear SMT (Satisfiability
Modulo Theories) constraint solver to obtain automation and accuracy.
Figure 7 shows the algorithm for computing the lower bound of a range. The computation
for the upper bound is symmetric. For each range to be computed, our technique first
computes an initial sound estimate of the range with interval arithmetic. It then performs
an initial quick check to test whether the computed first approximation bounds are already
tight. If not, it uses the first approximation as the starting point and then narrows down
the lower and upper bounds using a binary search. At each step of the binary search our
tool uses the nonlinear nlsat solver within Z3 [De Moura and Bjørner 2008; Jovanović and
de Moura 2012] to confirm or reject the newly proposed bound. The search stops when either
ACM Journal Name, Vol. V, No. N, Article A, Publication date: January YYYY.
A:10
E. Darulova and V. Kuncak
Z3 fails, i.e. returns unknown for a query or cannot answer within a given timeout, the
difference between subsequent bounds is smaller than a precision threshold, or the maximum
number of iterations is reached. This stopping criterion can be adjusted by the user.
Additional Constraints In our approach, since we are using Z3 to check the soundness of
bounds, we can assert the additional constraints and perform all checks with respect to all
additional and initial constraints. This is especially useful when taking into account branch
conditions from conditionals.
Optimizations Calling an SMT solver is fairly expensive so we want to minimize the number
of calls. The algorithm Figure 7 presents several direct knobs to do this: the maximum
number of iterations and the precision of the range estimate. Through our experiments
we have identified suitable default values, which seem to present a good trade-off between
accuracy and performance. In addition to these two parameters, if we are only interested in
the final range, we do not need to call Z3 and the algorithm in Figure 7 for every intermediate
expression. In principle, we could call Z3 only on the full expression, however, we found
that this resulted in suboptimal results as this expression very often was too complex. We
found a good compromise in calling Z3 only every 10 arithmetic operations. All of these
parameters can be adjusted by the user.
5. SOUNDLY ESTIMATING NUMERICAL ERRORS IN NONLINEAR EXPRESSIONS
Now we address the first challenge of error estimation for a loop-free nonlinear function
without branches: |f (x) − f˜(x̃)| where |x − x̃| ≤ λ, f : Rm → Rn and where the ranges for x
and x̃ are given by the precondition.
5.1. Error Estimation with Affine Arithmetic
Consider first roundoff errors only that is, we are interested in |f (x) − f˜(x)|, where the
input errors are zero. Our procedure executes the computation abstractly by computing an
interval and an affine form for each AST node:
(range: Interval,
err
ˆ :
AffineForm)
represents the real-valued range, and err
ˆ the accumulated worst-case errors, with
essentially one noise term for each roundoff error (together with artifacts from nonlinear
approximations). The actual finite-precision range is then given by range + [err],
ˆ where [err]
ˆ
denotes the interval represented by the affine form. For each computation step, we compute
the
range
(1) new range with our range procedure from section 4
(2) propagation of already accumulated errors
(3) new roundoff error, which is then added to the propagated affine form.
Since we compute the range at each intermediate node, we also check for possible overflows,
underflows, division by zero or negative square root errors without extra effort.
Propagation of Errors with Affine Arithmetic For linear operations, errors are propagated
with the standard rules of affine arithmetic. For multiplication, division and square root the
error propagation depends on the range of values, so that we have to adapt our computation
to use the ranges computed with our Z3-backed procedure. In the following, we denote the
real range of a variable x by [x] and its associated error by the affine form err
ˆ x . When
we write [x] ∗ err
ˆ y we mean that the interval [x] is converted into an affine form and the
multiplication is performed in affine arithmetic. Multiplication is computed as
x ∗ y = ([x] + err
ˆ x )([y] + err
ˆ y)
= [x] ∗ [y] + [x] ∗ err
ˆ y + [y] ∗ err
ˆ x + err
ˆ x ∗ err
ˆy +ρ
ACM Journal Name, Vol. V, No. N, Article A, Publication date: January YYYY.
Towards a Compiler for Reals
A:11
where ρ is the new roundoff error. Thus the first term contributes to the ideal range and the
remaining three to the error affine form. The larger the factors [x] and [y] are, the larger the
finally computed errors will be so that a tight range estimation is important for accuracy.
Division is computed as
x
1
= x ∗ = ([x] + err
ˆ x )([1/y] + errˆ1/y )
y
y
1
1
= [x] ∗ [ ] + [x] ∗ err
ˆ y1 + [ ] ∗ err
ˆ x + err
ˆ x ∗ err
ˆ y1 + ρ
y
y
For square root, we first compute an affine approximation of square root as in our previous
work [Darulova and Kuncak 2011]:
√
x=α∗x+ζ +θ
and then perform the affine multiplication term wise.
Roundoff Error Computation Roundoff errors for floating-point arithmetic are computed at
each computation step as
ρ
=
δ
* maxAbs(totalRange)
where δ is the machine epsilon, and added to err
ˆ as a fresh noise term. Note that this
roundoff error computation makes our error computation parametric in the floating-point
precision. Since we regard (and report) subnormal numbers as errors, this error abstraction
is sound. For fixed-point arithmetic, roundoff errors are computed as
ρ
= getFormat(totalRange, bitWidth).quantizationError
where the getFormat function returns the best fixed-point format [Anta et al. 2010] that can
accommodate the range. This computation is parametric in the bit-width.
5.2. Separation of Errors
We could use the affine arithmetic based procedure to track all errors,not only roundoff
errors, by simply adding the initial error as a fresh noise term at the beginning. Such an
approach treats all errors equally: the initial errors are propagated in the same way as
roundoff errors which are committed during the computation. We found, however, that the
over-approximation introduced by affine arithmetic for nonlinear computations increases
substantially as the magnitude of the noise terms (i.e. the errors) becomes larger. Instead,
we separate the total error as follows:
|f (x) − f˜(x̃)| = |f (x) − f (x̃) + f (x̃) − f˜(x̃)|
(2)
≤ |f (x) − f (x̃)| + |f (x̃) − f˜(x̃)|
The first term, |f (x) − f (x̃)|, captures the error on the result of f caused by the initial
error between x and x̃. The second term, |f (x̃) − f˜(x̃)|, covers the roundoff error committed
when evaluating f in finite precision, but note that we can now compute this roundoff
error on the same input x̃. Thus, we separate the overall error into the propagation of
existing errors, and the newly committed roundoff errors. We denote by σf : Rm → Rn the
function which returns the roundoff error committed when evaluating an expression f in
finite-precision arithmetic: σf (x̃) = |f (x̃) − f˜(x̃)|. We omit the subscript f , when it is clear
from the context. Further, g : Rm → Rn denotes a function which bounds the difference in
f , given a difference in its inputs: |f (x) − f (y)| ≤ g(|x − y|). When m, n > 1, the absolute
values are component-wise, e.g. g(|x1 − y1 |, . . . , |xm − ym |), but when it is clear from the
context, we will write g(|x − y|) for clarity. Thus, the overall numerical error is given by:
|f (x) − f˜(x̃)| ≤ g(|x − x̃|) + σ(x̃)
ACM Journal Name, Vol. V, No. N, Article A, Publication date: January YYYY.
(3)
A:12
E. Darulova and V. Kuncak
One alternative to Equation 2 would be to bound the error by |f (x) − f˜(x)| +|f˜(x) − f˜(x̃)|.
The first term now corresponds to roundoff errors, but the second requires bounding the
difference of f˜ over a certain input interval. In the separation that we have chosen, we
need to compute the difference over the real-valued f . Note that f is a simpler function
than its finite-precision counterpart, and its analysis is reusable across different concrete
implementations.
The function σ is instantiated with the affine arithmetic based procedure from subsection 5.1. Since roundoff errors are local, we found affine arithmetic suitable for this purpose.
In contrast, the propagation of existing errors (function g) depends highly on the steepness
of the function, so we want to capture as much global information, such as correlations
between variables, as possible. This is only feasible when looking at the function as a whole.
5.3. Propagation Errors
We instantiate Equation 3 with g(x) = K · x, i.e. |f (x) − f (y)| ≤ K|x − y| which bounds
the deviation on the result due to a difference in the input by a linear function in the input
errors. The constant K (or vector of constants Ki in the case of a multivariate function) is to
be determined for each function f individually, and is usually called the Lipschitz constant.
We will also use the in this context more descriptive name propagation coefficient. Note that
we need to compute the propagation coefficient K for the mathematical function f and not
its finite-precision counterpart f˜.
Error amplification or diminution depends on the derivative of the function at the value
of the inputs. The steeper the function, i.e. the larger the derivative, the more the errors are
magnified. For f : Rm → R we have
|f (x) − f (x̃)| ≤
m
X
i=1
Ki λi ,
where Ki = sup
x,x̃
∂f
∂wi
(4)
where λi are the initial errors and wi denote the formal parameters of f . This computation
naturally extends component-wise to multiple outputs. Thus, the propagation coefficients
are computed as a sound bound on the Jacobian.
We formally derive the computation of the propagation coefficients Ki for a multivariate
function f : Rm → R in the following. Let h : [0, 1] → R such that h(θ) := f (y + θ(z − y)).
d
Without loss of generality, assume y < z. Then h(0) = f (y) and h(1) = f (z) and dθ
h(θ) =
∇f (y + θ(z − y)) · (z − y). By the mean value theorem: f (z) − f (y) = h(1) − h(0) =
h0 (ζ), where ζ ∈ [0, 1].
|f (z) − f (y)| = |h0 (ζ)| = |∇f (y + ζ(z − y)) · (z − y)|
∂f
∂f
=
,...,
· (z − y) , s = y + ζ(z − y)
∂w1 s
∂wm s
∂f
∂f
· (z1 − y1 ) + · · · +
· (zm − ym )
=
∂w1
∂wm
m
X
∂f
≤
· |zi − yi |
(**)
∂wi
i=1
where the partial derivatives are evaluated at s = y + ζ(z − y) (which we omit for readability).
The value of s in (**) is constraint to be in s ∈ [y, z], so for a sound analysis we have to
determine the maximum absolute value of the partial derivative over [y, z]. y and z in our
application range over the values of x and x̃ respectively, so we compute the maximum
ACM Journal Name, Vol. V, No. N, Article A, Publication date: January YYYY.
Towards a Compiler for Reals
absolute value of
∂f
∂xi
A:13
over all possible values of x and x̃. With |yi − zi | ≤ λi we obtain
|f (x) − f (x̃)| ≤
m
X
Ki λi ,
where Ki = sup
x,x̃
i=1
∂f
∂wi
Bounding Partial Derivatives We compute the partial derivatives symbolically. Recall that
√
the arithmetic operations permitted are {+, −, ∗, /, }, which leaves the possibility of
discontinuities and undefined expressions. We detect these automatically during the bound
computation, so we do not need to make or check any assumptions on the derivatives
up-front.
We need to soundly bound the partial derivatives over all possible values of x and x̃.
Both interval and affine arithmetic suffer from possibly large over-approximations due to
nonlinearity and loss of correlations. Furthermore, they cannot take additional constraints
into account, for example from branch conditions (e.g. y < x) or user defined constraints on
the inputs. We use the range computation from section 4 which allows us to take these into
account, making the ranges computed much tighter.
Sensitivity to Input Errors Beyond providing a way to compute the propagated initial errors,
Equation 4 also makes an upper bound on the sensitivity of the function to input errors
explicit. The user can use this knowledge, for example, to determine which inputs need to be
determined more precisely, e.g. by more precise measurements or by using a larger number
of iterations of a numerical algorithm to find them. We report the values of K back to the
user.
5.4. Relationship with Affine Arithmetic
Both our presented propagation procedure and propagation using affine arithmetic perform
approximations. The question arises then, when is it preferable to use one over the other?
Our experience and experiments show empirically that for longer nonlinear computations,
error propagation based on Lipschitz continuity gives better results, whereas for shorter
and linear computations this is not the case. In this section, we present an analysis of this
phenomenon based on a small example.
Suppose we want to compute x ∗ y − x2 . For this discussion we consider propagation only
and disregard roundoff errors. We consider the case where x and y have an initial error of δx 1
and δy 2 respectively, where i ∈ [−1, 1] are the formal noise symbols of affine arithmetic.
Without loss of generality, we assume δx , δy ≥ 0. We first derive the expression for the error
with affine arithmetic and take the definition of multiplication from subsection 5.1. We
denote by [x] the evaluation of the real-valued range of the variable x.
The total range of x is then the real-valued range plus the error: [x] + δx 1 , where
1 ∈ [−1, 1]. Multiplying out, and removing the [x][y] − [x]2 term (since it is not an error
term), we obtain the expression for the error of x ∗ y − x2 :
([y]δx 1 + [x]δy 2 + δx δy 3 ) − (2[x]δx 1 + δx δx 4 )
= ([y] − 2[x])δx 1 + [x]δy 2 + δx δy 3 + δx δx 4
(5)
3 and 4 are fresh noise symbols introduced by the nonlinear approximation. Now we
compute the propagation coefficients:
∂f
= y − 2x
∂x
∂f
=x
∂y
so that the error is given by
[y + δy 2 − 2(x + δx 1 )] δx + [x + δx 1 ] δy
ACM Journal Name, Vol. V, No. N, Article A, Publication date: January YYYY.
(6)
A:14
E. Darulova and V. Kuncak
We obtain this expression by instantiating Equation (**) with the range expressions of x
and y. Note that the ranges used in the evaluation of the partial derivatives include the
errors. Multiplying out Equation 6 we obtain:
[y − 2x] δx + [x] δx + δx δy + δx δx + δx δx
(7)
With affine arithmetic, we compute ranges for propagation at each computation step, i.e.
in Equation 5 we compute [x] and [y] separately. In contrast, with our new technique, the
range is computed once, taking all correlations into account between the variables x and y.
It is these correlations that improve the computed error bounds. For instance, if we choose
x ∈ [1, 5] and y ∈ [−1, 2] and we, say, know that x < y, then by a step-wise computation we
obtain [y] − 2[x] = [−1, 2] − 2[1, 5] = [−11, 0] whereas taking the correlations into account, we
can narrow down the range of x to [1, 2] and obtain [y −2x] = [−1, 2]−2[1, 2] = [−5, 0]. Hence,
since we compute the maximum absolute value of these ranges for the error computation,
affine arithmetic will use the factor 11, whereas our approach will use 5.
But, comparing Equation 7 with Equation 5, we also see that one term δx δx is included
twice with our approach, whereas in the affine propagation it is only included once. We
conclude that a Lipschitz-based error propagation is most useful for longer computations
where it can leverage correlations. In other cases, we keep the existing affine arithmetic-based
technique. It does not require a two-step computation, so we want to use it for smaller
expressions. We remark that for linear operations the two approaches are equivalent.
5.5. Implementation
We have implemented Rosa in the Scala programming language. Internally, we use a rational
data type implemented on top of Java’s BigIntegers for all our computations. This lets us
avoid having to deal with roundoff errors ourselves, and easily interface with Z3 which also
uses rationals.
5.6. Comparison with State-of-the-Art
We are aware of two other tools which can automatically quantify numerical errors: Fluctuat [Goubault and Putot 2011] and FPTaylor [Solovyev et al. 2015].
Fluctuat is an abstract interpreter which uses affine arithmetic for both the ranges
of variables and for the error computation. In order to combat the over-approximations
introduced by affine arithmetic, Fluctuat can add constraints on noise terms [Ghorbal et al.
2010]. Further, Fluctuat uses Taylor approximations locally to handle division and square
root [Ghorbal et al. 2009], but the expansion is hard coded and does not consider the global
expression
Another technique employed by Fluctuat is interval subdivision, where the user can
designate up to two variables in the program whose ranges will be subdivided, analyzed
separately and the results then merged. This procedure works for floating- point arithmetic
as the decimal point is dynamic, however, for fixed-point arithmetic the global ranges are
needed at each point. Naturally, interval subdivision increases the runtime of the analysis,
especially for multivariate functions, and the optimal subdivision strategy may not always be
obvious. Interval subdivision can also be straight-forwardly added to Rosa - at a performance
penalty. We choose here to compare our SMT-based technique against Fluctuat with and
without subdivision to obtain a good comparison between the techniques. In the future, we
expect a combination of different techniques to work best.
Fluctuat also has a procedure for computing discretization errors, and can handle loops
either by computing fixpoints, if such exist, or by unrolling. Finally, Fluctuat also separates
errors similarly to our presented approach, although it does not treat the different part
fundamentally differently as we do. We want to note that our formalism has also enabled
the unified treatment of loops and discontinuities.
ACM Journal Name, Vol. V, No. N, Article A, Publication date: January YYYY.
Towards a Compiler for Reals
A:15
FPTaylor [Solovyev et al. 2015] is a recent tool for computing the roundoff errors of
nonlinear expressions, including transcendental functions. It relies similarly to Rosa on
Taylor series, but does the expansion with respect to errors, whereas we expand with respect
to inputs. Furthermore, FPTaylor uses a global optimization as the backend, which enables
the use of transcendental functions (Z3’s nlsat solver only supports arithmetic). FPTaylor
currently only supports error computation for straight-line computations in floating-point
arithmetic.
Like Rosa, both Fluctuat and FPTaylor also compute relative errors from absolute errors,
whenever the resulting range does not straddle zero.
Another framework that can be used for estimating numerical errors is the Frama-C
framework [2015] with the Gappa front-end [Boldo and Marché 2011]. Gappa works internally
with interval arithmetic and works best on precise properties when the user can provide
hints [Solovyev et al. 2015]. In this paper we want to focus on automated error estimation,
thus we only compare our results against those from Fluctuat and FPTaylor.
5.7. Experimental Results
We have chosen a number of benchmarks from the domains of scientific computing and
embedded systems [Anta et al. 2010; Woodford and Phillips 2012] to evaluate the accuracy
and performance of our technique. The tool and all benchmarks are open-source and available
at https://github.com/malyzajko/rosa.
We perform all tests in double floating-point precision as this is the precision supported
by both Fluctuat and FPTaylor. Rosa is currently the only tool that also supports fixedpoint arithmetic. In our experience, while the absolute errors naturally change with varying
precisions and data types, relative differences when comparing different tools on the same
precision data type remain similar. Experiments were performed on a desktop computer
running Ubuntu 14.04.1 with a 3.5GHz i7 processor and 16GB of RAM, and using the
unstable branch (as of 10 December 2014) of Z3.
Table I shows our experimental results in terms of accuracy (absolute errors computed)
and performance (running time of tool). All running times have been rounded up. We
consider three flavors of our benchmarks: inputs with roundoff errors only, inputs with initial
larger uncertainties and inputs with an additional nonlinear constraint.
Inputs with Roundoff In the first set of benchmarks in Table I we assume only roundoff
as the initial error on inputs. We compare against Fluctuat without and with subdivisions.
For the subdivisions, we uniformly chose 20 subdivisions for the two inputs where the
effect was largest. While choosing more is certainly possible, we found the running time
increasing rapidly and disproportionately with the accuracy gains. For FPTaylor we used
default settings with the improved rounding model, approximate optimization and the
branch and bound optimization procedure, which we believe are the most accurate settings.
The annotation ’(ref)’ marks benchmarks that are refactored. Rosa’s technique in general
benefits from such a refactoring (see subsection 5.4), which is supported by the experiments.
For Fluctuat this is also sometimes the case when subdivisions are used.
FPTaylor is able to compute the tightest error bounds on these benchmarks, but we
observe that the difference (except for the jet example) are in many cases not very large.
FPTaylor’s computation is also the most time consuming in the majority of cases.
Finally, we would like to remark that subdivisions cannot be directly extended to fixedpoint arithmetic as the determination of fixed-point formats and thus roundoff errors requires
the knowledge of the global ranges, i.e. the ranges valid over all input and not only over the
subdivision.
Inputs with Uncertainties The second set of benchmarks features inputs with uniform
uncertainty of 1e-11, aiming to compare the different tools ability to estimate the error
propagation accurately. The tools’ settings are the same as for the first set of testcases.
ACM Journal Name, Vol. V, No. N, Article A, Publication date: January YYYY.
A:16
E. Darulova and V. Kuncak
Table I. Absolute errors computed by Rosa, Fluctuat and FPTaylor for double-precision floating-point arithmetic.
(r) marks refactored benchmarks, (e) marks benchmarks with additional input errors
benchmark
Rosa
Fluctuat
Fluctuat
(subdiv)
FPTaylor
Rosa
Fluctuat
Fluctuat
(subdiv)
FPTaylor
3.90e-13
3.90e-13
4.08e-8
4.08e-8
3.65e-11
3.65e-11
7.97e-16
1.15e-15
3.21e-13
9.20e-14
9.26e-14
1.29e-13
1.34e-13
6.99e-14
7.03e-14
1.54e-13
1.40e-13
2.10e-11
1.88e-11
3.65e-11
3.65e-11
7.41e-16
1.09e-15
3.21e-13
2.21e-14
2.21e-14
2.87e-14
2.87e-14
1.34e-14
1.32e-14
1.35e-13
1.35e-13
1.17e-11
1.17e-11
3.61e-11
3.61e-11
5.52e-16
8.90e-16
2.87e-13
2.11e-14
2.11e-14
2.62e-14
2.62e-14
1.55e-14
1.55e-14
total
total (-jet)
8
7
95
77
1
1
2
1
1
5
2
2
1
4
2
209
37
1
1
1
1
1
1
1
1
1
1
1
1
1
1
1
15
13
2
2
2
2
2
2
1
1
1
2
2
2
2
2
2
27
23
7
7
12
12
6
5
6
4
7
8
8
6
7
7
7
109
85
5.45e-11
4.67e-4
1.82e-9
2.82e-9
1.24e-9
1.50e-7
2.09e-11
2.21e-11
1.37e-7
6.02e-10
6.14e-10
2.53e-10
1.50e-7
1.01e-11
1.82e-11
3.85e-8
4.61e-10
5.86e-10
3.32e-10
1.50e-7
1.00e-11
total
total (-jet)
13
76
2
1
2
2
2
98
22
1
1
1
1
1
1
1
7
6
2
2
2
2
2
2
1
13
11
7
13
8
10
8
6
6
58
45
1.09e-13
1.37e-11
4.08e-8
4.67e-4
3.65e-11
1.50e-7
8.66e-14
1.94e-9
1.45e-13
3.02e-9
7.32e-14
1.33e-09
4.84e-14
6.28e-12
1.88e-11
1.37e-7
3.34e-11
1.15e-7
2.21e-14
6.51e-10
2.44e-14
6.33e-10
9.50e-15
2.30e-10
1.57e-14
4.77e-13
1.48e-11
1.52e-11
6.78e-8
2.48e-14
4.59e-10
2.92e-14
4.84e-10
1.49e-14
2.76e-10
total
total (-jet)
7
12
84
81
13
13
3
2
2
2
5
5
229
64
1
1
1
1
1
1
1
1
1
1
1
1
12
10
2
2
2
2
2
2
2
2
2
2
2
2
24
20
5
11
1731
61
284
6
109
6
37
11
316
2577
846
with roundoff errors only
doppler
doppler (r)
jet
jet (r)
rigidBody
rigidBody (r)
sine
sineOrder3
sqroot
turbine1
turbine1 (r)
turbine2
turbine2 (r)
turbine3
turbine3 (r)
4.15e-13
2.42e-13
5.33e-9
4.91e-9
3.65e-11
3.65e-11
5.74e-16
9.96e-16
2.87e-13
5.99e-14
5.15e-14
7.68e-14
6.30e-14
4.62e-14
4.02e-14
with input errors
doppler (re)
jet (re)
turbine1 (re)
turbine2 (re)
turbine3 (re)
rigidBody (re)
sine (e)
1.83e-11
3.36e-7
4.61e-10
5.87e-10
3.33e-10
1.50e-7
1.00e-11
with input constraint
doppler (r)
doppler (re)
jet (r)
jet (re)
rigidBody (r)
rigidBody (re)
turbine1 (r)
turbine1 (re)
turbine2 (r)
turbine2 (re)
turbine3 (r)
turbine3 (re)
1.76e-14
4.67e-13
4.91e-9
3.36e-7
1.66e-11
8.84e-8
4.26e-14
4.61e-10
5.26e-14
5.87e-10
3.55e-14
3.33e-10
ACM Journal Name, Vol. V, No. N, Article A, Publication date: January YYYY.
Towards a Compiler for Reals
A:17
Except for the jet example, which is difficult for Z3, Rosa computes essentially as tight error
bounds as FPTaylor with a smaller running time.
Inputs with Nonlinear Constraint For the last set of benchmarks, we have constrained the
inputs with a nonlinear constraint of the form x ∗ x + y ∗ y + z ∗ z < c, where x, y, z are input
variables and c is a meaningful benchmark-specific constant. This constraint is representative
of constraints that cannot be captured by a linear technique like affine arithmetic. In Rosa,
this constraint can be specified naturally in the precondition. In Fluctuat, it is possible to
enclose the computation in an if-condition (if (constr) ...) and the affine terms will be
constrained with a linearized branch condition. We used the ’Constraints on noise symbols’
setting. FPTaylor provides syntax to specify additional constraints, however these are only
supported with Z3 as the backend, and hence without the improved rounding model. We
observe that no one tool consistently provides the most accurate error estimates, but that
FPTaylor’s technique turns out to be quite expensive in this case.
6. LOOPS
We have identified a class of loops for which the propagation of errors idea allows us to
express the numerical errors as a function of the number of iterations. Concretely, we assume
a single non-nested loop without conditional branches for which the ranges of variables are
bounded and fixed statically. We do not attempt to prove that ranges are preserved across
loop iterations; we leave the discovery of suitable inductive invariants that implies ranges
for future work. Our approach does not include all loops, but it does cover a number of
interesting patterns, including simulations of initial value problems in physics. We note that
the alternative for analyzing numerical errors in general nonlinear loops is unrolling, which,
as our experiments show, does not scale well.
6.1. General Error Propagation
Representing the computation of the loop body by f , we want to compute the overall
error after k-fold iteration f k of f : |f k (x) − f˜k (x̃)|. f, g and σ are now vector-valued:
f, g, σ : Rn → Rn , because we are nesting the potentially multivariate function f . In essence,
we want to compute the effect of iterating Equation 3.
Theorem: Let g be such that |f (x) − f (y)| ≤ g(|x − y|), it satisfies g(x + y) ≤ g(x) + g(y)
and is monotonic. Further, σ and λ satisfy σ(x̃) ≤ |f (x̃) − f˜(x̃)| and |x − x̃| ≤ λ. The absolute
value is taken component-wise. Then the numerical error after k iterations is given by
|f k (x) − f˜k (x̃)| ≤ g k (|x − x̃|) +
k−1
X
g i (σ(f˜k−i−1 (x̃)))
(8)
i=0
Thus, the overall error after k iterations can be decomposed into the initial error propagated
through k iterations, and the roundoff error from the ith iteration propagated through the
remaining iterations.
ACM Journal Name, Vol. V, No. N, Article A, Publication date: January YYYY.
A:18
E. Darulova and V. Kuncak
Proof: We show this by induction. The base case k = 1 is covered by our treatment of
straight-line computations (subsection 5.2). By adding and subtracting f (f˜k−1 (x̃))1 we get
k
|f (x)1 − f˜k (x̃)1 |
..
.
k
k
˜
|f (x)n − f (x̃)n |
k
|f (x)1 − f (f˜k−1 (x̃))1 |
|f (f˜k−1 (x̃))1 − f˜k (x̃)1 |
..
..
≤
+
.
.
k
k−1
k−1
k
˜
˜
˜
|f (x)n − f (f
(x̃))n |
|f (f
(x̃))n − f (x̃)n |
Applying the definitions of g and σ
k−1
|f
(x)1 − f˜k−1 (x̃)1 |
..
k−1
≤ g
+ σ(f˜ (x̃))
.
|f k−1 (x)n − f˜k−1 (x̃)n |
then using the induction hypothesis and monotonicity of g,
!
k−2
X
k−1 ~
i
k−i−1
≤g g
(λ) +
g (σ(f˜
(x̃))) + σ(f˜k−1 (x̃))
i=0
then using g(x + y) ≤ g(x) + g(y), we finally have
≤ g k (~λ) +
k−1
X
g i (σ(f˜k−i−1 (x̃))) + σ(f˜k−1 (x̃))
i=1
= g k (~λ) +
k−1
X
g i (σ(f˜k−i−1 (x̃)))
i=0
6.2. Closed Form Expression
We instantiate the propagation function g as before using propagation coefficients. Evaluating Equation 8 as given, with a fresh set of propagation coefficients for each iteration i
amounts to loop unrolling, but with a loss of correlation between each loop iteration. We
observe that when the ranges are bounded (as by our assumption), then we can compute
K as a matrix of propagation coefficients, and similarly obtain σ(f˜i ) = σ as a vector of
constants, both valid for all iterations. Then we obtain a closed-form for the expression of
the error:
|f k (x) − f˜k (x̃)| ≤ K k λ +
k−1
X
i=1
K iσ + σ = K k λ +
k−1
X
K iσ
i=0
where λ is the vector of initial errors. Denoting by I the identity matrix, if (I − K)−1 exists,
|f k (x) − f˜k (x̃)| ≤ K k λ + ((I − K)−1 (I − K k ))σ
We obtain K k with power-by-squaring and compute the inverse with the Gauss-Jordan
method with rational coefficients to obtain sound results (though a closed-form is not strictly
necessary for our purpose because we do know the number of iterations k).
Computing K and σ When the ranges of the variables of the loop are inductive, that is,
both the real-valued and the finite-precision values remain within the initial ranges, then
these are clearly the ranges for the computation of K and roundoffs σ. For loops, we require
ACM Journal Name, Vol. V, No. N, Article A, Publication date: January YYYY.
Towards a Compiler for Reals
A:19
the user to specify both the real-valued ranges of variables (e.g. a <= x && x <= b) as well
as the actual finite-precision ones (c <= ~x && ~x <= d, as in Figure 4). We also require that
the actual ranges always include the real ones ([a, b] ⊆ [c, d]), and we use the actual ranges
([c, d]) for the computation of K and σ. We believe that it is reasonable to assume that a
user writing these applications to have the domain knowledge to be able to provide these
specifications.
6.3. Handling Additional Sources of Errors
What if roundoff errors are not the only errors present? If the real-valued computation
given by the specification is also the ideal computation, we can simply add the errors in
the same way as roundoff errors. If the real-valued computation is, however, already an
approximation of some other unknown ideal function, say f∗ , it is not directly clear how our
error computation applies.
This may be the case, for example, for truncation errors due to a numerical algorithm. To
model such errors, let us suppose that we can compute (or at least overestimate) these by a
function τ : Rn → Rn , i.e. τf∗ (x) = |f∗ (x) − f (x)|.
In the following we consider the one-dimensional case n = 1 for simplicity of exposition,
but it generalizes as before to the n-dimensional case. We can apply a similar separation of
errors as before:
|f∗ (x) − f˜(x̃)|
≤ |f∗ (x) − f (x)| + |f (x) − f (x̃)| + |f (x̃) − f˜(x̃)|
= τ (x) + g(|x − x̃|) + σ(x̃)
which lets us decompose the overall error into the truncation error, the propagated initial
error and the roundoff error. If we now iterate, we find by a similar argument as before:
|f∗m (x) − f˜(x̃)|
≤ g m (|x − x̃|) +
m−1
X
g j τ (f∗m−j−1 (x)) + g j σ(f˜m−j−1 (x̃))
j=0
= g m (|x − x̃|) +
m−1
X
g j τ (f∗m−j−1 (x)) + σ(f˜m−j−1 (x̃))
j=0
The result essentially means that our previously defined method can also be applied to the
case when truncation (or similar) errors are present. We do not pursue this direction further
however, and leave a proper automated treatment of truncation errors to future work.
6.4. Experimental Results
We evaluate our technique on three benchmarks in Table II. We already presented the
pendulum benchmark in Figure 4. The mean benchmark computes a running average of
values in a range of [-1200, 1200]. The nbody benchmark is a 2-body simulation of Jupiter
orbiting around the Sun. For each benchmark we consider different number of iterations of
the loop and report the error for one of the loop’s variables. Fluctuat is not able to compute
a fixpoint for these benchmarks, as the errors keep growing with each iteration. Instead we
manually set the number of times the loop is unrolled. For the pendulum 50 benchmark
Rosa is able to compute a tighter error bound with a faster runtime. For larger numbers of
iterations, Fluctuat reports an error of ∞. This is also the result for the nbody benchmark.
For the mean benchmark, where the computation is less complex, Fluctuat can compute
tighter error bounds, at the expense of much longer analysis times. This illustrates that
our technique outperforms unrolling in Fluctuat for benchmarks that are highly nonlinear,
ACM Journal Name, Vol. V, No. N, Article A, Publication date: January YYYY.
A:20
E. Darulova and V. Kuncak
Table II. Absolute errors and runtimes for different benchmarks
and different number of loop iterations
benchmark
Absolute errors
Rosa
Fluctuat
pendulum 50
pendulum 100
pendulum 250
pendulum 500
pendulum 1000
mean 100
mean 500
mean 1000
mean 2000
mean 3000
mean 4000
nbody 50
nbody 100
2.21e-14
8.82e-14
2.67e-12
6.54e-10
3.89e-5
3.21e-7
1.62e-6
3.30e-6
4.51e-6
4.96e-6
5.12e-6
1.30e-11
1.35e-8
2.43e-13
9.92e-9
1.01e-8
1.01e-8
1.03e-8
1.05e-8
1.06e-8
-
Running times
Rosa
Fluctuat
8
8
8
8
8
5
6
7
4
4
5
794
776
47
1
5
27
158
392
734
-
whereas Fluctuat’s strategy may be used for cases where the nonlinearity is limited as is
the number of iterations. Note that Rosa’s runtime is independent of the loop’s number of
iterations.
7. DISCONTINUITIES
Recall the piece-wise jet engine approximation from Figure 5. Due to the initial errors on x
and y, the real-valued computation may take a different branch than the finite-precision one,
and thus produce a different result. We call this difference the discontinuity error.
We will assume that individual branch conditions are of the form e1◦e2, where ◦ ∈ {<, ≤
, >, ≥} and e1, e2 are arithmetic expressions. More complex conditions can be obtained by
nesting conditionals. We do not assume the function represented by the conditional to be
neither smooth nor continuous. We perform our analysis pairwise for each pair of paths in
the program. While this gives, in the worst-case, an exponential number of cases to consider,
we found that many of these paths are infeasible due to inconsistent branch conditions; such
infeasible paths are eliminated early.
7.1. Applying Separation of Errors
Using our previous notation, let us consider a function with a single branch statement like
in the example above and let f1 and f2 be the real-valued functions corresponding to the if
and the else branch respectively. Then, the discontinuity error is given by |f1 (x) − f˜2 (x̃)|,
i.e. the real computation takes branch f1 , and the finite-precision one f2 . The opposite case
is analogous. We again apply the idea of separation of errors:
|f1 (x) − f˜2 (x̃)| ≤ |f1 (x) − f1 (x̃)| + |f1 (x̃) − f2 (x̃)| + |f2 (x̃) − f˜2 (x̃)|
(9)
The individual components are
(1) |f1 (x) − f1 (x̃)|: the difference in f1 due to initial errors. We can compute this
difference with our propagation coefficients: |f1 (x) − f1 (x̃)| ≤ K|x − x̃|.
(2) |f1 (x̃) − f2 (x̃)|: the real-valued difference between f1 and f2 . We can bound
this value by the Z3-aided range computation from section 4.
(3) |f2 (x̃) − f˜2 (x̃)|: the roundoff error when evaluating f2 in finite-precision arithmetic. We use the procedure from subsection 5.1 as before.
ACM Journal Name, Vol. V, No. N, Article A, Publication date: January YYYY.
Towards a Compiler for Reals
A:21
We expect the individual parts to be easier to handle for the underlying SMT-solver since
we reduce the number of variables and correlations. We clearly introduce an additional
over-approximation, but we observed in our experiments that this is in general small. In
contrast, Fluctuat’s approach relies on constraints on the affine forms to capture the different
branch conditions [Goubault and Putot 2013]. A split of the total error into two parts is
also possible, e.g. as |f1 (x) − f˜2 (x̃)| ≤ |f1 (x) − f2 (x̃)| + |f2 (x̃) − f˜2 (x̃)|, which performs
one computation less. This split, combined with a precise constraint relating x to x̃ would
introduce one constraint with many correlations between variables [Darulova and Kuncak
2014]. Such a precise and complex relation overwhelms the SMT solver quickly, but bounding
the ranges without the correlation information yields unsatisfactory results.
7.2. Determining Ranges for x and x̃
As in the previous sections, it is crucial to determine the ranges of x, x̃ ∈ R over which to
evaluate the individual parts of Equation 9. A sound approach would be to simply use the
input ranges, but this would lead to unnecessary over-approximations. In general, not all
inputs can exhibit a divergence between the real-valued and the finite-precision computation.
They are determined by the branch conditions and the errors on the variables. Consider the
branch condition if (e1 < e2) and the case where the real-valued path takes the if-branch,
i.e. variable x satisfies e1 < e2 and x̃ satisfies e1 ≥ e2. The constraint for the finite- precision
variables x̃ is then
e1 + δ1 < e2 + δ2 ∧ e1 ≥ e2
where δ1 , δ2 are error intervals on evaluating e1 and e2 respectively. This constraint expresses
that we want those values which satisfy the condition e1 ≥ e2, but are “close enough” to
the boundary such that their corresponding ideal real value could take the other path. We
create such a constraint both for the variables representing finite-precision values (x̃), as
well as the real-valued ones x and use them as additional constraints when computing the
individual parts of Equation 9. The procedure for other branch conditions is analogous.
7.3. Experimental Results
We evaluate our technique on a number of benchmarks with discontinuities, which we
have either constructed by piece-wise approximating a more complex function or chosen
from Goubault and Putot [2013]. All the benchmarks’ source code is available online. We
compare our results in terms of accuracy and performance against Fluctuat. Fluctuat does
not check for discontinuity errors by default; we enable this analysis with the ’Unstable
test analysis’ option (this is the only way). Subdivisions, however, do not appear to work
with this setting. Table III summarizes our results. While Fluctuat is faster than Rosa,
Rosa is able to compute significantly tighter error bounds and, we believe, achieves a good
compromise between accuracy and performance.
8. RELATED WORK
To the best of our knowledge, Fluctuat [Goubault and Putot 2011; Goubault and Putot
2013] and FPTaylor [Solovyev et al. 2015] are most related to our work. We are not aware of
other tools or techniques that can soundly and automatically quantify numerical errors in
the presence of nonlinearity, branches and loops.
In the context of abstract interpretation, domains exist that are sound with respect to
floating-points and that can be used to prove the absence of runtime errors such as division by
zero [Blanchet et al. 2003; Miné 2004; Feret 2004; Chen et al. 2008; Ghorbal et al. 2009]. Feret
[2005] presents an abstract domain which associates the ranges with the iteration count,
similar to our proposed technique for loops. Martel [2002] considers the stability of loops,
by proving whether loops can asymptotically diverge. The problem that we are solving is
ACM Journal Name, Vol. V, No. N, Article A, Publication date: January YYYY.
A:22
E. Darulova and V. Kuncak
Table III. Absolute discontinuity errors computed and runtimes of Rosa
and Fluctuat
benchmark
Rosa
Fluctuat
cubicSpline
jetApprox
jetApprox (err)
jetApproxBadFit
jetApproxBadFit (err)
jetApproxGoodFit
jetApproxGoodFit (err)
linearFit
quadraticFit
quadraticFit (err)
quadraticFit2
quadraticFit2 (err)
simpleInterpolator
sortOfStyblinski
sortOfStyblinski (err)
squareRoot
squareRoot3
squareRoot3Invalid
styblinski
styblinski (err)
1.25e-15
0.0232
0.0242
0.8825
0.8852
0.0428
0.0450
0.6374
0.2548
0.2551
3.14e-9
0.0009
3.40e-5
1.0878
1.0982
0.0238
2.76e-9
3.93e-9
4.81e-8
0.0132
12.00
18.40
19.06
9.305
10.09
5.191
5.193
1.721
10.60
10.96
0.6321
0.7188
1.0e-5
27.07
28.82
0.0394
0.4289
0.4288
121.16
124.10
Rosa
Fluctuat
9
46
44
16
11
5
5
3
21
20
4
4
1
5
5
3
6
6
30
25
1
1
1
1
1
1
1
1
1
1
1
1
1
1
1
1
1
1
1
1
different, however, as we want to quantify the difference between the real-valued and the
finite-precision computation.
Floating-points have been formalized in the SMT-LIB format [Rümmer and Wahl 2010],
and approaches exist which deal with the prohibiting complexity of bit-precise techniques
via approximations [Brillout et al. 2009; Haller et al. 2012]. Efficient combination of theories
needed to express roundoff errors is non-trivial, and we are not aware of an approach that is
able to quantify the deviation of finite-precision computations with respect to reals. Floatingpoint precision assertions can also be proven using an interactive theorem prover [Boldo and
Marché 2011; Linderman et al. 2010; Ayad and Marché 2010; Harrison 2006]. These tools can
reason about ranges and errors of finite-precision implementations, but target specialized
and precise properties, which, in general, require an expert user and interactively guiding
the proof. Very tight error bounds have been shown by manual proof for certain special
computations, such as powers [Graillat et al. 2014]. Our work is on the other side of the
trade-off between accuracy and automation as well as generality.
Synthesis of specifically fixed-point arithmetic programs has also been an area of active
research, with different utilized techniques: simulation or testing [Mallik et al. 2007; Jha and
Seshia 2013], interval or affine arithmetic [Lee et al. 2006] or automatic differentiation [Gaffar
et al. 2004]. Some approaches try to optimize the bit-width whereas in our case we keep
it fixed, but provide a sound and accurate analysis, which could be used in combination
with an optimization technique, like e.g. Jha and Seshia [2013]. A similar approach to our
range estimation has been developed independently by Kinsman and Nicolici [2009] in the
context of fixed-point arithmetic. We also identify the potential of additional constraints
and develop optimizations to make the use of an SMT solver efficient enough. Further, our
techniques aim to be generally applicable to various finite-precision arithmetics.
Several approaches also exist to test the stability of numerical programs, e.g. by perturbation of low-order bits and rewriting [Tang et al. 2010], or by perturbing the rounding
modes [Scott et al. 2007]. Another common theme is to run a higher-precision program alongACM Journal Name, Vol. V, No. N, Article A, Publication date: January YYYY.
Towards a Compiler for Reals
A:23
side the original one. Benz et al. [2012] does so by instrumentation, Paganelli and Ahrendt
[2013] generates constraints which are then discharged with a floating-point arithmetic solver
and Chiang et al. [2014] developed a guided search to find inputs which maximize errors. Lam
et al. [2013] uses instrumentation to detect cancellation and thus loss of precision. Ivancic
et al. [2010] combines abstract interpretation with model checking to check the stability of
programs, tracking one input at a time. Majumdar et al. [2010] uses concolic execution to
find two sets of inputs which maximize the difference in the outputs. These approach are
based on testing, however, and cannot prove sound bounds.
It is natural to use the Jacobian for sensitivity analysis. Related to our work is a proof
framework using this idea for showing programs robust in the sense of k-Lipschitz continuity [Chaudhuri et al. 2011]. Note, however, that our approach does not require programs to be
continuous. Gazeau et al. [2012] relaxes the strict definition of robustness to programs with
specified uncertainties and presents a framework for proving while-loops with a particular
structure robust. Our work follows the philosophy of these approaches in leveraging Jacobians
of program paths, yet we explicitly incorporate the handling of roundoff errors in a fully
automated system.
9. CONCLUSION
We believe that numerical errors, such as roundoff errors, should not be an afterthought
and that programming language support is needed and possible to help scientists write
numerical code that does what it is expected to do. To this end, we presented, on one hand,
a real-valued specification language with explicit error annotations from which our tool Rosa
synthesizes finite-precision code that fulfills the given specification. On the other hand, we
presented a set of techniques based on unified principles which provides automated, efficient,
static error analysis which is crucial towards making such a compiler practical. We have
extensively evaluated these techniques against state-of-the-art tools and we believe they
represent an interesting compromise between accuracy and efficiency.
REFERENCES
2015. Frama-C Software Analyzers. http://frama-c.com/index.html. (2015).
Adolfo Anta, Rupak Majumdar, Indranil Saha, and Paulo Tabuada. 2010. Automatic Verification of Control
System Implementations. In EMSOFT.
Ali Ayad and Claude Marché. 2010. Multi-Prover verification of floating-point programs. In IJCAR.
David H. Bailey, Yozo Hida, Xiaoye S. Li, and Brandon Thompson. 2013. C++/Fortran-90 double-double
and quad-double package. http://crd-legacy.lbl.gov/~dhbailey/mpdist/. (2013).
Florian Benz, Andreas Hildebrandt, and Sebastian Hack. 2012. A dynamic program analysis to find floatingpoint accuracy problems. In PLDI.
Bruno Blanchet, Patrick Cousot, Radhia Cousot, Jérome Feret, Laurent Mauborgne, Antoine Miné, David
Monniaux, and Xavier Rival. 2003. A static analyzer for large safety-critical software. In PLDI. 196–207.
Sylvie Boldo and Claude Marché. 2011. Formal verification of numerical programs: from C annotated
programs to mechanical proofs. Mathematics in Computer Science (2011).
A. Brillout, D. Kroening, and T. Wahl. 2009. Mixed abstractions for floating-point arithmetic. In FMCAD.
69–76.
Swarat Chaudhuri, Sumit Gulwani, Roberto Lublinerman, and Sara Navidpour. 2011. Proving Programs
Robust. In ESEC/FSE.
Liqian Chen, Antoine Miné, and Patrick Cousot. 2008. A Sound Floating-Point Polyhedra Abstract Domain.
In APLAS.
Wei-Fan Chiang, Ganesh Gopalakrishnan, Zvonimir Rakamaric, and Alexey Solovyev. 2014. Efficient Search
for Inputs Causing High Floating-point Errors. In PPoPP.
Eva Darulova and Viktor Kuncak. 2011. Trustworthy Numerical Computation in Scala. In OOPSLA.
Eva Darulova and Viktor Kuncak. 2014. Sound Compilation of Reals. In POPL.
L. H. de Figueiredo and J. Stolfi. 2004. Affine Arithmetic: Concepts and Applications. Numerical Algorithms
37, 1-4 (2004).
ACM Journal Name, Vol. V, No. N, Article A, Publication date: January YYYY.
A:24
E. Darulova and V. Kuncak
Leonardo De Moura and Nikolaj Bjørner. 2008. Z3: an efficient SMT solver. In TACAS.
Jérôme Feret. 2004. Static Analysis of Digital Filters. In ESOP.
Jérôme Feret. 2005. The Arithmetic-Geometric Progression Abstract Domain. In VMCAI.
A.A. Gaffar, O. Mencer, and W. Luk. 2004. Unifying bit-width optimisation for fixed-point and floating-point
designs. In FCCM.
Ivan Gazeau, Dale Miller, and Catuscia Palamidessi. 2012. A non-local method for robustness analysis of
floating point programs. In QAPL.
Khalil Ghorbal, Eric Goubault, and Sylvie Putot. 2009. The Zonotope Abstract Domain Taylor1+. In CAV.
Springer Berlin Heidelberg.
Khalil Ghorbal, Eric Goubault, and Sylvie Putot. 2010. A Logical Product Approach to Zonotope Intersection.
In CAV.
Eric Goubault and Sylvie Putot. 2011. Static Analysis of Finite Precision Computations. In VMCAI.
Eric Goubault and Sylvie Putot. 2013. Robustness Analysis of Finite Precision Implementations. In APLAS.
Stef Graillat, Vincent Lefèvre, and Jean-Michel Muller. 2014. On the maximum relative error when computing
xn in floating-point arithmetic. Technical Report <ensl-00945033v2>. Laboratoire d’Informatique de
Paris 6, Inria Grenoble Rhône-Alpes.
L. Haller, A. Griggio, M. Brain, and D. Kroening. 2012. Deciding floating-point logic with systematic
abstraction. In FMCAD.
John Harrison. 2006. Floating-Point Verification using Theorem Proving. In Formal Methods for Hardware
Verification.
F. Ivancic, M.K. Ganai, S. Sankaranarayanan, and A. Gupta. 2010. Numerical stability analysis of floatingpoint computations using software model checking. In MEMOCODE.
Susmit Jha and Sanjit A. Seshia. 2013. Synthesis of Optimal Fixed-Point Implementations of Numerical
Software Routines. In Proc. Sixth International Workshop on Numerical Software Verification (NSV).
Dejan Jovanović and Leonardo de Moura. 2012. Solving Non-linear Arithmetic. In IJCAR.
A.B. Kinsman and N. Nicolici. 2009. Finite Precision Bit-Width Allocation using SAT-Modulo Theory. In
DATE.
Michael O. Lam, Jeffrey K. Hollingsworth, and G.W. Stewart. 2013. Dynamic floating-point cancellation
detection. Parallel Comput. 39, 3 (2013).
D. Lee, A. A. Gaffar, R. C. C. Cheung, O. Mencer, W. Luk, and G. A. Constantinides. 2006. AccuracyGuaranteed Bit-Width Optimization. IEEE Trans. on CAD of Integrated Circuits and Systems 25, 10
(2006).
Michael D. Linderman, Matthew Ho, David L. Dill, Teresa H. Meng, and Garry P. Nolan. 2010. Towards
program optimization through automated analysis of numerical precision. In CGO.
R. Majumdar, I. Saha, and Zilong Wang. 2010. Systematic Testing for Control Applications. In MEMOCODE.
A. Mallik, D. Sinha, P. Banerjee, and H. Zhou. 2007. Low-Power Optimization by Smart Bit-Width Allocation
in a SystemC-Based ASIC Design Environment. IEEE Trans. on CAD of Integrated Circuits and Systems
26, 3 (2007).
Matthieu Martel. 2002. Static Analysis of the Numerical Stability of Loops. In SAS.
Antoine Miné. 2004. Relational Abstract Domains for the Detection of Floating-Point Run-Time Errors. In
ESOP.
R.E. Moore. 1966. Interval Analysis. Prentice-Hall.
Gabriele Paganelli and Wolfgang Ahrendt. 2013. Verifying (In-)Stability in Floating-point Programs by
Increasing Precision, using SMT Solving. In SYNASC.
Philipp Rümmer and Thomas Wahl. 2010. An SMT-LIB Theory of Binary Floating-Point Arithmetic. In
Informal proceedings of 8th International Workshop on Satisfiability Modulo Theories (SMT) at FLoC.
N.S. Scott, F. Jézéquel, C. Denis, and J.-M. Chesneaux. 2007. Numerical ‘health check’ for scientific codes:
the CADNA approach. Computer Physics Communications (2007).
Alexey Solovyev, Charles Jacobsen, Zvonimir Rakamaric, and Ganesh Gopalakrishnan. 2015. Rigorous
Estimation of Floating-Point Round-off Errors with Symbolic Taylor Expansions. In FM.
Enyi Tang, Earl Barr, Xuandong Li, and Zhendong Su. 2010. Perturbing numerical calculations for statistical
analysis of floating-point program (in)stability. In ISSTA.
C. Woodford and C. Phillips. 2012. Numerical Methods with Worked Examples. Vol. 2nd. Springer.
ACM Journal Name, Vol. V, No. N, Article A, Publication date: January YYYY.
| 6cs.PL
|
Post-transcriptional
knowledge
in
pathway analysis increases the accuracy
of phenotypes classification
This manuscript has been published in Oncotarget Vol 7, No
34.
If you are using this work please refer to:
Alaimo, S., Giugno, R., Acunzo, M., Veneziano, D., Ferro, A., &
Pulvirenti, A. (2016). Post-transcriptional knowledge in pathway
analysis increases the accuracy of phenotypes classification.
Oncotarget. (http://dx.doi.org/10.18632/oncotarget.9788)
Post-transcriptional
knowledge
in
pathway analysis increases the accuracy
of phenotypes classification
Salvatore Alaimo 1 , Rosalba Giugno 2 , Mario Acunzo 3 , Dario Veneziano 3 , Alfredo Ferro 2 , and
Alfredo Pulvirenti 2 , *
1
Department of Mathematics and Computer Science and 2Department of Clinical and Experimental
Medicine, University of Catania, Catania, Italy,
3
Department of Molecular Virology, Immunology and Medical Genetics, Comprehensive Cancer Center, The
Ohio State University, Columbus, OH, USA.
Abstract
Motivation: Prediction of phenotypes from high-dimensional data is a crucial task in precision biology and
medicine. Many technologies employ genomic biomarkers to characterize phenotypes. However, such
elements are not sufficient to explain the underlying biology. To improve this, pathway analysis techniques
have been proposed. Nevertheless, such methods have shown lack of accuracy in phenotypes classification.
Results: Here we propose a novel methodology called MITHrIL (Mirna enrIched paTHway Impact anaLysis)
for the analysis of signaling pathways, which has built on top of the work of Tarca et al., 2009. MITHrIL
extends pathways by adding missing regulatory elements, such as microRNAs, and their interactions with
genes. The method takes as input the expression values of genes and/or microRNAs and returns a list of
pathways sorted according to their deregulation degree, together with the corresponding statistical
significance (p-values). Our analysis shows that MITHrIL outperforms its competitors even in the worst
case. In addition, our method is able to correctly classify sets of tumor samples drawn from TCGA.
Availability: MITHrIL is freely available at the following URL: http://alpha.dmi.unict.it/mithril/
Introduction
The prediction of phenotypes, such as diseases, or of responses to therapies from the large amount of
genotypic high-dimensional data obtained through Next-Generation Sequencing techniques is an extremely
important task in translational biology and precision medicine. However, the gap between current analysis
techniques and the ability to obtain accurate knowledge is broad.
High-throughput sequencing and gene profiling techniques are radically transforming medical research,
allowing the full monitoring of a biological system. The use of these technologies typically generates a list of
differentially expressed elements (i.e. genes or microRNAs) whose behavior varies significantly among the
phenotypes under examination.
Furthermore, compared to traditional gene expression extraction techniques (eg. Microarray), deep
sequencing methods, such as RNA-Seq, provide much larger lists of differentially expressed genes,
increasing, therefore, the complexity of the analysis. The common approach to simplify and make the
analysis of such data more fruitful consisted in grouping genes into smaller sub-sets according to some
relationship, leveraging on existing knowledge-bases such as ontologies or pathways. The analysis of this
*
To whom correspondence should be addressed.
type of data at the functional level is crucial since it allows a strong reduction of dimensionality, thus
providing greater insights on the biology of the phenomenon under study [1].
An extensive class of techniques known as Pathway Analysis [2] goes in this direction. In the past, such
term has been associated to the analysis of ontological terms, protein-protein interaction (PPI) networks, or
to the inference of gene regulatory networks from expression data. More recently, great interest has
shifted toward a class of methods called Knowledge base-driven pathway analysis [3]. Such methods
leverage on existing databases, such as the Kyoto Encyclopedia of Gene and Genomes (KEGG) [4, 5] or
Pathway Commons [6], to identify those pathways that may be affected by the expression changes in the
observed phenotype. Knowledge base-driven pathway analysis techniques can be grouped into three
generations of approaches [3]: i) Over-Representation Analysis (ORA); ii) Functional Class Scoring (FCS); iii)
Pathway Topology-based (PT).
First-generation methods statistically evaluate the number of altered genes in a pathway with respect to
the set of all analyzed genes. After filtering the resulting gene set of an expression assessment experiment,
ORA strategies [7-13] typically divide the list of genes according to the pathway each gene belongs to. By
applying an hypothesis test (i.e. hypergeometric, chi-square, or binomial) they are able to determine if the
number of such genes is over- or under-represented. These methods, however, have some major
limitations. Firstly, considering only the number of differentially expressed genes, while omitting their
expression, implies that the magnitude of their change be unimportant for pathway activity. Furthermore,
considering only statistically significant differential expression may exclude those genes whose coordinated
alteration may lead to remarkable effects, although their differential expression may not be statistically
significant. Finally, they consider individual genes and pathways, respectively, in an manner independent of
the surrounding biological context, eluding what truly happens in reality.
A first generation method, DIANA-miRPath [14], assesses the impact of miRNAs in biological processes by
identifying the pathways in which they are significantly involved. The software employs a functional
annotation of one or more miRNAs by means of a hypergeometric distribution, or an unbiased empirical
distribution, or a statistical meta-analysis. Moreover, it allows the identification of sub-sets of miRNAs
which significantly regulates a collection of pathway, on the basis of experimental data.
Second-generation methods compensated some of the disadvantages of ORA approaches. Typically FCS
methods compute a gene-level statistic from their expression levels, by means of a statistical approach (i.e.
ANOVA, Q-statistic, signal-to-noise ratio, t-test, or Z-score). Such a statistic is calculated considering all
genes in a pathway [15-21] and its statistical significance is estimated through an appropriate null
hypothesis [16, 22-24]. FCS methods avoid some of the limitations of the ORA approaches by ranking all
genes through their expression level and by considering the dependencies within a pathway. However, by
using gene expression only to sort genes, they do not take into account the fact that such values can
change pathways activity.
In order to overcome the disadvantages of FCS methods, the third class of techniques models a pathway as
a graph, considering its topology when computing scores. A thorough analysis of all PT-based approaches
has been provided in [25].
In Draghici et al. [26], an analytical technique called impact factor (IF) was introduced. The impact factor is
a pathway-level score that takes into account biological factors such as the magnitude of change in genes
expression, the type of interactions between genes, and their location in the pathway. In Draghici et al.
[26], each pathway is modeled as a graph in which nodes represent genes, while edges represent
interactions between them. Authors also define a gene-level statistic (called perturbation factor, PF) as a
linear function of the change in gene expression and the perturbation of its neighborhood. Such a statistic
is then combined for each element in a pathway, and a p-value is computed by means of exponential
distribution.
The analysis method presented by Draghici et al. [26], has been further improved by the SPIA algorithm [27]
which attenuates the dominant effect exercised by the change in expression within PFs computation, while
reducing the high rate of false positives when the input list of genes is small. SPIA uses a bootstrap
procedure to evaluate the significance of the observed perturbation in the pathway. All this is combined
with a p-value computed in ORA style to make a full assessment of the statistical significance of the
perturbation of each pathway.
To reduce the number of false positives, and to obtain a more significant analysis, Vaske et al. [28]
presented the PARADIGM algorithm, which has been further improved by [29]. PARADIGM is a method to
infer patient-specific genetic activity by incorporating information regarding interactions between genes
provided in a pathway. The method predicts the degree of alteration in the activity of a pathway by
employing a probabilistic inference algorithm. The authors show that their model obtains significantly more
reliable results than SPIA. However, Mitrea et al. [25] stated they could not reproduce the results reported
in Vaske et al. [28], despite the full cooperation of its authors.
However, both SPIA and PARADIGM completely ignore post-transcriptional regulatory interactions enacted
by miRNAs. To fill this gap, Calura et al. developed a new approach, Micrographite [30], which is able to
integrate pathway with predicted and validated miRNA-target interactions. The method, by performing a
topological analysis based on expression profiles, is able to identify significant gene circuits specific of a
phenotype. The main advantage of the methodology is the ability to accurately describe the cellular
scenario that led to the input expression data.
Here, we present MITHrIL (miRNA enriched pathway impact analysis), a technique that extends the method
in [26] and SPIA [27], by combining their effectiveness and improves the reliability of the results. The
strength of MITHrIL lies in the enrichment of pathways with information regarding microRNAs, posttranscriptional regulatory elements whose addition is clearly essential to the greater meaningfulness of the
results. Our method, starting from expression values of genes and/or microRNAs, returns a list of pathways
sorted according to the degree of their de-regulation, together with the corresponding statistical
significance (p-values), and a predicted degree of alteration for each endpoint (a pathway node whose
alteration, based on current knowledge, affects the phenotype in a specific way).
To evaluate our algorithm, we used the decoy pathway methodology introduced in Vaske et al. [28] on
expression datasets obtained from The Cancer Genome Atlas. We showed that adding information on the
otherwise missing regulatory elements proves to be pivotal in improving the reliability of pathway analysis
methodologies. As further evidence of the reliability of our pathway impact analysis method, we employed
our algorithm for the classification of phenotypes. The results highlight the ability of our methodology to
strongly reduce the dimensionality of the data while maintaining a very high classification quality.
Results
Biological soundness. We compared our methodology with PARADIGM [28], SPIA [27] and Micrographite
[30] by employing the technique defined in Vaske et al. 2010. The aim is to establish whether the ranking
computed with a pathway analysis algorithm is biologically significant. This is achieved by defining random
pathways (called decoy pathways) with the same topology as the real ones but randomly selected nodes.
All pathways are then evaluated by each algorithm, estimating the ability of each method to properly
separate decoy pathways from real ones by means of a receiver operating characteristic (ROC) curve. In
principle, a method that can correctly distinguish real pathways from decoys should yield biologically
significant results.
We performed comparisons between MITHrIL, SPIA, PARADIGM and Micrographite on a set of selected
cancer types (see Table 1). Such a comparison allowed us, by ranking the datasets according to
performance, to identify the single cancer type in which our algorithm had the lowest quality results,
namely, Lung squamous cell carcinoma (LUSC).
The results of the four methodologies were ranked as follows: PARADIGM according to the average number
of significant scores, as described in [28]; SPIA according to the adjusted p-value as obtained through their
software implementation; Micrographite according to the pathway prioritization phase; MITHrIL according
to the adjusted p-value and the accumulator. More precisely, in MITHril, all the results are sorted first by pvalue and, thereafter, in the presence of equal p-values their corresponding accumulator is taken into
account to determine their order. In Supplementary Figures 1-3, we present the results of the detailed
comparison in each TCGA dataset. Out analysis clearly shows that MITHrIL gives the best performances. As
further proof of the goodness of our methodology, we computed the average area under each ROC curve
(AUC). The results were summarized in Figure 1 (more details can be found in Supplementary Table 1). The
four boxes in the figure represent the AUC variability range for the four compared methodologies,
respectively.
Prediction of cancer types. We also evaluated our algorithm by assessing its performances in terms of
capability to predict the cancer type. To do this, we elected to train the PAMr [31] classification algorithm
and evaluated its performance by means of a 10-fold cross validation (CV) procedure. PAMr is an approach
devised to predict cancer class from gene expression profiling, based on an enhancement of the nearest
shrunken centroid classifier. The algorithm is able to identify subsets of genes that best characterize each
class. The technique is general and can be used in many other classification problems. The CV procedure
takes as input all the feature profiles of each patient, and divides them into 10 subsets, by balancing the
elements of each class in each subset. A subset is, then, removed (test set), and the classifier is trained on
the remaining nine sets (training set), in order to prioritize and select the features. Each profile in the test
set is then classified, and the results are used to estimate the error. The methodology is repeated so that
each subset is used once as the test set. The CV procedure was designed in order to remove overfitting and
overestimation of the results.
A reference classification was thus established by applying such a procedure to the Log-Fold-Change of
differentially expressed genes of our cancer cases. The rationale behind such a choice is to show that
pathway perturbation, which takes into account network structure, increases the biologically soundness of
the results with respect to a plain Log-Fold-Change-based approach, widely used as a gold standard. First
we computed all differentially expressed genes for each tumor type, obtaining a total of 17,326 genes that
appear to be de-regulated in at least one disease. Next, we calculated their Log-Fold-Change in each
sample, trained a classifier and verified its performance using the above described CV procedure. The
results (Table 2) demonstrate that such a classification is quite reliable since it yields a very small error.
Notice that Micrographite is not able to compute pathway ranking for a single sample, therefore we did not
perform any classification using such a method. Hence, we ran MITHrIL, SPIA and PARADIGM on all samples
of our set of selected cancer types, and trained three classification models using total accumulation scores.
As before, we performed a 10-fold cross validation and evaluated errors in each class (Table 2).
Furthermore, leveraging the ability of MITHrIL and PARADIGM to return the perturbation for each of 3,165
pathway endpoints, we trained additional classifiers based on such values. Since SPIA computes pathwaylevel statistic by means of a linear equation system, it could not return perturbations of endpoints.
Therefore we elected to use MITHrIL without miRNAs to establish the classification performances of
endpoints, when such elements are missing.
Our analysis clearly shows that performances are considerably improved over reference classification, by
taking into account endpoint perturbations (Table 2). Moreover, we can notice a significant dimensionality
reduction of our data, since by using perturbation of pathway endpoints, computed by means of Equation 1
(see the section Materials and Method), we are able to train PAMR on about 3165 genes (18% of the
number of differentially expressed ones).
Table 2 reports also the classifications based on MITHrIL pathway accumulators. We recall that accumulator
summarizes, with a single value, the general perturbation we observe within a pathway. Hence, as a further
effect this yields a stronger dimensionality reduction. Although we notice a slight increase in
misclassification error, compared to reference classification, it is important to highlight that we were able
to reduce to 237 the number of features on which PAMR classifier was trained. Pathway Accumulators
were computed according to Equation 5. The last two columns of Table 2 report the classification
performances obtained by SPIA accumulators and PARADIGM scores. All of this shows that the addition of
miRNA information is crucial in order to obtain more reliable results. Notice that we cannot deduce any
information about the performances of PARADIGM extended with microRNAs, since no implementation
with such a knowledge is available, and this goes beyond the scope of our paper.
To further highlight the biological relevance of endpoints, we performed a set of experiments with
randomly selected nodes within pathways (see third and fourth columns of Table 2). The results show that
the choice of endpoints is reasonable, since endpoints synthetize perturbations of upstream nodes.
Coherence of the prediction of pathway nodes state. As a further validation our methodology, we chose
to verify the percentage of endpoints for which a coherent prediction of the deregulation is obtained.
Initially, we applied MITHrIL with and without miRNAs to estimate the perturbations for each endpoint of
each sample (excluding the expression values of the endpoints in order to avoid introducing a bias in our
results). Subsequently, we computed the percentage of endpoints for which the sign of perturbation agrees
with that of the log-Fold Change. This validation estimates the reliability of the predictions of our method
and the importance of the addition of miRNA knowledge to our model. The results (Figure 2) highlight that
plugging quantitative information on miRNAs is crucial to establish a far more comprehensive and
meaningful estimation of phenotype activity. Therefore, using perturbation without miRNAs could be
misleading.
Discussion
In the last decade, miRNAs have ever more revealed to be crucial in the modulation of numerous cellular
pathways via the exertion of their important regulatory function when targeting key genes. Since the first
connection between miRNAs and cancer was made in 2002 [32], miRNA deregulation has been proven to
be indeed at the root of the pathogenesis of all cancers [33]. It suffices to consider, for instance, how the
deregulation of even a single miRNA is capable to cause cancer, as in the case of miR-155 which has been
shown to be responsible for the onset of Acute Lyphoblastic Leukemia/high-grade lymphoma in mouse
[34]. Additionally, the predominant roles played by miRs 21, 221 and 222 in several cancer types prove the
importance these small RNA molecules have in tumor pathogenesis and progression, while also being a
determining factor in drug resistance [33]. In light of this and many other evidences discovered in recent
years, the integration of miRNA expression when evaluating cancer pathway perturbation has become of
utmost importance. The proper consideration of the crucial effects yielded by the action of these small
non-coding RNA molecules on overall gene expression indeed contributes to a more comprehensive
depiction of the biological reality, providing a more accurate means for pathway assessment and
phenotype categorization. In fact, given the very important biological role played by miRNAs, integrating
their evaluation can greatly help in the discernment of even fine changes in the cellular gene expression
profile, which could make the difference between a normal and abnormal phenotypes, already at disease
onset.
Here we presented a novel knowledge base-driven pathway analysis methodology called MITHrIL. By
enriching KEGG pathways with experimentally validated interactions between genes and miRNAs, MITHrIL
is capable to clearly improve the reliability of pathway-based analysis of phenotypes.
Through the enrichment with miRNA information, MITHrIL can greatly improve predictions over SPIA,
PARADIGM, and Micrographite. In fact, while the other methodologies cannot properly distinguish between
decoy pathways and real ones, MITHrIL is capable of obtaining much better results. Even our worst case
had superior results than our three competitors. From a biological standpoint, the ability to distinguish
decoy pathways from real ones addresses the fundamental necessity to be able to properly interpret the
actual cellular mechanisms as possessing a biological criterion which is crucial to the life of the cell and not
the result of random phenomena.
Therefore, we focused our analysis on the ability of our methodology to synthesize information gained
from gene expression data and thus provide novel biological clues. For this purpose, by using the PAMR
algorithm, we performed different types of classification, taking as reference the results arising from the
classification based directly on Log-Fold-Changes. Our findings demonstrated the capability of MITHrIL to
synthesize biological information contained in the data, while yielding high classification accuracy.
Furthermore, MITHrIL greatly reduces the dimensionality of data of about 73 times compared to a naive
Log-Fold-Change based method. This significant dimensionality reduction may also make the analysis more
accurate since it could reduce the noise that can be introduced by the technologies used to gather
expression data from samples. Furthermore, when knowing the phenotype that is being analyzed, it is
possible to further reduce the number of dimensions by focusing only on those pathways that are known to
be somehow involved in it.
Compared to SPIA, MITHrIL can also return the perturbation computed for pathway endpoints, whose
subsequent analysis can lead to important additional insights about the biology underlying the phenomena
under study. Indeed, the proper evaluation of pathway endpoints can contribute to a far more accurate
phenotype assessment, as a more detailed diversification between pathological phenotypes at the pathway
level is reflected more at the endpoints rather than in any other node of the pathway network. This allows
to more effectively distinguish pathologies sharing even a very similar set of deregulated genes (as you
could more easily distinguish similar yet different trees more easily by confronting their leaves rather than
their roots). By leveraging on endpoint perturbation, we are also able to greatly reduce the misclassification
error, although we are able to reduce data dimensionality by only 5 times. This allows us to stress the fact
that gene perturbations are capable to discriminate among the pathological classes of our data.
Leveraging on the potential provided by miRNA enrichment in pathway analysis, MITHrIL represents a
bioinformatic resource capable of a far more accurate evaluation of pathway deregulation in cancer. This
can provide a decisive contribution to cancer research in terms of directing researchers more effectively,
reducing costs and time requirements. Specifically, MITHrIL can contribute to an earlier diagnosis, an early
and more accurate drug resistance assessment, as well as to more precise prognosis in terms of predicting
future disease development.
Nevertheless, our results can be further improved given that pathways are still incomplete, thus potentially
resulting in partial or erroneous conclusions. Future development in pathway analysis methodologies
should take into account additional regulatory elements, such as long non-coding RNA (lncRNAs), along
with epigenetic information, such as methylation patterns, variants, or copy number variants. Mutations
could be exploited by considering their impact on the modulation of 𝛽 function (see the section Materials
and Method for the definition of 𝛽), for example by assessing the difference in interaction strength by
means of free energy. We could also define the 𝛽 function by evaluating the correlation between patient
expression profiles and corresponding phenotype.
Materials and Methods
Pathway Enrichment Outline
Our methodology distinguishes itself from other pathway analysis techniques primarily for the use of KEGG
[4, 5] pathways enriched with microRNAs (miRNAs) and their interactions with genes.
In order to achieve this, we downloaded all validated inhibition interactions between miRNA and targets
from miRTarBase [35] and miRecords [36]. We also obtained interactions between transcription factors
(TFs) and miRNAs from TransmiR [37]. By taking into account TFs activating miRNA genes we are able to
increase the knowledge stored within each pathway. We then standardized all identifiers in their respective
databases to avoid duplicates. The mapping of miRNA identifiers was performed by using miRBase release
20 [38-42] as reference database. For each target, we performed a twofold mapping procedure: firstly,
each gene identifier has been converted to its Entrez one; then, by taking advantage of KEGG REST API, we
mapped each Entrez Id to the corresponding KEGG Id. This standardized list of interactions was, lastly,
filtered to remove all duplicates. Such a procedure allowed us to build a knowledge base of 10,537
experimentally validated interactions between 385 miRNAs and 3,080 genes.
Pathway enrichment was performed by defining a new type of nodes representing miRNAs in the pathway
notation, along with two types of directed edges, for miRNA-target inhibition interactions and TF-miRNA
interactions, respectively. The enrichment is thus performed automatically by adding to each pathway only
miRNAs that interact with at least one element within it.
Finally, in order to acquire information on which endpoints are contained in each pathway, we employed a
depth-first search algorithm (DFS) [43] to automatically mark which genes are located at the end of the
chains of reactions in each pathway. The search for endpoints in a pathway starts from a random node. The
DFS algorithm follows the interactions down to the nodes from which no other one can be reached
(putative endpoints). The procedure is repeated until all nodes have been analyzed. Putative endpoints are,
then, manually screened to determine if they are associated with phenotypic changes as stated on the
KEGG database. Only the latter are taken as pathway endpoint. An example of endpoints is reported in
Supplementary Figure 4.
Algorithm
Our methodology consists in an extension of Draghici et al. [26] and Tarca et al. [27]. It requires a
case/control expression data set from which statistically differentially expressed features have been
extracted (genes, miRNAs, or both). For such elements, the computation of their Log-Fold-Change is also
needed. Starting from such information, MITHrIL computes, for each gene in a pathway, a Perturbation
Factor (PF), which is an estimate of how much its activity is altered considering its expression and 1neighborhood. Positive (negative) values of PF indicate that the gene is likely activated (inhibited). By
appropriately combining each PF of a pathway, our algorithm is, therefore, able to calculate an Impact
Factor (IF) and an Accumulator (Acc). The IF of a pathway is a metric expressing how important are the
changes detected in the pathway, the greater the value, the most significant are the changes. The Acc
indicates the total level of perturbation in the pathway and the general tendency of its genes: positive Acc
values indicate a majority of activated genes (or inhibited miRNAs), while negative ones corresponds to an
abundance of inhibited genes (or activated miRNAs). To the Acc is also assigned a p-value which is an
estimate of the probability of getting such accumulator by chance. Finally, by applying the [44] method, we
estimate the false discovery rate and p-values are adjusted on multiple hypotheses.
More precisely, let 𝑛 be a node in pathway 𝑃. . Its perturbation factor, 𝑃𝐹 𝑛, 𝑃. can be defined as:
𝑃𝐹 𝑛, 𝑃. = Δ𝐸 𝑛 +
4∈6 7,89
;∈<
𝛽 𝑢, 𝑛
4,89 𝛽 𝑢, 𝑑
⋅ 𝑃𝐹 𝑢, 𝑃. ,
(1)
where Δ𝐸 𝑛 is the Log-Fold-Change computed for the node 𝑛, 𝑈 𝑛, 𝑃. and 𝐷 𝑛, 𝑃. are the set of
upstream and downstream nodes of 𝑛 in pathway 𝑃. respectively, and 𝛽 𝑢, 𝑛 is a function that indicates
the strength and type of interaction between genes 𝑢 and 𝑛. In particular, negative values of 𝛽 indicate an
inhibitory effect, while positive values an activating one. To ensure that the perturbation coming from an
upstream node is divided to its downstream ones, proportionally to the strength of their interactions,
without altering the total perturbation, a normalization is applied dividing by the sum of the weights
absolute values. By exploiting the methodology described in Draghici et al. [26] we compute an impact
factor, 𝐼𝐹 𝑃. , which reflects the importance of the changes observed in a pathway, as:
1
7∈89 𝑃𝐹 𝑛, 𝑃.
+
,
(2)
𝑝 𝑃.
Δ𝐸 ⋅ 𝑁;G 𝑃.
where 𝑝 𝑃. is the probability, calculated using an hyper-geometric distribution, of obtaining a number of
𝐼𝐹 𝑃. = log
differentially expressed nodes at least equal to the observed one in 𝑃. ; Δ𝐸 is the mean Log-Fold-Change in
𝑃. ; finally, 𝑁;G 𝑃. represents the number of differentially expressed nodes in the pathway.
Our methodology takes also advantage of the accumulation (or accumulator) as described by Tarca et al.
[27]. Such a methodology has been revised to take into account the addition of miRNAs. In order to do so,
first we need to compute two partial accumulators, 𝐴𝑐𝑐J.K 𝑃. and 𝐴𝑐𝑐LG7G 𝑃. , which take into account
the perturbation, respectively, of miRNAs and genes:
𝐴𝑐𝑐J.K 𝑃. =
𝑃𝐹 𝑚, 𝑃. − Δ𝐸 𝑚 ,
(3)
𝑃𝐹 𝑔, 𝑃. − Δ𝐸 𝑔 ,
(4)
J∈89M
𝐴𝑐𝑐LG7G 𝑃. =
L
P
L∈89
where 𝑃.J and 𝑃. are the sets of miRNAs and genes present in 𝑃. , respectively.
L
Therefore, in equations 3 and 4, we sum the perturbations of all miRNAs (𝑃.J ) and genes (𝑃. ) in pathway
𝑃. , addressing the dominant effect of the expression change in the PF computation by subtracting such
values. We can now compute total perturbation accumulation, 𝐴𝑐𝑐 𝑃. , which measures whether the
pathway is likely activated or inhibited. The introduction of miRNAs in our model addresses the necessity to
take into account the fact that an increased (decreased) expression of such elements results in an inhibition
(activation) of the pathway. We compute 𝐴𝑐𝑐 𝑃. as:
𝐴𝑐𝑐 𝑃. = 𝐴𝑐𝑐LG7G 𝑃. − 𝐴𝑐𝑐J.K 𝑃. − 𝐸 𝐴𝑐𝑐 𝑃. ,
(5)
where 𝐸 𝐴𝑐𝑐 𝑃. is an estimate of the expected value of the distribution of all accumulators computed for
pathway 𝑃. , as explained below.
P-value estimation is then performed by combining the Z-scores, computed through an inverse
Standardized Normal distribution, associated to two probabilistic terms: the first is the probability of
obtaining by chance a number of differentially expressed genes in the pathway at least equal to the
observed one, while the second consists in the probability of observing by chance an accumulator higher
than the computed one. The first term corresponds to 𝑝 𝑃. introduced in equation 2. The second term,
instead, has to be estimated through a permutation test. In such a test, we assign, to a random group of
genes in the pathway in question, a Log-Fold-Change selected randomly from the input ones, so as to
compute a random accumulator. The procedure is repeated several times and the final probability is
estimated as the ratio between the number of random accumulators greater than 𝐴𝑐𝑐 𝑃. and the number
of repetitions performed. In our experiments, the repetitions were set to 2000 in order to obtain maximum
precision up to two decimal places.
At this stage we are also able to estimate expected value 𝐸 𝐴𝑐𝑐 𝑃.
accumulators.
as the median value of the random
Therefore, the final result of our algorithm consists of a list of pathways along with their impact factor,
accumulator and adjusted p-values. Such list is sorted by p-value and Acc.
Expression Data Sources
To perform a comprehensive test of our algorithm, we exploited expression data provided by The Cancer
Genome Atlas (beginning of 2014). We downloaded all patient expression profiles of genes (RNASeqV2
obtained through platforms Illumina Genome Analyzer and Illumina HiSeq) and miRNAs (miRNASeq
obtained through platforms Illumina Genome Analyzer and Illumina HiSeq). The initial dataset was then
filtered by removing all patients for which one of the two types of expression was unavailable. We then
eliminated all tumor samples for which no healthy controls were available. By applying such a procedure,
we built a dataset of 3,053 expression profiles (2,721 case samples and 332 control samples) of patients
affected by 10 distinct tumor pathologies (see Table 1 for more details). Case samples were further divided
by disease stage.
To run our algorithm, we performed a differentially expressed genes analysis by using the RNASeq pipeline
based on Limma [45]. The expression matrices for each disease were firstly normalized by using the Voom
algorithm [46], then a linear model was trained with Limma and differentially expressed genes for each
stage of the disease were extracted along with their Log-Fold-Change. In our analysis we considered as
differentially expressed only those genes for which an adjusted p-value was lower than 0.01 as computed
by Limma.
In order to correctly ascertain PARADIGM performance, for each tumor sample we also downloaded and
processed copy number variation (CNV) as shown in Vaske et al. 2010 [28].
Performance Assessment
To compare our algorithm with other methodologies, PARADIGM [28], SPIA [27] and Micrographite [30], we
used the decoy pathway technique introduced in Vaske et al. 2010 [28]. For each pathway in our internal
database, we built a decoy one obtained by maintaining the same structure and substituting each gene (or
miRNA) with one randomly chosen from the set of all possible genes. As in Vaske et al. 2010 [28], all the
complexes and abstract processes were kept unchanged. After the execution of the three algorithms, the
pathways were classified by each method and the fraction of real pathways versus the total number of
pathways considered was computed. The higher the fraction of real pathways, the better the ability of an
algorithm to extract biologically sound results. Lastly, to achieve a fair comparison with SPIA, we chose the
same 𝛽 function as Tarca et al. 2009 [27]: 𝛽 𝑢, 𝑛 = 1 for all interactions that increase node expression
level, 𝛽 𝑢, 𝑛 = −1 for those that have the effect of decreasing node expression level, 𝛽 𝑢, 𝑔 = 0 for
irrelevant ones. However, the 𝛽 function introduces a huge concealed potential in MITHrIL, which paves
the way for possible future extensions.
Acknowledgements
The results shown here are in whole or part based upon data generated by the TCGA Research
Network: http://cancergenome.nih.gov/. V. D. was supported by Italian Foundation for Cancer Research (FIRC)
(16572).
Conflicts of Interest
The authors declare no conflict of interest.
References
1.
Glazko GV, Emmert-Streib F. Unite and conquer: univariate and multivariate approaches for finding
differentially expressed gene sets. Bioinformatics. 2009;25(18):2348-2354.
2.
Green ML, Karp PD. The outcomes of pathway database computations depend on pathway
ontology. Nucl Acids Res. 2006;34(13):3687-3697.
3.
Khatri P, Sirota M, Butte AJ. Ten years of pathway analysis: current approaches and outstanding
challenges. PLoS Comput Biol. 2012;8(2):e1002375.
4.
Kanehisa M, Goto S. KEGG: kyoto encyclopedia of genes and genomes. Nucl Acids Res.
2000;28(1):27-30.
5.
Kanehisa M, Goto S, Sato Y, Kawashima M, Furumichi M, Tanabe M. Data, information, knowledge
and principle: back to metabolism in KEGG. Nucl Acids Res. 2014;42(D1):D199-D205.
6.
Cerami EG, Gross BE, Demir E, Rodchenkov I, Babur Ö, Anwar N, Schultz N, Bader GD, Sander C.
Pathway Commons, a web resource for biological pathway data. Nucl Acids Res. 2011;39(suppl
1):D685-D690.
7.
Khatri P, Draghici S, Ostermeier GC, Krawetz SA. Profiling gene expression using onto-express.
Genomics. 2002;79(2):266-270.
8.
Drǎghici S, Khatri P, Martins RP, Ostermeier GC, Krawetz SA. Global functional profiling of gene
expression. Genomics. 2003;81(2):98-104.
9.
Berriz GF, King OD, Bryant B, Sander C, Roth FP. Characterizing gene sets with FuncAssociate.
Bioinformatics. 2003;19(18):2502-2504.
10.
Beißbarth T, Speed TP. GOstat: find statistically overrepresented Gene Ontologies within a group of
genes. Bioinformatics. 2004;20(9):1464-1465.
11.
Castillo-Davis CI, Hartl DL. GeneMerge—post-genomic analysis, data mining, and hypothesis testing.
Bioinformatics. 2003;19(7):891-892.
12.
Martin D, Brun C, Remy E, Mouren P, Thieffry D, Jacq B. GOToolBox: functional analysis of gene
datasets based on Gene Ontology. Genome biology. 2004;5(12):R101.
13.
Doniger SW, Salomonis N, Dahlquist KD, Vranizan K, Lawlor SC, Conklin BR, others. MAPPFinder:
using Gene Ontology and GenMAPP to create a global gene-expression profile from microarray data.
Genome Biol. 2003;4(1):R7.
14.
Vlachos IS, Zagganas K, Paraskevopoulou MD, Georgakilas G, Karagkouni D, Vergoulis T, Dalamagas
T, Hatzigeorgiou AG. DIANA-miRPath v3.0: deciphering microRNA function with experimental
support. Nucl Acids Res. 2015;43(W1):W460-W466. doi:10.1093/nar/gkv403.
15.
Kong SW, Pu WT, Park PJ. A multivariate approach for integrating genome-wide expression data and
biological knowledge. Bioinformatics. 2006;22(19):2373-2380.
16.
Tian L, Greenberg SA, Kong SW, Altschuler J, Kohane IS, Park PJ. Discovering statistically significant
pathways in expression profiling studies. Proc Natl Acad Sci USA. 2005;102(38):13544-13549.
17.
Jiang Z, Gentleman R. Extensions to gene set enrichment. Bioinformatics. 2007;23(3):306-313.
18.
Lu Y, Liu P-Y, Xiao P, Deng H-W. Hotelling's T2 multivariate profiling for detecting differential
expression in microarrays. Bioinformatics. 2005;21(14):3105-3113.
19.
Xiong H. Non-linear tests for identifying differentially expressed genes or genetic networks.
Bioinformatics. 2006;22(8):919-923.
20.
Hummel M, Meister R, Mansmann U. GlobalANCOVA: exploration and assessment of gene group
effects. Bioinformatics. 2008;24(1):78-85.
21.
Klebanov L, Glazko G, Salzman P, Yakovlev A, Xiao Y. A multivariate extension of the gene set
enrichment analysis. Journal of bioinformatics and computational biology. 2007;5(05):1139-1153.
22.
Goeman JJ, Bühlmann P. Analyzing gene expression data in terms of gene sets: methodological
issues. Bioinformatics. 2007;23(8):980-987.
23.
Ackermann M, Strimmer K. A general modular framework for gene set enrichment analysis. BMC
Bioinformatics. 2009;10(1):47.
24.
Efron B, Tibshirani R. On testing the significance of sets of genes. The annals of applied statistics.
2007:107-129.
25.
Mitrea C, Taghavi Z, Bokanizad B, Hanoudi S, Tagett R, Donato M, Voichiţa C, Drăghici S. Methods
and approaches in the topology-based analysis of biological pathways. Frontiers in physiology.
2013;4.
26.
Draghici S, Khatri P, Tarca AL, Amin K, Done A, Voichita C, Georgescu C, Romero R. A systems biology
approach for pathway level analysis. Genome Research. 2007;17(10):1537-1545.
27.
Tarca AL, Draghici S, Khatri P, Hassan SS, Mittal P, Kim J-S, Kim CJ, Kusanovic JP, Romero R. A novel
signaling pathway impact analysis. Bioinformatics. 2009;25(1):75-82.
28.
Vaske CJ, Benz SC, Sanborn JZ, Earl D, Szeto C, Zhu J, Haussler D, Stuart JM. Inference of patientspecific pathway activities from multi-dimensional cancer genomics data using PARADIGM.
2010;26(12):i237-i245. doi:10.1093/bioinformatics/btq182.
29.
Sedgewick AJ, Benz SC, Rabizadeh S, Soon-Shiong P, Vaske CJ. Learning subgroup-specific regulatory
interactions and regulator independence with PARADIGM. Bioinformatics. 2013;29(13):i62-i70.
30.
Calura E, Martini P, Sales G, Beltrame L, Chiorino G, D’Incalci M, Marchini S, Romualdi C. Wiring
miRNAs to pathways: a topological approach to integrate miRNA and mRNA expression profiles.
Nucl Acids Res. 2014;42(11):e96-e96. doi:10.1093/nar/gku354.
31.
Tibshirani R, Hastie T, Narasimhan B, Chu G. Diagnosis of multiple cancer types by shrunken
centroids of gene expression. Proceedings of the National Academy of Sciences. 2002;99(10):65676572.
32.
Calin GA, Dumitru CD, Shimizu M, Bichi R, Zupo S, Noch E, Aldler H, Rattan S, Keating M, Rai K,
others. Frequent deletions and down-regulation of micro-RNA genes miR15 and miR16 at 13q14 in
chronic lymphocytic leukemia. Proceedings of the National Academy of Sciences.
2002;99(24):15524-15529.
33.
Acunzo M, Romano G, Wernicke D, Croce CM. MicroRNA and cancer--A brief overview. Advances in
biological regulation. 2015;57:1-9.
34.
Costinean S, Sandhu SK, Pedersen IM, Tili E, Trotta R, Perrotti D, Ciarlariello D, Neviani P, Harb J,
Kauffman LR, others. Src homology 2 domain--containing inositol-5-phosphatase and CCAAT
enhancer-binding protein β are targeted by miR-155 in B cells of Eμ-MiR-155 transgenic mice. Blood.
2009;114(7):1374-1382.
35.
Hsu S-D, Lin F-M, Wu W-Y, Liang C, Huang W-C, Chan W-L, Tsai W-T, Chen G-Z, Lee C-J, Chiu C-M,
others. miRTarBase: a database curates experimentally validated microRNA--target interactions.
Nucl Acids Res. 2010:gkq1107.
36.
Xiao F, Zuo Z, Cai G, Kang S, Gao X, Li T. miRecords: an integrated resource for microRNA--target
interactions. Nucl Acids Res. 2009;37(suppl 1):D105-D110.
37.
Wang J, Lu M, Qiu C, Cui Q. TransmiR: a transcription factor--microRNA regulation database. Nucl
Acids Res. 2010;38(suppl 1):D119-D122.
38.
Griffiths-Jones S. The microRNA registry. Nucl Acids Res. 2004;32(suppl 1):D109-D111.
39.
Griffiths-Jones S, Grocock RJ, Van Dongen S, Bateman A, Enright AJ. miRBase: microRNA sequences,
targets and gene nomenclature. Nucl Acids Res. 2006;34(suppl 1):D140-D144.
40.
Griffiths-Jones S, Saini HK, van Dongen S, Enright AJ. miRBase: tools for microRNA genomics. Nucl
Acids Res. 2008;36(suppl 1):D154-D158.
41.
Kozomara A, Griffiths-Jones S. miRBase: integrating microRNA annotation and deep-sequencing
data. Nucl Acids Res. 2010:gkq1027.
42.
Kozomara A, Griffiths-Jones S. miRBase: annotating high confidence microRNAs using deep
sequencing data. Nucl Acids Res. 2013:gkt1181.
43.
Cormen TH, Leiserson CE, Rivest RL, Stein C, others. Introduction to Algorithms. Vol 2. MIT press
Cambridge; 2001.
44.
Benjamini Y, Yekutieli D. The control of the false discovery rate in multiple testing under
dependency. Annals of statistics. 2001:1165-1188.
45.
Ritchie ME, Phipson B, Wu D, Hu Y, Law CW, Shi W, Smyth GK. limma powers differential expression
analyses for RNA-sequencing and microarray studies. Nucl Acids Res. 2015:gkv007.
46.
Law CW, Chen Y, Shi W, Smyth GK. Voom: precision weights unlock linear model analysis tools for
RNA-seq read counts. Genome Biol. 2014;15(2):R29.
Tables
Code
Cancer Type
Control Samples Case Samples
BLCA
Bladder Urothelial Carcinoma
19
193
Stage I, II, III, IV
BRCA
Breast invasive carcinoma
86
642
Stage I, II, III, IV, X
COAD Colon adenocarcinoma
8
389
Stage I, II, III, IV
KICH
Kidney Chromophobe
25
66
Stage I, II, III, IV
KIRC
Kidney renal clear cell carcinoma
71
224
Stage I, II, III, IV
LUAD Lung adenocarcinoma
19
388
Stage I, II, III, IV
LUSC
37
247
Stage I, II, III, IV
PRAD Prostate adenocarcinoma
50
191
Category 6, 7, 8, 9, 10
READ Rectum adenocarcinoma
3
150
Stage I, II, III, IV
UCEC
Uterine Corpus Endometrial Carcinoma
14
231
Stage I, II, III, IV
All Samples
332
2721
Lung squamous cell carcinoma
Case Samples Categories
Table 1 List of cancer types extracted from The Cancer Genome Atlas (TCGA) with their codes, number of case and control samples,
and Subcategories.
Perturb. Of Random Nodes
Endpoints
Pathway-level Statistics
MITHrIL no MITHrIL MITHrIL PARADIGM MITHrIL SPIA
miRNA
no miRNA
Acc.
Acc.
PARADIGM
Scores
Data
Log-FC
MITHrIL
BLCA
3.11%
9.59%
6.58%
1.55%
2.60%
2.08%
12.95% 49.74%
82.38%
BRCA
1.86%
2.12%
3.97%
1.09%
2.00%
2.34%
13.08% 8.25%
73.05%
COAD
2.31%
0.00%
7.81%
0.00%
0.00%
3.10%
0.77%
0.00%
32.90%
KICH
3.03%
1.67%
3.03%
0.00%
0.00%
3.03%
4.54%
3.03%
31.81%
KIRC
3.12%
2.68%
2.77%
1.79%
2.68%
3.13%
5.80%
2.67%
35.26%
LUAD
4.89%
0.03%
8.61%
1.80%
2.06%
5.41%
4.38%
2.83%
64.43%
LUSC
6.07%
1.78%
6.92%
1.21%
2.02%
6.91%
5.26%
4.04%
71.54%
PRAD
0.00%
0.37%
1.26%
0.00%
0.00%
0.52%
2.61% 30.89%
18.94%
READ
3.33%
0.00%
9.40%
0.00%
0.00%
4.00%
0.66%
0.00%
96.66%
UCEC
1.73%
1.39%
1.13%
0.00%
0.43%
0.09%
4.32%
1.29%
46.32%
Total
2.90%
1.75%
5.38%
0.90%
1.50%
3.20%
6.40%
8.80%
57.60%
Table 2 Classification results of tumor samples in our dataset obtained training PAMR algorithm by means of Log-Fold-Change, SPIA
total accumulation, Paradigm scores, MITHrIL accumulators, and MITHrIL endpoint perturbations. Each element in the table
corresponds to the classification error for a specific cancer type using one algorithm. Despite the reference classification based on
Log-Fold-Change yields a low average error (2.90%), the employment of perturbations computed for each endpoint provides a
significant improvement in the classification accuracy.
Figures
Figure 1 Performances comparison between MITHrIL, SPIA, and PARADIGM by means of the average area under the ROC curves.
Each box in the figure represents the variability range of AUC values for a specific methodology.
Figure 2 Significance of the addition of miRNA in our model by means of a comparison of the percentages of correctly predicted
endpoints for each sample between our method with and without miRNAs. Each box in the figure represents the variability range of
the percentage of correctly predicted endpoints for the patients of a specific tumor type. A prediction is correct when the
deregulation observed in the original data correspond to the one inferred by our algorithm. Namely, the sign of an endpoint logFold-Change corresponds to the sign of its perturbation value.
| 5cs.CE
|
Gene Selection With
Guided Regularized Random Forest
arXiv:1209.6425v3 [cs.LG] 20 Jun 2013
Houtao Deng
Intuit, Mountain View, CA, USA
George Runger
Arizona State University, Tempe, AZ, USA
Abstract
The regularized random forest (RRF) was recently proposed for feature selection by building only one ensemble. In RRF the features are evaluated
on a part of the training data at each tree node. We derive an upper bound
for the number of distinct Gini information gain values in a node, and show
that many features can share the same information gain at a node with a
small number of instances and a large number of features. Therefore, in a
node with a small number of instances, RRF is likely to select a feature not
strongly relevant.
Here an enhanced RRF, referred to as the guided RRF (GRRF), is proposed. In GRRF, the importance scores from an ordinary random forest
(RF) are used to guide the feature selection process in RRF. Experiments on
10 gene data sets show that the accuracy performance of GRRF is, in general,
more robust than RRF when their parameters change. GRRF is computationally efficient, can select compact feature subsets, and has competitive
accuracy performance, compared to RRF, varSelRF and LASSO logistic regression (with evaluations from an RF classifier). Also, RF applied to the
features selected by RRF with the minimal regularization outperforms RF
applied to all the features for most of the data sets considered here. Therefore, if accuracy is considered more important than the size of the feature
This research was partially supported by ONR grant N00014-09-1-0656.
Email addresses: [email protected] (Houtao Deng), [email protected]
(George Runger)
∗
Preprint submitted to Elsevier
June 21, 2013
subset, RRF with the minimal regularization may be considered. We use the
accuracy performance of RF, a strong classifier, to evaluate feature selection
methods, and illustrate that weak classifiers are less capable of capturing the
information contained in a feature subset. Both RRF and GRRF were implemented in the “RRF” R package available at CRAN, the official R package
archive.
Keywords: classification; feature selection; random forest; variable
selection.
1. Introduction
Given a training data set consisting of N instances, P predictor variables/features Xi (i = 1, ..., P ) and the class Y ∈ {1, 2, ..., C}, the objective
of feature selection is to select a compact variable/feature subset without
loss of predictive information about Y . Note feature selection selects a subset of the original feature set, and, therefore, may be more interpretable than
feature extraction (e.g., principal component analysis [15] and partial least
squares regression [8]) which creates new features based on transformations
or combinations of the original feature set [13]. Feature selection has been
widely used in many applications such as gene selection [23, 30, 18] as it can
moderate the curse of dimensionality, improve interpretability [10] and avoid
the effort to analyze irrelevant or redundant features.
Information-theoretic measures such as symmetrical uncertainty and mutual information can measure the degree of association between a pair of
variables and have been successfully used for feature selection, e.g., CFS
(correlation-based feature selection) [12] and FCBF (fast correlation-based
filter) [29]. However, these measures are limited to two variables and do
not capture high-order interactions between variables well. For example, the
measures can not capture the exclusive OR relationship Y = XOR(X1 , X2 ),
in which neither X1 nor X2 is predictive individually, but X1 and X2 together
can correctly determine Y [14].
LASSO logistic regression [25] and recursive feature elimination with a
linear SVM (SVM-RFE) [11] are well-known feature selection methods based
on classifiers. These methods assume a linear relationship between the log
odds of the class and the predictor variables (LASSO logistic regression) or
between the class and the predictor variables (linear SVM). Furthermore,
before using these methods, one often needs to preprocess the data such as
2
transforming categorical variables to binary variables or normalizing variables
of different scales.
A random forest (RF)[3] classifier has been commonly used for measuring
feature importance [17]. An RF naturally handles numerical and categorical
variables, different scales, interactions and nonlinearities, etc. Although the
RF feature importance scores can be used to select K features with the
highest importance scores individually, there could be redundancy among
the K features. Consequently, the selected features subset can differ from
the best combination of K-features, that is, the best feature subset. Similarly,
Boruta [22], a method based on RF, aims to select a set of relevant features,
which is different from the objective of a relevant and also non-redundant
feature subset.
A feature selection method based on RF, varSelRF[27], has become popular. varSelRF consists of multiple iterations and eliminates the feature(s)
with the least importance score(s) at each iteration. Since eliminating one
feature at each iteration is computationally expensive, the authors considered
eliminating a fraction, e.g., 1/5, of the features at each iteration. However,
when there is a large number of features, many features are eliminated at
each iteration, and, thus, useful features, but with small importance scores,
can be eliminated.
The ACE algorithm [26] is another ensembles-based feature selection
method. It was shown to be effective, but it is more computationally demanding than the simpler approaches considered here. It requires multiple
forests to be constructed, along with multiple gradient boosted trees [6].
Recently, the regularized random forest (RRF) was proposed for feature
selection with one ensemble [4], instead of multiple ensembles [27, 26]. However, in RRF the features are evaluated on a part of the training data at each
tree node and the feature selection process may be greedy.
Here we analyze a feature evaluation issue that occurs in all the usual
splitting algorithms at tree nodes with a small number of training instances.
To solve this issue, we propose the guided RRF (GRRF) method, in which the
importance scores from an ordinary RF are used to guide the feature selection
process in RRF. Since the importance scores from an RF are aggregated from
all the trees based on all the training data, GRRF is expected to perform
better than RRF.
Section 2 presents previous work. Section 3 discusses the node sparsity
issue when evaluating features at tree nodes with a small number of training
instances. Section 4 describes the GRRF method. Section 5 presents and
3
discusses the experimental results. Section 6 concludes this work.
2. Background
2.1. Variable importance scores from Random Forest
A random forest (RF) [3] is a supervised learner that consists of multiple
decision trees, each of which grown on a bootstrap sample from the original
training data. The Gini index at node v, Gini(v), is defined as
Gini(v) =
C
X
c=1
p̂vc (1 − p̂vc )
where p̂vc is the proportion of class-c observations at node v. The Gini information gain of Xi for splitting node v, Gain(Xi , v), is the difference between
the impurity at the node v and the weighted average of impurities at each
child node of v. That is,
Gain(Xi , v) =
Gini(Xi , v) − wL Gini(Xi , v L ) − wR Gini(Xi , v R )
where v L and v R are the left and right child nodes of v, respectively, and wL
and wR are the proportions of instances assigned to the left and right child
nodes. At each node, a random set of mtry features out of P is evaluated,
and the feature with the maximum Gain(Xi , v) is used for splitting the node
v.
The importance score for variable Xi can be calculated as
Impi =
1 X
Gain(Xi , v)
ntree v∈S
(1)
Xi
where SXi is the set of nodes split by Xi in the RF with ntree trees. The
RF importance scores are commonly used to evaluate the contributions of
the features regarding predicting the classes.
2.2. Regularized Random Forest
The regularized random forest (RRF) [4] applies the tree regularization
framework to RF and can select a compact feature subset. While RRF is
4
Algorithm 1: Feature selection at node v.
input : F and λ
output: F
1
2
3
4
5
6
7
8
9
10
11
12
13
14
15
16
17
18
19
20
gain∗ ← 0, count ← 0, f ∗ ← −1, f1 ← {1, 2, 3, ..., P }, f2 ← ∅
while f1 6= ∅ do
m ← select(f1 ) //randomly select feature index m from f1
f2 ← {f2 , m} //add m to f2
f1 ← f1 − m //remove m from f1
gainR (Xm , v) ← 0
if Xm ∈ F then
gainR (Xm , v) ← gain(Xm , v) //calculate gainR for all variables in F
end
√
if Xm ∈
/ F and count < ⌈ P ⌉ then
gainR (Xm ) ← λ · gain(Xm ) //regularize the gain if the variable is not in F
count ← count + 1
end
if gainR (Xm , v) > gain∗ then
gain∗ ← gainR (Xm , v), f ∗ ← m
end
end
if f ∗ 6= −1 and f ∗ ∈
/ F then
F ← {F, f ∗ }
end
return F
built in a way similar to RF, the main difference is that the regularized
information gain, GainR (Xi , v), is used in RRF
(
λ · Gain(Xi , v) i ∈
/F
(2)
GainR (Xi , v) =
Gain(Xi , v)
i∈F
where F is the set of indices of features used for splitting in previous nodes
and is an empty set at the root node in the first tree. Here λ ∈ (0, 1] is
called the penalty coefficient. When i ∈
/ F , the coefficient penalizes the ith
feature for splitting node v. A smaller λ leads to a larger penalty. RRF uses
GainR (Xi , v) at each node, and adds the index of a new feature to F if the
feature adds enough predictive information to the selected features.
RRF with λ = 1, referred to as RRF(1), has the minimum regularization.
Still, a new feature has to be more informative at a node than the features
already selected to enter the feature subset. The feature subset selected by
RRF(1) is called the least regularized subset, indicating the minimal regularization from RRF.
Figure 1 illustrates the feature selection process. The nodes in the forest
are visited sequentially (from the left to the right and from the top to the
5
bottom). The indices of three distinct features used for splitting are added
to F in Tree 1, and the index of X5 is added to F in Tree 2. Algorithm 1
shows the feature selection process at each node.
F={}+1
F={1,3,4}
X1
X3
F={1,4}+3
F={1}+4
X4
F={1,3,4}+5
X3
X5
Tree 1
...
Tree 2
Figure 1: The feature selection procedure of RRF. The non-leaf nodes are marked with
the splitting variables. Three distinct features used for splitting are added to F in Tree 1,
and one feature X5 is added to F in Tree 2.
Similar to GainR (·), a penalized information gain was used to suppress
spurious interaction effects in the rules extracted from tree models [7]. The
objective of Friedman and Popescu [7] was different from the goal of a compact feature subset here. Also, the regularization used in Friedman and
Popescu [7] only reduces the redundancy in each path from the root node
to a leaf node, but the features extracted from such tree models can still be
redundant.
3. The Node Sparsity Issue
This section discusses an issue of evaluating features in a tree node with
a small number of instances, referred to as the node sparsity issue here. The
number of instances decreases as the instances are split recursively in a tree,
and, therefore, this issue can commonly occur in tree-based models.
In a tree node, the Gini information gain or other information-theoretic
measures are commonly used to evaluate and compare different features.
However, the measures calculated from a node with a small number of instances may not be able to distinguish the features with different predictive
information. In the following we establish an upper bound for the number
of distinct values of Gini information gain for binary classification.
Let D(f ) denote the number of distinct values of a function f (defined
over a specified range), N denote the number of instances at a non-leaf node
v and N ≥ 2 (otherwise it cannot be split). For simplicity, in the following
6
we also assume N is even. The following procedure is similar when N is odd.
Let N1 and N2 be the number of instances of class 1 and class 2, respectively,
at the node. Let L denote the number of instances at the left child node,
and let L1 and L2 denote the number of instances for class 1 and class 2
at the left child node, respectively, with similar notations R, R1 and R2 for
the right child node. Note L ≥ 1 and R ≥ 1. The notation L1 L2 denotes
the product function where L1 and L2 assume values in the feasible domain
0 ≤ L1 , L2 ≤ L and L1 + L2 = L. Then we have the following lemmas and
theorem.
Lemma 1. An upper bound of the number of distinct values of L1 L2 is ⌈(L+
1)/2⌉. That is, D(L1 L2 ) ≤ ⌈(L + 1)/2⌉, where ⌈·⌉ denotes the ceiling.
Proof. L1 has at most L + 1 values (L1 ∈{0,1,...,L}). Now let l1 l2 and m1 m2
be two realizations of L1 L2 , and let l1 l2 = m1 m2 , we have
(L − l1 )l1
= (L − m1 )m1
2
2
⇔
Ll1 − l1 − Lm1 + m1
=
0
⇔ L(l1 − m1 ) + (m1 + l1 )(m1 − l1 ) =
0
⇔
(l1 − m1 )(L − l1 − m1 )
=
0
⇔
l1 = m1 or l1 = L − m1
Therefore, we obtain an upper bound for the number of distinct values for
L1 L2 as D(L1 L2 ) <= ⌈(L + 1)/2⌉.
Lemma 2. D(L1 L2 /L) ≤ N(N + 2)/4 − 1.
Proof. For each L, L1 L2 /L has at most D(L1 L2 ) distinct values. Because
1 ≤ L ≤ N − 1, an upper bound for D(L1 L2 /L) is derived as follows
D(L1 L2 /L) ≤ ⌈2/2⌉ + ⌈3/2⌉ + ... + ⌈N/2⌉ = N(N + 2)/4 − 1
Lemma 3. D(L1 L2 /L + R1 R2 /R) ≤ N(N + 2)/4 − 1
Proof. Because L1 L2 /L and R1 R2 /R are symmetric, the upper bound is the
same as for one term. That is, D(L1 L2 /L + R1 R2 /R) ≤ N(N + 2)/4 − 1
Theorem 1. For binary classification, an upper bound of the number of
distinct information gain values is N(N + 2)/4 − 1.
7
Proof. The information gain for spliting node v is
Gain(v) = Gini(v) − wL Gini(v L ) − wR Gini(v R )
Because Gini(v) is a constant at node v, we only need to consider D(wL Gini(v L )+
wR Gini(v R )). For two classes
wL Gini(v L )
= (L/N)(L1 /L · L2 /L + L2 /L · L1 /L)
= 2(L/N)(L1 L2 )/L2
Then
wL Gini(v L ) + wR Gini(v R )
= 2(L/N)(L1 L2 )/L2 + 2(R/N)(R1 R2 )/R2
= (2/N)(L1 L2 /L) + (2/N)(R1 R2 /R)
Because N is a constant, we have D(wL Gini(v L )+wR Gini(v R )) = D(L1 L2 /L+
R1 R2 /R). According to Lemma 3, D(L1 L2/L + R1 R2 /R) ≤ N(N + 2)/4 −
1.
Note similar conclusions may be applied to other information-theoretic
measures such as the regularized information gain in Equation 2.
Consequently, when N is small, the number of distinct Gini information
gain values is small. For a large number of variables, many could have the
same Gini information gain. For example, at a node with 10 instances there
are at most 29 distinct Gini information gain values for binary classification
problems. For 1000 genes with two classes, there are at least 1000-29=971
genes having the same information gain as other genes. The number of
instances can be even smaller than 10 in a node of an RF or RRF as each
tree is grown completely.
In RRF, a feature with the maximum regularized information gain is
added at a node based on only the instances at that node. When multiple
features have the maximum regularized information gain, one of these features is randomly selected. As discussed above, at a node with only a small
number of instances less relevant, or redundant features may be selected. An
additional metric is useful to help distinguish the features. In the following
section, we introduce the guided regularized random forest that leverages the
importance scores calculated from an RF based on all the training data.
8
4. Guided Regularized Random Forest
The guided RRF (GRRF) uses the importance scores from a preliminary
RF to guide the feature selection process of RRF. Because the importance
scores from an RF are aggregated from all the trees of the RF based on all the
training data, GRRF may be able to handle the node sparsity issue discussed
in the previous section.
A normalized importance score is defined as
Imp′i =
Impi
maxPj=1 Impj
(3)
where Impi is the importance score from an RF (defined in Equation 1).
Here 0 ≤ Imp′i ≤ 1. Instead of assigning the same penalty coefficient to all
features in RRF, GRRF assigns a penalty coefficient to each feature. That
is,
(
λi Gain(Xi , v) Xi ∈
/F
(4)
GainR (Xi , v) =
Gain(Xi , v)
Xi ∈ F
where λi ∈ (0, 1] is the coefficient for Xi (i ∈ {1, ..., P }) and is calculated
based on the importance score of Xi from an ordinary RF. That is,
λi = (1 − γ)λ0 + γImp′i
(5)
where λ0 ∈ (0, 1] controls the degree of regularization and is called the base
coefficient, and γ ∈ [0, 1] controls the weight of the normalized importance
score and is called the importance coefficient. Note that RRF is a special case
of GRRF with γ = 0. Given λ0 and γ, a feature with a larger importance
score has a larger λi , and, therefore, is penalized less.
In our experiments we found the feature subset size can be effectively
controlled by changing either λ0 or γ, but changing the latter often leads
to better performance in terms of classification accuracy. To reduce the
number of parameters of GRRF, we fix λ0 to be 1 and consider γ as the only
parameter for GRRF. With λ0 = 1, we have
λi = (1 − γ) + γImp′i = 1 − γ(1 − Imp′i )
(6)
For a feature Xi that does not have the maximum importance score (Imp′i 6=
1), a larger γ leads to a smaller λi and, thus, a larger penalty on Gain(Xi , v)
when Xi has not been used in the nodes prior to node v. Consequently, γ is
essentially the degree of regularization. Furthermore, GRRF with γ = 0 is
equivalent to RRF(1), with the minimal regularization.
9
5. Experiments
We implemented the RRF and GRRF algorithms in the “RRF” R package
based on the “randomForest” package [17]. The “RRF” package is available
from CRAN (http://cran.r-project.org/), the official R package archive.
Work by [27] showed that the accuracy performance of RF is largely
independent
√ of the number of trees (between 1000 and 40000 trees), and
mtry = P is√often a reasonable choice. Therefore, we used ntree = 1000
and mtry = P for the classifier RF and the feature selection methods
RRF, GRRF and varSelRF. Guided by initial experiments, randomly sampling 63% of the data instances (the same as the default setting in the “randomForest” package) without replacement was used in RRF and GRRF. We
also evaluated two well-known feature selection methods: varSelRF available
in the “varSelRF” R package and LASSO logistic regression available in the
“glmnet” R package. Unless otherwise specified, the default values in the
packages were used. Also note λ is the parameter of RRF, and as discussed,
γ is considered as the only parameter of GRRF here.
Table 1: The number of groups identified, and the number of irrelevant or redundant
features selected for different algorithms.
Method
Average
Std. err
LASSO
4
0.000
# Groups
varSelRF
RRF
4.95
5
0.050
0.000
GRRF
4.95
0.050
# Irrelevant or Redundant Features
LASSO
varSelRF
RRF
GRRF
4.3
4.95
4.95
0.75
0.147
0.050
0.050
0.123
5.1. Simulated Data Sets
We start with a simulated data set generated by the following procedure. First generate 10 independent variables: X1 ,..., X10 , each is uniformly
distributed in the interval [0,1]. Variable Y is calculated by the formula
Y = 10 sin(πX1 X2 ) + 20(X3 − 0.5)2 + 10X4 + 5X5 + e
where e follows a standard normal distribution. Note the above data generation procedure was described in [5]. We simulated 1000 instances and
then calculated the median of Y as y. We then labeled class 2 to the instances with Y > y, and class 1 otherwise, so that it becomes a classification problem. Furthermore, we added five other variables with X11 =X1 ,
X12 =X2 , X13 =X3 , X14 =X4 , X15 =X5 . Consequently, the feature selection
solution is {(X1 |X11 ) & (X2 |X12 ) & (X3 |X13 ) & (X4 |X14 ) & (X5 |X15 )},
10
Table 2: Summary of the data sets.
Data set
adenocarcinoma
brain
breast.2.class
breast.3.class
colon
leukemia
lymphoma
nci60
prostate
srbct
Reference
[20]
[19]
[28]
[28]
[2]
[9]
[1]
[21]
[24]
[16]
# Examples
76
42
77
95
62
38
62
61
102
63
# Features
9868
5597
4869
4869
2000
3051
4026
5244
6033
2308
# classes
2
5
2
3
2
2
3
8
2
4
where & stands for “and” and (X1 |X11 ) indicates that one and only one
from this group should be selected. For example, for the data set considered
here, both (X1 , X2 , X3 , X4 , X5 ) and (X11 , X2 , X13 , X4 , X5 ) are correct solutions. However, (X1 , X3 , X4 , X5 ) misses the group (X2 |X12 ). Furthermore,
(X1 , X2 , X3 , X4 , X5 , X11 ) has a redundant variable because X11 is redundant
to X1 .
We simulated 20 replicates of the above data set, and then applied GRRF,
RRF and two well-known methods varSelRF and LASSO logistic regression
to the data sets. Here γ of GRRF was selected from {0.4,0.5,0.6}, λ of RRF
was selected from {0.6,0.7,0.8}, and the regularization parameter of LASSO
logistic regression was selected from {0.01,0.02,0.03,0.05,0.1,0.2}, all by 10fold CV.
The results are shown in Table 1. LASSO logistic regression identifies the
least number of groups on average (4). The methods varSelRF, RRF and
GRRF identify almost all the groups, but varSelRF and RRF select 4.95,
on average, irrelevant or redundant variables, and GRRF only selects 0.75
irrelevant or redundant variables. This experiment shows GRRF’s potential
in selecting a relevant and non-redundant feature subset.
5.2. Gene Data Sets
The 10 gene expression data sets analyzed by Uriarte and de Andres [27]
are considered in this section. The data sets are summarized in Table 2. For
each data set, a feature selection algorithm was applied to 2/3 of the instances
(selected randomly) to select a feature subset. Then a classifier was applied
to the feature subset. The error rate is obtained by applying the classifier to
the other 1/3 of the instances. This procedure was conducted 100 times with
different random seeds, and the average size of feature subsets and the average
error rate over 100 runs were calculated. In the experiments, we considered
11
Table 3: The number of features selected in the least regularized subset selected by RRF
(i.e., RRF(1)), the total number of original features (“All”), and the error rates of RF
applied to the least regularized subset and all the features. The win-lose-tie results of
“All” compared to RRF(1) are shown. Here “◦” or “•” represents a significant difference
at the 0.05 level, according to the paired t-test. RRF(1) uses many fewer features than
the original features, and wins on 7 data sets. There are significant differences for four
data sets.
adenocarcinoma
brain
breast.2.class
breast.3.class
colon
leukemia
lymphoma
nci
prostate
srbct
win-lose-tie
Number of features
RRF(1)
All
86
9868
97
5597
210
4869
253
4869
92
2000
24
3051
31
4026
197
5244
88
6033
51
2308
-
Average error rates
RRF(1)-RF
All-RF
0.158
0.159
0.159
0.170
0.352
0.371
0.397
0.415
0.162
0.158
0.053
0.064
0.018
0.006
0.332
0.321
0.082
0.109
0.027
0.032
3-7-0
◦
◦
•
◦
Table 4: The total number of original features (“All”) and the average number of features
(from 100 replicates for each data set) selected by different methods. All feature selection
methods are able to select a small number of features. Here GRRF(0.1) and RRF(0.9)
select a similar number of features, but it is shown later that GRRF(0.1) is more accurate
than RRF(0.9) (with an RF classifier) for most data sets.
adenocarcinoma
brain
breast.2.class
breast.3.class
colon
leukemia
lymphoma
nci
prostate
srbct
All
9868
5597
4869
4869
2000
3051
4026
5244
6033
2308
GRRF(0.1)
20
22
59
77
29
6
5
63
18
13
GRRF(0.2)
15
12
25
31
13
4
4
26
13
9
RRF(0.9)
23
27
60
78
27
6
4
61
19
14
varSelRF
4
28
7
12
4
2
81
60
6
34
LASSO
2
24
7
18
7
8
25
53
12
28
RRF, GRRF, varSelRF and LASSO logistic regression as the feature selection
algorithms, and random forest (RF) and C4.5 as the classifiers.
5.2.1. Feature selection and classification
First compare the accuracy of RF applied to all the features, denoted as
“All”, and the least regularized subset (i.e., features selected by RRF(1)).
The number of features and the error rates are shown in Table 3. The winlose-tie results of “All” compared to RRF(1) are also shown in the tables. To
investigate the statistical significance of these results, we applied the paired
12
Table 5: The error rates of RF applied to the feature subsets selected by GRRF(0.1),
GRRF(0.2), RRF(0.9), varSelRF and LASSO logistic regression, respectively (to three
decimal places). The win-lose-tie results of each competitor compared to GRRF(0.1) with
RF are shown. Here “◦” or “•” represents a significant difference between a method and
GRRF(0.1) with RF at the 0.05 level, according to the paired t-test. Here GRRF(0.1)
leads to competitive accuracy performance, compared to GRRF(0.2), RRF(0.9), LASSO
and varSelRF.
adenocarcinoma
brain
breast.2.class
breast.3.class
colon
leukemia
lymphoma
nci
prostate
srbct
win-lose-tie
GRRF(0.1)
-RF
0.169
0.214
0.345
0.387
0.175
0.080
0.067
0.389
0.085
0.064
-
GRRF(0.2)
-RF
0.168
0.259
0.359
0.403
0.186
0.093
0.076
0.452
0.085
0.072
1-9-0
◦
◦
◦
◦
◦
RRF(0.9)
-RF
0.160
0.234
0.367
0.410
0.190
0.091
0.098
0.405
0.101
0.074
1-9-0
◦
◦
◦
◦
◦
varSelRF
-RF
0.212
0.231
0.386
0.418
0.232
0.107
0.022
0.418
0.085
0.035
3-7-0
◦
◦
◦
◦
◦
•
◦
•
LASSO
-RF
0.189
0.259
0.366
0.400
0.180
0.076
0.009
0.396
0.088
0.007
3-7-0
◦
◦
◦
•
•
Table 6: The error rates of C4.5 applied to the feature subsets selected by GRRF(0.1),
GRRF(0.2), RRF(0.9), varSelRF and LASSO logistic regression, respectively. The winlose-tie results of each competitor compared to GRRF(0.1) with C4.5 are calculated. Here
“◦” or “•” represents a significant difference between a method and GRRF(0.1) with
C4.5 at the 0.05 level, according to the paired t-test. The GRRF methods and the other
methods have perform similarly in terms of the accuracy with C4.5. As expected, C4.5
has noticeably higher error rates than RF, shown in Table 5.
adenocarcinoma
brain
breast.2.class
breast.3.class
colon
leukemia
lymphoma
nci
prostate
srbct
win-lose-tie
GRRF(0.1)
-C4.5
0.260
0.438
0.402
0.497
0.274
0.133
0.126
0.626
0.157
0.235
-
GRRF(0.2)
-C4.5
0.251
0.428
0.411
0.494
0.262
0.157
0.125
0.636
0.138
0.208
7-3-0
◦
•
•
13
RRF(0.9)
-C4.5
0.241
0.418
0.412
0.491
0.298
0.138
0.126
0.658
0.154
0.207
5-5-0
◦
◦
•
varSelRF
-C4.5
0.248
0.394
0.413
0.481
0.273
0.146
0.152
0.641
0.115
0.209
6-4-0
•
◦
•
•
LASSO
-C4.5
0.218
0.476
0.391
0.482
0.242
0.166
0.138
0.646
0.144
0.194
6-4-0
•
◦
•
◦
•
t-test to the error rates of the two methods from 100 replicates for each data
set. The data sets with a significant difference at the 0.05 level are marked
with “◦” or “•” in the table.
The least regularized subset not only has many fewer features than “All”,
but also leads to better accuracy performance on 7 data sets out of 10, and
the error rates are significantly different on 4 data sets. Therefore, RRF(1)
not only improves interpretability by reducing the number of features, but
also can improve the accuracy performance of classification, even for RF, considered as a strong classifier capable of handling irrelevant and relevant variables [4]. It should also be noted that although the least regularized subset
is much smaller than the original feature set, the size may still be considered
large in some cases, e.g., more than 200 features for the breast.2.class data
set. GRRF and RRF with larger regularization, investigated in the following
experiments, are able to further reduce the number of features.
Next compare GRRF to RRF and two well-known methods: varSelRF
and LASSO logistic regression. The regularization parameter of LASSO logistic regression was selected from {0.01, 0.02, 0.03, 0.05, 0.1, 0.2} by 10-fold
CV. Here γ ∈ {0.1, 0.2} was used for GRRF (i.e., GRRF(0.1) or GRRF(0.2)),
and λ = 0.9 was used for RRF (i.e., RRF(0.9)). We used a fixed parameter
setting for GRRF or RRF, as the parameter sensitivity analysis in the following section shows a consistent trend that GRRF or RRF tends to select more
features and also tends to be more accurate, for a smaller γ or a larger λ. We
chose these parameters so that a reasonably small number of features can be
selected. One can also use cross-validation error to determine an appropriate
parameter value customized for each data set and potentially improve these
results.
The total number of original features and the average number of features
selected by each feature selection method are shown in Table 4. All the
feature selection methods are able to select a small number of features.
The average error rates of RF applied to all the features (“All”) and the
subsets selected by different feature selection methods are shown in Table
5. GRRF(0.1) with RF outperforms GRRF(0.2) with RF on 9 data sets out
of 10, 5 of which have significant differences at the 0.05 level. Even though
GRRF(0.1) selects more features than GRRF(0.2), the sizes of the feature
subsets are reasonably small (all less than 80 features).
GRRF(0.1) and RRF(0.9) select a similar number of features. However,
GRRF(0.1) with RF outperforms RRF(0.9) with RF on 9 data sets, 5 of
which have significant differences at the 0.05 level. Consequently, GRRF
14
selects stronger features than RRF, which is consistent with the simulated
experiments. According to the discussion in Section 3, a feature with the
maximum Gini information gain in a node with a small number instances
may not be truly strong. Yet RRF adds this feature to the subset.
GRRF(0.1) with RF outperforms varSelRF with RF on 7 data sets, 6
of which have significant differences. Therefore, GRRF(0.1) may be more
favorable than varSelRF for the data sets considered here. Also, as shown in
the following section, GRRF has a clear advantage over varSelRF in terms of
computational time. Furthermore, GRRF(0.1) with RF outperforms LASSO
logistic regression with RF on 7 data sets, 3 of which have significant differences. The accuracy performance may be improved by applying a logistic
regression model to the features selected by LASSO logistic regression. However, tree models like GRRF have a few desirable properties compared to
LASSO logistic regression: they can naturally handle mixed categorical and
numerical features, and multiple classes, etc.
The average error rates of C4.5 applied to all the features (“All”) and the
subsets selected by different feature selection methods are shown in Table 6.
It can be seen that the error rates of C4.5 are clearly higher than RF shown
in Table 5. Indeed, RF has been considered as a stronger classifier than
C4.5. Interestingly, the differences between the methods in terms of C4.5
are smaller than the RF results. As mentioned by [4], a relatively weaker
classifier is less capable of capturing information from data than a stronger
classifier. Consequently, a feature subset that includes strong features, but
misses the features of small contributions may not affect the accuracy of
C4.5 much, but can affect the accuracy of RF. A weak classifier should only
be used for evaluating a feature subset if that classifier is actually used for
classification after feature selection. However, if a strong classifier is used
for classification after feature selection, or the objective is to evaluate the
information contained in the feature subset, a strong classifier should be
considered [4].
5.2.2. Parameter Sensitivity and Computational Time
We investigated the performance of RRF and GRRF with different parameter settings: λ ∈ {0.1, 0.2, 0.3, 0.4, 0.5, 0.6, 0.7, 0.8, 0.9, 1} for RRF
and γ ∈ {0.9, 0.8, 0.7, 0.6, 0.5, 0.4, 0.3, 0.2, 0.1, 0} for GRRF. The parameter
values are arranged by the degree of regularization in a decreasing order for
both methods.
The sizes of the feature subsets, averaged over 100 replicates, for each
15
500 1000
500 1000
leukemia
lymphoma
nci
prostate
srbct
100
50
# features
5
1
0.1
0.2
0.3
0.4
0.5
0.6
0.7
0.8
0.9
1
0.9
0.8
base coefficient of RRF
0.7
0.6
0.5
0.4
0.3
0.2
0.1
0
0.1
0
importance coefficient of GRRF
0.6
error rate
0.0
0.2
0.4
0.6
0.4
0.0
0.2
leukemia
lymphoma
nci
prostate
srbct
adenocarcinoma
brain
breast.2.class
breast.3.class
colon
0.8
leukemia
lymphoma
nci
prostate
srbct
0.8
adenocarcinoma
brain
breast.2.class
breast.3.class
colon
1.0
(b)
1.0
(a)
error rate
leukemia
lymphoma
nci
prostate
srbct
adenocarcinoma
brain
breast.2.class
breast.3.class
colon
10
50
1
5
10
# features
100
adenocarcinoma
brain
breast.2.class
breast.3.class
colon
0.1
0.2
0.3
0.4
0.5
0.6
0.7
0.8
0.9
1
0.9
base coefficient of RRF
0.8
0.7
0.6
0.5
0.4
0.3
0.2
importance coefficient of GRRF
(c)
(d)
Figure 2: The performance of RRF for selected values of λ and GRRF for selected values
of γ. Figure 2(a) shows the number of features selected by RRF. A smaller λ leads to
fewer features. Figure 2(b) shows the number of features selected by GRRF. A larger γ
leads to fewer features. Figure 2(c) shows the error rates of RF applied to the feature
subsets selected by RRF for different λ. The error rates tend to decrease as λ increases.
Figure 2(d) shows the error rates of RF applied to the feature subsets selected by GRRF
for different γ. The error rates tend to decrease as γ decreases, but are reasonably robust.
16
140
80
60
t
bc
sr
te
i
ta
nc
os
pr
a
ly
m
ph
om
ia
n
lo
uk
em
le
ss
.3
st
br
ea
co
la
ss
.c
la
n
.2
br
ea
st
ar
oc
ad
en
.c
br
ai
ci
no
m
a
0
20
40
Time(sec.)
100
120
RF
GRRF
varSelRF
Figure 3: Computational time of RF, GRRF and varSelRF. GRRF is more computationally efficient than varSelRF. GRRF takes about twice as much as RF as it builds two
ensembles.
parameter setting of RRF and GRRF, are shown in Figures 2(a) and 2(b),
respectively. The number of features tends to increase as λ increases for
RRF, or as γ decreases for GRRF. The consistent trend illustrates that one
can control the size of the feature subset by adjusting the parameters.
The error rates, averaged over 100 replicates, for each parameter setting of
RRF and GRRF are shown in Figures 2(c) and 2(d), respectively. In general,
the error rates tend to decrease as λ increases for RRF, or as γ decreases for
GRRF. However, for most data sets, the error rates of GRRF seem to be
reasonably robust to the changes of γ. As mentioned, RRF(1) and GRRF(0)
are equivalent, and, therefore, they have similar number of features and error
rates for every data set (differing only by random selections).
The computational time of the RF-based methods: RF, GRRF and varSelRF
is shown in Figure 3. As expected, RF was fast for training on these data
sets. GRRF builds two ensemble models, and, thus, the computational time
of GRRF is only about twice as much as RF. However, varSelRF needs to
17
build multiple ensemble models, and the computational advantage of GRRF
over varSelRF is clear.
6. Conclusions
We derive an upper bound for the number of distinct Gini information
gain values in a tree node for binary classification problems. The upper
bound indicates that the Gini information gain may not be able to distinguish
features at nodes with a small number of instances, which poses a challenge
for RRF that selects a feature only using the instances at a node. Motivated
by this node sparsity issue, we propose an enhanced method called the guided,
regularized random forest (GRRF), in which a preliminary random forest is
used to generate the initial variable importance scores to guide the regularized
feature selection process of RRF. The importance scores from the preliminary
RF help GRRF select better features when many features share the same
maximal Gini information gain at a node. For the experiments here, GRRF
is more favorable than RRF, computationally efficient, selects a small set of
features, and has competitive accuracy performance.
Feature selection eliminates irrelevant or redundant features, but also may
eliminate features of small importance. This may not affect the performance
of a weak classifier which is less capable of capturing small information, but
may affect the performance of a strong classifier such as RF [4]. Still, we
found that the least regularized subset selected by RRF with the minimal
regularization produces better accuracy performance than the complete feature set.
Finally we note that although RRF and GRRF can be used as classifiers,
they are designed for feature selection. The trees in RRF and GRRF are not
built independently as the features selected in previous trees have an impact
on the trees built later. Therefore, as a classifier, RRF or GRRF may have
a higher variance than RF because the trees are correlated. Consequently,
in this work we applied RF on the feature subset selected by GRRF or RRF
for classification.
References
[1] A.A. Alizadeh, M.B. Eisen, R.E. Davis, C. Ma, I.S. Lossos, A. Rosenwald, J.C. Boldrick, H. Sabet, T. Tran, X. Yu, J.I. Powell, L. Yang, G.E.
18
Marti, T. Moore, J. Hudson, L. Lu, D.B. Lewis, R. Tibshirani, G. Sherlock, W.C. Chan, T.C. Greiner, D.D. Weisenburger, J.O. Armitage,
R. Warnke, R. Levy, W. Wilson, M.R. Grever, J.C. Byrd, D. Botstein, P.O. Brown, L.M. Staudt, Distinct types of diffuse large B-cell
lymphoma identified by gene expression profiling., Nature 403 (2000)
503–511.
[2] U. Alon, N. Barkai, D.A. Notterman, K. Gish, S. Ybarra, D. Mack,
A.J. Levine, Broad patterns of gene expression revealed by clustering
analysis of tumor and normal colon tissues probed by oligonucleotide
arrays, Proceedings of the National Academy of Sciences of the United
States of America 96 (1999) 6745–6750.
[3] L. Breiman, Random forests, Machine Learning 45 (2001) 5–32.
[4] H. Deng, G.C. Runger, Feature selection via regularized trees, in:
The 2012 International Joint Conference on Neural Networks (IJCNN),
IEEE.
[5] J.H. Friedman, Multivariate adaptive regression splines, The annals of
statistics (1991) 1–67.
[6] J.H. Friedman, Greedy function approximation: A gradient boosting
machine, Annals of Statistics 29 (2001) 1189–1232.
[7] J.H. Friedman, B.E. Popescu, Predictive learning via rule ensembles,
Annals of Applied Statistics 2 (2008) 916–954.
[8] P. Geladi, B. Kowalski, Partial least-squares regression: a tutorial, Analytica chimica acta 185 (1986) 1–17.
[9] T.R. Golub, D.K. Slonim, P. Tamayo, C. Huard, M. Gaasenbeek, J.P.
Mesirov, H. Coller, M.L. Loh, J.R. Downing, M.A. Caligiuri, C.D.
Bloomfield, E.S. Lander, Molecular classification of cancer: class discovery and class prediction by gene expression monitoring., Science 286
(1999) 531–537.
[10] I. Guyon, A. Elisseeff, An introduction to variable and feature selection,
Journal of Machine Learning Research 3 (2003) 1157–1182.
19
[11] I. Guyon, J. Weston, S. Barnhill, V. Vapnik, Gene selection for cancer classification using support vector machines, Machine Learning 46
(2002) 389–422.
[12] M.A. Hall, Correlation-based feature selection for discrete and numeric
class machine learning, in: Proceedings of the 17th International Conference on Machine Learning, pp. 359–366.
[13] A. Jain, D. Zongker, Feature selection: Evaluation, application, and
small sample performance, Pattern Analysis and Machine Intelligence,
IEEE Transactions on 19 (1997) 153–158.
[14] A. Jakulin, I. Bratko, Analyzing attribute dependencies, Knowledge Discovery in Databases: PKDD 2003 (2003) 229–240.
[15] I. Jolliffe, MyiLibrary, Principal component analysis, volume 2, Wiley
Online Library, 2002.
[16] J. Khan, J.S. Wei, M. Ringner, L.H. Saal, M. Ladanyi, F. Westermann,
F. Berthold, M. Schwab, C.R. Antonescu, C. Peterson, P.S. Meltzer,
Classification and diagnostic prediction of cancers using gene expression
profiling and artificial neural networks, Nature Medicine 7 (2001) 673–
679.
[17] A. Liaw, M. Wiener, Classification and regression by randomforest, R
News 2 (2002) 18–22.
[18] H. Liu, L. Liu, H. Zhang, Ensemble gene selection for cancer classification, Pattern Recognition 43 (2010) 2763–2772.
[19] S.L. Pomeroy, P. Tamayo, M. Gaasenbeek, L.M. Sturla, M. Angelo,
M.E. Mclaughlin, J.Y.H. Kim, L.C. Goumnerova, P.M. Black, C. Lau,
J.C. Allen, D. Zagzag, J.M. Olson, T. Curran, C. Wetmore, J.A. Biegel,
T. Poggio, S. Mukherjee, R. Rifkin, A. Califano, G. Stolovitzky, D.N.
Louis, J.P. Mesirov, E.S. Lander, T.R. Golub, Prediction of central nervous system embryonal tumour outcome based on gene expression., Nature 415 (2002) 436–442.
[20] S. Ramaswamy, K.N. Ross, E.S. Lander, T.R. Golub, A molecular signature of metastasis in primary solid tumors., Nature genetics 33 (2003)
49–54.
20
[21] D.T. Ross, U. Scherf, M.B. Eisen, C.M. Perou, C. Rees, P. Spellman,
V. Iyer, S.S. Jeffrey, M.V. de Rijn, M. Waltham, A. Pergamenschikov,
J.C. Lee, D. Lashkari, D. Shalon, T.G. Myers, J.N. Weinstein, D. Botstein, P.O. Brown, Systematic variation in gene expression patterns in
human cancer cell lines, Nature Genetics 24 (2000) 227–35.
[22] W. Rudnicki, M. Kursa, Feature selection with the boruta package, Journal of Statistical Software 36 (2010).
[23] R. Ruiz, J. Riquelme, J. Aguilar-Ruiz, Incremental wrapper-based gene
selection from microarray data for cancer classification, Pattern Recognition 39 (2006) 2383–2392.
[24] D. Singh, P.G. Febbo, K. Ross, D.G. Jackson, J. Manola, C. Ladd,
P. Tamayo, A.A. Renshaw, A.V. D’Amico, J.P. Richie, E.S. Lander,
M. Loda, P.W. Kantoff, T.R. Golub, W.R. Sellers, Gene expression
correlates of clinical prostate cancer behavior., Cancer Cell 1 (2002)
203–209.
[25] R. Tibshirani, Regression shrinkage and selection via the lasso, Journal
of the Royal Statistical Society. Series B (Methodological) (1996) 267–
288.
[26] E. Tuv, A. Borisov, G. Runger, K. Torkkola, Feature selection with
ensembles, artificial variables, and redundancy elimination, Journal of
Machine Learning Research 10 (2009) 1341–1366.
[27] R.D. Uriarte, S.A. de Andres, Gene selection and classification of microarray data using random forest, BMC Bioinformatics 7 (2006) 3+.
[28] L.J. van ’t Veer, H. Dai, M.J. van de Vijver, Y.D. He, A.A.M. Hart,
M. Mao, H.L. Peterse, K. van der Kooy, M.J. Marton, A.T. Witteveen,
G.J. Schreiber, R.M. Kerkhoven, C. Roberts, P.S. Linsley, R. Bernards,
S.H. Friend, Gene expression profiling predicts clinical outcome of breast
cancer, Nature 415 (2002) 530–536.
[29] L. Yu, H. Liu, Efficient feature selection via analysis of relevance and
redundancy, Journal of Machine Learning Research 5 (2004) 1205–1224.
[30] Z. Zhu, Y. Ong, M. Dash, Markov blanket-embedded genetic algorithm
for gene selection, Pattern Recognition 40 (2007) 3236–3248.
21
| 5cs.CE
|
Smart Meter Privacy via the Trapdoor Channel
Miguel Arrieta∗ and Iñaki Esnaola∗§
arXiv:1708.04429v1 [cs.IT] 15 Aug 2017
∗
Dept. of Automatic Control and Systems Engineering, University of Sheffield, Sheffield S1 3JD, UK
§
Dept. of Electrical Engineering, Princeton University, Princeton, NJ 08544, USA
{marrieta2, esnaola}@sheffield.ac.uk
Abstract—A battery charging policy that provides privacy
guarantees for smart meter systems with finite capacity battery
is proposed. For this policy an upper bound on the information
leakage rate is provided. The upper bound applies for general
random processes modelling the energy consumption of the user.
It is shown that the average energy consumption of the user
determines the information leakage rate to the utility provider.
The upper bound is shown to be tight by deriving the probability
law of a random process achieving the bound.
I. I NTRODUCTION
The increasing appeal for economical and environmentallyfriendly energy calls for more efficient energy generation,
distribution, and consumption [1]. The introduction of a digital
infrastructure into the traditional power system takes steps
towards this vision by providing a cyber layer that elevates
the existing power system to a cyberphysical system. This
advanced sensing and communication infrastructure envisioned
by the smart grid enables high resolution and real time management of the processes within the grid. This application layer
also enables dynamic energy pricing, shifting user demand to
match energy generation [2]. Moreover, the introduction of
energy consumption indicators through Smart Meters (SMs)
are reported to reduce the energy consumption of the user by
up to 15% by raising awareness of the energy cost [3].
While the high-resolution information provided by the
smart grid brings clear advantages it also raises privacy and
security concerns [4], [5], [6]. By analysing the consumption
profile of a user, techniques such as non-intrusive load monitoring (NILM) [7] track and recognise appliance usage patterns
[8], [9]. Human presence, usage of individual appliances [10],
[11], and even tuned TV channel [12] are among the list of
recognizable elements [13]. This privacy breach hinders the
implementation of some of the essential components of the
smart grid [4], [5]. Within this paradigm, SMs are central
components to the dilemma posed by the need for accurate
monitoring while providing privacy. In 2009 two bills law
aimed to enforce the usage of SMs were blocked by the Senate
of the Netherlands motivated by the privacy concerns that
emerge as a result of the increased penetration of SMs [14].
There is a growing body of literature addressing the conflict
between efficient energy monitoring and privacy brought forth
by the introduction of SMs. In [15], [16] obfuscation of the
knowledge that the utility provider (UP) has about the energy
consumption of the user is studied. Indeed, in the case in which
the SM readings are the only source of information available to
the UP, obfuscation yields some degree of privacy. Obfuscation
is achieved by several mechanisms, such as aggregating the
consumption of multiple users [15], compression of the energy
consumption sequences [16] or homomorphic encryption [17]
among others. A different approach to the problem arises in
the setting in which users have access to alternative energy
sources or energy storage devices [18]. In this case, the UP
has perfect knowledge of the energy provided to the user, but
the user employs the energy storing capability of the system
to dissociate the energy consumed by the appliances from
the energy provided by the UP. In [19], [20], [21] the case
in which the user is assumed to have an alternative energy
source with instantaneous power constraints is studied. In
[20], [22], [23], [24] the user is assumed to have access to a
rechargeable battery and the energy consumption is modelled
as an independent and identically distributed (i.i.d.) random
process.
In a practical setting, the energy consumption profiles
of users exhibit non-stationary statistical structures and are
not well described by memoryless random processes [23].
Moreover, information-theoretic privacy measures for random
processes that are not i.i.d. are still not well understood [25].
The privacy utility tradeoff is characterized for stationary
Gaussian energy consumption models in [26]. A first-order
time-homogeneous Markov chain is considered in [27]. In this
work we adopt a non-probabilistic framework by modelling
the energy management system (EMS) with a finite capacity
battery as a finite state channel without probabilistic structure.
Inspired by the code construction in [28] we propose an energy
charging policy and we characterize the privacy guarantees of
the strategy for general random processes. We also particularize the analysis to the case in which the average energy
consumption of the user is known. For this case we provide an
upper bound to the amount of information that the user leaks to
the UP and show that the average energy consumption governs
the privacy that is achievable by the user. In this paper vectors
are denoted by bold font, e.g. x, random variables are denoted
by upper-case, e.g. X, and vectors of n random variables are
denoted by super-indexing the size, e.g. X n .
II.
BATTERY S YSTEM M ODEL
We consider the energy management system (EMS) operating in discrete-time illustrated in Fig. 1. The energy consumption of the user is modelled as a discrete-time random
process X1 , X2 , . . . where Xi is a random variable taking
values in X = {0, 1, . . . , α}. In this setting Xi describes the
energy consumed at time instant i ∈ N. At each time instant,
the energy consumption of the user is satisfied by requesting
energy from the UP or by discharging the battery. This decision
is taken by the energy management unit (EMU) at each time
instant based on some power consumption performance and
privacy criteria. The energy requested from the UP is also a
discrete-time random process Y1 , Y2 , ... where Yi is a random
variable taking values in Y = {0, 1, ..., γ} and describing
the energy requested from the UP at time instant i ∈ N.
Fig. 1.
Fig. 2. [29] Diagram depicting the operation of a trapdoor channel with β = 1.
Diagram of an Energy Management System with finite battery.
When Yi > Xi , the excess energy is stored in the battery.
Alternatively, when Yi ≤ Xi , the energy deficit is obtained
from the battery. We assume the UP is able to satisfy the
energy consumption of the user even in the case when there
is no battery, i.e. γ ≥ α.
In the above model, the state Si taking values in S =
{0, 1, ..., β} describes the energy stored in the battery at time
i. The energy stored in the battery Si is a function of the
previous energy consumption X i , the energy request Y i , and
the initial state s0 ∈ S given by
Si = s0 +
i−1
X
(Yk − Xk ) .
(1)
k=0
Within this setting, a power outage occurs when Si +
Yi − Xi < 0, and energy is wasted when Si + Yi − Xi > β.
In the following we focus on EMUs that do not allow power
outages nor energy wasting but provide a quantifiable privacy
guarantee for the user. Given a particular realization x ∈ X n
of the random process X1 , X2 , . . . , Xn modelling the energy
consumption of the user up to time n, the set of energy
requests that the EMU can implement is limited by the
power outage and the energy waste constraints. The following
definition describes the set of energy requests that the EMU
can implement.
Definition 1: Given an energy consumption sequence x ∈
X n as the input of an EMU with a battery of capacity β,
the set of stable energy request sequences that avoid power
outages and energy waste is given by
Yβn (s0 , x) := {y ∈ Y n : si + yi − xi ∈ S for all i},
(2)
where si ∈ Si is the state of the battery at time i, determined
by x and y according to (1).
The task of the EMU is therefore to choose a particular
sequence in the Yβn (s0 , x) for a given power consumption
realization x. The structure of the particular choice determines
the policy implemented by the EMU and is captured by the
following definition.
Definition 2: Given an EMU with a battery of capacity β
the set of stable battery policies is the set of mappings between
the energy consumption sequences and the set of stable energy
request sequences given by
n
Pβ := {Pβ : S × X →
Yβn
(s0 , x)}.
(3)
Since Y n is known by the UP, the information about the energy
consumption of the user that the UP acquires via the energy
request is given by the mutual information I(X n ; Y n ). As in
[22] this motivates the following definition of privacy.
Definition 3: Given an EMU operating with the stable
battery policy Pβ , the information about the consumption of
the user, X n , that is leaked to the UP is the information
leakage rate given by
1
Lβ (X n , Pβ ) := I(X n ; Y n ).
(4)
n
In a SM privacy context, the aim of the EMU is to choose
a stable battery policy Pβ that minimizes the information
leakage rate, i.e. maximizes the privacy of the user. Note
that the information leakage depends on the joint probability
distribution of X n and Y n . In general, the evaluation of (4)
yields involved expressions that are difficult to evaluate [25].
For that reason previous results [19], [20], [22] tend to consider
simple probabilistic models, e.g. memoryless processes, to
evaluate the mutual information. In the remaining of the paper
we analyze the privacy guarantees for general discrete-time
random processes modelling the user consumption. To that
end, we model the EMS with a battery of capacity β as a nonprobabilistic finite-state channel [28]. The rationale for this
approach and the equivalence between the EMS and a nonprobabilistic channel are discussed in the following section.
A. Equivalence with the trapdoor channel
The trapdoor channel [28], [29] is defined as a box
containing b0 blue balls and β − b0 red balls. The operation
of the channel is depicted in Fig. 2. At time i a new ball Xi
coloured blue or red is thrown into the box. Immediately after,
one of the β + 1 balls inside the box is selected and taken out
of the box. Let Yi denote the ball extracted at time i. Following
this model, the number of blue balls inside the box at time i
is given by
bi = b0 +
i−1
X
bl (Xk ) −
k=0
i−1
X
bl (Yk ) ,
(5)
k=0
where the indicator function bl(·) equals 1 when its argument
is coloured blue and 0 otherwise. Similarly, the number of red
balls inside the box at time i is given by ri = β −bi . Replacing
bi = β − ri and b0 = β − r0 into (5) yields:
ri = r0 +
i−1
X
k=0
bl (Yk ) − bl (Xk ) .
(6)
The number of red balls inside the box is bounded between 0
and β. For a box of capacity β the set of stable output balls
Yβn (r0 , x) is defined as the set of outputs yn ∈ Y n than can
be pulled out of the box given an initial state r0 and an input
sequence x ∈ X n , i.e.
Yβn (r0 , x) = {y : ri + bl (yi ) − bl (xi ) ∈ R for all i}, (7)
where R = {0, ..., β}.
It is easy to see that for the case in which γ = α = 1
the EMS with a battery of capacity β described in Section II
is equivalent to the trapdoor channel of capacity β. The set
of balls inside the box determines the state of the trapdoor
channel, and similarly, the amount of energy stored in the
battery determines the state of the EMS channel. For the
case in which α = γ = 1 both systems are equivalent since
requesting energy from the grid corresponds to extracting a ball
from the trapdoor channel. Similarly, replacing a ball from the
trapdoor channel in (6) corresponds to charging the battery of
the EMS in (1).
III.
In this section, we provide bounds on the information leakage rate when no restrictions are imposed on the probability
law of X n . We first propose the construction of a stable
battery policy Pβ∗ and characterize an upper bound on the
information leakage rate Lβ induced by Pβ∗ and any arbitrary
random process X n . Furthermore, we show the tightness of
the upper bound by constructing a random process X̂ whose
leakage is tight with respect to the upper bound. Moreover,
the leakage rate induced by the random process X̂ is shown
to be independent of the employed battery policy Pβ ∈ Pβ .
This shows that the upper bound is tight with respect to the
minimum information leakage rate a stable battery policy Pβ
can guarantee for general random processes X n .
The approach to policy construction in this section is
similar to the code construction in [28] where a trapdoor
channel with a box of size β̂ is considered. Therein, at every
time instant a ball numbered 1, 2, ..., α̂ is introduced into the
box and one of the β̂ + 1 inside the box is extracted. In [28,
Section II] the case in which the box acts as a jammer trying to
obstruct the communication process between a sender inserting
the balls into the box and a receiver drawing the output balls
is studied. Therein, the ball extracted from the box is selected
in order to minimize the mutual information between the
input and the output. Note that the extraction criteria is not
probabilistic and is instead analyzed using combinatorial tools.
Remarkably, in [28, Proposition 1] it is shown that the Shannon
capacity Cβ̂ of such channels is lower bounded by
log α̂
β̂ + 1
.
(8)
Moreover, when α̂ = 2 the capacity is upper bounded by
Cβ ≤
1
β̂ + 1
.
A. Upper bound on the information leakage rate
We propose a battery policy based on the code construction
in [28]. The codebook proposed in the trapdoor channel context
is the counterpart of the battery policy in the smart meter case.
The proposed policy structures the energy request sequences
according to the output alphabet defined below.
Definition 4: Consider the set of codewords of length l
constructed by repetitions of 0 or α symbols, i.e. Ol =
{(0, 0, · · · , 0), (α, α, · · · , α)}. For n = lm, we define the
block repetition alphabet as the set Olm of sequences obtained
by the m-fold concatenation of codewords of length l. Specifically
Olm = Ol × Ol × ... × Ol .
(10)
We now define a stable policy that maps the energy consumption of the user to the output sequences constructed with the
block repetition alphabet Olm .
P RIVACY W ITH A N A RBITRARY E NERGY
C ONSUMPTION
Cβ ≥
the battery. However, the output is not required to contain the
same symbols as the input . For that reason, the approach in
[28] requires some modification but the main idea remains.
The derivation is presented in the next section.
(9)
In [28] the output is constrained to permutations of the
input sequence. In our setting, the sum of the output sequence
is bounded by the sum of the input sequence and the size of
Definition 5: A block battery policy Pβ∗ is a mapping of
the form
Pβ∗ : S × X n → Olm ∩ Yβl (s0 , x) .
(11)
Note that a block battery policy is nothing more than a strategy
to assign to each input sequence a stable energy request
sequence constructed with a block repetition alphabet. With
these definitions at hand we now provide the following privacy
guarantee.
Theorem 1: Consider an EMS with a battery of capacity
β and initial state s0 ∈ S . Let X n be a random process with
Xi taking values in X = {0, 1, ..., α} for i = 1, 2, . . . n and
Pβ∗ a block battery policy as described in Definition 5. Then
for l ≤ b(β + 1)/αc at least one policy Pβ∗ exists such that
Lβ X n , Pβ∗ ≤
1
.
b(β + 1)/αc
(12)
Proof: Notice that the information leakage rate is upper
bounded by
1
1
Lβ X n , Pβ∗ = I (X n ; Y n ) ≤ H (Y n ) .
(13)
n
n
Since Y n takes values in Olm and |Olm | = 2m the following
holds:
1
1
m
1
1
H (Y n ) ≤ log |Olm | = log (2m ) =
= . (14)
n
n
n
n
l
We now show that when l ≤ (β + 1)/α there exists at
least one block battery policy Pβ∗ for every initial state s0
and consumption x. To prove this we establish that for every
realization x and initial state s0 there exist an energy request
sequence determined by y ∈ Olm such that y belongs to the set
of stable energy requests Yβn (s0 , x). The strategy is to notice
that Olm ∩ Yβn (s0 , x) 6= ∅ for m = 1 and to then prove by
induction that the non-emptiness of the intersection holds for
m ≥ 1.
Theorem 2: Consider an EMS with a battery of capacity
β and initial state s0 . Let X̂ n be a random process taking
uniformly distributed values in Olm with l = d(β + 1)/αe.
Let Pβ be a stable battery policy. Then
1
Lβ X̂ n , Pβ =
.
(19)
d(β + 1)/αe
Upper bound given by Theorem 1
^ n given by Theorem 2
Information leakage rate for X
Information leakage rate - L-
0.6
0.5
0.4
Proof: We expand Lβ as
0.3
0.2
0.1
5
10
15
20
25
Normalized batery size - -=,
Fig. 3. Upper bound on the information leakage rate of an EMS as a function
of the ratio between the battery size and the peak power consumption.
The intersection
{(0, 0, · · · , 0), (α, α, · · · , α)} ∩
Yβl s0 , X l is non-empty if and only if either the sequence
(0, 0, · · · , 0) or (α, α, · · · , α) belong to Yβn (s0 , x). Jointly
with (2) this implies that either
si + 0 − xi ∈ S
(15)
si + α − xi ∈ S
(16)
or
holds for i ≤ l. In the first case, described in (15), we have
that 0 − xi ≤ 0 for i = 0, · · · , l − 1. Hence, the energy stored
in the battery, si , decreases monotonically. Therefore, all si
belong to S when si ≥ 0 on the last time step, i.e.,
0 ≤ s0 −
l−1
X
xi .
(17)
i=0
Similarly, in the case described by (16), we have that α −xi ≥
0 and the energy stored increases monotonically. It is then
sufficient to show that
s0 −
l−1
X
1
1
1
I (X n ; Y n ) = H (X n ) − H (X n |Y n ) .
(20)
n
n
n
When X n is uniformly distributed over the alphabet Olm it
yields
1
1
1
H (X n ) = m = .
(21)
n
n
l
We now show that the equivocation rate n1 H (X n |Y n ) is 0
when X n takes values in Olm with l > β/α. We prove by
induction that when the input realization x belongs to Olm with
l > β/α, the sets Yβn (s0 , x) of stable output words generated
by different consumption sequences are disjoint, i.e.
Yβn (s0 , x̂) ∩ Yβn (s0 , x) = ∅ for x̂ 6= x.
(22)
As a result, any request sequence y ∈ Yβn (s0 , x) unequivocally determines the generating input x. In other words, given
an output sequence y there is no uncertainty about the input
x, and therefore, the equivocation rate n1 H (X n |Y n ) is 0.
For m = 1 there are two possible inputs (0, 0, · · · , 0) and
(α, α, · · · , α). When x = (0, 0, · · · , 0) ∈ Ol1 the energy stored
in the battery at time l is given by
sl = s0 +
l−1
X
(yi − 0) .
(23)
i=0
Similarly, when x = (α, α, · · · , α) ∈ X l the energy stored in
the battery at time l is given by
zl = s0 +
l−1
X
(yi − α) .
(24)
i=0
xi ≤ β − αl.
(18)
Taking the difference between (23) and (24) yields:
i=0
When β − αl ≥ −1 every integer si satisfies at least one
of the inequalities given by (17) and (18). This ensures that
either (17) or (18) hold for every s0 ∈ S and x ∈ X l , and
therefore, the intersection Olm ∩ Yβn (s0 , x) is non-empty. This
completes the proof for m = 1. The induction for m ≥ 1 is
straightforward as the proof for m = 1 holds for every initial
state s0 .
The upper bound derived in Theorem 1 is depicted for
different battery sizes in Fig. 3. It is interesting to note that
the privacy guarantees increase significantly for small values of
β/α but the benefit vanishes as the size of the battery increases.
B. Tightness of the upper bound
We now study the tightness of the upper bound presented
in Theorem 1. To this end, we construct a random process
modelling the energy consumption of the user that is tight
with respect to the result in Theorem 1 for every battery policy
Pβ ∈ P β .
sl − zl =
l−1
X
α = lα.
(25)
i=0
When zl ∈ S we have that sl = zl + lα ≥ lα, showing
that for lα > β the events zl ∈ S and sl ∈ S do not
occur simultaneously. This implies that the set of output words
belonging to Yβn (s0 , (0, 0, · · · , 0)) and Yβn (s0 , (α, α, · · · , α))
is empty for every initial state s0 . Therefore the sets are disjoint
and n1 H (X n |Y n ) = 0. The proof for m > 1 follows by
induction and noticing that the proof above is valid for every
initial state s0 .
IV.
P RIVACY W ITH A N AVERAGE E NERGY C ONSTRAINT
The information leakage rate bounds provided in Section
III do not impose any moment restriction on the random
process modeling the energy consumption of the user. Indeed,
they depend only on the range of the energy consumption
and on the size of the battery. However, one of the most
widely used energy consumption metrics is the average energy
Information leakage rate - L-
0.6
0.5
0.4
0.3
0.2
7/, = 0.50
7/, = 0.25
7/, = 0.10
7/, = 0.05
7/, = 0.02
7/, = 0.01
0.6
Information leakage rate - L-
-/, = 1
-/, = 2
-/, = 4
-/, = 8
-/, = 16
0.5
0.4
0.3
0.2
0.1
0.1
0.1
0.2
0.3
0.4
0.5
0.6
0.7
0.8
0.9
5
Normalized average energy consumption - 7=,
10
15
20
Normalized batery size - -=,
Fig. 4. Upper bound on the information leakage rate of an EMS when
n → ∞ as a function of the average energy consumption of the user for
different values of the ratio between the battery size and the peak power
consumption.
Fig. 5. Upper bound on the information leakage rate of an EMS when
n → ∞ as a function of the ratio between the battery size and the peak
power consumption for different values of the average energy consumption of
the user
consumption over an arbitrary time interval. In fact, it is
common for SMs to display this information to the user.
In the following, we particularize the results in Theorem 1
and Theorem 2 to the case in which the average energy
consumption of the user is specified. Specifically, we analyze
the impact of the average energy consumption on the privacy
performance. We define the average energy consumption of
the random process X n as
" n−1 #
1X
Xi .
(26)
µn = E
n i=0
where (28) follows by applying the chain rule and (29) follows
from the fact that conditioning reduces entropy. Notice that
(29) is the entropy of m sequences Y l taking values in Ol ,
and therefore, the entropy of Y n is upper bounded by
h Pn−1 i
E n1 i=0 Yi
1
1
,
H (Y n ) ≤ H2
(30)
n
l
α
Note that since we do not impose any stationarity condition
on the random process X n , the average energy consumption
is a function of the time index n. This agrees with the nonstationary nature observed in energy consumption profiles of
users [23].
A. Upper bound on the information leakage rate
The following result provides an upper bound on the information leakage rate for random processes X n with average
energy consumption µn .
Theorem 3: Consider a battery system with capacity β and
initial state s0 . Let X n be a random process with average
energy consumption µn . Let Pβ∗ be a block battery policy, then
for l ≥ b(β + 1)/αc at least one policy Pβ∗ exists such that
µ −β
µ +β
max H2 nα n , H2 nα n
Lβ X n , Pβ∗ ≤
, (27)
b(β + 1)/αc
where H2 (p) = −p log2 p − (1 − p) log2 (1 − p) denotes the
binary entropy.
Proof: The entropy of a random process Y n taking values
in Olm is upper bounded by
m−1
1
1X
H(Y n ) =
H Yil , . . . , Y(i+1)l−1 |Y0 , . . . , Yil−1 (28)
n
n i=0
m−1
≤
1X
H Yil , . . . , Y(i+1)l−1 ,
n i=0
(29)
for the case in which each sequence Y l is independent and
identically distributed, i.e. with distribution
h P
i
n−1
1
E
Y
i=0 i
n
P Y l = (α, α, . . . , α) =
,
(31)
α
and
P Y l = (0, 0, . . . , 0) = 1 − P Y l = (α, α, . . . , α) . (32)
We now bound the average energy requested from the grid as
a function of the average energy consumption of the user and
the battery size. Dividing (1) by n and taking the expected
value yields
" n−1 #
" n−1 #
1X
1X
Sn − s0
E
Yi = E
Xi + E
,
(33)
n i=0
n i=0
n
or equivalently
" n−1 #
1X
β
β
Yi ≤ µn + .
µn − ≤ E
n
n i=0
n
(34)
Notice now that for l ≤ (β + 1)/α, and for every initial state
s0 ∈ S and input realization x ∈ X n there exists a sequence
y ∈ Olm such that y belongs to the set of stable energy
requests Yβn (s0 , x). This completes the proof.
The upper bound on the information leakage rate when the
average energy consumption of the user is known and n → ∞
is illustrated in Fig. 4 and Fig. 5. As expected, the binary
entropy term in Theorem 3 introduces concavity in the upper
bound as shown in Fig. 4. Interestingly, Fig. 5 shows that the
information leakage rate reduction as the size of the battery
increases is less significant for extreme values of the average
energy consumption.
B. Tightness of the upper bound
Proceeding in a similar fashion as in Section III we now
prove that the upper bound in Theorem 3 is tight for a certain
class of random processes modelling the energy consumption.
Theorem 4: Consider a battery system with capacity β and
initial state s0 . Let X̂ n be a random process with average
energy consumption µn and taking values in Olm with l =
d(β + 1)/αe. Let Pβ be a stable battery policy, then
µ
1
n
Lβ X̂ n , Pβ =
.
(35)
H2
d(β + 1)/αe
α
Proof: Borrowing from (30) the entropy rate of the
random process X n taking values in Olm is upper bounded
by
h Pn−1 i
E n1 i=0 Xi
1
1
n
,
(36)
H (X ) ≤ H2
n
l
α
with equality when the X l symbols forming X n are i.i.d. We
now recall that when X n takes values in Olm with l > β/α the
input xn can be uniquely determined from the output sequence
yn and H (X n |Y n ) = 0. We conclude the proof by selecting
l = d(β + 1)/αe.
V. C ONCLUSION
We have studied the information leakage rate of EMSs with
finite battery capacity for general random processes modelling
the energy consumption of the user. Inspired by the results
on permuting channels we have proposed a battery charging
policy with bounded information leakage rate for arbitrary
random processes. We have particularized the analysis to
the case in which the average energy consumption of the
user is known and we have concluded that extreme values
of the average energy consumption provide lower values of
information leakage to the utility provider.
[9]
[10]
[11]
[12]
[13]
[14]
[15]
[16]
[17]
[18]
[19]
[20]
[21]
[22]
R EFERENCES
[1]
[2]
[3]
[4]
[5]
[6]
[7]
[8]
A. Ipakchi and F. Albuyeh, “Grid of the future,” IEEE Power Energy
Mag., vol. 7, no. 2, pp. 52–62, Mar. 2009.
R. Hartway, S. Price, and C.K. Woo, “Smart meter, customer choice
and profitable time-of-use rate option,” Energy, vol. 24, no. 10, pp.
895–903, Oct. 1999.
G. Wood and M. Newborough, “Dynamic energy-consumption indicators for domestic appliances: environment, behaviour and design,”
Energy and Buildings, vol. 35, no. 8, pp. 821–841, 2003.
E. L. Quinn, “Privacy and the new energy infrastructure,” Available at
SSRN 1370731, 2008.
P. McDaniel and S. McLaughlin, “Security and privacy challenges in
the smart grid,” IEEE Security Privacy, vol. 7, no. 3, pp. 75–77, May
2009.
I. Rouf, H. Mustafa, M. Xu, W. Xu, R. Miller, and M. Gruteser,
“Neighborhood watch: Security and privacy analysis of automatic meter
reading systems,” in Proc. Conf. Computer and Commun. Security, New
York, NY, USA, 2012.
G. W. Hart, “Nonintrusive appliance load monitoring,” Proc. IEEE,
vol. 80, no. 12, Dec. 1992.
E. McKenna, I. Richardson, and M. Thomson, “Smart meter data:
Balancing consumer privacy concerns with legitimate applications,”
Energy Policy, vol. 41, no. C, pp. 807–814, 2012.
[23]
[24]
[25]
[26]
[27]
[28]
[29]
A. Molina-Markham, P. Shenoy, K. Fu, E. Cecchet, and D. Irwin,
“Private memoirs of a smart meter,” in Proc. Workshop Embedded
Sensing Systems Energy-Efficiency Building, New York, NY, USA,
2010.
M. Enev, S. Gupta, T. Kohno, and S. N. Patel, “Televisions, video
privacy, and powerline electromagnetic interference,” in Proc. Conf.
Computer Commun. Security, New York, NY, USA, 2011.
A. Prudenzi, “A neuron nets based procedure for identifying domestic
appliances pattern-of-use from energy recordings at meter panel,” in
IEEE Power Engineering Society Winter Meeting, 2002.
U. Greveler, B. Justus, and D. Loehr, “Forensic content detection
through power consumption,” in Proc. IEEE Int. Conf. Commun.,
Ottawa, ON, Canada, Jun. 2012.
H. Y. Lam, G. S. K. Fung, and W. K. Lee, “A novel method to construct
taxonomy electrical appliances based on load signaturesof,” IEEE Trans.
Consum. Electron., vol. 53, no. 2, pp. 653–660, May 2007.
C. Cuijpers and B. Koops, Smart Metering and Privacy in Europe:
Lessons from the Dutch Case, pp. 269–293, Springer Netherlands, 2012.
A. Bartoli, J. Hernandez-Serrano, M. Soriano, M. Dohler, A. Kountouris, and D. Barthel, “Secure lossless aggregation over fading and
shadowing channels for smart grid m2m networks,” IEEE Trans. Smart
Grid, vol. 2, no. 4, pp. 844–864, Dec. 2011.
S. R. Rajagopalan, L. Sankar, S. Mohajer, and H. V. Poor, “Smart meter
privacy: A utility-privacy framework,” in Proc. IEEE Int. Conf. Smart
Grid Commun., Brussels, Belgium, Oct. 2011.
F. Li, B. Luo, and P. Liu, “Secure information aggregation for smart
grids using homomorphic encryption,” in Proc. IEEE Int. Conf. Smart
Grid Commun., Oct. 2010.
L. Zhu, Z. Zhang, Z. Qin, J. Weng, and K. Ren, “Privacy protection
using a rechargeable battery for energy consumption in smart grids,”
IEEE Trans. Network, vol. 31, no. 1, pp. 59–63, Jan. 2017.
J. Gomez-Vilardebo and D. Gündüz, “Privacy of smart meter systems
with an alternative energy source,” in Proc. IEEE Int. Symp. Inf. Theory,
Istanbul, Turkey, Jul. 2013.
O. Tan, D. Gündüz, and H. V. Poor, “Smart meter privacy in the
presence of energy harvesting and storage devices,” in Proc. IEEE
Int. Conf. Smart Grid Commun., Tainan, Taiwan, Nov. 2012.
Z. Li and T. J. Oechtering, “Privacy on hypothesis testing in smart
grids,” in Proc. IEEE Inf. Theory Workshop, Jeju, South Korea, Oct.
2015.
D. Varodayan and A. Khisti, “Smart meter privacy using a rechargeable
battery: Minimizing the rate of information leakage,” in Proc. IEEE Int.
Conf. Acoust. Speech Sig. Proc., Prague, Czech Republic, May 2011.
G. Kalogridis, C. Efthymiou, S. Z. Denic, T. A. Lewis, and R. Cepeda,
“Privacy for smart meters: Towards undetectable appliance load signatures,” in Proc. IEEE Int. Conf. Smart Grid Commun., Gaithersburg,
MD, USA, Oct. 2010.
G. Kalogridis, R. Cepeda, S. Z. Denic, T. Lewis, and C. Efthymiou,
“Elecprivacy: Evaluating the privacy protection of electricity management algorithms,” IEEE Trans. Smart Grid, vol. 2, no. 4, pp. 750–758,
Dec. 2011.
I. Esnaola, A. M. Tulino, and H. V. Poor, “A statistical physics approach
to the wiretap channel,” in Proc. IEEE Int. Symp. Inf. Theory, Istanbul,
Turkey, Jul. 2013.
L. Sankar, S. Raj Rajagopalan, S. Mohajer, and H. V. Poor, “Smart
meter privacy: A theoretical framework,” IEEE Trans. Smart Grid, vol.
4, no. 2, pp. 837–846, Jun. 2013.
G. Giaconi and D. Gündüz, “Smart meter privacy with renewable energy
and a finite capacity battery,” in Proc. IEEE Int. Workshop Sig. Process.
Advances Wireless Commun., Edinburgh, UK, Jul. 2016.
R. Ahlswede and A. Kaspi, “Optimal coding strategies for certain
permuting channels,” IEEE Trans. Inf. Theory, vol. 33, no. 3, pp. 310–
314, May 1987.
R. B. Ash, Information Theory, Dover Publications Inc., New York,
1990.
| 7cs.IT
|
Where to Focus: Deep Attention-based Spatially Recurrent Bilinear Networks for
Fine-Grained Visual Recognition
Lin Wu‡ , Yang Wang†
The University of Queensland, Australia
†
The University of New South Wales, Kensington, Sydney, Australia
[email protected]; [email protected]
arXiv:1709.05769v1 [cs.CV] 18 Sep 2017
‡
Abstract
Fine-grained visual recognition typically depends
on modeling subtle difference from object parts.
However, these parts often exhibit dramatic visual variations such as occlusions, viewpoints, and
spatial transformations, making it hard to detect.
In this paper, we present a novel attention-based
model to automatically, selectively and accurately
focus on critical object regions with higher importance against appearance variations. Given an
image, two different Convolutional Neural Networks (CNNs) are constructed, where the outputs
of two CNNs are correlated through bilinear pooling to simultaneously focus on discriminative regions and extract relevant features. To capture spatial distributions among the local regions with visual attention, soft attention based spatial LongShort Term Memory units (LSTMs) are incorporated to realize spatially recurrent yet visually selective over local input patterns. All the above intuitions equip our network with the following novel
model: two-stream CNN layers, bilinear pooling
layer, spatial recurrent layer with location attention are jointly trained via an end-to-end fashion
to serve as the part detector and feature extractor, whereby relevant features are localized and extracted attentively. We show the significance of our
network against two well-known visual recognition
tasks: fine-grained image classification and person
re-identification.
1
Introduction
Fine-grained recognition such as identifying the species of
birds, models of aircrafts, identities of persons, is a challenging task since the realistic images between categories often
have subtle visual difference, and easily overwhelmed by nuisance factors such as the poses, viewpoints or illuminations.
Unlike general object recognition, fine-grained scenario can
be improved by learning critical regions of the objects, to
discriminate different subclasses meanwhile align the objects
from the same class [Huang et al., 2016a; Lin et al., 2015a;
Zhang et al., 2016; Wei et al., 2016].
lstm
Input
image
Two-stream convolutional
Spatially lstm with attention
softmax
feature extraction + bilinear pooling over the image
Figure 1: The attention model with spatially recurrent bilinear
features. The two-stream CNNs detect and extract features
of regions, and spatially LSTMs focus selectively on distinct
parts with visual attention. The white regions show what the
mode is attending to and the brightness indicates the strength
of focus. Best view in color.
1.1
Motivations
The great efforts have been paid for fine-grained recognition.
Motivated by the advances in training deep neural networks,
the recent typical work [Lin et al., 2015b; Wang et al., 2017a;
2015b; Ustinova et al., 2015] has largely improved that by
utilizing a principled bilinear CNNs to localize discriminative regions and model the appearances conditioned on them.
Specifically, the feature extraction is performed based on the
convolutional layers from two different CNN streams whose
outputs are multiplied using outer product on each region
(a.k.a bilinear pooling), on which sum-pooling over all regions is performed to derive global image descriptor. The
resulting orderless features are normalized, and fed into the
softmax layer for classification. However, such spatial relationship is disposed in bilinear combination. Essentially, convolutional layers are using sliding filters, and their outputs,
known as feature maps, which involve not only the strength of
the responses, but also their spatial positions. This indicates
that the matching of visual objects should follow its spatial
constraints. For instance, the region containing the head of a
bird should be compared with head region rather than feet.
For fine-grained recognition, humans often abstract discriminative features of these objects and then compare the
similarity among them to find the specific one, which can
be repeated many times with relative spatial distributions
(e.g.,multiple glimpses of each person on his/her hair, jacket,
and pants). As indicated [Rensink, 2000], one of the most curious facets of the human visual system is “attention”. Rather
than compress an entire image into static representation, “attention” enables salient features to dynamically come to the
forefront as needed. This is especially important when there
are a lot of clutters in an image, and also in the case of visually similar objects that are difficult to be distinguished.
1.2
Our approach
Inspired by the observations above, we propose a soft attention model with spatial recurrence for fine-grained visual
recognition, and show how our model dynamically pools convolutional features. Specifically, we propose a flexible solution by simulating the process of human visual system and
learning an end-to-end model from raw images to recurrently
localize discriminative parts of visual objects in spatial manipulations. The network is designed to nontrivially combine the strengths of bilinear CNNs and multi-dimensional
recurrent neural networks (MDRNNs) [Graves et al., 2007]
to produce spatially expressive representations of feature interactions on critical object regions. Hence, the learned deep
features help discriminate different subclasses (see Fig.1).
To this end, we implement a spatial variant of MDRNNs
with long-short term memory units [Graves and Schimidhuber, 2009], i.e.,spatial LSTMs, which naturally render bilinear
pooled features spatially-context aware due to its capacity to
capture long-range correlations. These recurrent connections
are dynamically pooled via our soft attention mechanism to
create flexible internal representations on focused object regions, yet robust to localized distortions along any combination of the input dimensions.
Contributions The major contributions are three-fold: (1)
We present a novel deep spatially recurrent model with visual
attention. The proposed method is based on spatial LSTMs
that benefits the bilinear features with spatial manipulations
to focus on the most relevant regions but also to render those
regions robust against potential transformations and local distortions; (2) We delivered an interesting intuitions on our network from visualizing “where” and “what”, the attention focused on through stochastic back-propagation of classification errors in an end-to-end fashion; (3) We show that using these features for visual recognition delivers better results
compared with spatial pyramid pooling [He et al., 2014]. Our
model is a powerful mechanism to learn to attend at the right
regions to extract relevant information for recognition, which
outperforms the state-of-the-arts throughout the typical visual
recognition tasks.
2
2.1
Related work
Fine-grained visual object recognition
A number of effective fine-grained recognition methods
have been developed in the literature [Branson et al., 2014;
Huang et al., 2016a; Lin et al., 2015a; Zhang et al., 2016;
Wu et al., 2017c; Wu and Wang, 2017b; Wu et al., 2016b;
2016a; 2017a; 2017b; Wei et al., 2016; Lin et al., 2015b;
Jaderberg et al., 2015; Zhang et al., 2014]. One pipeline
is to align the objects to eliminate pose variations and the
influence of camera positions, e.g.,[Branson et al., 2014;
Lin et al., 2015a]. Considering that the subtle difference
between fine-grained images mostly resides in the unique
properties of object parts, some approaches based on partbased representations [Zhang et al., 2014; Lin et al., 2015a;
Azizpour and Laptev, 2012] use both bounding boxes of the
body and part annotations during training to learn an accurate part localization model. The most related work to us
are bilinear CNNs (B-CNNs) [Lin et al., 2015b] and spatial
transformer network [Jaderberg et al., 2015]. B-CNNs introduce bilinear pooling upon the outputs of two different CNN
streams in order to separate part detector and feature extractor. Nonetheless, the resulting bilinear features are orderless
in which spatial relationship is disposed. In [Jaderberg et al.,
2015], a module of spatial transformation is embedded into
CNNs such that features from CNNs are invariant against a
variety of spatial transformations. However, the region detection and feature extraction are not fully studied. By contrast,
we enhance bilinear CNNs with visual attention in spatial recurrence which inherits the advantage of bilinear models and
learns features not only robust against potential spatial transformations but also useful in salient region modeling.
2.2
Long-short term memory with attention
Attention models add a dimension of interoperability by capturing where the model is focusing its attention when performing a particular task. For example, a recent work of Xu
et al.[Xu et al., 2015] used both soft attention and hard attention mechanism to generate image descriptions. Their model
actually looks at the respective objects when generating their
description. Building upon this work, Sharma et al.[Sharma
et al., 2016] developed recurrent soft attention based models
for action recognition and analyzed where they focus their
attention. However, that method is limited in learning representations in spatial constraints, which turn out to be crucial in visual recognition. In this paper, we propose recurrent
attention model to generate location dependent features by
learning to attend on spatial regions. To our best knowledge,
our work is the first of realizing attention mechanism to finegrained visual recognition.
3
Deep spatially recurrent bilinear model
with visual attention
The attention model can be presented in a quadruple M =
([gA , gB ], B, S, C), where gA and gB are feature functions,
B is bilinear pooling, S is a spatial recurrent function with
LSTM units, and C is a classification function. In our architecture, each image is first processed by two separate CNNs
(gA and gB ) to produce features at particular part locations.
Then a bilinear pooling B allows pairwise correlations between feature channels and part detectors. After that, spatial
LSTMs are used to model the spatial distribution of images
with visual attention and produce hidden states as feature representation that can be fed into classification function C. In
the following sections, we will explain each of the components in greater details.
3.1
Convolutional features
We consider two CNNs to extract features to produce features for bilinear pooling. Specifically, we use CNNs pretrained on the ImageNet dataset [Krizhevsky et al., 2012]:
L
𝐼𝐼 = 𝐸𝐸𝐿𝐿 [𝐵𝐵] ∈ 𝑅𝑅 𝐾𝐾×𝐾𝐾×𝐷𝐷�
K
K
B
be retained. To model the distribution of locations, we employ spatial LSTM [Graves and Schimidhuber, 2009], which
can be modified with location attention mechanism to achieve
adaptive visual attention. For each location (i, j) on a twodimensional grid of I, the operations performed by spatial
LSTM are given by
�
𝐷𝐷
r
l
ii,j = σ(Wxi I<i,j + Whi
hi,j−1 + Whi
hi−1,j )
Figure 2: The soft attention mechanism.
M-Net [Chatfield et al., 2014] and D-Net [Simonyan and Zisserman, 2015], truncated at the convolutional layer including
non-linearities as feature functions. The architectures of the
two CNNs are shown in Table 1. For notational simplicity, we
refer to the complete CNNs as a function, conv5 = gA (I),
conv50 = gB (I) for the two CNNs, that take an image
I as input and produces activations of the last convolution
(conv5/conv5 4 ) as output.
3.2
Bilinear pooling
In CNNs, a feature function is defined as a mapping that
takes an input image I at location L and outputs the feature
of determined size D, that is, g : I × L → R1×D . Let
gA ∈ Ra×a×DA and gB ∈ RK×K×DB denote the feature
outputs from the last convolution, where K × K is the feature
dimension, DA and DB denote respective feature channels.
Feature outputs from the two feature extractors are combined
at each location using the outer product, i.e.,bilinear pooling
operation of gA and gB at a location l:
B(l, I, gA , gB ) = gA (l, I)T gB (l, I).
(1)
Thus, the bilinear form B ∈ RK×K×D̂ (D̂ = DA DB ) allows
the outputs of two feature streams to be conditioned on each
other by considering all their pairwise interactions. Specifically, in our architecture, the outer product of the deep features from two CNN streams are calculated for each spatial
location, resulting in the quadratic number of feature maps.
Our intention here is to fuse two networks such that channel responses at the same position are put in correspondence.
To motivate this, consider the case of recognizing a bird, if a
filter in a CNN has responses to textures of some spatial location (head or wing), and the other network can recognize the
location, and their combination then discriminates this bird
species. To sequentially focus on different parts of the visual
object and extract relevant information, bilinear features are
filtered by location-dependent importance, which takes the
expectation of the whole 2-D features, i.e.,I = EL [B] where
L is a location matrix over K × K locations, which encodes
the strength of focus (defined in Eq. (3)).
3.3
Spatial LSTM with soft attention mechanism
The resulting bilinear features only allow feature interaction
at every location on the spatial grid of last convolution outputs. The statistical correlations among grids should be captured in order to make the model flexible in local feature
displacement and robust to potential spatial transformations.
Meanwhile, different parts of an object should be assigned
different weights from which more relevant informations can
l
l
l
fi,j
= σ(Wxf
I<i,j + Whf
hi−1,j )
r
r
r
fi,j = σ(Wxf I<i,j + Whf
hi,j−1 )
l
r
oi,j = σ(Wxo I<i,j + Who
hi−1,j + Who
hi,j−1 )
gi,j = tanh(Wxc I<i,j +
l
Whc
hi−1,j
ci,j = gi,j ii,j + ci,j−1
hi,j = tanh(ci,j oi,j )
r
fi,j
+
(2)
r
Whc
hi,j−1 )
+ ci−1,j
l
fi,j
where σ is the sigmoid function, indicates an element-wise
product. W are the weights connecting the layers of the neurons. Let Ii,j be the feature at location (i, j), and I<i,j designate the set of locations Im,n such at m < i or m = i
and n < j. In our model, we limit I<i,j to a smaller neighborhood surrounding the specific location, referred to casual
neighborhood. This is based on the assumption that each location is stationary to local displacement or shift invariance.
Deterministic soft attention The feature cube of I is computed by multiplying location attention matrix over bilinear
features B: I = EL [B]. This formulates a deterministic attention model by computing a soft attention weighted bilinear features. Specifically, the location attention is formulated
into a location dependent matrix, L, which is a softmax over
K × K locations. The location softmax is defined as follows,
T
exp(Uu,v
hi,j )
,
Pj
T
u=1
v=1 (Uu,v hi,j )
Lu,v = P (L = (u, v)|hi,j ) = Pi
(3)
where Uu,v are the weights mapping to the (u, v) element of
the location softmax, and L is a random variable which can
take 1-of-K 2 values. This softmax can be thought of as the
probability with which our model deems the corresponding
region in the input frame is important. After calculating these
probabilities, the soft attention mechanism [Bahdanau et al.,
2015; Xu et al., 2015] computes the expected values of the
input by taking expectation over the feature slices at different
regions (see Fig.2):
I = EP (Lu,v |hi,j ) [Bi,j ] =
j
i X
X
Lu,v Bu,v .
(4)
u=1 v=1
where B is the bilinear pooled feature tube, and Bu,v is the
(u, v) slice of the feature cube within the region Bi,j . This
corresponds to feeding in a soft weighted feature cube into
the system, and the whole model is smooth and differentiable
under the deterministic attention. Thus, learning end-to-end
is trivial by using standard back-propagation.
3.4
Loss function and attention penalty
In the training of our model, we use cross-entropy loss coupled with the doubly stochastic penalty regularization [Xu et
Table 1: CNN architectures. Each architecture contains 5 convolutional layers (conv 1-5). In M-Net, the details of each
convolutional layer are given in three sub-rows: the first specifies the number of convolution filters and their receptive field
size as “num × size × size”; the second indicates the convolution stride (“st.”) and spatial padding (“pad”); the third indicates
if Local Response Normalisation (LRN) is applied, and the max-pooling downsampling factor. In D-Net, each convolutional
layer has additional 1 × 1 convolution filters (e.g.,conv1 2 ), which can be seen as linear transformation of the input channels.
The convolution stride is fixed to 1 pixel, and padding is 1 pixel for 3 × 3 convolution layers.
Arch.
conv1
96 × 7 × 7
st. 2, pad 0
LRN, ×2 pool
conv2
256 × 5 × 5
st. 2, pad 1
LRN, ×2 pool
conv1 1 (64 × 3 × 3)
conv1 2 (64 × 1 × 1)
× 2 pool
conv2 1 (128 × 3 × 3)
conv2 2 (128 × 1 × 1)
× 2 pool
M-Net
D-Net
conv3
512 × 3 × 3
st. 1, pad 1
conv3 1 (256 × 3 × 3)
conv3 2 (256 × 1 × 1)
conv3 3 (256 × 1 × 1)
conv3 4 (256 × 1 × 1)
conv4
512 × 3 × 3
st. 1, pad 1
conv4 1 (512 × 3 × 3)
conv4 2 (512 × 1 × 1)
conv4 3 (512 × 1 × 1)
conv4 4 (512 × 1 × 1)
conv5
512 × 3 × 3
st. 1, pad 1
conv5 1 (512 × 3 × 3)
conv5 2 (512 × 1 × 1)
conv5 3 (512 × 1 × 1)
conv5 4 (512 × 1 × 1)
al., 2015], which encourages the model to pay equal attention to every part of the image. We imposePan additional constraint over the location softmax, so that
Lu,v ≈ τ where
2
τ ≥ K /D̂. Finally, the loss function is defined as follows:
Loss =
C
X
i=1
yi log ŷi + λ
X
XX 2
(τ − Lu,v )2 + γ
Θi,j , (5)
u,v
i
j
where y is the one hot label vector, ŷ is the vector of class
probabilities, which is computed via Eq.(6), C is the number
of output classes, λ is the attention penalty coefficient, γ is
the weight decay coefficient, and Θ represents all the model
parameters. The computation of ŷ is
ex·wj
P (ŷ = j|x) = PC
c=1
ex·wc
(6)
where x is the feature vector from the network, and P (ŷ =
j|x) is to predict the probability for the j-th class given x
over a combination of C linear functions. The gradients of
classification, recurrent layer of LSTM units, bilinear layer,
and two-stream CNNs with convolution, pooling, and nonlinear activations can be computed using the chain rule. We
use the following initialization strategy [Xu et al., 2015] for
the cell states, and hidden states of spatial LSTM for faster
convergence:
Figure 3: Examples from birds dataset (left), cars dataset
(middle), aircraft dataset (right), and examples from person re-identification dataset: VIPeR (left), CUHK03 dataset
(middle), and Market1501 (right).
the property of the proposed method in localizing object parts
and modeling the appearance conditioned on detected locations while being robust against a range of spatial transformations. Person re-identification shares much similar to finegrained categorizations where the matching process often resorts to the analysis of texture details and body parts to be
localized. Also, the spatial distribution among distinguished
parts is a helpful prior in recognizing identities.
4.1
Baselines
The following baselines are used in our experiments: (1)
CNN with fully-connected layers (FC-CNN): The input image is resized to 224 × 224 and mean-subtracted before prop!
!
K K
K K
agating it through the CNN. For fine-tuning, we replace the
1 XX
1 XX
Bu,v , h0 = finit,h
Bu,v
c0 = finit,c
1000-way classification layer trained on the ImageNet with
K2 u v
K2 u v
a k-way softmax layer where k is the number of classes in
(7)
target dataset. The layer before the softmax layer is used to
where finit,c and finit,h are two multi-layer perceptions and
extract features. (2) Fisher vectors with CNN features (FVthese values are used to calculate the first location matrix L
CNN): Following [Cimpoi et al., 2015], we construct a dewhich determines the initial input of I. Details about archiscriptor using FV pooling of CNN filter bank responses with
tecture and hyper-parameters are given in Section 4.2.
64 GMM components. FV is computed on the output of the
last convolution layer. Following [Lin et al., 2015b], the input
4 Experiments
images are resized to 448 × 448 and pool features in a singleTo evaluate the effectiveness of the proposed method, we conscale. (3) Fisher vectors with SIFT features (FV-SIFT): It
duct experiments by comparison to a variety of baselines and
is implemented by using dense SIFT features over 14 dense
state-of-the-arts on two applications: fine-grained image clasoverlapping 32 × 32 pixels regions with a step stride of 16
sification and person re-identification.
pixels in both direction. The features are PCA projected before learning a GMM with 256 components. (4) Bilinear
Why on fine-grained image categorization and person reCNN classification model (B-CNN) [Lin et al., 2015b]: This
identification? Fine-grained recognition tasks refer to test
method is to perform bilinear pooling on the output features
Table 2: Fine-grained categorization results. We report perimage accuracy on three datasets: CUB-200-2011, without
bounding-boxes on birds body part, aircrafts and cars.
Method
FV-SIFT
FC-CNN [M]
FC-CNN [D]
FV-CNN [M]
FV-CNN [D]
B-CNN [D,M]
B-CNN + SPP [D,M]
Krause et al.
Part-based R-CNN
Pose-normalized CNN
Spatial transformer
Chai et al.
Gosselin et al.
Ours [D,M]
Birds
18.8
58.8
70.4
64.1
74.7
84.1
86.9
82.0
73.9
75.7
84.1
89.7
Aircrafts
61.0
57.3
74.1
70.1
77.6
83.9
86.7
72.5
80.7
88.4
Cars
59.2
58.6
79.8
77.2
85.7
91.3
92.5
92.6
78.0
82.7
93.4
from two CNN streams. Then, orderless sum-pooling is employed to aggregate the bilinear features across the image. We
use the model initialized with a D-Net and a M-Net (B-CNN
[D,M]). The input images are resized to 448 × 448 and features are extracted using two networks before bilinear combination, sum-pooling and normalization. The D-Net produces
output 28 × 28 while M-Net has 27 × 27. Thus, a downsampling is conducted by dropping a row and a column from
D-Net outputs. (5) Bilinear CNN model with spatial pyramid
pooling [He et al., 2014] (B-CNN+SPP): To have fair comparison, we perform a 2-level pyramid [He et al., 2014]: 2×2
and 1×1 subdivisions. (6) The proposed method: Identical to
B-CNN, features are extracted using two CNNs with outputs
from the last convolutional layer (conv5+relu for M-Net and
conv54 +relu for D-Net), followed by bilinear pooling, spatial
recurrence with visual attention, and flattening.
4.2
Implementations
In all of our experiments, model architecture and hyperparameters are set using cross-validation. In particular, we
train a 2-layer spatial LSTM model for all datasets, where the
dimensionality of the hidden state and cell state are set to 512.
The attention penalty coefficient λ is set to be 1, weight decay γ = 10−5 , and use dropout rate of 0.5 at all non-recurrent
connections. In fact, the training of proposed model is finetuning the components of of pre-trained two CNNs, and spatial recurrence. To this end, we add a k-way softmax layer.
We adopt the two-step training procedure [Branson et al.,
2014] where we first train the last layer using logistic regression, followed by a fine-tuning the entire model using backpropagation for a number of epochs at a small learning rate
(η = 0.001). Once the fine-tuning is done, training and validation sets are combined to train one-vs-all linear SVMs on
the extracted features. In experiments, we employ two kinds
of data augmentation: flipping and shifting. For flipping, we
flip each sample horizontally to allow the model observe mirror images of the original images during training. For shifting, we shift each image by 5 pixels to the left, 5 pixels to
the right, and then further shift it by 3 pixels to the top, and 3
pixels to the bottom. This procedure makes the model more
robust to slight shifting of an object in an image. The shifting
was done without padding the borders of the images.
4.3
Results on fine-grained image classification
We conduct experiments on three fine-grained datasets: bird
species [Wah et al., 2011], aircrafts [Maji et al., 2013], and
cars [Krause et al., 2013]. Examples selected from the three
datasets are shown in Fig.3.
Birds species classification
The CUB-200-2011 dataset [Wah et al., 2011] contains
11,788 images of 200 bird species. All methods are evaluated in a protocol where the object bounding-boxes are not
provided in both training and testing phase. The comparison
results without bounding boxes are shown in Table 2. We can
see that FV-CNN[D] 74.7% and FV-CNN[M] 64.1% achieves
better results than FC-CNN [D] 70.4% and FC-CNN [M]
58.8%. This is mainly because FV-CNN pools local features
densely within the described regions, and therefore more apt
at describing local patch textures. Our method achieves the
best results compared with B-CNN and B-CNN+SPP in all
corresponding variants. More recent results are reported by
Krause et al.[Krause et al., 2015] where 82% accuracy is
achieved by leveraging more accurate CNN models to train
part detectors in a weakly supervised manner. Part based RCNN [Zhang et al., 2014] and pose-normalized CNN [Branson et al., 2014] also perform well on this dataset with accuracy of 73.9% and 75.7%, respectively. However, the two
methods are performing a two-step procedure on part detection and CNN based classifier. A competing accuracy of
84.1% is achieved by spatial transformer networks [Jaderberg
et al., 2015] while this method only models the spatial transformation locally.
Aircraft classification
The Fine-Grained Visual Classification of Aircraft (FGVCAircraft) dataset [Maji et al., 2013] consists of 10,000 images of 100 aircraft variants. The task involves discriminating
variants such as Boeing 737-300 from Boeing 737-400, and
thus the difference are very subtle, where sometimes one may
be able to distinguish them by counting the number of windows in the model. In this dataset, airplanes tend to occupy
a large portion of the whole image and appear in a relatively
clear background. Comparison results are reported in Table
2, from which it can be seen that the results of trends are similar to those in birds dataset. It is notable that FV-SIFT performs remarkably better (61.0%) and outperforms FC-CNN
[M] (57.3%). In comparison to state-of-the-art approaches,
the two best performing methods [Gosselin et al., 2014] and
[Chai et al., 2013] achieve 80.7% and 72.5%, respectively.
Our method outperforms these approaches by a significant
margin. It indicates that spatial pooling is vital to image categorization due to its robustness to local feature displacement.
Car model classification
The cars dataset [Krause et al., 2013] contains 16,185 images of 196 classes. The data is split into 8,144 training images and 8,041 testing images, where each class has been split
roughly in a 50-50 split. Classes are typically at the level of
Table 3: Rank-1, -5, -10, -20 recognition rate of different methods on the VIPeR, CUHK03 dataset, and mAP on Market1501.
Method
JointRe-id [Ahmed et al., 2015]
SDALF [Farenzena et al., 2010]
SalMatch [Zhao et al., 2013]
PatchStructure [Shen et al., 2015]
SCSP [Chen et al., 2016]
LocalMetric [Huang et al., 2016b]
FPNN [Li et al., 2014]
Multi-region [Ustinova et al., 2015]
Ours [D,M]
R=1
34.80
19.87
30.16
34.80
53.54
42.30
56.11
VIPeR
R = 5 R = 10
63.32
74.79
38.89
49.37
52.00
62.50
68.70
82.30
82.59
91.49
70.99
85.23
87.29
93.66
Make, Model, Year, e.g.,2012 Tesla Model S or 2012 BMW
M3 coupe. Cars in the dataset are smaller and appear in a
more cluttered background, and thus, challenging object and
part localization. Once again the proposed method consistently outperforms all baselines with [D, D] model achieving
93.4% accuracy. Krause et al.[Krause et al., 2015] achieves
92.6%, and methods of [Gosselin et al., 2014] and [Chai et
al., 2013] achieve 82.7% and 78.0%. Our spatial bilinear
model has a clear advantage over these models by attentively
selecting features from regions for matching.
4.4
Results on person re-identification
We perform experiments on three benchmarks: VIPeR [Gray
et al., 2007], and CUHK03 [Li et al., 2014]. The VIPeR
data set contains 632 individuals taken from two cameras
with arbitrary viewpoints and varying illumination conditions. The 632 person’s images are randomly divided into two
equal halves, one for training and the other for testing. The
CUHK03 data set includes 13,164 images of 1360 pedestrians. This dataset provides both manually labeled and detected
pedestrian bounding boxes, and we report results on labeled
data set. The dataset is randomly partitioned into training,
validation, and test with 1160, 100, and 100 identities, respectively. The Market-1501 data set contains 32,643 fully annotated boxes of 1501 pedestrians, making it the largest person
re-id dataset to date. Each identity is captured by at most six
cameras. The dataset is randomly divided into training and
testing sets, containing 750 and 751 identities, respectively.
The evaluation metric we use is Cumulative Matching Characteristic (CMC). This evaluation is performed ten times, and
the average results are reported.
Comparison to state-of-the-art approaches
Comparative experiments with state-of-the-art methods are
conducted, and results are reported in Table 3. It can be seen
that our approach outperforms all competitors consistently on
the two benchmarks on rank-1 recognition accuracy. Compared with some approaches that consider pre-defined spatial
distribution among body parts to improve matching such as
SCSP [Chen et al., 2016], PatchStructure [Shen et al., 2015],
SDALF [Farenzena et al., 2010], and SalMatch [Zhao et al.,
2013], our model is more beneficial to person re-identification
by jointly performing feature extraction and spatial manipulation. Compared with deep learning approaches with computation on local patch region difference i.e.,JointRe-id [Ahmed
R = 20
82.45
65.73
75.60
91.80
96.65
94.25
96.97
R=1
54.74
5.60
20.65
63.87
65.23
CUHK03
R = 5 R = 10
86.42
91.50
23.45
36.09
51.32
68.74
89.25
94.33
90.95
95.73
R = 20
97.31
51.96
83.06
97.05
98.52
Market1501
R=1
mAP
20.53
8.20
51.90
26.35
45.58
54.67 34.33
Figure 4: Attending to distinct regions. Best view in color.
et al., 2015], FPNN [Li et al., 2014], and LocalMetric [Huang
et al., 2016b], our method learns features from critical parts,
which helps discriminate different persons with subtle differences. Also, our method achieves performance gain over
multi-region based bilinear models [Ustinova et al., 2015]
which manually partition body parts on which bilinear features are computed, whereas the proposed network can localize distinct patches with spatial attention and select features
automatically for matching.
Learning to attend Visualizing the attention learned by the
model allows us to interpret the output of the model. In this
sense, our model is more flexible by attending salient regions.
The input to two CNNs is resized to 448 × 448, and with five
convolution with max pooling layers and bilinear pooling, we
get an output dimension of 28 × 28. In order to visualize the
attention weights for the soft model, we upsample the weights
by a factor of 25 = 32 and apply a Gaussian filter emulate
the large receptive field size. Fig. 4 shows the model learns
alignments that agree very strongly with human intuition.
5
Conclusion
In this paper, we present a deep recurrent soft attention based
model for fine grained visual recognition and analyzed where
they focus their attention. The proposed model tends to recognize important elements in visual objects that have subtle appearance difference, and thus achieve performance improvement. The impressive soft attention models are found
to be computationally expensive since they require all the
features to perform dynamic pooling. In future, we plan
to explore some hard attention solution to sample locations
over input image, as well as kernel approximation to bilinear features and multi-view features [Wang et al., 2015d;
2016b; Wu and Wang, 2017a; Wu et al., 2013b; 2015;
2013a; Wang et al., 2017b; Wang and Wu, 2017a; 2017b;
Wang et al., 2016a; 2015a; 2014a; 2015c; 2014b; 2014c],
which can reduce the computational cost of our model.
References
[Ahmed et al., 2015] Ejaz Ahmed, Michael Jones, and Tim K.
Marks. An improved deep learning architecture for person reidentification. In CVPR, 2015.
[Azizpour and Laptev, 2012] Hossein Azizpour and Ivan Laptev.
Object detection using strongly-supervised deformable part models. In ECCV, 2012.
[Bahdanau et al., 2015] Dzmitry Bahdanau, Kyunghyun Cho, and
Yoshua Bengio. Neural machine translation by jointly learning
to align and translate. In ICLR, 2015.
[Branson et al., 2014] Steve Branson, Grant Van Horn, Serge Belongie, and Pietro Perona. Bird species categorization using pose
normalized deep convolutional nets. In BMVC, 2014.
[Chai et al., 2013] Yuning Chai, Victor Lempitsky, and Andrew
Zisserman. Symbiotic segmentation and part localization for finegrained categorization. In ICCV, 2013.
[Chatfield et al., 2014] Ken Chatfield, Karen Simonyan, Andrea
Vedaldi, and Andrew Zisserman. Return of the devil in the details: Delving deep into convolutional nets. In BMVC, 2014.
[Chen et al., 2016] Dapeng Chen, Zejian Yuan, Badong Chen, and
Nanning Zhang. Similarity learning with spatial constraints for
person re-identification. In CVPR, 2016.
[Cimpoi et al., 2015] M. Cimpoi, S. Maji, and A. Vedaldi. Deep
filter banks for texture recognition and description. In CVPR,
2015.
[Farenzena et al., 2010] M. Farenzena, L. Bazzani, A. Perina,
V. Murino, and M. Cristani.
Person re-identification by
symmetry-driven accumulation of local features. In CVPR, 2010.
[Gosselin et al., 2014] P.-H. Gosselin, N. Murray, H. Jegou, and
F. Perronnin. Revisiting the fisher vector for fine-grained classification. Pattern Recognition Letter, 49:92–98, 2014.
[Graves and Schimidhuber, 2009] Alex Graves and Jurgen
Schimidhuber. Offline handwriting recognition with multidimensional recurrent neural networks. In NIPS, 2009.
[Graves et al., 2007] Alex Graves, Santiago Fernandez, and Jurgen
Schimidhuber. Multidimensional recurrent neural networks. In
International conference on artificial neural networks, 2007.
[Gray et al., 2007] D. Gray, S. Brennan, and H. Tao. Evaluating appearance models for recognition, reacquisition, and tracking. In
Proc. Int’l. Workshop on Perf. Eval. of Track. and Surv’l., 2007.
[He et al., 2014] Kaiming He, Xiangyu Zhang, Shaoqing Ren, and
Jian Sun. Spatial pyramid pooling in deep convolutional networks for visual recognition. In ECCV, 2014.
[Huang et al., 2016a] Shaoli Huang, Zhe Xu, Dacheng Tao, and
Ya Zhang. Part-staced cnn for fine-grained visual categorization.
In CVPR, 2016.
[Huang et al., 2016b] Siyuan Huang, Jinwen Lu, Jie Zhou, and
Anil K. Jain. Nonlinear local metric learning for person reidentification. In CVPR, 2016.
[Jaderberg et al., 2015] Max Jaderberg, Karen Simonyan, Andrew
Zisserman, and Koray Kavukcuoglu. Spatial transformer networks. In NIPS, 2015.
[Krause et al., 2013] Jonathan Krause, Michael Stark, Jia Deng,
and Li Fei-Fei. 3d object representations for fine-grained categorization. In 4th IEEE Workshop on 3D Representation and
Recognition, at ICCV, 2013.
[Krause et al., 2015] J. Krause, H. Jin, J. Yang, and L. Fei-Fei.
Fine-grained recognition without part annotations. In CVPR,
2015.
[Krizhevsky et al., 2012] Alex Krizhevsky, Ilya Sutskever, and Geoffrey E. Hinton. Imagenet classification with deep convolutional
neural networks. In NIPS, 2012.
[Li et al., 2014] Wei Li, Rui Zhao, Xiaoou Tang, and Xiaogang
Wang. Deepreid: Deep filter pairing neural network for person
re-identification. In CVPR, 2014.
[Lin et al., 2015a] Di Lin, Xiaoyong Shen, Cewu Lu, and Jiaya Jia.
Deep lac: Deep localization, alignment, and classification for
fine-grained recognition. In CVPR, 2015.
[Lin et al., 2015b] Tsung-Yu Lin, Aruni RoyChowdhury, and
Subhransu Maji. Bilinear cnn models for fine-grained visual
recognition. In ICCV, 2015.
[Maji et al., 2013] Subhransu Maji, Esa Rahtu, Juho Kannala,
Matthew Blaschko, and Andrea Vedaldi. Fine-grained visual
classification of aircraft. In arXiv preprint arXiv:1306.5151,
2013.
[Rensink, 2000] Ronald A Rensink. The dynamic representation of
scences. VIsual cognition, 7(1-3):17–42, 2000.
[Sharma et al., 2016] Shikhar Sharma, Ryan Kiros, and Ruslan
Salakhutdinov. Action recognition using visual attention. In
ICLR Workshop, 2016.
[Shen et al., 2015] Yang Shen, Weiyao Lin, Junchi Yan, Mingliang
Xu, Jianxin Wu, and Jingdong Wang. Person re-identification
with correspondence structure learning. In ICCV, 2015.
[Simonyan and Zisserman, 2015] Karen Simonyan and Andrew
Zisserman. Very deep convolutional networks for large-scale image recognition. In ICLR, 2015.
[Ustinova et al., 2015] Evgeniya Ustinova, Yaroslav Ganin, and
Victor Lempitsky. Multiregion bilinear convolutional neural networks for person re-identification. In arXiv:1512.05300, 2015.
[Wah et al., 2011] Catherine Wah, Steve Branson, Peter Welinder,
Pietro Perona, and Serge Belongie. The caltech-ucsd birds-2002011 dataset. In Technical Report CNS-TR-2011-001, CalTech,
2011.
[Wang and Wu, 2017a] Yang Wang and Lin Wu. Beyond lowrank representations: Orthogonal clustering basis reconstruction
with optimized graph structure for multi-view spectral clustering.
arXiv preprint arXiv:1708.02288, 2017.
[Wang and Wu, 2017b] Yang Wang and Lin Wu. Multi-view spectral clustering via structured low-rank matrix factorization. arXiv
preprint arXiv:1709.01212, 2017.
[Wang et al., 2014a] Yang Wang, Xuemin Lin, Lin Wu, Wenjie
Zhang, and Qing Zhang. Exploiting correlation consensus: Towards subspace clustering for multi-modal data. In ACM Multimedia, 2014.
[Wang et al., 2014b] Yang Wang, Xuemin Lin, Qing Zhang, and
Lin Wu. Shifting hypergraphs by probabilistic voting. In
PAKDD, pages 234–246, 2014.
[Wang et al., 2014c] Yang Wang, Jian Pei, Xuemin Lin, Qing
Zhang, and Wenjie Zhang. An iterative fusion approach to graphbased semi-supervised learning from multiple views. In PAKDD,
pages 162–173, 2014.
[Wang et al., 2015a] Yang Wang, Xiaodi Huang, and Lin Wu. Clustering via geometric median shift over riemannian manifolds. Information Sciences, 220:292–305, 2015.
[Wang et al., 2015b] Yang Wang, Xuemin Lin, Lin Wu, and Wenjie Zhang. Effective multi-query expansions: Robust landmark
retrieval. In ACM Multimedia, pages 79–88, 2015.
[Wang et al., 2015c] Yang Wang, Xuemin Lin, Lin Wu, Wenjie
Zhang, and Qing Zhang. Lbmch: Learning bridging mapping
for cross-modal hashing. In ACM SIGIR, 2015.
[Wang et al., 2015d] Yang Wang, Xuemin Lin, Lin Wu, Wenjie
Zhang, Qing Zhang, and Xiaodi Huang. Robust subspace clustering for multi-view data by exploiting correlation consensus.
IEEE Trans. Image Processing, 24(11):3939–3949, 2015.
[Wang et al., 2016a] Yang Wang, Xuemin Lin, Lin Wu, Qing
Zhang, and Wenjie Zhang. Shifting multi-hypergraphs via collaborative probabilistic voting. Knowledge and Information Systems, 46(3):515–536, 2016.
[Wang et al., 2016b] Yang Wang, Wenjie Zhang, Lin Wu, Xuemin
Lin, Meng Fang, and Shirui Pan. Iterative views agreement: An
iterative low-rank based structured optimization method to multiview spectral clustering. In IJCAI, 2016.
[Wang et al., 2017a] Yang Wang, Xuemin Lin, Lin Wu, and Wenjie Zhang. Effective multi-query expansions: Collaborative deep
networks for robust landmark retrieval. IEEE Trans. Image Processing, 26(3):1393–1404, 2017.
[Wang et al., 2017b] Yang Wang, Wenjie Zhang, Lin Wu, Xuemin
Lin, and Xiang Zhao. Unsupervised metric fusion over multiview data by graph random walk-based cross-view diffusion.
IEEE Trans. Neural Networks and Learning Systems, 28(1):57–
70, 2017.
[Wei et al., 2016] Xiushen Wei, Chenwei Xie, and Jianjin Wu.
Mask-cnn: localizaing parts and selecting descriptors for finegrained image recognition. In arXiv: 1605.06878, 2016.
[Wu and Wang, 2017a] Lin Wu and Yang Wang. Robust hashing
for multi-view data: Jointly learning low-rank kernelized similarity consensus and hash functions. Image Vision Comput, 57:58–
66, 2017.
[Wu and Wang, 2017b] Lin Wu and Yang Wang. Structured deep
hashing with convolutional neural networks for fast person reidentification. arXiv:1702.04179, 2017.
[Wu et al., 2013a] Lin Wu, Yang Wang, and John Shepherd. Efficient image and tag co-ranking: a bregman divergence optimization method. In ACM Multimedia, 2013.
[Wu et al., 2013b] Lin Wu, Yang Wang, John Shepherd, and Xiang Zhao. Max-sum diversification on image ranking with nonuniform matroid constraints. Neurocomputing, 118:10–20, 2013.
[Wu et al., 2015] Lin Wu, Xiaodi Huang, Chengyuan Zhang, John
Shepherd, and Yang Wang. An efficient framework of bregman divergence optimization for co-ranking images and tags in
a heterogeneous network. Multimedia Tools and Applications,
74(15):5635–5660, 2015.
[Wu et al., 2016a] Lin Wu, Chunhua Shen, and Anton Van den
Hengel. Personnet: Person re-identification with deep convolutional neural networks. arXiv:1601.07255, 2016.
[Wu et al., 2016b] Lin Wu, Yang Wang, and Shirui Pan. Exploiting attribute correlations: A novel trace lasso-based weakly supervised dictionary learning method. IEEE Trans. Cybernetics,
2016.
[Wu et al., 2017a] Lin Wu, Chunhua Shen, and Anton van den Hengel. Deep linear discriminant analysis on fisher networks: A hybrid architecture for person re-identification. Pattern Recognition, 65:238–250, 2017.
[Wu et al., 2017b] Lin Wu, Yang Wang, Junbin Gao, and Xue Li.
Deep adaptive feature embedding with local sample distributions
for person re-identification. Pattern Recognition, 2017.
[Wu et al., 2017c] Lin Wu, Yang Wang, Xue Li, and Junbin Gao.
What-and-where to match: Deep spatially multiplicative integration networks for person re-identification. arXiv:1707.07074,
2017.
[Xu et al., 2015] Kelvin Xu, Jimmy Ba, Ryan Kiros, Kyunghyun
Cho, Aaron Courville, Ruslan Salakhutdinov, Richard Zemel,
and Yoshua Bengio. Show, attend and tell: Neural image caption generation with visual attention. In ICML, 2015.
[Zhang et al., 2014] Ning Zhang, Evan Shelhamer, Yang Gao, and
Trevor Darrell. Part-based r-cnns for fine-grained category detection. In ECCV, 2014.
[Zhang et al., 2016] Yu Zhang, Xiushen Wei, Jianxin Wu, Jianfei
Cai, and Jiangbo Lu. Weakly supervised fine-grained categorization with part-based image representation. IEEE Transactions on
Image Processing, 25(4):1713–1725, 2016.
[Zhao et al., 2013] Rui Zhao, Wanli Ouyang, and Xiaogang Wang.
Person re-identification by salience matching. In ICCV, 2013.
| 1cs.CV
|
1
arXiv:1605.06696v1 [math.AC] 21 May 2016
ON THE STRAIGHTENING LAW FOR MINORS OF A MATRIX
RICHARD G. SWAN
In memory of Gian Carlo Rota
Abstract. We give a simple new proof for the straightening law of Doubilet,
Rota, and Stein using a generalization of the Laplace expansion of a determinant.
1. Introduction
The straightening law was proved in [5] by Doubilet, Rota, and Stein generalizing
earlier work of Hodge [6]. Since then a number of other proofs have been given
[4, 2, 1]. The object of the present paper is to offer yet another proof of this result
based on a generalization of the Laplace expansion of a determinant. This proof has
the advantage (to some!) of not requiring any significant amount of combinatorics,
Young diagrams, etc. On the other hand, for the same reason, it does not show the
interesting relations between the straightening law and invariant theory but these
are very well covered in the above references and in [3]. For completeness, I have
also included a proof of the linear independence of the standard monomials.
2. Laplace Products
Let X = (xij ) be an m × n matrix where 1 ≤ i ≤ m and 1 ≤ j ≤ n. If
A ⊆ {1, . . . , m}, B ⊆ {1, . . . , n}, and A and B have the same number of elements
we define X(A|B) to be the minor determinant of X with row indices in A and
column indices in B. I will usually just write (A|B) for X(A|B) when it is clear
what X is. I will write |A| for the number of elements in A. We set X(A|B) = 0 if
|A| 6= |B|.
e
e
P I will write A for the complement {1, . . . , m} − A and B for {1, . . . , n} − B. Also
A will denote the sum of the elements of A.
Definition
2.1. If m = n, we define the Laplace product X{A|B} to be X{A|B} =
P
P
e B).
e
(−1) A+ B (A|B)(A|
If X is understood, I will just write {A|B} for X{A|B}. This notation is, of
course, for this paper only and is not recommended for general use.
The terminology comes from the Laplace expansion
X
X
{A|T }
{S|B} =
(1)
det X =
|T |=|A|
|S|=|B|
where A and B are fixed.
The following lemma explains the sign in Definition 2.1.
e×B
e and let yij = 0
Lemma 2.2. Let yij = xij if (i, j) lies in A × B or in A
otherwise. Then X{A|B} = det(yij ).
1
I would like to thank Darij Grinberg for many corrections to an earlier version of this paper
1
2
RICHARD G. SWAN
Proof. Rearrange the rows and columns of Y = (yij ) so that those with indices in
A and B lie in the upper left hand corner. The resulting matrix has determinant
P
P
e B).
e The sign of the permutation of rows and columns is (−1) A+ B by
(A|B)(A|
the next lemma.
Lemma 2.3. Let A = {a1 < · · · < ap } be a subset of {1, . . . , n} and let {c1 <
· · · < cq } = {1, . . . , n} − A. Then
the sign of the permutation taking {1, . . . , n} to
P
{a1 , . . . , ap , c1 , . . . , cq } is (−1) (ai −i) .
Proof. Starting with {1, . . . , n} move a1 to position 1, then a2 to position 2, etc.,
each time keeping thePremaining elements in their given order. The number of
transpositions used is (ai − i).
The Laplace expansion (1) gives us a non-trivial relation between the Laplace
products of X. This suggests looking for more general relations of the form
X
ai {Si |Ti } = 0
(2)
i
with constant ai .
Since {A|B} is multilinear in the rows of X it will suffice to check a relation (2)
for the case in which the rows of X are all of the form 0, . . . , 0, 1, 0, . . . , 0. Since
{A|B} is also multilinear in the columns of X, all terms of (2) will be 0 unless
there is a 1 in each column. Therefore it will suffice to check a relation (2) for the
case where X is a permutation matrix, X = P (σ −1 ) = (δσi,j ). To do this we first
compute {A|B} for this X.
Lemma 2.4. If X = (δσi,j ) is a permutation matrix P (σ −1 ) then {A|B} = sgn σ
if σA = B and {A|B} = 0 otherwise.
This is immediate from Lemma 2.2 and the fact that det X becomes 0 if any
entry 1 is replaced by 0.
P
products (with conCorollary 2.5. A relation i ai {Si |Ti } = 0 between Laplace
P
stant ai ) holds if and only if for each σ in Sn we have
ai = 0 over those i with
σSi = Ti .
Theorem 2.6. For given A and B we have
X
X
{A|V } =
{U |B}.
V ⊆B
U⊇A
Proof. As above, it will suffice to check this when X is a permutation matrix,
X = P (σ −1 ) = (δσi,j ). For this X the right hand side is sgn σ if σA ⊆ B and
otherwise is 0. The same is true of the left hand side.
In particular, we recover the Laplace expansion (1) by setting A = ∅ or by setting
B = {1, . . . , n}.
We will show later that, for generic X, all linear relations between Laplace
products are consequences of those in Theorem 2.6.
Corollary 2.7. For given A, B, and for C ⊆ B we have
X
X
(−1)|W | {U |B − W }.
{A|V } =
C⊆V ⊆B
U⊇A
W ⊆C
STRAIGHTENING LAW
Proof. By Theorem 2.6, the right hand side is
X
X
X
(−1)|W |
{A|V } =
W ⊆C
V ⊆B−W
X
3
(−1)|W | {A|V }
V ⊆B W ⊆C−V
which is equal to the left hand side since the inner sum is 0 unless C ⊆ V .
e here denotes the complement of A.
Recall that A
Corollary 2.8. For given A, B,
X
X
f
(−1)|W | {U |W } =
{A|Ve }.
V ⊆B
U⊇A
W ⊇B
Proof. Set B = {1, . . . , n} in Corollary 2.7, getting
X
X
f }.
{A|V } =
(−1)|W | {U |W
V ⊇C
U⊇A
W ⊆C
e Ve , W
f where B is now the B given in Corollary 2.8.
Replace C, V , W by B,
3. Straightening Laplace products
If S = {s1 < · · · < sp } and T = {t1 < · · · < tq } are subsets of {1, · · · , n},
we define a partial ordering S ≤ T as in [3] to mean p ≥ q and sν ≤ tν for all
ν ≤ q. Equivalently S ≤ T if and only if |S ∩ {1, · · · , r}| ≥ |T ∩ {1, · · · , r}| for all
1 ≤ r ≤ n. Note that S ⊇ T implies S ≤ T and S ⊃ T implies S < T .
As above let Se be the complement {1, . . . , n} − S. For want of a better terminology, I will say that S is good if S ≤ Se and that S is bad otherwise.
The following theorem is our straightening law for Laplace products.
P
Theorem 3.1. For any {A|B} we have {A|B} =
±{Ai |Bi } where Ai ≤ A,
Bi ≤ B and Ai and Bi are good.
Note that some of the (Ai , Bi ) may be equal.
Proof. By induction on the set of pairs of subsets of {1, · · · , n} partially ordered
by (A, B) ≤ (C,P
D) if A ≤ C and B ≤ D, it is sufficient to prove that if A is bad,
then {A|B} = P ±{Ai |Bi } with Ai < A and Bi ≤ B, and, similarly, if B is bad
then {A|B} =
±{Ai |Bi } with Ai ≤ A and Bi < B. Each statement implies the
other by transposing our matrix.
Suppose first that |A| = |B| < n/2. Note both A and B are bad in this case. By
Corollary 2.8 we have
X
±{U |W } = 0.
U⊇A
W ⊇B
since the other side in Corollary 2.8 is 0 because |A| < n/2 < |Ve |. One term of
this sum is ±{A|B} while all other nonzero terms have the form ±{U |W } where
U < A and W < B.
In the remaining case |A| = |B| ≥ n/2 we use an argument similar to that of
e = {j1 < · · · < jq }.
Hodge [6]. Suppose B is bad. Let B = {i1 < · · · < ip } and let B
Since B is bad and q ≤ p, we have iν > jν for some ν which we choose minimal.
Let D = B ∪ {j1 , . . . , jν } and let C = {j1 < · · · < jν < iν < · · · < ip }. Apply
4
RICHARD G. SWAN
Corollary 2.7 with D in place of B. The left hand side is 0 since |A| = p < |C| = p+1
so we get
X
(−1)|W | {U |D − W } = 0.
U⊇A
W ⊆C
The term with U = A and W = {j1 , . . . , jν } is ±{A|B}. This is the only term of the
form ±{U |B} since |U | = |B| and U ⊇ A. In the remaining terms we have U ≤ A
since U ⊇ A. In these terms D −W 6= B so if W ⊆ {j1 < · · · < jν } then D −W ⊃ B
and therefore D − W < B. In any case we have |D − W | = |U | ≥ |A| = |B|. If W
contains some iµ then D − W is obtained from B by removing some (and at least
one) of the elements {iν < · · · < ip } and replacing them by at least as many of the
smaller elements {j1 < · · · < jν }. This operation does not decrease the size of the
sets (D − W ) ∩ {1, · · · , r} for r ≥ 1 so |(D − W ) ∩ {1, · · · , r}| ≥ |B ∩ {1, · · · , r}|.
Therefore D − W < B.
4. The straightening law for minors
We now use a simple trick to generalize Theorem 3.1 to the case of products of
any two minors of a rectangular matrix X = (xij ) where 1 ≤ i ≤ m and 1 ≤ j ≤ n.
Recall that (S|T ) is the minor of X with row indices in S and column indices in
T . We set (S|T ) = 0 if |S| 6= |T |. As above we write (S ′ , S ′′ ) ≤ (T ′ , T ′′ ) if S ′ ≤ T ′
and S ′′ ≤ T ′′ .
The following theorem is the straightening law for minors.
Theorem 4.1. If (S ′ , T ′ ) (S ′′ , T ′′ ) then
X
(S ′ |T ′ )(S ′′ |T ′′ ) =
±(Si′ |Ti′ )(Si′′ |Ti′′ )
where (Si′ , Ti′ ) < (S ′ , T ′ ) and (Si′ , Ti′ ) ≤ (Si′′ , Ti′′ ).
We need two lemmas for the proof. By an order preserving map I mean one
satisfying f (a) ≤ f (b) if a ≤ b.
Lemma 4.2. Let U ′ and U ′′ be finite subsets of a totally ordered set and let k =
|U ′ | + |U ′′ |. Then there is an order preserving map f : K = {1, . . . , k} → U ′ ∪ U ′′
and disjoint subsets K ′ and K ′′ of K with K = K ′ ⊔ K ′′ such that
(a): f maps K ′ isomorphically onto U ′ and K ′′ isomorphically onto U ′′
(b): If a, b ∈ K, f (a) = f (b), and a < b, then a ∈ K ′ and b ∈ K ′′ .
Proof. Let L′ and L′′ be disjoint sets in 1 − 1 correspondence with U ′ and U ′′ . Let
L = L′ ⊔ L′′ and let f : L → U ′ ∪ U ′′ map L′ bijectively onto U ′ and L′′ bijectively
onto U ′′ . Define an ordering on L by setting a < b if f (a) < f (b) or if f (a) = f (b)
with a ∈ L′ and b ∈ L′′ . It is easy to check that this defines a total ordering
preserved by f . Since L is a totally ordered finite set of order k it is isomorphic to
K = {1, . . . , k} and we can substitute K for L letting K ′ and K ′′ correspond to L′
and L′′ . (a) is clear and (b) follows from the definition of the ordering/
If f is order preserving and is injective on P and Q then P ≤ Q implies f (P ) ≤
f (Q) since the injectivity guarantees that |f (P )| = |P | and similarly for Q.
Lemma 4.3. In the situation of Lemma 4.2 if P is a subset of K on which f is
injective then P < K ′ implies that f (P ) < f (K ′ ) = U ′ .
STRAIGHTENING LAW
5
Proof. It is clear that f (P ) ≤ f (K ′ ). We must show that f (P ) 6= f (K ′ ). Suppose
f (P ) = f (K ′ ). Injectivity shows that |P | = |f (P )| = |f (K ′ )| = |K ′ |. Let P =
{u1 < · · · < up } and K ′ = {v1 < · · · < vp }. Since f is order preserving and
injective on P and on K we have f (P ) = {f (u1 ) < · · · < f (up )} and f (K ′ ) =
{f (v1 ) < · · · < f (vp )} Now uν ≤ vν for all ν since P < K ′ but f (uν ) = f (vν ) for
all ν since f (P ) = f (K ′ ). Since P < K ′ , uν = vν can’t hold for all ν otherwise P
and K ′ would be equal, so for some ν we have uν < vν . But since f (uν ) = f (vν ),
Lemma 4.2(b) shows that vν must lie in K ′′ which is a contradiction.
Proof of Theorem 4.1. We can assume that |S ′ | = |T ′ | and |S ′′ | = |T ′′ | since otherwise the left hand side is 0. Let k = |S ′ | + |S ′′ | = |T ′ | + |T ′′ | and apply Lemma 4.2
to U ′ = S ′ and U ′′ = S ′′ getting an order preserving map ϕ : I → S ′ ∪ S ′′ with
I = I ′ ⊔ I ′′ (disjoint union), I ′ mapping isomorphically to S ′ , and I ′′ mapping isomorphically to S ′′ . Similarly define ψ : J → T ′ ∪ T ′′ with J = J ′ ⊔ J ′′ . Note that
I = J = {1, . . . , k}. We call them I and J to distinguish their use as row and
column indices.
Define a k × k matrix Y indexed by I and J by setting yij = xϕ(i)ψ(j) . Then, for
P ⊆ I and Q ⊆ J, we have Y (P |Q) = X(ϕ(P )|ψ(Q)) if ϕ|P and ψ|Q are injective
while Y (P |Q) = 0 otherwise since two rows
P or columns will be equal.
By Theorem 3.1 we have Y {I ′ |J ′ } = ±Y {Ii′ |Ji′ } which we can write as
X
(3)
Y (I ′ |J ′ )Y (I ′′ |J ′′ ) =
±Y (Ii′ |Ji′ )Y (Ii′′ |Ji′′ )
where Ii′′ = I − Ii′ = Iei′ and Ji′′ = J − Ji′ = Jei′ . By Theorem 3.1 we see that Ii′ ≤ I ′ ,
Ji′ ≤ J ′ and that Ii′ and Ji′ are good so that Ii′ ≤ Ii′′ and Ji′ ≤ Ji′′ . By omitting all
0 terms in (3) we can insure that ϕ is injective on all Ii′ and all Ii′′ and that ψ is
injective on all Ji′ and Ji′′ .
Let Si′ = ϕ(Ii′ ), Si′′ = ϕ(Ii′′ ) and similarly for T . Because of the injectivity we
can write (3) (with the 0 terms removed) as
X
(4)
(S ′ |T ′ )(S ′′ |T ′′ ) =
±(Si′ |Ti′ )(Si′′ |Ti′′ )
where (Si′ , Ti′ ) ≤ (S ′ , T ′ ) and (Si′ , Ti′ ) ≤ (Si′′ , Ti′′ ).
If I ′ and J ′ are good then I ′ ≤ I ′′ and J ′ ≤ J ′′ which implies that S ′ ≤ S ′′ and
′
T ≤ T ′′ contrary to the hypothesis. Therefore one of I ′ and J ′ , say I ′ , must be
bad. Since I ′ is bad, Ii′ is good, and Ii′ ≤ I ′ , we have Ii′ < I ′ . By Lemma 4.3 it
follows that Si′ < S ′ showing that (Si′ , Ti′ ) < (S ′ , T ′ ). The same argument applies
if J ′ is bad.
Remark 4.4. Suppose A = A′ ⊔ A′′ where A′ and A′′ are disjoint and ϕ : A → B is
onto. Suppose ϕ is injective on A′ and on A′′ . Let B ′ = ϕ(A′ ) and B ′′ = ϕ(A′′ ).
Then B ′ ∩ B ′′ = {x| |ϕ−1 (x)| = 2} is independent of A′ and A′′ . We can think
of B = ϕ(A) as a set with multiplicities where the multiplicity of a point x is the
order of ϕ−1 (x). If we think of B ′ and B ′′ as sets with all points of multiplicity 1
then B ′ ∪ B ′′ = B as sets with multiplicities. Applying this remark to the maps
ϕ : I → S ′ ∪ S ′′ and ψ : J → T ′ ∪ T ′′ defined in the proof of Theorem 4.1 we
see that the sets given by our proof of Theorem 4.1 satisfy Si′ ∪ Si′′ = S ′ ∪ S ′′ and
Ti′ ∪ Ti′′ = T ′ ∪ T ′′ as multisets.
6
RICHARD G. SWAN
5. Standard monomials
We say that a product (A1 |B1 ) · · · (Ar |Br ) of the minors of a matrix X is a
standard monomial if A1 ≤ A2 ≤ · · · ≤ Ar and B1 ≤ B2 ≤ · · · ≤ Br . We regard
two standard monomials which only differ by factors of the form (∅|∅) as identical.
The following is an easy consequence of Theorem 4.1.
Corollary 5.1. Any polynomial in the entries of X is a linear combination of
standard monomials in the minors of X.
Proof. Since xij = ({i}|{j}), it is clear that any such polynomial is a linear combination of products of the minors of X. We show that any product (A1 |B1 ) · · · (Ar |Br )
with r factors is a linear combination of standard monomials with r factors by induction on r and on (A1 , B1 ) in the finite partially ordered set of pairs of subsets
of {1, . . . , m} and {1, . . . , n}. By induction on r we can assume that (A2 , B2 ) ≤
· · · ≤ (Ar , Br ). If (A1 , B1 ) ≤ (A
P2 , B2 ) or r = 1 we are done. If not, Theorem 4.1
shows that (A1 |B1 )(A2 |B2 ) = ±(Ci |Di )(Pi |Qi ) where (Ci , Di ) < (A1 , B1 ) so we
are done by induction on (A1 , B1 ).
Remark 5.2. It follows from Remark 4.4 that if we write (A1 |B1 ) · · · (Ar |Br ) as a
(i)
(i)
(i)
(i)
linear combination of standard monomials (A1 |B1 ) · · · (Ari |Bri ) then
S
S (i)
S
S (i)
= j Bj for all i, counting multiplicities. In other
j Bj
j Aj and
j Aj =
words the two sides have the same content in the sense of [2].
To conclude, we give a proof of the following theorem which is rather similar to
the proof in [4] but which uses no combinatorial constructions. By a generic matrix
we mean one whose entries are distinct indeterminates.
Theorem 5.3. If X is a generic matrix, the standard monomials in the minors of
X are linearly independent.
Before giving the proof we review some results about ordering monomials. Given
a totally ordered set of indeterminates, we can order the monomials in these indeterminates as follows. If x is an indeterminate and m is such a monomial write
ordx m for the number of times x occurs in m. If m1 and m2 are monomials define
m1 > m2 to mean ordx m1 > ordx m2 for some x while ordy m1 = ordy m2 for all
y > x. It is easy to check that this defines a total ordering on the set of monomials.
Lemma 5.4. If u1 , u2 , · · · , uk and v1 , v2 , · · · , vk are monomials with ui ≤ vi for
all i and ui < vi for some i then u1 u2 · · · uk < v1 v2 · · · vk .
It is sufficient to show a < b implies ac < bc and then replace the ui ’s by the vi ’s
one by one.
It follows that if f and g are linear combinations of monomials, the leading
monomial of f g is the product of the leading monomials of f and g.
Proof of Theorem 5.3. We specialize X to a matrix of the form X = Y Z where Y
is a generic m × N matrix, Z is a generic N × n matrix and N is sufficiently large.
By the classical Binet–Cauchy theorem we have
X
(5)
X(A|B) =
Y (A|S)Z(S|B).
S
This just expresses the functoriality of the exterior product:
^
^
^
(6)
(X) = (Y ) (Z).
STRAIGHTENING LAW
7
By omitting 0 terms in equation (5) we can assume that |A| = |B| = |S| = p. In the
(ν)
(ν)
situation of equation (5) let Y = (yi )1≤i≤m, 1≤ν≤N and Z = (zj )1≤ν≤N 1≤j≤n
(ν)
(ν)
be generic matrices where the indeterminates yi and zj are all distinct. Then
PN
(ν) (ν)
X = Y Z has entries xij = ν=1 yi zj . We order the indeterminates as follows:
(1)
(1)
(2)
(1)
(2)
y1 > · · · > ym
> z1 > · · · > zn(1) > y1 > · · · > ym
> ···
(7)
and order the monomials in these indeterminates as described above.
If A = {a1 < · · · < ap } and S = {s1 < · · · < sp }, then
X
p)
(8)
Y (A|S) =
.
±ya(sσ11 ) · · · ya(sσp
σ∈Sp
(s )
(s )
(s )
(s )
(s )
(s )
p
The leading monomial of Y (A|S) is ya11 · · · yapp since ya11 · · · yapp > yaσ11 · · · yaσp
for σ 6= 1. To see this, let i be least such that σi 6= i. Then σi > i so
(sp )
(s )
(s )
(s )
(s )
(s )
(s )
while
yaσii < yaii . Choose x = yaii . Then ordx ya11 · · · yapp > ordx yaσ11 · · · yaσp
(sp )
(sp )
(s1 )
(s1 )
ordz ya1 · · · yap = ordz yaσ1 · · · yaσp for z > x.
Similarly, if B = {b1 < · · · < bp }, then the leading monomial of Z(S|B) is
(s )
(s1 )
zb1 · · · zbpp . The various terms on the right hand side of (5) have leading mono-
(s )
(s ) (s )
(s )
mials ya11 · · · yapp zb1 1 · · · zbpp . Of these, the one with si = i for all i is the largest
by Lemma 5.4. Therefore, if N ≥ |A| = |B|, the leading monomial of X(A|B) is
(p)
(1)
(p)
(1)
y(A)z(B) where y(A) = ya1 · · · yap and z(B) = zb1 · · · zbp . It follows that the
leading term of (A1 |B1 ) · · · (Ar |Br ) is y(A1 ) · · · y(Ar )z(B1 ) · · · z(Br ).
Now if A1 ≤ A2 ≤ · · · ≤ Ar , and if N ≥ |A1 |, we can recover A1 from M =
y(A1 ) · · · y(Ar ) as follows. Let Ai = {ai1 < · · · < aipi }. Then
(9)
M=
Y
j
···
ya(1)
j1
Y
···
ya(s)
js
j
Since A1 ≤ A2 ≤ · · · ≤ Ar we have a1s ≤ a2s ≤ · · · , and we see that a1s is the
(s)
least c such that yc occurs in M . Note that p1 ≥ p2 ≥ · · · so that a1s will exist if
ais does.
Since M/y(A1 ) = y(A2 ) · · · y(Ar ) we see that M determines A2 , A3 , etc. if
N ≥ |A1 |. Similarly z(B1 ) · · · z(Br ) determines B1 , B2 , etc. Therefore the leading
monomials of the standard monomials (A1 |B1 ) · · · (Ar |Br ) with N ≥ |A1 | and N ≥
|B1 | are all distinct and the theorem follows since N can be arbitrarily large.
Corollary 5.5. For a generic square matrix X, all linear relations between the
Laplace products of X are consequences of those in Theorem 2.6.
Proof. By Theorem 3.1 the space of all Laplace products of X is spanned by those of
e B)
e is a standard
the form {A|B} with A and B good. For these {A|B} = ±(A|B)(A|
monomial so these {A|B} are linearly independent. Since the only relations needed
to prove Theorem 3.1 are those of Theorem 2.6, the result follows.
References
[1] C. Akin, D. Buchsbaum, and J. Weyman, Schur functors and Schur complexes, Adv. in Math.
44(1982), 207–278.
[2] C. De Concini, D. Eisenbud, and C. Procesi, Young diagrams and determinantal varieties,
Inv. Math. 56(1980), 129–165.
8
RICHARD G. SWAN
[3] C. De Concini, D. Eisenbud, and C. Procesi, Hodge Algebras, Astérisque 91(1982).
[4] J. Desarmenien, J. P. S. Kung, and G. C. Rota, Invariant theory, Young bitableaux, and
combinatorics, Adv. in Math. 27(1978), 63–92.
[5] P. Doubilet, G. C. Rota, and J. Stein, On the foundations of combinatorial theory IX: combinatorial methods in invariant theory, Studies in Appl. Math. 53(1974), 185–216.
[6] W. V. D. Hodge, Some enumerative results in the theory of forms, Proc. Camb. Phil. Soc.
39(1943), 22–30.
Department of Mathematics, The University of Chicago, Chicago, IL 60637
E-mail address: [email protected]
| 0math.AC
|
Constant Factor Approximation
for ATSP with Two Edge Weights
Ola Svensson∗
[email protected]
Jakub Tarnawski†
[email protected]
László A. Végh‡
[email protected]
arXiv:1511.07038v2 [cs.DS] 4 Sep 2017
Abstract
We give a constant factor approximation algorithm for the Asymmetric Traveling Salesman Problem on shortest path metrics of directed graphs with two different edge weights.
For the case of unit edge weights, the first constant factor approximation was given recently by Svensson. This was accomplished by introducing an easier problem called LocalConnectivity ATSP and showing that a good solution to this problem can be used to obtain
a constant factor approximation for ATSP. In this paper, we solve Local-Connectivity ATSP
for two different edge weights. The solution is based on a flow decomposition theorem for
solutions of the Held-Karp relaxation, which may be of independent interest.
1
Introduction
The traveling salesman problem — one of finding the shortest tour of n cities — is one of the
most classical optimization problems. Its definition dates back to the 19th century and since
then a large body of work has been devoted to designing “good” algorithms using heuristics,
mathematical programming techniques, and approximation algorithms. The focus of this work
is on approximation algorithms. A natural and necessary assumption in this line of work that
we also make throughout this paper is that the distances satisfy the triangle inequality: for any
triple i, j, k of cities, we have d(i, j)+d(j, k) ≥ d(i, k) where d(·, ·) denotes the pairwise distances
between cities. In other words, it is not more expensive to take the direct path compared to a
path that makes a detour.
With this assumption, the approximability of TSP turns out to be a very delicate question
that has attracted significant research efforts. Specifically, one of the first approximation algorithms (Christofides’ heuristic [Chr76]) was designed for the symmetric traveling salesman
problem (STSP) where we assume symmetric distances (d(i, j) = d(j, i)). Several works (see
e.g. [FGM82, AGM+ 10, GS11, AG15, Sve15]) have addressed the more general asymmetric
traveling salesman problem (ATSP) where we make no such assumption.
However, there are still large gaps in our understanding of both STSP and ATSP. In fact,
for STSP, the best approximation algorithm remains Christofides’ 3/2-approximation algorithm
from the 70’s [Chr76]. For the harder ATSP, the state of the art is a O(log n/ log log n)approximation algorithm by Asadpour et al. [AGM+ 10] and a recent O(poly log log n)-estimation
algorithm1 by Anari and Oveis Gharan [AG15]. On the negative side, the best inapproximability results only say that STSP and ATSP are hard to approximate within factors 123/122
∗
École Polytechnique Fédérale de Lausanne. Supported by ERC Starting Grant 335288-OptApprox.
École Polytechnique Fédérale de Lausanne.
‡
London School of Economics. Supported by EPSRC First Grant EP/M02797X/1.
1
An estimation algorithm is a polynomial-time algorithm for approximating/estimating the optimal value
without necessarily finding a solution to the problem.
†
1
and 75/74, respectively [KLS15]. Closing these gaps is a major open problem in the field of
approximation algorithms (see e.g. “Problem 1” and “Problem 2” in the list of open problems
in the recent book by Williamson and Shmoys [WS11]). What is perhaps even more intriguing
about these questions is that we expect that a standard linear programming (LP) relaxation,
often referred to as the Held-Karp relaxation, already gives better guarantees. Indeed, it is
conjectured to give a guarantee of 4/3 for STSP and a guarantee of O(1) (or even 2) for ATSP.
An equivalent formulation of STSP and ATSP from a more graph-theoretic point of view
is the following. For STSP, we are given a weighted undirected graph G = (V, E, w) where
w : E → R+ and we wish to find a multisubset F of edges of minimum total weight such that
(V, F ) is connected and Eulerian. Recall that an undirected graph is Eulerian if every vertex
has even degree. We also remark that we use the term multisubset as the solution F may use
the same edge several times. An intuitive point of view on this definition is that G represents
a road network, and a solution is a tour that visits each vertex at least once (and may use a
single edge/road several times). The definition of ATSP is similar, with the differences that the
input graph is directed and the output is Eulerian in the directed sense: the in-degree of each
vertex equals its out-degree. Having defined the traveling salesman problem in this way, there
are several natural special cases to consider. For example, what if G is planar? Or, what if all
the edges/roads have the same length, i.e., if G is unweighted?
For planar graphs, we have much better algorithms than in general. Grigni, Koutsoupias
and Papadimitriou [GKP95] first obtained a polynomial-time approximation scheme for STSP
restricted to unweighted planar graphs, which was later generalized to edge-weighted planar
graphs by Arora et al. [AGK+ 98]. More recently, ATSP on planar graphs (and more generally
bounded genus graphs) was shown to admit constant factor approximation algorithms (first by
Oveis Gharan and Saberi [GS11] and later by Erickson and Sidiropoulos [ES14] who improved
the dependency on the genus).
In contrast to planar graphs, STSP and ATSP remain APX-hard for unweighted graphs
(ones where all edges have identical weight) and, until recently, there were no better algorithms
for these cases. Then, in a recent series of papers, the approximation guarantee of 3/2 was
finally improved for STSP restricted to unweighted graphs. Specifically, Oveis Gharan, Saberi
and Singh [GSS11] first gave an approximation guarantee of 1.5−ǫ; Mömke and Svensson [MS11]
proposed a different approach yielding a 1.461-approximation guarantee; Mucha [Muc12] gave
a tighter analysis of this algorithm; and Sebő and Vygen [SV14] significantly developed the
approach to give the currently best approximation guarantee of 1.4. Similarly, for ATSP, it
was only very recently that the restriction to unweighted graphs could be leveraged: the first
constant approximation guarantee for unweighted graphs was given by Svensson [Sve15]. In
this paper we make progress towards the general problem by taking the logical next step and
addressing a simple case left unresolved by [Sve15]: graphs with two different edge weights.
Theorem 1.1. There is an O(1)-approximation algorithm for ATSP on graphs with two different edge weights.
The paper [Sve15] introduces an “easier” problem named Local-Connectivity ATSP, where
one needs to find an Eulerian multiset of edges crossing only sets in a given partition rather than
all possible sets (see next section for definitions). It is shown that an “α-light” algorithm to this
problem yields a (9 + ε)α-factor approximation for ATSP. For unweighted graphs (and slightly
more generally, for node-induced weight functions2 ) it is fairly easy to obtain a 3-light algorithm
for Local-Connectivity ATSP; the difficult part in [Sve15] is the black-box reduction of ATSP
to this problem. Note that [Sve15] easily gives an O(wmax /wmin )-approximation algorithm in
2
For ATSP, we can think of a node-weighted graph as an edge-weighted graph where the weight of an edge
(u, v) equals the node weight of u.
2
general if we take wmax and wmin to denote the largest and smallest edge weight, respectively.
However, obtaining a constant factor approximation even for two different weights requires
substantial further work.
In Local-Connectivity ATSP we need a lower bound function
lb : V → R+ on the vertices.
P
∗
The natural choice for node-induced weights is lb(v) =
e∈δ+ (v) w(e)xe . With this weight
function, every vertex is able to “pay” for the incident edges in the Eulerian subgraph we are
looking for. This choice of lb does not seem to work for more general weight functions, and we
need to define lb more “globally”, using a new flow theorem for Eulerian graphs (Theorem 2.1).
In Section 1.2, after the preliminaries, we give a more detailed overview of these techniques
and the proof of the theorem. Our argument is somewhat technical, but it demonstrates the
potential of the Local-Connectivity ATSP problem as a tool for attacking general ATSP.
Finally, let us remark that both STSP [PY93, BK06] and ATSP [Blä04] have been studied in
the case when all distances are either 1 or 2. That restriction is very different from our setting,
as in those cases the input graph is complete. In particular, it is trivial to get a 2-approximation
algorithm there, whereas in our setting – where the input graph is not complete – a constant
factor approximation guarantee already requires non-trivial algorithms. (In our setting, we can
still think about the metric completion, but it will usually have more than two different edge
weights.)
1.1
Notation and preliminaries
We consider an edge-weighted directed graph G = (V, E, w) with w : E → R+ . For a vertex
subset S ⊆ V we let δ+ (S) = {(u, v) ∈ E : u ∈ S, v ∈ V \ S} and δ− (S) = {(u, v) ∈ E : u ∈
V \ S, v ∈ S} denote the sets of outgoing and incoming edges, respectively. For two vertex
subsets X, Y ⊆ V , we let δ(X, Y ) = {(u, v) ∈ E : u ∈ X \ Y, v ∈ Y \ X}. For a subset of edges
−
+
−
′
′
+
′
E ′ ⊆ E, we use δE
′ (S) = δ (S) ∩ E and δE ′ (S) = δ (S) ∩ E . We also let C(E ) = (G̃1 , . . . , G̃k )
′
denote the set of weakly connected components of the graph (V, E ); the vertex set V will always
be clear from the context. For a directed graph G̃ we use V (G̃) to denote its vertex set and
E(G̃) the edge set. For brevity, P
we denote the singleton set {v} by v (e.g. δ+ (v) = δ+ ({v})),
and we use the notation x(F ) = e∈F xe for a subset F ⊆ E of edges and a vector x ∈ RE . For
a multiset F , we have 1F denote the indicator vector of F , which has a coordinate for each edge
e with value equal to the number of copies of e in F . For the case of two edge weights, we use
0 ≤ w0 < w1 to denote the two possible values, and partition E = E0 ∪ E1 so that w(e) = w0
if e ∈ E0 and w(e) = w1 if e ∈ E1 . We will refer to edges in E0 and E1 as cheap and expensive
edges, respectively.
We define ATSP as the problem of finding a connected Eulerian subgraph of minimum
weight. As already mentioned in the introduction, this definition is equivalent to that of visiting
each city exactly once (in the metric completion) since we assume the triangle inequality. The
formal definition is as follows.
ATSP
Given: An edge-weighted (strongly connected) digraph G = (V, E, w).
P
Find: A multisubset F of E of minimum total weight w(F ) = e∈F w(e) such that (V, F )
is Eulerian and connected.
3
Held-Karp Relaxation. The Held-Karp relaxation has a variable xe ≥ 0 for every edge in
G. The intended meaning is that xe should equal the number of times e is used in the solution.
The relaxation LP(G) is defined as follows:
X
minimize
w(e)xe
e∈E
subject to x(δ+ (v)) = x(δ− (v))
v ∈ V,
+
x(δ (S)) ≥ 1
∅=
6 S ( V,
x ≥ 0.
(LP(G))
The first set of constraints says that the in-degree should equal the out-degree for each vertex,
i.e., the solution should be Eulerian. The second set of constraints enforces that the solution
is connected; they are sometimes referred to as subtour elimination constraints. Finally, we
remark that although the Held-Karp relaxation has exponentially many constraints, it is wellknown that we can solve it in polynomial time either by using the ellipsoid method with a
separation oracle or by formulating an equivalent compact (polynomial-size) linear program.
We will use x∗ to denote an optimal solution to LP(G) of value OPT, which is a lower bound
on the value of an optimal solution to ATSP on G.
Local-Connectivity ATSP. The Local-Connectivity ATSP problem can be seen as a twostage procedure. In the first stage, the input is an edge-weighted digraph G = (V, E, w) and the
output is a “lower bound” function lb : V → R+ on the vertices such that lb(V ) ≤ OPT. In the
second stage, the input is a partition of the vertices, and the output is an Eulerian multisubset
of edges which crosses each set in the partition and where the ratio of weight to lb of every
connected component is as small as possible. We now give the formal description of the second
stage, assuming the lb function is already computed.
Local-Connectivity ATSP
Given: An edge-weighted digraph G = (V, E, w), a function lb : V → R+ with lb(V ) ≤
OPT, and a partitioning V = V1 ∪ V2 ∪ . . . ∪ Vk of the vertices.
Find: A Eulerian multisubset F of E such that
|δF+ (Vi )| ≥ 1 for i = 1, 2, . . . , k
and
w(G̃)
is minimized.
G̃∈C(F ) lb(G̃)
max
P
Here we used the notation that for a connected component G̃ of (V, F ), w(G̃) = e∈E(G̃) w(e)
P
(summation over the edges) and lb(G̃) = v∈V (G̃) lb(v) (summation over the vertices). We say
that an algorithm for Local-Connectivity ATSP is α-light on G if it is guaranteed, for any
partition, to find a solution F such that for every component G̃ ∈ C(F ),
w(G̃)
≤ α.
lb(G̃)
P
In [Sve15], lb is defined as lb(v) = e∈δ+ (v) w(e)x∗e ; note that lb(V ) = OP T in this case.
We remark that we use the “α-light” terminology to avoid any ambiguities with the concept
4
of approximation algorithms (an α-light algorithm does not compare its solution to an optimal
solution to the given instance of Local-Connectivity ATSP).
Perhaps the main difficulty of ATSP is to satisfy the connectivity requirement, i.e., to select
an Eulerian subset F of edges which connects the whole graph. Local-Connectivity ATSP
relaxes this condition – we only need to find an Eulerian set F that crosses the k cuts defined
by the partition. This makes it intuitively an “easier” problem than ATSP. Indeed, an αapproximation algorithm for ATSP (with respect to the Held-Karp relaxation) is trivially an
α-light algorithm for Local-Connectivity ATSP for an arbitrary lb function with lb(V ) = OP T :
just return the same Eulerian subset F as the algorithm for ATSP; since the set F connects
the graph, we have maxG̃∈C(F ) w(G̃)/ lb(G̃) = w(F )/ lb(V ) ≤ α. Perhaps more surprisingly,
the main technical theorem of [Sve15] shows that the two problems are equivalent up to small
constant factors.
Theorem 1.2 ([Sve15]). Let A be an algorithm for Local-Connectivity ATSP. Consider an
ATSP instance G = (V, E, w), and let OPT denote the optimum value of the Held-Karp relaxation. If A is α-light on G, then there exists a tour of G with value at most 5α OPT. Moreover,
for any ε > 0, a tour of value at most (9 + ε)α OPT can be found in time polynomial in the
number n = |V | of vertices, in 1/ε, and in the running time of A.
In other words, the above theorem says that in order to approximate an ATSP instance G, it is
sufficient to devise a polynomial-time algorithm to calculate a lower bound lb and a polynomial
time algorithm for Local-Connectivity ATSP that is O(1)-light on G with respect to this lb
function. Our main result is proved using this framework.
1.2
Technical overview
Singleton partition. Let us start by outlining the fundamental ideas of our algorithm and
comparing it to [Sve15] for the special case of Local-Connectivity
ATSP when all partition
P
classes Vi are singletons. For unit weights, the choice lb(v) = e∈δ+ (v) w(e)x⋆e = x⋆ (δ+ (v)) in
[Sve15] is a natural one: intuitively, every node is able to pay for its outgoing edges. We can
thus immediately give an algorithm for this case: just select an arbitrary integral solution z to
⋆ +
the circulation problem with node capacities 1 ≤ z(δ+ (v))
P ≤ ⌈x (δ (v))⌉. Then for any v we
+
⋆
+
⋆
+
have z(δ (v)) ≤ x (δ (v)) + 1 ≤ 2x (δ (v)) and hence e∈δ+ (v) w(e)ze ≤ 2 lb(v), showing that
z is a 2-light solution.
The same choice of lb does not seem to work in the presence of two different edge costs.
Consider
a case when every expensive edge carries only a small fractional amount of flow. Then
P
⋆
e∈δ+ (v) w(e)xe can be much smaller than the expensive edge cost w1 , and thus the vertex v
would not be able to “afford” even a single outgoing expensive edge. To resolve this problem,
we bundle small fractional amounts of expensive flow, channelling them to reach a small set
of terminals. This is achieved via Theorem 2.1, a flow result which might be of independent
interest. It shows that within the fractional Held-Karp solution x⋆ , we can send the flow from
an arbitrary edge set E ′ to a sink set T with |T | ≤ 8x⋆ (E ′ ); in fact, T can be any set minimal
for inclusion such that it can receive the total flow from E ′ . We apply this theorem for E ′ = E1 ,
the set of expensive edges; let f be the flow from E1 to T , and call elements of T terminals.
Now, whenever an expensive edge is used, we will “force” it to follow f to a terminal in T ,
where it can be paid for. Enforcement is technically done by splitting the vertices into two
copies, one carrying the f flow and the other the rest. Thus we obtain the split graph Gsp and
split fractional optimal solution x⋆sp .
The design of the split graph is such that every walk in it which starts with an expensive edge
must proceed through cheap edges until it reaches a terminal before visiting another expensive
5
edge. In our terminology, expensive edges create “debt”, which must be paid off at a terminal.
Starting from an expensive edge, the debt must be carried until a terminal is reached, and no
further debt can be taken in the meantime. The bound on the number of terminals guarantees
that we can assign a lower bound function lb with lb(V ) ≤ OPT such that (up to a constant
factor) cheap edges are paid for locally, at their heads, whereas expensive edges are paid for
at the terminals they are routed to. Such a splitting easily solves Local-Connectivity ATSP
for the singleton partition: find an arbitrary integral circulation zsp in the split graph with an
upper bound zsp (δ+ (v)) ≤ ⌈2x⋆sp (δ+ (v))⌉ on every node, and a lower bound 1 on whichever copy
of v transmits more flow. Note that 2x⋆sp is a feasible fractional solution to this problem. We
map zsp back to an integral circulation z in the original graph by merging the split nodes, thus
obtaining a constant-light solution.
Arbitrary partitions. Let us now turn to the general case of Local-Connectivity ATSP,
where the input is an arbitrary partition V = V1 ∪ . . . ∪ Vk . For unit weights this is solved
in [Sve15] via an integer circulation problem on a modified graph. Namely, an auxiliary node
Ai is added to represent each partition class Vi , and one unit of in- and outgoing flow from
Vi is rerouted through Ai . In the circulation problem, we require exactly one in- and one
outgoing edge incident to Ai to be selected. When we map the solution back to the original
graph, there will be one incoming and one outgoing arc from every set Vi (thus satisfying the
connectivity requirement) whose endpoints inside Vi violate the Eulerian condition. In [Sve15]
every Vi is assumed to be strongly connected, and therefore we can “patch up” the circulation
by connecting the loose endpoints by an arbitrary path inside Vi . This argument easily gives a
3-light solution.
Let us observe that the strong connectivity assumption is in fact not needed for the result in
[Sve15]. Indeed, given a component Vi which is not strongly connected, consider its decomposition into strongly connected (sub)components, and pick a Ui ⊆ Vi which is a sink (i.e. it has
no edges outgoing to Vi \ Ui ). We proceed by rerouting 1 unit of flow through a new auxiliary
vertex just as in that algorithm, but we do this for Ui instead. This guarantees that Ui has at
least one outgoing edge in our solution, and that edge must leave Vi as well.
Our result for two different edge weights takes this observation as the starting point, but the
argument is much more complicated. We will find an integer circulation in a graph based on
the split graph Gsp , and for every 1 ≤ i ≤ k, there will be an auxiliary vertex Ai representing a
certain subset Ui ⊆ Vi . These sets Ui will be obtained as sink components in certain auxiliary
graphs we construct inside each Vi . This construction is presented in Section 3.2; we provide a
roadmap to the construction at the beginning of that section.
2
The Flow Theorem
In this section we prove our main flow decomposition result. As indicated in Section 1.2, we
will use it to channel the flow from the expensive edges E1 to a small set of terminals T (where
|T | ≤ 8w(E1 )). We will use the theorem stated below by moving the tail of every edge in E1
to a new vertex s. If w(E1 ) ≥ 1, then the constraints of the Held-Karp relaxation guarantee
condition (1). The details of the reduction are given in Lemma 3.4.
Theorem 2.1. Let D = (V ∪ {s}, E) be a directed graph, let c : E → R+ be a nonnegative
capacity vector, and let s be a source node with no incoming edges, i.e., δ− (s) = ∅. Assume that
for all ∅ =
6 S ⊆ V we have
c(δ− (S)) ≥ max{1, c(δ+ (S))}.
(1)
6
Consider a set T ⊆ V such that there exists a flow f ≤ c of value c(δ+ (s)) from the source s to
the sink set T , and T is minimal subject to this property.3 Then |T | ≤ 8c(δ+ (s)).
The proof of this theorem can be skipped on first reading, as the algorithm in Section 3 only
uses it in a black-box manner.
Proof. Fix a minimal set T and denote k = |T |. Our goal is to prove that k ≤ 8c(δ+ (s)). We
know that there exists a flow of value c(δ+ (s)) from s to T . For any such flow f we define
its imbalance sequence to be the sequence of values z(t) = f (δ− (t)) − f (δ+ (t)) ∈ R+ for all
t ∈ T sorted in non-increasing order. We select the flow f which maximizes the imbalance
sequence (lexicographically). We write T = {t1 , . . . , tk } so that z(t1 ) ≥ z(t2 ) ≥ . . . ≥ z(tk );
denote zi = z(ti ) for brevity. By minimality of T we have z(tk ) > 0. The following is our main
technical lemma.
Lemma 2.2. Let ℓ be the number of t ∈ T with z(t) ≥ 14 , i.e., z1 ≥ ... ≥ zℓ ≥
zk . Then we have
ℓ
X
1
zi .
(k − ℓ) ≤
4
1
4
> zℓ+1 ≥ ... ≥
i=1
In other words, the number of terminals with small imbalance is not much more than the sum
of large imbalances.
Assuming this lemma, the main theorem follows immediately, since we have 14 k = 41 ℓ + 14 (k −
P
P
ℓ) ≤ ℓi=1 zi + ℓi=1 zi ≤ 2c(δ+ (s)), i.e., k ≤ 8c(δ+ (s)).
The remainder of this section is devoted to the proof of the technical lemma. Let us first
give an outline. We analyze the residual capacity (with respect to f ) of certain cuts that must
be present due to the lexicographic property. First of all, there must be a saturated cut (that is,
one of 0 residual in-capacity) A ⊆ V containing all large terminals (i.e., those with imbalance
at least 1/4) but no small ones (Claim 3). Next, consider an arbitrary small terminal i. Also
by the maximality property, it is not possible to increase the value of zi to 1/4 by rerouting
flow from other small terminals to i. Hence there must be a cut Bi , disjoint from A, which
contains ti as the only terminal and has residual in-capacity less than 1/4−zi in D\A (Claim 4).
As an illustration of the argument, let us assume that these sets Bi are pairwise disjoint. It
follows from (1) that the residual in-capacity of Bi is at least 1 − zi (Claim 2). Hence every
set Bi must receive 3/4 units of its residual in-capacity
from A. On the other hand, (1) upperPℓ
bounds the residual out-capacity of A by 2 i=1 zi (Claim 1). These together give a bound
Pℓ
3
i=1 zi . Recall however that we assumed that all sets Bi are disjoint. Since these
4 (k − ℓ) ≤ 2
sets may in fact overlap, the proof needs to be more careful: instead of sets Bi , we argue with
the sets Bi \ (∪j6=i Bj ) (nonempty as containing ti ), and the union of pairwise intersections B ∗ ;
thus instead of 3/8, we get a slightly worse constant 1/4.
Proof of Lemma 2.2. First note that the claim is trivial if ℓ = k, so assume ℓ < k. For an
−
arc e = (u, v), we let ←
e = (v, u) denote the reverse arc. We define the residual graph Df =
−
−
(V ∪ {s}, Ef ) with Ef = {e ∈ E : f (e) < c(e)} ∪ {e : ←
e ∈ E, f (←
e ) > 0}. The residual capacity
−
for the first set of arcs is defined as cf (e) = c(e) − f (e), and for the second set as cf (e) = f (←
e ).
−
+
For any set X ⊆ V , and a disjoint Y ⊆ V , let δf (X), δf (X) and δf (X, Y ) denote the capacities
3
That is, the maximum flow value from s to any proper subset T ′ ( T is smaller than c(δ + (s)).
7
of the respective cuts in the residual graph Df of f , i.e.,
δf− (X) =
X
cf (e) = c(δ− (X)) − f (δ− (X)) + f (δ+ (X)),
e∈δ− (X)
δf+ (X) =
X
cf (e) = c(δ+ (X)) − f (δ+ (X)) + f (δ− (X)),
e∈δ+ (X)
δf (X, Y ) =
X
cf (e) = c(δ(X, Y )) − f (δ(X, Y )) + f (δ(Y, X)).
e∈δ(X,Y )
The next two claims derive simple bounds from (1) on the residual in-and out-capacities of cuts.
P
Claim 1. If X ⊆ V , then δf− (X) + 2 t∈T ∩X z(t) ≥ δf+ (X).
Proof.
δf− (X) − δf+ (X) = c(δ− (X)) − c(δ+ (X)) − 2 f (δ− (X)) − f (δ+ (X))
X
X
= c(δ− (X)) − c(δ+ (X)) − 2
z(t) ≥ −2
z(t).
t∈T ∩X
t∈T ∩X
The equality is by flow conservation. The inequality is by (1). The claim follows.
Claim 2. Consider X ⊆ V such that X ∩ T = {ti } for some 1 ≤ i ≤ k. Then δf− (X) ≥ 1 − zi .
Proof.
δf− (X) = c(δ− (X)) − f (δ− (X)) − f (δ+ (X)) ≥ 1 − zi .
Here we used (1) and the flow conservation f (δ− (X)) − f (δ+ (X)) = zi (as the single sink
contained in X is ti ).
The next claim shows that the large terminals can be separated from the small ones by a
cut of residual in-degree 0. This follows from the lexicographically maximal choice, and is not
a particular property of the threshold 1/4 (it remains true if we replace ℓ by any 1 ≤ j ≤ k).
Claim 3. There exists a set A ⊆ V with A ∩ T P
= {t1 , ..., tℓ } (i.e., A contains exactly the large
terminals) such that δf− (A) = 0, and δf+ (A) ≤ 2 ℓi=1 zℓ .
Proof. If ℓ = 0, then we can choose A = ∅. So assume 0 < ℓ < k. Consider the maximum flow
in the residual graph Df from the source set {tℓ+1 , ..., tk } to the sink set {t1 , ..., tℓ }. If its value
is positive, then there exists a path P in Df from ti to tj for some i > ℓ and j ≤ ℓ (without loss
of generality it contains no other terminals). Set ε = min{z(ti ), min(u,v)∈P cf (u, v)} > 0. Then
the s-T flow f ′ = f + ε · 1P has a lexicographically larger imbalance sequence than f because
zi is increased without decreasing any other of the large imbalances, a contradiction. So there
must be a cut A ⊆ V with A ∩ T = {t1 , ..., tℓ } and δf− (A) = 0.4 The second part follows by the
first via Claim 1.
Claim 4. For any ℓ + 1 ≤ i ≤ k (i.e., ti is a small terminal) there exists a set Bi ⊆ V \ A such
that Bi ∩ T = {ti } and δf− (Bi ) − δf (A, Bi ) < 41 − zi .
4
Note that s ∈
/ A since Df contains a path from tk to s.
8
Proof. If k = ℓ + 1, then we can choose Bk = V \ A; we have δf− (Bk ) − δf (A, Bk ) = 0 since all
arcs leaving the source s are saturated. So assume k − ℓ ≥ 2. Consider the maximum flow from
the source set {tℓ+1 , ..., tk } \ {ti } to the sink ti in the graph Df \ (A ∪ {s}).
If its value is at least 14 − zi , then we will get a contradiction by increasing the imbalance
of ti to at least 41 without changing any of the large imbalances. Namely, let g be a flow from
{tℓ+1 , ..., tk } \ {ti } to ti of value 41 − zi . Consider the vector f + g. There are two possible cases:
• If f + g is still an s-T flow, i.e., if for all j we have g(δ+ (tj )) − g(δ− (tj )) ≤ z(tj ), then it
has a lexicographically larger imbalance sequence than f , a contradiction.
• Otherwise pick the maximum α > 0 such that f + αg is still an s-T flow, i.e., for all j we
have α (g(δ+ (tj )) − g(δ− (tj ))) ≤ z(tj ), with equality for some j. This means that f + αg
is an s-T flow where at least one terminal tj has zero imbalance, i.e., it can be removed
from the set T , contradicting its minimality.
So there must be a cut Bi ⊆ V \ (A ∪ {s}) such that Bi ∩ T = {ti } and
1
− zi > δf (V \ (A ∪ {s} ∪ Bi ), Bi ) = δf− (Bi ) − δf (A, Bi ) − δf (s, Bi ).
4
The claim follows by δf (s, Bi ) = 0, which holds since all edges in δ+ (s) are saturated in f .
P
The argument uses the bound δf+ (A) ≤ 2 ℓi=1 zℓ and the fact that all the Bi ’s must receive
a large part of their residual in-degrees from A. Since the sets Bi overlap, we have to take their
intersections into account. Let us therefore define
[
B ⋆ :=
(Bi ∩ Bj ) ⊆ V \ (A ∪ {s})
i,j>ℓ,i6=j
as the set of vertices contained in at least two sets Bi . Let α⋆ := δf (A, B ⋆ ), αi := δf (A, Bi \B ⋆ ),
and βi := δf (B ⋆ , Bi \ B ⋆ ) for each ℓ + 1 ≤ i ≤ k.
Claim 5. For each ℓ + 1 ≤ i ≤ k we have
3
4
< αi + βi .
Proof. Note that (Bi \ B ⋆ ) ∩ T = {ti }, and thus δf− (Bi \ B ⋆ ) ≥ 1 − zi by Claim 2. From this we
can see
1 − zi ≤ δf− (Bi \ B ⋆ )
≤ δf− (Bi ) − δf (A, Bi ) + δf (A, Bi \ B ⋆ ) + δf (B ⋆ , Bi \ B ⋆ )
<
1
− zi + αi + βi ,
4
where the second inequality follows because an edge entering Bi \ B ⋆ either enters Bi from
outside of A, or enters Bi \ B ⋆ from B ⋆ , or enters Bi \ B ⋆ from A.
For the residual in-degree of the set B ⋆ , we apply the trivial bound
δf− (B ⋆ )
k
X
1
≤ δf (A, B ) +
δf− (Bi ) − δ(A, Bi ) ≤ α⋆ + (k − ℓ).
4
⋆
i=ℓ+1
The last estimate is by the choice of the sets Bi in Claim 4. For the residual out-degree, we get
δf+ (B ⋆ ) ≥
k
X
i=ℓ+1
k
X
3
βi > (k − ℓ) −
αi ,
4
i=ℓ+1
9
using Claim 5. Applying Claim 1 to B ⋆ and noting that B ⋆ ∩ T = ∅ gives δf− (B ⋆ ) ≥ δf+ (B ⋆ ).
Putting Claim 3 and the above two bounds together, we conclude that
2
ℓ
X
zi ≥
δf+ (A)
⋆
≥α +
i=1
Lemma 2.2 now follows.
3
k
X
i=ℓ+1
1
αi ≥ (k − ℓ).
2
(2)
Algorithm for Local-Connectivity ATSP
We prove our main result in this section. Our claim for ATSP follows from solving LocalConnectivity ATSP:
Theorem 3.1. There is a polynomial-time 100-light algorithm for Local-Connectivity ATSP on
graphs with two edge weights.
Together with Theorem 1.2, this implies our main result:
Theorem 3.2. For any graph with two edge weights, the integrality gap of its Held-Karp relaxation is at most 500. Moreover, we can find an 901-approximate tour in polynomial time.
The factor 500 comes from 5 · 100. In Theorem 1.2, we select ε such that (9 + ε) · 100 ≤ 901. Our
proof of Theorem 3.1 proceeds as outlined in Section 1.2. In Section 3.1, we give an algorithm
for calculating lb and define the split graph which will be central for finding light solutions. In
Section 3.2, we then show how to use these concepts to solve Local-Connectivity ATSP for any
given partitioning of the vertices.
Recall that the edges are partitioned into the set E0 of cheap edges and the set E1 of
expensive edges. Set x⋆ to be an optimal solution to the Held-Karp relaxation. We start by
noting that the problem is easy if x⋆ assigns very small total fractional value to expensive
edges. In that case, we can easily reduce the problem to the unweighted case which was solved
in [Sve15].
Lemma 3.3. There is a polynomial-time 6-light algorithm for Local-Connectivity ATSP for
graphs where x⋆ (E1 ) < 1.
Proof. If x⋆ (E1 ) = 0, then just apply the standard 3-light polynomial-time algorithm for unweighted graphs [Sve15]. So suppose that 0 < x⋆ (E1 ) < 1. Then clearly the graph (V, E0 ) is
strongly connected, i.e. every pair of vertices is connected by a directed path of cheap edges
(of length at most n − 1). Thus each expensive edge (u, v) can be replaced by such a u-v-path
P (u, v). Let us obtain a new circulation x′ from x⋆ by replacing all expensive edges in this way,
i.e.,
X
x⋆(u,v) · 1P (u,v) .
x′ = x⋆ |E0 +
(u,v)∈E1
To bound the cost of
x′ ,
note that
x⋆ (E0 )
= x⋆ (E) − x⋆ (E1 ) > n − 1 > (n − 1)x⋆ (E1 ) and thus
w(x′ ) ≤ w0 · x⋆ (E0 ) + (n − 1)w0 · x⋆ (E1 ) ≤ 2w0 · x⋆ (E0 ) ≤ 2w(x⋆ ).
By construction, x′ is a feasible solution for the Held-Karp relaxation and supp(x′ ) ⊆ E0 .
Therefore we can use it in the standard 3-light polynomial-time algorithm for the unweighted
graph (V, E0 ). Together with the bound w(x′ ) ≤ 2w(x⋆ ) this gives a 6-light algorithm.
For the rest of this section, we thus assume x⋆ (E1 ) ≥ 1. Our objective is to define a function
lb : V → R+ such that lb(V ) ≤ OPT = w(x⋆ ) and then show how to, given a partition
V = V1 ∪ ... ∪ Vk , find an Eulerian set of edges F which crosses all Vi -cuts and is O(1)-light with
respect to the defined lb function.
10
3.1
Calculating lb and constructing the split graph
First, we use our flow decomposition technique to find a small set of terminals T such that it
is possible to route a certain flow f from endpoints of all expensive edges to T . Next, we use f
and T to calculate the function lb and to construct a split graph Gsp , where each vertex of G is
split into two.
Finding terminals T and flow f . We use Theorem 2.1 to obtain a small-enough set of
terminals T and a flow f which takes all flow on expensive edges to this set T . More precisely,
we have the following corollary of Theorem 2.1.
Lemma 3.4. There exist a vertex set T ⊆ V and a flow f : E → R+ from source set {tail(e) :
e ∈ E1 } to sink set T of value x⋆ (E1 ) such that:
• |T | ≤ 8x⋆ (E1 ),
• f ≤ x⋆ ,
• f saturates all expensive edges, i.e., f (e) = x⋆e for all e ∈ E1 ,
• for each t ∈ T , f (E0 ∩ δ+ (t)) = 0 and f (δ− (t)) > 0.
Moreover, T and f can be computed in polynomial time.
Proof. We construct G′ to be G with a new vertex s, where the tail of every expensive edge
is redirected to be s. Formally, V (G′ ) = V ∪ {s} and E(G′ ) = E0 ∪ {(s, head(e)) : e ∈ E1 }.
The capacity vector c is obtained from x⋆ by just following this redirection, i.e., for any edge
e′ ∈ E(G′ ) we define c(e′ ) = x⋆e , where e ∈ E(G) is taken to be the preimage of e′ in G.
Clearly c(δ+ (s)) = x⋆ (E1 ), and s ∈ V (G′ ) has no incoming edges. To see that condition (1)
of Theorem 2.1 is satisfied, recall that for every ∅ =
6 S ( V (G) we have x⋆ (δ+ (S)) = x⋆ (δ− (S)) ≥
1; redirecting the tail of some edges to s can only reduce the outdegree or increase the indegree
of S, i.e., c(δ− (S)) ≥ x⋆ (δ− (S)) = x⋆ (δ+ (S)) ≥ c(δ+ (S)). This gives condition (1) for all
sets S ( V (G′ ) − s; for S = V (G′ ) − s, note that c(δ− (S)) = x⋆ (E1 ) ≥ 1 = max{1, 0} =
max{1, c(δ+ (S))} since we assumed that x⋆ (E1 ) ≥ 1.
From Theorem 2.1 we obtain a vertex set T with |T | ≤ 8x⋆ (E1 ) and a flow f ′ : E(G′ ) → R+
from s to T of value x⋆ (E1 ) with f ′ ≤ c. We can assume f ′ (δ+ (t)) = 0 for all t ∈ T : in a
path-cycle decomposition of f ′ we can remove all cycles and terminate every path at the first
terminal it reaches. The flow f is obtained by mapping f ′ back to G, i.e., taking each f (e) to
be f ′ (e′ ), where e′ is the image of e. Note that f ′ must saturate all outgoing edges of s, so f
saturates all expensive edges. For the last condition, the part f (E0 ∩ δ+ (t)) = 0 is implied by
f ′ (δ+ (t)) = 0, and for the part f (δ− (t)) > 0, note that if f (δ− (t)) = 0, then we could have
removed t from T .
Note that such a set T can be found in polynomial time: starting from T = V (for which
the required flow exists: consider f ′ = c|δ+ (s) , the restriction of c to δ+ (s)), we remove vertices
from T one by one until we obtain a minimal set T such that there exists a flow of value x⋆ (E1 )
from s to T .
Definition of lb.
defined as follows:
We set lb : V → R+ to be a scaled-down variant of lb : V → R+ which is
(
w0 · x⋆ (δ− (v))
lb(v) :=
w0 · x⋆ (δ− (v)) + w1 · ⌈f (δ− (t))⌉
11
if v ∈
/ T,
if v ∈ T.
The definition of lb is now simply lb(v) = lb(v)/10. The scaling-down is done so as to satisfy
lb(V ) ≤ OPT (see Lemma 3.5). Clearly we have lb(v) ≥ w0 for all v ∈ V and lb(t) ≥ w1 + w0 ≥
w1 for terminals t ∈ T .
The intuition behind this setting of lb is that we want to pay for each expensive edge e ∈ E1
in the terminal t ∈ T which the flow f “assigns” to e. Indeed, in the split graph we will reroute
flow (using f ) so as to ensure that any path which traverses e must then visit such a terminal t
to offset the cost of the expensive edge. As for the total cost of lb, note that if we removed the
rounding from its definition, then we would get lb(V ) ≤ 2 OPT, since
X
X
w0 ·
x⋆ (δ− (v)) + w1 ·
f (δ− (t)) = w0 · x⋆ (E0 ) + w0 · x⋆ (E1 ) + w1 · x⋆ (E1 ) ≤ 2w(x⋆ )
v∈V
v∈T
P
(here we used that f is of value x⋆ (E1 ) = t∈T f (δ− (t))). So, similarly to the 3-light algorithm
for unweighted metrics in [Sve15], the key is to argue that rounding does not increase this cost
too much. For this, we will take advantage of the small size of T . Details are found in the proof
of the following lemma.
Lemma 3.5. lb(V ) ≤ 10 · OPT .
Proof. The bound follows from elementary calculations:
X
X
lb(V ) = w0 ·
x⋆ (δ− (v)) + w1 ·
⌈f (δ− (t))⌉
v∈V
≤ w0 ·
X
t∈T
⋆
−
x (δ (v)) + w1 ·
v∈V
⋆
X
t∈T
⋆
f (δ− (t)) + 1
≤ w0 · x (E) + w1 · (x (E1 ) + |T |)
≤ w(x⋆ ) + w1 · 9x⋆ (E1 )
≤ 10w(x⋆ )
(recall that |T | ≤ 8x⋆ (E1 ) by Lemma 3.4).
Construction of the split graph. The next step is to reroute flow so as to ensure that all
expensive edges are “paid for” by the lb at terminals. To this end, we define a new split graph
and a split circulation on it (see also Fig. 1 for an example).
Definition 3.6. The split graph Gsp is defined as follows. For every v ∈ V we create two
copies v 0 and v 1 in V (Gsp ). For every cheap edge (u, v) ∈ E0 :
• if x⋆ (u, v) − f (u, v) > 0, create an edge (u0 , v 0 ) in E(Gsp ) with x⋆sp (u, v) = x⋆ (u, v) −
f (u, v),
• if f (u, v) > 0, create an edge (u1 , v 1 ) in E(Gsp ) with x⋆sp (u, v) = f (u, v).
For every expensive edge (u, v) ∈ E1 we create one edge (u0 , v 1 ) in E(Gsp ) with x⋆sp (u, v) =
f (u, v). Finally, for each t ∈ T we create an edge (t1 , t0 ) in E(Gsp ) with x⋆sp (t1 , t0 ) = f (δ− (t)).
The new edges are weighted as follows: images of edges in E0 have weight w0 , the images
of edges in E1 have weight w1 , and the new edges (t1 , t0 ) have weight 0. Let us denote the new
weight function by wsp .
Vertices v 0 will be called free vertices and vertices v 1 will be called debt vertices. Edges
entering a free vertex will be called free edges, and those entering a debt vertex will be called
debt edges.
12
A fundamental consequence of our construction is the following.
Fact 3.7. Consider any path P in Gsp such that the first edge of P is a debt edge and the
last one is a free edge or an expensive edge. Then P must go through a terminal, i.e., it must
contain an edge (t1 , t0 ) for some t ∈ T .
We also have the following two properties by design.
Fact 3.8. The vector x⋆sp : E(Gsp ) → R+ is a circulation in Gsp .
Proof. Intuitively, this is because our rerouting corresponds to taking every path v0 , v1 , ..., vk in
a path-cycle decomposition of f (where (v0 , v1 ) is an expensive edge and vk ∈ T is a terminal)
and, rather than placing it on the “free level” in Gsp (i.e., mapping it to v00 , v10 , ..., vk0 ), instead
1 , v 1 , v 0 . We can think of the edge (v 0 , v 1 ) as incurring a
routing it as follows: v00 , v11 , v21 , ..., vk−1
0 1
k k
debt, and of the edge (vk1 , vk0 ) as discharging this debt.
Now we proceed to give a formal proof. For all v ∈ V we will prove that x⋆sp (δ+ (v 0 )) =
⋆
xsp (δ− (v 0 )) and x⋆sp (δ+ (v 1 )) = x⋆sp (δ− (v 1 )).
Suppose v = t ∈ T . Then x⋆sp (δ− (t1 )) = f (δ− (t)). Also x⋆sp (δ+ (t1 )) = x⋆sp (t1 , t0 ) = f (δ− (t))
because of the property that f (E0 ∩ δ+ (t)) = 0. For t0 we have
x⋆sp (δ− (t0 )) = f (δ− (t)) + (x⋆ − f )(E0 ∩ δ− (t))
= f (E1 ∩ δ− (t)) + x⋆ (E0 ∩ δ− (t))
= x⋆ (δ− (t))
and
x⋆sp (δ+ (t0 )) = (x⋆ − f )(E0 ∩ δ+ (t)) + x⋆ (E1 ∩ δ+ (t))
= x⋆ (E0 ∩ δ+ (t)) + x⋆ (E1 ∩ δ+ (t))
= x⋆ (δ+ (t))
where the second inequality again follows by f (E0 ∩ δ+ (t)) = 0. This implies that x⋆sp (δ− (t0 )) =
x⋆ (δ− (t)) = x⋆ (δ+ (t)) = x⋆sp (δ+ (t0 )).
We turn to the case v ∈
/ T . Note that since all expensive edges are saturated by f , and their
tails are exactly the sources of f , we have
f (E0 ∩ δ+ (v)) = f (E0 ∩ δ− (v)) + x⋆ (E1 ∩ δ− (v)).
(3)
Since the left hand side is equal to x⋆sp (δ+ (v 1 )) and the right hand side is x⋆sp (δ− (v 1 )), we have
proved our claim for v 1 . For v 0 , note that the incoming edges of v0 are all cheap, and the
incoming flow is
x⋆sp (δ− (v 0 )) = (x⋆ − f )(E0 ∩ δ− (v)).
As for outgoing flow, v 0 is also the tail of all expensive edges whose tail in G was v, i.e.,
x⋆sp (δ+ (v 0 )) = (x⋆ − f )(E0 ∩ δ+ (v)) + x⋆ (E1 ∩ δ+ (v))
= x⋆ (δ+ (v)) − f (E0 ∩ δ+ (v))
= x⋆ (δ− (v)) − f (E0 ∩ δ− (v)) − x⋆ (E1 ∩ δ− (v))
= x⋆ (E0 ∩ δ− (v)) − f (E0 ∩ δ− (v))
= (x⋆ − f )(E0 ∩ δ− (v))
= x⋆sp (δ− (v 0 )),
where the third equality is by (3).
13
Gsp and x⋆sp
G and f
a0
1/3
a
1/3
1/3
1/3
1/3
e
b
0
2/3
d1
1/3
1/3
2/3
1/3
c
g
c
a1
2/3
1/3
1/3
d0
1/3
1/3
b0
1/3
0
1/3
d
1/3
e0
1/3
2/3
b1
e1
2/3
2/3
g0
0
1/3
1
c1
1/3
g1
1
Figure 1: An example of the construction of Gsp and x⋆sp from G, x⋆ and f .
Here x⋆ is 1/3 for the expensive edges (depicted as thick) and 2/3 for the
remaining (cheap) edges; the set T of terminals of the flow f is depicted in
black.
Fact 3.9. For each proper subset U ⊂ V we have x⋆sp (δ− ({v 0 , v 1 : v ∈ U })) = x⋆sp (δ+ ({v 0 , v 1 :
v ∈ U })) ≥ 1. In other words, the image of every cut in G is still crossed by at least one unit
of x⋆sp .
Proof. This follows since x⋆ (δ+ (U )) = x⋆ (δ− (U )) ≥ 1, and contracting all pairs v 0 , v 1 would
yield back G and x⋆ .
3.2
Solving Local-Connectivity ATSP
Now our algorithm is given a partition V = V1 ∪ ... ∪ Vk of the original vertex set. The objective
is to output a set of edges F which crosses all Vi -cuts and is O(1)-light with respect to our lb
function.
We are aiming for a similar construction as in the unit-weight case: based on the split graph
Gsp , we construct an integer circulation problem with an auxiliary vertex Ai representing a
certain subset Ui ⊆ Vi for every 1 ≤ i ≤ k. We then map its solution back to the original
graph and patch up the loose endpoints inside every Ui by a path. However, we have to account
for the following difficulties: (i) an edge leaving Ui should also leave Vi ; (ii) debt should not
disappear inside Ui : if the edge entering it carries debt but the edge leaving does not, we must
make sure this difference can be charged to a terminal in Ui ; (iii) the path used inside Ui must
pay for all expensive edges it uses.
All three issues can be appropriately tackled by defining an auxiliary graph inside Vi . Edges
of the auxiliary graph represent paths containing one expensive edge and one terminal (which
can pay for themselves); however, these paths may not map to paths in the split graph. We
select the subset Ui ⊆ Vi as a sink component in the auxiliary graph. For convenience, Figs. 2
and 3 give an overview of the different steps, graphs and flows used by our algorithm.
Construction of auxiliary graphs and modification of split graph. Our first step is to
construct an auxiliary graph for each component Vi . The strong-connectivity structure of this
graph will guide our algorithm.
Definition 3.10. The auxiliary graph Gaux
is a graph with vertex set Vi and the following edge
i
set: for u, v ∈ Vi , (u, v) ∈ E(Gaux
)
if
any
of
the
following three conditions is satisfied:
i
• there is a cheap edge (u, v) ∈ E0 ∩ G[Vi ] inside Vi , or
14
name
x⋆
x⋆sp
x′sp
′
ysp
ysp
y
F
graph
G
Gsp
G′sp
G′sp
Gsp
G
G
circulation
yes
yes
yes
yes
no
no
yes
integral
no
no
no
yes
yes
yes
yes
obtained from the previous by
solving LP (Section 1.1)
splitting (Definition 3.6)
redirecting edges to Ai
rounding to integrality (Lemma 3.12)
redirecting edges back from Ai
mapping back to G
adding walks Pi
Figure 2: This table summarizes the various circulations and pseudo-flows
that appear in our algorithm, in order.
• there is a u-v-path in G[Vi ] whose first edge is expensive and all other edges are cheap,
and v ∈ T is a terminal – we then call the edge (u, v) ∈ E(Gaux
i ) a postpaid edge – or
• there is a u-v-path in G[Vi ] whose last edge is expensive and all other edges are cheap, and
u ∈ T is a terminal – we then call the edge (u, v) ∈ E(Gaux
i ) a prepaid edge.
Define the preimage of such an edge (u, v) ∈ E(Gaux
i ) to be a shortest path inside Vi as above
(in the first case, a single edge).
Now, for each i consider a decomposition of Gaux
into strongly connected components.5 Let
i
Ui ⊆ Vi be the vertex set of a sink component in this decomposition. That is, there is no edge
aux is constructed based only on
from Ui to Vi \ Ui in the auxiliary graph Gaux
i . Note that Gi
the original graph G and not the split graph Gsp . However, we will solve Local-Connectivity
ATSP by solving an integral circulation problem on G′sp : a modification of the split graph Gsp ,
described as follows.
For each i, define Uisp = {v 0 , v 1 : v ∈ Ui } ⊆ V (Gsp ) to be the set of vertices in the split graph
corresponding to Ui . (Note that Uisp may not be strongly connected in Gsp .) We are going to
reroute part of the x⋆sp flow going in and out of Uisp to a new auxiliary vertex Ai . While the
3-light algorithm for unit-weight graphs rerouted flow from all boundary edges of a component
Ui (see Section 1.2), here we will be more careful and choose only a subset of boundary edges
of Uisp to be rerouted.
To this end, select a subset of edges Xi− ⊆ δ− (Uisp ) with x⋆sp (Xi− ) = 1/2 such that either
all edges in Xi− are debt edges, or all are free edges. This is possible since x⋆sp (Uisp ) ≥ 1 by
Fact 3.9.6
We define the set of outgoing edges Xi+ ⊆ δ+ (Uisp ) to be, intuitively, the edges over which
the flow that entered Uisp by Xi− exits Uisp . That is, consider an arbitrary cycle decomposition
of the circulation x⋆sp , and look at the set of cycles containing the edges in Xi− . We define Xi+
as the set of edges on these cycles that first leave Uisp after entering Uisp on an edge in Xi− ;
clearly, x⋆sp (Xi+ ) = 1/2.7
Let gi denote the flow on these cycles connecting the heads of edges in Xi− and the tails of
edges in Xi+ . We will use the following claim later in the construction.
Fact 3.11. Assume all edges in Xi− are debt edges but e ∈ Xi+ is a free edge or an expensive
edge. Then there exists a path in Gsp [Ui ] between a vertex t0 (for some terminal t ∈ T ) and the
tail of e, made up of only cheap edges.
5
Note that we decompose the vertex set Vi , but with respect to the edge set E(Gaux
), not E(G[Vi ]).
i
To obtain exactly 1/2, we might need to break an edge up into two copies, dividing its x⋆sp -value between
them appropriately, and include one copy in Xi− but not the other; we omit this for simplicity of notation, and
assume there is such an edge set with exactly x⋆sp (Xi− ) = 1/2.
7
Again, we might need to split some edges into two copies.
6
15
Gaux
i
G[Vi ]
Ui
Gsp and x⋆sp on Vi
G′sp and x′sp on Vi
Xi+
1
1
.5
Ai
.5
.5
1
.5
.5
Xi+
.5
.5
.5
.5
.5 .5
.5
.5
.5
1
.5
.5
.5
.5
1
Xi−
.5
.5
.5
.5
.5
Xi−
′ on V
G′sp and ysp
i
G and F on Vi
Xi+
0
0
0
1
Ai
0
0
0
1
0
Xi−
0
1
1
1
0
Xi+
Xi−
Figure 3: An illustration of the different graphs and flows (restricted to a
single component Vi ) used in our algorithm for Local-Connectivity ATSP. The
black vertices depict terminals. In the split graphs, we depict debt vertices
by squares and free vertices by circles.
Proof. Consider the cycle in the cycle decomposition that contains e; it enters Ui on a debt
edge. Using Fact 3.7, this cycle fragment must contain an edge of the form (t1 , t0 ); pick the last
such edge. All edges that follow are free and cheap.
We now transform Gsp into a new graph G′sp and x⋆sp into new circulation x′sp as follows.
For every set Vi in the partition we introduce a new auxiliary vertex Ai and redirect all edges
in Xi− to point to Ai and those in Xi+ to point from Ai . We further subtract the flow gi inside
Uisp ; hence the resulting vector x′sp will be a circulation, with x′sp (δ− (Ai )) = 1/2. If Xi− is a set
of free edges, then we will say that Ai is a free vertex, otherwise we say that it is a debt vertex.
Transforming x′sp into an integral flow and obtaining our solution F .
we round x′sp to integrality while respecting degrees of vertices:
In the next step
′ on G′ satisfying the following condiLemma 3.12. There exists an integral circulation ysp
sp
tions:
′ (δ − (v)) ≤ ⌈2x⋆ (δ − (v))⌉ for each v ∈ V (G ),
• ysp
sp
sp
16
′ (δ − (A )) = 1 for each i.
• ysp
i
′ can be found in polynomial time.
Such a circulation ysp
Proof. The bounds are integral, and there exists a fractional circulation which satisfies them,
namely 2x′sp .
′ into an Eulerian set of edges F in the original graph G. We can
We will now transform ysp
think of this as a three-stage process.
First, we map all edges adjacent to the auxiliary vertices Ai back to their preimages in Gsp ,
′ an integral pseudo-flow y
obtaining from ysp
sp in Gsp . (We use the term pseudo-flow as now,
some vertices may not satisfy flow conservation.)
Second, we contract the two copies v 0 and v 1 of every vertex v ∈ V , thus mapping all edges
back to their preimages in G. (We remove all edges (t1 , t0 ) for t ∈ T .) This creates an integral
pseudo-flow y in G.
′ , now (in y) in each component U
Since the in- and out-degree of Ai were exactly 1 in ysp
i
there is a pair of vertices ui , vi which are the head and tail, respectively, of the mapped-back
edges adjacent to Ai . These are the only vertices where flow conservation in y can be violated.8
As the third step, to repair this, we route a walk Pi from ui to vi , as described below. Our
Eulerian set of edges F ⊆ E which we finally
Preturn is the integral pseudo-flow y plus the union
(over i) of all such walks Pi , i.e., 1F = y + i 1Pi .
It remains to describe how we route these paths. Fix i. Recall that Ui is strongly connected
in Gaux
i . We distinguish two cases:
′ (in G′ ) is a debt edge, then select a
• If Ai is a free vertex or the edge exiting Ai in ysp
sp
shortest ui -vi -path in Gaux
,
map
each
edge
of
this
path
to
its preimage path (see Definii
tion 3.10) and concatenate them to obtain a ui -vi -walk Pi in Vi .9
′ (in G′ ) is a free edge, then by Fact 3.11
• If Ai is a debt vertex but the edge exiting Ai in ysp
sp
there is a terminal t inside Ui , with a path from t to vi using only cheap edges.10 Proceed
as above to obtain a ui -t-walk and then append this cheap t-vi -path to it, obtaining a
ui -vi -walk Pi in Vi .
This concludes the description of the algorithm. In Sections 3.2.1 and 3.2.2 we prove that
the returned Eulerian set of edges F has the properties we desire, i.e.,
e
e of (V, F ) we have w(G)
e ≤ 10 · lb(G).
Lemma 3.13. For every connected component G
Lemma 3.14. For every component Vi we have |δF+ (Vi )| ≥ 1.
Lemmas 3.5 and 3.13 together prove that our algorithm is 100-light with respect to lb.
3.2.1
Bounding the cost – proof of Lemma 3.13
e to be a connected component of (V, F ). We want to prove the
For this section, let us fix G
e ≤ 10 · lb(G).
e
following lightness claim: w(G)
Intuitively, our solution is cheap because we built the split graph Gsp so as to ensure that
any circulation in Gsp which roughly respects the degree bounds given by x⋆sp has low cost, and
8
It is violated unless ui = vi .
Note that this walk may exit Ui , but it will stay inside Vi .
10
Map the path given by Fact 3.11 from Gsp to G.
9
17
because our pseudoflow y is not too far from being a circulation (in particular, the cost of walks
Pi can be accounted for). We make this argument precise below.
First, recall that the edges in F are edges in the integral pseudo-flow y and the edges of the
e if Pi is contained in G
e (note that a walk
walks P1 , P2 , . . . , Pk . For a walk Pi , we write Pi ⊆ G
e
e
is either contained in or disjoint from G because G is a connected component of (V, F )). Hence
X
X
e =
w(G)
ye w(e) +
w(Pi ).
(4)
e
e∈E(G)
e
i:Pi ⊆G
P
We start by analyzing the first term: e∈E(G)
e ye w(e). Recall that y is obtained by mapping
esp denote the pre-image of G
e in Gsp . It will be convenient to define
ysp from Gsp to G. Let G
the debt of ysp as the number of “unpaid” expensive edges in this component:
X
X
ysp (e) −
ysp (t1 , t0 ).
debt =
e
t∈V (G)∩T
e sp ):wsp (e)=w1
e∈E(G
esp is the difference between the ysp -flow on all expensive edges inside
That is, the debt of ysp on G
esp (those edges “incur debt”) and the ysp -flow on all edges (t1 , t0 ) inside G
esp (those edges
G
P
“discharge debt”). Note also that, by the definition of y from ysp , e∈E(Gesp ):wsp (e)=w1 ysp (e) =
P
y(e).
e
e∈E(G):w(e)=w
1
By the construction of Gsp we have the following upper bound on debt.
+
′
Lemma 3.15. Let e−
i and ei be the incoming and outgoing edges of Ai in ysp . For any i,
define
(
+
1 if e−
i is a debt edge and ei is a free edge,
bad(i) =
0 otherwise.
Then
debt ≤
X
bad(i)
e
i:Pi ⊆G
Moreover, if ysp is a circulation, then debt = 0.
Proof. In Gsp , the only edges of the form (u0 , v 1 ) are the expensive edges, and the only edges
of the form (u1 , v 0 ) are edges (t1 , t0 ) with t ∈ T . Hence, if ysp is a circulation, then
X
e
ysp (e) = ysp (δ− ({v 1 : v ∈ V (G)}))
esp ):wsp (e)=w1
e∈E(G
e
= ysp (δ+ ({v 1 : v ∈ V (G)}))
=
X
ysp (t1 , t0 ),
e
t∈V (G)∩T
which shows that debt = 0 in this case.
Now, if ysp is not a circulation, then one can imagine transforming it into one by connecting
+
the “dangling” endpoints of e−
i and ei in Gsp using a new “virtual” edge for each i; this edge
+
would be from a debt vertex to a free vertex if and only if e−
i is a debt edge and ei is a free edge.
Hence, by the calculations above, debt is upper-bounded by the number of such new edges in
our component. (We do not have equality as the new edges may also be introduced from free
vertices to debt vertices.)
18
P Using the above lemma, degree bounds and basic calculations, we upper-bound the term
e ye w(e):
e∈E(G)
Lemma 3.16. We have
P
e ye w(e)
e∈E(G)
e
≤ 6 lb(G).
Proof. We begin by writing:
X
X
ye + w1
ye w(e) ≤ w0
e
e∈E(G)
e
e∈E(G)
X
= w0
e
e∈E(G)
= w0
X
X
ysp (e)
esp ):wsp (e)=w1
e∈E(G
ye + w1 debt +
X
e
t∈V (G)∩T
X
y(δ− (v)) + w1
ysp (t1 , t0 )
′
ysp
(t1 , t0 ) + w1 · debt .
e
t∈V (G)∩T
e
v∈V (G)
We bound the first term. Note that for each v ∈ V we have y(δ− (v)) = ysp (δ− ({v 0 , v 1 }))
′ (δ − (v j )) ≤ ⌈2x⋆ (δ − (v j ))⌉
and x⋆ (δ− (v)) = x⋆sp (δ− ({v 0 , v 1 })). Lemma 3.12 guarantees that ysp
sp
′ (δ − ({v 0 , v 1 })) ≤ 2x⋆ (δ − (v)) + 2, because:
for j ∈ {0, 1}. This implies that ysp
• If v ∈ T , then v 1 has only one outgoing edge in Gsp , which goes to v 0 , and thus for any cir′ (δ − ({v 0 , v 1 })) =
culation c in Gsp we have c(δ− ({v 0 , v 1 })) = c(δ− (v 0 )). It follows that ysp
′
−
0
⋆
−
0
⋆
−
0
⋆
−
0
1
ysp (δ (v )) ≤ ⌈2xsp (δ (v ))⌉ ≤ 2xsp (δ (v ))+1 = 2xsp (δ ({v , v }))+1 = 2x⋆ (δ− (v))+1.
• If v 6∈ T , then there are no edges between v 0 and v 1 in Gsp , and thus for any circulation c in Gsp we have c(δ− ({v 0 , v 1 })) = c(δ− (v 0 )) + c(δ− (v 1 )). It similarly follows
′ (δ − ({v 0 , v 1 })) = y ′ (δ − (v 0 )) + y ′ (δ − (v 1 )) ≤ ⌈2x⋆ (δ − (v 0 ))⌉ + ⌈2x⋆ (δ − (v 1 ))⌉ ≤
that ysp
sp
sp
sp
sp
−
⋆
2xsp (δ (v 0 )) + 2x⋆sp (δ− (v 1 )) + 2 = 2x⋆sp (δ− ({v 0 , v 1 })) + 2 = 2x⋆ (δ− (v)) + 2.
′ except for the edges redirected from the auxiliary vertices in G
As ysp is the same as ysp
sp
(which may increase the in-degree of a vertex by at most 1),
′
y(δ− (v)) = ysp (δ− ({v 0 , v 1 })) ≤ ysp
(δ− ({v 0 , v 1 })) + 1 ≤ 2x⋆ (δ− (v)) + 3 ≤ 5x⋆ (δ− (v)),
where we used x⋆ (δ− (v)) ≥ 1 for the last inequality. Therefore
X
X
X
x⋆ (δ− (v)).
y(δ− (v)) ≤ 5 · w0
ye = w0
w0
e
v∈V (G)
e
v∈V (G)
e
e∈E(G)
We bound the second term similarly:
′
′
ysp
(t1 , t0 ) = ysp
(δ− (t1 )) ≤ ⌈2x⋆sp (δ− (t1 ))⌉ ≤ 2⌈x⋆sp (δ− (t1 ))⌉ = 2⌈f (δ− (t))⌉.
Plugging both in we get:
X
X
x⋆ (δ− (v)) + 2 · w1
ye w(e) ≤ 5 · w0
e
e∈E(G)
e
v∈V (G)
≤ 5 · w0
X
x⋆ (δ− (v)) + w1
⌈f (δ− (t))⌉ + w1 · debt
e
t∈(G)∩T
X
e
t∈V (G)∩T
e
v∈V (G)
e + w1 · debt .
= 5 lb(G)
X
19
⌈f (δ− (t))⌉ + w1 · debt
(For the final equality, recall the definition of the function
P lb.) We will be done if we show that
e
w1 · debt ≤ lb(G). By Lemma 3.15 we have debt ≤ i:Pi ⊆Ge bad(i). Whenever bad(i) = 1,
Fact 3.7 implies that Pi contains a terminal, and so lb(Pi ) ≥ w1 . Consequently,
X
X
w1 · debt ≤
w1 ≤
lb(Pi ) ≤ lb(G̃).
e
i:Pi ⊆G,bad(i)=1
e
i:Pi ⊆G,bad(i)=1
In the last inequality we used that the different Pi ’s are disjoint subsets of G̃ (each Pi is contained
in Vi ). The statement follows.
It remains to argue that the cost of the walks Pi can be accommodated.
Lemma 3.17. For every i we have w(Pi ) ≤ 4 · lb(Pi ).
Proof. This holds because in Pi , each vertex has low indegree and expensive edges can be offset
against terminals. Concretely, we have the following two claims:
Claim 6. For each v ∈ V (Pi ) we have |δP−i (v)| ≤ 4. (In particular, |E(Pi )| ≤ 4|V (Pi )|.)
(see
Proof. This follows from the fact that we select a shortest ui -vi -path (or ui -t-path) in Gaux
i
Definition 3.10). This implies that each vertex v appears as an internal vertex on at most one
preimage path of a prepaid edge (otherwise we could shortcut the path in Gaux
i ). Same applies
for postpaid edges. And similarly, each v appears as head of at most one preimage path of any
kind (single cheap edge, prepaid or postpaid). This means that v has indegree at most 3 on the
walk created from preimages of edges in Gaux
i . The path from t to ui can contribute a fourth
incoming edge.
Claim 7. The number of expensive edges on Pi is at most twice the number of terminals on Pi ,
i.e., |E1 ∩ E(Pi )| ≤ 2 · |T ∩ V (Pi )|.
Proof. Expensive edges appear only in preimage paths of prepaid or postpaid edges, one per
such preimage path, and such a path also contains a terminal (as its head or tail). A terminal
can only appear as head of one prepaid and as tail of one postpaid edge (otherwise, again, we
could shortcut the path in Gaux
i ).
Having these two claims, we can bound the cost:
w(Pi ) ≤ w0 · |E(Pi )| + w1 · |E1 ∩ E(Pi )|
≤ w0 · 4|V (Pi )| + w1 · 2|T ∩ V (Pi )|
≤ 4 (w0 · |V (Pi )| + w1 · |T ∩ V (Pi )|)
≤ 4 · lb(Pi ).
Now we have all the tools needed to prove our lightness claim, i.e., to upper-bound (4) by
e
10 · lb(G):
X
X
e =
w(G)
ye w(e) +
w(Pi )
e
e∈E(G)
e +
≤ 6 · lb(G)
X
e
i:Pi ⊆G
4 · lb(Pi )
e
i:Pi ⊆G
e + 4 · lb(G)
e
≤ 6 · lb(G)
e
= 10 · lb(G),
where line 2 is by Lemmas 3.16 and 3.17, and in line 3 we use that all walks Pi are disjoint and
e This completes the proof of Lemma 3.13.
in G.
20
3.2.2
Crossing the cuts – proof of Lemma 3.14
In this section we prove that our solution F ⊆ E crosses each component Vi . Let us recall that
aux from
Ui is a sink strongly-connected component of the graph Gaux
i , so there is no edge in Gi
Ui to Vi \ Ui . However, there can be such edges in G and therefore it might not be the case that
the edge in F which leaves Ui also leaves Vi . We will however argue that F does contain some
edge leaving Vi . Assume towards a contradiction that this is not true, i.e., that Vi is a union of
connected components of the graph (V, F ).
′ (in G′ ) with tail A ; let e
Fix i. Consider the (only) edge in ysp
i
sp be its image in ysp (in
sp
Gsp ) and e its image in y (in G). We have esp ∈ Xi+ ; the tail of e is vi ∈ Ui , and the head of e
is in Vi \ Ui (since we assumed that no edge of F leaves Vi ).
The following claim is the first, simplest example of how we use the structure of Gaux
to
i
reason about our solution F .
Claim 8. Any edge from Ui to Vi \ Ui in G is expensive. (Therefore e is expensive, and esp is
a debt edge.)
Proof. If there was such a cheap edge, it would also appear in Gaux
i . However, there is no edge
in Gaux
from
U
to
V
\
U
.
i
i
i
i
Define (Vi \ Ui )sp = {v 0 , v 1 : v ∈ Vi \ Ui } ⊆ V (Gsp ). We can now start traversing ysp inside
(Vi \ Ui )sp like an Eulerian graph (since ysp satisfies flow conservation in (Vi \ Ui )sp ), starting
from esp , until we return to Uisp .
Claim 9. We will not reach a terminal before returning to Uisp . (Therefore we will return on
a debt edge e′sp .)
Proof. If there was a terminal t ∈ T ∩ (Vi \ Ui ) such that there is a path from head(esp ) (a debt
vertex) to t1 in Gsp [Vi ], then (assuming without loss of generality that there is no other terminal
on this path) all edges of this path are cheap. Therefore e together with the image of this path
in G would give rise to a postpaid edge (tail(e), t) from Ui to Vi \ Ui in Gaux
i , a contradiction.
Since head(esp ) is a debt vertex and we do not visit a terminal, we will return on a debt edge.
Now we distinguish two cases. Suppose that the debt edge e′sp on which we return to Uisp is
′ (in G′ ) with head A . This means that we can
not the image in Gsp of the (only) edge in ysp
i
sp
sp
sp
keep following y inside Ui until we exit Ui again via some edge e′′sp (maybe e′′sp = esp ). By
Claim 8, e′′sp is expensive. So we have followed a path (a segment of y) which contains a debt
edge and later an expensive edge; this means (see Fact 3.7) that it must also contain a terminal
in between. Pick the last such terminal t; then the segment of y between t0 and tail(e′′sp ) must
consist of cheap (free) edges. This gives rise to a prepaid edge (t, head(e′′sp )) from Ui to Vi \ Ui
in Gaux
i , a contradiction.
So suppose instead that the debt edge e′sp on which we return to Uisp is, in fact, the image in
′ (in G′ ) with head A . This means that e′ ∈ X − . Recall that by
Gsp of the (only) edge in ysp
i
sp
sp
i
definition, Xi− consists only of debt or only of free edges, and therefore every edge in Xi− must
be a debt edge. By Fact 3.11, Gsp [Ui ] contains a path from a terminal t0 via cheap edges to the
expensive edge esp . Again, the image of this path in G gives rise to a prepaid edge (t, head(esp ))
from Ui to Vi \ Ui in Gaux
i , a contradiction. This concludes the proof of Lemma 3.14.
21
Acknowledgment
An earlier version of this paper has appeared in the proceedings of IPCO 2016 (the 18th Conference on Integer Programming and Combinatorial Optimization). This work has been published in Mathematical Programming Series B. The final publication is available at Springer via
http://dx.doi.org/10.1007/s10107-017-1195-7.
References
[AG15]
Nima Anari and Shayan Oveis Gharan. Effective-resistance-reducing flows and
asymmetric TSP. In Proceedings of the 56th Annual IEEE Symposium on Foundations of Computer Science (FOCS). IEEE, 2015.
[AGK+ 98] Sanjeev Arora, Michelangelo Grigni, David R. Karger, Philip N. Klein, and Andrzej
Woloszyn. A polynomial-time approximation scheme for weighted planar graph
TSP. In Proceedings of the Ninth Annual ACM-SIAM Symposium on Discrete Algorithms, SODA 1998, pages 33–41, 1998.
[AGM+ 10] Arash Asadpour, Michel X. Goemans, Aleksander Madry, Shayan Oveis Gharan,
and Amin Saberi. An O(log n/ log log n)-approximation algorithm for the asymmetric traveling salesman problem. In Proceedings of the Twenty-First Annual ACMSIAM Symposium on Discrete Algorithms, SODA 2010, pages 379–389, 2010.
[BK06]
Piotr Berman and Marek Karpinski. 8/7-approximation algorithm for (1, 2)-TSP.
In Proceedings of the Seventeenth Annual ACM-SIAM Symposium on Discrete Algorithms, SODA, pages 641–648, 2006.
[Blä04]
Markus Bläser. A 3/4-approximation algorithm for maximum ATSP with weights
zero and one. In Approximation, Randomization, and Combinatorial Optimization,
Algorithms and Techniques, 7th International Workshop on Approximation Algorithms for Combinatorial Optimization Problems, APPROX, pages 61–71, 2004.
[Chr76]
Nicos Christofides. Worst-case analysis of a new heuristic for the travelling salesman
problem. Technical report, DTIC Document, 1976.
[ES14]
Jeff Erickson and Anastasios Sidiropoulos. A near-optimal approximation algorithm
for asymmetric TSP on embedded graphs. In 30th Annual Symposium on Computational Geometry (SOCG), page 130, 2014.
[FGM82]
Alan M. Frieze, Giulia Galbiati, and Francesco Maffioli. On the worst-case performance of some algorithms for the asymmetric traveling salesman problem. Networks,
12(1):23–39, 1982.
[GKP95]
Michelangelo Grigni, Elias Koutsoupias, and Christos H. Papadimitriou. An approximation scheme for planar graph TSP. In 36th Annual Symposium on Foundations
of Computer Science, FOCS 1995, pages 640–645, 1995.
[GS11]
Shayan Oveis Gharan and Amin Saberi. The asymmetric traveling salesman problem
on graphs with bounded genus. In Proceedings of the twenty-second annual ACMSIAM symposium on Discrete Algorithms, pages 967–975. SIAM, 2011.
22
[GSS11]
Shayan Oveis Gharan, Amin Saberi, and Mohit Singh. A randomized rounding
approach to the traveling salesman problem. In IEEE 52nd Annual Symposium on
Foundations of Computer Science, FOCS 2011, pages 550–559, 2011.
[KLS15]
Marek Karpinski, Michael Lampis, and Richard Schmied. New inapproximability
bounds for TSP. J. Comput. Syst. Sci., 81(8):1665–1677, 2015.
[MS11]
Tobias Mömke and Ola Svensson. Approximating graphic TSP by matchings. In
IEEE 52nd Annual Symposium on Foundations of Computer Science, FOCS 2011,
pages 560–569, 2011.
[Muc12]
Marcin Mucha. 13/9-approximation for graphic TSP. In 29th International Symposium on Theoretical Aspects of Computer Science, STACS 2012, pages 30–41,
2012.
[PY93]
Christos H. Papadimitriou and Mihalis Yannakakis. The traveling salesman problem
with distances one and two. Math. Oper. Res., 18(1):1–11, 1993.
[SV14]
András Sebő and Jens Vygen. Shorter tours by nicer ears: 7/5-approximation for
the graph-TSP, 3/2 for the path version, and 4/3 for two-edge-connected subgraphs.
Combinatorica, 34(5):597–629, 2014.
[Sve15]
Ola Svensson. Approximating ATSP by relaxing connectivity. In Proceedings of the
56th Annual IEEE Symposium on Foundations of Computer Science (FOCS), 2015.
URL: http://arxiv.org/abs/1502.02051.
[WS11]
David P. Williamson and David B. Shmoys. The Design of Approximation Algorithms. Cambridge University Press, 2011.
23
| 8cs.DS
|
On interestingness measures of formal concepts
arXiv:1611.02646v2 [cs.AI] 19 Apr 2017
Sergei O. Kuznetsov, Tatiana Makhalova
National Research University Higher School of Economics
Kochnovsky pr. 3, Moscow 125319, Russia
[email protected], [email protected]
Abstract. Formal concepts and closed itemsets proved to be of big importance for knowledge discovery, both as a tool for concise representation of association rules and a tool for clustering and constructing
domain taxonomies and ontologies. Exponential explosion makes it difficult to consider the whole concept lattice arising from data, one needs to
select most useful and interesting concepts. In this paper interestingness
measures of concepts are considered and compared with respect to various aspects, such as efficiency of computation and applicability to noisy
data and performing ranking correlation.
Formal Concept Analysis intrestingess measures closed itemsets
1
Introduction and Motivation
Formal concepts play an important role in knowledge discovery, since they can be
used for concise representation of association rules, clustering and constructing
domain taxonomies, see surveys [40, 41, 34]. Most of the difficulties in the application of closed itemsets (or, a more common name, attribute sets) to practical
datasets are caused by the exponential number of formal concepts. It complicates both the process of the model construction and the analysis of the results
as well. Recently, several approaches to tackle these issues were introduced. For a
dataset all closed sets of attributes ordered by inclusion relation make a lattice,
below we use the terminology of concept lattices and Formal Concept Analysis [21]. In terms of FCA, a concept is a pair consisting of an extent (a closed set
of objects, or transactions in data mining terms) and an intent (the respective
closed set of attributes, or itemset).
We propose to divide existing approaches to simplifying concept mining into
the groups presented in Figure 1. Methods of the first group compute the concept
lattice built on simplified data. The most general way to get smaller lattices is
to reduce the size of datasets preserving the most important information. For
this purpose such methods as Singular Value Decomposition [15], Non-negative
Matrix Decomposition [44] and more computationally efficient k-Means [16],
fuzzy k-Means [18], agglomerative clustering of objects based on a similarity
function of weighted attributes [17] are used. Another way to reduce the lattice
size was proposed in [19]. This approach aims at the reduction in the number of
incomparable concepts by making slight changes of the context.
Approaches to concept mining simplification
Preprocessing:
modification
of the original data
Dimension
reduction
Computing:
approximate
concepts (clusters)
Changing of
data entries
Computing
a subset of concepts
Application of
an antimonotonic
function
Postprocessing:
application of indices
to concepts
Modification
of a closure
operator
Fig. 1: Classification of methods for concept mining simplification
Computing approximate concepts, or so-called bi-/triclusters, is becoming
increasingly more ubiquitous nowadays. In general, biclustering refers to performing simultaneous row–column clustering of real-valued data. Biclusters itself
can be defined in different ways [36], e.g. as a submatrix with constant values,
with constant values on rows or columns, with coherent values, etc. In the case
of binary data bicluster is defined as a submatrix with high density (i.e., the proportion of 1s). In [24] authors presented a set of evaluation criteria for triclusters,
which are density, coverage, diversity, noise tolerance and cardinality. As it was
shown in experiments, the optimality by some criteria imposes non-optimality
w.r.t. other criteria and the optimization by a particular criteria underlies a
clustering algorithm itself. Put differently, in the case of multimodal clustering
the choice of criteria for cluster evaluation defines the choice of a clustering
algorithm rather than selection of a subset of computed clusters.
Methods from “Computing a subset of concepts” class aim at simplifying
the analysis of a concept lattice by computing a subset of concepts. It can be
done, e.g., by defining an antimonotonic function w.r.t. the size of concept extent.
Computing concepts with extents exceeding a threshold was proposed in [35] and
studied in relation to frequent itemsets mining in [45], where the authors propose
an algorithm for building iceberg lattices. In this case some interesting rare
concepts can be discarded, thus, it yields fewer results. From this perspective,
the most promising approach is applying ∆-stability. Σoϕια algorithm (Sofia) for
computing a subset of most ∆-stable concepts was proposed in [10]. ∆-stability
will be considered in Section 3.
The most well-reasoned way to build a subset of concepts is to modify the
closure operator and to restrict the creation of new concepts by involving background knowledge. In [8] attribute-dependency formulas (AD-formulas) were introduced. The authors define mandatory attributes for particular attributes of a
context. If an attribute is included in a closed attribute set without its mandatory attributes the concepts will not be generated. In [5] an analogous approach
based on weights of attributes was proposed. The interestingness of concepts is
estimated using an aggregation function (average, maximum or minimum), which
is applied to a generator, a minimal generator, or to an intent. The authors also
note the correspondence between the numerical approach and AD-formulas.
Some restricted closure operators naturally arise from analyzed data and
subject area. Carpineto and Romano [13] considered the document–term relation, where objects are documents and terms used in documents are attributes.
The authors proposed to modify the closure operator using a hierarchy of terms
as follows: two different attributes are considered as equivalent if they have a
common ancestor that is a more general term.
In the last decade some polynomial-time algorithms for computing Galois
sub-hierarchies were proposed, see [9, 2].
The main idea of methods from the latter group is to assess interestingness
of concepts by means of interestingness indices. This approach does not have
disadvantages of the previously described methods, namely, getting concepts approximately corresponding to original objects or attributes, missing interesting
concepts due to early termination of an algorithm, a costly preprocessing procedure requiring the involvement of domain experts or another reliable source of
background knowledge. This class, however, has its own drawbacks, for example,
the exponential complexity of the lattice computation can be aggravated by high
complexity of index computation. The most promising approach from this point
of view is to compute a subset of concepts using the idea of antimonotonicity
of an index. When the index value is antimonotonic w.r.t. the order relation of
the concept lattice, one can start computing from the top concept and proceed
top-down until the concepts with the lowest possible value given by a threshold.
In this paper we focus on a thorough study of concept indices. We leave statistical significance tests of concepts beyond this study, some basic information
on statistical approaches to itemsets mining can be found in [48]. The paper
is naturally divided into three parts. First, we investigate the main features of
indices and discuss the intuition behind them, as well as their applicability in
practice with regard to their computation cost. Second, we provide a framework
for the development of new indices and give detailed information on the basic
metrics and operations that can be used to create new indices. Third, we describe
the results of the comparative study of indices regarding the following aspects:
estimation of interestingness concepts, approximation of intractable indices and
non-sensitivity to noise.
The rest of the paper is organized as follows. Section 2 briefly introduces
the main definitions of Formal Concept Analysis. Section 3 is devoted to the description of indices. The main index features are given in Section 3.1. Section 3.2
provides the basic information on indices and respective formulas. In Section 3.3
we use the known indices and their features to build a concept lattice, then we
reveal the most interesting concepts (i.e. groups of indices) w.r.t. certain indices.
In Section 4 we propose guidelines for the development of new indices based on
indices for arbitrary sets of attributes. We discuss the approaches to measuring
interestingness and provide the basic metrics and operations that can be used for
index construction. Section 5 focuses on a comparative study of the indices w.r.t.
the most important tasks: selection of interesting concepts (Section 5.1), approximation of intractable indices (Section 5.2) and noise reduction (Section 5.3). In
Section 6 we conclude and discuss the future work.
2
Formal Concept Analysis: Basic Definitions
Here we briefly recall the main definitions of FCA [21]. Given a finite set of
objects G and a finite set of attributes M , we consider an incidence relation
I ⊆ G × M so that (g, m) ∈ I if object g ∈ G has attribute m ∈ M . A formal
0
context is a triple (G, M, I). The derivation operators (·) are defined for A ⊆ G
and B ⊆ M as follows:
A0 = {m ∈ M | gIm for all g ∈ A}
B 0 = {g ∈ G | gIm for all m ∈ B}
A0 is the set of attributes common to all objects of A and B 0 is the set of
0
objects sharing all attributes from B. The double application of (·) is a closure
00
operator, i.e., (·) is extensive, idempotent and monotone. Sets A ⊆ G, B ⊆ M ,
such that A = A00 and B = B 00 are said to be closed.
A (formal) concept is a pair (A, B), where A ⊆ G, B ⊆ M and A0 = B,
0
B = A. A is called the (formal) extent and B is called the (formal) intent of
the concept (A, B).
A concept lattice (or Galois lattice) is a partially ordered set of concepts,
the order 6 is defined as follows: (A, B) ≤ (C, D) iff A ⊆ C (D ⊆ B), a pair
(A, B) is a subconcept of (C, D) and (C, D) is a superconcept of (A, B). Each
finite lattice has the highest element with A = G, called the top element, and
the lowest element with B = M , called the bottom element.
a1
stability
×
∆l
×
∆h (collapse
×
index)
stab2N OE
×
stab2OE
×
stab2OIE
×
robustness ×
a2 a3 a4 a5 a6 a7 a8 a9 a10 a11
×
×
× × × × ×
× ×
×
× × × × ×
×
×
×
×
×
×
×
×
× ×
stability, ∆l , ∆h , stab2N OE ,
stab2OE ,stab2OIE , robustness
{a1 , a2 }
×
× × ×
× ×
× × ×
× ×
× × ×
× ×
×
×
× ×
× ×
a1 : (designed) for closed subsets, a2 : not applicable to arbitrary subsets, a3 : based on comparison
to other attribute subsets, a4 : based on comparison
to neighboring attribute subsets, a5 : size-based, a6 :
monotonic–antimonotonic w.r.t. order on attribute
sets, a7 : using a tuning parameter, a8 : polynomial
complexity, a9 : linear complexity, a10 : cubic and
higher complexity, a11 : computable in one-pass over
data
stability,
robustness,
{a1 , a2 , a10 }
∆l , ∆h , stab2N OE ,
stab2OE ,stab2OIE , robustness
a1 , a2 , a3 , a4 , a6
∆l , ∆h , stab2OIE ,
robustness
a1 , a2 , a3 , a4 , a6 , a11
robustness
a1 , a2 , a3 , a4 , a6 , a7 , a10 , a11
∆l , ∆h , stab2N OE ,
stab2OE ,stab2OIE
a1 , a2 , a3 , a4 , a5 , a6 , a8 , a9
∆l , ∆h ,stab2OIE
a1 , a2 , a3 , a4 , a5 , a6 , a8 , a9 , a11
–
a1 , a2 , a3 , a4 , a5 , a6 , a8 , a9 , a10 a11
Fig. 2: Formal context of subsampling-based indices (on the left) and the corresponding concept lattice (on the right)
Example 1 Let us consider a formal context given in Fig. 2. Group of
subsampling-based indices for formal concepts is a set of objects, their essential
features are attributes. The corresponding concept lattice (see Fig. 2, on the
right) consists of 8 concepts.
3
Closed Itemset Indices
The application of indices offers advantages over other approaches to reducing of
the number of analyzed concepts. For example, it provides a thorough study of a
concept lattice without recomputing the lattice and ensures easily interpretable
results by preserving objects and attributes corresponding exactly to the original
data. However, the exponential time complexity of lattice computation makes
attracts attention to antimonotonic indices, since they allow generating only a
subset of formal concepts.
In this section we propose a hierarchical clustering of the existing indices for
closed set of attributes using the FCA-based approach. We describe the most
important index features and use them to build a concept lattice of indices: we
consider a formal context of indices and their features, and then compute the
ordered set of formal concepts. Upon that we apply some of the described indices
to identify the most interesting (top-ranked) concepts.
3.1
Description of Index Features
Some of the indices originally designed to assess concepts (“designed for closed
subsets”) can not be applied to an arbitrary set of attributes (“not applicable
to arbitrary subsets”), whereas others were originally designed as measures for
arbitrary attribute sets. In our study we consider only two indices for arbitrary
sets of attributes, i.e. support/frequency, due to their importance in concept
mining.
There exist indices which deal with each attribute of an intent (or a whole
context) separately and assess the contribution of these attributes to a concept
(“using all of the single attributes of a context”). Another essential feature of
the concept indices is the use of other concepts (“based on comparison to other
attribute subsets”). The concepts are selected based on a particular condition
or index values, i.e., frequent, (dis)similar to some other concepts, etc. In this
case an index is comprised of relative values or differences between absolute
values of the selected concepts and the assessed concept. We additionally add
only one predicate “is neighbor” (“based on comparison to neighboring attribute
subsets”), since the direct neighbors of a concept are highly important both for
computing closed itemsets and for concept mining.
As it was noted previously, support has a special place in concept mining.
We define “support-based (for subsets of attributes)” and “support-based (for
single attributes)” features to characterize whether an index is based on support.
Support may be regarded as probability of an attribute or a set of attributes,
along with it we also consider conditional probability (“using conditional (joint)
probability”). “Size-based” indices use the size of an intent or an extent, the total
number of attributes or objects, thus, they do not require to assess particular
elements of a context.
As it was mentioned before, the antimonotonicity is one of the most important index properties, since it allows one to solve the problem of exponential
complexity of algorithms for generating concepts. The time complexity is referred
through the following attributes: “polynomial complexity”, “sublinear complexity”, “linear complexity”, “quadratic complexity”, “cubic and higher complexity”.
These groups are defined w.r.t. the size of a context, i.e., |G||M | under the assumption that the lattice has been computed. The possibility to compute indices
in “one-pass over data” is relevant to complexity problem, but mostly it refers
to the memory complexity and analysis of streaming data.
Some indices have a tuning parameter (“using a tuning parameter”) that either changes the concept rankings or is used as a binarization threshold (“boolean
value”).
3.2
Description of Indices
In our study we use the indices described below.
Stability Stability indices for formal concepts were introduced in [30, 31] and
modified in [33]. For a formal concept (A, B) the (integral) intentional stability
index is the probability that B will remain closed when removing a subset of
objects from extent A with equal probability.
The intentional stability is defined as follows:
| {C ⊆ A|C 0 = B} |
2|A|
Intentional stability measures overfitting, i.e., the dependence of an intent
(i.e.,a closed pattern) on observations (objects of the respective extent). Extensional stability is defined dually.
In [11] it was noted that stability indices are antimonotonic w.r.t. chain of
projections. This class of antimonotonicity contains previously known classes [46]
and allows one to generate k most stable concepts with delay polynomial in k
and the input size.
Stabi (A, B) =
Estimates of stability
Logarithmic Stability The problem of computing stability is #P -complete [31],
which hampers its application in practice where one usually faces large-sized
contexts. In [4] it was proposed to use Monte Carlo approximation of stability,
in [12] combination of Monte Carlo and an upper-bound estimate is described.
The authors use the logarithmic scale of stability (the concept rankings remain
the same):
LStab (c) = −log2 (1 − Stab (c))
It allows one to deal with the problem described in [25]: closeness of stability
values to 1 for large contexts.
The bounds of stability are given by
X
∆min (c) − log2 (|M |) ≤ −log2
2−∆(c,d) ≤ LStab (c) ≤ ∆min (c) ,
d∈DD(c)
where ∆min (c) = mind∈DD(c) ∆ (c, d), DD (c) is the set of all direct descendants
of c in a lattice and ∆ (c, d) is the size of the set-difference between extents of
formal concepts c and d.
In [32] more precise upper bounds of the logarithmic stability were described.
These estimates are based on two direct neighbors. The formulas are given below.
Max-disjoint-extents upper bound (stab2N OE ) The estimate uses two lower neighbors, one of them has the maximal extent among all the lower neighbors, the
second one has the maximal extent among the rest of neighbors and does not
share elements with the first one:
(∅)
LStab (c) ≤ ∆min (c) + ∆min (c, ∆min (c))
(∅)
where ∆min (c) = mind∈DD(c) ∆ (c, d), ∆min (c, ∆min (c)) = d1 + d2 , where
d1 = mind∈DD(c) ∆ (c, d) ,
d2 = mind∈DD {∆ (c, d) |d1 ∩ d = ∅} .
Max-distinguished-extents upper bound (stab2OE ) To compute this index the first
neighbor is selected in the same way, while the second one meets the following
condition: its extent has the maximal number of objects not included in the
extent of the first one.
LStab (c) ≤ |d1 | + |d2 | − |d1 ∩ d2 |
where d2 = mind∈DD(c) ∆ (c, d) |d2 − d1 = maxd∗ ∈DD(c) d∗ − d1 and
d1 = mind∈DD(c) ∆ (c, d).
Max-extent upper bound (stab2OIE ) The estimate takes into account two different maximal extends:
LStab (c) ≤ |d1 | + |d2 | − |d1 ∩ d2 |
where d1 = mind∈DD(c) (c, d), d2 = mind∈DD (c) and d1 6= d2 . But such a
“greedy” strategy can give an underestimated upper bound.
Stability Indices The notion of stability was revised and considered with regard to estimating concept-based hypotheses in [31]. Level-wise stability indices are studied. For formal concept c = (A, B) stability index of the jth
γ (c)
level (2 ≤ j ≤ n − 1) is defined as follows: Jj (c) = jn , where n = |A|,
j
0
γj (c) = | {Y ⊂ A | |Y | = j, Y = B} |.
Pn−1
Integral stability index is defined as JΣ (c) = i=2 Ji .
In this study we also consider integral stability indices of the jth level (2 ≤
j ≤ n − 1):
Pj
– Minor-set-based integral stability: JΣj (c) = i=2 Ji
Pn−1
– Major-set-based integral stability:JΣj (c) = i=j Ji .
Integral stability of the jth level may be regarded as the approximation of
the stability index. The survey of the problem of best approximation is given in
Section 5.
Robustness Robustness [46] of an (arbitrary) set of attributes of a dataset has
been introduced to estimate the probability that a pattern would still be generated if some transactions (rows) of the dataset are removed. To compute this
index one needs to generate 2|G| subsamples, where G is the set of transactions
(objects). Some classes of itemsets allow computing this index using an exact
formula instead of subsampling the data. Here we use the formula for closed
sets of attributes. For formal concept (A, B) this is the probability that B will
remain closed when removing an object (row) from extent A with probability
1 − α. Hence, stability index can be considered as an instantiation of robustness
for α = 0.5. So, for formal concept c = (A, B) the robustness is given as follows:
X
|B |−|Bc |
|A |−|Ad |
r (c, α) =
(−1) d
(1 − α) c
d≤c
Proposition 1. Stability of concept (A, B) is equal to its robustness for α = 0.5.
Proof
In [43] it was noted that stability can be computed recursively by the traversal
of covering relation (i.e., graph of the diagram) of the concept lattice from the
bottom concept upwards as follows:
Stab(A, B) =
| {C ⊆ A | C 0 = B} |
σ(A, B)
=
,
2|A|
2|A|
P
where σ(A, B) = 2|A| − (C,D)<(A,B) σ(C, D).
In [4] the following relation of stability to Möbius function µ [22] was shown:
X
| {C ⊆ A | C 0 = B} | =
2|C| µ((C, D), (A, B))
(C,D)≤(A,B)
Using the Möbius function
lattice, the formula takes the folP of the concept
|C|
lowing form: σ(A, B) =
µ ((C, D), (A, B)) and stability can
(C,D)≤(A,B) 2
alternatively be represented as
X
Stab(A, B) =
2|C|−|A| µ ((C, D), (A, B)) .
(C,D)≤(A,B)
Robustness of concept (A, B) is computed as
X
ak = {
P|A|
k=0
ak (1 − α)k , where
e(D, B) | |A| − |C| = k},
(C,D)≤(A,B)
and
e(D, B) =
−
1
P
(C,D)<(E,F )≤(A,B)
if D = B
e(F, B) otherwise.
This formula gives Möbius function, thus robustness can be rewritten as
r((A, B), α) =
X
(C,D)≤(A,B)
µ ((C, D), (A, B)) (1 − α)|A|−|C| .
Replacing 1 − α by 0.5 we obtain the formula for stability, which completes
the proof.
Concept Probability In [29] it was noticed that some interesting concepts with
the small number of objects (i.e., small extents) usually have low stability values.
The concept probability was proposed to get rid of this bias. The definition of
concept probability from [29] is equivalent to the concept probability introduced
earlier by R. Emilion [20].
The probability that an arbitrary object has all attributes from set B is
defined as follows:
Y
pB =
pm
m∈B
Concept probability is defined as the probability of B being closed:
00
p (B = B ) =
n
X
k=0
0
00
p (|B | = k, B = B ) =
n
X
k=0
"
Ckn pkB
n−k
(1 − pB )
Y
m∈B
/
1−
pkm
#
where n = |G|.
Concept probability aggregates three probabilistic components: the occurrence of each attribute from B in all k objects, the absence of at least one
attribute from B in other objects and the absence of other attributes shared by
all k objects.
Separation The index was introduced in [29] to estimate the specificity of the
object–attribute relation of a concept with respect to the formal context. It is
defined as a part of the area covered by a formal concept among all nonzero
elements in the rows and columns corresponding to the formal concept.
s (A, B) = P
0
g∈A |g | +
|A||B|
P
m∈B
|m0 | − |A||B|
Frequency (support) It is one of the most popular measures in the theory of
pattern mining. Frequency arises from the assumption that the most “interesting” concepts are frequent ones:
supp (A, B) =
|A|
|G|
The support provides an efficient level-wise algorithm of semilattice computing
since it exhibits antimonotonicity (a priori property [1, 37]):
B1 ⊂ B2 → supp (B1 ) ≥ supp (B2 ) .
In this study we say that a set of attributes is frequent if its support exceeds a
certain threshold. So, the frequency of an attribute set means that it is frequent.
Monocle The Monocle [47] is a method that defines concept weights based on
a subset of concepts. The weight function has the following form:
!
X
X
w(c, H) = |A| +
NG (g, H) · |B| +
NM (m, H) ,
g∈A
m∈B
where NG (g, H) = |{(A, B) | (A, B) ∈ H, g ∈
/ A}| is the number of concepts in H
not containing an object g, NM (m, H) is defined similarly for attributes, H ⊂ L.
The weight function is monotone w.r.t. the size of H.
δ-Tolerance Closed Frequent Itemsets (δ-TCFIs) δ-TCFIs [14] uses the
subset of lower neighbors (direct descendants) in the concept lattice to assess
concept interestingness. The main idea is to select relatively frequent concepts
with respect to their direct descendants. Concept c = (A, B) is δ-TCFI iff it
meets the following condition: ∀d = (C, D), such that |D| = |B| + 1, supp(D) ≥
(1 − δ) · supp(B), where δ ∈ [0, 1] is a tolerance factor.
Margin-closed itemset The index was proposed by Moerchen et al. [38]. A
margin closed itemset has no supersets with almost the same support, by definition, it satisfies the following expression: X ∈ F I & ∀X 0 ∈ F I : X ⊂ X 0 ⇒
supp(X 0 )
supp(X) ≤ 1 − α, where F I is a set of frequent itemsets.
As can be seen from the formulas, margin-closed itemset and δ-TCFIs indices
are very close to each others. They are based on the same indices computed on
different subsets of concepts.
In our study we consider the relaxed formulation of the margin-closed itemset
index, namely, we consider the ratio of the maximal extent size among all direct
predecessors of a concept to the extent size of the concept itself.
Belohlavek and Trnecka in [6, 7] investigated the group of so-called “basic
level” measures. It is a psychology-motivated approach that was designed to
formalize the existing psychological approach to defining the basic level of concepts [39]. The group is comprised of similarity- and predictability-based indices,
cue validity, category feature collocation and category utility indices.
Similarity approach (S) The similarity approach to the assessment of belonging to a basic level was proposed in [42] and subsequently formalized and
applied to FCA in [6]. This index is the combination of three fuzzy functions that
correspond to formalized properties outlined by E. Rosch [42]: high cohesion of
concepts, considerably greater cohesion with respect to upper neighbors and a
slightly less cohesion with respect to lower neighbors. The membership degree
of the basic level is defined as follows:
∗∗
BLS = coh∗∗ (A, B) ⊗ coh∗∗
un (A, B) ⊗ cohln (A, B) ,
where coh is a fuzzy function that corresponds to the conditions defined above,
⊗ is t-norm [28].
A cohesion function is based on the pairwise similarity of objects from an
extent. To assess similarity between two objects authors use simple matching
coefficient or Jaccard similarity:
simSM C (B1 , B2 ) =
|B1 ∩ B2 | + |M − (B1 ∪ B2 ) |
;
|M |
simJ (B1 , B2 ) =
|B1 ∩ B2 |
.
|B1 ∪ B2 |
A cohesion function is one of the following aggregation functions:
P
0
0
{x1 ,x2 }⊆A,x1 6=x2 sim... (x1 , x2 )
a
coh... (A, B) =
|A| (|A| − 1) /2
0
0
cohm
... (A, B) = min sim... (x1 , x2 )
x1 ,x2 ∈A
The Rosch’s properties for upper and lower neighbors take the following forms:
P
∗
∗
c∈U N (A,B) coh... (c) /coh... (A, B)
a∗
coh...,un (A, B) = 1 −
|U N (A, B) |
P
∗
∗
c∈LN (A,B) coh... (A, B) /coh... (c)
coha∗
...,ln (A, B) =
|LN (A, B) |
cohm∗
...,un (A, B) = 1 −
cohm∗
...,ln (A, B) =
max
c∈U N (A,B)
min
c∈LN (A,B)
coh∗... (c) /coh∗... (A, B)
coh∗... (A, B) /coh∗... (c)
where U N (A, B) and LN (A, B) are upper and lower neighbors of formal concept
(A, B) respectively.
As the authors noted, experiments revealed that the type of a cohesion function does not affect the result, while the choice of a similarity measure can greatly
affect the outcome. More than that, in some cases upper (lower) neighbors may
have higher (lower) cohesion than the formal concept itself (for example, the
boundary cases, when a neighbors’s extent / intent is comprised of identical
rows / columns). To tackle this issue of non-monotonic neighbors w.r.t. a sim∗∗
ilarity function the authors proposed to take coh∗∗
...,ln and coh...,un as 0, if the
rate of non-monotonic neighbors is larger than a threshold.
∗∗
∗∗
Below, we use the following notation: SSM
C and SJ , where stars are replaced
by the type of cohesion functions for neighbors and objects, respectively. SM C
and J stand for simple matching coefficient or Jaccard similarity, respectively.
Predictability approach (P) Predictability [7] of a formal concept is computed in a way quite similar to the previous one. In this approach a cohesion
function is replaced by the predictability function:
∗∗
P (A, B) = pred∗∗ (A, B) ⊗ pred∗∗
un (A, B) ⊗ predln (A, B) .
From this point of view, concepts are close to basic level if there are only few
attributes outside B contained in objects from A:
E (I [hx, yi ∈ I] |I [x ∈ A]) = −
pred (A, B) = 1 −
X
y∈M −B
|A ∩ y 0 |
|A ∩ y 0 |
log
|A|
|A|
E (I [hx, yi ∈ I] |I [x ∈ A])
.
|M − B|
Cue Validity (CV), Category Feature Collocation (CFC), Category
Utility (CU) The following measures are based on the conditional probability
of object g ∈ A given y ⊆ g 0 [7]:
CV (A, B) =
X
P (A|y 0 ) =
y∈B
y∈B
X |A ∩ y 0 | |A ∩ y 0 |
|y 0 |
|A|
y∈M
y∈M
"
2 0 2 #
i |A| X
Xh
|A ∩ y 0 |
|y |
2
0
0 2
CU (A, B) = p (A)
p (y |A) − p (y ) =
−
|G|
|y 0 |
|G|
CF C (A, B) =
y∈M
X
X |A|
|y 0 |
p (A|y 0 ) p (y 0 |A) =
y∈M
CV deals with the probability of an extent given attributes from an intent,
CFC takes into account the relation between all attributes from a context and
an intent of a formal concept, and CU characterizes how much an attribute from
an intent is specific for a given concept rather than for a formal context [49].
3.3
Concept Lattice of Indices for Closed Itemsets
In this part of the paper we propose a classification of the existing indices for
formal concepts. The classification can be done w.r.t. several features, thus a
lot of different index classifications may be suggested. FCA provides a universal
framework for discovering domain structure by constructing inclusion-ordered
overlapping clusters with various degrees of generality.
We represent the indices described in Section 3.2 and their features from
Section 3.1 as a formal context of size 20 × 19. The objects and attributes of the
formal context are indices and their features, respectively. The context is given
in Table 1, the corresponding lattice consists of 73 formal concepts.
We applied the described indices to the lattice to discover interesting concepts
and consider top-8 concepts (about 10% of all concepts) by values of the following indices: probability, separation, monocle, margin-closed, frequency, stability,
stability estimates, CV, CFC, CU, predictability, similarity approach (SJaa and
aa
SSM
C ), robustness with α = 0.3. Below, the most frequent concept in the topranked groups are listed. We obtained top-ranked concepts with a singleton as
an extent or an intent for margin-closed and probability, respectively. The most
interesting groups of indices (concepts) with their frequencies (i.e., the rate of
top-8 groups where a concept has been included) are listed below.
Frequency = 0.4:
Extent: ∆l , ∆h , stab2OIE , δ-TCFIs.
Intent: for closed subsets, not applicable to arbitrary subsets, based on comparison to other attribute subsets, based on comparison to neighboring attribute
subsets, size-based, (anti)monotonic, polynomial complexity, linear complexity,
computable in one-pass over data.
Extent: ∆l , ∆h , stab2N OE , stab2OE , stab2OIE , margin-closed*, δ-TCFIs
Intent: for closed subsets, not applicable to arbitrary subsets, based on comparison to other attribute subsets, based on comparison to neighboring attribute
subsets, size-based, (anti)monotonic, polynomial complexity, linear complexity.
The most interesting concepts by stability and its estimates.
Frequency = 0.33:
Extent: ∆l , ∆h , stab2N OE , stab2OE , stab2OIE , similarity, predictability,
margin-closed*, δ-TCFIs.
Intent: for closed subsetes, not applicable to arbitrary subsets, based on comparison to other attribute subsets, based on comparison to neighboring attribute
subsets, polynomial complexity.
The most interesting concepts by CV and separation.
Table 1: A formal context of indices for formal concepts and their key features
Index name
(the number
a1
of section)
stability (3.2)
×
∆l (3.2)
×
∆h (collapse
×
index)(3.2)
stab2N OE (3.2)
×
stab2OE (3.2)
×
stab2OIE (3.2)
×
robustness (3.2)
×
probability (3.2) ×
separation (3.2)
×
support (3.2)
frequency (3.2)
monocle (3.2)
×
δ-TCFIs (3.2)
×
margin×
closed (3.2)
margin×
closed* (3.2)
similarity (3.2)
×
predictability (3.2) ×
CV (3.2)
×
CFC (3.2)
×
CU (3.2)
×
a2 a3 a4 a5 a6 a7 a8 a9 a10 a11 a12 a13 a14 a15 a16 a17 a18 a19
×
×
× ×
×
×
×
×
×
×
× ×
×
×
×
×
×
×
×
×
×
× ×
× ×
× ×
×
×
×
×
×
×
×
× ×
×
×
×
×
×
×
×
×
×
×
×
×
×
×
×
×
×
×
×
×
×
× ×
× × × × ×
× × ×
×
× × × ×
×
×
×
×
× ×
×
× × ×
×
×
×
× ×
× × ×
×
×
×
×
×
×
×
×
× ×
×
×
×
×
×
× ×
×
×
×
× ×
× ×
×
×
×
×
×
×
×
×
×
×
×
×
×
×
×
a1 : (designed) for closed subsets, a2 : not applicable to arbitrary attribute subsets,
a3 : applicable to arbitrary attribute subsets, a4 : using all of the single attributes
of a context, a5 : based on comparison to other attribute subsets, a6 : based on
comparison to neighboring attribute subsets, a7 : support-based (for subsets of
attributes), a8 : support-based (for single attributes), a9 : size-based, a10 : using
conditional (joint) probability, a11 : monotonic–antimonotonic w.r.t. order on attribute sets, a12 : using a tuning parameter, a13 : boolean value, a14 : polynomial
complexity, a15 : sublinear complexity, a16 : linear complexity, a17 : quadratic complexity, a18 : cubic and higher complexity, a19 : computable in one-pass over data
Extent: ∆l , ∆h , stab2N OE , stab2OE , stab2OIE , separation, margin-closed,
margin-closed*, δ-TCFIs.
Intent: for closed subsets, size-based.
The most interesting concepts by frequency.
Frequency = 0.25:
Extent: ∆l , ∆h , stab2N OE , stab2OE , stab2OIE , robustness, margin-closed*,
δ-TCFIs.
Intent: for closed subsets, not applicable to arbitrary subsets, based on comparison to other attribute subsets, based on comparison to neighboring attribute
subsets, (anti)monotonic.
Extent: stability, ∆l , ∆h , stab2N OE , stab2OE , stab2OIE , probability, robustness, separation, similarity, predictability, CV, CFC, CU, margin-closed,
margin-closed*, monocle, δ-TCFIs.
Intent: for closed.
Extent: ∆l , ∆h , stab2N OE , stab2OE , stab2OIE , robustness, similarity, predictability, margin-closed*, δ-TCFIs.
Intent: for closed subsets, not applicable to arbitrary subsets, based on comparison to other attribute subsets, based on comparison to neighboring attribute
subsets.
The most interesting concepts by monocle.
Extent: ∆l , ∆h , stab2N OE , stab2OE , stab2OIE , robustness, CV, CFC, marginclosed*, δ-TCFIs.
Intent: for closed subsets, (anti)monotonic.
The most interesting concepts by CFC.
Extent: ∆l , ∆h , stab2N OE , stab2OE , stab2OIE , probability, margin-closed*,
δ-TCFIs.
Intent: for closed subsets, not applicable to arbitrary subsets, polynomial
complexity, linear complexity.
The most interesting concepts by CU.
Extent: support, frequency, δ-TCFIs.
Intent: support-based (for attribute sets), (anti)monotonic, polynomial complexity, computable in one-pass over data.
The most interesting concepts by robustness, α = 0.3.
Extent: CV, CFC.
Intent: for closed subsets, using conditional probability, (anti)monotonic,
polynomial complexity, linear complexity, computable in one-pass over data.
The most interesting concepts by predictability, similarity (Basic Level Metrics).
As can be seen from the results of the concept rankings, most of the selected
indices (accordingly, the rankings) include either ∆-indices or support-based
measures. Both types of concepts have almost the same intent that causes the
large number of quite similar concepts.
4
Generic Approach to the Assessment of Closed
Itemsets
In this section we propose guidelines for applying arbitrary itemset indices to
closed ones. It should be noted that the problem of the arbitrary itemset assessment was thoroughly investigated in the literature and a lot of different indices
for arbitrary sets of attributes and associative rules were proposed. We will not
consider in this study statistical-based approaches to patterns and rules assessment (the basic information on the approach can be found in [48]). Some indices
for associative rules can be applied directly to evaluate sets of attributes, for
example, lif t given
Q in Table 2 takes the following form for itemset assessment:
lif t(A) = P (A)/ a∈A P (a). Moreover, a rule-assessment measure may be applied to attribute set within the rule-based approach. In this case, all possible
rules are generated from the examined attribute set and then an aggregation
function is used to obtain a new index (see details in [48]).
All these indices can be applied directly to formal concepts to assess them
either as subsets of objects or as subsets of attributes. However, the closeness of
attribute sets can be exploit to construct new indices. The purpose of this section
is to provide the main ideas of how to adopt arbitrary itemset indices (including
rule-based measures) to assess both dimensions of concepts simultaneously.
To develop a new index one needs to answer the following questions:
– Which concepts are interesting (which patterns we are looking for)?
– Which index will be chosen as the basic measure?
– Which operations will be used to aggregate the values of the basic measure?
The starting point is the personal perspective on concept interestingness. We
define the following ones:
– Some “internal”/“external” properties of the concept itself. In this case homogeneous components are considered.
– Impact of the element gathering. Here, a measure is applied to a set and its
elements: f (set) and {f (element) | element ∈ set}.
– Impact of a condition: f (parameter | condition) and f (parameter).
– Concept pureness. This approach is based on the comparison of features
computed on a concept and out of the concept: f (parameter) and f (U \
parameter)), where U is the domain of the parameter.
– Stability to random changes in data. Resampling strategy is applied to a set:
{resamplingα (set)}, where α is a noise rate.
The most popular basic measures are given in Table 2 to make the paper
self-contained. More detailed information about the listed indices can be found
in [23, 3, 27]. In the case of formal concepts, indices for associative rules, can be
used not only in the ways mentioned above, but also to compare a concept to
others similar concepts, i.e., to discover some relatively interesting concepts in
a neighborhood.
To combine values of basic measures an aggregation function is used. If a new
index measures a relative property of a concepts, then aggregation and comparison operators are applied in any order, i.e., compare(aggregateel∈S (el), d) or
aggregateel∈S (compare(el, d)), where S is a set of peer elements, d is a distinguished value to compare.
It is worth noting that both real-valued and fuzzy functions can be used
to aggregate values. If the values are in the interval [0, 1] (i.e., they can be
considered as probabilistic components), a fuzzy aggregation function should be
applied. Some binary fuzzy operators (t-norm or its dualization s-norm related
the equation t(v1 , v2 ) = 1−s(1−v1 , 1−v2 )) and real-valued aggregation functions
are given in Table 3. In the case of poor data quality a robust aggregation
function, as median, can be applied to data. A comparison function can take
one of the following forms: ∗ − ∗, ∗/∗, log(∗/∗).
Example 2 Let us consider how the proposed guidelines can be used for
designing a new index. We will give examples to show that the existing indices
conform to the presented approach and to demonstrate how new indices can be
constructed step-by-step using this scheme.
Cue Validity (or CV):
CV (A, B) =
X
P (A|y 0 ) =
y∈B
X |A|
.
|y 0 |
y∈B
– Notion of interestingness: a concept (A, B) is interesting, if the objects from
its extent A are more specific for the set of attributes B. We study only a
subset of attributes, i.e., properties of the concept itself.
– Basic measure: conditional probability.
– Aggregation operator (applied to homogeneous elements): sum.
A new index:
Index1 (A, B) =
min
(C,D)∈U A((A,B))
(P S(D) − P S(B)) ,
where
U N ((A, B) is the set of direct ancestors of (A, B), and P S(B) = P (B) −
Q
m∈B P (m).
– Notion of interestingness: a concept (A, B) is interesting if it differs from
more general (in the descriptive sense) concepts. We study only direct ancestors in a lattice.
– Basic measure: Piatetsky-Shapiro.
– Aggregation operator (applied to non-homogeneous elements): comparison
of a concept to each of its ancestors, aggregation by taking the minimal
value.
Table 2: Indices for arbitrary itemsets (associative rules A → B)
Index
accuracy
added value/change of support
certainty factor
collective strength
conditional probability
conviction
cosine
Gini index
information gain
Formula
P (AB) + P (¬A¬B)
P (B|A) − P (B)
(P (B|A) − P (B)) / (1 − P (B))
P (AB)+P (¬B|¬A)
(A)P (B)P (¬A)P (¬B)
· 1−P
P (A)P (B)+P (¬A)P (¬B)
1−P (AB)−P (¬B|¬A)
P (B|A)
P (A)P (¬B)/P
(A¬B)
p
P (AB)/ P (A)P (B)
P (A) P (B|A)2 + P (¬B|A)2 +
2
+P (¬A) P (B|¬A) + P (¬B|¬A)2 − P (B)2 − P (¬B)2
P (AB)
log P (A)P
(B)
P (AB) log PP(B|A)
+ P (A¬B) log PP(¬B|A)
(B)
(¬B)
P
(AB)/
(P
(A)
+
P
(B)
−
P
(AB))
p
pP (AB)(P (B|A)P (B)),
Klosgen
P (AB) max(P (B|A)P (B), P (A|B)P (A))
Laplace correction
(N (AB) + 1) /(N (A) + 2)
least contradiction
(P (AB) − P (A¬B)) /P (B)
leverage
P (B|A) − P (A)P (B)
lift
P (AB)/ (P (A)P (B))
Loevinger
1 − P (A)P (¬B)/P (A¬B)
P
P P
P (Ai Bj )
/ − i P (Ai ) log2 P (Ai )
normalized mutual information i j P (Ai Bj ) · log2 P (Ai )P
(Bj )
odd multiplier
(P (AB)P (¬B)) / (P (B)P (A¬B))
Example and
1 − P (A¬B)/P (AB)
counter example Rate
odds ratio
P (AB)P (¬A¬B)/ (P (A¬B)P (¬BA))
P (AB)
one-way support
P (B|A) log2 P (A)P
(B)
(P (AB)−P (A)P (B))2
(P (¬AB)−P (¬A)P (B))2
|G|
+
+
P (¬A)P (B)
P (A)P (B)
Pearson’s χ2
(P (A¬B)−P (A)P (¬B))2
(P (¬A¬B)−P (¬A)P (¬B))2
+|G|
+
P (A)P (B)
P (¬A)P (¬B)
J-measure
Jaccard
P (AB) − P (A)P (B)
P (B|A)/P (B|¬A)
P (AB)/P (A¬B)
P (AB)
P (AB) log2 P (A)P
(B)
P (AB)−P (A)P (B)
Linear Correlation Coefficient √
Piatetsky-Shapiro
relative risk
Sebag-Schoenaue
two-way support
P (A)P (B)P (¬A)P (¬B)
Zhang
(P (AB) − P (A)P (B)) / max (P (AB)P (¬B), P (B)P (A¬B))
Table 3: Aggregation functions
Real-valued
Ordinal (for sorted
Numerical
values X(1) , X(2) , . . . , X(n) )
X (n−1) , if n is odd
P
2
Sum
Median
v
0.5(X n2 + X n2 +1 ),
v∈V
if n is even
P
v//|V
|
Arithmetic mean
Maximum
X(n)
v∈V
q
Q
|V |
Geometric mean
Minimum
X(1)
v∈V v
P
Harmonic mean
|V |/( v∈V v −1 )
Midrange
0.5(X(1) + X(n) )
Fuzzy
T-norms
S-norms
min(v1 , v2 ),
max(v1 , v2 ),
if max(v1 , v2 ) = 1 Drastic sum
if min(v1 , v2 ) = 0
Drastic product
0, otherwise
1, otherwise
Bounded difference max(0, v1 + v2 − 1) Bounded sum
min(1, v1 + v2 )
v1 v2
v1 +v2
Einstein product
Einstein sum
2−(v1 +v2 −v1 v2 )
1+v1 v2
Algebraic product v1 v2
Probabilistic sum v1 + v2 − v1 v2
v2
v1 +v2 −2v1 v2
Hamacher product v1 +vv21 −v
Hamacher sum
v
1−v1 v2
1 2
Minimum
min(v1 , v2 )
Maximum
max(v1 , v2 )
A new index:
Index2 (A, B) = |M \ B|
X
1/P (m|B).
m∈M \B
– Notion of interestingness: a concept (A, B) is interesting if it stands out from
the context. We study out-of-concept elements of a context.
– Basic measure: conditional probability.
– Aggregation operator (applied to homogeneous elements): harmonic mean.
5
Quality of Indices
Application of indices in practice is aimed mostly at addressing the following
goals: selecting a subset from the whole set of concepts or computing only a
fragment of a concept lattice. The first goal then is attained by selecting the most
interesting concepts from the set of concepts or detecting “original” concepts
computed on a noisy context. The second goal is to reduce the computation cost
by constructing only the most interesting concepts.
5.1
Rank Correlation
To realize how similar the indices are, we examined the similarity of the concept
rankings by values of indices. The pairwise similarity of indices is measured by
the Kendall tau correlation coefficient [26], this coefficient takes into account not
only an absolute rank, but also a relative position.
We randomly generated 4 groups of formal concepts with the following densities (the rate of the “1” in the context): 0.1, 0.2, 0.3, 0.4. Each group consisted
of 100 formal contexts with the number of attributes ranged between 10 and 50,
and the number of objects varying from 40 to 80.
The standard deviation of within-group pairwise correlation is quite small
(not more then 0.05) and does not depend on the context density (p-values of
Levene’s test is less than 0.05). Since the studied values cannot be assumed to be
normally distributed (based on D’Agostino and Pearson’s test), the Wilcoxon’s
test was used to compare mean values of the Kendall tau coefficient.
The averaged values of the pairwise Kendall tau coefficients are presented in
Table 4. Among investigated indices several groups of correlated measures stood
out, but only few values of pairwise correlation are statistically stable (have the
same average value w.r.t. context densities).
One of the groups of correlated indices corresponds to approximate robustness with different values of parameters. It allows us to conclude that relative
importance of concepts is mostly preserving regardless of α. Another class of indices that utilizes similarity and predictability approaches (Basic Level Metrics
group) yields the second group. It should be noted that the highly correlated indices of this group are the measures based on similarity approach with the same
cohesion function and with different aggregation functions for sub/superordinate
concepts. The conclusion agrees with the results presented in [6].
The other groups of correlated indices are highlighted in Table 4. Stability
(robustness) is the most complex index for calculation, hence it is important to
identify more easily computable indices and use them instead. In our experiments
we found that logarithmic estimates of stability (∆l , ∆h ) and Max-distinguishedextents upper bound (stab2OE ) are strongly correlated with stability, highly
correlated robustness indices belong to the same complexity class as stability.
5.2
Integral Stability Indices: Best Approximation of Stability
Here, we consider the problem of stability approximation w.r.t. different levels
of integral stability index, i.e., the number of levels that are used to compute
the index. Since the extent size of concepts varies in the range from 0 to |G|, for
each concept we took the level of stability depending on the size of set A, to be
precise, the level is defined by formula [rate ∗ |A|], where 0 < rate < 1.
To find the best rate we used randomly generated formal contexts (described
in the previous subsection) of different densities: 0.1, 0.2, 0.3 and applied simple
linear regression. The integral stability index of the jth level was taken as a
regressor and the stability as a dependent variable.
Scatter plots of the studied indices are given in Figure 3. As we can see from
the diagrams, taking too small rates does not allow us to estimate stability, since
for most concepts small-sized subsets Y ⊂ A, such that Y 0 = B do not exist.
The first local maximum of the coefficients of determination for model
stability = A ∗ stabilityΣratio + B
CU
CFC
CV
P
SJmm
SJma
SJam
SJaa
mm
SSM
C
ma
SSM
C
am
SSM
C
aa
SSM
C
marg-clos∗
support
separation
probability
rob0,8
rob0,5
rob0,3
rob0,1
stab2OIE
stab2OE
stab2N OE
∆h
∆l
stability
0,88 0,73 0,48 0,63 0,54 0,86 0,94 1,00 0,97 0,27 0,18 0,32 0,12 0,17 0,16 0,05 0,05 0,18 0,18 0,07 0,08 -0,02 0,05 0,06 0,07
∆l
0,55 0,34 0,50 0,40 0,88 0,93 0,88 0,83 0,15 0,13 0,18 0,20 0,16 0,16 0,00 0,00 0,13 0,13 0,01 0,02 -0,06 0,00 -0,02 -0,05
∆h
0,66 0,67 0,65 0,53 0,64 0,73 0,76 0,55 0,28 0,65 -0,07 0,16 0,16 0,15 0,15 0,29 0,30 0,19 0,21 0,08 0,16 0,27 0,34
stab2N OE
0,88 0,95 0,32 0,41 0,48 0,51 0,67 0,36 0,80 -0,47 0,34 0,36 0,35 0,35 0,53 0,53 0,45 0,47 0,30 0,21 0,38 0,50
stab2OE
0,94 0,47 0,56 0,63 0,66 0,60 0,32 0,72 -0,37 0,38 0,38 0,39 0,39 0,54 0,55 0,47 0,48 0,34 0,18 0,29 0,40
stab2OIE
0,38 0,47 0,54 0,57 0,65 0,36 0,78 -0,45 0,36 0,37 0,37 0,37 0,53 0,54 0,46 0,47 0,32 0,21 0,35 0,47
rob0.1
0,91 0,86 0,81 0,14 0,12 0,17 0,22 0,17 0,16 0,02 0,02 0,12 0,13 0,02 0,03 -0,04 -0,02 -0,05 -0,05
rob0.3
0,94 0,89 0,21 0,15 0,26 0,16 0,17 0,16 0,04 0,03 0,16 0,16 0,05 0,06 -0,03 0,02 0,01 0,02
rob0.5
0,97 0,27 0,18 0,32 0,12 0,17 0,16 0,05 0,05 0,18 0,18 0,07 0,08 -0,02 0,05 0,06 0,07
rob0.8
0,30 0,19 0,36 0,09 0,16 0,16 0,06 0,05 0,19 0,19 0,08 0,09 -0,01 0,07 0,08 0,10
probability
0,27 0,81 -0,51 0,22 0,24 0,29 0,30 0,44 0,44 0,35 0,38 0,28 -0,02 0,31 0,36
separation
0,40 -0,27 0,25 0,27 0,17 0,17 0,24 0,25 0,26 0,25 0,14 0,40 0,25 0,18
support
-0,62 0,27 0,29 0,32 0,33 0,53 0,53 0,43 0,45 0,28 0,23 0,48 0,61
∗
marg-clos
-0,26 -0,28 -0,37 -0,38 -0,47 -0,47 -0,50 -0,51 -0,39 -0,19 -0,36 -0,53
aa
SSM C
0,91 0,58 0,57 0,61 0,61 0,50 0,51 0,52 0,07 0,03 0,16
am
SSM
0,57 0,57 0,62 0,63 0,50 0,50 0,51 0,08 0,05 0,17
C
ma
SSM
0,97 0,59 0,59 0,68 0,69 0,73 0,00 0,01 0,16
C
mm
SSM
0,59 0,59 0,67 0,68 0,73 0,00 0,01 0,16
C
SJaa
0,97 0,61 0,62 0,55 0,11 0,17 0,36
SJam
0,61 0,62 0,55 0,11 0,17 0,35
ma
SJ
0,93 0,65 0,12 0,11 0,30
mm
SJ
0,66 0,11 0,12 0,29
P
-0,05 -0,03 0,12
CV
0,35 0,39
CFC
0,42
Table 4: The averaged Kendall tau coefficient for indices. Statistically equal
average values are in bold type.
Fig. 3: The dependence of stability values on integral stability values
corresponds to rate = 0.4 (0.78 and 0.79 for context densities 0.2 and 0.3, respectively). Thus, this value can be used as the most suitable one for computing
estimates of stability by integral stability index.
5.3
Noise Filtering
In practice one usually faces noisy data. Even a small noise rate can result in
exponential explosion of the number of formal concepts [29]. In this connection,
we study the ability of indices to select original concepts from the set of concepts
computed on a noisy context. We took 5 formal concepts, the lattices of the first
four of them have quite simple structure (see Figure 4) and the last one is a
fragment of Mushroom dataset 1 consisted of 500 objects and 14 attributes.
[a]
[b]
[c]
[d]
Fig. 4: The concept lattices of formal contexts with 300 objects and 6 attributes
(a - c), with 400 objects and 4 attributes (d)
We added different amount of noise into the contexts as follows: each entry
changes with given probability (noise rate). For “noisy” lattices we identified
original concepts and considered a binary classification problem. We computed
the AUC (Area Under the Receiver Operator Curve) from each index separately.
The averaged AUCs within groups corresponding the same datasets with different noise rates and within groups with the same noise rate for the described
above contexts are given in Figure 5.
As can be seen in Figure 5, the index quality mostly depends on lattice structure rather than on noise rate. Cue Validity (CV), Category Feature Collocation
(CFC), Category Utility (CU) and separation have the highest AUC on particular datasets, but the quality considerably changes depending on the lattice
structure. For instance, AUC of CFC varies from 0.67 to 0.95. The most stable
results correspond to robustness with α ∈ {0.3, 0.5, 0.8}, the estimates of stability are able to distinguish most of the original concepts (the AUC is greater
than 0.7).
Almost all indices are stable with respect to different noise rates. The poor
data quality has the strong impact on the estimates of stability: the quality of
these indices drops down as the noise rate increases. Since these measures are
based on the elements of a lattice, the more noise is introduced, the more noisy
1
https://archive.ics.uci.edu/ml/datasets/Mushroom
Fig. 5: The averaged AUC within groups with the same noise rate (dotted lines)
and corresponding to the same datasets (lines a-d)
concepts are involved in computation. The quality of similarity-based indices
(Basic Level group) is close to random guessing, which makes them inapplicable
to the analysis of noisy data.
The experiments allow us to conclude that the CV, CU, CFC, separations
and robustness are the most suitable for the analysis of noisy data.
6
Conclusion
In this paper we have presented some results on formal concept indices. We
have also analyzed the existing indices for closed itemsets in Data Mining. We
defined the main features of the indices and proposed their classification using an
FCA-based approach. We have also provided some examples of most interesting
groups of indices selected by means of some of the studied indices.
We have given basic ideas of adapting indices for arbitrary itemsets to closed
ones. We have suggested to utilize bimodal nature of the concepts to get new
indices on the basis of the indices for arbitrary set of attributes.
An important part of the study is devoted to practical application of indices.
We have performed a comparative study of indices in the context of the following tasks: selection of the most interesting concepts, approximation of the
exponentially computable indices (stability) and filtering noisy data. The results of our experiments allow us to distinguish groups of correlated indices,
thus these results can be used to reduce computational complexity of concept
mining by choosing easily computable indices among the correlated ones. Another important aspect of our study is identification of indices that can be used
for the analysis of noisy data. It was shown that the noise filtering quality of
indices depends more on the structure of the concept lattice than on the noise
rate. The strongest dependence of the filtering quality on noise rate corresponds
to the estimates of stability. Even a small proportion of noise added to data
can significantly change the lattice structure, which results in biased values of
neighbor-based indices.
As possible directions of future work, we can propose the study of indices for
approximate concepts, e.g., biclusters, as well as indices for multimodal concepts.
Another important application of indices as tools for selecting hypothesis can be
examined within the framework of classification and learning.
Acknowledgments
This paper was prepared within the framework of the Basic Research Program
at the National Research University Higher School of Economics (HSE) and supported within the framework of a subsidy by the Russian Academic Excellence
Project ’5-100’.
List of indices
Used notation
stability (3.2)
Description
It estimates the probability of the closeness of
an attribute set after removing some elements
from the object set.
∆l (3.2)
∆-measure that is based on direct descendents
of a concept.
∆h (3.2)
∆-measure that is based on the maximal extent.
stab2N OE (max-disjoint-extents ∆-measure that is based on two maximal disupper bound, 3.2)
joint extents.
stab2OE
(max-distinguished- ∆-measure that is based on the maximal extent
extents upper bound, 3.2)
and the extent with maximal disjoint subset.
stab2OIE (max-extent upper ∆-measure that is based on two maximal exbound, 3.2)
tents.
robustness (3.2)
It estimates the probability that an attribute set
will be arose from the data, if some transactions
are removed.
probability (3.2)
It estimates the probability that a set of attributes is closed, if all attributes are independent and the probability of an attribute is assumed to be equal to its frequency.
separation (3.2)
The ratio of the area covered by a formal concept to all nonzero elements in rows and columns
corresponding to the concept.
frequency, support (3.2)
The frequency of an attribute set (if all attributes are assumed to be independent).
monocle (3.2)
The index assigns weights to concepts by taking
into account the size of a context and the number of the concepts that do not contain objects
and attributes contained to the concept.
δ-TCFIs (3.2)
It selects relatively frequent concepts with respect to their direct descendants.
margin-closed (3.2)
The index characterizes whether a frequent
intent have well-distinguishable support w.r.t.
other frequent intents, supersets of the intent.
S∗∗∗ (similarity, 3.2)
A fuzzy combination of three cohesion functions
which correspond to the following conditions:
high cohesion of concepts, considerably greater
cohesion with respect to upper neighbors and a
slightly less cohesion with respect to lower neighbors. The bottom star is replaced by a similarity function defined on the two sets of attributes
that correspond to particular objects (SM C and
J denotes simple matching coefficient and Jaccard similarity, respectively). The first and the
second top stars are replaced by cohesion functions defined on the set of neighbouring concepts
and objects, respectively.
P (predictability, 3.2)
Defined the same way as the previous one, the
only difference is the cohesion function is replaced by a predictability function.
CV (cue validity, 3.2)
The sum of the conditional probabilities of objects from an extent given attributes from an
intent.
CFC (category feature colloca- Characterizes specificity of all attributes of a
tion, 3.2)
context for given set of objects.
CU (category utility, 3.2)
Characterizes specificity each context attribute
for a given concept.
References
References
1. Agrawal, R., Srikant, R., et al.: Fast algorithms for mining association rules. In:
Proc. 20th int. conf. very large data bases, VLDB. vol. 1215, pp. 487–499 (1994)
2. Arévalo, G., Berry, A., Huchard, M., Perrot, G., Sigayret, A.: Performances of
Galois sub-hierarchy-building algorithms. In: International Conference on Formal
Concept Analysis, pp. 166–180. Springer (2007)
3. Azevedo, P.J., Jorge, A.M.: Comparing rule measures for predictive association
rules. In: Machine Learning: ECML 2007, pp. 510–517. Springer (2007)
4. Babin, M., Kuznetsov, S.: Approximating concept stability. In: Domenach, F., Ignatov, D., Poelmans, J. (eds.) Formal Concept Analysis. Lecture Notes in Computer
Science, vol. 7278, pp. 7–15. Springer Berlin Heidelberg (2012)
5. Belohlavek, R., Macko, J.: Selecting important concepts using weights. In:
Valtchev, P., Jschke, R. (eds.) Formal Concept Analysis, Lecture Notes in Computer Science, vol. 6628, pp. 65–80. Springer Berlin Heidelberg (2011)
6. Belohlavek, R., Trnecka, M.: Basic level of concepts in formal concept analysis. In:
Domenach, F., Ignatov, D., Poelmans, J. (eds.) Formal Concept Analysis, Lecture
Notes in Computer Science, vol. 7278, pp. 28–44. Springer Berlin Heidelberg (2012)
7. Belohlavek, R., Trnecka, M.: Basic level in formal concept analysis: Interesting
concepts and psychological ramifications. In: Proceedings of the Twenty-Third International Joint Conference on Artificial Intelligence. pp. 1233–1239. IJCAI ’13,
AAAI Press (2013)
8. Belohlavek, R., Vychodil, V.: Formal concept analysis with background knowledge:
attribute priorities. Systems, Man, and Cybernetics, Part C: Applications and
Reviews, IEEE Transactions on 39(4), 399–409 (2009)
9. Berry, A., Huchard, M., McConnell, R., Sigayret, A., Spinrad, J.: Efficiently computing a linear extension of the sub-hierarchy of a concept lattice. In: Ganter, B.,
Godin, R. (eds.) Formal Concept Analysis, Lecture Notes in Computer Science,
vol. 3403, pp. 208–222. Springer Berlin Heidelberg (2005)
10. Buzmakov, A., Kuznetsov, S., Napoli, A.: Sofia: How to Make FCA Polynomial?
In: Proceedings of the 4th International Conference on What Can FCA Do for Artificial Intelligence? - Volume 1430. pp. 27–34. FCA4AI’15, CEUR-WS.org, Aachen,
Germany, Germany (2015)
11. Buzmakov, A., Kuznetsov, S.O., Napoli, A.: Fast generation of best interval patterns for nonmonotonic constraints. In: Machine Learning and Knowledge Discovery in Databases, pp. 157–172. Springer (2015)
12. Buzmakov, A., Kuznetsov, S., Napoli, A.: Scalable estimates of concept stability.
In: Glodeanu, C., Kaytoue, M., Sacarea, C. (eds.) Formal Concept Analysis, Lecture Notes in Computer Science, vol. 8478, pp. 157–172. Springer International
Publishing (2014)
13. Carpineto, C., Romano, G.: A lattice conceptual clustering system and its application to browsing retrieval. Machine Learning 24, 95–122 (1996)
14. Cheng, J., Ke, Y., Ng, W.: Delta-tolerance closed frequent itemsets. In: Data Mining, 2006. ICDM’06. Sixth International Conference on. pp. 139–148. IEEE (2006)
15. Cheung, K., Vogel, D.: Complexity reduction in lattice-based information retrieval.
Information Retrieval 8(2), 285–299 (2005)
16. Dhillon, I., Modha, D.: Concept decompositions for large sparse text data using
clustering. Machine Learning 42(1-2), 143–175 (2001)
17. Dias, S.M., Vieira, N.: Reducing the size of concept lattices: The JBOS approach.
In: Proceedings of the 7th International Conference on Concept Lattices and Their
Applications, Sevilla, Spain, October 19-21, 2010. pp. 80–91 (2010)
18. Dobša, J., Dalbelo-Bašić, B.: Comparison of information retrieval techniques: latent semantic indexing and concept indexing. Journal of Inf. and Organizational
Sciences 28(1-2), 1–17 (2004)
19. Düntsch, I., Gediga, G.: Simplifying contextual structures. In: Kryszkiewicz, M.,
Bandyopadhyay, S., Rybinski, H., Pal, S.K. (eds.) Pattern Recognition and Machine Intelligence, Lecture Notes in Computer Science, vol. 9124, pp. 23–32.
Springer International Publishing (2015)
20. Emilion, R.: Concepts of a discrete random variable. In: Brito, P., Cucumel, G.,
Bertrand, P., de Carvalho, F. (eds.) Selected Contributions in Data Analysis and
Classification, pp. 247–258. Studies in Classification, Data Analysis, and Knowledge Organization, Springer Berlin Heidelberg (2007)
21. Ganter, B., Wille, R.: Contextual attribute logic. In: Tepfenhart, W.M., Cyre, W.
(eds.) Conceptual Structures: Standards and Practices, Lecture Notes in Computer
Science, vol. 1640, pp. 377–388. Springer Berlin Heidelberg (1999)
22. Gauss, C.F.: Disquisitiones Arithemeticae, Arthur A. Clarke (English translator)
(corrected 2nd ed.). Springer-Verlag, New York (1986)
23. Geng, L., Hamilton, H.J.: Interestingness measures for data mining: A survey. ACM
Computing Surveys (CSUR) 38(3), 9 (2006)
24. Ignatov, D.I., Gnatyshak, D.V., Kuznetsov, S.O., Mirkin, B.G.: Triadic formal concept analysis and triclustering: searching for optimal patterns. Machine Learning
101(1-3), 271–302 (2015)
25. Jay, N., Kohler, F., Napoli, A.: Analysis of social communities with iceberg and
stability-based concept lattices. In: Medina, R., Obiedkov, S. (eds.) Formal Concept
Analysis, Lecture Notes in Computer Science, vol. 4933, pp. 258–272. Springer
Berlin Heidelberg (2008)
26. Kendall, M.G.: A new measure of rank correlation. Biometrika pp. 81–93 (1938)
27. Kirchgessner, M., Leroy, V., Amer-Yahia, S., Mishra, S.: Testing interestingness
measures in practice: A large-scale analysis of buying patterns. arXiv preprint
arXiv:1603.04792 (2016)
28. Klement, E.P., Mesiar, R., Pap, E.: Triangular norms. Springer Netherlands (2000)
29. Klimushkin, M., Obiedkov, S., Roth, C.: Approaches to the selection of relevant
concepts in the case of noisy data. In: Kwuida, L., Sertkaya, B. (eds.) Formal
Concept Analysis, Lecture Notes in Computer Science, vol. 5986, pp. 255–266.
Springer Berlin Heidelberg (2010)
30. Kuznetsov, S.O.: Stability as an estimate of degree of substantiation of hypotheses
derived on the basis of operational similarity. Nauchn. Tekh. Inf., Ser. 2 (12), 21–29
(1990)
31. Kuznetsov, S.O.: On stability of a formal concept. Annals of Mathematics and
Artificial Intelligence 49(1-4), 101–115 (2007)
32. Kuznetsov, S.O., Makhalova, T.P.: Concept interestingness measures: a comparative study. In: Proceedings of the Twelfth International Conference on Concept
Lattices and Their Applications. pp. 59–72 (2015)
33. Kuznetsov, S.O., Obiedkov, S., Roth, C.: Reducing the representation complexity
of lattice-based taxonomies. In: Conceptual Structures: Knowledge Architectures
for Smart Applications, pp. 241–254. Springer Berlin Heidelberg (2007)
34. Kuznetsov, S.O., Poelmans, J.: Knowledge representation and processing with formal concept analysis. Wiley Interdisciplinary Reviews: Data Mining and Knowledge Discovery 3(3), 200–215 (2013)
35. Kuznetsov, S.O.: Interpretation on graphs and complexity characteristics of a
search for specific patterns. Nauchno-Tekhnicheskaya Informatsiya Ser. 2 pp. 37–45
(1989)
36. Madeira, S.C., Oliveira, A.L.: Biclustering algorithms for biological data analysis:
a survey. IEEE/ACM Transactions on Computational Biology and Bioinformatics
(TCBB) 1(1), 24–45 (2004)
37. Mannila, H., Toivonen, H., Verkamo, A.I.: Efficient algorithms for discovering association rules. In: KDD-94: AAAI workshop on Knowledge Discovery in Databases.
pp. 181–192 (1994)
38. Moerchen, F., Thies, M., Ultsch, A.: Efficient mining of all margin-closed itemsets
with applications in temporal knowledge discovery and classification by compression. Knowledge and Information Systems 29(1), 55–80 (2011)
39. Murphy, G.L.: The big book of concepts. MIT press (2002)
40. Poelmans, J., Ignatov, D.I., Kuznetsov, S.O., Dedene, G.: Formal concept analysis
in knowledge processing: A survey on applications. Expert systems with applications 40(16), 6538–6560 (2013)
41. Poelmans, J., Kuznetsov, S.O., Ignatov, D.I., Dedene, G.: Formal concept analysis
in knowledge processing: A survey on models and techniques. Expert systems with
applications 40(16), 6601–6623 (2013)
42. Rosch, E.: Principles of categorization. Hillsdale, NJ: Lawrence Erlbaum (1978)
43. Roth, C., Obiedkov, S., Kourie, D.: Towards concise representation for taxonomies
of epistemic communities. In: Concept Lattices and Their Applications, pp. 240–
255. Springer (2008)
44. Snásel, V., Polovincak, M., Abdulla, H.M.D., Horak, Z.: On concept lattices and
implication bases from reduced contexts. In: ICCS Supplement
45. Stumme, G., Taouil, R., Bastide, Y., Pasquier, N., Lakhal, L.: Computing iceberg
concept lattices with titanic. Data Knowl. Eng. 42(2), 189–222 (Aug 2002)
46. Tatti, N., Moerchen, F., Calders, T.: Finding robust itemsets under subsampling.
ACM Transactions on Database Systems (TODS) 39(3), 20 (2014)
47. Torim, A., Lindroos, K.: Sorting concepts by priority using the theory of monotone
systems. In: Conceptual Structures: Knowledge Visualization and Reasoning, pp.
175–188. Springer (2008)
48. Zaki, M.J., Meira Jr, W., Meira, W.: Data mining and analysis: fundamental concepts and algorithms. Cambridge University Press (2014)
49. Zeigenfuse, M.D., Lee, M.D.: A comparison of three measures of the association
between a feature and a concept. In: Proceedings of the 33rd Annual Conference
of the Cognitive Science Society. pp. 243–248 (2011)
| 2cs.AI
|
arXiv:1712.04034v1 [cs.CL] 11 Dec 2017
Learning Robust Dialog Policies in Noisy
Environments
Maryam Fazel-Zarandi
Amazon Inc, USA
[email protected]
Shang-Wen Li
Amazon Inc, USA
[email protected]
Jin Cao
Amazon Inc, USA
[email protected]
Jared Casale
Amazon Inc, USA
[email protected]
Peter Henderson
Amazon Inc, USA
[email protected]
David Whitney
Brown University
[email protected]
Alborz Geramifard
Amazon Inc, USA
[email protected]
Abstract
Modern virtual personal assistants provide a convenient interface for completing
daily tasks via voice commands. An important consideration for these assistants
is the ability to recover from automatic speech recognition (ASR) and natural
language understanding (NLU) errors. In this paper, we focus on learning robust
dialog policies to recover from these errors. To this end, we develop a user simulator which interacts with the assistant through voice commands in realistic scenarios
with noisy audio, and use it to learn dialog policies through deep reinforcement
learning. We show that dialogs generated by our simulator are indistinguishable
from human generated dialogs, as determined by human evaluators. Furthermore,
preliminary experimental results show that the learned policies in noisy environments achieve the same execution success rate with fewer dialog turns compared
to fixed rule-based policies.
1
Introduction
Modern speech-based assistants, such as Amazon Alexa, Apple Siri, and Google Assistant, enable
users to complete daily tasks such as shopping, setting reminders, and getting answers to factual
questions using vocal commands. Such human-like interfaces create a rich experience for users by
enabling them to complete many tasks hands-free and eyes-free in a conversational manner. Depending on how complicated a task is, multiple rounds of conversation may be needed for the assistant to
fully understand user requests. To achieve this, these interfaces require the design of complex dialog
policies which can generate appropriate responses to queries and steer the conversation.
Designing dialog policies in current speech-based assistants is challenging and time-consuming. In
particular, responding to a user can be difficult since the selection of system actions is conditioned on
the task, the user goals and preferences, and the dialog history. Additionally, environment noise and
ambiguous user utterances present added complexity. Communicating with speech-based assistants
always involves a noisy communication channel and the assistant should be able to recover from
automatic speech recognition (ASR) and natural language understanding (NLU) errors by confirming
its recognition results or eliciting more information from the user. To address these issues, both
rule-based strategies and statistical modeling techniques have been developed [27]. Rule-based
NIPS 2017 Workshop on Conversational AI: Today’s Practice and Tomorrow’s Potential, Long Beach, CA, USA.
systems are costly to design and maintain, and generally assume the dialog state is fully observable
with only a limited account of errors and uncertainties. Statistical modeling techniques, on the other
hand, are able to generalize to such unknowns, but rely on the availability of huge amounts of training
data. Some statistical-modeling techniques formulate dialog management as a reinforcement learning
(RL) problem [16]. In this formulation, the goal is to adjust a parameterized dialog policy in order to
maximize the expected cumulative reward over the course of a dialog. To reduce the cost of collecting
the large amount of data needed by such models, researchers have focused on the development of
user simulators which interact with the dialog policy (e.g., [6, 8, 9, 12, 27, 29]). Most of the existing
simulators are domain dependent [9] and lack robustness against noise in real user utterances in order
to recover from ASR/NLU errors [19].
In this paper, we focus on learning dialog policies in noisy environments where the assistant needs to
recover from ASR/NLU errors. The key contribution of this paper is the development of a realistic
user simulator which interacts with the assistant through voice commands. The simulator is trained
on real-world data and takes dialog context into account. Furthermore, while other works simulate
NLU errors at the intent- or slot-level [18], our simulator is able to use artificial noise in audio signals
to induce a similar corpus-specific ASR word-error rate (WER). This error can then propagate to
realistic intent- or slot-level errors. Using dialogs collected from MovieBot [25], an Alexa Skill [14]
that converses with users about existing, new, and forthcoming movies, we show that our simulator
generates dialogs that are indistinguishable from human generated dialogs. Using the simulator, we
leverage deep RL to learn robust dialog policies that overcome upstream ASR/NLU errors in a noisy
environment. Experimental results show that the learned policies achieve the same execution success
rate with fewer dialog turns compared to fixed rule-based policies in noisy environments.
2
Related Work
A dialog system can be formalized as a Markov Decision Process (MDP) [16]. An MDP is a tuple
<S, A, P, R, γ> of states, actions, transition probability function, reward, and discount factor. In the
formulation of RL, at each time step t, the agent observes a state st ∈ S and selects an action at ∈ A
according to its policy (π : S → A). After performing the selected action, the agent receives the next
state st+1 and a scalar reward rt . The trajectory restarts after the agent reaches a terminal state. With
this formalization, RL can be used to find the optimal dialog policy (e.g., [26, 34, 15, 13]). In this
context, at each turn the system acts based on its understanding of what the user has said, and the
reward function is modeled in terms of different dimensions such as per-interaction user satisfaction,
accomplishment of the task, efficiency of interaction, dialog duration, etc.
Recently, deep RL has also been applied to the problem of dialog management. Cuayahuitl [7]
presented an implementation of DQN [23, 24]. Zhao and Eskenazi [36] similarly used a deep
recurrent Q-network (DRQN) to learn both state tracking and dialog policy for task-oriented dialogs.
Fatemi et al. [10] compared different deep RL techniques on a restaurant-finding dialog problem.
They showed that DDQN [32] converges faster than DQN for this domain.
An important challenge in using RL for learning dialog policies is creating realistic user simulators
that can generate natural conversations similar to a human user [27]. Existing approaches for
simulating users can be categorized into two groups: intent-level and utterance-level simulation. An
intent-level simulator abstracts a conversation as a sequence alternating between user intents and slots
and bot actions, where an intent indicates the user’s intention (e.g., getting the plot) and slots represent
information about a particular entity (e.g., movie title). The goal is to predict the next user intent based
on the dialog history (e.g., [8, 11, 18, 28]). These methods require extensive work to manually design
the user goal or agenda, which makes them less generalizable. Additionally, it is difficult to simulate
real-world recognition errors. Utterance-level user simulators, on the other hand, attempt to generate
natural language user responses. To this end, template-based natural language generation is usually
adopted to output utterances based on handcrafted rules and predicted user intents [18]. Inspired
by recent studies in sequence-to-sequence machine translation [31], end-to-end dialog simulators
generate utterances with minimal feature engineering [9, 5]. However, these techniques often produce
generic responses [17]. Additionally, they have to be trained with a significant amount of dialogs,
and thus they do not work well for domain-specific applications.
Given that the dialog state is not fully observable due to ASR/NLU errors and uncertainty over a user’s
goals, researchers have also investigated the use of Partially Observable Markov Decision Processes
2
(POMDPs) [35]. However, these systems are complex and exact policy learning for POMDPs is
intractable [35]. Rule-based policies, on the other hand, use ASR/NLU confidence scores to recover
from errors. However, asking too many clarifying questions can frustrate users and in some scenarios
it might be more beneficial to execute on available information. Additionally, ASR/NLU components
are not static and as such policies with fixed confidence levels need to be constantly updated.
3
User Simulation
To develop a user simulator that generates voice responses to the agent, we adopt a hybrid of language
model based intent prediction and template-based utterance generation. The overall architecture of
the user simulator is illustrated in Figure 1.
The user simulator first generates the next intent and slot type based on dialog context. Dialog context
includes previous intent and slot type pairs in the conversation, as well as text responses from the
agent encoded with word2vec [22]. We model intent generation as a language modeling problem and
use recurrent neural networks (RNNs) to predict the trajectory of intents. In this formulation, each
possible intent and slot type pair form a token in the vocabulary, and every training dialog becomes a
training intent sequence. For example, the sequence for the conversation with MovieBot in Table 1 is
[Start, GetGenreMoviesIntent+genre, GetNextActorIntent, GetRatingIntent, End].
Given the predicted intent and slot type, the user simulator uniformly samples one utterance from
the templates and fills the utterance with slot values uniformly sampled from a lexicon of the
corresponding slot type. For example, given intent GetDirectorIntent and slot type movie_title, a
sample utterance from the template is "who is the director of {movie_title}?". {movie_title} is
then replaced uniformly by a value from the lexicon of movie titles. When the simulator is responding
to a clarifying question, it answers yes or no to a confirmation question, but repeats the original
utterance for elicitation.
Providing the sampled text directly to the MovieBot NLU model would result in practically zero
errors as the sample utterances are used in training the NLU model via the ASK portal. Therefore,
we take the approach of generating a spoken audio signal from the text, injecting noise into the
audio and then passing the audio to ASR and NLU models of the MovieBot Skill. This process
simulates the real-world sources of error such as background noise in a person’s home. Concretely, we
synthesize an audio signal with the text-to-speech (TTS) service provided by Amazon Polly [1] and
contaminate the clean TTS output using PyAcoustics [20] to achieve a desired signal-to-noise ratio
(SNR). Contamination is performed by filtering a white noise signal using a Fast Fourier transform
applied to the original TTS output. Given our choice of architecture, any noise injection strategy may
be developed without impacting the rest of the experimental setup. For example, in the future we plan
User
Simulator
Conversational
Agent
Text Response
TTS
Add Noise
ASR + NLU
Dialog Policy
Simulated user utterance
User Utterance Simulated
Generator
intent + slot
Intents, Slots
Text
Response
Intent/Slot
Generator
Response
Generator
Action
Previous Intents/Slots
Figure 1: User simulator and conversational agent interaction. We use the text response generated by
the agent before sending it to TTS to eliminate the need for ASR/NLU on the user simulator side.
3
Table 1: Example of a conversation with MovieBot. The intent indicates the user’s intention and slots
(e.g., genre) represent information about a particular entity.
Speaker
User
Bot
User
Bot
User
Bot
User
Bot
Utterance
A science fiction movie [intent = GetGenreMoviesIntent, genre = science fiction]
How about Arrival, starring Amy Adams. If you want, you can say who else is in it.
Who else is in it? [intent = GetNextActorIntent]
The character of Ian Donnelly was played by Jeremy Renner.
Is it any good? [intent = GetRatingIntent]
It is rated 8.4 on I.M.D.b., based on about 70000 votes.
Thank you [intent = AMAZON.StopIntent]
Thank you for using MovieBot.
to explore the impact of using different kinds of noise such as background noise, speaker-dependent
speech variations, or multiple concurrent speakers. In this particular set of experiments, we tune the
injected white noise to a level that causes the ASR WER of the simulator to match that of MovieBot’s
deployed ASR model.
This approach to simulating users has a number of benefits. By predicting intents first, the model
becomes highly accurate and data efficient at the intent level due to the smaller number of intents
compared to words in the language. By using RNNs, our simulator learns long-term dependencies
between dialog context and intents from data to predict consistent user behavior. Additionally,
using handcrafted templates ensures that the generated utterances are of high quality, while the TTS
component with synthesized noise allows us to simulate real-world noisy environments.
3.1
Implementation and Evaluation
We used a corpus of 966 MovieBot dialogs consisting of 6349 user turns to train the intent generation
model with and without agent responses. The corpus is collected from real user interactions with
MovieBot and annotated at the utterance level. For example, a user request such as "How old is
Tom Hanks?" is annotated as intent=GetAgeIntent and slot PersonName=Tom Hanks. For intents
and slot types, we added an embedding layer initialized with small random values. We used a single
layer network followed by softmax, and for the hidden units we used two variants: vanilla RNN and
Gated Recurrent Unit (GRU) [4]. The output of the network is a sequence of vectors, each with total
number of intent/slot pair elements. The optimal parameters were found using grid search (Appendix
A). For all models we set the learning rate to 0.001, and used Google’s trained word2vec model for
encoding agent responses. As a baseline, we used a bigram model. Note that this model only takes
the intent and slot type inputs. With this setting, we found that the vanilla RNN models have the
lowest perplexity [30] scores across various trials (Table 2).
We also conducted a user study to evaluate if participants could distinguish simulator generated
dialogs from real human dialogs. The participants were each presented 10 dialogs via a web interface
and asked to determine if each dialog was generated by a human or a simulated user. Optionally, the
participants could also provide feedback as to why they made their decision. For each participant,
five dialogs were sampled from the annotated MovieBot data, and the first utterance of each was
used to generate five dialogs with the simulator. We used two models for the simulated data, namely,
the bigram model and the vanilla RNN model (see Appendix B for example of simulated dialogs).
The difference between the true positive rate (human dialog correctly identified as human) and false
positive rate (simulated user incorrectly identified as human) is of particular interest: more individuals
identifying simulated user dialogs as human indicates better user simulation.
Table 2: Perplexity results for the Bigram, RNN, and GRU models. Note that a lower value indicates
better performance.
Model
Intent and Slot Only
Intent and Slot + Bot Response
4
Bigram
16.270
−
RNN
7.675
7.274
GRU
7.769
7.892
Table 3: Results of the user study for distinguishing between real and simulated dialogs. The
difference between identifying human as human and simulator as human is small.
Number of Subjects
Total Number of Dialogs
Accuracy (Human/RNN + Bigram Models)
Precision (Human/RNN + Bigram Models)
Recall (Human/RNN + Bigram Models)
F1-Score (Human/RNN + Bigram Models)
Human Identified as Human
Simulator Identified as Human (RNN + Bigram Models)
Simulator Identified as Human (RNN Model)
Simulator Identified as Human (Bigram Model)
All Participants
Participants Who
Left Feedback
31
309
50.49%
50.29%
57.14%
53.50%
13
129
58.14%
57.14%
62.50%
59.70%
57.14%
56.13%
59.21%
53.16%
154 dialogs
155 dialogs
76 dialogs
79 dialogs
62.50%
46.15%
58.82%
32.26%
64 dialogs
65 dialogs
34 dialogs
31 dialogs
In total, 31 individuals from the Amazon Alexa team participated in the study and evaluated 309
dialogs. The results are reported in Table 3. Overall results indicate that the difference between
identifying human as human (57.14%) and simulator as human (56.13%) is not statistically significant
(p-value = 0.858). Additionally, the RNN model is performing slightly better than the bigram model.
Upon further inspection, we found a weak correlation (Pearson’s correlation coefficient: 0.38, p-value:
0.036) between how participants rated a dialog and whether or not they left feedback. The results
for the 13 participants who left feedback is reported separately in the table, clearly showing the
RNN (58.82% of simulated users identified as human) significantly (p-value = 0.033) outperforms
the bigram model (32.26% of simulated users identified as human). We hypothesize that this is due
to bigram dialogs being longer on average, and the participants who left feedback read the dialogs
more carefully. Criteria for selecting a dialog as a real human as indicated in the participant feedback
included: the expression of emotions (e.g., the user gets mad when the bot doesn’t answer correctly),
logical progression of the dialog, specific goals that humans have regardless of the dialog system’s
response, and naturalness of the dialogs. On the other hand, top criteria for identifying simulated
users included: asking a question without context, referring to non-existent entities in the dialog,
random responses by the user, and the user not getting frustrated. These comments illustrate that
better context tracking and use of sentiments would further improve user simulation. Nevertheless,
given the small difference between true positive and false positive results, the user simulator can
sufficiently replace real users for dialog policy learning.
4
Dialog Policy Learning
We used the proposed user simulator to train dialog policies through deep RL with the goal of conversational error recovery – that is, recovering from ASR/NLU errors through necessary clarifications.
The components of the dialog policy learning problem are as follows:
• S: The state is composed of 1) the hypothesis intent and slot, 2) the ASR and NLU intent
and slot confidence scores, and 3) the previous bot action.
• A: We constrain the action space to three actions that are critical for conversational error
recovery: execute, confirm, and elicit. execute accesses an IMDb database and answers
the user’s question. confirm and elicit are used to recover from ASR/NLU errors. confirm
clarifies the intent and/or the slot with the user (e.g., "Do you want movies directed by
Christopher Nolan?"), whereas elicit asks the user for missing information (e.g., "Which
movie are you talking about?").
• R: Table 4 summarizes the reward function. Here, ref indicates the reference intent and
slot as generated by the user simulator, and hyp refers to what the bot understands. If
hyp 6= ref , this indicates that an NLU error was induced by added noise to the audio
signal. The environment gives a large positive reward for executing correctly, a large
negative reward for executing incorrectly, and smaller negative rewards for confirm and
elicit. A successful execution is when the reference and hypothesis intent and slots are
equal. Clarification, although sometimes necessary, can frustrate users, hence the small
5
Table 4: Reward Function
a
s
r
execute(hyp)
execute(hyp)
confirm
elicit
terminate (after correct execute)
terminate (after incorrect execute)
hyp = ref
hyp 6= ref
∗
∗
∗
∗
+10
−12
−3
−6
+1
−5
negative rewards for clarifying actions. Additionally, if the user leaves the conversation
after a successful execution, the bot receives a small positive reward, but if the user leaves
after an incorrect execution, the bot receives a medium negative reward. The reason we give
a small positive reward to termination after a successful execution is that in the absence
of an explicit goal we assume the user received the requested information and decided to
terminate the dialog. All specific values were assigned through empircal analysis of user
interactions with the simulator.
For the deep RL experiments we used DQN and Dueling DDQN [33], with a fully-connected
Multi-Layer Perceptron (MLP) to represent the deep Q-network. The input of the network is the
concatenated embedding vector of intent and slot, the vector of confidence scores, and previous bot
action. Additionally, we tuned a window size to include previous dialog turns as input. The hidden
layers use a rectifier nonlinearity, and the output layer is a fully connected layer with linear activation
function and a single output for each valid action. We trained the agents using an -greedy policy
with decreasing linearly from 1 to 0.1 over 100,000 steps. We ran each method 30 times for 150,000
steps, and in each run, after every 10,000 steps, we sampled 30 dialog episodes with no exploration
to evaluate performance. All methods used the same set of random seeds and the best parameters
were empirically found for each method using Hyperopt [2]. Additionally, we experimented with two
fixed policies: 1) execute only and 2) execute with confirm and elicit. For the second fixed policy, we
learned global ASR and NLU confidence score thresholds for MovieBot by optimizing the simulated
agent’s average return (Table 5).
Table 5: Fixed Policy (MovieBot - Execute/Confirm/Elicit)
4.1
a
s
execute(hyp)
elicit
confirm
scoreASR ≥ 0.34 AND scoreN LU ≥ 0.56
scoreASR < 0.06
otherwise
Experimental Results
For our experiments, we varied the amount of noise such that we match the ASR WER of the
MovieBot corpus. This was achieved when SNR was set to 2.8. Figure 2 shows the simulation results
(see Appendix C for learned parameters). The Y-axis in the four subfigures is the average reward per
turn, the average discounted return, the average number of turns per dialog, and the execution success
rate of each technique, respectively. The X-axis in all four figures shows the number of learning
steps. Shaded areas highlight the standard error of the mean. RL approaches have the same dialog
episodes with the user with fewer clarifying questions, and as such result in around 40% relative gain
in average rewards per turn over the best rule-based method (Figure 2a). Furthermore, they achieve
the same success rate (Figure 2d) with 2 fewer turns per dialog on average (Figure 2c).
To better understand the advantage of RL approaches, Figure 3 illustrates what a user experiences in
order to obtain information in terms of number of turns needed to execute for each agent, where 1
means execute without any clarifying question. This figure illustrates that RL approaches learn to ask
fewer clarifying questions while achieving a similar success rate. In other words, the percentage of
unsuccessful executions stays the same with 6.2% for the Fixed Policy, 7.2% (±1.3%) for DQN, and
5.0% (±1.7%) for Dueling DDQN, with 0.2 less turns per execution.
6
(a)
(b)
(c)
(d)
Figure 2: Results during policy learning process. a) Average reward per turn, b) average discounted
return, c) average number of turns per dialog, and d) execution success rate. Shaded areas highlight
the standard error of the mean. RL approaches achieve our goal of recovering from ASR/NLU errors
without frustrating the user and result in around 40% relative gain in average rewards per turn over
the best rule-based method, with the same success rate and 2 fewer turns per dialog.
To gain insight into the learned policy’s properties, we conducted a qualitative analysis to compare
100 dialogs generated by Dueling DDQN and Fixed Policy agents. To achieve a fair comparison, we
fixed the random number generator so that the user simulator generated identical sequences of intents
and utterances to interact with both agents. Overall, 28% of the dialogs were identical. Considering
the remaining dialogs, two main reasons for the differences can be identified. First, the DDQN
agent can decide not to clarify based on the combination of specific intent and confidence scores.
Table 6 shows one dialog turn between the simulator and the two agents, along with the ground truth
of the simulated utterance and intent (first column of the table), the hypothesis from MovieBot’s
ASR/NLU component (second column), and the confidence scores of the hypotheses with a range
between 0 and 1 (third column). In this example, since the ASR score is relatively low (0.239), the
Figure 3: The number of turns needed in order to execute for Fixed Policy (Execute/Confirm/Elicit),
DQN and Dueling DDQN when testing on the learned policies. RL approaches learn to ask fewer
clarifying questions while achieving the same success rate.
7
Table 6: One dialog turn between a simulated user and Fixed Policy (MovieBot - Execute/Confirm/Elicit) agent, and Dueling DDQN agent, along with the ground truth of the utterance
and intent (first column of the table), the hypothesis from the ASR/NLU components (second column),
and the confidence scores of the hypotheses with a range between 0 to 1 (third column). The DDQN
agent can decide not to clarify based on the combination of specific intent and confidence scores.
Fixed Policy
USER: recommend a popular movie
BOT: Do you want popular movies? <confirm>
Dueling DDQN
USER: recommend a popular movie
BOT: A popular movie is It, a Drama Horror film, starring Bill Skarsgård. <execute>
reference
hypothesis
confidence score
recommend a popular movie
popular movies
0.239
GetPopularMoviesIntent
GetPopularMoviesIntent
0.769
Fixed Policy agent decides to ask a clarifying question (denoted as <confim>). However, the low
ASR score comes from the misrecognition of words which have small effect on the recognition of
the intent as indicated by the high NLU confidence score. As such, the DDQN agent learns that
the ASR confidence score is less important in this case and the agent executes the command and
provides the requested information. Secondly, the DDQN agent learns that elicitation brings more
information than confirmation since the user simulator repeats the original utterance. 82% of the
clarifying questions asked by the DDQN agent are elicitation, as opposed to 61% for the Fixed Policy.
5
Conclusion
In this paper, we demonstrate a method for learning robust dialog policies despite upstream NLU/ASR
errors caused by noisy environments. The user simulator presented here is able to mimic realistic
conversations, nearly indistinguishable from human dialogs, in the context of a real-world application
– the MovieBot Alexa Skill. While other simulators inject noise via stochastic processes at the intent
or slot level [18], our simulator is able to leverage audio signals and white noise to train policies in a
production-like environment. Using this simulator, we apply deep RL to learn dialog policies which
achieve the same success rate with fewer dialog turns compared to fixed rule-based policies in noisy
environments. Based on feedback from the participants in our user study, an interesting area of future
work is to use sentiment to further improve the user simulator. That is, user sentiment can be sampled
to generate polarized target utterances to simulate user frustration or satisfaction. Furthermore, we
plan to investigate other dimensions of dialog to improve our simulator and RL-based dialog policy
learning methods. These include per-interaction user satisfaction and efficiency of interactions to
model the reward function. Lastly, we plan to explore the impact of using different kinds of noise on
the learned policies.
Acknowledgments
We thank Jason Pazis and other members of the Conversational AI group at Amazon for helpful
discussions and comments. We also thank our anonymous reviewers for insightful feedback.
Appendix
A
Intent Generation Hyper-parameters
Number of Hidden Nodes per Layer
Intent and Slot Embedding Size
Number of Epochs
RNN Model
RNN Model + Bot Response
GRU Model
GRU Model + Bot Response
30
50
100
90
50
30
10
70
10
10
70
10
8
B
Example of simulated dialogs identified as human dialogs
Bigram Simulator Dialog
User
Bot
User
Bot
User
Bot
User
Bot
User
Bot
User
Bot
C
RNN Simulator Dialog
ummm comedy
I’d recommend Kingsman: The Golden Circle, a comedy action film starring Taron
Egerton
Look for something else.
I really liked Moana, a comedy adventure
film starring Auli’i Cravalho.
What’s my favorite genre?
Sci-fi, because it has been in 25 percent of
your genre-related requests.
Give me a blooper from the movie inception.
Here is an interesting inception goof. *gives
goof*
How does John Wick rate?
John Wick has a rating of 7.3 on IMDb.,
based on about 300000 votes.
Cancel
Thank you for using MovieBot.
User
Bot
ummm comedy
I think you might like Cars 3, a comedy adventure film, starring Owen Wilson.
User
Bot
How long is the movie?
The movie duration is 1 hour and 42 minutes.
User
Bot
Who wrote the soundtrack?
I don’t know how to handle that yet. You can
also say, can children watch this movie.
What is a popular romantic comedy movie?
I’d recommend Blind, a romance drama film
starring Demi Moore.
Tell me about it.
Here is a summary of blind from IMDb.
*gives plot summary*
That is enough.
Thank you for using MovieBot.
User
Bot
User
Bot
User
Bot
Learning parameters for DQN and Dueling DDQN for SNR of 2.8
Number of Hidden Layers
Number of Hidden Nodes per Layer
Embedding Size
Dropout
Learning Rate
Experience Replay Size
Window Size
Discount Factor
Target Model Update Interval
DQN
Dueling DDQN
2
32
5
0.5
0.0001
10,000
2
0.97
12,000
3
128
30
0.0
0.00001
15,000
8
0.97
8,000
References
[1] Amazon Polly. https://aws.amazon.com/polly
[2] Bergstra, J.; Yamins, D.; Cox, D. D. (2013). Hyperopt: A Python Library for Optimizing
the Hyperparameters of Machine Learning Algorithms, Computational Science & Discovery, 8(1).
https://github.com/hyperopt/hyperopt.
[3] Chandramohan, S.; Geist, M.; Lefevre, F.; Pietquin, O. (2011). "User Simulation in Dialogue Systems using
Inverse Reinforcement Learning,” in Proceedings of Interspeech.
[4] Cho, B. van Merrienboer, D. Bahdanau, and Y. Bengio. (2014). On the properties of neural machine
translation: Encoder-decoder approaches. In Proceedings of the Eighth Workshop on Syntax, Semantics and
Structure in Statistical Translation.
[5] Crook, P. A.; Marin, A. (2017). Sequence to Sequence Modeling for User Simulation in Dialog Systems.
In Proceedings of the 18th Annual Conference of the International Speech Communication Association
(INTERSPEECH 2017), 1706–1710.
[6] Cuayahuitl, H.; Renals, S.; Lemon, O.; Shimodaira, H. (2005). Human-computer dialogue simulation using
hidden Markov models. In Automatic Speech Recognition and Understanding, 2005 IEEE Workshop on (pp.
290-295).
[7] Cuayahuitl, H. (2016). SimpleDS: A Simple Deep Reinforcement Learning Dialogue System, in Proceedings
of International Workshop on Spoken Dialogue Systems (IWSDS).
9
[8] Eckert, W.; Levin, E.; Pieraccini, R. (1997). User modeling for spoken dialogue system evaluation. In
Automatic Speech Recognition and Understanding, 1997. Proceedings., 1997 IEEE Workshop on (pp. 80-87).
IEEE.
[9] El Asri, L.; He, J.; Suleman, K. (2016). A Sequence-to-Sequence Model for User Simulation in Spoken
Dialogue Systems, Interspeech, 1151–1155.
[10] Fatemi, M.; Asri, L. E.; Schulz, H.; He, J.; Suleman, K. (2016). Policy networks with two-stage training
for dialogue systems, in the Annual SIGdial Meeting on Discourse and Dialogue (SIGDIAL).
[11] Frampton, M.; Lemon, O. (2006). Learning more effective dialogue strategies using limited dialogue move
features. In Proceedings of the 21st International Conference on Computational Linguistics and the 44th
annual meeting of the Association for Computational Linguistics, Association for Computational Linguistics,
185–192.
[12] Georgila, K.; Henderson, J.; Lemon, O. (2006). User Simulation for Spoken Dialogue Systems: Learning
and Evaluation, in Proceedings of Interspeech.
[13] Georgila, K.; Traum, D. R. (2011). Reinforcement learning of argumentation dialogue policies in negotiation, in INTERSPEECH, 2073–2076.
[14] Kumar, A.; Gupta, A.; Chan, J.; Tucker, S.; Hoffmeister, B.; Dreyer, M. (2017). Just ASK: Building an
Architecture for Extensible Self-Service Spoken Language Understanding. arXiv:1711.00549 [cs.CL]
[15] Lee, S.; Eskenazi, M. (2012). Pomdp-based let’s go system for spoken dialog challenge, in Proceedings of
Spoken Language Technology Workshop (SLT), IEEE, 61–66.
[16] Levin, E.; Pieraccini, R.; Eckert, W. (2000). A Stochastic Model of Human-Machine Interaction for
Learning Dialog Strategies, IEEE Transactions on Speech and Audio Processing, 1, 11–23.
[17] Li, J.; Monroe, W.; Ritter, A.; Galley, M.; Gao, J.; Jurafsky, D. (2016). Deep Reinforcement Learning for
Dialogue Generation, in Proceedings of the 2016 Conference on Empirical Methods in Natural Language
Processing (EMNLP).
[18] Li, X.; Lipton, Z. C.; Dhingra, B.; Li, L.; Gao, J.; Chen, Y. N. (2016). A user simulator for task-completion
dialogues. arXiv preprint arXiv:1612.05688.
[19] Li, X.; Chen, Y. N.; Li, L.; Gao, J.; Celikyilmaz, A. (2017). End-to-End Task-Completion Neural Dialogue
Systems, in Proceedings of The 8th International Joint Conference on Natural Language Processing.
[20] Mahrt, T. (2016). PyAcoustics. https://github.com/timmahrt/pyAcoustics.
[21] Mikolov, T; Karafiat, M.; Burget, L.; Cernocky, J; Khudanpur, S. (2010). Recurrent neural network based
language model, in Proceedings of INTERSPEECH, vol. 2010, no. 9. International Speech Communication
Association, 1045–1048.
[22] Mikolov, T.; Sutskever, I.; Chen, K.; Corrado, G.; Dean, J. (2013). Distributed Representations of Words
and Phrases and their Compositionality, in Proceedings of NIPS.
[23] Mnih, V.; Kavukcuoglu, K.; Silver, D.; Graves, A.; Antonoglou, I.; Wierstra, D.; Riedmiller,M. (2013).
Playing atari with deep reinforcement learning, in Proceesinds of NIPS Deep Learning Workshop.
[24] Mnih, V.; Kavukcuoglu, D.; Silver, A. A.; Rusu, J.; Veness, M. G.; Bellemare, A.; Graves, M.; Riedmiller,
A. K.; Fidjeland, G.; Ostro-vski, S.; Petersen, C.; Beattie, A.; Sadik, I.; Antonoglou, H.; King, D.; Kumaran,
D.; Wierstra, D.; Legg, S.; Hassabis, D. (2015). Human-level control through deep reinforcement learning,
Nature, 518(7540), 529–533.
[25] MovieBot. https://www.amazon.com/Amazon-MovieBot/dp/B01MRKGF5W
[26] Singh, S.; Litman, D.; Kearns, M.; Walker, M. (2002). Optimizing dialogue management with reinforcement
learning: Experiments with the njfun system, Journal of Artificial Intelligence Research, 105–133.
[27] Schatzmann, J.; Weilhammer, K.; Stuttle, M.; Young, S. (2006). A survey of statistical user simulation
techniques for reinforcement-learning of dialogue management strategies. The knowledge engineering
review, 21(2), 97–126.
[28] Schatzmann, J.; Young, S. (2009). The hidden agenda user simulation model. IEEE transactions on audio,
speech, and language processing, 17(4), 733–747.
[29] Scheffler, K.; Young, S. (2002). Automatic learning of dialogue strategy using dialogue simulation
and reinforcement learning. In Proceedings of the second international conference on Human Language
Technology Research, Morgan Kaufmann Publishers Inc, 12–18.
[30] Serban, I. V.; Sordoni, A.; Bengio, Y.; Courville, A.; Pineau, J. (2016). Building End-To-End Dialogue
Systems Using Generative Hierarchical Neural Network Models, in Proceedings of the Thirtieth AAAI
Conference on Artificial Intelligence, 3776–3783.
[31] Sutskever, I.; Vinyals, O.; Le, Q. V. (2014). Sequence to sequence learning with neural networks. In
Advances in neural information processing systems, pp. 3104–3112.
10
[32] van Hasselt, H.; Guez, A.; Silver, D. (2015). Deep Reinforcement Learning with Double Q-Learning, in
Proceedings of AAAI 2016, arXiv:1509.06461.
[33] Wang, Z.; Schaul, T; Hessel, M.; van Hasselt, H.; Lanctot, M.; de Freitas, N. (2016). Dueling Network
Architectures for Deep Reinforcement Learning, in Proceedings of the 33rd International Conference on
Machine Learning, vol 48.
[34] Williams, J. D.; Young, S. (2007). Partially observable markov decision processes for spoken dialog
systems, Computer Speech & Language, 21(2), 393–422.
[35] Young, S.; Gasic, M.; Thomson, B.; and Williams, J. D. (2013). POMDP-based statistical spoken dialog
systems: A review, IEEE, 101(5), 1160–1179.
[36] Zhao, T.; Eskenazi, M. (2016). Towards end-to-end learning for dialog state tracking and management
using deep reinforcement learning, in Proceedings of the Meeting of the Special Interest Group on Discourse
and Dialogue (SIGDIAL).
11
| 2cs.AI
|
Causality in the Semantics of Esterel: Revisited
MohammadReza Mousavi
Department of Computer Science, Eindhoven University of Technology,
P.O. Box 513, NL-5600 MB Eindhoven, The Netherlands
We re-examine the challenges concerning causality in the semantics of Esterel and show that they pertain to the known issues in the semantics of Structured Operational Semantics with negative premises.
We show that the solutions offered for the semantics of SOS also provide answers to the semantic
challenges of Esterel and that they satisfy the intuitive requirements set by the language designers.
1
Introduction
Esterel [Ber99, PBEB07] is an imperative synchronous language used for the specification and programming of embedded systems. Esterel is based on the synchronous hypothesis, i.e., instantaneous reaction
to signals and immediate propagation of signals in each time-instant. The combination of the imperative
programming style and the synchronous hypothesis in Esterel has led to semantic challenges addressed
in the literature [BG92, Ber99, Tin00, Tin01, TdS05, PB02]. In this paper, we present the main semantic
challenge posed by Esterel, namely, the issue of causality. We show that it is reminiscent of the semantic
challenges [Gro93, BG96, Gla04] in Structured Operational Semantics [AFV01] (esp. in the setting with
negative premises; the same challenges were encountered before in logic programming [AB94]). We
then show that using the known solutions for the latter simplifies the presentation of the semantics of the
former substantially and leads to the desired intuitive properties set forth by the language designers.
The rest of this paper is organized as follows. In Section 2, we present a brief overview of the Esterel
language and its intuitive semantics. Section 3 introduces Structured Operational Semantics and notions
of semantics and well-definedness associated with SOS specifications. Section 4 connects these two
worlds by first presenting an SOS specification for Esterel and then studying the notions of semantics
and well-definedness for the given specification. There, we show that certain notions of semantics for
SOS formalize the intuitive criteria given by the language designers. Section 5 concludes the paper and
presents directions for future research.
2
Esterel and Its Semantics: A Cook’s Tour
The abstract syntax of Esterel is given by the grammar in Figure 1.
A short introduction to the intuitive semantics of each of these constructs follows. In this grammar,
0 stands for the terminated process. Emitting signal s is denoted by emit s, which is instantaneously
p, q ::= 0 | emit s | pres s ? p q end |
p ; q | p || q | sign s in p end |
1 | susp p when s | trap t in p end | exit t | loop p end
Figure 1: The Abstract Syntax of Esterel
B. Klin, P. Sobociński (Eds.):
6th Workshop on Structural Operational Semantics (SOS’09)
EPTCS 18, 2010, pp. 32–45, doi:10.4204/EPTCS.18.3
c M.R. Mousavi
This work is licensed under the Creative Commons
Attribution-Noncommercial-No Derivative Works License.
M.R. Mousavi
33
visible to all parts of the system (and may in turn cause more signals to be emitted). Reacting to present
and absent signals is done via the if-then-else construct pres s ? p q end, where if s is currently
present (emitted by some other part of the system), p is executed, otherwise if s is absent q is executed.
The combination of synchronous assumption, i.e., instantaneous propagation of signals, and checking
for absence/presence of signals leads to semantic complications, presented shortly. Parallel composition
of p and q is denoted by p || q. Process sign s in p end encapsulates s in p, i.e., declares s local to
p. Another way of reacting to signals is by using the suspend construct susp p when s, which initially
acts as p, but after one synchronous round will stop p as soon as signal s is emitted (suspension may
happen after a number of rounds). Process 1 stands for a process that passes one unit of time and then
terminates. One can define traps (exit points, exception handlers) to which a program can jump to by
trap t in p end. The actual jump (raising the exception) is performed by executing exit t. A program
can engage in a loop by means of loop p end (it can either keep on executing in the loop or exit the loop
using exit t).
An Esterel program is usually suffixed by a header declaring input and output signals. The syntax of
this header is of the form input i; output o; and we assume that the set of input and output variables in
a program is disjoint from the set of its local signals. Moreover, to unclutter the syntax, we assume fixed
sets ι, ω and λ , respectively, of input, output and local variables. We pick typical members i, i0 , i0 , . . . ∈ ι,
o, o0 , o0 , . . . ∈ ω and s, s0 , s0 ∈ λ . This way, one does not need to consider the input and out declaration
anymore since input and output (and local) variables are recognized by their names. In some cases output
and local variables can be treated uniformly, in which case we denote them by x, x0 , x0 ∈ ω ∪ λ .
To study the semantic challenge concerning causality it suffices for us to look at the first two rows of
our grammar. The other constructs, e.g., traps and time passing, are semantically interesting on their own
but are treated satisfactorily in the literature and are orthogonal to the causality problems addressed here.
Hence, in the remainder, we focus on the subset of Esterel given in the first two lines of our grammar and
only in passing mention how to include time and traps in our presented semantics.
A causality relation between events s and s0 (signals in this case), means that the presence and absence of s directly influences the presence or absence of s0 . For example, consider the following Esterel
program:
P0 pres i ? emit s 0 end ; pres s ? 0 emit o end
In the above program, there is a causality chain starting from the input variable i to the local variable s
and from s to the output variable o, namely the presence of i determines the presence of s and eventually
leads to the absence of o, while the absence of s (caused by the absence of i), determines the presence of
o. Using the syntax of Esterel one can easily write programs with cyclic dependencies (e.g., s is present
if and only if s is present) or even worse, cyclic dependencies of a paradoxical nature (e.g., s is present
if and only if s is absent). To illustrate these issues in Esterel, consider the following simple programs,
which are all due to [Ber99]. These programs are canonical examples of different issues concerning
causality in Esterel programs.
P1 pres s ? emit s 0 end
Program P1 relies on the presence of s in order to emit signal s. The logical semantics of Esterel
rejects this program on the ground that it has two “models”. The first one is by assuming that
s is present, which leads to a justification of this assumption by emitting s. The other one is by
assuming that s is absent, which is supported by that 0 does not emit (denies emitting) signal s.
In each synchronous round, the “model” of an Esterel program is defined by a global status, which
defines the status (presence/absence) of signals in this round. A global status of a program is called
34
Causality in the Semantics of Esterel
coherent when the presence/absence of signals are determined consistently by the emit statements
in the program [Ber99]:
The global status of a program is logically coherent iff at least one emit statement is
executed for each signal assumed present and no emit statement is executed for each
signal assumed absent.
For example, program P1 has two logically coherent global statuses, namely presence of s and absence of s, as motivated above. The basis for rejecting program P1 is called “logical determinism”
and is defined as follows [Ber99].
A program is logically deterministic if it has at most one logically coherent global
status.
P2 pres s ? 0 emit s end
Program P2 relies on the absence of s in order to emit signal s. According to the logical semantics
of Esterel, the above-given program has no logically coherent global status. Assuming that s is
absent leads to s being emitted and hence, incoherency. Likewise, assuming that s is present
requires emission of s, which is only justified when s is absent.
The basis for rejecting program P2 is called “logical reactivity” and is defined as follows [Ber99].
A program is logically reactive if it has at least one logically coherent global status.
The conjunction of logical determinism and logical reactivity is called logical coherency and is the
main well-definedness criterion for the logical semantics of Esterel.
P3 pres s ? emit s emit s end
The program above has only one logically coherent global status, namely that s is present. This
global status is also coherent since assuming the presence of s leads to emitting it and moreover, it
is not logically coherent to assume the absence of s, because it leads to its emission. Hence, as far
as logical coherency is concerned this program is accepted and the logical semantics defines the
semantics sketched above for this program.
However, the semantics of Esterel used for its compiler, called the constructive semantics [Ber99,
PBEB07], has further constraints which lead to the rejection of the above program. In this paper,
we consider the issue of causality in both variants of the semantics and hence, also study the issue
of constructiveness defined below.
A program is constructive, if for each signal, it either proves its presence (must emit
the signal) or proves its absence (cannot emit it).
Program P3 is rejected by the above criterion since it can neither prove the emission of s (its only
possible proof is cyclic since relies on the assumption that s is emitted), nor can it coherently prove
its absence, since to prove the absence of s it should prove that neither of the two emit statements
can be executed, thus it should prove that s can neither be present nor absent.
P4
pres s0 ? emit s0 0 end ||
pres s0 ? pres s1 ? 0 emit s1 end 0 end
Note that P4 is logically coherent, since its only logically coherent global status is that both s0 and
s1 are absent. To check its constructiveness, let us focus on the emission of s0 . It definitely does
M.R. Mousavi
35
not have to emit s0 , since the only reason for emitting s0 is the emit s0 statement in the left-hand
side of the parallel composition, which is guarded by the check on the presence of s0 . Hence, the
only proof for emitting s0 is cyclic. But it can potentially emit s0 (because it contains an emit s0
statement) and the only way to make sure that s0 cannot be emitted is to prove that the guard for
emit s0 never becomes true, i.e., we need again to show that s0 cannot be emitted, which is also a
cyclic reasoning. Hence, we conclude that P4 is not constructive because it neither must emit s0 ,
nor it can deny its emission.
P5 pres s0 ? emit s1 0 end ; emit s0
The above program is logically coherent and its unique logically coherent global status is that
both s0 and s1 are present. However, it is again rejected by the constructive semantics of Esterel.
The reason is that in order to reach the emit statement for s0 , we should first make sure that the
first statement has a well-defined semantics in this context, i.e., it either takes the if branch or the
else branch and then terminates. However, giving a constructive proof for the transition of the
conditional requires a constructive proof for the emission of s0 . This is another instance of the
cyclic proof phenomenon rejected by the constructive semantics.
3
Structured Operational Semantics
Structural Operational Semantics (SOS) was originally proposed by Plotkin [Plo04] as a syntax-directed
and compositional way of defining semantics. Gradually, SOS has gained popularity and by now has
become a de facto standard in defining operational semantics. This popularity has called for a richer
syntax for SOS deduction rules and thus, in some applications, SOS deduction rules lost their structural,
i.e., inductive, nature. Some authors then decided to use the same acronym for Structured Operational
Semantics [AFV01, GV92]. With the richer syntax of SOS rules, one can write deduction rules whose
meaning is not clear any more.
Example 1 Examples of cyclic rules are the deduction rules (r1) and (r2) given below.
s
s
p −→ p
(r1)
s
p −→ p
(r2)
p9
s
p −→ p
The reader may already note the curious similarity between program P1 and deduction rule (r1) on
one hand and program P2 and deduction rule (r2) on the other hand. Moreover, program P3 resembles
the combination of (r1) and (r2). These similarities materialize as formal definitions in the remainder
of this paper.
To formalize the syntax and semantics of SOS, we first formalize the concepts of formulae and
(transition) formulae.
Definition 2 (Signature and (sub)terms) We let V represent an infinite set of variables. A signature Σ
is a set of function symbols (operators), each with a fixed arity. An operator with arity zero is called a
constant. We define the set T(Σ) of terms over Σ as the smallest set satisfying the following constraints.
• A variable x ∈ V is a term.
• If f ∈ Σ has arity n and t1 , . . . ,tn are terms, then f (t1 , . . . ,tn ) is a term.
36
Causality in the Semantics of Esterel
We write t1 ≡ t2 if t1 and t2 are syntactically equal. The function vars : T(Σ) → 2V gives the set of
variables appearing in a term. The set C(Σ) ⊆ T(Σ) is the set of closed terms, i.e., terms that contain
no variables. A substitution σ is a function of type V → T(Σ). We extend the domain of substitutions to
terms homomorphically. If the range of a substitution lies in C(Σ), we say that it is a closing substitution.
A term s is considered a subterm of itself; if s is a subterm of ti , then s is also a subterm of
f (t0 , . . . ,ti , . . . ,tn−1 ), for each s,ti ∈ T(Σ), 0 ≤ i < n, and n-ary f ∈ Σ. The set of subterms of a term
t are denoted by subterms(t).
Next, we formalize the syntax of SOS in terms of Transition System Specifications.
Definition 3 (Transition System Specifications (TSS)) A transition system specification is a triplet
(Σ, L, D) where
• Σ is a signature.
l
l
• L is a set of labels. If l ∈ L, and t,t 0 ∈ T(Σ) we say that t −→t 0 is a positive formula and t 9 (also
l
l
denoted by ¬t −→ ) and t 9t 0 are negative formulae. A formula, typically denoted by φ , ψ, φ 0 , φi ,
. . . is either a negative formula or a positive one.
• D is a set of deduction rules, i.e., tuples of the form (Φ, φ ) where Φ is a set of formulae and φ
is a positive formula. We call the formulae contained in Φ the premises of the rule and φ the
conclusion.
We write vars(r) to denote the set of variables appearing in a deduction rule (r). We say a formula is
closed if all of its terms are closed. Substitutions are also extended to formulae and sets of formulae in
the natural way.
A deduction rule (Φ, φ ) is typically written as Φ
φ . For a deduction rule r, we write conc(r) to denote
its conclusion and prem(r) to denote its premises. A set of positive closed formulae is called a transition
relation. Given a transition relation T , L0 -labeled transitions of closed term p, denoted by T ↓ (p, L0 ) is
the subset of T containing all formulae in T that have p as their source and some l ∈ L0 as their label. A
TSS is supposed to define a transition relation but for the TSSs such as those given by deduction rules
(r0) and (r1), it is not clear what the associated transition relation is. Several proposals are given in the
literature, of which [Gla04] gives a comprehensive overview and comparison. In this paper, we shall use
some of these proposals to define the semantics of Esterel. In order to facilitate the presentation of these
proposals, we need two auxiliary definitions, namely contradiction and contingency, which are given
below.
l
l
l
Definition 4 (Contradiction and Consistency) Formula t −→t 0 is said to contradict both t 9 and t 9t 0 ,
and vice versa. Φ is consistent w.r.t. Ψ, denoted by Φ Ψ, when for each positive formula ψ ∈ Ψ, it
holds that ψ ∈ Φ and for each negative formula ψ ∈ Ψ, there is no φ ∈ Φ such that φ contradicts ψ.
l
In the remainder, we only use negative formulae of the form t 9 in our specifications. We now have
all the necessary ingredients to present different proposals for the semantics of TSSs. The first proposal
is the following notion of supported model, which is a slight modification of the definition in [Gla04]
(restricting it to particular sets of terms and labels).
Definition 5 (Supported Model) Given A TSS, a transition relation T is a supported model for a set
P ⊆ C(Σ) of closed terms and a set L0 ⊆ L of labels, when
Φ
l
1. for each q, q0 ∈ T(Σ) and l ∈ L if q −→ q0 ∈ T , then there exists a deduction rule
and a substiφ
l
tution σ such that σ (φ ) = q −→ q0 and T Φ, and
M.R. Mousavi
37
2. for each p ∈ P and l ∈ L0 , p0 ∈ T(Σ), if there exists a deduction rule
l
Φ
and a substitution σ such
φ
l
that σ (φ ) = p −→ p0 and T Φ, then p −→ p0 ∈ T .
A transition relation T is a supported model for a TSS when it is a supported model for C(Σ) and L.
Note that in the above definition and throughout the rest of the paper, we only consider the “immediate transitions” of p as its semantics. One can adapt the above definitions (and the subsequent ones) to
consider the “transition system” associated with p as its semantics. For the subset of Esterel considered
in this paper, these two notions lead to the same conclusion concerning the well-definedness and the
semantics of a program.
Semantics 1 (Unique Supported Model Semantics) Given a set P ∈ C(Σ) of closed terms and a set
L0 ⊆ L of labels a TSS is meaningful w.r.t. P and L0 when it has a unique supported model for P and
L0 ; the transition system associated with P and L0 is the unique supported model for P and L0 . A TSS
is meaningful when it has a unique supported model; the transition relation associated with a TSS is its
unique supported model.
To illustrate these concepts, we give a few simple TSSs and study their supported models.
Example 6 Consider the deduction rules given in Example 1.
Consider the TSS comprising only deduction rule (r1). This TSS is not meaningful (w.r.t. {p} and
s
{s}) according to Semantics 1 because it has two supported models, namely 0/ and {p −→ p}.
Also according to Semantics 1, the TSS comprising only deduction rule (r2) is not meaningful (w.r.t.
{p} and {s}) either, because it has no supported model. Particularly, T = 0/ is not a supported model
s
s
because it follows from the right-to-left implication of Definition 5 that p −→ p ∈ T . T = {p −→ p} is
s
not a supported model either since the only deduction rule providing a reason for p −→ p ∈ T is (r2) but
s
it does not hold that T p0 9 .
The TSS comprising both (r1) and (r2) is indeed meaningful and its associated transition relation is
s
T = {p −→ p}. Transition relation T is indeed a supported model since (r1) now provides a reason for
s
p −→ p ∈ T . Moreover T 0 = 0/ is not a supported model for this TSS because it then follows from (r2)
s
and the right-to-left implication of Semantics 1 that p −→ p ∈ T 0 .
If one takes the transition system of a program as a formalization of its global state, then the TSS
comprising of deduction rule (r1) is rejected because it has no coherent global state and the TSS with
only (r2) is rejected because it does equivocally define a coherent global state.
This suggests that Semantics 1 provides a suitable formalization for logical coherency. Next, we give
a formalization of constructiveness in terms of supported proofs and denials.
Definition 7 (Supported Proofs) A TSS T provides a supported proof for a formula φ , denoted by
T `s φ , when there is a well-founded upwardly branching tree with formulae as nodes and of which
• the root is labelled by φ ;
• if a node is labelled by a positive formula ψ and the nodes above it form the set K then
instance of a deduction rule in T .
K
ψ
is an
• if a node is labelled by a negative formula ψ, and the nodes above it form the set K, then for each
instance of a deduction rule ψKii in T such that ψi contradicts ψ, there exists a formula ψi0 ∈ K
contradicting a formula in Ki .
38
Causality in the Semantics of Esterel
Semantics 2 (S-Complete Semantics) A TSS is s-complete for a set of closed terms P when for each
formula φ with a p ∈ P as its source, either φ or a formula contradicting it has a supported proof. A TSS
is s-complete when it is s-complete for the set C(Σ) of all closed terms and its transition relation is the
set of positive formulae, for which it provides supported proofs.
The following theorem is taken from [Gla04], which shows that constructiveness is indeed stronger
than logical coherency.
Theorem 8 A program (TSS) is s-complete only if it is meaningful according to Semantics 1 (has a
unique supported model) and its associated transition system (unique supported model) coincides with
the set of all positive formulae with a supported proof.
Another useful property of supported proofs is their consistency [Gla04], stated below.
Theorem 9 The notion of supported proof is consistent, i.e., for each formula φ with a supported proof,
its negation does not have a supported proof.
Next, we re-examine the TSS of Example 6 using our new notion of semantics.
Example 10 The two TSSs comprising only (r1) and only (r2) are both rejected by Semantics 2, as
s
s
well, since neither p −→ p, nor p 9 can be proven from either of them. (This is also an immediate
consequence of Theorem 8.)
s
In the case of the TSS comprising only (r1), any attempt to build a supported proof for p −→ p has
s
the same formula as its premise. Moreover, p 9 cannot be proven because its proof tree should prove a
s
s
s
negation of a premise of (r1), i.e., again p 9 . In other words, both p −→ p and p 9 only have cyclic,
and thus unsupported, proofs.
s
s
Similarly, in the case of the TSS comprising only (r2), neither p −→ p, nor p 9 have a supported
proof.
Consider the TSS comprising both (r1) and (r2); it does have a unique supported model T =
s
s
{p −→ p} but it is not s-complete and is thus rejected by Semantics 2. Any proof for p −→ p or its
negation leads to a cycle, i.e., repeating the node below in the node above, and are thus not supported.
Again drawing an analogy with Esterel programs, Semantics 2 requires the existence of a “constructive” (supported) proof for presence/absence of signals and thus rejects a program which uses both
possibilities for a signal in order to establish its own presence.
4
Structured Operational Semantics for Esterel
Our semantic specification of Esterel is presented in Figures 2 and 3. The state of the SOS comprises the
syntax of the program currently being executed (defined by the grammar in Figure 1). The semantics is
supposed to define two predicate, pXI,c , p ↑I,c,s , respectively, where the former means that p terminates
with input evaluation I and under context c (if p is part of program c), and the latter means that p emits
signal s (in the present time-instant) under the same assumptions. (A predicate formula can be formally
interpreted as a transition formula with a dummy right-hand-side; in our case one can take 0 to be the
dummy target of all predicate formulae, i.e., read p ↑I,c,s and pXI,c as p ↑I,c,s 0 and pXI,c 0, respectively.)
In addition to the two predicates, the semantics is supposed to define a transition relation of the form
I,c,s
p −→ p0 , which denotes that program p emits signal s under input evaluation I and context c. Next, we
briefly describe the deduction rules in Figures 2 and 3 and then show how they formalize the intuitive
properties of Esterel programs discussed before. In all labels (of predicates and transitions) of Figures 2
M.R. Mousavi
(e0)
(p0)
39
(s0)
emit x ↑I,c,x
p ↑I,c,x
p || q ↑I,c,x
(p1)
(f2)
p ↑I,c,x
(s1)
p ; q ↑I,c,x
q ↑I,c,x
p || q ↑I,c,x
i+ ∈ I
(f0)
pXI,c
q ↑I,c,x
p ; q ↑I,c,x
c ↑I,c,s
I,c,x0
(s2)
p ↑I,c,x
pres s ? p q end ↑I,c,x
p ↑I,c,x
(f3)
pres i ? p q end ↑I,c,x
i− ∈ I
p −→ p0
(f1)
p0 XI,c
q ↑I,c,x
p ; q ↑I,c,x
¬c ↑I,c,s
q ↑I,c,x
pres s ? p q end ↑I,c,x
q ↑I,c,x
pres i ? p q end ↑I,c,x
0
(en0)
p[s00 /s] ↑I,c,s
0
sign s in p end ↑I,c,s [s/s
00 ]
s00 fresh in p and r
Figure 2: Structured Operational Semantics for Esterel (Part I: Signal Emission)
and 3, I ⊆ {i+ , i− | i ∈ ι} such that for each i ∈ ι, either i− ∈ I or i+ ∈ ι (but not both), c ∈ C(Σ), i ∈ ι,
x ∈ ω ∪ λ and s, s0 , s00 ∈ λ .
In Figure 2, (e0) states that emit s can emit signal s under any arbitrary input evaluation and context.
Deduction rules (s0), (s1) and (s2) describe when a sequential composition emits a signal, namely, when
either the first component of the composition emits it, or when the first component terminates (possibly
after a transition) and the second component emits the signal.
The notions of termination and transition are defined in Figure 3. A parallel composition emits a
signal if one of its components emits the signal, which is captured by deduction rules (p0) and (p1).
An if-then-else constructs emits a signal, if either, according to deduction rule (f0), the local signal in
its condition is emitted and the if-branch emits the signal or, according to deduction rule (f1), the local
signal in the condition cannot be emitted and the else-branch is taken. Deduction rules (f2) and (f3) take
care of the case where the condition is an input signal. In such cases, the condition is checked against
the given input evaluation. A program p with a local signal s can emit a signal s0 , if p with a fresh signal
s00 substituted for s can emit s0 (but if s0 is s, then p should be able to emit s00 ).
In Figure 3, the concept of termination is defined through the predicate XI,c , in a straightforward
manner. Exceptions are deduction rules (if4) and (if5), which rely on (the impossibility of) the emission
of the condition signal for proving termination. In Figure 3, the deduction rules specifying a transition
relation are almost identical to their counterparts in Figure 2. The most notable exceptions are deduction
rules (seq0) to (seq4) and (par0) to (par3), which should consider all possible combinations of simultaneous transitions and individual transitions with (non-)termination in order to record the right target for
the transition.
One advantage of our approach to the semantics of Esterel presented in [Ber99, Tin00, Tin01] is that
we can capture both the logical semantics and constructive semantics of Esterel using the same TSS (by
using two generic notions of semantics for TSS already known in the literature). Another advantage is
that it establishes a clear link between, respectively, the logical and the constructive approaches to Esterel
semantics, on the one hand and the model- and proof-theoretic semantics of TSSs on the other hand.
Definition 11 (Logical Semantics of Esterel) An Esterel program p is logically coherent if the above
given TSS is meaningful according to Semantics 1 for subterms(p) and (predicates and) transitions
I,p,x
labeled {XI,p , ↑I,p,x , −→ }. The semantics of p is the set of above-mentioned predicates and transitions
40
Causality in the Semantics of Esterel
(nil)
I,c,x
(seq0)
(seq2)
p −→ p0
I,c,x0
p0 XI,c
I,c,x
q −→ q0
(seq1)
I,c,x
I,c,x
I,c,x
q −→ q0
(seq3)
I,c,x
p ; q −→ q0
pXI,c
(if0)
p −→ p0
p −→ p0
I,c,x
I,c,x
I,c,x
(par3)
I,c,x
p || q −→ q0
c ↑I,c,s
p −→ p0
(if2)
I,c,x
I,c,x
c ↑I,c,s
pXI,c
pres s ? p q endXI,c
(if6)
i− ∈ I
qXI,c
p || q −→ p0
pXI,c
pres i ? p q endXI,c
(if3)
(if5)
(if7)
qXI,c
p ; qXI,c
I,c,x0
q −→ q0
(par4)
pXI,c
qXI,c
p || qXI,c
I,c,x
q −→ q0
¬c ↑I,c,s
I,c,x
pres s ? p q end −→ q0
i− ∈ I
I,c,x
q −→ q0
I,c,x
pres i ? p q end −→ q0
¬c ↑I,c,s
qXI,c
pres s ? p q endXI,c
i+ ∈ I
qXI,c
pres i ? p q endXI,c
c,s0
(enc0)
pXI,c
I,c,x0
I,c,x
(if1)
(seq4)
p || q −→ p0 || q0
I,c,x
p −→ p0
pres i ? p q end −→ p0
(if4)
p −→ p0
I,c,x
p −→ p0
pres s ? p q end −→ p0
i+ ∈ I
qXI,c
I,c,x
(par1)
q −→ q0
I,c,x0
p ; q −→ p0
q −→ q0
I,c,x0
p0 XI,c
p ; q −→ q0
I,c,x
p || q −→ p0 || q0
q −→ q0
p −→ p0
p0 XI,c
I,c,x0
I,c,x
(par0)
I,c,x
emit x −→ 0
p ; q −→ q0
pXI,c
(par2)
(em)
0XI,c
p[s00 /s] −→ p0
c,s0 [s/s00 ]
sign s in p end −→ sign s in p0 [s/s00 ] end
(enc1)
p[s00 /s]Xc
sign s in p endXI,c
s00 fresh in p and r
Figure 3: Structured Operational Semantics for Esterel (Part II: Transition and Termination)
M.R. Mousavi
41
associated with subterms(p).
Next, we show that Definition 11 indeed satisfies the intuition behind logical coherency by reexamining the examples introduced in Section 2.
Example 12 Consider program P0, recalled below.
P0 pres i ? emit s 0 end ; pres s ? 0 emit o end
It is straightforward to check that the following is the semantics of P0:
{P0 ↑{i
+ },P0,s
, P0 ↑{i
− },P0,o
{i+ },P0,i
{i− },P0,o
{i− },P0,s
{i+ },P0,s
{i+ },P0,s
{i− },P0,s
, P0 −→ 0, P0 −→ 0, emit s −→ 0, emit s −→ 0,
emit o −→ 0, emit o −→ 0, 0X{i+ },P0 , 0X{i− },P0 }.
Consider program P1 quoted below.
1
P1 pres s ? emit s 0 end
0,P1,s
/
0,P1,s
/
/
/
It has two supported models, namely {P1 ↑0,P1,s
, P1 −→ 0, emit s ↑0,P1,s
, emit s −→ 0, 0X0,P1
}
/
0,P1,s
/
/
, 0X0,P1
}. Hence, P1 is not meaningful according to Semantics 1.
and {emit s −→ 0, emit s ↑0,P1,s
/
Consider program P2 recalled below.
P2 pres s ? 0 emit s end
/
Program P2 does not have any supported model: Assume, towards a contradiction, that P2 ↑0,P2,s
is in the purported supported model of T . It then follows from item 1 in Definition 5 that there exists a
/
deduction rule whose conclusion can match P2 ↑0,P2,s
and whose premises are consistent with T . The
only candidates are (f0) and (f1); we analyze both cases below and show that they both lead to a
contradiction.
/
/
(f0) The premises of the instance of (f0) are P2 ↑0,P2,s
and 0 ↑0,P2,s
. It follows from item 1 of Definition
5 that both predicates should be in T and hence, item 1 again applies to both predicates and in
/
particular to 0 ↑0,P2,s
. Hence, there should exist a deduction rule whose conclusions matches with
the above predicate. A simple syntactic check on the deduction rules of Figures 2 and 3 reveals
that none of the conclusions can be unified with the above predicate and hence a contradiction
follows.
/
/
(f1) The premises of the instance of (f0) are ¬P2 ↑0,P2,s
and emit s ↑0,P2,s
, both of which should be in
0,P2,s
/
T . Again item 1 of Definition 5 applies and thus, ¬P2 ↑
should be consistent with T , or in
/
other words, P2 ↑0,P2,s
∈
/ T , which contradicts our initial assumption.
The next program to consider is P3, quoted below.
P3 pres s ? emit s emit s end
Program P3 is indeed meaningful and has the following unique supported model.
0,P3,s
/
0,P3,s
/
/
/
{P3 ↑0,P3,s
, P3 −→ 0, emit s ↑0,P3,s
, emit s −→ 0, 0X0,P3
}.
/
0,P3,s
/
Note that P3 ↑
(and/or the transition of P3) cannot be removed from the supported model; to see
/
this, it follows from item 2 of Definition 5 and deduction rule (e0) that emit s ↑0,P3,s
∈ T , and following
0,P3,s
/
the same reasoning and deduction rule (f1), we have that P3 ↑
∈ T.
Program P4 is considered logically coherent but not constructive by the language designers. Next,
we show that this intuition is indeed supported by our formal definitions.
1 For each program, we choose ι, ω and λ , respectively, to comprise only the input, output and local variables mentioned in
the program at hand. This allows us to focus only on the possibly relevant part of I when considering supported models.
42
Causality in the Semantics of Esterel
P4
pres s0 ? emit s0 0 end ||
pres s0 ? pres s1 ? 0 emit s1 end 0 end
Program P4 has a unique supported model, given below.
0,P3,s
/
1
0,P3,s
/
0
/
/
1 , emit s −→ 0, 0X
0 , emit s −→ 0, emit s ↑0,P3,s
}.
{emit s0 ↑0,P3,s
0,P3
/
1
1
0
Note that neither emission of s0 , nor s1 cannot be present in a supported model. First, concerning s1 ,
suppose that s1 can be emitted, then it follows from item 1 of Definition 5 that there should be a deduction
rule supporting this emission. This can only be due to (p1) and thus, the right-hand-side component of
the parallel composition. This component, in turn can only emit s1 (due to deduction rules (f0) and then
(f1)) if s0 is present and s1 is absent under the same context. The latter contradicts our assumption.
Similarly, suppose that the supported model contains a predicate (or transition) to the effect that s0 can
be emitted. We already know that no predicate for emitting s1 can be in the supported model. Hence,
it follows from successive application of item 2 of Definition 5 using deduction rules (f0), (f1) and (e0)
that s1 can be emitted under the same context, which is already shown to lead to contradiction.
P5 pres s0 ? emit s1 0 end ; emit s0
Program P5 is also meaningful and has a unique supported model, given below.
0,P5,s
/
0
0,P5,s
/
1
/
/
0 , P5 ↑0,P5,s
1 , P5 −→ 0, P5 −→ 0,
{P5 ↑0,P5,s
0,P5,s
/
1
/
1 , pres s ? emit s 0 end −→ 0,
pres s0 ? emit s1 0 end ↑0,P5,s
0
1
0,P3,s
/
0
0,P3,s
/
1
/
/
0 , emit s −→ 0, emit s ↑0,P5,s
1 , emit s −→ 0, 0X
emit s0 ↑0,P5,s
}.
0
1
1
0,P5
/
Note that none of the predicates or transitions concerning the emission of s0 and s1 can be omitted
from the supported model. If the predicate (transition) concerning the emission of s0 is omitted then the
first component of sequential composition terminates and hence s0 should be emitted due to the second
component. Since s0 should always be emitted, the emission of s1 is guaranteed by the first component
of sequential composition.
Definition 13 (Constructive Semantics of Esterel) An Esterel program p is constructive if for each sigI,p,s
I,c,s
nal s and each input evaluation I either p ↑I,p,s and p −→ p0 (for some p0 ) or ¬p ↑I,p,s and p 9 has a
supported proof and moreover, either pXI,p or ¬pXI,p has a supported proof.
To illustrate this semantics and identify its differences with the logical semantics, we reconsider those
programs whom are considered non-constructive but logically coherent in Section 2.
Example 14 Consider program P3. This program is both intuitively and formally shown to be logically coherent. Moreover, in Section 2, we introduced this program as a canonical example of a non/
constructive program. Next, we show that it is also formally non-constructive since neither P3 ↑0,P3,s
0,P3,s
/
/
nor ¬P3 ↑0,P3,s
have a supported proof (a similar reasoning shows that neither P3 −→ p0 for any p0
0,P3,s
/
/
nor P3 9 have a supported proof). Suppose P3 ↑0,P3,s
has a supported proof, then its proof is either
due to (f0) or (f1). In the former case, the nodes placed above our proof obligation are P3 ↑I,P3,s and
/
emit s ↑0,P3,s
. While the latter has a supported proof (due to (e0)), the former was our original proof
obligation, thus, it only remains to check the alternative option due to (f1). The premises of (f1) are
/
then ¬P3 ↑I,P3,s and emit s ↑0,P3,s
. Again the latter formula has a supported proof but the former is the
negation of our proof obligation and thanks to Theorem 9, we know that if ¬P3 ↑I,P3,s has a supported
proof then P3 ↑I,P3,s cannot have a supported proof. Similarly, if ¬P3 ↑I,P3,s has a supported proof, then
a negation of a premise of all deduction rules that can match P3 ↑I,P3,s must have a supported proof.
M.R. Mousavi
43
These two rules are again (f0) and (f1). The negation of the common premise of these two rules, i.e.,
/
emit s ↑0,P3,s
cannot have a supported proof (following Theorem 9, because the premise itself has a supported proof). Hence a negation of both P3 ↑I,P3,s and ¬P3 ↑I,P3,s should have supported proofs, which
is again impossible due to Theorem 9.
/
0 , nor its negation have a supported proof.
Program P4 is not constructive since neither P4 ↑0,P4,s
The only possible proof for the emission predicate can be due to (p0) or (p1). The case for (p1) does
not lead to a supported proof since the right-hand-side does not contain any emit statement for s0 . If
/
0 which was to be proven. The
the supported proof is due to (p0), then it should hold that P4 ↑0,P4,s
0,P4,s
/
0
does not have a supported proof, either. Since then a negation
negation of the predicate, i.e., P4 ↑
of a premise of (p0) and (p1) should have a supported proof. The negation of the only premise of (p0) is
/
0 , which in turn means that a negation of a premise of (f0) or (f1) must
pres s0 ? emit s0 0 end ↑0,P4,s
/
0 , but we were seeking a proof of
have a supported proof. Consider (f0), its two premises are P4 ↑0,P4,s
0,P4,s
/
0
, whose negation cannot be proven.
its negation and emit s0 ↑
/
0 nor its negation are
Program P5 is not constructive, either. We next show that neither P5 ↑0,P5,s
0,P5,s
/
0 is due to one of the rules (s0) to (s2).
provable. The purported supported proof for predicate P5 ↑
Next, we analyze each case and show that it leads to a contradiction.
/
0 . This, in turn, can be either due to
(s0) Then, it should hold that pres s0 ? emit s1 0 end ↑0,P5,s
/
0,
(f0) or (f1). If the predicate is due to (f0), then we should have a supported proof for P5 ↑0,P5,s
0,P5,s
/
0
which was to be proven. If the proof is due to (f1), then ¬P5 ↑
should have a supported proof,
which is impossible due to Theorem 9.
(s1) Then, it should hold that pres s0 ? emit s1 0 endX0,P5
. This termination can be due to ei/
/
0 or
ther (if4) or (if5). None of these two are possible since otherwise, respectively, P5 ↑0,P5,s
0,P5,s
/
0
¬P5 ↑
should have a supported proof.
0
0
0
0 p for some s and p . This transition
(s2) Then, it should hold that pres s0 ? emit s1 0 endX0,P5,s
/
is due to either (if0) or (if1). Again, both cases lead to a contradiction due to a similar reasoning
as in item (s0).
As a side note, the common intuition and the similarities between deduction rules of Figures 2 and 3
may suggest that we can replace deduction rules of Figure 2 with the following rule (or even do without
the emission predicates and make the same changes in the deduction rule for if-then-else statements in
Figure 3):
I,c,x
p −→ p0
(emit)
p ↑I,c,x
This change leads to a much more restrictive semantics, which is unable to provide supported proofs
for transitions of perfectly acceptable programs such as the following:
P6 pres s ? emit o 0 end || emit s
To see this, the reader may try to prove that P6 can emit signal o using deduction rule (par0). The
proof of the premise of (par0) then should rely on (if0) and hence due to deduction rule (emit), we need
to prove that s can be emitted (for the if-then-else to be able to take a transition). In turn, this can only
be due to (par1). But to apply (par1), we need to know that the left-hand-side component can take a
transition (in order to record its target), which is what we wanted to prove initially. This cycle is broken
in our semantics, by deduction rule (p1) which only considers one of the two components to infer the
44
Causality in the Semantics of Esterel
emission of s0 (without trying to record the target of the transition). The following proof illustrates why
this program is indeed constructive.
/
emit s ↑0,P6,s
/
P6 ↑0,P6,s
0,P6,o
/
emit o −→ 0
0,P6,o
/
pres s ? emit o 0 end −→ 0
0,P6,s
/
emit s −→ 0
0,P6,o
/
P6 −→ 0 || 0
5
Conclusions and Future Work
In this paper, we presented a link between the intuitive notions of logical coherency and constructiveness
in the semantics of Esterel on the one hand, and the formal notions of supported models and supported
proofs in the semantics of Structured Operational Semantics, on the other hand. By means of several
canonical examples from the literature, we showed that our formal definitions indeed capture the intuitive
criteria put forward by the language designers.
Several formalizations of these two intuitive criteria exist in the literature. For example [Ber99,
PBEB07] present three formalizations of constructive semantics of Esterel. In [Tin00, Tin01] another
formalization of constructive semantics of Esterel is presented and is proven to coincide with one of the
notions in [Ber99]. A rigorous comparison between all these notions and the ones presented in this paper
remains as a topic for future research.
In the semantics presented in this paper, we abstracted from the issues of exceptions (traps), loops
and time. We expect that one can include these aspects without any substantial change in the semantics
presented in this paper using the modular semantics approach of [Mos04, MN08]. This remains as
another interesting exercise for the future.
Acknowledgements. Inspiring discussions with Jean-Pierre Talpin and Paul Guernic are gratefully
acknowledged. The author would like to thank the anonymous reviewers of SOS 2009 for their insightful
reviews.
References
[AB94]
Krzysztof R. Apt and Roland N. Bol. Logic programming and negation: A survey. Journal of Logic
Programming (JLAP), 19/20:9–71, 1994.
[AFV01]
Luca Aceto, Willem Jan (Wan) Fokkink, and Chris Verhoef. Structural operational semantics. In Jan A.
Bergstra, Alban Ponse, and Scott A. Smolka, editors, Handbook of Process Algebra, Chapter 3, pages
197–292. Elsevier Science, Dordrecht, The Netherlands, 2001.
[Ber99]
Gérard Berry. The Constructive Semantics of Pure Esterel. 1999. Draft version, available from:
ftp://ftp-sop.inria.fr/meije/esterel/papers/constructiveness3.ps.gz.
[BG92]
Gérard Berry and Georges Gonthier. The Esterel synchronous programming language: Design, semantics, implementation. Science of Computer Programming (SCP), 19(2):87–152, 1992.
[BG96]
Roland N. Bol and Jan Friso Groote. The meaning of negative premises in transition system specifications. Journal of the ACM (JACM), 43(5):863–914, September 1996.
[Gla04]
Robert Jan (Rob) van Glabbeek. The meaning of negative premises in transition system specifications
II. Journal of Logic and Algebraic Programming (JLAP), 60-61:229–258, 2004.
M.R. Mousavi
45
[Gro93]
Jan Friso Groote. Transition system specifications with negative premises. Theoretical Computer
Science (TCS), 118(2):263–299, 1993.
[GV92]
Jan Friso Groote and Frits W. Vaandrager. Structured operational semantics and bisimulation as a
congruence. Information and Computation (I&C), 100(2):202–260, October 1992.
[MN08]
Peter D. Mosses and Mark J. New. Implicit propagation in structural operational semantics. In Proceedings of the 5th Workshop on Structural Operational Semantics (SOS’08), pages 78–92, 2008.
[Mos04]
Peter D. Mosses. Modular structural operational semantics. Journal of Logic and Algebraic Programming (JLAP), 60-61:195–228, 2004.
[PB02]
Dumitru Potop-Butucaru. Optimizations for Faster Simulation of Esterel Programs. PhD thesis, École
des Mines de Paris, CMA, Paris, France, 2002.
[PBEB07] Dumitru Potop-Butucaru, Stephen A. Edwards, and Gérard Berry. Compiling Esterel. Springer-Verlag,
2007.
[Plo04]
Gordon D. Plotkin. The origins of structural operational semantics. Journal of Logic and Algebraic
Programming (JLAP), 60:3–15, 2004.
[TdS05]
Olivier Tardieu and Robert de Simone. Loops in Esterel. ACM Transactions on Embedded Computing
Systems (ACM TECS), 4:708–750, 2005.
[Tin00]
Simone Tini. Structural Operational Semantics for Synchronous Languages. PhD thesis, Dipartimento
di Informatica, Università degli Studi di Pisa, Pisa, Italy, 2000.
[Tin01]
Simone Tini. An axiomatic semantics for Esterel. Theoretical Computer Science (TCS), 269(1-2):231–
282, 2001.
| 6cs.PL
|
Enumerating submodules invariant under an
endomorphism
Tobias Rossmann
arXiv:1606.00760v1 [math.NT] 1 Jun 2016
Fakultät für Mathematik, Universität Bielefeld, D-33501 Bielefeld, Germany
We study zeta functions enumerating submodules invariant under a given endomorphism of a finitely generated module over the ring of (S-)integers of a number
field. In particular, we compute explicit formulae involving Dedekind zeta functions
and establish meromorphic continuation of these zeta functions to the complex plane.
As an application, we show that ideal zeta functions associated with nilpotent Lie
algebras of maximal class have abscissa of convergence 2.
1 Introduction
Zeta functions derived from endomorphisms. Throughout, rings are assumed to be
commutative and unital. We say that a ring R has polynomial submodule growth if
the following holds for every finitely generated R-module M : for each m > 1, the number
of submodules of additive index m of M is finite and polynomially bounded as a function
of m. Recall that R is semi-local if it contains only finitely many maximal ideals.
Theorem 1.1 ([20, Thm 1]). Let R be a ring which is finitely generated over Z or
semi-local with finite residue fields. Then R has polynomial submodule growth if and only
if it has Krull dimension at most 2.
Let R be a ring with polynomial submodule growth, let M be a finitely generated
R-module, and let A ∈ EndR (M ). For m > 1, let am (A, R) denote the number of
A-invariant R-submodules U 6 M with |M : U | = m. We define a zeta function
ζA,R (s) :=
∞
X
am (A, R)m−s
m=1
and we let αA,R < ∞ denote its abscissa of convergence; it is well-known that αA,R is
precisely the degree of polynomial growth of the partial sums a1 (A, R) + · · · + am (A, R)
as a function of m.
2010 Mathematics Subject Classification. 11M41, 15A04, 15A21, 17B30
Keywords. Zeta functions, invariant submodules, nilpotent matrices, ideal growth, submodule growth.
This work is supported by the DFG Priority Programme “Algorithmic and Experimental Methods in
Algebra, Geometry and Number Theory” (SPP 1489).
The zeta functions ζA,R (s) belong to the larger theory of subobject zeta functions; for
a recent survey of the area, see [25]. Indeed, using the terminology from [17], ζA,R (s) is
∞
P
the submodule zeta function ζR[A]yM (s) of the enveloping algebra R[A] :=
R · Ai ⊂
i=0
EndR (M ) of A acting on M .
The main results of this article, Theorems A–D, constitute a rather exhaustive analysis
of the zeta functions ζA,R (s) in the cases that R is the ring of (S-)integers of a number
field or a (generic) completion of such a ring. In particular, our findings provide further
evidence in support of the author’s general conjectures on submodule zeta functions
stated in [17, §8].
Related work: invariant subspaces. The study of subspaces invariant under an endomorphism has a long history. For a finite-dimensional vector space V over the real or
complex numbers and A ∈ End(V ), Shayman [21] investigated topological properties
of the compact analytic space SA of A-invariant subspaces of V . In particular, if A is
nilpotent, then he found the subspace SA (d) ⊂ SA of d-dimensional A-invariant subspaces
of V to be connected but usually singular.
For an arbitrary ground field F and a fixed number n, Ringel and Schmidmeier [16]
studied the category of triples (V, U, T ), where V is a finite-dimensional vector space
over F , T ∈ EndF (V ) satisfies T n = 0, and U 6 V is F -invariant. While their point of
view is rather different from ours, we would like to point out that they found the case of
exponent n > 7 to involve instances of so-called “wild” representation type.
Ideal zeta functions. In our study of the zeta functions ζA,R (s), we will frequently
encounter another special case of submodule zeta functions, namely ideal zeta functions.
Let R be a ring with polynomial submodule growth and let A be a possibly non-associative
R-algebra whose underlying R-module is finitely generated. We write I /R A to indicate
that I is a two-sided ideal of A which is also an R-submodule. The ideal zeta function
(cf. [11]) of A is
X
ζA (s) :=
|A : I|−s .
I/R A
|A:I|<∞
For example, the ideal zeta function of the ring of integers of a number field k is precisely
the Dedekind zeta function of k. In particular, the ideal zeta function of Z is the Riemann
zeta function ζ(s). As explained in [17, Rem. 2.2(ii)], ideal zeta functions are in fact a
special case of the submodule zeta functions discussed below.
Global setup, Euler products, and growth rates. For the remainder of this article,
let k be a number field with ring of integers o.
Let Vk denote the set of non-Archimedean places of k. For v ∈ Vk , let kv be the v-adic
completion of k and let ov be its valuation ring. For S ⊂ Vk , let
\
oS =
ov ∩ k
v∈Vk \S
2
be the usual ring of S-integers of k.
In the following, we investigate ζA,R (s), where A ∈ EndR (M ) and R = ov or R = oS for
v ∈ Vk or a finite set S ⊂ Vk , respectively. The techniques that we use are predominantly
local and valid for almost all places of k (i.e. for all but finitely many places); the exclusion
of a finite number of exceptional places is common and frequently unavoidable in the
theory of subobject zeta functions.
If M is a finitely generate oS -module, then M ⊗oS ov is a free ov -module for almost all
v ∈ Vk \ S. We thus lose little by henceforth assuming that M = onS and A ∈ Mn (oS ),
where Mn (R) denotes the algebra of n × n matrices over a ring R. Note that if A ∈ Mn (k),
then A ∈ Mn (ov ) for almost all v ∈ Vk . In order to exclude trivialities, unless otherwise
stated, we always assume that n > 0. Being instances of submodule zeta functions, the
zeta functions ζA,oS (s) admit natural Euler product factorisations.
Proposition (Cf. [17, Lemma 2.3]). Let A ∈ Mn (oS ) for finite S ⊂ Vk . Then
Y
ζA,oS (s) =
ζA,ov (s).
v∈Vk \S
The following is a consequence of deep results of du Sautoy and Grunewald on subobject
zeta functions expressible in terms of what they call “cone integrals”.
Theorem 1.2 (Cf. [7, §4]). Let A ∈ Mn (oS ) for finite S ⊂ Vk . Then:
(i) The abscissa of convergence αA,oS of ζA,oS (s) is a rational number.
(ii) ζA,oS (s) admits meromorphic continuation to {s ∈ C : Re(s) > αA,oS − δ} for some
δ > 0. This continued function is regular on the line Re(s) = αA,oS except for a
pole at s = αA,oS .
(iii) Let βA,oS denote the multiplicity of the pole of (the meromorphic continuation of )
ζA,oS (s) at αA,oS . Then there exists a real constant cA,oS > 0 such that
a1 (A, oS ) + · · · + am (A, oS ) ∼ cA,oS · mαA,oS (log m)βA,oS −1 .
where f (m) ∼ g(m) signifies that f (m)/g(m) → 1 as m → ∞.
Matrices, polynomials, and partitions. Prior to stating our main results, we need to
establish some notation and recall some terminology. By a partition of an integer
n > 0, we mean a non-increasing sequence λ = (λ1 , . . . , λr ) of positive integers with
n = λ1 + · · · + λr ; for background, we refer to [12]. We write |λ| := n, len(λ) := r,
and λ−1 := λr . We write λ ` n to signify that λ is a partition of n. For i > 0, define
σi (λ) := λ1 +· · ·+λi . For 1 6 j 6 |λ|, let λ−1 (j) be the unique
number i ∈ {1, . . . , len(λ)}
−1
with σi−1 (λ) < j 6 σi (λ); equivalently, λ (j) = min i ∈ {1, . . . , len(λ)} : j 6 σi (λ) .
The dual partition of λ is denoted by λ∗ . Thus, if |λ| > 0, then λ∗ = (µ1 , . . . , µt ),
where t = λ1 and µi = # i ∈ {1, . . . , len(λ)} : λi > i .
3
For a monic polynomial f = X m + am−1 X m−1 + · · · + a0 , let
0
1
..
..
.
.
C(f ) =
0
1
−a0 . . . −am−2 −am−1
be its companion matrix. Let A ∈ Mn (k). It is well-known that there are monic
irreducible polynomials f1 , . . . , fe ∈ k[X] and partitions λ1 , . . . , λe of positive integers
n1 ,. . . , ne such that n = deg(f1 )n1 + · · · + deg(fe )ne and A is similar to its (primary)
rational canonical form
λ
λ
λ
λ
e len(λe )
1,len(λ1 )
diag C f1 1,1 , . . . , C f1
, . . . . . . , C fe e,1 , . . . C fe ,
over k. We call ((f1 , λ1 ), . . . , (fe , λe )) an elementary divisor vector of A over k; any
two elementary divisor vectors of A coincide up to reordering.
Main results. Recall that k is a number field with ring of integers o. Throughout,
pv ∈ Spec(o) denotes the prime ideal corresponding to a place v ∈ Vk and qv = |o/pv |
denotes the residue field size of kv . Our global main result is the following.
Theorem A. Let S ⊂ Vk be finite and A ∈ Mn (oS ). Let ((f1 , λ1 ), . . . , (fe , λe )) be an
elementary divisor vector of A over k. Write ki = k[X]/(fi ). Let oi denote the ring of
integers of ki . Let Si = {w ∈ Vki : ∃v ∈ S.w | v} and write oi,Si := (oi )Si . Then the
following hold:
(i) There are finitely many places w1 , . . . , w` ∈ Vk \ S and associated rational functions
W1 , . . . , W` ∈ Q(X) such that
ζA,oS (s) =
`
Y
u=1
−s
Wu (qw
)
u
×
|λi |
e Y
Y
ζoi,Si (λ∗i )−1 (j) · s − j + 1 .
(1.1)
i=1 j=1
In particular, ζA,oS (s) admits meromorphic continuation to the complex plane.
(ii) The abscissa of convergence αA,oS of ζA,oS (s) satisfies αA,oS = max len(λi ) ∈ N.
16i6e
(iii) Let I := i ∈ {1, . . . , e} : len(λi ) = αA,o
PS . Then the multiplicity βA,oS of the pole
of ζA,oS (s) at αA,oS satisfies βA,oS =
λi,−1 .
i∈I
As we will see, part (i) is in fact a consequence of a similar formula (5.1) which is
−s ) in (1.1)
valid for almost all local zeta functions ζA,ov (s). The exceptional factors Wu (qw
u
cannot, in general, be omitted, see Example 5.5 below.
We note that the special case A = 0n in Theorem A is consistent with the well-known
formula ζoS (s)ζoS (s − 1) · · · ζoS (s − (n − 1)) for the zeta function enumerating all finiteindex submodules of onS . We further note that the shape of the right-hand side of (1.1) is
4
rather similar to that of Solomon’s formula [22, Thm 1] for the zeta function enumerating
submodules of finite index of a ZG-lattice for a finite group G.
Local functional equations under “inversion of the residue field size” are a common,
but not universal, phenomenon in the theory of subobject zeta functions; see [23, 24]. For
an extension of number fields k 0 /k and v ∈ Vk , let gv (k 0 ) denote the number of places of
k 0 which divide v.
Theorem B. Let A ∈ Mn (k) and let ((f1 , λ1 ), . . . , (fe , λe )) be an elementary divisor
vector of A over k. Write µi := λ∗i . Then, for almost all v ∈ Vk ,
e
λ
e
i1
P
P
P
e
|λi |
P
deg(f
)
−
deg(f
)
jµ
s
(
)
i
i
ij
2
|λi | ·gv (k[X]/(fi ))
i=1
i=1
j=1
ζA,ov (s)
= (−1)i=1
· qv
· ζA,ov (s).
qv →qv−1
(1.2)
The operation of inverting qv can be interpreted using (5.1) or, in far greater generality,
in terms of suitable explicit formulae as in [23]. We note that in the special case that
(A − a1n )n = 0 for some a ∈ k, the functional equation (1.2) follows from [24, Thm 1.2]
(see [24, Rem. 1.5]).
It is natural to ask what properties of A can be inferred from its associated zeta
functions. We will make frequent use of the following elementary observation.
Lemma. Let A, B ∈ Mn (k). Suppose that k[A] and k[B] are similar (i.e. GLn (k)conjugate). Then for almost all v ∈ Vk , ζA,ov (s) = ζB,ov (s).
The following is another consequence of our explicit formulae.
Theorem C. Let A ∈ Mn (k) and B ∈ Mm (k) be nilpotent. The following are equivalent:
(i) n = m and A and B are similar.
(ii) For almost all v ∈ Vk , ζA,ov (s) = ζB,ov (s).
(iii) There exists a finite S ⊂ Vk such that A and B both have entries in oS and such
that ζA,oS (s) = ζB,oS (s).
The nilpotency condition in Theorem C cannot, in general, be omitted, see Remark 5.8.
The author previously conjectured [17, §8.3] that generic local submodule zeta functions
associated with nilpotent matrix algebras have a simple pole at zero. In the present case,
our explicit formulae allow us to deduce the following.
Theorem D. Let A ∈ Mn (k). Then for almost all v ∈ Vk , ζA,ov (s) has a pole at zero.
Moreover, the following are equivalent:
(i) For almost all v ∈ Vk , ζA,ov (s) has a simple pole at zero.
(ii) There exists a ∈ k with (A − a1n )n = 0.
5
Behaviour at zero in general—a conjecture. We use this opportunity to state a
generalisation of our conjecture on the behaviour at zero of local submodule zeta functions
(see [17, Conj. IV and §8.3]); this generalisation disposes of the mysterious nilpotency
assumption found in its precursor.
For a ring R with polynomial submodule growth, a finitely generated R-module M , and
Ω ⊂ EndR (M ), the submodule zeta function ζΩyM (s) is the Dirichlet series enumerating
Ω-invariant R-submodules of finite index of M (cf. [17, Def. 2.1(ii)]).
Let V be a finite-dimensional vector space over k and let A ⊂ Endk (V ) be an associative,
unital subalgebra. Let rad(A) denote the (nil)radical of A. By the Wedderburn-Malcev
Theorem [6, Thm 72.19], there exists a subalgebra S ⊂ A such that A = rad(A) ⊕ S
as vector spaces (whence S ≈k A/ rad(A) is semisimple); moreover, S is unique up
to conjugacy under (1 + rad(A)) 6 A× . Choose o-forms V ⊂ V , A ⊂ Endo (V) and
S ⊂ Endo (V) of V , A, and S, respectively. We write Xv := X ⊗o ov in the following.
Conjecture E. For almost all v ∈ Vk ,
ζAv yVv (s)
ζSv yVv (s)
= 1.
s=0
This conjecture reduces to the behaviour predicted in [17, §8.3] in the “nilpotent case”
A = rad(A) ⊕ k1V . In order to make Conjecture E more explicit, we recall Solomon’s
formula for ζSv yVv (s). Let S = S1 ⊕ · · · ⊕ Sr be the Wedderburn decomposition of the
semisimple algebra S (so that each Si is simple). Let Wi be a simple Si -module and
decompose V = V1 ⊕ · · · ⊕ Vr , where Vi is isomorphic to Wimi and S acts diagonally on
V . Let ki be the centre of Si and let oi be the ring of integers of ki . Finally, let ei be the
Schur index of the central simple ki -algebra Si and define ni by dimki (Ai ) = n2i .
Theorem 1.3 ([22, §4]). For almost all v ∈ Vk ,
ζSv yVv (s) =
r m
i ei Y
Y
Y
ζoi,w (ni s − j + 1).
(1.3)
i=1 j=1 w∈Vki
w|v
The special case A = k[α] (α ∈ Endk (V )) of Conjecture E follows from Theorem 1.3
and Theorem 5.1 below.
For a more abstract interpretation of Conjecture E, note that we may identify S acting
on V with A/ rad(A) acting (faithfully) on the semi-simplification of V as an A-module
(i.e. the direct sum of the composition factors of V as an A-module).
Overview. In order to derive Theorems A–D, we proceed as follows. In §2, we reduce the
computation of ζA,oS (s) to the case that the minimal polynomial of A over k is a power
of an irreducible polynomial. In §3, we then further reduce to the case that A is nilpotent.
The heart of this article, §4, is then devoted to the explicit determination of ζA,ov (s) for
nilpotent A and almost all v ∈ Vk ; as a by-product, in Theorem 4.4, we compute the ideal
zeta function of the 2-dimensional ring Z[[X]]. We then combine our findings and derive
Theorems A–D in §5. Finally, as an application, in §6, we use Theorem A to compute the
abscissae of convergence of some (largely unknown) submodule and ideal zeta functions.
6
Acknowledgment
I would like to thank Christopher Voll for interesting discussions.
Notation
Throughout, N = {1, 2, . . . } and δij denotes the Kronecker symbol. The symbol “⊂”
indicates not necessarily proper inclusion. We use ≈R to denote both the similarity
of matrices over R and the existence of an R-isomorphism. Matrices act by rightmultiplication on row vectors. Matrix sizes are indicated by single subscripts for square
matrices and double subscripts in general; in particular, 1n and 0m,n denote the n × n
identity and m × n zero matrix, respectively.
We say that a property depending on S holds for sufficiently large finite S ⊂ Vk , if
there exists a finite S0 ⊂ Vk such that the property holds for all finite S ⊂ Vk with
S ⊃ S0 . Given v ∈ Vk , we write | · |v for the v-adic absolute value on kv with |π|v = qv−1
for π ∈ pv \ p2v .
By a p-adic field, we mean a finite extension K of the p-adic numbers Qp for some
prime p. We let OK denote the valuation ring of K and write qK for the residue field size
of K. Furthermore, νK and | · |K denote the additive valuation and absolute value on K,
−1
respectively, normalised such that any uniformiser π satisfies νK (π) = 1 and |π|K = qK
.
When the reference to K is clear, we occasionally omit the subscript “K”.
2 Reduction to the case of a primary minimal polynomial
By the following, up to enlarging S, we may reduce the computation of ζA,oS (s) to the
case where the minimal polynomial of A over k is primary (i.e. a power of an irreducible
polynomial).
Proposition 2.1. Let A ∈ Mn (k). Let f = f1 · · · fe be a factorisation of the minimal
polynomial f of A over k into a product of pairwise coprime monic polynomials fi ∈ k[X].
Let Ai ∈ Mni (k) denote the matrix of A acting on Ker(fi (A)) with respect to an arbitrary
k-basis. Then for almost all v ∈ Vk ,
ζA,ov (s) =
e
Y
ζAi ,ov (s).
i=1
Proof. It is well-known that k n = Ker(f1 (A)) ⊕ · · · ⊕ Ker(fe (A)) is an A-invariant decomposition into subspaces of dimensions n1 , . . . , ne , say, and fi is the minimal polynomial of
Ai . We may thus assume that A = diag(A1 , . . . , Ae ). By the Chinese remainder theorem,
for each i = 1, . . . , e, there exists gi ∈ k[X] with gi ≡ δij mod fj for j = 1, . . . , e. Hence,
gi (A) = diag(δi1 1n1 , . . . , δie 1ne ) ∈ k[A]. Choose a finite set S ⊂ Vk with Ai ∈ Mni (oS )
and gi ∈ oS [X] for i = 1, . . . , e.
Let v ∈ Vk \ S. Write V := onv . The block diagonal shape of A yields an A-invariant
decomposition V = V1 ⊕ · · · ⊕ Ve into free ov -modules of ranks n1 , . . . , ne . Note that A
acts as Ai on each Vi and that each gi (A) acts as the natural map V Vi ,→ V . Let
7
U 6 V be an ov -submodule. If U is A-invariant, then it decomposes as U = U1 ⊕ · · · ⊕ Ue
for Ai -invariant submodules Ui 6 Vi . We conclude that (U1 , . . . , Ue ) 7→ U1 ⊕ · · · ⊕ Ue
defines a bijection from
n
o
(U1 , . . . , Ue ) : Ui 6ov Vi and Ui Ai 6 Ui for i = 1, . . . , e
onto the set of A-invariant submodules of V whence ζA,ov (s) = ζA1 ,ov (s) · · · ζAe ,ov (s).
3 Reduction to the case of a nilpotent matrix
Recall that C(f ) denotes the companion matrix of a polynomial f . Given a partition
λ = (λ1 , . . . , λr ), let
N(λ) := diag(C(X λ1 ), . . . , C(X λr )).
Suppose that the minimal polynomial of A ∈ Mn (k) is a power of an irreducible
polynomial f ; we then say that A is (f -)primary. The elementary divisors of A are
f λ1 , . . . , f λr for a unique partition λ = (λ1 , . . . , λr ) of n/ deg(f ). We call λ the type
of A.
For an extension k 0 /k of number fields and S ⊂ Vk , define
Dk0 /k (S) = {w ∈ Vk0 : ∃v ∈ S.w | v}.
P
Hence, using the notation from Theorem B, #Dk0 /k (S) =
gv (k 0 ).
v∈S
In this section, we prove the following.
Theorem 3.1. Let f ∈ k[X] be monic and irreducible. Let A ∈ Mn (k) be an f -primary
matrix of type λ. Let k 0 = k[X]/(f ), and let o0 be the ring of integers of k 0 . Then for
almost all v ∈ Vk ,
Y
ζN(λ),o0w (s).
ζA,ov (s) =
w∈Vk0
w|v
Hence, for all sufficiently large finite S ⊂ Vk , setting S 0 = Dk0 /k (S).
ζA,oS (s) = ζN(λ),o0 0 (s).
S
Remark 3.2. In [21, §3], the study of the variety of subspaces invariant under an
endomorphism of a finite-dimensional real or complex vector space is reduced to the
case of a nilpotent endomorphism. Shayman proceeds by first reducing to the case of a
primary endomorphism ([21, Thm 2]) and our Proposition 2.1 proceeded along the same
lines. In his setting, the minimal polynomial of a primary endomorphism is a power of
a linear or quadratic irreducible and he considers these cases separately. His reasoning
is similar to arguments employed in our proof of Theorem 3.1 below. We may regard
the factorisation of ζA,ov (s) obtained by combining Proposition 2.1 and Theorem 3.1
as an arithmetic analogue of the factorisation of the space of A-invariant subspaces in
[21, Thm 3]. In [21, §4], Shayman then proceeds to study invariant subspaces of nilpotent
matrices in Jordan normal form. For our purposes, a slightly different normal form,
introduced in §4.1, will prove advantageous.
8
Our proof of Theorem 3.1 requires some preparation.
3.1 A generalised Jordan normal form for primary matrices
Let ⊗ denote the usual Kronecker product [aij ] ⊗ B = [aij B] of matrices. The following
result is a special case of the “separable Jordan normal form” in [14, §6.2]; it can also be
obtained by restriction of scalars from the usual Jordan normal form of an f -primary
matrix over a minimal splitting field of f over k.
Proposition 3.3. Let f ∈ k[X] be monic and irreducible of degree d. Let A ∈ Mn (k) be
f -primary of type λ. Write m := n/d. Then A ≈k 1m ⊗ C(f ) + N(λ) ⊗ 1d .
Lemma 3.4. Let f ∈ k[X] be monic and irreducible of degree d, λ ` m > 0, and
A = 1m ⊗ C(f ) + N(λ) ⊗ 1d . Then 1m ⊗ C(f ) = diag(C(f ), . . . , C(f )) ∈ k[A].
Proof. Write γ := C(f ) and e := λ1 ; note that X e is the minimal polynomial of
N(λ) over every field. We may naturally regard A as an m × m matrix over the
field k 0 := k[γ]. Moreover, we may identify k 0 = k[1m ⊗ C(f )] as k-algebras. Thus,
k[A, 1m ⊗ C(f )] = k 0 [γ1m + N(λ)] = k 0 [N(λ)] whence the k-dimension of k[A, 1m ⊗ C(f )]
is |k 0 : k|e = de. As f e is the minimal polynomial of A over k, the number de is also the
k-dimension of k[A] whence the claim follows.
Regarding the transition from the number field k to the local ring ov , we note that the
enveloping algebras of companion matrices take the expected forms over UFDs.
Lemma 3.5. Let R be a UFD and let f ∈ R[X] be monic. Then evaluation at C(f )
induces an isomorphism R[X]/(f ) ≈R R[C(f )].
Proof. Let K denote the field of fractions of R. The kernel of the natural map R[X] →
R[C(f )] is I := R[X] ∩ f K[X] and, clearly, f R[X] ⊂ I. Let h ∈ I so that h = f g for
some g ∈ K[X]. By [4, Thm 7.7.2], there exists a ∈ K × with af, a−1 g ∈ R[X]. As f is
monic (hence primitive), a ∈ A whence g = a(a−1 g) ∈ R[X] and h ∈ f R[X].
3.2 Properties of S-integers and their completions
Lemma 3.6. Let k 0 /k be an extension of number fields. Let o0 be the ring of integers
of k 0 . Let S ⊂ Vk be finite and S 0 = Dk0 /k (S). Then o0 ⊗o oS ≈o o0S 0 .
Proof. The following argument is taken
Q from [5]: if h is the class number of k and
a ∈ o generates the principal ideal v∈S phv , then oS = o[1/a]. We conclude that
o0 ⊗o oS = o0 [1/a] = o0S 0 .
Lemma 3.7. Let f ∈ k[X] be monic and irreducible. Let k 0 = k[X]/(f ) with ring of
integers o0 . Then the following holds for all sufficiently large finite S ⊂ Vk :
(i) oS [X]/(f ) ≈oS o0S 0 , where S 0 = Dk0 /k (S).
9
(ii) ov [X]/(f ) ≈ov
Q
o0w for v ∈ Vk \ S.
w∈Vk0
w|v
Proof. We freely use the exactness of localisation and completion; see [10, Prop. 2.5,
Thm 7.2]. Let S0 ⊂ Vk be finite with f ∈ oS0 [X]. If S ⊃ S0 , then oS0 [X]/(f ) ⊗oS0 oS ≈oS
oS [X]/(f ). As oS0 [X]/(f ) and o0 both become isomorphic to k 0 after base change to k,
for sufficiently large finite S ⊃ S0 , oS [X]/(f ) ≈oS o0S 0 by Lemma 3.6. This proves the
first part. For the second part, first note that, using (i) and Lemma 3.6,
ov [X]/(f ) ≈ov oS [X]/(f ) ⊗oS ov ≈ov o0S 0 ⊗oS ov ≈ov o0 ⊗o ov .
(3.1)
Write o(v) := ov ∩ k for the v-adic valuation ring of k. It is easy to see that we may
naturally identify o0 ⊗o o(v) with the integral closure of o(v) in k 0 . The key observation
here is that if a ∈ k 0 is a root of a monic polynomial f (X) ∈ o(v) [X], then there exists
m ∈ o with v(m) = 0 and ma ∈ o0 . Indeed, as in the proof of Lemma 3.6, we find m ∈ o
such that for all w ∈ Vk , w(m) > 0 if and only if some coefficient c of f (X) satisfies
w(c) < 0. By replacing m by a suitable power, we can ensure that all coefficients of
mf (X) belong to o whence ma is integral over o and thus belongs to o0 .
isomorphism
k 0 ⊗k kv ≈kv
QWe0 conclude (see [13, Ch. II, §8, Exerc. 4]) that the canonical
Q 0
0
kw ([13, Ch. II, Prop. 8.3]) induces an isomorphism o ⊗o ov ≈ov
ow . Part (ii) thus
w|v
w|v
follows from the latter isomorphism and (3.1).
3.3 Proof of Theorem 3.1
Recall that am (A, R) denotes the number of A-invariant R-submodules of Rn of index m,
where A ∈ Mn (R).
Proposition 3.8. Let R1 , . . . , Rr be rings with polynomial submodule growth.
(i) R := R1 × · · · × Rr has polynomial submodule growth.
(ii) (Cf. [22, Lem. 1].) Let A ∈ Mn (R) and let Ai denote the image of A under the
map Mn (R) → Mn (Ri ) induced by the projection R → Ri . Then am (A, R) =
am (A1 , R1 ) · · · am (Ar , Rr ) for each m ∈ N. Thus, ζA,R (s) = ζA1 ,R1 (s) · · · ζAr ,Rr (s).
Proof. Decompose Rn = R1n × · · · × Rrn with R acting diagonally on Rn . Multiplication by
ei = (δ1i , . . . , δni ) ∈ R acts as the natural map Rn → Rin → Rn . Given an Ri -submodule
Ui 6 Rin for i = 1, . . . , r, we obtain an R-submodule U = U1 × · · · × Ur of Rn and it is
easy to see that every R-submodule of Rn is of this form in a unique way. Evidently, U
has finite index in Rn if and only if each Ui has finite index in Rin . Part (i) is immediate
and (ii) follows since A acts as Ai on Rin .
Proof of Theorem 3.1. Assuming that the finite set S ⊂ Vk is sufficiently large, we can
make the following assumptions for all v ∈ Vk \ S:
(NOR) A = 1m ⊗ C(f ) + N(λ) ⊗ 1d ∈ Mn (ov ) for d = deg(f ) and λ ` m (Proposition 3.3).
10
(DIA) 1m ⊗ C(f ) ∈ ov [A] (Lemma 3.4).
Q 0
(INT) ov [X]/(f ) ≈ov
ow (Lemma 3.7).
w∈Vk0
w|v
Let v ∈ Vk \ S. First note that as an ov -module, ov [C(f )] is freely generated by
(1d , C(f ), . . . , C(f )d−1 ). It follows easily that onv is free of rank m as an ovQ
[C(f )]-module.
Using Lemma 3.5,(INT) allows us to identify ov [C(f )] = ov [X]/(f ) = w|v o0w =: Rv .
Thanks to (NOR), we may then regard A as an m × m matrix over Rv . It follows from
(DIA) that A-invariant ov -submodules of onv coincide with A-invariant Rv -submodules
of Rvm . Using (DIA) once more, the latter Rv -submodules are precisely those invariant
under A − C(f ) · 1m = N(λ). Therefore, ζA,ov (s) = ζN(λ),Rv (s). Noticing that the
0
(0, 1)-matrix N(λ)
Q is preserved by each projection Rv → ow , Proposition 3.8 shows that
ζN(λ),Rv (s) =
ζN(λ),o0w (s) which concludes the proof.
w|v
4 The case of a nilpotent matrix
Let λ ` n. Recall the definitions of λ−1 (j) from the introduction and of N(λ) from §3.
Definition. Wλ (X, Y ) = 1/
n
Q
−1 (j)
1 − X j−1 Y λ
∈ Q(X, Y ).
j=1
Equivalently, Wλ (X, Y ) = 1/
len(λ)
λi
Q Q
1 − X σi−1 (λ)+j−1 Y i . This section is devoted
i=1 j=1
to proving the following.
Theorem 4.1. Let λ ` n and let K be a p-adic field. Then
−s
ζN(λ∗ ),OK (s) = Wλ (qK , qK
).
Prior to giving a proof of Theorem 4.1, we record a few consequences.
Corollary 4.2. Let A ∈ Mn (k) be nilpotent of type λ (see §3). Then for all sufficiently
large finite sets S ⊂ Vk ,
ζA,oS (s) =
n
Y
ζoS (λ∗ )−1 (j) · s − j + 1 .
j=1
If A ∈ Mn (o) and A ≈o N(λ), then we may take S = ∅.
As an application, we can determine the ideal zeta function of Z[X]/(X n ). Recall that
ζ(s) denotes the Riemann zeta function.
Corollary 4.3. For every prime p,
ζZp [X]/(X n ) (s) = 1/
n
Y
(1 − pj−1−js ).
j=1
11
In particular,
ζZ[X]/(X n ) (s) =
n
Y
ζ(js − j + 1).
j=1
Proof. The matrix of multiplication by X acting on Z[X]/(X n ) with respect to the basis
(1, X, . . . , X n−1 ), i.e. the companion matrix of X n , is precisely N((n)).
Remark. The subalgebra zeta functions of Zp [X]/(X n ) are known only for n 6 4 and
sufficiently large primes p. Moreover, the author’s computation of these zeta functions
for n = 4 relied on fairly involved machine calculations; see [18, §9.2]. (The formula for
ζZp [X]/(X 4 ) (s) in [18] takes up about a page in total.)
Subobject zeta functions over rings other than oS or ov have received little attention
so far. We obtain the following.
Theorem 4.4.
(i) Z[[X]] has polynomial submodule growth.
(ii) ζZ[[X]] (s) =
∞
Q
ζ(js − j + 1) for Re(s) > 1.
j=1
Proof. It is well-known that the maximal ideals of Z[[X]] are precisely of the form (X, p)
for a rational prime p. It follows that X acts nilpotently on every Z[[X]]-module of
finite length. Hence, if U 6Z[[X]] Z[[X]]d has finite index, then U contains X n Z[[X]]d
for some n > 1. As Z[[X]] is Noetherian, U thus corresponds to a Z[X]-submodule of
Z[X]d /X n Z[X]d . In particular, (i) follows since Z[X] has polynomial submodule growth
by Theorem 1.1. Moreover, Corollary 4.3 implies the identity in (ii) on the level of formal
Dirichlet series.
In order to establish
Q∞(absolute) convergence, let s > 1 be real. By well-known facts
on infinite products,
j=1 ζ(js − j + 1) converges (absolutely) if and only if the same is
P
true of F (s) := ∞
(ζ(js
− j + 1) − 1). Using the non-negativity of the coefficients of
j=1
each Dirichlet series ζ(js − j + 1), we obtain
F (s) =
∞ X
∞
X
nj−1 (nj )−s =
j=1 n=2
gn n−s ,
n=2
where
gn := n ·
∞
X
X
m>2,j>1
n=mj
1
.
m
We see that for N > 2,
N
X
n=2
gn 6 N
X
m>2,j>1
mj 6N
N
X
1
2 log N
6N
= O(N (log N )2 ) = O(N 1+ε )
m
m
m=2
for every ε > 0. In particular, F (s) and ζZ[[X]] (s) both converge for Re(s) > 1.
12
Remark 4.5. Note, in particular, that ζZ[[X]] (s) has an essential singularity at s = 1 and
therefore does not admit meromorphic continuation beyond its abscissa of convergence.
This illustrates that Theorem 1.2(ii) does not carry over to general ground rings with
polynomial submodule growth.
In order to prove Theorem 4.1, we employ the p-adic integration machinery from
[11]. For a ring R, let Trn (R) denote the R-algebra of upper triangular n × n-matrices
over R. Recall that an element of a ring is regular if it is not a zero divisor. Write
Trreg
n (R) = {x ∈ Trn (R) : det(x) ∈ R is regular}. For a p-adic field K, let µK denote the
Haar measure on K n with µK (OnK ) = 1.
Proposition
4.6 ([11, §3]). Let K be a p-adic field and A ∈ Mn (OK ). Define VK (A) :=
reg
x ∈ Trn (OK ) : OnK xA ⊂ OnK x to be the set of upper-triangular n × n matrices
over OK whose rows span an A-invariant OK -submodule of finite index of OnK . Then
Z
s−1
s−n
−1 −n
ζA,OK (s) = (1 − qK )
|x11 |K
|x22 |s−2
(4.1)
K · · · |xnn |K dµK (x).
VK (A)
Strategy. In order to prove Theorem 4.1, we proceed as follows. First, in §4.1, we define
a matrix A(λ) which is similar (over Z) to N(λ∗ ) so that ζN(λ∗ ),OK (s) = ζA(λ),OK (s).
As we will see in §4.2, the advantage of A(λ) over N(λ∗ ) is that the sets VK (A(λ))
in Proposition 4.6 exhibit a natural, recursive structure. Specifically, we will define
dλ := (λ2 , . . . , λlen(λ) ) and find that VK (A(λ)) can be described in terms of VK (A(dλ))
and membership conditions for generic vectors in generic sublattices. In §4.3, the geometry
of such membership conditions is elucidated by means of suitable (birational) changes of
coordinates. Finally, in §4.4, we combine all these ingredients and prove Theorem 4.1.
4.1 A dual normal form for nilpotent matrices
Definition. Let λ = (λ1 , . . . , λr ) ` n > 0. Define dλ := (λ2 , . . . , λr ). We recursively
define A(λ) ∈ Mn (Z) as follows:
(i) If r 6 1, define A(λ) = 0n .
(ii) If r > 1, define
0λ1
A(λ) =
1 λ2
0λ1 −λ2 ,λ2
A(dλ)
In other words,
13
0λ1 ,λ3 +···+λr
.
(4.2)
0
λ1
A(λ) =
1 λ2
0λ1 −λ2 ,λ2
0 λ2
1λ3
0λ2 −λ3 ,λ3
..
.
..
.
..
.
1λr
0λr−1 −λr ,λr
0λr
(4.3)
By the following, the A(λ) parameterise similarity classes of nilpotent matrices.
Proposition 4.7. A(λ∗ ) and N(λ) are conjugate by permutation matrices.
Proof. Let T (λ) be the Young diagram of λ and let V (λ) be the Z-module freely generated
by the cells of T ; we use “English notation” for T (λ) and draw each row underneath its
predecessor (if any). Define Θ(λ) to be the endomorphism of V (λ) (acting on the right)
which sends each cell to its right neighbour if it exists and to zero otherwise.
We consider two orderings on the cells of T (λ) and describe the associated matrices
representing Θ(λ). The horizontal order is defined by traversing the cells of T (λ) from
left to right within each row, proceeding from top to bottom. Clearly, N(λ) is the matrix
of Θ(λ) with respect to this order.
The vertical order is obtained by traversing the cells of T (λ) from top to bottom within
each column, proceeding from left to right. Write µ := λ∗ , say µ = (µ1 , . . . , µ` ). We
now show by induction on ` that the matrix of Θ(λ) with respect to the vertical order is
A(µ)—it then follows, in particular, that A(µ) and N(λ) are conjugate as claimed.
If ` 6 1, then Θ(λ) = 0 and A(µ) = 0 so let ` > 1. Let t1 , . . . , tn be the cells of T (λ)
according to the vertical order. Then ti Θ(λ) = tµ1 +i for 1 6 i 6 µ2 and ti Θ(λ) = 0 for
µ2 < i 6 µ1 . Let λ̃ := (dµ)∗ and Ṽ := Ztµ1 +1 ⊕ · · · ⊕ Ztn . We may naturally identify the
endomorphism of Ṽ induced by Θ(λ) with Θ(λ̃) acting on V (λ̃); the defining basis of Ṽ
is then ordered vertically. By induction, the matrix of Θ(λ) acting on Ṽ with respect to
the basis (tµ1 +1 , . . . , tn ) is therefore A(dµ) whence the claim follows from the recursive
description of A(µ) in (4.2).
For |λ| > 0, let B(λ) ∈ M|λ|,|dλ| (Z) denote the matrix obtained by deleting the first λ1
columns of A(λ). The following consequence of (4.3) will be useful below.
Lemma 4.8. B(λ) contains precisely λ1 zero rows and by deleting these, the |dλ| × |dλ|
identity matrix is obtained.
4.2 Recursion
In this subsection, we give a recursive description of VK (A(λ)) (see Proposition 4.6).
14
Lemma 4.9. Let λ = (λ1 , . . . , λr ) ` n and let X be the generic upper triangular n × n
matrix. Partition X in the form
XλI 2
∗
X = 0λ1 −λ2 ,λ2 XλII1 −λ2
0
X̄λ1 ,|dλ|
0
X|dλ|
,
where subscripts are added to denote block sizes. Then
XI
X̄B(dλ)
0
0
XA(λ) = λ1
.
0
0
X A(dλ)
Proof. This follows easily from (4.2).
By Lemmas 4.8–4.9, the λ1 × |dλ| submatrix obtained by considering the first λ1 rows
of XA(λ) and then deleting the first λ1 columns is of the form
x1,1 . . . x1,λ2 ∗ . . . ∗
..
.. . . ..
..
.
. .
.
.
λ
xλ2 ,λ2 ∗ . . . ∗
(4.4)
X :=
,
.. . . ..
. .
.
∗ ... ∗
where the entries marked “∗” indicate unspecified but distinct variables taken from X̄.
Corollary 4.10. Let λ ` n and let K be a p-adic field. For x ∈ Trn (K), define x0 and
xλ by specialising X 0 and X λ from Lemma 4.9 and (4.4), respectively, at x. Then
n
|dλ| 0
λ
VK (A(λ)) = x ∈ Trreg
n (OK ) : (i) each row of x belongs to OK x and
o
(ii) x0 ∈ VK (A(dλ)) .
(4.5)
Proof. Let x ∈ Trreg
n (OK ). Clearly, x ∈ VK (A(λ)) if and only if every row of xA(λ) is
contained in the OK -span of the rows of x. By Lemma 4.9 and since det(x) 6= 0, the
first λ1 rows of xA(λ) satisfy this condition if and only if every row of xλ is contained
in the OK -span of the rows of x0 . Similarly, the rows numbered λ1 + 1, . . . , n of xA(λ)
are contained in the OK -span of x if and only if each row of x0 A(dλ) is contained in the
OK -span of x0 or, equivalently, if x0 ∈ VK (A(dλ)).
4.3 Characterising submodule membership
Condition (i) in (4.5) leads us to investigate pairs (x, y) ∈ Rn × Trn (R) (where R is a
ring) such that x is contained in the row span of y over R. In this subsection, we study
the set of all such pairs (x, y) in the case that R = OK for a p-adic field K.
15
We write An = Spec(Z[X1 , . . . , Xn ]) and Trn = Spec(Z[Yij : 1 6 i 6 j 6 n]). Let
En (R) := (x, y) ∈ Rn × Trn (R) : x ∈ Rn y .
(4.6)
We identify An × Trn = Spec(Z[X1 , . . . , Xn , Y11 , . . . , Y1n , Y22 , . . . , Ynn ]). Define
Cn := (α, ω) ∈ Rn>0 × Trn (R>0 ) : ωii 6 αi for 1 6 i 6 n .
(4.7)
For a p-adic field K, we extend νK to families of elements of K via νK (a1 , . . . , am ) =
(νK (a1 ), . . . , νK (am )) and write
n
o
Cn (K) := (x, y) ∈ K n × Trn (K) : (νK (x), νK (y)) ∈ Cn ⊂ OnK × Trreg
n (OK ).
The following lemma will play a key role in our proof of Theorem 4.1. It shows that
away from sets of measure zero, a suitable Z-defined change of coordinates (defined
independently of K) transforms En (OK ) into Cn (K).
Lemma 4.11. There exist
• closed subschemes Vn , Vn0 ⊂ An × Trn of the form fn = 0 and fn0 = 0, respectively,
where fn , fn0 ∈ Z[X, Y ] are non-zero non-units, and
• an isomorphism ϕn : (An × Trn ) \ Vn → (An × Trn ) \ Vn0
such that the following conditions are satisfied:
0
K
(i) For each p-adic field K, ϕK
n (En (OK ) \ Vn (OK )) = Cn (K) \ Vn (OK ), where ϕn
denotes the map induced by ϕn on K-points.
(ii) The Jacobian determinant of ϕn is identically 1.
(iii) ϕn commutes with (the restriction to its domain of ) the projection of An × Trn
onto Trn and (the restriction of ) the projection onto the first coordinate of An .
Example (n = 2). Let K be a p-adic field; we drop the subscripts “K” in the following.
Let x, y, a, b, c ∈ O and suppose that
x(ay
− bx)abc 6= 0. Define y 0 := y − xa b ∈ K and
note that y 0 6= 0. Then (x, y) ∈ O2 · a0 cb if and only if ν(a) 6 ν(x) and (x, y) − xa (a, b) =
(0, y 0 ) ∈ O(0, c); the latter condition
toν(c) 6 ν(y 0 ) and implies that y 0 ∈ O.
is equivalent
We see that the map ((x, y), a0 cb ) 7→ ((x, y 0 ), a0 cb ) has the properties of ϕ2 stated in
Lemma 4.11.
Proof of Lemma 4.11. We proceed by induction. For n = 1, we let f1 = f10 = X1 Y11 and
define ϕ1 to be the identity. Clearly, (i)–(iii) are satisfied.
Let n > 1 and suppose that ϕn−1 with the stated properties has been defined. Let K
be a p-adic field and let (x, y) ∈ K n × Trn (K) with x1 y11 6= 0. We again drop the
subscripts “K”. Gaussian elimination shows that (x, y) ∈ En (O) if and only if the
following conditions are satisfied:
(a) xi , yij ∈ O for 1 6 i 6 j 6 n,
16
(b)
x1
y11
∈ O, and
(c) x2 −
x1
y11 y12 , . . . , xn
−
x1
y11 y1n
h i
∈ On−1 · yij
26i6j6n
.
We will now simplify (c) using a change of coordinates. For 2 6 j 6 n, let x0j :=
xj − yx111 y1j . Write x01 := x1 and x0 := (x01 , . . . , x0n ). Note that (x, y) 7→ (x0 , y) is
an automorphism of the complement of Y11 = 0 in An × Trn and that the Jacobian
determinant of this map is identically 1.
Assuming that yij ∈ O for 1 6 i 6 j 6 n and yx111 ∈ O, we see that xj ∈ O if and only
if x0j ∈ O. Hence, (x, y) ∈ En (O) if and only if (b) and the following two conditions are
satisfied:
(a’) x0i , yij ∈ O for 1 6 i 6 j 6 n,
h i
(c’) (x02 , . . . , x0n ) ∈ On−1 · yij
26i6j6n
.
After excluding suitable hypersurfaces, our inductive hypothesis allows us to perform
another change of coordinates, replacing x02 , . . . , x0n by x002 , . . . , x00n , say, such that (x, y) ∈
En (K) if and only if the following conditions are satisfied:
(a”) x00i , yij ∈ O for 1 6 i 6 j 6 n (where x001 := x01 = x1 ) and
(c”) ν(yii ) 6 ν(x00i ) for 1 6 i 6 n;
note that (b) is implied by the case i = 1 of (c”).
For (i), assuming that the product of all x00i and yij is non-zero, conditions (a”) and (c”)
are both satisfied if and only if (x00 , y) ∈ Cn (K), where x00 := (x001 , . . . , x00n ). The change
of coordinates x 7→ x00 is defined over Z, does not depend on K, and, does not modify
the x1 - or y-coordinate, as required for (iii); part (ii) follows since ϕn is defined as a
composite of maps, the Jacobian determinant of each of which is identically 1.
Remark 4.12. It follows from Lemma 4.11(ii) that the change of variables afforded
by ϕn does not affect p-adic measures. Moreover, it is well-known that if 0 6= f ∈
OK [X1 , . . . , Xn ], then the zero locus of f in OnK has measure zero. We conclude that
Vn and Vn0 in Lemma 4.11 are without relevance for the computation of the integral in
Proposition 4.6.
4.4 Final steps towards Theorem 4.1
By combining Corollary 4.10 and Lemma 4.11, we may reduce the computation of the
integral in Proposition 4.6 for A = A(λ) to a purely combinatorial problem.
Proposition 4.13. Let λ = (λ1 , . . . , λr ) ` n and let K be a p-adic field. Then
−1 −n
ζA(λ),OK (s) = (1 − qK
)
Z
Vλ (OK )
17
n
Y
i=1
|xi |s−i
K dµ(x),
(4.8)
n(n+1)/2
where Vλ (OK ) consists of those x ∈ OK
satisfying the following divisibility conditions, where the yi,j,` below denote distinct variables among the xn+1 , . . . , xn(n+1)/2 :
• For 2 6 i 6 r and 1 6 j 6 λi ,
xσi−1 (λ)+j xσi−2 (λ)+j , yi,j,1 , . . . , yi,j,j−1 .
• For 3 6 i 6 r and σi−1 (λ) < j 6 n,
xj yi,j,n+1 , . . . , yi,j,n+λi−2 .
Remark. Since the yi,j,` do not appear in the integrand in the right-hand side of (4.8),
it is of no consequence precisely which of the xn+1 , . . . , xn(n+1)/2 each yi,j,` refers to
provided that distinct triples (i, j, `) yield different yi,j,` .
Proof of Proposition 4.13. If r 6 1, the claim is trivially true so let r > 2.
As our first step, we combine Corollary 4.10 and Lemma 4.11 in order to transform
the membership condition (i) in (4.5) into the given divisibility conditions for i = 2 and
i = 3, respectively; here, x1 , . . . , xn correspond to the diagonal entries x11 , . . . , xnn in
Proposition 4.6. This transformation does not affect the integrand in (4.1) thanks to
condition (iii) in Lemma 4.11.
Subsequent steps then recursively apply the same procedure in order to express the
condition x0 ∈ VK (A(dλ)) in Corollary 4.10 in terms of the stated divisibility conditions,
taking into account the evident shifts of variable indices. Crucially, in doing so, none of
the diagonal coordinates x1 , . . . , xn will ever be modified, again thanks to condition (iii)
in Lemma 4.11. Therefore, the divisibility conditions obtained during earlier steps will
never be altered by subsequent ones. The claim thus follows by induction.
Proof of Theorem 4.1. We once again omit subscripts “K” in the following. Moreover,
we will make repeated use of the identity
Z
Z
r
s
|x| |y| dµ(x, y) = |x|r+s+1 |y|s dµ(x, y)
(4.9)
{(x,y)∈O2 :x|y}
O2
which follows by performing a change of variables
y = xy 0 on the left-hand side. We will
R
s
furthermore use the well-known identity O |x| dµ(x) = (1 − q −1 )/(1 − q −s−1 ).
By repeatedly applying (4.9), we can eliminate all the yi,j,` variables and rewrite (4.8)
as an integral over On . In order to record the effect of this procedure on the integrand, we
use λ to index x1 , . . . , xn as follows. Let f (i, j) := σi−1 (λ) + j and, for x = (x1 , . . . , xn ),
write xij := xf (i,j) . Define
n
o
Uλ (O) := x ∈ On : xi,j | xi−1,j for 2 6 i 6 r and 1 6 j 6 λi
18
Proposition 4.7 and repeated applications of (4.9) to (4.8) show that
Z
−1 −n
ζN(λ∗ ),O (s) = ζA(λ),O (s) = (1 − q )
Fλ (x) dµ(x),
Uλ (O)
where
Fλ (x) =
λi
r Y
Y
s−f (i,j)
xij
×
=
x1j
xij
j−1
×
s−j
×
j=1
λi
r Y
Y
s−(λi−1 +1)
xij
λi
r Y
r Y
Y
xij
λa−2
a=3 i=a j=1
i=2 j=1
i=1 j=1
λ1
Y
λi
r Y
Y
;
i=2 j=1
P
the second equality follows since s − f (i, j) + j − 1 + ia=3 λa−2 = s − (λi−1 + 1) for
2 6 i 6 r and 1 6 j 6 λi . Another sequence of applications of (4.9) can be used to
remove the divisibility conditions in Uλ (O), yielding
(1 − q
−1 n
) ζA(λ),O (s) =
Z Y
λ1
On
=
x1j
j=1
×
λi
r Y
Y
s−j+i−1+
xij
i−1
P
(s−(λa +1))
a=1
dµ(x)
i=2 j=1
Z Y
λi
r Y
On
s−j
xij
is−(σi−1 (λ)+j)
dµ(x)
i=1 j=1
= (1 − q −1 )n ·
λi
r Y
Y
1 − q −is+σi−1 (λ)+j−1
−1
i=1 j=1
= (1 − q
−1 n
) · Wλ (q, q −s ).
5 Proofs of Theorems A–D
At the heart of our proofs of Theorems A–D lies the following local version of Theorem A.
Theorem 5.1. Let S ⊂ Vk be finite and A ∈ Mn (oS ). Let ((f1 , λ1 ), . . . , (fe , λe )) be an
elementary divisor vector of A over k. Write ki = k[X]/(fi ). Let oi denote the ring of
integers of ki . Then for almost all v ∈ Vk ,
ζA,ov (s) =
|λi |
e Y
Y
Y
ζoi,w (λ∗i )−1 (j) · s − j + 1 .
(5.1)
i=1 j=1 w∈Vki
w|v
Proof. Combine Proposition 2.1, Theorem 3.1, and Theorem 4.1.
The following is a consequence of Proposition 4.6 and well-known rationality results
from p-adic integration.
19
Proposition 5.2 (Cf. [11, §3]). Let K be a p-adic field and let A ∈ Mn (OK ). Then
−s
ζA,OK (s) ∈ Q(qK
). Hence, ζA,OK (s) admits meromorphic continuation to all of C.
In order to deduce parts (ii)–(iii) of Theorem A, we will use the following corollary to
the detailed analysis of analytic properties of subobject zeta functions in [7].
Lemma 5.3. Let S 0 ⊂ Vk be finite, S ⊂ S 0 , and let A ∈ Mn (oS ). Then αA,oS = αA,oS0
and βA,oS = βA,oS0 .
Proof. We first argue that αA,ov < αA,oS for each v ∈ Vk \S. The zeta function ζA,oS (s+n)
is an Euler product of cone integrals (cf. Proposition 4.6) in the sense of [7, Def. 4.2];
cf. [7, Cor. 5.6]. Using the notation from [7], by [7, Cor. 3.4] (which is correct despite
a minor, fixable mistake in [7, Prop. 3.3], see [1, Rem. 4.6]), it follows that each αA,ov
for v ∈ Vk \ S is a number of the form n − Bj /Aj for j = 1, . . . , q. Hence, by combining
[7, Cor. 4.14, Lem. 4.15], for each v ∈ Vk \ S,
1 − Bk
= αA,oS .
Ak
Q
Clearly, 0 < αA,oS0 6 αA,oS . Define F (s) = v∈S 0 \S ζA,ov (s) so that ζA,oS (s) =
F (s)ζA,oS0 (s) for all s ∈ C with Re(s) > αA,oS − δ and some constant δ > 0 (see
Theorem 1.2). By the above, every real pole of F (s) is less than αA,oS . Since F (s) is a
non-zero Dirichlet series with non-negative coefficients, we conclude that F (αA,oS ) > 0.
In particular, since ζA,oS (s) has a pole at αA,oS , the same is true of ζA,oS0 (s) whence
αA,oS0 > αA,oS . Moreover, F (αA,oS ) > 0 clearly also implies that βA,oS = βA,oS0 .
αA,ov < n + max
k=1,...,q
Remark 5.4.
(i) The corresponding statement for subalgebra and submodule zeta functions (proved
in the same way) is certainly well-known to experts in the area. Unfortunately, it
does not seem to have been spelled out in the literature. For a similar statement in
the context of representation zeta functions, see [2, Thm 1.4].
(ii) While in [7] only the case k = Q, S = ∅ is discussed, their arguments carry over to
the present setting in the expected way (cf. [1] and [9, §4]).
Proof of Theorem A. Part (i) follows from Theorem 5.1 and Proposition 5.2. Let µ ` n.
We now determine the largest real pole, α say, and its multiplicity, β say, of
Z(s) :=
n
Y
ζoS (µ−1 (j) · s − j + 1).
j=1
Write r = len(µ). Since ζoS (s) has a unique pole at 1 (with multiplicity 1) and ζoS (s0 ) 6= 0
for real s0 > 1,
α = max
16j6n
j
σi−1 (µ) + j
σi (µ)
= max max
= max
= µ1 = len(µ∗ ),
16i6r
µ−1 (j) 16i6r 16j6λi
i
i
20
σi+1 (µ)
i+1
1. Next, β is precisely the number of i ∈ {1, . . . , r} with µ1 = σi (µ)
i
the largest ` > 1 with µ1 = . . . = µ` . In other words, β = µ∗−1 .
where the penultimate equality follows since iµi+1 6 σi (µ) and thus
σi (µ)
i
>
for
1 6 i 6 r−
or,
equivalently,
Parts (ii)–(iii) of Theorem A now follow from Lemma 5.3 and the observation that
Z(s) > 0 for s > α.
−s ) in Theorem A is in
Example 5.5. The presence of the exceptional factors Wu (qw
u
general unavoidable. For a simple example, let a ∈ o be non-zero and define A = 00 a0 .
Using Proposition 4.6, a simple computation reveals that for v ∈ Vk ,
(1−s)(v(a)+1)
ζA,ov (s) =
1 − qv1−2s + qv
1 − qv1−s
· (qv−s − 1)
· ζov (s)ζov (2s − 1);
(5.2)
note that ζA,ov (s) = ζov (s)ζov (2s − 1) whenever v(a) = 0. We further note that the
exceptional factor in (5.2) in fact belongs to Z[qv−s ] and is thus regular at s = 1; this is
consistent with the general fact that for subobject zeta functions, each local abscissa of
convergence is strictly less than the associated global one (see the proof of Lemma 5.3).
Finally note the failure of (1.2) for the finitely many v ∈ Vk with v(a) > 0.
Remark 5.6. In view of a conjecture of Solomon proved by Bushnell and Reiner [3], it
is natural to ask if the Wu ∈ Q(X) in Theorem A are in fact always elements of Z[X].
Proof of Theorem B. The claim follows by combining Theorem 5.1 and the following
simple observation. Let k 0 /k be an extension of number fields, let o0 be the ring of
integers of k 0 , and let v ∈ Vk be unramified in k 0 . If w ∈ Vk0 divides v, define f(w/v) by
f(w/v)
qw = qv
. Define
Y
Y
−1
Zv (s) =
ζo0w (s) =
1 − qv−f(w/v)s
.
w∈Vk0
w|v
w∈Vk0
w|v
Then, recalling the definition of gv (k 0 ) from p. 5 and using
P
f(w/v) = |k 0 : k|,
w|v
Zv (s)
0
qv →qv−1
0
= (−1)gv (k ) qv−|k :k|s · Zv (s).
0
Lemma 5.7. Let S ⊂ Vk be finite. Let Z(s) and
series
Q Z (s) be two Dirichlet
Q with finite
abscissae of convergence. Suppose that Z(s) = v∈Vk \S Zv (s) and Z0 (s) = v∈Vk \S Z0v (s),
where each Zv (s) and Z0v (s) is a series in qv−s with non-negative real coefficients. Suppose
that Z(s) = Z0 (s) and that W (X, Y ), W 0 (X, Y ) ∈ Q(X, Y ) satisfy Zv (s) = W (qv , qv−s )
and Z0v (s) = W 0 (qv , qv−s ) for almost all v ∈ Vk \ S. Then W (X, Y ) = W 0 (X, Y ).
Proof.
of S.
in the
of the
Let S0 be the set of rational primes which are Q
divisible by at least one element
For a rational prime p 6∈ S0 , define Zp (s) = v∈Vk ,v|p Zv (s) and define Z0p (s)
same way. Assuming that Z(s) = Z0 (s), it is well-known that the coefficients
Dirichlet series Z(s) and Z0 (s) coincide. We conclude that Zp (s) = Z0p (s) for
21
p 6∈ S0 . By Chebotarev’s density theorem, there exists an infinite set of rational primes
P such that each p ∈ P splits completely in k. Writing d = |k : Q|, for almost all
p ∈ P , we thus have W (p, p−s )d = Zp (s) = Z0p (s) = W 0 (p, p−s )d which easily implies
W (X, Y )d = W 0 (X, Y )d . Thus, W (X, Y )/W 0 (X, Y ) is a dth root of unity in R(X, Y )
and hence in R, for the latter is algebraically closed in the former (see [4, Prop. 11.3.1]).
The non-negativity assumptions on the coefficients of Zv (s) and Z0v (s) as series in qv−s
now imply W (X, Y ) = W 0 (X, Y ).
Proof of Theorem C. The implications “(i)⇒(ii)⇒(iii)” in Theorem C are obvious. Suppose that (iii) holds. Let λ and µ be the types of the matrices A and B, respectively.
By Theorem 5.1 and the preceding lemma, Wλ (X, Y ) = Wµ (X, Y ). It is easy to see that
the binomials 1 − X a Y b for a > 0 and b > 1 freely generate a free abelian subgroup of
Q(X, Y )× . Hence, λ = µ and A and B are similar.
Remark 5.8. If A is nilpotent and α ∈ k × , then A and A + α1n give rise to the same
local and global zeta functions without A and A + α1n being similar. In general, equality
of local and global zeta functions associated with non-nilpotent matrices A and B does not
suffice to even conclude that the algebras k[A] and k[B] are similar. We give two examples
to illustrate this behaviour, the first being arithmetic and the second of combinatorial
origin.
(i) By [15], there are monic irreducible polynomials f, g ∈ Z[X] of the same degree
such that the number fields Q[X]/(f ) and Q[X]/(g) are non-isomorphic but have
the same Dedekind zeta functions; moreover, as explained in [15, §1], every rational
prime has the same “splitting type” in each of these two number fields. Consequently,
ζC(f ),Zp (s) = ζC(g),Zp (s) for almost all primes p
(ii) Recall the definition of Wλ from §4. A simple calculation shows that
W(2,2,1) · W(3,1) = W(2,2) · W(3,1,1) .
Let a, b ∈ k × be distinct and choose A, B ∈ M9 (k) to have elementary divisor vectors
((X − a, (3, 2), (X − b, (2, 1, 1))) and ((X − a, (2, 2)), (X − b, (3, 1, 1))), respectively.
Then k[A] and k[B] are not similar but ζA,ov (s) = ζB,ov (s) for almost all v ∈ Vk .
Remark 5.9. We further note that even for nilpotent A, the family of associated
functional equations (1.2) in Theorem B does not determine A up to similarity; an
example is given by two nilpotent 7 × 7-matrices with types (3, 1, 1, 1, 1) and (2, 2, 2, 1),
respectively.
Proof of Theorem D. By Theorem 5.1, ζA,ov (s) has a pole at zero for almost all v ∈ Vk .
Moreover, again for almost all v ∈ Vk , this pole is simple if and only if e = 1 and almost
all places of k remain inert in k[X]/(f1 ); the latter condition is equivalent to f1 being
linear.
22
6 Applications
6.1 Submodules for unipotent groups
Let S ⊂ Vk be finite, let M be a finitely generated oS -module, and let Ω ⊂ EndoS (M).
We let αΩyM denote the abscissa of convergence of ζΩyM (s). As a special case (cf.
[17, Rem. 2.2(ii)])), given a possibly non-associative oS -algebra A whose underlying
oS -module is finitely generated, we let αA denote the abscissa of convergence of its ideal
zeta function ζA (s), as defined in the introduction. We now illustrate how Theorem A can
sometimes be used to determine αΩyM or αA without computing the corresponding zeta
function. The key observation is that if ω ∈ Ω, then αΩyM 6 αω,o ; by Theorem A(ii),
the latter number can be easily read off from an elementary divisor vector of ω ⊗oS k.
We let Un denote the group scheme of upper unitriangular n × n matrices. For
λ = (λ1 , . . . , λr ) ` n, we regard Uλ := Uλ1 × · · · × Uλr as a subgroup scheme of Un
via the natural diagonal embedding. The case len(λ) = 1 of the following provides an
affirmative answer to [18, Question 9.7].
Proposition 6.1. Let λ ` n. Then αUλ (o)yon = len(λ).
Proof. Using the characterisation of Um (k) as the centraliser of a maximal flag of
subspaces of k m , we see that on contains an Uλ (o)-invariant submodule N such that Uλ (o)
acts trivially on on /N and on /N ≈o olen(λ) . We conclude that αUλ (o)yon > len(λ). For
an upper bound, note that (1 + N(λ)) ∈ Uλ (o) whence αUλ (o)yon 6 αN(λ),o = len(λ).
For |λ| 6 5 and almost all v ∈ Vk , explicit formulae for ζUλ (ov )yonv (s) have been
obtained by the author (see [18, §9.4] and the database included with [19]); the only
unknown case for len(λ) = 6, namely λ = (6), seems out of reach at present. In addition
to their global abscissae of convergence, the ζUλ (ov )yonv (s) are known to generically satisfy
local functional equations under inversion of qv by [24, §5.2].
6.2 Lie algebras of maximal class
Let g be a finite-dimensional Lie k-algebra. For finite S ⊂ Vk , by an oS -form of g, we
mean a Lie oS -algebra g whose underlying module is free and such that g ⊗oS k ≈k g.
Let g = g1 ⊃ g2 ⊃ · · · be the lower central series of g. Recall that g has maximal
class if g is nilpotent of class dimk (g) − 1. Equivalently, g has maximal class if and only
if dimk (g1 /g2 ) = 2 and dimk (gi /gi+1 ) = 1 for 1 6 i 6 dimk (g) − 1.
Proposition 6.2. Let g be an oS -form of a non-abelian finite-dimensional Lie k-algebra
of maximal class. Then αg = 2.
A proof of Proposition 6.2 using Theorem A will be given below.
We note that Proposition 6.2 is consistent with explicit calculations carried out for
specific Z-forms of the Lie algebras M3 ,M4 ,M5 , and Fil4 of maximal class and dimension
at most 5 over the rationals; see [8, Ch. 2].
23
Lemma 6.3. Let S ⊂ Vk be finite. Let g be an oS -form of a nilpotent Lie k-algebra of
finite dimension n. Let A be the enveloping unital associative algebra of ad(g) within
EndoS (g).
(i) For each ϕ ∈ A, there exists c ∈ oS with (ϕ − c1g )n = 0; thus, ϕ ⊗oS k is primary.
(ii) Let ϕ ∈ A have type λ over k. Then αg 6 len(λ).
Proof. The first part follows from Engel’s theorem and the second part is then an
immediate consequence of Theorem A(ii).
Lemma 6.4. Let g be an (n + 2)-dimensional non-abelian Lie k-algebra of maximal class.
Then there exists a k-basis (x1 , x2 , y1 , . . . , yn ) of g such that [x1 , x2 ] = y1 , [x1 , yi ] = yi+1
for 1 6 i 6 n − 1, and [x1 , yn ] = 0.
L
Proof. Consider the graded Lie algebra i>1 gi /gi+1 associated with g. We claim that
there exists an element a ∈ g/g2 such that [a, · ] maps gi /gi+1 onto gi+1 /gi+2 for each
i > 1. To see that, first note that [g/g2 , gi /gi+1 ] = gi+1 /gi+2 for each i > 1. Let (u, v)
be a k-basis of g/g2 . Then [u, v] spans g2 /g3 . Moreover, if wi spans gi /gi+1 , then the
image of at least one of [u, wi ] and [v, wi ] spans gi+1 /gi+2 . Consequently, we may take
a = u + cv for almost all c ∈ k.
Given a as above, choose b ∈ g/g2 such that (a, b) is a basis of g/g2 . Let x1 , x2 ∈ g
be preimages of a and b, respectively. Then, if we define y1 = [x1 , x2 ] and yi+1 = [x1 , yi ],
we obtain a basis (x1 , x2 , y1 , . . . , yn ) of the desired form.
Proof of Proposition 6.2. By Lemma 5.3 and Remark 5.4(i), we are free to enlarge S as
needed. In particular, we may assume that g/g2 ≈oS o2S whence αg > 2 follows. Moreover,
we may assume that g possesses an oS -basis (x1 , x2 , y1 , . . . , yn ) as in by Lemma 6.4. The
matrix of [x1 , · ] with respect to the basis (x2 , y1 , . . . , yn , x1 ) is precisely N((n + 1, 1))
whence αg 6 2 follows from Theorem A.
References
[1] N. Avni, B. Klopsch, U. Onn, and C. Voll, Representation zeta functions of compact p-adic analytic
groups and arithmetic groups, Duke Math. J. 162 (2013), no. 1, 111–197.
[2]
, Arithmetic Groups, Base Change, and Representation Growth, Geom. Funct. Anal. 26
(2016), no. 1, 67–135.
[3] C. J. Bushnell and I. Reiner, Zeta functions of arithmetic orders and Solomon’s conjectures, Math.
Z. 173 (1980), no. 2, 135–161.
[4] P. M. Cohn, Basic algebra, Springer-Verlag London, Ltd., London, 2003. Groups, rings and fields.
[5] B. Conrad, The lattice of S-integers. See http://math.stanford.edu/~conrad/248APage/handouts.
[6] C. W. Curtis and I. Reiner, Representation theory of finite groups and associative algebras, Pure
and Applied Mathematics, Vol. XI, Interscience Publishers, a division of John Wiley & Sons, New
York-London, 1962.
[7] M. P. F. du Sautoy and F. J. Grunewald, Analytic properties of zeta functions and subgroup growth,
Ann. of Math. (2) 152 (2000), no. 3, 793–833.
[8] M. P. F. du Sautoy and L. Woodward, Zeta functions of groups and rings, Lecture Notes in
Mathematics, vol. 1925, Springer-Verlag, Berlin, 2008.
24
[9] D. H. Dung and C. Voll, Uniform analytic properties of representation zeta functions of finitely
generated nilpotent groups (2015). To appear in Trans. Amer. Math. Soc. arXiv:1503.06947.
[10] D. Eisenbud, Commutative algebra, Graduate Texts in Mathematics, vol. 150, Springer-Verlag, New
York, 1995. With a view toward algebraic geometry.
[11] F. J. Grunewald, D. Segal, and G. C. Smith, Subgroups of finite index in nilpotent groups, Invent.
Math. 93 (1988), no. 1, 185–223.
[12] I. G. Macdonald, Symmetric functions and Hall polynomials, The Clarendon Press, Oxford University
Press, New York, 1979. Oxford Mathematical Monographs.
[13] J. Neukirch, Algebraic number theory, Grundlehren der Mathematischen Wissenschaften, vol. 322,
Springer-Verlag, Berlin, 1999.
[14] C. Norman, Finitely generated abelian groups and similarity of matrices over a field, Springer
Undergraduate Mathematics Series, Springer-Verlag, London, 2012.
[15] R. Perlis, On the equation ζK (s) = ζK 0 (s), J. Number Theory 9 (1977), no. 3, 342–360.
[16] C. M. Ringel and M. Schmidmeier, Invariant subspaces of nilpotent linear operators. I, J. Reine
Angew. Math. 614 (2008), 1–52.
[17] T. Rossmann, Computing topological zeta functions of groups, algebras, and modules, I, Proc. Lond.
Math. Soc. (3) 110 (2015), no. 5, 1099–1134.
[18]
[19]
, Computing local zeta functions of groups, algebras, and modules (preprint) (2016).
arXiv:1602.00919.
, Zeta, version 0.3, 2016. See http://www.math.uni-bielefeld.de/~rossmann/Zeta/.
[20] D. Segal, On the growth of ideals and submodules, J. London Math. Soc. (2) 56 (1997), no. 2, 245–263.
[21] M. A. Shayman, On the variety of invariant subspaces of a finite-dimensional linear operator, Trans.
Amer. Math. Soc. 274 (1982), no. 2, 721–747.
[22] L. Solomon, Zeta functions and integral representation theory, Advances in Math. 26 (1977), no. 3,
306–326.
[23] C. Voll, Functional equations for zeta functions of groups and rings, Ann. of Math. (2) 172 (2010),
no. 2, 1181–1218.
[24]
, Local functional equations for submodule zeta functions associated to nilpotent algebras of
endomorphisms (preprint) (2016). arXiv:1602.07025.
[25]
, Zeta functions of groups and rings – recent developments, Groups St Andrews 2013, 2016,
pp. 469–492.
25
| 4math.GR
|
Journal of Machine Learning Research ()
Submitted ; Published
A k-core Decomposition Framework for Graph Clustering
Christos Giatsidis
[email protected]
LIX, École Polytechnique
Palaiseau, 91120, France
Fragkiskos D. Malliaros
[email protected]
LIX, École Polytechnique
Palaiseau, 91120, France
arXiv:1607.02096v1 [cs.SI] 7 Jul 2016
Nikolaos Tziortziotis
[email protected]
LIX, École Polytechnique
Palaiseau, 91120, France
Charanpal Dhanjal
[email protected]
TSI, Télécom ParisTech
Paris, 75013, France
Emmanouil Kiagias
[email protected]
LIX, École Polytechnique
Palaiseau, 91120, France
Dimitrios M. Thilikos
[email protected]
CNRS, LIRMM, University of Athens
Montpellier, 34095, France and
Athens, 15784, Greece
Michalis Vazirgiannis
[email protected]
LIX, École Polytechnique
Palaiseau, 91120, France
Editor: -
Abstract
Graph clustering or community detection constitutes an important task for investigating
the internal structure of graphs, with a plethora of applications in several domains. Traditional techniques for graph clustering, such as spectral methods, typically suffer from high
time and space complexity. In this article, we present CoreCluster, an efficient graph
clustering framework based on the concept of graph degeneracy, that can be used along with
any known graph clustering algorithm. Our approach capitalizes on processing the graph
in an hierarchical manner provided by its core expansion sequence, an ordered partition
of the graph into different levels according to the k-core decomposition. Such a partition
provides an efficient way to process the graph in an incremental manner that preserves its
clustering structure, while making the execution of the chosen clustering algorithm much
faster due to the smaller size of the graph’s partitions onto which the algorithm operates.
An experimental analysis on a multitude of real and synthetic data demonstrates that our
approach can be applied to any clustering algorithm accelerating the clustering process,
while the quality of the clustering structure is preserved or even improved.
Keywords: Graph clustering, community detection, k-core decomposition, graph degeneracy, graph mining
©.
Giatsidis, Malliaros, Tziortziotis, Dhanjal, Kiagias, Thilikos and Vazirgiannis
1. Introduction
Detecting clusters or communities in graphs constitutes a cornerstone problem with many
applications in several disciplines. Characteristic application domains include social and
information network analysis, biological networks, recommendation systems and image segmentation. Due to its importance and multidisciplinary nature, the problem of graph clustering has received great attention from the research community and numerous algorithms
have been proposed (see Fortunato, 2010, for a survey in the area).
Spectral clustering (e.g. Ng et al., 2001; von Luxburg, 2007) is one of the most sophisticated methods for capturing and analyzing the inherent structure of data and can have
highly accurate results on different data types such as data points, images, and graphs.
Nevertheless, spectral methods impose a high cost of computing resources both in time and
space regardless of the data on which it is going to be applied (Fortunato, 2010). Other
well-known approaches for community detection are the ones based on modularity optimization (Newman and Girvan, 2004; Clauset et al., 2004), stochastic flow simulation (Satuluri
and Parthasarathy, 2009) and local partitioning methods (Fortunato, 2010). In any case,
scalability is still a major challenge in the graph clustering task, especially nowadays with
the significant increase of the graphs’ size.
Typically, the methodologies for scaling up a graph clustering method can be divided into
two main categories: (i) algorithm-oriented and (ii) data-oriented. The former considers the
algorithm of interest and appropriately optimizes, whenever is possible, the “parts” of the
algorithm responsible for scalability issues. Prominent examples here are the fast modularity
optimization method (Clauset et al., 2004) and the scalable flow-based Markov clustering
algorithm (Satuluri and Parthasarathy, 2009). The latter constitutes a widely used class
of methodologies relied on sampling/sparsification techniques. In this case, the size of
the graph onto which the algorithm will operate is reduced, by disregarding nodes/edges.
However, possible useful structural information of the graph (i.e., nodes/edges) is ignored
in this approach.
In this paper, we propose CoreCluster, a graph clustering framework that capitalizes
on the notion of graph degeneracy, also known as k-core decomposition (Seidman, 1983).
The main idea behind our approach is to combine any known graph clustering algorithm
with an easy-to-compute, clustering-preserving hierarchical representation of the graph –
as produced by the k-core decomposition – towards a scalable graph clustering tool. The
k-core of a graph is a maximal size subgraph where each node has at least k neighbors
in the subgraph (we say that k is the rank of such a core). The maximum k for which a
graph contains a k-core is known as its degeneracy. We refer to this core as “the densest
core”. Intuitively, the k-core of such a graph is located in its “densest territories”. Based on
this idea, we show that the densest cores of a graph are roughly maintaining its clustering
structure and thus constitute good starting points (seed subgraphs) for computing it. Given
the fact that the size of the densest core of a graph is orders of magnitude smaller than that
of the original graph, we apply a clustering algorithm starting from its densest core and then,
on the resulting structure, we incrementally cluster the rest of the nodes in the lower rank
cores in decreasing order, following the hierarchy produced by the k-core decomposition.
At a high level, the main contributions of this paper are three-fold:
2
CoreCluster: A core decomposition framework for graph clustering
• Clustering Framework : We introduce CoreCluster, a scalable degeneracy-based
graph clustering framework, that can be used along with any known graph clustering
algorithm. We show how CoreCluster utilizes the k-core decomposition of a graph
in order to (i) select seed subgraphs for starting the clustering process and (ii) expand
the already formed clusters or create new ones.
• Scalability and Accuracy Analysis: We discuss analytically the ability of CoreCluster to scale-up, describing its expected running time. We also justify why the k-core
decomposition provides the direction under which to perform clustering incrementally.
More specifically, we demonstrate that the k-core structure captures the clustering
properties of a graph, thus being able to indicate good seed subgraphs for a clustering algorithm. Furthermore, we derive upper bounds about the difference between
eigenspaces of successive cores. Finally, we show that the cluster expansion process
of the CoreCluster framework is closely related to the minimization of the graph
cut criterion.
• Experimental Analysis: We perform an extensive experimental evaluation regarding
the efficiency and accuracy of the CoreCluster framework. A large set of experiments were conducted both on synthetic and real-world graphs. This is to evaluate on
ground truth information from the synthetic datasets and on the quality of the clusters from the real-world graphs. The empirical results show that the time complexity
is improved by 3-4 orders of magnitude (compared to a baseline algorithm), especially
for large graphs. Moreover, in the case of graphs with inherent community structure,
the quality of the results is maintained or even improved. In additionally, the initial
experimentation of Giatsidis et al. (2014) is extended here to a wide range of algorithms in order to identify the properties of the graphs and algorithms that optimize
the performance our framework. The CoreCluster framework was introduced on
the basis of studies of degeneracy in real social graphs (Giatsidis et al., 2011b). Given
the interdisciplinary application of graphs, we do not hold claims of universal applicability (i.e., not all types of graphs have the same properties). We conduct here an
extensive evaluation of CoreCluster in order to outline detectable properties that
indicate the conditions under which the proposed framework is useful or meaningless.
The remainder of this paper is organized as follows. Section 2 discusses related work
and Section 3 introduces some preliminary notions. Section 4 formally introduces the
CoreCluster framework. The theoretical analysis and the computational complexity
of the proposed framework are presented in Sections 5 and 6, respectively. Finally, our
extended empirical analysis is presented in Section 7 and we conclude with a discussion
about the advantages of the CoreCluster framework along with suggestions for future
work in Section 8.
2. Related Work
In this section, we review the related work regarding the graph clustering problem, approaches for scaling-up graph clustering and applications of the k-core decomposition.
Graph clustering. The problem of community detection and graph clustering has been extensively studied from several points of view. Some well-known approaches include spectral
3
Giatsidis, Malliaros, Tziortziotis, Dhanjal, Kiagias, Thilikos and Vazirgiannis
clustering (e.g., Ng et al., 2001; Shi and Malik, 2000; White and Smyth, 2005), modularity
optimization (e.g., Newman, 2004; Clauset et al., 2004), multilevel graph partitioning (e.g.,
Metis, Karypis and Kumar, 1998), flow-based methods (Satuluri and Parthasarathy, 2009),
hierarchical methods (Newman and Girvan, 2004) and many more. A very informative
and comprehensive review over the different approaches can be found in Fortunato (2010).
Also, Lancichinetti et al. (2008) have conducted a comparative analysis on the performance
of some of the most recent algorithms, in artificial data produced by their parameterized
generator of benchmark graphs. In our work, we use the same graph generator as in Lancichinetti et al. (2008) to evaluate our framework. Another recent empirical comparison of
community detection algorithms has been performed by Leskovec et al. (2010). There, due
to lack of ground-truth data, the evaluation of the produced clusters is achieved applying
quality measures, such as conductance. As we will present during our experimental analysis (see Section 7), we follow a similar practice to evaluate our framework in the case of
real-world graph data.
Scaling-up graph clustering. The efficiency of graph clustering can be improved in
various ways. Two well-known approaches are the ones of sampling and sparsification. In
the case of spectral clustering, sampling-based approaches include the Nyström method
(Kumar et al., 2009) and randomized SVD algorithm (Drineas et al., 2004). The approach
of Kumar et al. (2009) capitalizes on the Nyström column-sampling method (Williams and
Seeger, 2001) which is an efficient technique to generate low-rank matrix approximations.
More specifically, a novel technique has been proposed that follows a non uniform sampling
of columns of the affinity matrix. Although, Nyström method suffers from high time and
memory complexity (Yan et al., 2009). Drineas et al. (2004) proposed the randomized
SVD algorithm that essentially samples a number of columns of the Gram matrix with
probability proportional to their norms, and performs Principal Component Analysis on
the selected features. Nevertheless, the randomized SVM algorithm may need to sample
a large number of columns in order to get a sufficient small appproximation error in some
cases. In (Yan et al., 2009) a fast approximate algorithm for spectral clustering has been
presented where two different preprocessors have been used in order two reduce the size of
the initial data structure. The first one is the classical k-means algorithm while the other
one is the Random Projection tree. In a nutshell, neighbouring data points correspond into
a set of local representative points and then the spectral algorithm is executed only on the
reduced set of the representative points.
Concerning graph sampling, the goal is to produce a graph of smaller size (nodes and
edges), preserving a set of desired graph properties (e.g., degree distribution, clustering
coefficient) (Leskovec and Faloutsos, 2006). The work by Maiya and Berger-Wolf (2010),
presents a method, based on the notion of expansion properties, to sample a subgraph that
preserves the community structure, i.e., contains representative nodes of the communities.
Then, the community membership of the nodes that do not belong to the sample can be
expressed as an inference problem. Unlike the previous methods that sample both nodes
and edges, the graph sparsification algorithm presented in Satuluri et al. (2011) reduces only
the number of edges (focusing on inter-community edges) in order improve the running time
of a clustering algorithm. In contrast to the aforementioned methodologies, our approach
keeps the structure of the graph intact, without excluding any structural information from
the clustering process.
4
CoreCluster: A core decomposition framework for graph clustering
k-core decomposition. The k-core decomposition constitutes a well-established approach
for identifying particular subsets of a graph, called k-cores, in an hierarchical way, starting
from external vertices and moving to more central ones. The k-cores are important structures in graph theory and their study goes back to the 60s (Erdős, 1963). Seidman (1983)
first applied the k-core decomposition to study the cohesion of social networks. Since then,
k-core decomposition has been applied in several graph-related tasks, such as graph visualization (Alvarez-Hamelin et al., 2005; Zhang and Parthasarathy, 2012), dense subgraph
discovery (Andersen and Chellapilla, 2009) and as an edge ordering criterion for graph coarsening (Abou-Rjeili and Karypis, 2006). Some very recent studies include the extension of
the notion of k-cores to directed graphs (Giatsidis et al., 2011a) and core decomposition in
massive graphs (Cheng et al., 2011).
3. Preliminaries
Given an undirected graph G, we denote by V (G) and E(G) ⊂ V (G) × V (G) the sets of
its vertices and edges, respectively. Given a set S ⊆ V (G), we denote by G[S] the induced
subgraph of G that is obtained if we remove from it all vertices that do not belong in S.
We also define as n and m the number of vertices and edges of graph G, respectively, i.e.,
n = |V (G)| and m = |E(G)|. The neighbourhood of a vertex v ∈ V (G) is denoted by NG (v)
and contains all vertices of G that are adjacent to v. The degree degG (v) of a vertex v in
G, is the number of edges incident to the vertex and is equal to |NG (v)|. The minimum
degree of graph G, denoted by δ(G), is the minimum degree of the vertices in G, i.e.,
δ(G) = min{degG (v) | v ∈ V (G)}.
Given a non-negative integer k, we define the k-core of a graph G, denoted by Gk ,
corek (G), as the maximum size subgraph of G with minimum degree k, where k is the rank
of the core. In a nutshell, in the k-core of a graph G every vertice is adjacent with at least k
vertices. The degeneracy of a graph G is the maximum k for which G contains a non-empty
core. In a more formal way, the degeneracy of a graph is denoted by δ ∗ (G) and is given as:
δ ∗ (G) = max{δ(H) | H ⊆ G}.
(3.1)
Intuitively, the dense cores of a graph, i.e., those whose ranks are close to δ ∗ (G), may serve
as starting seeds for any clustering algorithm as they are expected to preserve the clustering
structure of the original graph. Later in this paper, we make this statement more precise,
providing the necessary theoretical as well as experimental justification.
For a given graph G with δ ∗ (G) = k, we define its core expansion sequence as a sequence
of vertex sets {Vk , Vk−1 , ..., V0 }, computed in a recursive top-down way as follows:
V (corek (G)),
if i = k,
Vi =
(3.2)
V (corei (G)) \ Vi+1 , if i = k − 1, . . . , 0.
We refer to the sets of a core expansion sequence as layers, with set Vi being its ith layer
(see Fig. 1 for an illustrative example).
In general terms, detecting the i-core of a graph G is an easy process (Batagelj and
Zaversnik, 2003): just remove vertices, and edges incident to them, of degree less than
5
Giatsidis, Malliaros, Tziortziotis, Dhanjal, Kiagias, Thilikos and Vazirgiannis
core0(G)
core1(G)
core2(G)
core3(G)
core4(G)
Figure 1: A graph G of degeneracy 4 and its cores. The different colors express the partition
of the vertices of the graph to layers V4 , V3 , V2 , V1 , and V0 . Fat-edges indicate parts of a
clustering of the graph.
i until this is not possible any more. It can be easily verified that O(k · n) steps are
required for the computation of the k-core decomposition. Typically, the degeneracy of
real world graphs is small, making the computation of the core expansion sequence an easy
computational task.
4. Graph Clustering with Degeneracy
In this section, we describe in detail the proposed graph clustering framework. CoreCluster capitalizes on the concept of degeneracy to improve the efficiency of graph clustering.
The main idea behind our approach is that the k-core decomposition preserves the clustering structure of a graph and therefore the “best” k-core subgraph can be used as good
starting point for a clustering method. At the same time, the decomposition provides an
hierarchical organization of the nodes in the graph, that can serve as a “guide” for the
clustering process.
4.1 The CoreCluster Framework
One of the main attributes of the proposed framework, which constitutes our initial motivation, is its ability to accelerate the application of computationally complex clustering
algorithms without any significant loss in their accuracy. Suppose we have a clustering algorithm that takes as input a graph G and outputs a partition of V (G) into a number of sets
that form a clustering of G. As our method could be applied at any clustering algorithm,
we do not further specify the attributes of such an algorithm and we, abstractly, name it
Cluster. Furthermore, we assume that Cluster algorithm runs in O(n3 ) steps. This assumption coincides with the computational complexity of the spectral clusetering algorithm
6
CoreCluster: A core decomposition framework for graph clustering
Algorithm 1 CoreCluster(G)
Input: A graph G.
Output: A partition of V (G) into clusters.
1: k := δ ∗ (G)
2: q := 0
3: Let Vk , . . . , V0 be the core expansion sequence of G
4: Let Gi be the i-core of G, for i = 0, . . . , k
5: Sk = Vk
6: Ak = {C1k , . . . , Cρkk } = Cluster(G[Sk ])
7: for i = k − 1 to 0 do
8:
Si =Select(Gi , Ak ∪ . . . ∪ Ai+1 , Vi )
9:
Ai = (C1i , . . . , Cρi i ) = Cluster(G[Si ])
10: end for
11: Return Ak ∪ · · · ∪ A0
proposed by Ng et al. (2001) and utilized in our initial experiments as well as during the
theoretical analysis of our methodology. In the totality of our experiments (see Section 7),
we study more than the computational complexity and we utilize various algorithms in the
place of Cluster including “fast” algorithms.
In a nutshell, the CoreCluster framework applies the Cluster algorithm at the highest k-core of the graph and then it iterates from the highest to the lowest core applying the
following strategy:
i. assign based on a simple criterion all the nodes that can be assigned to the already
existing clusters, and
ii. apply the Cluster algorithm to the remaining nodes (in order to create new clusters).
A sketch of CoreCluster framework is given in pseudocode in Alg. 1. Initially,
CoreCluster performs k-core decomposition to obtain the core expansion sequence of
the graph. Then, algorithm Cluster is applied to the k-core subgraph, creating the first
set of ρk clusters. The procedure Select (see Section 4.3, for a detailed description) takes
as input the, so far, created clusters, i.e., the sets in Fi+1 = Ak ∪ · · · ∪ Ai+1 and the ith
layer Vi , and tries to assign each of the vertices of Vi in some cluster in Ak ∪ · · · ∪ Ai+1 .
After this update, the procedure Select returns the unassigned vertices. The choice of the
selection procedure considers the way under which the vertices of Vi are adjacent with the
vertices of the clusters in Ak ∪ · · · ∪ Ai+1 . This selection can be done by using several
heuristic approaches, and next we describe such a procedure. Afterwards, the Cluster is
applied to the unassigned vertices returned from the Select procedure. The whole procedure
is implemented for each layer (Vi ) in a recursive manner, starting from the highest level.
4.2 Spectral Clustering
It should be stressed that the CoreCluster can be essentially seen as a “meta-algorithmic
procedure” that can be applied to any graph clustering algorithm, accelerating the execution
and preserving the clustering structure of the latter. The discussion that follows in the
7
Giatsidis, Malliaros, Tziortziotis, Dhanjal, Kiagias, Thilikos and Vazirgiannis
experimental analysis (Section 7) of the paper argues that this indeed can accelerate a highdemanding clustering algorithm without any significant expected loss in its performance.
The spectral clustering algorithm, proposed by Ng et al. (2001), is one of the most
important and well-known algorithms in data mining and constitutes the initial motivation
of the proposed framework, since its computational complexity restrict its applicability to
large graph networks. Under this prism, we use spectral clustering as the baseline and the
basis for the theoretical analysis of the CoreCluster framework (algorithm Cluster).
The central idea of the spectral clustering algorithm is to keep the top ρ eigenvectors of the
normalized Laplacian matrix, defined as:
L = D−1/2 (D − W)D−1/2 = I − D−1/2 WD−1/2 ,
(4.1)
where W ∈ Rn×n and D ∈ Rn×n are the adjacency and the degree matrix, respectively. The
adjacency matrix gives the similarity between pairs of graph vertices, such that Wij > 0
if there is an edge from vertex i to vertex j and zero (Wij = 0) otherwise. If the graph is
undirected, as happens in our case, the adjacency matrix is symmetric (Wij = Wji ). The
degree P
matrix is a diagonal matrix containing the degrees of the nodes on its diagonal, e.g.,
Dii = j Wij . Afterward, assign the corresponding eigenvectors as columns to a matrix
Y ∈ Rn×ρ and apply k-means clustering on the rows of Y . Thereafter, each row of the new
formed matrix corresponds to a data point, while the parameter ρ identifies the number of
clusters that we are looking for.
It has been shown that the performance of the k-means is strongly affected by the initialization scheme (Bubeck et al., 2012). Until now, various approaches have been proposed
for the initialization of k-means algorithm (Arthur and Vassilvitskii, 2007; Redmond and
Heneghan, 2007; Lu et al., 2008). During our empirical analysis (Section 7), we are using the
k-means++ algorithm proposed by Arthur and Vassilvitskii (2007), instead of the classical
k-means algorithm, taking advantage of performing better seeding during the initialization
process. Moreover, since we desire to have an automatic choice of ρ, we define it by noticing
the “sudden drop” in the eigenvalues as it is suggested by Polito and Perona (2001).
4.3 Selection procedure
In this section, we give a detailed description of the selection procedure (Select) presented
in Line 8 of Alg. 1. The procedure takes as input the, so far, created clustering Fi+1 =
Ak ∪ · · · ∪ Ai+1 and the vertex set Vi . Then, this procedure assigns some of the vertices of
Vi to the clusters in Fi+1 and outputs the remaining ones.
For simplicity purposes, we will call the tuple (G, F, V ) as a candidate triple, where G
is a graph and F ∪ {V } is a partition of V (G). Thus, given a candidate triple (G, F, V ), we
define the following property over the vertices of V :
Pα,β (v) = ∃C ∈ F :
|NG (v) ∩ V (C)|
≥α
|NG (v)|
and |NG (v)| ≥ β,
(4.2)
where α > 0.5 and β is a positive integer. Notice that, as α > 0.5, the truth of Pα,β (v) can
be certified by a unique set C in F. We call such a set the certificate of verticle v.
The selection procedure first (lines 1–4) tries to assign vertices of V to clusters of F
using the criterion of the property Pα,β that assigns a vertex to a cluster only if the vast
8
CoreCluster: A core decomposition framework for graph clustering
Algorithm 2 Select(G, F, V )
Input: A candidate triple (G, F, V )
Output: A subset S of V and a partition F 0 of V (F) ∪ (V \ S).
1: while Pα,β (v) is true for some v ∈ V do
2:
F ← (F \ {C}) ∪ {C ∪ {v}} where C is the certificate of v.
3:
V ← V \ {v}.
4: end while
5: V 1= NG (V (F))
6: V 2 = V \ (V 1 ∪ F)
7: if V 2 is either empty or an independent set of G, then
8:
F ← Assign(G, F, V, V 1 )
9:
F ←Assign(G, F, V, V 2 )
10:
return ∅
11: else
12:
return V 1 ∪ V 2 .
13: end if
majority of its neighbours belong in this cluster. The quantification of this “vast majority”
criterion is done by the constants α and β. The vertices that cannot be assigned at an
already existing cluster are partitioned into two groups: (i) V 1 contains those that have
neighbours in vertices that are already classified in the clusters of F and (ii) V 2 contains
the rest. As the vertices in V 2 have no neighbours in the clusters, they have at least k
neighbours out of them, it is most likely that they may not enter to any existing cluster in
the future, unless, possibly, they are completely disjoint. If this is not the case, a further
(milder) classification is attempted by the assign procedure that first classifies the vertices
in V 1 in the existing clusters and then we do the same for the vertices in V 2 . It is worth
noting that the last selection procedure has been found useful in our experiments, especially
in the case of cores with low rank (where many independent vertices may appear).
The procedure assign is a heuristic that classifies each vertex to the cluster that has the
majority of its neighbours. Before presenting the Assign routine, given by Alg. 3, let us
first give some definitions. Given a candidate triple (G, F, V ) and a vertex v ∈ V , we define
as span(v) = max{|NG (v) ∩ V (C)| | C ∈ F }. We also define argspan(v) as a minimum
size C ∈ F with the property that |NG (v) ∩ V (C)| = span(v).
A number of illustrative examples about the choices of the selection procedure can be
extracted by Fig. 2, that is an instance of a graph derived from our experimental data (in
particular from the D1 dataset see Sec. 7.1.1). In the particular graph, the 7-core (i.e., the
graph delimited by the red square) consists exclusively of the vertices of the grey cluster.
The set V6 contains all the vertices of the 6-core that are not in the 7-core. All but two
of these vertices are sparsely connected with the grey cluster, while they exhibit a strong
interconnection between them. Therefore, they form the red cluster, while the remaining
two vertices, that have all of their neighbours in the gray cluster, become a member of it.
A similar assignment to the gray and red clusters is happening for the vertices in V5 , i.e.,
the vertices in the 5-core that do not belong in the 6-core. Finally, the remaining vertices
of the graph (the vertices in V4 ) either clearly belong to the existing clusters or are forming
9
Giatsidis, Malliaros, Tziortziotis, Dhanjal, Kiagias, Thilikos and Vazirgiannis
Algorithm 3 Assign(G, F, V, S)
Input: A candidate triple (G, F, V ) and a subset S of V
Output: A partition F
1: while S 6= ∅ do
2:
` = max{span(v) | v ∈ S}
3:
L = {v ∈ S | span(v) = `}
4:
for each v ∈ L do
5:
C 0 = C ∪ {v}, where C = argspan(v)
6:
S ← S \ {v}
7:
F ← (F \ {C}) ∪ {C 0 }
8:
end for
9: end while
10: return F
two new clusters, the ping and the purple ones. The experimental analysis indicates that a
similar clustering behaviour characterizes the entire data set and that it is captured by the
CoreCluster framework.
Figure 2: An example of the operation of the CoreCluster procedure for a portion of
a graph used through our experimental study. This graph consists of the core sequence
members V7 , V6 , V5 , and V4 that are included in the black, green (#), blue (∗), and red (s)
squares respectively. Vertices of different colors correspond to different clusters.
10
CoreCluster: A core decomposition framework for graph clustering
5. Theoretical Analysis of the CoreCluster’s Quality
The intuition behind the CoreCluster framework is that the core expansion sequence
Vk , Vk−1 , . . . , V0 gives a good sense of direction on how to perform clustering in an incremental way. After that, the procedure considers Vk−1 as the remaining vertices of the
(k − 1)-core Gk−1 , and tries to assign them one by one to the already existing clusters
C1k , . . . , Cρkk . The vertices for which an assignment to an already existing cluster is not
possible, form the set Sk−1 and the Cluster is now applied on G[Sk−1 ]. As the algorithm
continues, the existing clusters grow up and the vertices for which this is not possible, are
grouped to new clusters. The fact that this procedure approximates satisfactorily the result
of the application of Cluster to the whole graph is justified by the observation that the
early i-cores (i.e., i-cores where i is close to k) are already dense, and therefore sufficiently
coherent, to provide a good starting clustering that will expand well because of the selection
criterion. In fact, the subgraphs obtained by the k-core decomposition, provide an (1/2)approximation algorithm for the Densest-Subgraph problem (Andersen and Chellapilla,
2009).
In the remaining of this section, a number of theoretical results are presented which confirms that the core expansion sequence offers an reliable way for performing any clustering
methodology incrementally. Our theoretical analysis could be divided on three main axis:
1. Initially, we justify why the k-core subgraph can be supposed as a seed subgraph
for any clustering algorithm, as nodes with a high clustering coefficient are usually
retained at the highest k-core subgraph.
2. Afterwards, based on the perturbation theory, we demonstrate that the difference
between eigenspaces of successive cores is upper bounded. Under this prism, we are
able to analyze the conditions under which the eigenspaces of successive cores vary
significantly with each other.
3. Last but not least, we highlight that the cluster expansion process, which assigns the
vertices of Vi to an already existing cluster by using the selection criterion, is closely
related to minimizing the graph cut criterion.
5.1 Quality of the maximal core as a starting seed
We claim that the decomposition identifies subgraphs that progressively correspond to the
most central regions and connected parts of the graph. Here we show that nodes with high
clustering coefficient in G, are more likely to“survive” at the highest k-core subgraph by
the pruning (k-core decomposition) procedure. At this point, let us describe the clustering
coefficient (Watts and Strogatz, 1998) measure which will be used later through the analysis
of our framework. The clustering coefficient measures the density of triangles in the graph
and indicates the degree to which nodes tend to cluster together. More specifically, the
local clustering coefficient of a vertex is defined as the ratio of the links that the vertex has
with the vertices of its neighbourhood to the total number of edges that could exist among
the vertices within its neighbourhood. On the other hand, the global clustering coefficient
is defined as the mean of the local clustering coefficients over all graph vertices. Our claim
is based on the following theorem.
11
Giatsidis, Malliaros, Tziortziotis, Dhanjal, Kiagias, Thilikos and Vazirgiannis
Theorem 1 (Gleich and Seshadhri (2012)) Let G be a graph with heavy-tailed degree
distribution, and let CG be the (global) clustering coefficient of G. Then, there exists a
dε
k-core in G for k ≥ CG max , where ε < 1 is a constant such that most edges are incident
2
to a node with degree at least dεmax (typically ε = 2/3), where dmax is the maximum degree
of the nodes.
The above theorem implies that graphs with heavy-tailed degree distribution and high
global clustering coefficient CG , have large degeneracy. Next we present our claim for the
relationship between the local clustering coefficient Cv , ∀v ∈ V (G) and the k-core subgraph
justifying the selection of the k-core as good seed subgraph in the clustering procedure.
Corollary 2 Let G be a graph with heavy-tailed degree distribution. The contribution of
each node v ∈ V (G) to the k-core decomposition of the graph is proportional to the local
clustering coefficient Cv .
Proof The global clustering coefficient CG of the entire graph is given by the average of
1P
the local clustering coefficients Cv , ∀v ∈ V (G), i.e., CG =
Cv , where n = |V (G)|.
n v
Then, according to the Theorem 1 we get that:
!
!
dεmax
1X
dεmax
1 dεmax X
k ≥ CG
=
Cv
=
Cv
2
n v
2
n 2
v
| {z }
γ
⇒ k≥γ
X
Cv ,
v
where parameter γ captures global characteristics of the graph (that depend on the total
number of nodes and the maximum degree).
Therefore, nodes with high clustering coefficient (in the original graph) are more likely
to be found in the best (max-core) k-core (k = δ ∗ (G)) subgraph, since they tend to be
more robust to the degeneracy process. Thus, the k-core subgraph can be used as good
starting point (seed subgraph) for the clustering task. Furthermore, we have experimentally
validated the above claim (see Appendix A).
5.2 The Eigenspaces of k-cores
The spectral clustering algorithm of (Ng et al., 2001) is based on the eigenvectors of a
Laplacian matrix and here we study the eigenspaces of successive cores in conjunction
with the CoreCluster approach. The key question we are interested in is under which
conditions the ρ largest eigenvectors of the Laplacian L(i) for the ith core are similar to
the corresponding eigenvectors of L(i+1) for i = 0, . . . , k − 1? If the eigenspaces are indeed
similar then it follows that clustering on a core is a good basis for clustering the entire
graph.
To answer this question we defer to perturbation theory (Stewart and Sun (1990)) so
that we can examine the change of the eigenvalues along with the eigenvectors of a matrix
12
CoreCluster: A core decomposition framework for graph clustering
when a small perturbation is added, corresponding to moving from one core to the next.
In the case of spectral clustering, consider that the perturbation between the Laplacian
matrices (L(i) and L(i+1) ) of successive cores is given by ∆L(i) = L(i) − L(i+1) .
The key result that we will use, is presented by Davis and Kahan (1970), and bounds
the difference between eigenspaces of Hermitian matrices under perturbations.
Theorem 3 (Davis and Kahan (1970)) Let A ∈ Rn×n be a Hermitian matrix with
spectral resolution given by [X1 X2 ]> A[X1 X2 ] = diag(L1 , L2 ), where [X1 , X2 ] is unitary
with X1 ∈ Rn×ρ . Let Z ∈ Rn×ρ have orthonormal columns, M be any Hermitian matrix of
order ρ and define the residual matrix as R = AZ−ZM. Let λ(A) represent the set of eigenvalues of A and suppose that λ(M) ⊂ [α, β] and for some δ > 0, λ(L2 ) ⊂ R\[α − δ, β + δ].
Then, for any unitary invariant norm, we have:
ksinΘ(R(X1 ), R(Z))k ≤
kRk
,
δ
(5.1)
where R(X1 ) is the column space of matrix X1 .
The spectral resolution of A is a transformation of it using the unitary matrix [X1 X2 ]
such that X1 spans the invariant subspace of A. The function sinΘ(·, ·) corresponds to
a diagonal matrix of dimension ρ, composed of sines of the canonical angles between two
subspaces. Roughly speaking, this corresponds to the angles between the bases of the
subspaces and measures how the two subspaces differ (see Stewart and Sun, 1990, for
details).
Then, the following lemma can be easily derived.
Lemma 4 Let A and à = A + E be two n × n Hermitian matrices whose eigendecompositions are given by A = UΛU> and à = ŨΛ̃Ũ> , with corresponding sets of eigenvalues
λ1 ≤ λ2 ≤ · · · ≤ λn and λ̃1 ≤ λ̃2 ≤ · · · ≤ λ̃n . Then the following results hold, where U1 and
Ũ1 are the matrices of the first ρ eigenvectors of A and Ã:
q
P (ρ) >
i λi (E E)
ksinΘ(R(U1 ), R(Ũ1 ))kF ≤
,
(5.2)
λ̃ρ+1 − λρ
σmax (E)
ksinΘ(R(U1 ), R(Ũ1 ))k2 ≤
,
(5.3)
λ̃ρ+1 − λρ
where λ(ρ) (·) is the largest ρ eigenvalues of a matrix, σmax (·) is the largest singular value
of a matrix and λ̃ρ+1 > λρ .
Proof First, note that we can express the eigendecompositions of A and à as:
Λ1 0
A = UΛU> = [U1 U2 ]
[U1 U2 ]>
0 Λ2
Λ̃1 0
à = ŨΛ̃Ũ> = [Ũ1 Ũ2 ]
[Ũ1 Ũ2 ]> .
0 Λ̃2
13
(5.4)
(5.5)
Giatsidis, Malliaros, Tziortziotis, Dhanjal, Kiagias, Thilikos and Vazirgiannis
Then, we can apply Theorem 3 by setting X1 = Ũ1 , X2 = Ũ2 , L1 = Λ̃1 , L2 = Λ̃2 , M = Λ1
and Z = U1 . Computing the residual matrix R gives:
R = ÃU1 − U1 Λ1 = AU1 + EU1 − U1 Λ1 = EU1 .
P (ρ) >
>
The Frobenius norm of the residual is kRk2F = tr(U>
1 E EU1 ) ≤
i λi (E E) where we
have used the Rayleigh-Ritz theorem to bound the trace (Lutkepohl, 1997). With the spectral norm we have kEU1 k2 = σmax (EU1 ) ≤ σmax (E) using a similar process. It is easy to
show that δ is computed as λ̃ρ+1 − λρ provided λ̃ρ+1 > λρ .
It follows that we want to choose a ρ to make the “eigengap” λ̃ρ+1 − λρ as large as
possible whilst having a residual matrix with small maximum eigenvectors. The eigengap
corresponds to the quality of the clusters (see von Luxburg, 2007, for example), while the
perturbation represents the residual eigenvalues whilst moving from one Laplacian to the
next. In the following part of this subsection we will elaborate on the perturbation by
showing how it relates to edges in a graph.
It is worth noting that when we add vertices by going from a core to a lower core, the
dimensions of the eigenspace corresponding to the new vertices will be zero in the higher
core. Therefore, we restrict the analysis of the eigenspace to existing dimensions. In other
words, we are interested in the following matrix:
∆L(i) = L̂(i) − L(i+1) ,
L̂(i)
L(i)
where
is the submatrix of
corresponding to the vertices in the (i +
individual elements of the normalized Laplacian are computed as:
if x = y and deg(vx ) 6= 0
1
1
√
−
if x 6= y and (vx , vy ) ∈ E(G)
Lxy =
deg(vx )deg(vy )
0
otherwise
(5.6)
1)th
core. The
(5.7)
so we have Lxy ∈ [−1, 1] for all x, y. Note that off-diagonal elements are non-positive. This
allows us to introduce our main result.
Theorem 5 Let L(i) ∈ Rni ×ni be the Laplacian matrix corresponding to the ith core of a
graph, and L̂(i) be the submatrix of L(i) corresponding to the vertices in the (i + 1)th core.
Similarly, E (i) is the set of edges in the ith core, and Ê (i) is the subset of E (i) with one
vertex in the (i + 1)th core. Then the following results hold with U1 and Ũ1 as the matrices
of the first ρ eigenvectors of L̂(i) and L(i+1) :
p
|Ji ||i − k|
ksinΘ(R(U1 ), R(Ũ1 ))kF ≤
,
(5.8)
ik(λ̃ρ+1 − λρ )
√
ni+1 |i − k|
ksinΘ(R(U1 ), R(Ũ1 ))k2 ≤
,
(5.9)
ik(λ̃ρ+1 − λρ )
S
where Ji = (vx ,vy )∈∆E (i) NGi (vx , vy ) is the number of changes in the Laplacian, ∆E (i) =
Ê (i) − E (i+1) is the change in the edges, NG (vx , vy ) is the set of edges adjacent to and
including the edge (vx , vy ) in G and λ̃ρ+1 > λρ .
14
CoreCluster: A core decomposition framework for graph clustering
Proof The change in Laplacian elements between cores can be bounded for x 6= y as
1
1
∆L(i)
+q
xy = − p
degGi (vx )degGi (vy )
degGi+1 (vx )degGi+1 (vy )
≥
i−k
,
ik
where Gi = corei (G). It follows that the Frobenius norm of the change in Laplacians is
bounded as:
X
i−k 2
2
(∆L(i)
)
≤
|J
|
k∆L(i) k2F =
.
i
xy
ik
x,y
P
Note also that k∆L(i) k2F = j λ2j (∆L(i) ). The spectral norm can be written in terms of
maximising the Rayleigh quotient:
uT (∆L(i) )2 u
uT u
i−k 2
≤ ni+1
,
ik
k∆L(i) k22 = max
where u ∈ Rni+1 . Putting the pieces together with Lemma 4 gives the required result.
According to this theorem, we can readily show the scenarios in which moving from
the ith core to the (i + 1)th core does not significantly alter the eigenspaces. The i−k
ik
term implies that for a fixed degeneracy k the core transitions close to k result in small
eigenspace perturbations. In the case of Frobenius norm, the eigenspaces grow apart at a
rate proportional to the neighbourhood of edge changes within the (i + 1)th core. On the
other hand, for the spectral norm we see that the eigenspaces diverge according to the size
(ni+1 ) of the (i + 1)th core.
5.3 Cluster Expansion
The analysis so far has considered changes between cores however an important part of
the CoreCluster algorithm is how clusters are grown in a greedy fashion according to
Alg. 2. In this section we show a close relationship between the selection criterion and the
the notion of graph cuts, which motivate many spectral clustering approaches. To see the
connection, first we define the cut of a graph as follows:
X
Wij ,
(5.10)
cut(C1 , C2 ) =
vi ∈C1 ,vj ∈C2
and it corresponds to the number of edges between disjoint subsets C1 , C2 ⊂ V (G). Based
on this definition the cut for a partition into ρ clusters is defined as:
cut(C1 , . . . , Cρ ) =
ρ
X
i=1
15
cut(Ci , C̄i ),
(5.11)
Giatsidis, Malliaros, Tziortziotis, Dhanjal, Kiagias, Thilikos and Vazirgiannis
where C̄ is the set of vertices in G that are not in C. Consider adding a new vertex v to
the graph, then the change in the cut term corresponding to candidate cluster Cj is given
by
∆cutj
= cut(Cj0 , C̄ 0 j ) − cut(Cj , C̄j )
= cut(Cj , C̄j ) + |NG (v) ∩ V (C̄j )| − cut(Cj , C̄j )
= |NG (v) ∩ V (C̄j )|
= |NG (v)| − |NG (v) ∩ V (Cj )|,
where Cj0 = Cj ∪ {v}. The connection to the selection criteria of Eq. 4.2 can be seen by
noting that the minimization of ∆cutj corresponds to choosing a cluster j for the vertex v
which maximises |NG (v) ∩ V (Cj )|. In this sense, the values of α|NG (v)| and β can be seen
respectively as confidence thresholds on the number of neighbours in the chosen cluster and
the degree of the vertex at hand.
6. Complexity
The proposed CoreCluster procedure executes the graph clustering algorithm (Cluster(G))
on the subgraphs induced by the subsets Sk , . . . , S0 . Each Si corresponds to a subset of
vertices of Vi that cannot be assigned to already existing clusters according to the selection
procedure (see Step 8 of Alg. 1). Thus, the selection step makes it possible for some of the
vertices in Vi to be incorporated in an already created cluster, reducing the burden of the
computation of Cluster(G[Si ]).
In this way, it becomes apparent that the speed up of the algorithm is attributed to the
fact that the CoreCluster(G) runs in k + 1 disjoint subgraphs of G instead of G itself.
As the i-th selection phase requires O(|V (Gi )|3 ) steps, we conclude that the running time
of CoreCluster(G) is bounded by
X
X
O(|Si |3 ) ≤
O(|Vi |3 ) ≤ O(k · n3max ),
(6.1)
i=k,...,0
i=k,...,0
where nmax = max{|Vk |,. . . ,|V0 |}. In the above bound, the first equality holds only in the extremal case where no selection occurs during the selection phases. Clearly, the general bound
in Eq. 6.1 is the best possible when |Vk |,. . . ,|V0 | tend to be equally distributed (which would
accelerate the running time by a factor of k 2 ). According to the first inequality of Eq. 6.1
the running time of the algorithm
is proportionaloto (k + 1) · n3max , where
n
n nmax = max{|Sko|,
|V (G)|
. . . , |S0 |}. Let ρG = max |Si | | i = 0, . . . , k and µG = max |V|V(G)|
| i = 0, . . . , k .
i|
Therefore, we observe that ρG ≥ µG . Notice that the discrepancy between ρG and µG
is a measure of the acceleration of the algorithm because of the selection phases.
Concluding, the acceleration of CoreCluster is upper bounded by
X
i=k,...,0
k
3
O(|Si | ) = O 3 · n .
ρG
3
16
CoreCluster: A core decomposition framework for graph clustering
This estimation is purely theoretical and its purpose is to expose the general complexity
contribution of our algorithmic machinery. In practice, the acceleration can be much better
and this also depends on the heuristics that are applied at the selection phase.
7. Experimental Evaluation
The CoreCluster framework is intended to be used with algorithms of high computational
complexity as its’ primal purpose is to “lower” the cost of such algorithms and apply them
only on the “important parts” of a graph. This can be demonstrated simply by utilizing an
expensive algorithm (see Section 4.2) as we do in our experiments. The computational cost
becomes obviously smaller as the sum of the parts (in processing time) is smaller than the
total for algorithms with high complexity (e.g., O(n3 )).
The degeneracy behaviour of graphs might differ along with their other properties (e.g.,
density, etc.). Thus, in this section, we introduce additional experiments to study the
effectiveness of CoreCluster framework under different scenarios. The purpose of these
experiments is to identify in a clear manner the properties of the graphs and the utilized
algorithms that best optimize its’ performance.
7.1 Datasets description
While real networks are the objective, actual datasets lack ground truth which leaves only
evaluation metrics of the quality of clustering as an option and not direct comparison.
On the other hand, artificial networks offer ground truth and a large variety of properties
that can be parameterized to produce different “types” of networks. The evaluation of our
framework is conducted on both real and artificial networks in order to have complete and
decisive results.
7.1.1 Artificial Networks
We exploit the graph generator by Fortunato Lancichinetti et al. (2008) to produce graphs
with a clustering structure which is available to the tester (ground truth). This graph
generator provides a wide range of input parameters. We used the parameters presented on
Table 1 and tuned them accordingly, by considering various combinations in order to get a
wide range of graphs with different features. Thus, the testing of our approach is credible
as it is evaluated in essentially hundreds of graphs with different properties and quality of
the clustering structure. The parameters used are: N is the size of the graph, maxd is the
maximum node degree, mind is the minimum node degree and µ is the mixing parameter
representing the overlapping between clusters, i.e., each node shares a fraction 1 − µ of
its links with the other nodes of its community and a fraction µ with the other nodes of
the network. Graphs produced by the generator contain inherent clusters and the cluster
assignment is offered by the generator, enabling thus usage of these data sets for evaluating
graph clustering algorithms. In order to ensure a thorough and robust evaluation, for each
of the combinations among the parameters, we generated 10 distinct graphs.
Table 1 depicts the various parameters’ values for the three main different settings in
our experiments. It becomes apparent that the most important parameter is the mind , as
it is the one differentiating the overall density of the graph.
17
Giatsidis, Malliaros, Tziortziotis, Dhanjal, Kiagias, Thilikos and Vazirgiannis
D1
D2
D3
maxd
(node max degree)
10%, 30%,
10%, 30%, 50%
200 edges
mind
(node min degree)
∼5
(the absolutely minimum)
7
20
µ
(mixing parameter)
1% – 43%
(in 7 equal steps)
3% – 43%
(in 6 equal steps)
3% – 43%
(in 6 equal steps)
N
(graph size in nodes)
100−3600
100−3600
100−3600
Table 1: Parameters’ values for the artificial graphs.
7.1.2 Real Networks
We also perform evaluations to a subset of the Facebook 100 dataset provided by Traud
et al. (2011). This is a collection of friendship networks of Facebook from 2005, for 100
US Universities (i.e., 100 individual networks). The evaluations were not performed to
the full extend of this dataset as hardware limitation did not allow us to evaluate, with
spectral clustering, networks with more than 13K nodes (the CoreCluster framework
could handle much larger networks). About half of the networks from this dataset were
used for the final evaluation.
7.2 Evaluation Metrics
To evaluate the clustering results we use different metrics for the artificial and the real
(Facebook 100 ) data.
• Artificial networks: Given the ground truth, we measure the quality of the clustering results in terms of the widely used Normalized Mutual Information (NMI,
Manning et al., 2008). NMI measures how “clear” each found cluster is in relevance
to the known clusters (taking into account the size of the cluster as well).
Let Ω = {ω1 , ω2 , . . . , ωk } be the set of detected clusters and C = {c1 , c2 , . . . , cj } the
set of known clusters (ground-truth). Then, NMI is given as follows:
NMI(Ω, C) =
I(Ω, C)
,
[H(Ω) + H(C)]/2
where I is the mutual information:
X X |ωk ∩ cj |
N |ωk ∩ cj |
I(Ω; C) =
log
,
N
|ωk ||cj |
k
(7.1)
(7.2)
j
and H is the Entropy:
H(Ω) = −
X |ωk |
k
18
N
log
|ωk |
N
.
(7.3)
CoreCluster: A core decomposition framework for graph clustering
• Facebook: Since the networks of this dataset lack ground truth, we choose to evaluate
the results with the evaluation criterion of conductance. Given a graph G and a cut
(S, S), conductance is defined as:
P
i∈S,j ∈S
/ Aij
φ(S) =
,
min (a(S), a(S))
P
P
where Aij are the entries in the adjacency matrix A of G and a(S) = i∈S j∈G Aij .
Informally, conductance measures (for a cluster) the ratio of internal to external connectivity. It has been used widely to examine clustering quality (e.g., Leskovec et al.
(2010)) and has a simple and intuitive definition.
Both NMI and Conductance measures takes values in the range of (0, 1). It is also important
to be noticed that:
1. For NMI, higher values are better.
2. For Conductance, lower values are indicating better clustering quality.
7.3 Algorithms for Additional Evaluation
Apart from the spectral clustering algorithm proposed by Ng et al. (2001), we expand the
study of our framework to a wide variety of algorithms. The purpose of this extended
study is to evaluate the clustering performance of CoreCluster various properties these
algorithms might display. At the same time, we examine whether the properties of the
graphs affect the framework in a consistent manner (i.e., the performance is good for the
same “type” of graphs across different algorithms). The latter is justified through this
extended analysis and establishes clear conditions for the use CoreCluster.
We considered algorithms indiscriminately of their computational complexity, focusing
more on how they approach the clustering problem (i.e., what they try to optimize). In the
following, we list in sort the clustering algorithms that have been considered in our analysis.
• InfoMap (Rosvall et al., 2009): An information theoretic approach that tackles the
problem of graph clustering as a compression problem. For random walks over a graph
the probability of flow is converted to a description. By minimizing the description
length of the nodes in the flow’s path, InfoMap optimizes the network’s partitions.
• Leading Eigenvector (Newman, 2006): In order to optimize modularity, this method
uses the eigenvector corresponding to the largest eigenvalue of the modularity matrix.
While this is also a spectral method, the calculation of the leading eigenvector is much
more efficient as it can be done with the power iteration method.
• Fast Greedy (Clauset et al., 2004): This is another modularity optimization algorithm that works in a greedy manner. It follows an hierarchical agglomerative process,
starting from each node being a unique community and joining communities that produce the highest modularity.
• MCL (Van Dongen, 2000): The Markov Cluster (MCL) algorithm follows a simple
iteration of two steps over the transition probability matrix T of graph : i) raise T to
19
Giatsidis, Malliaros, Tziortziotis, Dhanjal, Kiagias, Thilikos and Vazirgiannis
an integer power and ii) raise the values in the cells of the latter resulting matrix in
a real valued power p. The two steps “represent”: i) the probability that a random
walker takes a specific path in p steps and ii) enhancing the probability that two nodes
are connected in with “greater strength” for larger values from step a.
• Metis (Karypis and Kumar, 1998): Metis relies on a multilevel approach where the
graph is coarsened (vertices are unified and edge weights are adjusted accordingly).
Then follows the partitioning of the coarsened graph and the projection of the clusters
back to the original one. It is a well established algorithm with low computational
cost that has been shown to be very effective.
• MultiLevel (Blondel et al., 2008): The term MultiLevel is used to refer to the greedy
optimization algorithm introduced by Blondel et al. (2008). This approach is similar
to Fast Greedy algorithm with the difference that, at each iteration, joined vertices
are replaced with single nodes (similarly to metis).
• SpinGlass (Reichardt and Bornholdt, 2006): A cross-discipline approach, that interprets the community structure of a graph as the spin configuration with the minimum
energy of the spin glass (spin glass: disordered magnet, spin: magnetic orientation,
the spin states represent the community indices). In essence, this algorithm deals
with the adhesion between vertices which is the equivalent of edge betweenness and
provides a different approach to the work of Clauset et al. (2004)
• Walktrap (Pons and Latapy, 2005): The main intuition behind this algorithm is that
short random walks will tend to stay in the same communities. Thus the transition
probability between two vertices will be high if they are in the same community (for
short random walks and adjusting for the degree of a vertex). By using this property
as distance, clustering can be applied with an agglomerative approach (for efficiency).
7.4 Performance
7.4.1 Artificial Data, NMI comparison
As mentioned on the previous sections, we wish to evaluate the performance of our framework in multiple aspects. We start here with the artificial data, comparing the performance
in relevance to the NMI metric. This comparison is separate for each of the three groups
of the artificial data. Moreover, we compare across the parameters of the data generator
(Table 1) as those affect crucial aspects of the graph’s structure (density, cluster clarity,
etc.). These parameters also affect the degeneracy properties of the graph (number of nodes
in the maximum k-core, number of core partitions etc.).
We organize the comparisons per algorithm. For each of the aforementioned clustering algorithms, we display the evaluation of the “original version” (which we will refer
to as Baseline in all comparisons) in comparison to the evaluation for the corresponding
CoreCluster “combination” (CoreCluster utilizing the Baseline in place of the Cluster algorithm as described in Section 4).
Thus for each algorithm we have three figures (one for each of the aforementioned artificial data sets - Section 7.1.1) displaying the NMI clustering evaluation metric in relevance
to the cluster size as well as the mixing parameter (that affects the “clarity” of the clusters).
20
CoreCluster: A core decomposition framework for graph clustering
Baseline
CoreCluster
1.0
0.8
0.6
NMI 0.4
0.2
0.0
0.0
mix 0.1 0.2
ing
par 0.3
am
ete 0.4 0.5
r
4
1.0
0.8
3
1
2
size
graph
0.6
0.4
0.2
x10e3
0.0
0
Baseline
CoreCluster
1.0
0.8
0.6
NMI 0.4
0.2
0.0
0.0
mix 0.1 0.2
ing
par 0.3
am
ete 0.4 0.5
r
4
1.0
0.8
3
1
2
size
graph
0.6
0.4
0.2
x10e3
0.0
0
Baseline
CoreCluster
1.0
1.0
0.8
0.6
NMI 0.4
0.2
0.0
0.0
mix 0.1 0.2
ing
par 0.3
am
ete 0.4 0.5
r
4
0.8
3
1
2
size
graph
0.6
0.4
0.2
x10e3
0.0
0
(a) Spectral Clustering
Baseline
CoreCluster
1.0
0.8
0.6
NMI 0.4
0.2
0.0
0.0
mix 0.1 0.2
ing
par 0.3
am
ete 0.4 0.5
r
4
1.0
0.8
3
1
2
size
graph
0.6
0.4
0.2
x10e3
0.0
0
Baseline
CoreCluster
1.0
0.8
0.6
NMI 0.4
0.2
0.0
0.0
mix 0.1 0.2
ing
par 0.3
am
ete 0.4 0.5
r
4
1.0
0.8
3
1
2
size
graph
0.6
0.4
0.2
x10e3
0.0
0
Baseline
CoreCluster
1.0
0.8
0.6
NMI 0.4
0.2
0.0
0.0
mix 0.1 0.2
ing
par 0.3
am
ete 0.4 0.5
r
4
1.0
0.8
3
1
2
size
graph
0.6
0.4
0.2
x10e3
0.0
0
(b) InfoMap
Baseline
CoreCluster
1.0
0.8
0.6
NMI 0.4
0.2
0.0
0.0
mix 0.1 0.2
ing
par 0.3
am
ete 0.4 0.5
r
4
1.0
0.8
3
1
2
size
graph
0.6
0.4
0.2
x10e3
0.0
0
Baseline
CoreCluster
1.0
0.8
0.6
NMI 0.4
0.2
0.0
0.0
mix 0.1 0.2
ing
par 0.3
am
ete 0.4 0.5
r
4
1.0
0.8
3
1
2
size
graph
0.6
0.4
0.2
x10e3
0.0
0
Baseline
CoreCluster
1.0
0.8
0.6
NMI 0.4
0.2
0.0
0.0
mix 0.1 0.2
ing
par 0.3
am
ete 0.4 0.5
r
4
1.0
0.8
3
1
2
size
graph
0.6
0.4
0.2
x10e3
0.0
0
(c) Leading Eigenvector
Baseline
CoreCluster
1.0
0.8
0.6
NMI 0.4
0.2
0.0
0.0
mix 0.1 0.2
ing
par 0.3
am
ete 0.4 0.5
r
4
1.0
0.8
3
1
2
size
graph
0.6
0.4
0.2
x10e3
0.0
0
Baseline
CoreCluster
1.0
0.8
0.6
NMI 0.4
0.2
0.0
0.0
mix 0.1 0.2
ing
par 0.3
am
ete 0.4 0.5
r
4
1.0
0.8
3
1
2
size
graph
0.6
0.4
0.2
x10e3
0.0
0
Baseline
CoreCluster
1.0
0.8
0.6
NMI 0.4
0.2
0.0
0.0
mix 0.1 0.2
ing
par 0.3
am
ete 0.4 0.5
r
4
1.0
0.8
3
1
2
size
graph
0.6
0.4
0.2
x10e3
0.0
0
(d) Fast Greedy
Baseline
CoreCluster
1.0
0.8
0.6
NMI 0.4
0.2
0.0
0.0
mix 0.1 0.2
ing
par 0.3
am
ete 0.4 0.5
r
4
1.0
0.8
3
1
2
size
graph
0.6
0.4
0.2
x10e3
0.0
0
Baseline
CoreCluster
1.0
0.8
0.6
NMI 0.4
0.2
0.0
0.0
mix 0.1 0.2
ing
par 0.3
am
ete 0.4 0.5
r
4
1.0
0.8
3
1
2
size
graph
0.6
0.4
0.2
x10e3
0.0
0
Baseline
CoreCluster
1.0
1.0
0.8
0.6
NMI 0.4
0.2
0.0
0.0
mix 0.1 0.2
ing
par 0.3
am
ete 0.4 0.5
r
4
0.8
3
1
2
size
graph
0.6
0.4
0.2
x10e3
0.0
0
(e) MCL
Figure 3: Clustering quality comparisons in terms of NMI for the three artificial graphs
(D1, D2, and D3) (Part I).
We start this comparison at Figure 3 with the Spectral Clustering algorithm. In this
case, we observe that the CoreCluster solution performs quite close to the Baseline and
21
Giatsidis, Malliaros, Tziortziotis, Dhanjal, Kiagias, Thilikos and Vazirgiannis
Baseline
CoreCluster
1.0
0.8
0.6
NMI 0.4
0.2
0.0
0.0
mix 0.1 0.2
ing
par 0.3
am
ete 0.4 0.5
r
4
1.0
0.8
3
1
2
size
graph
0.6
0.4
0.2
x10e3
0.0
0
Baseline
CoreCluster
1.0
0.8
0.6
NMI 0.4
0.2
0.0
0.0
mix 0.1 0.2
ing
par 0.3
am
ete 0.4 0.5
r
4
1.0
0.8
3
1
2
size
graph
0.6
0.4
0.2
x10e3
0.0
0
Baseline
CoreCluster
1.0
1.0
0.8
0.6
NMI 0.4
0.2
0.0
0.0
mix 0.1 0.2
ing
par 0.3
am
ete 0.4 0.5
r
4
0.8
3
1
2
size
graph
0.6
0.4
0.2
x10e3
0.0
0
(f) Metis
Baseline
CoreCluster
1.0
0.8
0.6
NMI 0.4
0.2
0.0
0.0
mix 0.1 0.2
ing
par 0.3
am
ete 0.4 0.5
r
4
1.0
0.8
3
1
2
size
graph
0.6
0.4
0.2
x10e3
0.0
0
Baseline
CoreCluster
1.0
0.8
0.6
NMI 0.4
0.2
0.0
0.0
mix 0.1 0.2
ing
par 0.3
am
ete 0.4 0.5
r
4
1.0
0.8
3
1
2
size
graph
0.6
0.4
0.2
x10e3
0.0
0
Baseline
CoreCluster
1.0
0.8
0.6
NMI 0.4
0.2
0.0
0.0
mix 0.1 0.2
ing
par 0.3
am
ete 0.4 0.5
r
4
1.0
0.8
3
1
2
size
graph
0.6
0.4
0.2
x10e3
0.0
0
(g) MultiLevel
Baseline
CoreCluster
1.0
0.8
0.6
NMI 0.4
0.2
0.0
0.0
mix 0.1 0.2
ing
par 0.3
am
ete 0.4 0.5
r
4
1.0
0.8
3
1
2
size
graph
0.6
0.4
0.2
x10e3
0.0
0
Baseline
CoreCluster
1.0
0.8
0.6
NMI 0.4
0.2
0.0
0.0
mix 0.1 0.2
ing
par 0.3
am
ete 0.4 0.5
r
4
1.0
0.8
3
1
2
size
graph
0.6
0.4
0.2
x10e3
0.0
0
Baseline
CoreCluster
1.0
0.8
0.6
NMI 0.4
0.2
0.0
0.0
mix 0.1 0.2
ing
par 0.3
am
ete 0.4 0.5
r
4
1.0
0.8
3
1
2
size
graph
0.6
0.4
0.2
x10e3
0.0
0
(h) SpinGlass
Baseline
CoreCluster
1.0
0.8
0.6
NMI 0.4
0.2
0.0
0.0
mix 0.1 0.2
ing
par 0.3
am
ete 0.4 0.5
r
4
1.0
0.8
3
1
2
size
graph
0.6
0.4
0.2
x10e3
0.0
0
Baseline
CoreCluster
1.0
0.8
0.6
NMI 0.4
0.2
0.0
0.0
mix 0.1 0.2
ing
par 0.3
am
ete 0.4 0.5
r
4
1.0
0.8
3
1
2
size
graph
0.6
0.4
0.2
x10e3
0.0
0
Baseline
CoreCluster
1.0
0.8
0.6
NMI 0.4
0.2
0.0
0.0
mix 0.1 0.2
ing
par 0.3
am
ete 0.4 0.5
r
4
1.0
0.8
3
1
2
size
graph
0.6
0.4
0.2
x10e3
0.0
0
(i) Walktrap
Figure 3: Clustering quality comparisons in terms of NMI for the three artificial graphs
(D1, D2, and D3) (Part II).
in fact outperforms it in many occasions. We further make an initial note on the effect of
graph density to the quality of the results for our framework. Thus, we can remark that
the largest gap between the Baseline and CoreCluster across all algorithms is observed
for the D3 dataset, which has the biggest density.
This is an expected result as the global density of the graph affects the maximum k-core
number and the partitions generated by the k-core decomposition algorithm. Graphs that
are very dense will have a high maximum core number but there will be no partition of
significance. In simpler terms, when all the nodes are highly connected, they are all likely
to belong in high-core partitions and subsequently the biggest portion of the graph could
22
CoreCluster: A core decomposition framework for graph clustering
belong to the maximum core. The case of the D3 dataset displays this point as it was
intentionally generated with high density. On the contrary, D1 was generated as a case
closer to the density of real graphs and D2 was generated as an intermediate case. We see
consistently along all experiments that the general graph density affects more or less the
same our framework. The cases in D3 where CoreCluster performs in a similar manner
with the Baseline are of those with low mixing parameter. The reason for this is twofold:
1. The clusters are better separated from each other.
2. While the minimum and the maximum node degree are the main parameters that
affect directly the density of the generated graph, the mixing parameter contributes
as well. Thus, for low mixing parameters, we can get lower densities as well in all of
the datasets.
Our framework is designed based on analysis of real networks and their Degeneracy
(Giatsidis et al., 2011b,c) and for this reason datasets with “abnormal” behaviour in the
Degeneracy aspect are expected to give bad evaluations. Intuitively, in order to provide
acceleration to the “input” algorithm (Cluster), the size of corei should be (significantly)
smaller than the size of corei−1 . Otherwise, the incremental nature of our framework would
not make sense.
Moving to the rest of the algorithms, we see similarly that CoreCluster performs
quite close to the Baseline for most cases. In addition to the cases of high density, we
would like to study in detail other parameters that also affect the divergence between the
two algorithms(the Baseline and CoreCluster). For this reason in the following sections,
we examine in depth properties of both the graph and the Baseline algorithms in order to
clearly outline the evaluation of the CoreCluster framework.
It is also worth noting that the SpinGlass algorithm is missing evaluations for the larger
graphs in the Baseline part. Its execution was conducted with the implementation provided
by igraph1 and it was riddled with abruptly terminations on these datasets. Only a small
portion of the data was processed successfully with just the Baseline and for this reason
we present only the parts that were covered thoroughly (i.e., more than one graph per
combination passed the processing with SpinGlass).
7.4.2 Analysis: Graph and Algorithm Properties - Artificial Data
In this part of our work, we present a thorough analysis of the evaluations in regards to the
following two aspects:
• Performance analysis in relevance to the coverage of the maximum-core:
How much of the complete graph exists in the max-core versus the NMI from CoreCluster.
• “Final Performance” vs “Performance at the Maximum-Core”: The relation of the performance of the Baseline on the max-core versus that of CoreCluster
at the entire graph.
Similarly to the performance evaluations presented in Section 7.4.3, the aforementioned
evaluations are conducted per Dataset (D1,D2,D3).
1. http://igraph.org/c/
23
Giatsidis, Malliaros, Tziortziotis, Dhanjal, Kiagias, Thilikos and Vazirgiannis
How to use the information bellow
Most tools and algorithms in the Machine Learning domain require a lot of experimentation
and fine tuning before deployment. Since CoreCluster is a meta algorithmic framework,
we are limiting the additional complexity in this process by clearly marking properties that
are easy and efficient to detect before the deployment of our framework. Instead of general
remarks on the performance, the next two sections will provide a clear roadmap to anyone
who wishes to apply CoreCluster on any data.
Performance analysis in relevance to the coverage of the maximum-core The
degeneracy properties of a graph can vary greatly in two main aspects. The first one is the
k-core number of the graph, i.e., the maximum of interconnectivity for which we can get a
non empty part of the graph. The other one is the number of nodes for all the possible cores
from 0 to k. As graphs with communities are complex structures with dense and sparse
areas, it would be disorienting to evaluate all possible properties for a graph in relevance
to Degeneracy and the performance of CoreCluster. For example, while density affects
the core numbers, a graph with a high k-core value could be sparse as well. For this reason,
we focus on a key aspect that can portray the graph in a major manner. This aspect is the
percentage of the graph that is maintained in the maximal k-core. We call this dimension
“max-core coverage” and we examine across it how well our framework performs (in terms
of NMI).
In Figure 4, we see the comparison between the max-core coverage and the NMI of our
framework in the entire graph (referred here as final NMI). From this comparison, there
is clearly a negative correlation between the two. This correlation is consistent across all
algorithms which points out that it is a property of the graph that affects the performance
of CoreCluster. In simple terms, our framework performs better when the max-core
contains a small percentage of the total graph (and performs worse in the opposite case).
This makes sense in an intuitive manner as well as the reasons for having a high percentage of the graph in the max-core coincide with those of having bad or no clustering
structures. Regardless of the general density of the graph, the main reason for getting a
high “max-core coverage”, is that the distribution of edges among the vertices is close to
uniform.
“Final Performance” vs “Performance at the Maximum-Core” Our proposed
framework works in an incremental manner starting a cluster from the most cohesive part
and adding on top of that. Following this hierarchical manner it is susceptible to errors
propagating from the starting basis to the final clustering result. Thus, it is important to
evaluate an algorithm at the most cohesive part in order to anticipate the behaviour at the
entirety of the graph.
In this section, we would like to study properties of the algorithm and in particular the
behaviour of our framework in regards with the starting core (which is the application of the
Baseline on the max-core). Specifically we would like to see how the “starting” performance
at the maximum core affects the final performance in the entire graph. The performance at
the maximum core is evaluated in the same manner as it would be for the entire graph but
only for the sub-graph of the maximum core. We assume that:
• the maximum core is the input graph,
24
CoreCluster: A core decomposition framework for graph clustering
D1
1.0
0.9
0.8
0.7
0.6
0.5
0.4
0.3
0.2
0.1
0.0
D2
D3
final NMI
Percentage over all Graphs
0.200
0.175
0.150
0.125
0.100
0.075
0.050
0.025
0.0
0.5
1.0 0.0
0.5
1.0 0.0
0.5
1.0
max-core coverage (percentage of the graph)
1.0
0.9
0.8
0.7
0.6
0.5
0.4
0.3
0.2
0.1
0.0
D2
D3
(a) Spectral Clustering
0.5
1.0 0.0
0.5
1.0 0.0
0.5
1.0
1.0
0.9
0.8
0.7
0.6
0.5
0.4
0.3
0.2
0.1
0.0
0.5
1.0 0.0
0.5
1.0 0.0
0.5
0.5
1.0 0.0
0.5
1.0
D1
(c) Fast Greedy
D2
D3
0.16
0.14
0.12
0.10
0.08
0.06
0.04
0.02
final NMI
Percentage over all Graphs
0.18
0.16
0.14
0.12
0.10
0.08
0.06
0.04
0.02
0.0
1.0 0.0
Percentage over all Graphs
D3
final NMI
1.0
0.9
0.8
0.7
0.6
0.5
0.4
0.3
0.2
0.1
0.0
0.5
max-core coverage (percentage of the graph)
(b) Leading Eigenvector
D2
D3
0.14
0.12
0.10
0.08
0.06
0.04
0.02
0.0
max-core coverage (percentage of the graph)
D1
D2
final NMI
final NMI
Percentage over all Graphs
0.14
0.12
0.10
0.08
0.06
0.04
0.02
0.0
D1
1.0
0.9
0.8
0.7
0.6
0.5
0.4
0.3
0.2
0.1
0.0
Percentage over all Graphs
D1
1.0
0.0
max-core coverage (percentage of the graph)
0.5
1.0 0.0
0.5
1.0 0.0
0.5
1.0
max-core coverage (percentage of the graph)
(d) MCL
(e) Metis
Figure 4: Performance analysis in relevance to the coverage of the maximum-core for the
three artificial graphs (D1, D2, and D3) (Part I)
25
Giatsidis, Malliaros, Tziortziotis, Dhanjal, Kiagias, Thilikos and Vazirgiannis
D3
0.5
1.0 0.0
0.5
1.0 0.0
0.5
1.0
D2
1.0
0.9
0.8
0.7
0.6
0.5
0.4
0.3
0.2
0.1
0.0
0.5
1.0 0.0
0.5
1.0 0.0
0.5
1.0 0.0
0.5
1.0 0.0
0.5
1.0
D1
(g) SpinGlass
D2
D3
0.135
0.120
0.105
0.090
0.075
0.060
0.045
0.030
0.015
final NMI
Percentage over all Graphs
0.18
0.16
0.14
0.12
0.10
0.08
0.06
0.04
0.02
0.0
0.5
max-core coverage (percentage of the graph)
D3
final NMI
1.0
0.9
0.8
0.7
0.6
0.5
0.4
0.3
0.2
0.1
0.0
(f) MultiLevel
D3
0.135
0.120
0.105
0.090
0.075
0.060
0.045
0.030
0.015
0.0
max-core coverage (percentage of the graph)
D1
D2
final NMI
final NMI
Percentage over all Graphs
0.16
0.14
0.12
0.10
0.08
0.06
0.04
0.02
0.0
D1
1.0
0.9
0.8
0.7
0.6
0.5
0.4
0.3
0.2
0.1
0.0
Percentage over all Graphs
D2
Percentage over all Graphs
D1
1.0
0.9
0.8
0.7
0.6
0.5
0.4
0.3
0.2
0.1
0.0
1.0
0.0
max-core coverage (percentage of the graph)
0.5
1.0 0.0
0.5
1.0 0.0
0.5
1.0
max-core coverage (percentage of the graph)
(h) InfoMap
(i) Walktrap
Figure 4: Performance analysis in relevance to the coverage of the maximum-core for the
three artificial graphs (D1, D2, and D3) (Part I)
• the cluster membership is maintained,
and we simply apply any of the given clustering algorithms. The maximum core contains
a very dense part of the graph (Andersen and Chellapilla, 2009). Moreover, it contains the
most cohesive parts of the clusters in respect to interconnectivity (i.e., a sub-graph where
the number of neighbours k is -at least- the maximal one for all).
In our analysis of the results on the artificially generated data, we noticed that almost
50% of the data contain only one cluster. Many graph clustering algorithms are implemented
or originally designed with the assumption that the graph contains multiple clusters. In
the current state of our framework, we apply the clustering algorithm in the maximum core
without conducting any analysis on the number of the clusters. While the assumption that
there will always be more than one cluster makes sense in general cases, in this specific
scenario we see two distinctive behaviors.
Therefore, in order to study the effect of the performance at the maximum core, we
partition our cases to data that have more that one cluster at the maximum core and data
that have only one. For the first case, we report the relation of the NMI at the maximum
26
CoreCluster: A core decomposition framework for graph clustering
D1
1.0
0.9
0.8
0.7
0.6
0.5
0.4
0.3
0.2
0.1
0.0
D2
D3
final NMI
Percentage over all Graphs
0.18
0.16
0.14
0.12
0.10
0.08
0.06
0.04
0.02
1.0
0.9
0.8
0.7
0.6
0.5
0.4
0.3
0.2
0.1
0.0
D2
1.0 0.0
0.5
D3
1.0
0.9
0.8
0.7
0.6
0.5
0.4
0.3
0.2
0.1
0.0
1.0 0.0
0.5
1.0 0.0
max-core NMI
0.5
1.0
0.0
(b) Leading Eigenvector
D2
D3
0.0
1.0
0.9
0.8
0.7
0.6
0.5
0.4
0.3
0.2
0.1
0.0
0.5
1.0 0.0
0.5
1.0 0.0
max-core NMI
0.5
D1
D2
D3
0.050
0.045
0.040
0.035
0.030
0.025
0.020
0.015
0.010
0.005
0.5
D1
1.0 0.0
0.5
1.0 0.0
max-core NMI
(c) Fast Greedy
D2
0.5
1.0
D3
0.045
0.040
0.035
0.030
0.025
0.020
0.015
0.010
0.005
final NMI
Percentage over all Graphs
0.16
0.14
0.12
0.10
0.08
0.06
0.04
0.02
final NMI
1.0
0.9
0.8
0.7
0.6
0.5
0.4
0.3
0.2
0.1
0.0
D1
1.0
final NMI
Percentage over all Graphs
final NMI
0.5
0.5
(a) Spectral Clustering
0.036
0.032
0.028
0.024
0.020
0.016
0.012
0.008
0.004
0.0
1.0 0.0
max-core NMI
Percentage over all Graphs
D1
0.5
Percentage over all Graphs
0.0
1.0
0.0
(d) MCL
0.5
1.0 0.0
0.5
1.0 0.0
max-core NMI
0.5
1.0
(e) Metis
Figure 5: “Final Performance” vs “Performance at the Maximum-Core” for the three artificial graphs (D1, D2, and D3) (Part I).
27
Giatsidis, Malliaros, Tziortziotis, Dhanjal, Kiagias, Thilikos and Vazirgiannis
D3
1.0
0.9
0.8
0.7
0.6
0.5
0.4
0.3
0.2
0.1
0.0
D1
1.0 0.0
0.5
1.0 0.0
max-core NMI
(f) MultiLevel
D2
0.5
1.0
0.0
0.0
D3
1.0
0.9
0.8
0.7
0.6
0.5
0.4
0.3
0.2
0.1
0.0
0.16
0.14
0.12
0.10
0.08
0.06
0.04
0.02
0.5
1.0 0.0
0.5
1.0 0.0
max-core NMI
0.5
D3
0.036
0.032
0.028
0.024
0.020
0.016
0.012
0.008
0.004
0.5
D1
1.0 0.0
0.5
1.0 0.0
max-core NMI
(g) SpinGlass
D2
0.5
1.0
D3
0.064
0.056
0.048
0.040
0.032
0.024
0.016
0.008
final NMI
final NMI
1.0
0.9
0.8
0.7
0.6
0.5
0.4
0.3
0.2
0.1
0.0
0.5
Percentage over all Graphs
0.0
D2
final NMI
final NMI
Percentage over all Graphs
0.056
0.048
0.040
0.032
0.024
0.016
0.008
D1
Percentage over all Graphs
D2
Percentage over all Graphs
D1
1.0
0.9
0.8
0.7
0.6
0.5
0.4
0.3
0.2
0.1
0.0
1.0
0.0
(h) InfoMap
0.5
1.0 0.0
0.5
1.0 0.0
max-core NMI
0.5
1.0
(i) Walktrap
Figure 5: “Final Performance” vs “Performance at the Maximum-Core” for the three artificial graphs (D1, D2, and D3) (Part II).
core versus the NMI on the entire graph. In the second case, since the evaluation at the
maximum core is meaningless, we report only the final evaluation on the entire graph.
Starting with Figure 5 (multiple clusters at the maximum core), we see a consistent
behaviour:
1. There is an obvious correlation between the performance at the maximum core and
that at the entire graph.
2. Combinations of datasets and clustering algorithms that CoreCluster was very
close to the Baseline, are cases where the latter was performing relatively well on the
max-core.
3. The performance at D1 dataset is consistently better. This validates that D1 contains
structures which most algorithms would identify as clusters.
One example, of the second observation, is the case of the algorithm “Leading EigenVector”
in D1 where the performance of our framework is identical to that of the Baseline. At the
same time, Figure 5 indicates that this algorithm was really good at detecting the clusters
28
leading
eigenvector
spinglass
infoMap
label
propagation
metis
multilevel
spectral
clustering
fast
greedy
walktrap
edge
betweenness
mcl
D1
D2
D3
0.16
0.14
0.12
0.10
0.08
0.06
0.04
0.02
Percentage over all Graphs
Algorithm
CoreCluster: A core decomposition framework for graph clustering
0.0 0.2 0.4 0.6 0.8 1.0 0.0 0.2 0.4 0.6 0.8 1.0 0.0 0.2 0.4 0.6 0.8 1.0
NMI
Figure 6: Final NMI: one cluster at maximum core.
at the maximum core. On the other hand, “InfoMap” is an example of an algorithm that
had “failures” at the maximum core which would explain the greater divergence.
Additionally, we examine Figure 6 to provide further insight. With “Walktrap” as a
counter example, we see that this algorithm behaves quite similarly to “Leading EigenVector” in the case where we have multiple clusters in the maximum core. But the NMI
evaluation is not that similar. We attribute this to the other 50% of the data where we have
only one cluster (at the maximum core). In this case, “Walktrap” performs distinctively
worse. Upon close examination this algorithm was less capable at detecting that there is
only one cluster in the maximum core (in comparison to other algorithms). We can clearly
see the result of this in Figure 6.
7.4.3 Real Data, Conductance comparison
The previous sections offered a great detail on the inner workings of our framework and gave
an easy set of techniques to determine how appropriate is the application of our framework
without having to perform exhaustive evaluations. Given that the artificial data range over
a wide variety of properties, the results on the performance (NMI) might not always reflect
the performance on real data. Simply put, the artificial data were intentionally created to
stress test our framework in even extreme cases.
In this section, we apply CoreCluster into a real dataset along with all of the aforementioned algorithms, and we evaluate its performance with the well know conductance
metric. In this manner, we provide a complete picture of how well our framework performs
and we justify its’ usefulness with real and concrete examples.
In Figure 7, the comparison between the aforementioned “Baselines” (Section 7.3) and
the corresponding CoreCluster “combination” is presented. Given that the Facebook
dataset is composed by 100 different graphs, we summarize the clustering performance in
the following two aspects:
1. Conductance per cluster size: The upper “parts” of the sub-figures in Figure 7 represents the comparison of the clusters found (along all the graphs) in terms of conductance (lower values are better). While conductance reflects in both a practical
and intuitive manner the quality of a cluster, the size of the clusters found plays
29
Giatsidis, Malliaros, Tziortziotis, Dhanjal, Kiagias, Thilikos and Vazirgiannis
1.0
CoreCluster
Baseline
Bin size
Conductance
0.8
0.6
80
70
60
50
0.4
40
0.2
30
0.0
20
Baseline
10
core
cluster
10 1
0
10 4
10 2
10 3
Cluster size (number of nodes)
(a) Spectral Clustering
CoreCluster
Baseline
Bin size
0.6
0.4
0.2
0.0
Baseline
core
cluster
10 1
10 4
10 2
10 3
Cluster size (number of nodes)
32
0.2
24
0.0
16
Baseline
10 1
2400
1800
1200
600
10 4
10 2
10 3
Cluster size (number of nodes)
CoreCluster
Baseline
0.8
Conductance
3000
core
cluster
0
10 4
10 2
10 3
Cluster size (number of nodes)
1.0
4200
0.4
Baseline
8
core
cluster
4800
3600
0.0
40
(c) Fast Greedy
0.6
0.2
48
0.4
10 1
Bin size
Bin size
Conductance
0.8
CoreCluster
Baseline
56
0.6
(b) Leading Eigenvector
1.0
CoreCluster
Baseline
0.8
Conductance
Conductance
0.8
1.0
90
80
70
60
50
40
30
20
10
0
Bin size
1.0
0
0.6
(d) MCL
150
100
0.2
75
0.0
50
Baseline
25
core
cluster
10 4
10 2
10 3
Cluster size (number of nodes)
(e) Metis
Figure 7: Facebook conductance (Part I).
30
175
125
0.4
10 1
200
0
CoreCluster: A core decomposition framework for graph clustering
CoreCluster
Baseline
0.6
120
0.8
105
90
75
0.4
60
0.2
45
0.0
30
Baseline
15
core
cluster
10 1
1.0
Bin size
Bin size
Conductance
0.8
135
Conductance
1.0
10 2
10 4
10 3
0
0.6
0.4
0.2
0.0
Baseline
core
cluster
10 1
Cluster size (number of nodes)
(f) MultiLevel
Bin size
0.6
0.4
0.2
0.0
Baseline
core
cluster
10 1
10 4
10 2
10 3
Cluster size (number of nodes)
1.0
2250
2000
1750
1500
1250
1000
750
500
250
0
CoreCluster
Baseline
0.8
Conductance
Conductance
0.8
CoreCluster
Baseline
10 4
10 2
10 3
Cluster size (number of nodes)
(g) SpinGlass
Bin size
1.0
45
40
35
30
25
20
15
10
5
0
CoreCluster
Baseline
0.6
(h) InfoMap
105
90
75
0.4
60
0.2
45
0.0
30
Baseline
15
core
cluster
10 1
120
10 4
10 2
10 3
Cluster size (number of nodes)
0
(i) Walktrap
Figure 7: Facebook conductance (Part II).
an important role (e.g., we can’t consider a clustering “good” if there is one “huge”
cluster with high conductance and many “small” ones with low). On an additional
note, we excluded from the comparison small non connected components as those are
trivial to be detected by any algorithm and don’t provide additional information for
the comparison.
2. Frequency of the bin (bin size): For presentation purposes, the aforementioned comparison is done by binning the cluster sizes in logarithmic bins. This is to avoid noise
in the plot and to provide a better picture on the average behaviour. In this point, we
evaluate the frequency of the bins to further investigate the clustering “tendencies”
of the Baselines and their combinations. On the plots we refer to the frequency of the
bin as Bin size.
With the exception of “MultiLevel”, we see that CoreCluster either outperforms the
corresponding Baselines or has an almost identical quality of results. Regarding the sizes
of the clusters, we see from the frequency of the bins that some of the Baselines (Spinglass,
InfoMap, Metis, MCL,) are more prone to a specific range of sizes. On the other hand,
CoreCluster framework maintains a good clustering quality while being able to detect
31
Giatsidis, Malliaros, Tziortziotis, Dhanjal, Kiagias, Thilikos and Vazirgiannis
cluster in all possible ranges indiscriminately of the “input” algorithm. Evidently, this is
important as it is not an attribute that all algorithms can demonstrate.
7.4.4 Execution Time
As a final part of our evaluation, we present the average execution time of the Baselines
in comparison to the corresponding one when CoreCluster is applied. We once more
clarify that not all of the Baseline algorithms are appropriate for “acceleration” with this
framework. As mentioned at Section 4, CoreCluster is designed to accelerate a computationally expensive algorithm. For the cases of “complex” algorithms, we can see in
Figures 8 and 9 the examples of Spectral Clustering, SpinGlass, Walktrap and MCL algorithms. In these examples, we see orders of magnitude in the difference between the
Baseline and CoreCluster that increases exponentially for greater graph sizes.
In the case of the algorithms that are not as computationally expensive there is the added
overhead to compute the core structure with the k-core algorithm. This overhead makes
the CoreCluster slower but we see that the difference is great as the k-core algorithm
can be linear to the number of edges (Batagelj and Zaversnik, 2003). This means that the
additional cost will not increase immensely for greater graph sizes but, as we also observe
in the plots, it will remain within a practically logical range.
Additionally, we note that CoreCluster does not require the full graph in memory
as it processes parts of the graph incrementally. Added with the fact that CoreCluster
can outperform the Baseline (e.g., Metis for conductance at Figure 7 ) we can envision
CoreCluster being applicable with “fast” algorithms as well.
8. Conclusions
In this paper, we have presented in full detail a framework for optimizing the efficiency
of graph clustering. This framework, is capitalizing on the structures produced by the kcore decomposition algorithm. The intuition is that the extreme k-cores maintain the most
crucial parts of the clustering structure of the original graph. Our main contribution is the
CoreCluster framework that processes the graph in an incremental manner and in parts
of small size (compared to the entire graph).
We have described how this framework scales-up from an analytical view and why the kcore structure offers an appropriate partition with increased “importance” (to the clustering
structure) of vertices. We displayed both of these in an experimental manner as well with a
multitude of evaluations. In more detail, we have presented an exhaustive set of experiments
on real and artificial data which:
• Displayed the acceleration of computationally expensive algorithms while maintaining
high cluster quality.
• Through thorough analysis, described a set of properties -easy to identify and detectfor the proper usage of our framework.
Moreover, we provided an indirect review of the used algorithms and offered insights on
their individual performance.
As future work, we plan to further extend and improve our framework in the following
aspects:
32
CoreCluster: A core decomposition framework for graph clustering
10 3
Time (sec)
10 2
10 1
10 0
10 -1
Baseline
CoreCluster
10 -2
10 -30.0 0.5 1.0 1.5 2.0 2.5 3.0 3.5 4.0
1e3
Graph size
Baseline
CoreCluster
(a) Spectral Clustering
10 0
10 -1
10 0
Time (sec)
Time (sec)
10 1
10 -1
10 -2
Baseline
CoreCluster
10 -20.0 0.5 1.0 1.5 2.0 2.5 3.0 3.5 4.0
1e3
Graph size
10 -30.0 0.5 1.0 1.5 2.0 2.5 3.0 3.5 4.0
1e3
Graph size
(b) Leading Eigenvector
10 2
Baseline
CoreCluster
10 0
Time (sec)
Time (sec)
10 1
(c) Fast Greedy
10 -1
10 -2
10 -1
Baseline
CoreCluster
10 -2
10 -30.0 0.5 1.0 1.5 2.0 2.5 3.0 3.5 4.0
1e3
Graph size
10 -30.0 0.5 1.0 1.5 2.0 2.5 3.0 3.5 4.0
1e3
Graph size
(d) MCL
(e) Metis
Figure 8: Artificial Graphs execution time (Part I).
33
Giatsidis, Malliaros, Tziortziotis, Dhanjal, Kiagias, Thilikos and Vazirgiannis
10 -1
10 3
Time (sec)
Time (sec)
10 2
10 -2
Baseline
CoreCluster
10 1
10 0
Baseline
CoreCluster
10 -1
10 -30.0 0.5 1.0 1.5 2.0 2.5 3.0 3.5 4.0
1e3
Graph size
10 -20.0 0.5 1.0 1.5 2.0 2.5 3.0 3.5 4.0
1e3
Graph size
(f) MultiLevel
(g) SpinGlass
10 1
10 0
10 0
Time (sec)
Time (sec)
10 -1
10 -1
10 -2
10 -2
Baseline
CoreCluster
10 -30.0 0.5 1.0 1.5 2.0 2.5 3.0 3.5 4.0
1e3
Graph size
Baseline
CoreCluster
10 -30.0 0.5 1.0 1.5 2.0 2.5 3.0 3.5 4.0
1e3
Graph size
(h) Info Map
(i) Walktrap
Figure 8: Artificial Graphs execution time (Part II).
• Optimize the heuristics and clustering assignments our framework uses when not
utilizing the provided clustering algorithm.
• Merging adjacent cores. On a few hand picked occasions, we noticed that the framework would have been more effective if adjacent cores were merged as one. This is
mainly for cores that were only marginally different. It could potentially increase the
efficiency of execution and the quality of the results if a second pass over the cores
merged them appropriately.
• Extend it for directed graphs. For the application of the same model in a directed
network, one would have to take into account that degeneracy creates partitions in
two different dimensions (incoming and outgoing vertices). Exploring all possible
partitions in this case becomes a computation issue by itself.
Finally, we provide our implementation of CoreCluster for the use of further experiments and its’ application to real world scenarios. The full implementation along with
instructions on how to utilize the algorithms described in this paper can be found at:
removedwillbeinthepublishedversion
34
CoreCluster: A core decomposition framework for graph clustering
10 5
Time (sec)
10 4
Baseline
CoreCluster
10 3
10 2
10 1
0.0 0.5 1.0 1.5 2.0 2.5 3.0 3.5 4.0 4.5
1e4
Graph size
(a) Spectral Clustering
10 3
Time (sec)
Time (sec)
10 2
10 2
10 1
Baseline
CoreCluster
10 1
0.0 0.5 1.0 1.5 2.0 2.5 3.0 3.5 4.0 4.5
1e4
Graph size
(b) Leading Eigenvector
0.0 0.5 1.0 1.5 2.0 2.5 3.0 3.5 4.0 4.5
1e4
Graph size
(c) Fast Greedy
10 3
Time (sec)
Time (sec)
10 2
10 2
10 1
Baseline
CoreCluster
Baseline
CoreCluster
10 1
10 0
Baseline
CoreCluster
10 -1
0.0 0.5 1.0 1.5 2.0 2.5 3.0 3.5 4.0 4.5
1e4
Graph size
(e) Metis
0.0 0.5 1.0 1.5 2.0 2.5 3.0 3.5 4.0 4.5
1e4
Graph size
(d) MCL
Figure 9: Facebook execution time (Part I).
35
Giatsidis, Malliaros, Tziortziotis, Dhanjal, Kiagias, Thilikos and Vazirgiannis
10 2
Time (sec)
Time (sec)
10 2
10 1
10 0
10 1
Baseline
CoreCluster
Baseline
CoreCluster
10 0
0.0 0.5 1.0 1.5 2.0 2.5 3.0 3.5 4.0 4.5
1e4
Graph size
(f) MultiLevel
0.0 0.5 1.0 1.5 2.0 2.5 3.0 3.5 4.0 4.5
1e4
Graph size
(g) SpinGlass
10 2
Time (sec)
Time (sec)
10 3
10 2
10 1
Baseline
CoreCluster
10 1
0.0 0.5 1.0 1.5 2.0 2.5 3.0 3.5 4.0 4.5
1e4
Graph size
(h) Info Map
Baseline
CoreCluster
0.0 0.5 1.0 1.5 2.0 2.5 3.0 3.5 4.0 4.5
1e4
Graph size
(i) Walktrap
Figure 9: Facebook execution time (Part II).
Appendix A. Empirical Analysis of the CoreCluster’s Quality
As described in the main part, the intuition behind the CoreCluster framework is that
the core expansion sequence Vk , Vk−1 , . . . , V0 gives a good sense of direction on how to do
clustering in an incremental way. Moreover, the observation is that the early i-cores (i.e.,
i-cores where i is close to k) are already dense, and therefore sufficiently coherent, to provide
a good starting clustering that will expand well because of the selection criterion.
Figure 10 depicts the above intuition about the datasets used in our experiments (as they
are described in the main document). More specifically, it represents the average clustering
coefficient of a k-core subgraph with regards to the core index value k (normalized by the
maximum k of each graph). This indicates that, overall, a clearer clustering can be given
at the maximum k-core. As a reminder, the reader may see the properties of the artificial
data presented on Table 1.
The particular results, along with the theoretical justification (described in the main
document, Section 5), provide the conclusion that the highest ranked k-core will most
likely have the nodes of the graph that belong to the largest clusters and at the same the
clustering structure will be the best under the measure of the clustering coefficient. As the
36
CoreCluster: A core decomposition framework for graph clustering
0.9
Clustering Coefficient
0.85
0.8
0.75
0.7
0.65
0.6
0.55
0.5
0.1
0.2
0.3
0.4
0.5
0.6
0.7
Core (normalized)
0.8
0.9
1
Figure 10: Average clustering coefficient compared to k-core index normalized by the maximum value.
CoreCluster framework starts from that subgraph (the maximal k-core), it will perform
the best into separating the clusters of this subgraph and then it will move one the ”next
best” subset of nodes and edges as they are given from the core expansion sequence.
References
Amine Abou-Rjeili and George Karypis. Multilevel algorithms for partitioning power-law
graphs. In IPDPS, pages 124–124, 2006.
José I. Alvarez-Hamelin, Luca Dall’asta, Alain Barrat, and Alessandro Vespignani. k-core
decomposition: a tool for the visualization of large scale networks. arXiv, 2005.
Reid Andersen and Kumar Chellapilla. Finding dense subgraphs with size bounds. In
WAW, pages 25–37, 2009.
David Arthur and Sergei Vassilvitskii. k-means++: the advantages of careful seeding. In
SODA, pages 1027–1035, 2007.
Vladimir Batagelj and Matjaz Zaversnik. An o(m) algorithm for cores decomposition of
networks. CoRR, 2003.
Vincent D Blondel, Jean-Loup Guillaume, Renaud Lambiotte, and Etienne Lefebvre. Fast
unfolding of communities in large networks. Journal of statistical mechanics: theory and
experiment, 2008(10):P10008, 2008.
S. Bubeck, M. Meila, and U. von Luxburg. How the initialization affects the stability of the
k-means algorithm. ESAIM: Probability and Statistics, 16:436–452, 2012.
J. Cheng, Yiping Ke, Shumo Chu, and M.T. Ozsu. Efficient core decomposition in massive
networks. In ICDE, pages 51–62, 2011.
37
Giatsidis, Malliaros, Tziortziotis, Dhanjal, Kiagias, Thilikos and Vazirgiannis
Aaron Clauset, M. E. J. Newman, and Cristopher Moore. Finding community structure in
very large networks. Phys. Rev. E, 70(6):066111, Dec 2004. doi: 10.1103/PhysRevE.70.
066111.
Chandler Davis and W. M. Kahan. The rotation of eigenvectors by a perturbation. iii.
SIAM Journal on Numerical Analysis, 7(1):1–46, 1970.
P. Drineas, A. Frieze, R. Kannan, S. Vempala, and V. Vinay. Clustering large graphs via
the singular value decomposition. Mach. Learn., 56(1-3), 2004. ISSN 0885-6125.
Paul Erdős. On the structure of linear graphs. Israel J. Math., 1:156–160, 1963. ISSN
0021-2172.
Santo Fortunato. Community detection in graphs. Physics Reports, 486(3-5), 2010.
Christos Giatsidis, Dimitrios M. Thilikos, and Michalis Vazirgiannis. D-cores: Measuring
collaboration of directed graphs based on degeneracy. In ICDM, pages 201–210, 2011a.
Christos Giatsidis, Dimitrios M Thilikos, and Michalis Vazirgiannis. D-cores: Measuring
collaboration of directed graphs based on degeneracy. In Data Mining (ICDM), 2011
IEEE 11th International Conference on, pages 201–210. IEEE, 2011b.
Christos Giatsidis, Dimitrios M Thilikos, and Michalis Vazirgiannis. Evaluating cooperation
in communities with the k-core structure. In Advances in Social Networks Analysis and
Mining (ASONAM), 2011 International Conference on, pages 87–93. IEEE, 2011c.
Christos Giatsidis, Fragkiskos Malliaros, Dimitrios M Thilikos, and Michalis Vazirgiannis.
Corecluster: A degeneracy based graph clustering framework. In AAAI’14: 28th Conference on Artificial Intelligence, 2014.
David F. Gleich and C. Seshadhri. Vertex neighborhoods, low conductance cuts, and good
seeds for local community methods. In KDD, pages 597–605, 2012.
George Karypis and Vipin Kumar. A fast and high quality multilevel scheme for partitioning
irregular graphs. SIAM J. Sci. Comput., 20(1):359–392, December 1998. ISSN 1064-8275.
Sanjiv Kumar, Mehryar Mohri, and Ameet Talwalkar. On sampling-based approximate
spectral decomposition. In ICML’09, pages 553–560, 2009.
Andrea Lancichinetti, Santo Fortunato, and Filippo Radicchi. Benchmark graphs for testing
community detection algorithms. Physical Review E, 78, 2008. doi: 10.1103/PhysRevE.
78.046110.
Jure Leskovec and Christos Faloutsos. Sampling from large graphs. In KDD, pages 631–636,
2006. ISBN 1-59593-339-5.
Jure Leskovec, Kevin J. Lang, and Michael Mahoney. Empirical comparison of algorithms
for network community detection. In WWW, pages 631–640, 2010.
J.F. Lu, J.B. Tang, Z.M. Tang, and J.Y. Yang. Hierarchical initialization approach for
k-means clustering. Pattern Recognition Letters, 29(6):787 – 795, 2008. ISSN 0167-8655.
38
CoreCluster: A core decomposition framework for graph clustering
Helmut Lutkepohl. Handbook of matrices. Computational Statistics and Data Analysis, 2
(25):243, 1997.
Arun S. Maiya and Tanya Y. Berger-Wolf. Sampling community structure. In WWW, pages
701–710, 2010.
Christopher D. Manning, Prabhakar Raghavan, and Hinrich Schütze. Introduction to information retrieval. Cambridge University Press, 2008.
M. E. J. Newman. Fast algorithm for detecting community structure in networks. Phys.
Rev. E, 69(6):066133, Jun 2004. doi: 10.1103/PhysRevE.69.066133.
M. E. J. Newman and M. Girvan. Finding and evaluating community structure in networks.
Physical Review E, 69, 2004. doi: 10.1103/PhysRevE.69.026113.
Mark EJ Newman. Finding community structure in networks using the eigenvectors of
matrices. Physical review E, 74(3):036104, 2006.
Andrew Y. Ng, Michael I. Jordan, and Yair Weiss. On spectral clustering: Analysis and an
algorithm. In NIPS, pages 849–856, 2001.
Marzia Polito and Pietro Perona. Grouping and dimensionality reduction by locally linear
embedding. In NIPS, pages 1255–1262, 2001.
Pascal Pons and Matthieu Latapy. Computing communities in large networks using random
walks. In Computer and Information Sciences-ISCIS 2005, pages 284–293. Springer, 2005.
Stephen J. Redmond and Conor Heneghan. A method for initialising the k-means clustering
algorithm using kd-trees. Pattern Recognition Letters, 28(8):965–973, 2007.
Jörg Reichardt and Stefan Bornholdt. Statistical mechanics of community detection. Physical Review E, 74(1):016110, 2006.
Martin Rosvall, Daniel Axelsson, and Carl T Bergstrom. The map equation. The European
Physical Journal Special Topics, 178(1):13–23, 2009.
Venu Satuluri and Srinivasan Parthasarathy. Scalable graph clustering using stochastic
flows: applications to community discovery. In KDD, pages 737–746, 2009.
Venu Satuluri, Srinivasan Parthasarathy, and Yiye Ruan. Local graph sparsification for
scalable clustering. In SIGMOD, pages 721–732, 2011.
Stephen B. Seidman. Network structure and minimum degree. Social Networks, 5(3):269–
287, 1983.
Jianbo Shi and Jitendra Malik. Normalized cuts and image segmentation. IEEE Trans.
Pattern Anal. Mach. Intell., 22(8):888–905, 2000. ISSN 0162-8828.
GW Stewart and Ji-Guang Sun. Matrix perturbation theory (computer science and scientific
computing), 1990.
39
Giatsidis, Malliaros, Tziortziotis, Dhanjal, Kiagias, Thilikos and Vazirgiannis
Amanda L. Traud, Peter J. Mucha, and Mason A. Porter. Social structure of facebook
networks. CoRR, 2011.
Stijn Van Dongen. A cluster algorithm for graphs. Report-Information systems, (10):1–40,
2000.
Ulrike von Luxburg. A tutorial on spectral clustering. Statistics and Computing, 17(4):
395–416, 2007. ISSN 0960-3174.
D. J. Watts and S. H. Strogatz. Collective dynamics of’small-world’networks. Nature, 393
(6684), 1998.
Scott White and Padhraic Smyth. A spectral clustering approach to finding communities
in graph. In SDM, 2005.
Christopher K. I. Williams and Matthias Seeger. Using the nyström method to speed
up kernel machines. In Advances in Neural Information Processing Systems 13, pages
682–688. MIT Press, 2001.
Donghui Yan, Ling Huang, and Michael I. Jordan. Fast approximate spectral clustering. In
Proceedings of the 15th ACM SIGKDD International Conference on Knowledge Discovery
and Data Mining, KDD ’09, pages 907–916. ACM, 2009.
Yang Zhang and Srinivasan Parthasarathy. Extracting analyzing and visualizing triangle
k-core motifs within networks. In ICDE, pages 1049–1060, 2012.
40
| 8cs.DS
|
arXiv:1709.05266v1 [math.LO] 15 Sep 2017
DIMENSION 1 SEQUENCES ARE CLOSE TO RANDOMS
NOAM GREENBERG, JOSEPH S. MILLER, ALEXANDER SHEN,
AND LINDA BROWN WESTRICK
Abstract. We show that a sequence has effective Hausdorff dimension 1 if
and only if it is coarsely similar to a Martin-Löf random sequence. More
generally, a sequence has effective dimension s if and only if it is coarsely
similar to a weakly s-random sequence. Further, for any s ă t, every sequence
of effective dimension s can be changed on density at most H ´1 ptq ´ H ´1 psq
of its bits to produce a sequence of effective dimension t, and this bound is
optimal.
The theory of algorithmic randomness defines an individual object in a probability space to be random if it looks plausible as an output of a corresponding random
process. The first and the most studied definition was given by Martin-Löf [ML66]:
a random object is an object that satisfies all “effective” probability laws, i.e., does
not belong to any effectively null set. (See [DH10, UVS13, She15] for details; we
consider only the case of uniform Bernoulli measure on binary sequences, which
corresponds to independent tossings of a fair coin.) It was shown by Schnorr and
Levin (see [Sch72, Sch73, Lev73]) that an equivalent definition can be given in
terms of description complexity: a bit sequence X P 2ω is Martin-Löf (ML) random
if and only if the prefix-free complexity of its n-bit prefix X æn is at least n ´ Op1q.
(See [LV93, UVS13, She15] for the definition of prefix-free complexity and for the
proof of this equivalence; one may use also monotone or a priori complexity.) This
robust class also has an equivalent characterization based on martingales that goes
back to Schnorr [Sch71].
The notion of randomness is in another way quite fragile: if we take a random
sequence and change to zero, say, its 10th, 100th, 1000th, etc. bits, the resulting sequence is not random, and for a good reason: a cheater that cheats once in a while is
still a cheater. To consider such sequences as “approximately random”, one option
is to relax the Levin-Schnorr definition by replacing the Op1q term in the complexity characterisation of randomness by a bigger opnq term, thus requiring that
limnÑ8 KpX æn q{n “ 1. Such sequences coincide with the sequences of effective
Hausdorff dimension 1. (Effective Hausdorff dimension was first explicitly introduced by Lutz [Lut00]. It can be defined in several equivalent ways via complexity,
via natural generalizations of effective null sets, and via natural generalizations of
martingales; again, see [DH10, UVS13, She15] for more information.)
Another approach follows the above example more closely: we could say that
a sequence is approximately random if it differs from a random sequence on a set
The first and fourth authors were supported by a Rutherford Discovery Fellowship from the
Royal Society of NZ. The second author was partially supported by grant #358043 from the
Simons Foundation. The third author was partially supported by RaCAF ANR-15-CE40-0016-01
grant.
1
2
N. GREENBERG, J.S. MILLER, A. SHEN, AND L.B. WESTRICK
of density 0. Our starting point is that this also characterizes the sequences of
effective Hausdorff dimension 1.
To set notation, for n ě 1, we let d be the normalised Hamming distance
n
on t0, 1u , the set of binary strings of length n:
# tk : σpkq ‰ τ pkqu
;
n
and we also denote by d the Besicovitch distance on Cantor space 2ω (the space of
infinite binary sequences), defined by
dpσ, τ q “
dpX, Y q “ lim sup dpX æn , Y æn q,
nÑ8
where Z æn stands for the n-bit prefix of Z. If dpX, Y q “ 0, then we say that X
and Y are coarsely equivalent.1
Theorem 1.7. A sequence has effective Hausdorff dimension 1 if and only if it is
coarsely equivalent to a ML-random sequence.
In Section 2, we generalize this result to sequences of effective dimension s in
various ways. Because a sequence X having effective dimension s implies that the
prefix-free complexity of its n-bit prefix X æn is at least sn ´ opnq, it is natural to
consider the weakly s-randoms, those sequences X such that KpX æn q ě sn ´ Op1q.
Theorem 2.5. Every sequence of effective Hausdorff dimension s is coarsely equivalent to a weakly s-random.
Along the way to proving this, we pass through the question of how to raise the
effective dimension of a given sequence while keeping density of changes at a minimum. If dpX, Y q “ 0, then dimpXq “ dimpY q; so sequences of effective Hausdorff
dimension s ă 1 cannot be coarsely equivalent to a ML random sequence. It is
natural then to ask, what is the minimal distance required between any sequence
and a random? By Theorem 2.5, it is equivalent to ask about distances between sequences of dimension s and dimension 1; and naturally generalising, to ask, for any
0 ď s ă t ď 1, about distances between sequences of dimension s and dimension t.
We start with a naive bound. For any X, Y P 2ω ,
| dimpY q ´ dimpXq| ď HpdpX, Y qq.
This is our Proposition 3.1. Here Hppq “ ´pp log p ` p1 ´ pq logp1 ´ pqq is the binary
entropy function defined on r0, 1s. The binary
entropy
function is used to measure
` ˘
ř
the size of Hamming balls. If V pn, rq “ kďnr nk is the size of a Hamming ball of
radius r ă 1{2 in 2n , then
Hprqn ´ Oplog nq ď logpV pn, rqq ď Hprqn
(see [MS77, Cor. 9, p. 310]).
In Proposition 3.5, we will see that this bound is tight, in the sense that if s ă t
then there are X, Y P 2ω with dimpXq “ s, dimpY q “ t and dpX, Y q “ H ´1 pt ´ sq.
Note that for H ´1 we take the branch which maps r0, 1s to r0, 1{2s.
Bounding the distance from an arbitrary dimension s sequence to the nearest
dimension t sequence requires more delicate analysis. For example, fix 0 ă s ă
t ď 1. If X is Bernoulli H ´1 psq-random, then its dimension is s. But its density
of 1s is H ´1 psq. If dimpY q ě t then the density of 1s in Y is at least H ´1 ptq, so
1One place this is defined is in [JS12], where it is called “generic similarity”.
DIMENSION 1 SEQUENCES ARE CLOSE TO RANDOMS
3
dpX, Y q ě H ´1 ptq ´ H ´1 psq. Note that H ´1 ptq ´ H ´1 psq ě H ´1 pt ´ sq, so this is
a sharper bound, and it is tight:
Theorem 4.1. For every sequence X with dimpXq “ s, and every t P ps, 1s, there
is a Y with dimpY q “ t and dpX, Y q ď H ´1 ptq ´ H ´1 psq.
In particular, for t “ 1, in light of Theorem 1.7, we obtain
Theorem 2.1. For every X P 2ω there is a ML-random sequence Y such that
dpX, Y q ď 1{2 ´ H ´1 pdimpXqq.
(We however prove Theorem 2.1 first, and elaborate on its proof to obtain Theorem 2.5 and then Theorem 4.1.)
We can also ask, starting from an arbitrary random, how close is the nearest
sequence of dimension s guaranteed to be? For example, a typical construction
of a sequence of effective dimension 1{2 starts with a random and replaces all the
even bits with 0. The distance between the resulting pair is 1{4, less than the
1{2 ´ H ´1 p1{2q « .4 needed to get to a Bernoulli random, but more than the
H ´1 p1 ´ 1{2q « .1 lower limit from Proposition 3.1. Here, the latter bound is tight:
Theorem 3.3. For any Y P 2ω there is some X P 2ω such that dimpXq ď s and
dpX, Y q ď H ´1 p1 ´ sq.
These results mean that in general, the distance from an arbitrary dimension 1
sequence to the nearest dimension s sequence is quite a bit less than the distance
from an arbitrary dimension s sequence to the nearest dimension 1 sequence.
Finally, we mention that for t ă 1, the bound given by Proposition 3.1 for
the required distance to a sequence of dimension s ă t is not optimal; we show
below (Proposition 3.4) that for the case t “ 1{2 and s “ 0, there are Y P 2ω
of dimension 1{2 with no X P 2ω of dimension 0 within distance H ´1 p1{2q. Here
decreasing information is not so simple due to the possibility of redundancy of
information.
Each of these infinitary results have finite versions. Examples of similar finite
theorems previously appeared in [BFNV04]. The finite versions are proved using
either Harper’s Theorem, a result of finite combinatorics; or estimates on covering
Hamming space by balls of a given radius. We adapt those methods, together with
some convexity arguments, to prove our results.
These results exist in the context of a larger set of questions on the general
theme of asking whether every sequence of high effective dimension is obtainable
by starting with a random and “messing it up”. If a random were messed up only
slightly to produce a sequence of high effective dimension, it might be possible to
computably extract a random sequence back out. It was shown in [BDS09, FHP` 11]
that if X has positive packing dimension, then X computes a sequence of packing
dimension at least 1 ´ ε for each ε ą 0. On the other hand, the second author
showed that for any left-c.e. α P p0, 1q, there is a sequence of effective dimension α
which does not compute any sequence of effective dimension greater than α [Mil11],
and in [GM11] it was shown that there is a sequence of dimension 1 that does not
compute any random. Therefore, the symmetric differences X∆Y which we find
here are not, in general, computable. For more references on this type of question,
see [DH10, Section 13.7].
4
N. GREENBERG, J.S. MILLER, A. SHEN, AND L.B. WESTRICK
n
To set notation, for a binary string σ of length n (we write σ P t0, 1u ) let
Kpσq
;
n
then for an infinite binary sequence X P 2ω ,
dimpσq “
dimpXq “ lim inf dimpX æn q.
n
We can similarly define conditional dimension:
dimpσ | τ q “
Kpσ | τ q
.
|σ|
1. Dimension 1 sequences and randoms
In this section, we prove Theorem 1.7. Let P be the set of random sequences
with deficiency 0:
P “ tY : p@nq KpY æn q ě nu .
This P is not empty. Given a dimension 1 sequence X, we will build a Y P P such
that dpX, Y q “ 0.
Let P be the set of extendible strings of P : the prefixes of elements of P . The
following lemma tells us that every string σ in P has many extensions in P of length
2|σ|. An analogous lemma for supermartingales rather than prefix-free complexity
was proved by Merkle and Mihailovic [MM04, Rmk. 3.1]. They gave a clean presentation of the Kučera–Gács theorem. Gács made use of a similar lemma ([Gác86,
Lem. 1]), which guaranteed a sufficient number of extensions in Π01 classes.
Lemma 1.1. Every σ P P has at least 2|σ|´Kp|σ|q´Op1q extensions in P of length 2|σ|.
Proof. We prove that there is a k such that every σ P P has at least 22|σ|´Kpσq´k
extensions of length 2|σ|. This is enough, since Kpσq ď |σ| ` Kp|σ|q ` Op1q.
Suppose that some σ P P has fewer than 22|σ|´Kpσq´k extensions in P (but has
some, otherwise σ cannot be in P). Then each extension (denoted by σ 1 ) has small
complexity:
Kpσ 1 q ď Kpσq ` Kpσ 1 |σ, Kpσqq ` Op1q ď
ď Kpσq ` Kpσ 1 |k, σ, Kpσqq ` Oplog kq ď
ď Kpσq ` 2|σ| ´ Kpσq ´ k ` Oplog kq “ 2|σ| ´ k ` Oplog kq
The first inequality is the formula for the complexity of a pair. In the second one we
add k to the condition; the additional Oplog kq term appears. For the third one, if
we know k, σ, and Kpσq, then we can wait until fewer than 22|σ|´Kpσq´k candidates
for σ 1 remain (the set P is co-c.e.), and then specify each remaining candidate by its
ordinal number using a 2|σ|´Kpσq´k bit string; this is a self-delimiting description
since its length is known from the condition.
For each such σ 1 , we have Kpσ 1 q ě 2|σ|. Therefore, k ´ Oplog kq ď 0. So if such
a σ exists, then k is bounded. Equivalently, if k is sufficiently large, then there is
no such σ, i.e., each σ must have at least 22|σ|´Kpσq´k extensions.
n
For sets of strings A, B Ď t0, 1u , we let dpA, Bq “ min tdpσ, τ q : σ P A, τ P Bu.
We let dpσ, Aq “ dptσu, Aq.
Harper’s theorem ([Har66], see also [FF81]) says that among all subsets A, B Ď
n
t0, 1u of fixed sizes, a pair with maximal distance is obtained by taking spheres
DIMENSION 1 SEQUENCES ARE CLOSE TO RANDOMS
5
with opposite centres of 0n and 1n . Here a sphere centred at σ is a set C that (for
some k) contains the Hamming ball of radius k{n centred at σ and is also contained
in the ball of radius pk ` 1q{n with the same centre.2
n
Harper’s Theorem. For any sets A, B Ď t0, 1u , there are spheres Â, B̂, centred
at 0n and 1n respectively, such that |A| “ |Â|, |B| “ |B̂| and dpA, Bq ď dpÂ, B̂q.
A first application, useful for us, is the following.
n
Lemma 1.2. For every ε ą 0 there is a q ă 1 such that for any n and any A Ď t0, 1u
of size at least 2nq , there are at most 2nq strings σ P t0, 1un such that dpσ, Aq ą ε.
n
n
Proof. For a given A Ď t0, 1u , let B “ tσ P t0, 1u : dpσ, Aq ą εu. We need to
show that A and B cannot both contain at least 2nq elements, for an appropriate
choice of q. Note that if |A| ě 2nq and |B| ě 2nq , where q “ Hp1{2 ´ ε{2q,
then the inner radii of the spheres  and B̂ from Harper’s Theorem are at least
1{2 ´ ε{2 ´ Op1q{n, because each sphere is an intermediate set between two balls
whose radii differ by 1{n. Therefore, dpA, Bq ď ε ` Op1q{n. Note that H is strictly
increasing on r0, 1{2s and Hp1{2q “ 1, so q ă 1.
To get rid of the error term Op1q{n that appears because of discretisation, we
can decrease ε in advance. Then the statement is true for all sufficiently large n.
To make it true for all n, we choose q so close to 1 that the statement is vacuous
for small n; it is guaranteed if 2nq ą 2n ´ 1.
These tools (Harper’s theorem and the entropy bound) were used in [BFNV04] to
prove results on increasing the Kolmogorov complexity of finite strings by flipping
a limited number of bits. As an example of this technique, consider the following
“finite version” of Theorem 1.7: for any ε ą 0 there is a q ă 1 such that for
n
sufficiently large n, for any string σ P t0, 1u of dimension at least q (i.e., Kpσq ě
n
nq), there is a random string τ P t0, 1u (i.e., Kpτ q ě n) such that dpσ, τ q ď ε.
Here is the argument using Lemma 1.2. The set of random strings has size at least
2n´1 , and is co-c.e.; so once we see that a string σ is one of the fewer than 2qn
many strings that are at least ε-away from each random string, we can give it a
description of length essentially nq.
A naive plan for the infinite version is to repeat this construction for longer and
longer consecutive blocks of bits of a given sequence X of dimension 1, finding closer
and closer extensions in a Π01 class of randoms. This fails because the opponent
X can copy the extra information that we pump into Y , erasing our gains. For
example, if X begins with a very large string of 0s, we must begin Y with a random
string to stay in P . Then X (which must have dimension 1 eventually) could bring
its initial segment complexity as close to 1/2 as it likes by copying Y from the
beginning onto its upcoming even bits. Then Y can never take advantage of X’s
complexity to get closer to X, since that would cause Y to repeat information.
We cannot overcome this problem by taking huge steps (so large that X runs out
of things to copy and must show us new information) because X can still use a
similar strategy to ensure that the density of symmetric difference is large near the
beginning of each huge interval, driving up the lim supn dpX æn , Y æn q even as we
keep dpX æn , Y æn q low for n on the interval boundary.
Our solution is to not do an initial segment construction but rather use compactness to let ourselves change our mind about our initial segment whenever the
2This is, admittedly, an unusual use of “sphere”; we adopt it from [FF81].
6
N. GREENBERG, J.S. MILLER, A. SHEN, AND L.B. WESTRICK
opponent seems to take advantage of its extra information. Lemma 1.3 below shows
how to do this.
Let E be the set of all finite sequences ε̄ “ pε1 , ε2 , . . . , εm q such that ε1 “ 1 and
εk`1 equals either εk or εk {2 for all k ă m. For m ě 1, binary sequences σ, τ of
length 2m , and ε̄ P E of length m, we write
σ „ε̄ τ
if for all k P t1, . . . , mu,
`
˘
d σ ær2k´1 ,2k q , τ ær2k´1 ,2k q ď εk .
So we compare the second halves of the strings for k “ m, the second quarters for
k “ m ´ 1, and so on. (The 0th bits of σ and τ are ignored.)
Lemma 1.3. For every ε ą 0 there is an s ă 1 such that for sufficiently large m, for
every ε̄ P E of length m, and for all binary strings σ, ρ of length 2m , if
(1) pε̄, εq P E (i.e., ε P tεm , εm {2u),
(2) dimpρ | σq ě s,3 and
(3) there is a τ P P of length 2m such that τ „ε̄ σ,
then there is a ν P P of length 2m`1 such that ν „pε̄,εq σρ.
Note that the guaranteed ν need not be an extension of τ .
n
Proof. Let n “ 2m . For a given σ and ε̄, let A be the set of all strings η P t0, 1u
such that for some τ̂ P P X t0, 1un we have τ̂ „ε̄ σ and τ̂ η P P. The set A is co-c.e.
(given σ and ε̄). Let q be given by Lemma 1.2 (for ε). Now apply Lemma 1.1 to
n
any τ̂ P P X t0, 1u : since Kpnq{n Ñ 0 as n Ñ 8, the size of A (and even its part
that corresponds to this specific τ̂ ) is at least 2nq (for sufficiently large m).
n
Let B be the set of strings π P t0, 1u such that dpπ, Aq ě ε. The set B is c.e.,
and Lemma 1.2 guarantees that the size of B is at most 2nq . This implies that each
string in B can be given a description (conditioned on σ) of length nq ` m ` Op1q
bits; m bits are used to specify ε̄. Set s ą q. Then since m “ log n, for sufficiently
large m we have dimpπ | σq ă s for all π P B. So ρ R B. This means that there
is some η P A such that dpη, ρq ď ε. Let τ̂ witness that η P A. Then ν “ τ̂ η is as
required.
We finish our preparation with three easy observations.
Lemma 1.4. Let X, Y P 2ω and suppose that
lim dpX ær2m´1 ,2m q , Y ær2m´1 ,2m q q “ 0.
mÑ8
Then dpX, Y q “ 0.
ω
Lemma 1.5. Let X P 2 and suppose that dimpXq “ 1. Then
KpX ær2m ,2m`1 q | X æ2m q
“ 1.
2m
Proof. The complexity of pairs formula shows that
lim
mÑ8
KpX æ2m`1 q “ KpX æ2m q ` KpX ær2m ,2m`1 q | X æ2m q ` op2m q;
the sum can be (almost) maximal only if both terms are (almost) maximal.
Lemma 1.6. If dpX, Y q “ 0, then dimpXq “ dimpY q.
3Recall that dimpρ | σq “ Kpρ | σq{|ρ|.
DIMENSION 1 SEQUENCES ARE CLOSE TO RANDOMS
7
Theorem 1.7. Let X P 2ω . Then dimpXq “ 1 if and only if there is a ML random
Y P 2ω such that dpX, Y q “ 0.
Proof. One direction is immediate from Lemma 1.6. For the other direction, assume
that dimpXq “ 1. Let
sm “ dim pX ær2m ,2m`1 q | X æ2m q.
Define an infinite sequence ε̄ “ pε1 , ε2 , . . . q such that:
‚ ε1 “ 1 and εk`1 P tεk , εk {2u for all k,
‚ limkÑ8 εk “ 0, and
‚ for all m, the triple εm , sm , m satisfies the conclusion of Lemma 1.3 (so if
sm slowly converges to 1, we need a sequence εm that slowly converges to
0).
We then let
!
)
2m
Qm “ ν P P X t0, 1u
: ν „pε1 ,...,εm q X æ2m .
By induction, Lemma 1.3 shows that for all m the set Qm is nonempty. Note also
that all elements of Qm have a prefix from Qm´1 . By compactness, there is a Y P P
such that Y æ2m P Qm for all m. By Lemma 1.4, such Y is as required.
2. Dimension s sequences and randoms
In this section, we look at the distance between dimension s sequences and two
kinds of randoms. We consider first the density of symmetric difference required to
change a sequence of dimension s into a ML random. Extending that argument,
we then show that every sequence of dimension s is coarsely similar to a weakly
s-random.
Theorem 2.1. For every X P 2ω there is a ML-random sequence Y such that
dpX, Y q ď 1{2 ´ H ´1 pdimpXqq.
In the introduction we saw that Theorem 2.1 is optimal by considering the case
when X is a Bernoulli H ´1 psq-random sequence.
The proof of Theorem 2.1 requires several modifications to the work we did in
the previous section. For one thing, Lemma 1.4 fails to generalize, as a positive
upper density of X∆Y may be greater partway through these intervals than at their
boundaries. This could be improved by shortening the intervals. On the other hand,
to make something like Lemma 1.5 true for dimension s we would need to increase
the size of the intervals; shortening them only makes it fail even worse. The solution
is to go ahead and shorten the intervals, but instead of trying to approach a given
target symmetric distance slowly and directly, we let the local symmetric difference
rise and fall in accordance with the rise and fall of the conditional complexity of
each new chunk. Then a convexity argument will let us conclude that the distance
between X and Y is small enough.
Letting the size of the intervals grow quadratically achieves a happy medium. In
2
this and all following constructions, the
ř jth2chunk has size j 2 and the first j chunks
have concatenated length of nj :“ iăj i , so that nj ` j “ nj`1 . There was
plenty of freedom in choice of chunk growth rate, but we must choose something
satisfying these conditions: We need nj`1 ´ nj ! nj so that the impact of new
chunks on the density of symmetric difference and effective dimension goes to zero
in the limit. And although we could get away with somewhat less, we also want
8
N. GREENBERG, J.S. MILLER, A. SHEN, AND L.B. WESTRICK
j log j ! nj`1 ´ nj in order to have room to fit a description of δ, which we define
next.
To replace the set E of descending sequences of εi , we use the set D of finite
sequences δ “ pδ1 , . . . , δm q such that each δj is a fraction of the form k{j for some
positive integer k ď j. As before, we write σ „δ τ if for each j,
dpσ ærnj ,nj`1 q , τ ærnj ,nj`1 q q ă δj .
By Theorem 1.7, it suffices to construct Y of effective dimension 1. So instead
of staying inside a tree of randoms, in this and all following constructions we stay
in trees of the following type.
Given a sequence t̄ of numbers in the unit interval, let
Pt̄ “ tY : p@iq dimpY ærni ,ni`1 q | Y æni q ě ti u.
Let Pt̄ denote the set of initial segments of Pt̄ . For this first argument, we let t̄ be
1̄, the sequence of all 1s. By Lemma 2.4 below, if Y P P1̄ , then dimpY q “ 1.
The following lemma plays the role of Lemma 1.2 and Lemma 1.3.
Lemma 2.2. For every ε ą 0, there is an N such that for every j ě N , σ P t0, 1unj ,
j2
ρ P t0, 1u , δ̄ P D, and δ “ k{j, if δ ą ε and
(1) dimpρ | σq ě Hp1{2 ` ε ´ δq and
(2) σ „δ τ for some τ P P1̄ ,
then there is a ν P P1̄ such that σρ „pδ,δq ν.
Proof. The sets A and B are defined the same as in the proof of Lemma 1.3:
j2
A “ tη P t0, 1u
: Dτ̂ P P1̄ pτ̂ „δ̄ σ and τ̂ η P P1̄ qu,
j2
and B “ tπ P t0, 1u : dpπ, Aq ě δu. Now use Harper’s Theorem to show that
2
because |A| ą 2j ´1 , we have log |B| ď Hp1{2´δqj 2. Therefore, relative to σ, codes
for elements of B can be given which have length Hp1{2 ´ δqj 2 ` Opj log jq, where
Opj log jq bits are used to describe δ̄. The difference Hp1{2 ` ε ´ δq ´ Hp1{2 ´ δq is
bounded away from zero by a fixed amount that does not depend on δ. Therefore,
there is an N large enough that codes for elements of B are shorter than Hp1{2 `
ε ´ δqj 2 , where the choice of N does not depend on δ.
In the construction itself, we define an infinite sequence pε1 , ε2 , . . . q so that εi Ñ 0
and N “ i witnesses Lemma 2.2 for εi . Then we define
(1)
sj “ dim pX ærnj ,nj`1 q | X ænj q
and finally δj “ 1{2`εj ´H ´1 psj q`Op1{jq to obtain the sequence pδ1 , δ2 , . . . q to be
used for the definition of Qj and the proof that the Qj are nonempty. This produces
Y P P1̄ with X „δ̄ Y . It remains only to examine the relationship between the sj ,
the δj , dimpXq, and dpX, Y q for X „δ Y . The following variations on Lemma 1.4
and Lemma 1.5 are well-known.
Lemma 2.3. For X, Y P 2ω , letting δi “ dpX ærni ,ni`1 q , Y ærni ,ni`1 q q,
j´1
1 ÿ 2
δi i .
dpX, Y q “ lim sup
jÑ8 nj i“1
Proof. Since nj grows slowly enough, both are equal to lim sup dpX ænj , Y ænj q.
jÑ8
DIMENSION 1 SEQUENCES ARE CLOSE TO RANDOMS
9
Lemma 2.4. For X P 2ω , letting sj be defined as in (1),
dimpXq “ lim inf
jÑ8
j´1
1 ÿ 2
si i .
nj i“1
Proof. Since nj grows slowly enough, it suffices to show
Kpσq “
j´1
ÿ
si i2 ` opnj q
i“1
considering only σ of length nj . One direction is immediate; the other uses j
applications of symmetry of information and the fact that j log j ! nj (since we
condition only on X ænj and not its code, each application of symmetry of information introduces an error of up to log nj « log j.)
In our case, with X „δ Y , we have
δi “ 1{2 ` εi ´ H ´1 psi q ě dpX ærni ,ni`1 q , Y ærni ,ni`1 q q.
so
j´1
j´1
1 ÿ
1 ÿ 2
δi i “ lim sup
p1{2 ´ H ´1 psi qqi2
dpX, Y q ď lim sup
jÑ8 nj i“1
jÑ8 nj i“1
because εi Ñ 0. By Lemma 2.4 and the concavity of 1{2 ´ H ´1 pxq, this is bounded
by 1{2 ´ H ´1 pdimpXqq, as required. This finishes the proof of Theorem 2.1.
This method can be extended to answer a question that was suggested to us
by M. Soskova. Recall that a sequence X is called weakly s-random if it satisfies
KpX æn q ě sn ´ Op1q.
Theorem 2.5. Every sequence of effective Hausdorff dimension s is coarsely equivalent to a weakly s-random.
Simply staying in a tree of s-randoms (direct generalization of Theorem 1.7) does
not provide a strong enough lower bound on the number of possible extensions to
use Harper’s theorem, since we now need for more than half of the possible strings
of a certain length to be available at each stage. To ensure this, we instead require
our constructed sequence to buffer its complexity above the level needed to be
s-random. Such a strategy was not a possibility in the 1-random case.
In the case of finite strings, Harper’s Theorem can be used to show, essentially,
that if x is a string of dimension s, then by changing it on ε fraction of its bits, its
dimension can be increased to
M ps, εq :“ Hpminp1{2, H ´1 psq ` εqq.
We adapt the standard method to prove the following lemma, Case 2 of which is
identical to Lemma 2.2.
n
Lemma 2.6. For all ε, there is an N such that for all j ě N , all σ P t0, 1u j and
j2
ρ P t0, 1u , all δ̄ P D, t̄ P D of length j, and all δ, if
(1) pδ̄, δq P D and δ ě ε and
(2) there is a τ P Pt̄ with τ „δ̄ σ,
then, letting s “ dimpρ | σq and t “ M ps, δ ´ εq ` Op 1j q, there is a ν P Ppt̄,tq with
ν „pδ̄,δq σρ.
10
N. GREENBERG, J.S. MILLER, A. SHEN, AND L.B. WESTRICK
Proof. Like before, let
2
A “ tη P t0, 1uj : Dτ̂ „δ̄ σ pτ̂ η P Ppt̄,tq qu
j2
and B “ tπ P t0, 1u : dpπ, Aq ě δu.
Case 1. Suppose that 0 ď s ď Hp1{2 ´ δ ´ ε{2q. The upper bound is chosen so
that t is bounded below 1 for s in this interval. Since τ exists, by considering only
2
2
extensions of it, we can bound |A| ą 2j ´ 2tj . Letting q “ M ps, δ ´ ε ` ε{4q (note
this q is chosen so that t ă q ă 1), for sufficiently large j we have
(2)
2
|A| ą 2j ´ V pj 2 , H ´1 pqqq,
where V pn, rq is the size of a sphere of radius rn in 2n (this uses the lower bound
for V pn, rq mentioned in the introduction). How large j has to be for this bound to
hold depends on the size of q ´ t, which in general varies with s and δ. But since
q ą t for all s, δ with ε ď δ ď 1 and s in the closed interval associated to this case,
compactness allows us to bound q ´ t away from 0 by a quantity that depends only
on ε. Assuming j is large enough for (2) to hold, Harper’s Theorem tells us that
log |B| ď HpH ´1 pqq ´ δqj 2 “ HpH ´1 psq ´ ε ` ε{4qj 2
so B is either empty, or everything in B can be compressed (relative to σ) to
length HpH ´1 psq ´ 3ε{4qj 2 ` Opj log jq, (where the Opj log jq is enough to code
the parameters δ̄ and t̄ needed to define B as a c.e. set), and for large enough j
the code length is less than sj 2 . How large j has to be depends on the difference
s ´ HpH ´1 psq ´ 3ε{4q, but this can again be bounded in a way that depends only
on ε. So for sufficiently large j, if the conditions are satisfied, then ρ R B, and the
lemma holds.
Case 2. Suppose that Hp1{2 ´ δ ` ε{2q ď s ď 1. Because τ exists, |A| ě
2
2j ´1 , so log |B| ď Hp1{2 ´ δqj 2 (this is true regardless of the choice of s). Either
B is empty, or this allows the construction of a similar code, with the needed
largeness of j determined by s ´ Hp1{2 ´ δq, which is bounded away from 0 by
Hp1{2 ´ δ ` ε{2q ´ Hp1{2 ´ δq for all s ě Hp1{2 ´ δ ` ε{2q. For δ P rε, 1{2s,
this bound is strictly positive, so by compactness there is a uniform lower bound
depending only on ε.
Choosing N large enough that j ě N works for both cases finishes the proof.
Now given an X with dimpXq “ s, define the sequence sj as in (1). Suppose
also that an infinite sequence ε̄ Ñ 0 is given, decreasing slowly enough that N “ i
satisfies the previous lemma for εi . Define the infinite sequence δ̄ by δi “ 2εi , and
define t̄ by letting tj be M psj , δj ´εj q “ M psj , εj q rounded up to the nearest rational
with denominator j. The previous lemma together with the usual compactness
argument provides a sequence Y P Pt̄ with dpX, Y q “ 0 and thus dimpY q “ s.
We claim that by that by appropriately slow choice of ε̄, we can guarantee that Y
comes out weakly s-random.
The idea is that while ε is held fixed above 0, by changing ε fraction of each
new chunk, we make Y behave like a sequence of dimension strictly greater than
s for as long as we like. This allows the production and maintenance of a buffer
of extra complexity which is used to smooth out the bumps in our construction—
the logarithmic factors from the use of the complexity of pairs formula and the
possibility for a mid-chunk decrease in the complexity of Y .
DIMENSION 1 SEQUENCES ARE CLOSE TO RANDOMS
11
Specifically, using the complexity of pairs formula repeatedly, and considering
worst case mid-chunk complexities, we have
KpY ænj `k q ě
j
ÿ
ti i2 ´ Opj log jq
i“0
2
where k ă j . So Y will be s-random if we can arrange that eventually
j
ÿ
ti i2 ´ Opj log jq ą spnj ` j 2 q,
i“0
which is what the next lemma guarantees.
Lemma 2.7. For any constant c and any infinite sequence s̄ P D, let
s “ lim inf
jÑ8
j
1 ÿ 2
si i .
nj i“0
Then there is an infinite ε̄ P E with εi Ñ 0, and a constant b, such that for all j,
j
ÿ
M psi , εi qi2 ´ cj 2 ą snj ´ b.
i“0
Proof. Observe that for any fixed ε, there is a d ą 0 such that
M px, εq ě d ` p1 ´ dqx.
(On the closed interval r0, Hp1{2 ´ εqs, M px, εq ą x, so by compactness, on this
interval M px, εq ą x ` d for some d. Outside this interval, M px, εq “ 1.) By this
bound,
j
j
ÿ
ÿ
si i 2 .
M psi , εqi2 ě dnj ` p1 ´ dq
i“1
i“0
Since s ă 1, there is a δ such that d ` p1 ´ dqps ´ δq ą s ` δ. So for large enough
N , we have for each j ą N ,
j
ÿ
M psi , εqi2 ě dnj ` p1 ´ dqps ´ δqnj ą ps ` δqnj ą snj ` cj 2 .
i“1
All this was done for fixed ε, so rename this N as Nε . The sequence we want is
defined by letting ε0 “ 1, and then let εj “ εj´1 {2 if j ą Nεj´1 {2 , and otherwise
εj “ εj´1 . The constant b is chosen to absorb any irregularity that occurs for
j ď N1 .
Setting c large enough, choosing ε̄ according to Lemma 2.7, and constructing
Y using ε̄ with Lemma 2.6 produces a Y which is coarsely similar to X and with
KpY æn q ą sn ´ b. This completes the proof of Theorem 2.5.
3. Intermezzo: decreasing dimension
In the next section, we will generalise Theorem 2.1 to increasing dimension from s
to some t ă 1. Now we discuss decreasing dimension. As discussed in the introduction, in one case we can meet the following naive bound.
Proposition 3.1. For any X, Y P 2ω ,
| dimpY q ´ dimpXq| ď HpdpX, Y qq.
12
N. GREENBERG, J.S. MILLER, A. SHEN, AND L.B. WESTRICK
Proof. Let s “ dimpXq and δ “ HpdpX, Y qq. For infinitely many n we have
KpY æn q ď ps ` εqn ` Hpδ ` εqn ` Oplog nq, where ε is arbitrary. The first term
comes from the inequality KpX æn q ď ps ` εqn which holds infinitely often. The
second term comes from a description of the symmetric difference X æn ∆Y æn ,
using the fact that eventually dpX æn , Y æn q ă δ ` ε to bound the number of
possible symmetric differences to 2nHpδ`εq .
Therefore, if dimpY q “ 1 and dimpXq “ s, then dpX, Y q ě H ´1 p1 ´ sq. We show
that in this case, the bound is tight. This will be based on the finite case:
Lemma 3.2. For any string y of length n and a given radius r, there is a string x
in Bpy, rq with dimpxq ď 1 ´ Hprq ` Oplog n{nq.
Lemma 3.2 can be proved using tools from the Vereshchagin-Vitányi theory
[VV10], which is surveyed in [VSar]. This is essentially the first part of Theorem
8 of [VSar], modified using the following facts. First, the class of Hamming balls
satisfies the conditions of Theorem 8 (see [VSar, Proposition 28]). Second, if V pn, rq
is the size of a Hamming ball of radius r contained in 2n , then logpV pn, rqq “
Hprqn ˘ Oplog nq (see [VSar, Remark 11]). Finally, the complexity of a Hamming
ball is within Oplog nq of the complexity of its center.
Another way to obtain Lemma 3.2 is by considering covering Hamming space
by balls. For any n and r P p0, 1{2q,
Ť let κpn, rq be the smallest cardinality of a
n
n
set C Ď t0, 1u such that t0, 1u “ xPC Bpx, rq. Delsarte and Piret ([DP86], see
[CHLL97, Thm.12.1.2]) showed that
κpn, rq ă 1 ` n2n ln 2{V pn, rq,
where recall from the introduction that V pn, rq is the size of the Hamming space of
radius r. As mentioned above, log V pn, rq ě Hprqn ´ Oplog nq (for fixed r), whence
logpκpn, rqq ď p1 ´ Hprqqn ` Oplog nq;
Lemma 3.2 follows by finding a witness for κpn, rq and giving appropriately short
descriptions to all the strings in that witness.
Now the counterpart to Proposition 3.1 follows.
Theorem 3.3. For any Y P 2ω there is some X P 2ω such that dimpXq ď s and
dpX, Y q ď H ´1 p1 ´ sq.
Note that if Y is random, then it must be the case that dimpXq “ s.
Proof. Let δ “ H ´1 p1 ´ sq. Given any Y , we build X by initial segments. Split
Y into chunks by cutting it at the locations nj as before. By Lemma 3.2, for each
chunk y from Y , find a chunk x in Bpy, δq with dimpxq ď s ` Oplog n{nq, where
n “ |y|, and append it to X. Then dpX, Y q ď δ, and dimpXq ď s, because each
chunk satisfies these (with Op logn n q error in the latter case), and Lemma 2.3 and
Lemma 2.4 apply.
On the other hand, Proposition 3.1 is not always optimal. This can be demonstrated with a simple error correcting code.
Proposition 3.4. There is a sequence Y P 2ω of dimension 1{2 such that dimpXq ą
0 for all X with dpX, Y q ď H ´1 p1{2q.
DIMENSION 1 SEQUENCES ARE CLOSE TO RANDOMS
13
Proof. if Y is the join of a random with itself pt “ 1{2q, and if s “ 0, suppose X is
such that dpX, Y q ď H ´1 p1{2q. Then Y æn can be given a description of length
KpX æn q ` H ´1 p1{2qn ` n{4 ` opnq.
Here the description first provides X æn . Then for each i such that Xp2iq ‰ Xp2i`1q
(there are at most H ´1 p1{2qn such i), it gives Y p2iq. Then it gives a description of
´1
ti : Xp2iq “ Xp2i ` 1q ‰ Y p2iqu, a subset of n{2 which has size at most H 2p1{2q n,
and therefore a description of length HpH ´1 p1{2qq n2 . Since H ´1 p1{2q ` 1{4 ă 1{2,
and for all n we have KpY æn q ě n{2, we have dimpXq ą 0.
In a weaker sense, however, the bound from Proposition 3.1 is always optimal.
Proposition 3.5. For all s ă t there are X, Y P 2ω with dimpXq “ s, dimpY q “ t
and dpX, Y q “ H ´1 pt ´ sq.
Proof. This is similar to the proof of Theorem 3.3, once we obtain the analogous
finite case. Let r “ H ´1 pt ´ sq. Find a witness C for κpn, rq; then
Ť find some
D Ď C of log-size sn ` Oplog nq maximising the size of SpDq “ xPD Bpx, rq.
Because we can guarantee to cover at least |D|
|C| of the space with SpDq, we have
log |SpDq| ě nt ´ Oplog nq. This gives:
Lemma 3.6. There are at least 2nt nOp1q many strings y P t0, 1un which are H ´1 pt´
sq-close to a string of dimension at most s ` Oplog n{nq.
Sequences X and Y as promised by Proposition 3.5 are constructed by chunks of
size i2 , as above. Having defined X ænj and Y ænj , we choose xj “ X ærnj ,nj`1 q and
yj “ Y ærnj ,nj`1 q so that dimpxj q ď s ` Oplog n{nq, dimpyj | Y ænj q ě t ´ Oplog n{nq
(where n “ j 2 ), and dpxj , yj q ď r.
4. Dimension s sequences and dimension t sequences
Finally we ask what density of changes are needed to turn a dimension s sequence
into a dimension t sequence (where t ą s). By the results of this paper, it is
equivalent to ask for the density of changes needed to turn a weakly s-random into
a weakly t-random.
Theorem 2.1 can be generalized. Recall that Bernoulli randoms show that the
following bound is optimal.
Theorem 4.1. For every sequence X with dimpXq “ s, and every t ą s, there is
a Y with dimpY q “ t and dpX, Y q ď H ´1 ptq ´ H ´1 psq.
In analogy with the proof of Theorem 2.1, in which the relative complexity of
each chunk of X was raised to 1 via whatever distance was necessary, one might
first consider raising the complexity of each chunk up to t. This fails because of a
failure of concavity. Given an individual chunk whose relative complexity si is less
than t, the density of changes needed to bring it up to t is δi “ H ´1 ptq ´ H ´1 psi q.
But when si ą t, no changes are needed, so we should choose δi “ 0. However, the
resulting function (mapping si to δi ) is not concave. So a tricky X could cause this
strategy to use distance greater than H ´1 ptq ´ H ´1 psq.
We use one of two different strategies, with the chosen strategy depending on
the particular s ă t pair. The first strategy is simple: raise the complexity of
each chunk as much as possible while staying within distance δ “ H ´1 ptq ´ H ´1 psq
14
N. GREENBERG, J.S. MILLER, A. SHEN, AND L.B. WESTRICK
of the given chunk. This strategy clearly produces a Y with dpX, Y q ď δ, but
showing that dimpY q ě t takes a little work and requires the assumption that
p1 ´ sqg 1 psq ď p1 ´ tqg 1 ptq, where gptq “ H ´1 ptq.
The second strategy is informed by the following reasoning. If for some j, we
have dimpX ænj q « s, then we should hope to have arranged that dimpY ænj q ě t,
since if we do so, then we have made the effective dimension of Y large enough. If
we have achieved this for Y ænj and then X’s next chunk is relatively random, we
can make Y ’s next chunk relatively random for free, so we may as well do so. This
has the effect that pdimpX ænj`1 q, dimpY ænj`1 qq lies on or above the line connecting
ps, tq to p1, 1q. In this case, our strategy is to use whatever density of changes are
necessary to keep pdimpX ænj`1 q, dimpY ænj`1 qq on or above this line. Then it
is clear that dimpY q ě t, but showing dpX, Y q ď δ requires a little work and the
assumption that p1 ´ tqg 1 ptq ď p1 ´ sqg 1 psq, complementary to the assumption under
which the first strategy works.
Proof of Theorem 4.1. Given X, let si be defined as in (1), and define ŝj by
ŝj “
j´1
1 ÿ 2
si i .
nj i“1
For notational convenience, define gptq “ H ´1 ptq.
Case 1. Suppose that p1 ´ sqg 1 psq ď p1 ´ tqg 1 ptq. We will construct Y P Pt̄ ,
where t̄ is the infinite sequence defined by
ti “ M psi , δq ` Op1{iq.
(The small extra factor is just to round ti up to a fraction of the form k{i.) Let εi Ñ
0 be an infinite sequence decreasing slowly enough that N “ i satisfies Lemma 2.6
for εi . Let δ̄ be defined by δi “ δ ` εi . By that lemma and the usual compactness
argument, there is Y P Pt̄ with dpX, Y q ď δ. We must show that if Y P Pt̄ , then
dimpY q ě t.
Let rpxq “ M px, δq, so that ti ě rpsi q. By Lemma 2.4,
dimpY q ě lim inf
j
j´1
1 ÿ
rpsi qi2 .
nj i“0
We would like it if rpxq were convex, so that we could conclude that if ŝj « s,
řj´1
then n1j i“0 rpsi qi2 ě rpŝj q « rpsq “ t. But it is not convex, so we use a convex
approximation. Let ℓpxq be the tangent line to rpxq at s. Note that since r is
increasing, ℓ has positive slope. Below we will show that ℓpxq ď rpxq on r0, 1s.
Assuming this, we can finish the argument. Whenever ŝj ď s, we have ℓpŝj q “
ℓpsq ˘ ε “ t ˘ ε, with ε vanishing in the limit infimum. On the other hand, if ŝj ą s,
then since ℓ is increasing and each ti ě ℓpsi q, we have
j´1
1 ÿ
rpsi qi2 ě ℓpŝj q ą ℓpsq “ t,
nj i“0
as required.
It remains to show that ℓpxq ď rpxq. The proof of the following lemma is
elementary but not short; here we state and use it, but delay a proof sketch to the
end of the section.
DIMENSION 1 SEQUENCES ARE CLOSE TO RANDOMS
15
Lemma 4.2. There is a point z P p0, 1q such that rpxq is convex on p0, zq and concave
on pz, 1q.
We can assume t ă 1 (if t “ 1, we are covered by Theorem 2.1), so in a neighborhood of s, rpxq “ Hpgpxq ` δq. We claim that r is convex at s. Consider the slope
of r at s. We have r1 pxq “ H 1 pgpxq ` δqg 1 pxq, so (using that H 1 pgpxqq “ 1{g 1 pxq)
r1 psq “ H 1 pgptqqg 1 psq “ g 1 psq{g 1 ptq.
By the case assumption, r1 psq is less than the slope of the line connecting ps, tq to
p1, 1q. Since r1 psq is also the slope of ℓ, it follows that ℓp1q ď 1, a fact we use later.
If r were concave already at s (and therefore also onwards), its graph would lie
below this line for all x ą s, so we would have rpxq ă 1 for all x P ps, 1q. But when
gpxq ` δ “ 1{2, rpxq “ 1, and this happens for some x P ps, 1q. Therefore, rpxq is
convex at s, so s ď z.
By convexity of r on p0, zq, ℓpxq ď rpxq on r0, zs. Also, ℓp1q ď 1 by the case
assumption, and rp1q “ 1. Since ℓpzq ď rpzq and ℓp1q ď rp1q, and since ℓ is linear
on rz, 1s while r is concave on that interval, ℓpxq ď rpxq on that interval. Therefore,
ℓpxq ď rpxq on r0, 1s, completing the proof of this case.
Case 2. Suppose that p1 ´ tqg 1 ptq ď p1 ´ sqg 1 psq. Let
t´s
1´t
x`
.
1´s
1´s
This is the line containing ps, tq and p1, 1q. We will construct Y P Pt̄ , where t̄ is the
infinite sequence defined by
ℓpxq “
ti “ ℓpsi q ` Op1{iq.
(Again, the extra factor is just to round ti up to a fraction of the form k{i.) By
linearity, if each psi , ti q is on or above the graph of ℓ, then
¸
˜ j´1
j´1
1 ÿ 2 1 ÿ 2
si i ,
ti i
nj i“1
nj i“1
is also. By Lemma 2.4, the limit infimum of the first coordinate is dimpXq, and
the limit infimum of the second coordinate is a lower bound for dimpY q. So if
dimpXq “ s, then dimpY q ě t as required. Considering also the density of changes,
we need to find Y P Pt̄ with dpX, Y q ď gptq ´ gpsq.
Let ε̄ Ñ 0 be an infinite sequence decreasing slowly enough that N “ i satisfies
Lemma 2.6 for εi . Let δ̄ be defined by
δi “ gpti q ´ gpsi q ` εi .
Observe that
M psi , δi ´ εq ` Op1{iq “ ti .
This is the complexity increase guaranteed by Lemma 2.6, so by that lemma and
the usual compactness argument, there is Y P Pt̄ with
dpX, Y q ď lim sup
j
Now, letting
j´1
1 ÿ 2
δi i .
nj i“1
ppxq “ gpℓpxqq ´ gpxq,
16
N. GREENBERG, J.S. MILLER, A. SHEN, AND L.B. WESTRICK
by the uniform continuity of g on r0, 1s, we may additionally assume that εi
decreases slowly enough that |x ´ y| ă 1{i implies |gpxq ´ gpyq| ă εi , so that
δi ď ppsi q ` 2εi . Since εi Ñ 0,
j´1
1 ÿ
ppsi qi2 .
dpX, Y q ď lim sup
nj i“1
j
The proof of the concavity of p is elementary but not short. We just use the
lemma here and give a sketch of the proof at the end of this section.
Lemma 4.3. The function p is concave on r0, 1s.
By the concavity of p, for each j we have
j´1
1 ÿ
ppsi qi2 ď ppŝj q.
nj i“1
When ŝj ď s, we have ppŝj q “ ppsq ˘ ε “ δ ˘ ε, with ε vanishing in the limit
supremum. But when ŝj ą s, we still need to bound ppŝj q ď δ “ ppsq. Here we use
the case assumption. We claim p is decreasing on rs, 1s, because p1 pxq ď 0 is true
exactly when
1´t
´ g 1 pxq ď 0,
g 1 pℓpxqq
1´s
which is satisfied when x “ s by assumption, and therefore satisfied for x ą s
because p is concave. Therefore, ppŝj q ď δ for ŝj ą s, so dpX, Y q ď δ.
Now we sketch the lemmas about convexity and concavity used above. The
proofs use only undergraduate calculus, mostly of a single variable.
Proof sketch of Lemma 4.2. Given rpxq “ M px, δq, we need to show there is a
z P p0, 1q such that r is convex on p0, zq and concave on pz, 1q. It suffices to
consider only what happens on the interval r0, Hp1{2 ´ δqs, since r is increasing and
rpxq “ 1 for x ě Hp1{2 ´ δq, continuing the concavity begun at z. For x in this
interval,
r2 pxq “ H 2 pgpxq ` δqpg 1 pxqq2 ` H 1 pgpxq ` δqg 2 pxq
Since these functions all reference gpxq, we make the substitution y “ gpxq. Note
that y P r0, 1{2 ´ δs. With this substitution,
r2 pxq “ H 2 py ` δq{pH 1 pyqq2 ` H 1 py ` δqp´H 2 pyqq{pH 1 pyqq3
Multiplying by lnp2qpH 1 pyqq3 and dividing by H 2 pyqH 2 py ` δq, r2 pxq shares its sign
with
wpyq “ f py ` δq ´ f pyq
where f pyq “ yp1 ´ yq log2 p1{y ´ 1q. When y “ 0, wpxq “ f pδq ą 0 (since δ ă
1{2). When y “ 1{2 ´ δ, wpxq “ ´f p1{2 ´ δq ă 0. (Perhaps neither r2 nor f
is strictly defined in these places, but the limits exist). So to show that r2 pxq is
positive on p0, zq and negative on pz, Hp1{2 ´ δqq for some z in p0, Hp1{2 ´ δqq,
since g 1 pxq is positive, it suffices to show that w is strictly decreasing. Equivalently,
f 1 py ` δq ´ f 1 pyq is negative; it suffices to show that f 1 pyq is strictly decreasing on
p0, 1{2q; equivalently f 2 pyq is strictly negative on this interval. But
f 2 pyq “ ´p1 ´ 2yq{pln 2py ´ y 2 qq ´ 2 log2 p1{y ´ 1q,
which is negative, so we are done.
DIMENSION 1 SEQUENCES ARE CLOSE TO RANDOMS
17
Proof sketch of Lemma 4.3. Letting ppxq “ gpℓpxqq ´ gpxq for ℓ with slope in r0, 1s,
we show that p is concave. Since ℓp1q “ 1, we write ℓpxq “ ax ` 1 ´ a. We need to
show
p2 pxq “ g 2 pax ` 1 ´ aqa2 ´ g 2 pxq
is non-positive. When a “ 1, p2 pxq ” 0, so defining kpa, xq “ g 2 pax ` 1 ´ aqa2 ´
2
g 2 pxq, it suffices to show that Bk
Ba pa, xq is non-negative for pa, xq P p0, 1s . We have
Bk
pa, xq “ g 3 pax ` 1 ´ aqpx ´ 1qa2 ` 2ag 2 pax ` 1 ´ aq,
Ba
which, with the substitution y “ gpax ` 1 ´ aq, simplifies to
`
˘
Bk
pa, xq “ a g 3 pHpyqqpHpyq ´ 1q ` 2g 2 pHpyqq .
Ba
So it suffices to show that this function is non-negative for all y P rgp1 ´ aq, 1{2s.
Expanding the computation of g 3 and g 2 , we have
Bk
pa, xq “ f pyq{papH 1 pyqq5 p1 ´ yq2 y 2 pln 2q2 q
Ba
where f pyq, which has the same sign as Bk{Ba, is
˙˙
ˆ
ˆ
1
´1
pHpyq ´ 1q
f pyq “ 3 ´ pln 2qp1 ´ 2yq log2
y
˙˙2
ˆ
ˆ
1
´1
` 2pln 2qyp1 ´ yq log2
y
It suffices to show that f pyq ě 0 on rgp1 ´ aq, 1{2s. Since f p1{2q “ 0, it suffices to
show that f 1 pyq ď 0. One may check that f 1 p1{2q “ 0, so it suffices to show that
f 2 pyq ě 0. We have f 2 pyq “ phpyq ` yp1 ´ yqq{py 2 p1 ´ yq2 q, where
hpyq “ lnp2 ´ 2yq ´ yp2 ´ 3y ` 2y 2 q lnp1{y ´ 1q.
It suffices to show that hpyq ě 0. Since hp1{2q “ 0, it suffices to show that h1 pyq ď 0.
One may check h1 p1{2q “ 0, so it suffices to show that h2 pyq ě 0. We have
h2 pyq “
3p1 ´ yqy lnp1{y ´ 1q ` p1 ´ 2yq
ě 0,
2yp1 ´ yqp1 ´ 2yq
completing the proof.
References
[BDS09]
Laurent Bienvenu, David Doty, and Frank Stephan. Constructive dimension and Turing
degrees. Theory Comput. Syst., 45(4):740–755, 2009.
[BFNV04] Harry Buhrman, Lance Fortnow, Ilan Newman, and Nikolay Vereshchagin. Increasing Kolmogorov complexity. Technical Report TR04-081, Electronic Colloquium on
Computational Complexity, 2004.
[CHLL97] Gérard Cohen, Iiro Honkala, Simon Litsyn, and Antoine Lobstein. Covering codes, volume 54 of North-Holland Mathematical Library. North-Holland Publishing Co., Amsterdam, 1997.
[DH10]
Rodney G. Downey and Denis R. Hirschfeldt. Algorithmic randomness and complexity.
Theory and Applications of Computability. Springer, New York, 2010.
[DP86]
Philippe Delsarte and Philippe Piret. Do most binary linear codes achieve the Goblick
bound on the covering radius? IEEE Trans. Inform. Theory, 32(6):826–828, 1986.
[FF81]
P. Frankl and Z. Füredi. A short proof for a theorem of Harper about Hamming-spheres.
Discrete Mathematics, 34:311–313, 1981.
18
N. GREENBERG, J.S. MILLER, A. SHEN, AND L.B. WESTRICK
[FHP` 11] Lance Fortnow, John M. Hitchcock, A. Pavan, N. V. Vinodchandran, and Fengming
Wang. Extracting Kolmogorov complexity with applications to dimension zero-one
laws. Inform. and Comput., 209(4):627–636, 2011.
[Gác86]
Péter Gács. Every sequence is reducible to a random one. Inform. and Control, 70(23):186–192, 1986.
[GM11]
Noam Greenberg and Joseph S. Miller. Diagonally non-recursive functions and effective
Hausdorff dimension. Bull. Lond. Math. Soc., 43(4):636–654, 2011.
[Har66]
L. H. Harper. Optimal numberings and isoperimetric problems on graphs. J. Combinatorial Theory, 1:385–393, 1966.
[JS12]
Carl G. Jockusch, Jr. and Paul E. Schupp. Generic computability, Turing degrees, and
asymptotic density. J. Lond. Math. Soc. (2), 85(2):472–490, 2012.
[Lev73]
L. A. Levin. On the notion of a random sequence. Soviet Math. Dokl., 14(5):1413–1416,
1973.
[Lut00]
Jack H. Lutz. Gales and the constructive dimension of individual sequences. In Automata, languages and programming (Geneva, 2000), volume 1853 of Lecture Notes in
Comput. Sci., pages 902–913. Springer, Berlin, 2000.
[LV93]
Ming Li and Paul Vitányi. An introduction to Kolmogorov complexity and its applications. Texts and Monographs in Computer Science. Springer-Verlag, New York, 1993.
[Mil11]
Joseph S. Miller. Extracting information is hard: a Turing degree of non-integral effective Hausdorff dimension. Adv. Math., 226(1):373–384, 2011.
[ML66]
Per Martin-Löf. The definition of random sequences. Information and Control, 9:602–
619, 1966.
[MM04]
Wolfgang Merkle and Nenad Mihailović. On the construction of effectively random
sets. J. Symbolic Logic, 69(3):862–878, 2004.
[MS77]
F. J. MacWilliams and N. J. A. Sloane. The theory of error-correcting codes. I. NorthHolland Publishing Co., Amsterdam-New York-Oxford, 1977. North-Holland Mathematical Library, Vol. 16.
[Sch71]
Claus Peter Schnorr. A unified approach to the definition of random sequences. Mathematical Systems Theory, 5(3):246–258, sep 1971.
[Sch72]
Claus Peter Schnorr. The process complexity and effective random tests. In Proceedings
of the 4th STOC, Denver, Colorado, pages 168–176. ACM, 1972.
[Sch73]
C.-P. Schnorr. Process complexity and effective random tests. J. Comput. System Sci.,
7:376–388, 1973. Fourth Annual ACM Symposium on the Theory of Computing (Denver, Colo., 1972).
[She15]
Alexander Shen. Around Kolmogorov complexity: Basic Notions and Results. In
Vladimir Vovk, Harris Papadopoulos, and Alexander Gammerman, editors, Measures
of Complexity. Festschrift for Alexey Chervonenkis, chapter 7, pages 75–116. Springer
International Publishing, 2015.
[UVS13] Vladimir A. Uspensky, Nikolai K. Vereshchagin, and Alexander Shen. Kolmogorov
complexity and algorithmic randomness (Russian title: Kolmogorovskaya slozhnost i
algoritmichskaya sluchainost). MCCME, 2013.
[VSar]
Nikolai Vereshchagin and Alexander Shen. Algorithmic statistics: forty years later. To
appear.
[VV10]
Nikolai K. Vereshchagin and Paul M. B. Vitányi. Rate distortion and denoising of individual data using Kolmogorov complexity. IEEE Trans. Inform. Theory, 56(7):3438–
3454, 2010.
DIMENSION 1 SEQUENCES ARE CLOSE TO RANDOMS
19
(N. Greenberg) School of Mathematics, Statistics and Operations Research, Victoria University of Wellington, Wellington, New Zealand
E-mail address: [email protected]
URL: http://homepages.mcs.vuw.ac.nz/~greenberg/
(J.S. Miller) Department of Mathematics, University of Wisconsin–Madison, 480
Lincoln Dr., Madison, WI 53706, USA
E-mail address: [email protected]
URL: http://www.math.wisc.edu/~jmiller/
(A. Shen) LIRMM, CNRS & University of Montpellier, 161 rue Ada, 34095, Montpellier, France. Supported by RaCAF ANR grant.
E-mail address: [email protected]
URL: http://www.lirmm.fr/~ashen
(L.B. Westrick) Department of Mathematics, University of Connecticut, 341 Mansfield Rd U-1009, Storrs, CT 06269, USA
E-mail address: [email protected]
URL: http://www.math.uconn.edu/~westrick/
| 7cs.IT
|
Robust Low-Rank Matrix Estimation
arXiv:1603.09071v4 [math.ST] 24 Jul 2017
Andreas Elsener and Sara van de Geer
Seminar for Statistics
ETH Zurich
8092 Zurich
Switzerland
e-mail: [email protected]
e-mail: [email protected]
Abstract: Many results have been proved for various nuclear norm penalized estimators of the uniform sampling matrix completion problem.
However, most of these estimators are not robust: in most of the cases the
quadratic loss function and its modifications are used. We consider robust
nuclear norm penalized estimators using two well-known robust loss functions: the absolute value loss and the Huber loss. Under several conditions
on the sparsity of the problem (i.e. the rank of the parameter matrix) and
on the regularity of the risk function sharp and non-sharp oracle inequalities for these estimators are shown to hold with high probability. As a
consequence, the asymptotic behavior of the estimators is derived. Similar
error bounds are obtained under the assumption of weak sparsity, i.e. the
case where the matrix is assumed to be only approximately low-rank. In all
our results we consider a high-dimensional setting. In this case, this means
that we assume n ≤ pq. Finally, various simulations confirm our theoretical
results.
MSC 2010 subject classifications: Primary 62J05, 62F30; secondary
62H12.
Keywords and phrases: Matrix completion, robustness, empirical risk
minimization, oracle inequality, nuclear norm, sparsity.
1. Introduction
1.1. Background and Motivation
Netflix, Spotify, Apple Music, Amazon and many other on-line services offer an
almost infinite amount of songs or films to their users. Clearly, a single person
will never be able to watch every film or to listen to every available song. For
this reason, an elaborate recommendation system is necessary in order to allow
the users to choose content that already match his or her preferences. Many
models and estimation methods have been proposed to address this question.
Matrices provide an appropriate way of modelling this problem. Imagine that
the plethora of films/songs is identified with the rows of a matrix, call it B ∗ ,
and the users with its columns. One entry of the matrix corresponds to the
rating given to film “i” (row) by user “j” (column). This matrix will have many
missing entries. These entries are bounded and we can expect the rows of B ∗ to
be very similar to each other. It is therefore sensible to assume that B ∗ has a
low rank. The challenge is now to predict the missing ratings/fill in the empty
1
Elsener and van de Geer/Robust Low-Rank Matrix Estimation
2
entries of B ∗ . Define for this purpose the set of observed (possibly noisy) entries
A := {(i, j) ∈ {1, . . . , p} × {1, . . . , q} : the (noisy) entry
(1.1)
∗
Aij of B is observed} ,
where p is the number of films/songs and q the number of users. Our estimation
problem can therefore be stated in the following way: for B ∈ B ⊂ Rp×q we
minimize Rn (B), subject to rank(B) = s.
In this special case we have that
B = B ∈ Rp×q |kBk∞ ≤ η ,
(1.2)
where η is for instance the mean highest rating, and Rn is some convex empirical
error measure that is defined by the data, e.g.
Rn (B) =
1 X
(Aij − Bi,j )2 .
|A|
(1.3)
(i,j)∈A
Since the rank of a matrix is not convex we use the nuclear norm as its convex
surrogate. This leads us to a relaxed convex optimization problem. For B ∈ B
we
minimize Rn (B), subject to kBknuclear ≤ τ,
for some τ > 0. The model described above can be considered as a special case
of the trace regression model.
In the trace regression model (see e.g. Rohde and Tsybakov (2011)) one considers the observations (Xi , Yi ) satisfying
Yi = trace(Xi B ∗ ) + εi , i = 1, . . . , n,
(1.4)
where εi are i.i.d. random errors. The matrices Xi are so-called masks. They
are assumed to lie in
χ = ek (q)el (p)T : 1 ≤ k ≤ q, 1 ≤ l ≤ p ,
(1.5)
where ek (q) is the q-dimensional k-th unit vector and el (p) is the p-dimensional
l-th unit vector. We will assume that the Xi are i.i.d. with
1
P Xikj = 1 = 1 − P Xikj = 0 =
pq
for all i ∈ {1, . . . , n}, k ∈ {1, . . . q}, and j ∈ {1, . . . , p}. However, we point
out that it is not necessary for our estimators to know this distribution. This
knowledge will only be used in the proofs of the theoretical results.
The trace regression model together with the space χ and the distribution
on χ is equivalent to the matrix completion case. The entries of the vector Y
can be identified with the observed entries as those in the matrix A.
Elsener and van de Geer/Robust Low-Rank Matrix Estimation
3
From this, it can be seen that we are in a high-dimensional setting since the
number of observations n must be smaller than or equal to the total number of
entries of A. The setup described above is then called uniform sampling matrix
completion. A very similar setup was first considered in Srebro, Rennie and
Jaakkola (2005) and Srebro and Shraibman (2005).
As in the standard regression setting, parameter estimation in the trace
regression model is also done via empirical risk minimization. Using the Lagrangian form for B ∈ B we
minimize Rn (B) + λkBknuclear ,
Pn
where Rn (B) = 1/n i=1 ρ(Yi − trace(Xi B)), ρ is a convex loss function and
λ > 0 is the tuning parameter. The loss function is often chosen to be the
quadratic loss (or one of its modifications) as in Koltchinskii, Lounici and Tsybakov (2011); Negahban and Wainwright (2011, 2012); Rohde and Tsybakov
(2011) and many others. In Lafond (2015) the case of an error distribution belonging to an exponential family is considered. As long as the errors are assumed
to be light tailed as it is the case for i.i.d. Gaussian errors the least squares estimator will perform very well. However, the ratings are heavily subject to frauds
(e.g. by the producer of a film). It is necessary to take this fact into account
also in the estimation procedure. One might also be interested in estimating the
median or another quantile of the ratings. For this purpose, M-estimators based
on different losses than the quadratic loss are usually chosen.
1.2. Proposed estimators
In this paper, we consider the absolute value loss and the Huber loss. The first
robust estimator is then given by
n
B̂ := arg min
B∈B
1X
|Yi − trace (Xi B)| + λ kBknuclear .
n i=1
(1.6)
Using the Huber loss we can define
n
B̂H := arg min
B∈B
1X
ρH (Yi − trace (Xi B)) + λ kBknuclear ,
n i=1
(1.7)
where the function
ρH (x) :=
x2 ,
if |x| ≤ κ
2κ |x| − κ2 , if |x| > κ
defines the Huber loss function. The tuning parameter κ > 0 is assumed to be
given for our estimation problem. The possible values for the Huber parameter
κ depend on the distribution of the errors as shown in Lemma 2.1. In practice,
one usually estimates κ and λ with methods such as cross-validation. Notice
that it could happen that the estimators defined in Equations 1.6 and 1.7 are
Elsener and van de Geer/Robust Low-Rank Matrix Estimation
4
not unique since the objective functions are not strictly convex. As will be
shown, the rates depend on the Lipschitz constants of the loss functions and
on η. Typically, the Lipschitz constants of the absolute value loss as well as of
the Huber loss induce smaller constants in the rates compared to the Lipschitz
constant of the truncated quadratic loss.
The “target” is defined as
B 0 := arg min R(B),
B∈B
where R(B) = ERn (B) is the theoretical risk. It has to be noticed that the
matrix B 0 is not necessarily equal to the matrix B ∗ . Our main interest lies in
the theoretical analysis of the above estimators. The estimators should mimic
the sparsity behavior of an oracle B. In the case of the absolute value loss we
will prove a non-sharp oracle inequality. Whereas for the Huber loss, thanks to
its differentiability, we are able to derive a sharp oracle inequality.
Assuming B = B 0 in Corollary 3.1 the upper bound is typically of the form
R(B̂H ) − R(B 0 ) . λ2 pqs0 ,
(1.8)
where . means that some multiplicative constants (depending on the tuning
parameter κ) are omitted.
The assumptions for this kind of results are mainly based on the regularity of the absolute value and the Huber loss. Moreover, the properties of the
nuclear norm, which are very similar to those of the `1 -norm for vectors, will
be exploited. In addition, we use the sparsity behavior induced by the nuclear
norm to infer the behavior of weakly sparse estimators. This takes into account
that a matrix could have few very large singular values and many small, but
not exactly zero singular values:
q
X
B ∈ B0 ∈ B :
Λrj ≤ ρrr , Λ1 , . . . , Λq the singular values of B 0 ,
j=1
where 0 < r < 1 and ρrr is some reasonably small constant.
1.3. Related Literature
A first study with robust matrix estimation was made in Chandrasekaran et al.
(2011) in a setting with no missing entries. In order to avoid identifiability
issues the authors introduce “incoherence” conditions on the low-rank component. These conditions make sure that the low-rank component itself is not too
sparse. The locations of the corruptions are assumed to be fixed. In the context
of Principal Component Analysis (PCA) which is a special case of the matrix regression model robustness was investigated in Candès et al. (2011). The authors
assume that the matrix to be estimated is decomposed in a low-rank matrix and
a sparse matrix. In contrast to Chandrasekaran et al. (2011) the non-zero entries
Elsener and van de Geer/Robust Low-Rank Matrix Estimation
5
of the sparse matrix are assumed to be drawn randomly following a uniform distribution. Following this line of research Li (2011) apply these conditions to the
matrix completion problem with randomly observed entries. In a parallel work
Chen et al. (2013) consider the case where the indices of the observed entries
may be simultaneously both random and deterministic. In these papers only
noiseless robust matrix completion is considered.
Cambier and Absil (2016) study computational aspects of robust matrix completion (in the previously mentioned setting). A method relying on Riemannian
optimization is proposed. The authors assume the rank of the matrix to be
estimated to be known.
In Foygel et al. (2011) weighted nuclear norm penalized estimators with (possibly nonconvex) Lipschitz continuous loss functions are studied from a learning
point of view. The partially observed entries are assumed to follow a possibly
non-uniform distribution on the set χ. In contrast, our derivations rely among
other properties on the convexity of the risk (i.e. the margin conditions).
Noisy robust matrix completion was first investigated in Klopp, Lounici and
Tsybakov (2016). The authors assume that the truth A∗ is decomposed in a
low-rank matrix and a sparse matrix where the low-rank matrix contains the
“parameters of interest” and the sparse matrix contains the corruptions. In addition, every observation is corrupted by independent and centered subgaussian
noise. The largest entries of both the low-rank and sparse matrices are assumed
to be bounded (e.g. by the maximal possible rating). Their model is as follows:
(Xi , Ỹi ), i = 1, . . . N satisfy
Ỹi = trace(Xi A∗ ) + ξi , i = 1, . . . , N,
(1.9)
where A∗ = L∗ + S ∗ with L∗ a low-rank matrix and S ∗ a matrix with entrywise
sparsity. Columnwise sparsity is also considered but in view of a comparison of
this and our approach we prefer to restrict to entrywise sparsity. The masks
Xi are assumed to lie in the set χ (1.5) and to be independent of the noise
ξi for all i. The set of observed indices is assumed to be the union of two
disjoint components Ξ and Ξ̃. The set Ξ corresponds to the non-corrupted noisy
observations (i.e. only entries of L∗ plus ξi ). The entries corresponding to these
observations of S ∗ are zero. The set Ξ̃ contains the indices of the observations
that are corrupted by a (nonzero) entry of S ∗ . It is not known if an observation
comes from Ξ or Ξ̃. The estimator given in Klopp, Lounici and Tsybakov (2016)
is
(
(L̂, Ŝ) ∈ arg min
kLk∞ ≤η
kSk∞ ≤η
N
1 X
(Ỹi − trace(Xi (L + S)))2 + λ1 kLknuclear + λ2 kSk1
N i=1
)
. (1.10)
In contrast to the previously mentioned papers on robust matrix completion,
we consider (possibly heavy-tailed) random errors that affect the observations
but not the truth.
Elsener and van de Geer/Robust Low-Rank Matrix Estimation
6
1.4. Organization of the paper
The paper is organized in the following way. We state the main assumptions
that are used throughout the paper in Section 2. Then, the nuclear norm, its
properties, and its similarities to the `1 -norm are discussed. To bound the empirical process part resulting from the matrix completion setting we make use
of the results in Section B.3 of the Supplementary Material (Appendix B). In
Section 3 the main theorems are presented: the (deterministic) sharp and nonsharp oracle inequalities. In Section 3.3 we present the applications of these
results to the case of Huber loss and absolute value loss. The asymptotics and
the applications to weak sparsity are presented in Section 4. Finally, to verify
the theoretical findings, Section 5 presents some simulations. The Student t distribution with three degrees of freedom is considered as an error distribution.
2. Preliminaries
In this section the assumptions on the loss functions, the risk, and the distribution of the errors are presented. In particular, Assumptions 1-3 below are on
the curvature of the (theoretical) risk. They are used to derive the deterministic sharp and non-sharp oracle inequalities. It is important to notice that the
curvature of the risk mainly depends on the properties of the distribution of the
errors. Assumptions 4 and 5 below will be shown to be sufficient for Assumptions
2 and 3 to hold, respectively.
Furthermore, we also discuss the properties of the nuclear norm. Thanks
to the penalization term in the objective functions the optimization problems
become computationally tractable. We also highlight the commonalities of the
vector `1 -norm and the nuclear norm for matrices.
2.1. Assumptions on the risk and the distribution of the errors
The first assumption is about the loss function.
Assumption 1. Let ρ be the loss function. We assume that it is Lipschitz
continuous with constant L, i.e. that for all x, y ∈ R
|ρ(x) − ρ(y)| ≤ L|x − y|.
(2.1)
The next two assumptions ensure the identifiability of the parameters by
requiring a sufficient convexity of the risk around the target.
Assumption 2. One-point-margin condition. There is an increasing
strictly convex function G with G(0) = 0 such that for all B ∈ B
R (B) − R B 0 ≥ G kB − B 0 kF ,
where R is the theoretical risk function.
Elsener and van de Geer/Robust Low-Rank Matrix Estimation
7
Assumption 3. Two-point-margin condition. There is an increasing
strictly convex function G with G(0) = 0 such that for all B, B 0 ∈ B we have
T
R (B) − R (B 0 ) ≥ trace Ṙ (B 0 ) (B − B 0 ) + G (kB − B 0 kF ) ,
where R is the theoretical risk function and [Ṙ(B 0 )]kl =
∂
0
∂Bkl R(B)|B=B .
Assumption 1 is crucial when it comes to the application of the Contraction
Theorem which in turn allows us to apply the dual norm inequality to find
a bound for the random part of the oracle bounds. Assumptions 2 and 3 are
essential in the proofs of the (deterministic) results. In particular, in addition
to the differentiability of the empirical risk Rn , Assumption 3 is responsible
for the sharpness of the first oracle bound that will be proved. The margin
conditions are strongly related to the shape of the distribution function and the
corresponding density of the errors.
For the specific application to the Huber loss and absolute value loss estimators we show that mild conditions on the distribution of the errors ensure a
sufficient curvature of the risk for both loss functions under study.
Assumption 3 holds under a weak condition on the distribution function of
the errors:
Assumption 4. Assume that there exists a constant C1 > 0 such that the
distribution function F with density with respect to Lebesgue measure f of the
errors fulfills
F (u + κ) − F (u − κ) ≥ 1/C12 , for all |u| ≤ 2η and κ ≤ 2η.
(2.2)
Lemma 2.1. Assumption 4 implies Assumption 3 with G(u) = u2 /(2C12 pq).
The following assumption guarantees that Assumption 2 holds.
Assumption 5. Suppose ε1 , . . . , εn are i.i.d. with median zero and density f
with respect to Lebesgue measure. Assume that for C2 > 0
f (u) ≥
1
, for all |u| ≤ 2η.
C22
(2.3)
Lemma 2.2. Assumption 5 implies Assumption 2 with G(u) = u2 /(2C22 pq).
Another important fact is that when the distribution of the errors is assumed
to be symmetric around zero B ∗ = B 0 . This phenomenon is discussed in Section
4 of the Supplement.
2.2. Properties of the nuclear norm
The regularization by the nuclear norm plays a similar role as the `1 -norm in the
Lasso (Tibshirani, 1996). We illustrate the similarities and differences of these
types of regularizations. In view of the oracle inequalities and in order to keep
the notation as simple as possible we discuss the properties of the nuclear norm
Elsener and van de Geer/Robust Low-Rank Matrix Estimation
8
of the oracle. The oracle is typically a value B that takes an up-to-constants
optimal trade-off between approximation error and estimation error. In what
follows, B is called “the oracle” although its choice is flexible.
Consider the singular value decomposition of the oracle B with rank s? given
by
B = P ΛQT ,
(2.4)
where P is a p × q matrix, Q a q × p matrix and Λ a q × q diagonal matrix
containing the ordered singular values Λ1 ≥ · · · ≥ Λq . Then the nuclear norm is
given by
q
X
kBknuclear =
Λi (B) = kΛ(B)k1 ,
(2.5)
i=1
by interpreting Λ(B) ∈ Rq as the vector of singular values. The penalization
with the nuclear norm induces sparsity in the singular values, whereas the penalization with the vector `1 -norm of the parameters in linear regression induces
sparsity directly in the parameters. On the other hand, the rank plays the role
of the number of non-zero coefficients in the Lasso setting, namely
s? = kΛ(B)k0 .
(2.6)
One main ingredient of the proofs of the oracle inequalities is the so-called
triangle property as introduced in van de Geer (2001). This property was used
in e.g. Bühlmann and van de Geer (2011) to prove non-sharp oracle inequalities.
For the `1 -norm the triangle property follows from its decomposability. For the
nuclear norm the triangle property as it is used in this work depends on the
features of the oracle B. For this reason we notice that for any positive integer
s ≤ q the oracle can be decomposed in
B = B+ + B−,
B+ =
s
X
Λk Pk QTk ,
B− =
k=1
q
X
Λk Pk QTk .
k=s+1
The matrix B is called “active” part of the oracle B, whereas the matrix B −
is called the “non-active” part. The singular value decomposition of B + is given
by
T
B + = P + ΛQ+ .
+
We observe that the integer s is not necessarily the rank of the oracle B.
The choice of s is free. One may choose a value that trades off the roles of
the “active” part B + and “non-active” part B − , see Lemma 4.1. The following
lemma is adapted from Lemma 7.2 and Lemma 12.5 in van de Geer (2016).
Lemma 2.3. Let B + ∈ Rp×q be the active part of the oracle B. Then we have
for all B 0 ∈ Rp×q with
√ + +T 0
T
T
T
0
s P P B
+ B 0 Q+ Q+
+ P + P + B 0 Q+ Q+
Ω+
B + (B ) :=
F
F
and
0
Ω−
B + (B ) :=
T
I − P +P +
T
B 0 I − Q+ Q+
nuclear
F
Elsener and van de Geer/Robust Low-Rank Matrix Estimation
9
that
0
+
0
B + nuclear − kB 0 knuclear ≤ Ω+
− Ω−
B+ B − B
B + (B ) .
We then say that the triangle property holds at B + .
−
0
p×q
In particular, since Ω+
B + (B ) = 0, we have for any B ∈ R
(2.7)
−
0
0
−
kBknuclear − kB 0 knuclear ≤ Ω+
B + (B − B) − ΩB + (B − B) + 2 B
nuclear .
(2.8)
Moreover, we have
−
k·knuclear ≤ Ω+
(2.9)
B + + ΩB + .
−
−
From now on, we write Ω+ = Ω+
B + and Ω = ΩB + . Equation 2.8 is proved
in Appendix A.
Hence, the property that our estimators should mimic is not the rank of the
oracle but rather the fact that the “non-active” part is zero under the semi-norm
induced by the active part.
Moreover, we define the norm Ω as
Ω := Ω+ + Ω− .
Remark 1. Notice that the semi-norms Ω+ and Ω− form a complete pair,
meaning that Ω := Ω+ + Ω− is a norm.
The estimation error in several different norms can thus be “computed” in
general (semi-)norms.
A tail bound for the maximal singular value of a finite sum of matrices lying
in the set χ defined in Equation (1.5) is given in the following theorem. For
this purpose, we first need to define the Orlicz norm of a random variable. Let
Z ∈ R be a random variable and α ≥ 1 a constant. Then the Ψα −Orlicz norm
is defined as
α
(2.10)
kZkΨα := inf {c > 0 : E exp [|Z| /cα ] ≤ 2}
Theorem 2.1 ( Proposition 2 in Koltchinskii, Lounici and Tsybakov (2011) ). Let
X1 , . . . , Xn be i.i.d. q × p matrices that satisfy for some α ≥ 1 (and all i)
EXi = 0, kΛmax (Xi )kΨα =: K < ∞.
Define
(
2
S := max Λmax
n
X
i=1
!
EXi XiT
/n, Λmax
n
X
!
EXiT Xi
)
/n .
i=1
Then for a constant C and for all t > 0
!
r
n
X
t + log(p + q)
P Λmax
Xi /n ≥ CS
n
i=1
K
t + log(p + q)
+C log1/α
≤ exp(−t).
S
n
This theorem is used with the tail summation property of the expectation in
the derivations of the tail bounds in Section 3 of the Supplementary Material.
Elsener and van de Geer/Robust Low-Rank Matrix Estimation
10
3. Oracle inequalities
We first give two deterministic sharp and non-sharp oracle inequalities. The connection to the empirical process parts and to the specific loss functions follows
in Subsection 3.3. Let B 0 = arg min R(B 0 ) be the target. It is assumed that
B 0 ∈B
q ≤ p.
3.1. Sharp oracle inequality
Here, we assume that the loss function is differentiable and Lipschitz continuous. The next lemma gives a connection between the empirical risk and the
penalization term.
Lemma 3.1 (adapted from Lemma 7.1 in van de Geer (2016)). Suppose that
Rn is differentiable. Then for all B ∈ B
− trace Ṙn (B̂)T (B − B̂) ≤ λkBknuclear − λkB̂knuclear .
The following theorem is inspired by Theorem 7.1 in van de Geer (2016). In
contrast to this theorem, we need to bound the empirical process part differently. In view of the application to the matrix completion problem, we assume
a specific bound on the empirical process.
Theorem 3.1. Suppose that Assumptions 1 and 3 hold, that the loss function
is differentiable, and let H be the convex conjugate of G. Assume further for all
B 0 ∈ B that for λε > 0 and λ∗ > 0
T
trace
Ṙn (B 0 ) − Ṙ (B 0 ) (B − B 0 ) ≤ λε Ω(B 0 − B) + λ∗ .
Take λ > λε . Let 0 ≤ δ < 1 be arbitrary, and define
λ := λ − λε ,
λ := λε + λ + δλ
Then
δλΩ+ (B̂ − B) + δλΩ− (B̂ − B) + R(B̂) − R(B)
√
≤ H(λ3 s) + 2λ B − nuclear + λ∗ .
In the proof of this theorem the differentiability of the loss function and
Assumption 3 are crucial. Without this property an additional term arising
from the one-point-margin condition would appear in the upper bound. This
term would then lead to a non-sharp bound.
Elsener and van de Geer/Robust Low-Rank Matrix Estimation
11
3.2. Non-sharp oracle inequality
Instead of bounding an empirical process term depending on the derivative of the
empirical and theoretical risks we need to consider differences of these functions.
Theorem 3.2. Suppose that Assumptions 1 and 2 hold. Let H be the convex
conjugate of G. Suppose further that for λε > 0, λ∗ > 0, and all B 0 ∈ B
|[Rn (B 0 ) − R(B 0 )] − [Rn (B) − R(B)]| ≤ λε Ω(B 0 − B) + λ∗ .
(3.1)
Let 0 < δ < 1, take λ > λε and define
λ = λ + λε ,
λ = λ − λε .
(3.2)
Then,
√
δλΩ(B̂ − B) ≤ 2H(λ(1 + δ)3 s)
+ 2 λ∗ + (R(B) − R(B 0 )) + 4λkB − knuclear
and
R(B̂) − R(B) ≤
√
1
2H(λ(1 + δ)3 s) + λ∗ + 2(R(B) − R(B 0 ))
δ
+2λkB − knuclear + λ∗ + 2λkB − knuclear .
It has to be noticed that the above bound is “good” only if R(B) − R(B 0 ) is
already small. The main cause for the non-sharpness is Assumption 2 that leads
to an additional term in the upper bound of the inequality.
3.3. Applications to specific loss functions
We now apply the deterministic sharp and non-sharp oracle inequalities to the
case of the Huber loss and absolute value loss, respectively. We assume in both
cases that the distribution of the errors is symmetric around 0 so that B 0 = B ∗ .
This is discussed in detail in Section B.4.
Huber Loss-sharp oracle inequality
We first consider the case that arises by choosing the Huber loss. Theorem 3.1
together with Lemma 2.1 and the first claim of Lemma 3.2 in the Supplement
imply the following corollary. It is useful to notice that the Lipschitz constant
of the Huber loss is 2κ.
Corollary 3.1. Let B = B + + B − where B + and B − are defined in Equation
2.4. Let Assumption 4 be satisfied.
For a constant C0 > 0 let
s
!
p
√
log(p + q)
log(p + q)
λε = 2(4η + 2κ) (8C0 + 2)
+ 8C0 log(1 + q)
nq
n
Elsener and van de Geer/Robust Low-Rank Matrix Estimation
12
and λ∗ = 8η(4η + 2κ)p log(p + q)/(3n) + λε .
Assume that λ > λε . Take 0 ≤ δ < 1,
λ := λ − λε and λ := λε + λ + δλ
√
Choose j0 := dlog2 (7q pqη)e and define
(3.3)
α = (j0 + 2) exp(−p log(p + q)).
Then we have with probability at least 1 − α that
δλΩ+ (B̂H − B) + δλΩ− (B̂H − B) + R(B̂H ) − R(B)
pqC12 λ̄2 9s
+ 2λ B − nuclear + λ∗ .
2
Assumption 4 guarantees that the risk function is sufficiently convex. From
this assumption we also obtain a bound for the possible values of the tuning
parameter κ. We can also see that the results hold for errors with a heavier
tail than the Gaussian. The choice of the noise level λε and consequently of the
tuning parameter λ results from the the probability inequalities for the empirical
process in Section B.3. The quantity λ∗ is also a consequence of the bound on
the empirical process part. However, it does not affect the asymptotic rates.
≤
Absolute value loss - non-sharp oracle inequality
The next corollary is an application to the case of the absolute value loss.
Theorem 3.2 combined with Lemma 2.2 and the second claim of Lemma 3.2 in
the Supplement lead to the following corollary. The Lipschitz constant in this
case is 1.
Corollary 3.2. Let the oracle B be as in 2.4. Suppose that Assumption 5 is
satisfied. For a constant C0 > 0 let
s
!
p
√
log(p + q)
log(p + q)
λε = 2 (8C0 + 2)
+ 8C0 log(1 + q)
nq
n
and λ∗ = 8ηp log(p + q)/(3n) + λε . Take 0 < δ < 1 and λ > λε . Choose
√
j0 := dlog2 (7q pqη)e and define
α = (j0 + 2) exp(−p log(p + q)).
Then we have with probability at least 1 − α that
δλΩ(B̂ − B)
2
2
≤ 6C 2 λ (1 + δ) pqs + 2λ∗ + 2(R(B) − R(B 0 )) + 4λkB − knuclear
and
R(B̂) − R(B) ≤
1h 2 2
2
6C2 λ (1 + δ) pqs + λ∗ + 2(R(B) − R(B 0 ))
δ
i
+ 2λkB − knuclear + λ∗ + 2λkB − knuclear .
Elsener and van de Geer/Robust Low-Rank Matrix Estimation
13
Also in this case, the choices of λε and λ∗ are a consequence of the probability
bounds.
4. Asymptotics and Weak sparsity
The results in Section 3 are valid for finite values of the dimension of the matrix
p, q, the rank, and the number of observed entries n. A question that is answered
in this section is how the estimation errors of the proposed estimators behave
when n, p, and q are allowed to grow.
As mentioned in Negahban and Wainwright (2012), practical reasons motivate the assumption that the matrix B 0 is not exactly low-rank but only
approximately. In relation to the matrix completion problem one observes that
the ratings given by the users are unlikely to be exactly equal but rather very
similar. This translates to a matrix that is not low-rank. However, it is sensible to assume that the matrix is almost low-rank. The notion of weak sparsity
quantifies this assumption by assuming that for some 0 < r < 1 and ρ > 0
q
X
Λ0k
r
=: ρrr ,
(4.1)
k=1
where Λ01 , . . . , Λ0q are the singular values of B 0 . For r = 0 we have under the
convention that 00 = 0 that
q
q
X
0 X
Λ0k =
1{Λ0 >0} = s0 ,
k
k=1
k=1
where s0 is the rank of B 0 . The following lemma gives a bound of the non-active
part of the matrix B that appears in the oracle bounds.
Lemma 4.1. For σ > 0 we may take
B − nuclear ≤ σ 1−r ρrr ,
(4.2)
and
s ≤ σ −r ρrr .
We first consider the asymptotic behavior of our estimators in the case of an
exactly low-rank matrix and deduce from this the asymptotics for the case of
an approximately low-rank matrix.
4.1. Asymptotics
4.1.1. sharp
n
, and therefore using
By Corollary 3.1, assuming that q log(1 + q) = o log(p+q)
the choice for the noise level
s
log(p + q)
λε
nq
Elsener and van de Geer/Robust Low-Rank Matrix Estimation
14
we obtain
R(B̂H ) − R(B 0 ) ≤ R(B) − R(B 0 )
+ OP
ps log(p + q)
+
n
s
!
log(p + q)
1 + B − nuclear .
nq
(4.3)
We choose for simplicity the oracle to be the matrix B 0 itself with s0 =
rank(B 0 ). Then, we make use of the two point margin condition that is shown
to hold in Lemma 2.1.The resulting rate is then given by
2
0 2
2 4 p qs0 log(p + q)
kB̂H − B kF = OP (4η + 2κ) C1
,
(4.4)
n
where κ is the Huber parameter and C1 is the constant from Lemma 2.1.
Remark 2. The rate (4.4) depends on η as in Koltchinskii, Lounici and Tsybakov (2011) and on the Lipschitz constant of the loss function which is typically
smaller than η. If C12 = O(η), the constant in front of the rate is of order O(η 4 ).
This is a “worst-case” scenario that shows the cost that is paid when allowing
for very general error distributions as in our case. We emphasize that in this
case the distribution of the errors is not required to have a density.
In addition to the rate obtained for the Frobenius norm, we are also able to
derive rates for the estimation error measured in nuclear norm. From Corollary
3.1 and Equation 2.9 under the previous conditions it follows that
s
!
log(p + q)
0
2
.
(4.5)
kB̂H − B knuclear = OP C1 pqs0
nq
4.1.2. non-sharp
By
Corollary 3.2 it
is
known
that
the
assumption
p
n
q log(1 + q) = o log(p+q)
leads to the choice λε log(p + q)/nq. Therefore, we have
R(B̂) − R(B 0 )
= OP
ps log(p + q)
+ R (B) − R B 0 +
n
s
!
log(p + q)
−
(1 + B nuclear ) .
nq
(4.6)
What can be observed comparing the rates in Equations (4.3) and (4.6) is the
presence of the additional term R(B) − R(B 0 ) in the non-sharp case in contrast
to the sharp case. We choose again the oracle to be the matrix B 0 itself. By
Elsener and van de Geer/Robust Low-Rank Matrix Estimation
15
the one point margin condition derived in Lemma 2.2 we see that the rate of
convergence in this case is given by
2
0 2
4 p qs0 log(p + q)
kB̂ − B kF = OP C2
,
(4.7)
n
where the constant C2 comes from Lemma 2.2.
Remark 3. If C22 = O(η) a comparison with the rates obtained in Koltchinskii, Lounici and Tsybakov (2011) shows that the rates agree. In contrast to
the rate obatined for the Huber loss (Equation 4.4), the distribution of the errors is assumed to have a density. This leads to a constant of order O(η 2 ) in
a “worst-case” scenario. It is a natural consequence of the stronger assumption
on the distribution of the errors. This is comparable to the constant obtained in
Koltchinskii, Lounici and Tsybakov (2011).
In analogy to the previous case, we are able to derive a rate for the estimation
error measured in nuclear norm:
s
!
log(p + q)
0
2
.
(4.8)
kB̂ − B knuclear = OP C2 pqs0
nq
The rates are indeed very slow but this is not surprising given that per entry
the number of observations is about n/(pq). The price to pay for the estimation
of the reduced number of parameters ps0 is given by the term log(p + q).
4.2. Weak sparsity
In what follows, the asymptotic behavior of the proposed estimators is discussed
when applied to an estimation problem where one aims at estimating a matrix
that is not exactly low-rank. With Lemma 4.1 and the rates given in the previous
section we are able to derive an explicit rate also for the approximatley low-rank
case. For this purpose, we assume that Equation 4.1 holds.
4.2.1. Huber estimator
The following corollary gives rates for the estimation error of the Huber estimator when used for estimation of a not exactly low-rank matrix.
n
Corollary 4.1. With q log(1 + q) = o log(p+q)
we choose
s
λε
log(p + q)
.
nq
We then have
B̂H − B
0
2
F
1−r
2
2 4 p q log(p + q)
ρrr .
= OP (η + κ) C1
n
Elsener and van de Geer/Robust Low-Rank Matrix Estimation
16
4.2.2. Absolute value estimator
Using the oracle inequality under the weak sparsity assumption we obtain the
following result.
n
we choose
Corollary 4.2. With q log(1 + q) = o log(p+q)
s
λε
log(p + q)
.
nq
Then we have for the Frobenius norm of the estimation error
2
1−r
2
p q log(p + q)
= OP
B̂ − B 0
ρrr .
n
F
(4.9)
5. Simulations
In this section, the robustness of the Huber estimator 1.7 is empirically demonstrated. In Subsection 5.3, the Huber estimator is compared with the estimator
proposed in Klopp, Lounici and Tsybakov (2016) under model 1.4 and 1.9 with
each Student t and standard Gaussian noise. The sample size ranges in all
simulations for all dimensions considered here from 3p log(p)s0 to pq. Between
minimal and maximal sample size there are in each case 10 points. To illustrate
the rate derived in Section 4 we compute the error kB̂H − B 0 k2F for different
dimensions of the problem under increasing number of observations.
To compute the solution of the optimization problem 1.7 functions from the
Matlab library cvx (CVX Research, 2012) were used.
Throughout this section the error is assumed to have the following shape
1
Error =
B̂H − B 0
pq
2
F
p
q
2
1 XX
0
=
B̂Hij − Bij
.
pq i=1 j=1
(5.1)
To verify the robustness of the estimator 1.7 and the rate of convergence that
was derived in Section 4 we use the Student t distribution with 3 degrees of
freedom. Every point in the plots corresponds to an average of 25 simulations.
The value of the tuning parameter is set to
s
log(p + q)
.
λ=2
nq
A comparison with λε from Corollary 3.1 indicates that λ is rather small. For
the settings we consider in this section we found that this value for λ is more
appropriate. As done in Candès and Plan (2010), for a better comparison between the error curves of our estimator and the oracle rate in Equation (4.4)
this rate was multiplied with 1.68 in the case of Student t distributed errors and
with 1.1 in the case of Gaussian errors.
Elsener and van de Geer/Robust Low-Rank Matrix Estimation
17
5.1. t-distributed and gaussian errors
The variance of the Student t distribution with ν > 2 degrees of freedom is
given by
ν
Var (εi ) =
, for εi ∼ tν .
(5.2)
ν−2
Figure 1a shows a comparison between the Huber estimator 1.7 with the estimator that uses the quadratic loss in the case of Student t with 3 degrees of
freedom distributed errors. As expected, the estimator that uses the quadratic
loss is not robust against the corrupted entries. On the other hand, we can see in
Figure 1b that the Huber estimator performs almost as well as the quadratic loss
estimator in the case of Gaussian errors with variance 1. In agreement with the
theory, the rate of the estimator is very close to the oracle rate for sufficiently
large sample sizes. The value of κ that we used in the simulations is 1.345. The
maximal rating η is chosen to be η = 10.
(a)
(b)
Fig 1: Comparison of the Huber and quadratic loss in the case of Student t
distributed (left figure) and Gaussian errors (right figure). We choose p = q = 80
and s0 = 2. The dot-dashed red line corresponds to 1.68 multiplied with the
oracle bound derived in Equation 4.4 for the exact low-rank case.
5.2. Changing the problem size
In order to confirm/verify the theoretical results, we proceed similarly to what
was done in Negahban and Wainwright (2011) and Negahban and Wainwright
(2012) in the corresponding cases. Here, we consider three different problem
sizes: p, q ∈ {30, 50, 80}. In Figure 2a we observe that as the problem gets
harder, i.e. as the dimension of the matrix increases, also the sample size needs
to be larger. Figure 2b shows that by rescaling the sample size by n/(3ps0 log(p))
Elsener and van de Geer/Robust Low-Rank Matrix Estimation
18
the rate of convergence agrees very well with the theoretical one. It is assumed
that the rank of the matrices is s0 = 2 for all cases. Every point corresponds to
an average of 25 simulations. The maximal rating η and the tuning parameter
κ are chosen as before.
(a)
(b)
Fig 2: Three different problem sizes are considered: p = q ∈ {30, 50, 80}. The
rank is fixed to s0 = 2 in both cases. In Panel (b) it can be seen that the
rate of convergence corresponds approximately to the theoretical one derived
in Equation 4.4. The dot-dashed line is the oracle. It was multiplied by 1.68 in
order to fit our curves.
5.3. Comparison with a low-rank + sparse estimator
In this subsection, we compare the performance of the Huber estimator 1.7 with
the performance of the low-rank matrix estimator proposed by Klopp, Lounici
and Tsybakov (2016) 1.10. We first compare the estimators B̂H and L̂ with the
observations Yi generated according to the model 1.4 with standard Gaussian
and Student t with 3 degrees of freedom distributed errors. Equation (21) in
Klopp, Lounici and Tsybakov (2016) suggests that the tuning parameters are
chosen as follows:
s
log(p + q)
log(p + q)
λ1 = 2
, λ2 = 2
,
nq
n
where λ1 and λ2 are the tuning parameters of the estimator 1.10. Also in this case
it has to be noticed that the tuning parameters are smaller than the theoretical
values given in their paper.
In Figure 3a the Huber estimator 1.7 is compared with the low-rank plus
sparse estimator 1.10 under the model 1.4 with i.i.d. Student t noise with 3 degrees of freedom. As expected, these estimators perform comparably well under
Elsener and van de Geer/Robust Low-Rank Matrix Estimation
(a)
19
(b)
Fig 3: The left panel shows the Huber estimator 1.7 and the low-rank plus
sparse estimator 1.10 under the model 1.4 with Student t noise with 3 degrees
of freedom. The right panel shows the same estimators under the same model
with standard Gaussian noise.
the trace regression model 1.4. In Figure 3b the same estimators are compared
under the model 1.4 with i.i.d. standard Gaussian noise. Also in this case, we see
that both estimators achieve approximately the same error. These observations
are not surprising since the theoretical analysis of Section 3 could be carried
over by adapting the (semi-)norms to the different penalization.
We now consider the model proposed in Klopp, Lounici and Tsybakov (2016)
where around 5% of the observed entries are taken to be only one rating. This
is the case of malicious users who systematically rate only one particular movie
with the same rating. We refer to Section 2.3 of Klopp, Lounici and Tsybakov
(2016) for more details on this setting. In Figure 4a we see that the Huber
estimator outperforms the low-rank plus sparse estimator with Student t noise
with 3 degrees of freedom. This might be due to the quadratic loss function and
to the choice of the tuning parameters. In Figure 4b where Gaussian noise is
considered we observe that both estimators perform almost equally well.
6. Discussion
In this paper, we have derived sharp and non-sharp oracle inequalities for two
robust nuclear norm penalized estimators of the noisy matrix completion problem. The robust estimators were defined using the well-known Huber loss for
which the sharp oracle inequality has been derived and the absolute value loss
for which we have shown a non-sharp oracle inequality. For both types of oracle
inequalities we proved a general deterministic result first and added then the
part arising from the empirical process. We have also shown how to apply the
Elsener and van de Geer/Robust Low-Rank Matrix Estimation
(a)
20
(b)
Fig 4: The left panel shows the Huber estimator 1.7 and the low-rank plus
sparse estimator 1.10 under the model 1.9 with Student t noise with 3 degrees
of freedom. The right panel shows the same estimators under the same model
as on the left panel but with standard Gaussian noise.
oracle inequalities to the case where we only assume weak sparsity, i.e. approximately low-rank matrices. It is worth pointing out that our estimators do not
require the distribution on the set of matrices (1.5) to be known in contrast to
e.g. Koltchinskii, Lounici and Tsybakov (2011). In our case, the distribution on
the set of matrices (1.5) is only needed in the theoretical analysis. The proofs
of the oracle inequalities rely on the properties of the nuclear norm, and for the
empirical process part on the Concentration, Symmetrization, and Contraction
Theorems. A main tool in this context was also the bound on the largest singular value of a matrix with finite Orlicz norm. Our simulations, in the case
of the Huber loss, showed a very good agreement with the convergence rates
obtained by our theoretical analysis. We saw that the oracle rate is attained up
to constants in presence of non-Gaussian noise and that the robust estimation
procedure outperforms the quadratic loss function.
It is left to future research to establish a sharp oracle inequality also for
the case of a non-differentiable robust loss function. The Contraction inequality
used in this paper for the Huber loss requires that also the derivative of the loss
is Lipschitz continuous. This is not the case for the absolute value loss. Thanks
to the convexity of the loss function it might be possible to derive a sharp result
also for this case.
Elsener and van de Geer/Robust Low-Rank Matrix Estimation
21
Appendix A: Proofs of main results
Proof of Inequality 2.8. Using the triangle property at B + with B 0 = B + we
obtain
0 = kB + knuclear − kB + knuclear ≤ Ω+ (B + − B + ) −Ω− (B + )
{z
}
|
=0
⇒ Ω− (B + ) = 0.
By the triangle property at B + with B 0 = B = B + + B − we have that
kB + knuclear − kBknuclear
= kB + knuclear − kB + + B − knuclear
≤ Ω+ (B − ) − Ω− (B + + B − )
= −Ω− (B + + B − ).
By the triangle inequality it follows using Ω− (B + ) = 0 that
Ω− (B + + B − ) ≥ Ω− (B − ) − Ω− (B + ) = Ω− (B − ).
Therefore, we have
kB + knuclear − kB + + B − knuclear ≤ −Ω− (B − ),
and by the triangle inequality
kB + knuclear − kB + + B − knuclear ≥ −kB − knuclear ,
which gives
Ω− (B − ) ≤ kB − knuclear .
For an arbitrary B we have again by the triangle inequality
kBknuclear − kB 0 knuclear ≤ kB + knuclear + kB − knuclear − kB 0 knuclear .
Applying the triangle property at B + we find that
kBknuclear − kB 0 knuclear ≤ Ω+ (B + − B 0 ) − Ω− (B 0 ) + kB − knuclear .
Apply now twice the triangle inequality (first inequality) to find that
kBknuclear − kB 0 knuclear ≤ Ω+ (B − B 0 ) + Ω+ (B − ) − Ω− (B − B 0 )
+ Ω− (B) + kB − knuclear
≤ Ω+ (B − B 0 ) − Ω− (B − B 0 )
+ 2kB − knuclear ,
where it was used that Ω(B − ) = 0 and that Ω− (B) ≤ Ω− (B − ) ≤ kB − knuclear .
Elsener and van de Geer/Robust Low-Rank Matrix Estimation
22
Proof of Lemma 3.1. Let B ∈ B. Define for 0 < t < 1
B̃t := (1 − t)B̂ + tB.
Since B is convex we have that B̃t ∈ B for all 0 < t < 1. Since B̂ is the minimizer
of the objective function and by the convexity of the objective function we have
Rn (B̂) + λkB̂knuclear ≤ Rn (B̃t ) + λkB̃t knuclear
≤ Rn (B̃t ) + (1 − t)λkB̂knuclear + tλkBknuclear .
Finally, we can conclude that
Rn (B̂) − Rn (B̃t )
≤ λkBknuclear − λkB̂knuclear .
t
Letting t → 0 the claim follows.
Proof of Theorem 3.1. The first order Taylor expansion of R at B̂ is given by
R(B) = R(B̂) + trace Ṙ(B̂)T (B − B̂) + Rem(B̂, B).
(A.1)
Then it follows that
R(B̂) − R(B) + Rem(B̂, B) = − trace Ṙ(B̂)T (B − B̂) .
(A.2)
Case 1
If
trace Ṙ(B̂)T (B − B̂)
≥ δλΩ+ (B̂ − B) + δλΩ− (B̂ − B) − 2λkB − knuclear − λ∗ ,
(A.3)
then by the two-point-margin condition 3 we find that
R(B) − R(B̂) ≥ trace Ṙ(B̂)T (B − B̂) + G(kB − B̂kF ).
(A.4)
Which implies that
R(B) − R(B̂) ≥ δλΩ+ (B̂ − B) + δλΩ− (B̂ − B) − 2λkB − knuclear
− λ∗ + G(kB − B̂kF )
|
{z
}
≥0
≥ δλΩ (B̂ − B) + δλΩ− (B̂ − B) − 2λkB − knuclear − λ∗ .
+
Case 2
Assume in the following that
trace Ṙ(B̂)T (B − B̂)
≤ δλΩ+ (B̂ − B) + δλΩ− (B̂ − B) − 2λkB − knuclear − λ∗ ,
(A.5)
Elsener and van de Geer/Robust Low-Rank Matrix Estimation
23
By the two-point inequality (Lemma 3.1) we have that
− trace Ṙn (B̂)T (B − B̂) ≤ λkBknuclear − λkB̂knuclear ,
(A.6)
which implies that
0 ≤ trace Ṙn (B̂)T (B − B̂) + λkBknuclear − λkB̂knuclear .
(A.7)
Hence,
− trace Ṙ(B̂)T (B − B̂) + δλΩ+ (B̂ − B) + δλΩ− (B̂ − B)
≤ trace (Ṙn (B̂) − Ṙ(B̂))T (B − B̂) + δλΩ+ (B̂ − B) + δλΩ− (B̂ − B)
+ λkBknuclear − λkB̂knuclear
≤ λε Ω(B̂ − B) + λ∗ + δλΩ+ (B̂ − B) + δλΩ− (B̂ − B)
+ λkBknuclear − λkB̂knuclear
≤ λε Ω+ (B̂ − B) + λε Ω− (B̂ − B) + λ∗ + δλΩ+ (B̂ − B)
+ δλΩ− (B̂ − B) + λΩ+ (B̂ − B) − λΩ− (B̂ − B) + 2λkB − knuclear
= λΩ+ (B̂ − B) − (1 − δ)λΩ− (B̂ − B) + 2λkB − knuclear + λ∗ .
Therefore, by Equation A.5
Ω− (B̂ − B) ≤
λ
Ω+ (B̂ − B).
(1 − δ)λ
We then have by the convex conjugate inequality
√
Ω+ (B̂ − B) ≤ kB̂ − BkF 3 s
√
≤ H 3 s + G(kB̂ − BkF ).
Which implies that
− trace Ṙ(B̂)T (B − B̂) + λΩ− (B̂ − B) + δλΩ+ (B̂ − B)
= R(B̂) − R(B) + Rem(B̂, B) + λΩ− (B̂ − B) + δλΩ+ (B̂ − B)
√
≤ H λ3 s + G(kB̂ − BkF ) + 2λkB − knuclear + λ∗
√
≤ H λ3 s + Rem(B̂, B) + 2λkB − knuclear + λ∗ .
Proof of Theorem 3.2. We start the proof with the following inequality using
the fact that B̂ is the minimizer of the objective function.
Rn (B̂) + λkB̂knuclear ≤ Rn (B) + λ kBknuclear
(A.8)
Elsener and van de Geer/Robust Low-Rank Matrix Estimation
24
Then, by adding and subtracting R(B̂) on the left hand side and R(B) on the
right hand side we obtain
h
i
R(B̂) − R(B) ≤ − (Rn (B̂) − R(B̂) − (Rn (B) − R(B))
+ λkBknuclear − λkB̂knuclear .
Applying Assumption 3.1, the definition of Ω, and Lemma 2.8 we obtain
R(B̂) − R(B) ≤ λε Ω(B̂ − B) + λ∗ + λkBknuclear − λkB̂knuclear
≤ λε Ω+ (B̂ − B) + λε Ω− (B̂ − B) + λ∗
+ λΩ+ (B̂ − B) − λΩ− (B̂ − B) + 2λkB − knuclear
= (λε + λ)Ω+ (B̂ − B) − (λ − λε )Ω− (B̂ − B) + λ∗ + 2λkB − knuclear .
Since later on we apply Assumption 2 we subtract on both sides of the above
inequality R(B 0 ).
R(B̂) − R(B 0 ) + λΩ− (B̂ − B)
≤ R(B) − R(B 0 ) + λΩ+ (B̂ − B) + λ∗ + 2λkB − knuclear .
(A.9)
It is then useful to make the following case distinction that allows us to obtain
an upper bound for the estimation error.
Case 1
λ∗ + R(B) − R(B 0 ) + 2λkB − knuclear , then
If λΩ+ (B̂ − B) ≤ (1−δ)
δ
δλΩ+ (B̂ − B) ≤ (1 − δ) λ∗ + R(B) − R(B 0 ) + 2λkB − knuclear
≤ λ∗ + R(B) − R(B 0 ) + 2λkB − knuclear .
By multiplying Equaiton A.9 on both sides with δ we arrive at
δλΩ− (B̂ − B) ≤ λ∗ + R(B) − R(B 0 ) + 2λkB − knuclear .
Therefore,
δ(λΩ+ (B̂ − B) + λΩ− (B̂ − B))
≤ 2λ∗ + 2(R(B) − R(B 0 )) + 4λkB − knuclear .
(A.10)
And since
λ < λ,
we conclude that
δλ(Ω+ + Ω− )(B̂ − B) ≤ 2λ∗ + 2(R(B) − R(B 0 )) + 4λkB − knuclear .
Case 2
If λΩ+ (B̂ − B) ≥
(1−δ)
δ
λ∗ + R(B) − R(B 0 ) + 2λkB − knuclear , then
R(B̂) − R(B 0 ) + λΩ− (B̂ − B) ≤ λΩ+ (B̂ − B) + λΩ+ (B̂ − B)
δ
.
(1 − δ)
Elsener and van de Geer/Robust Low-Rank Matrix Estimation
25
This implies
h
i
(1 − δ) R(B̂) − R(B 0 ) + (1 − δ)λΩ− (B̂ − B) ≤ λΩ+ (B̂ − B).
And finally we conclude that
Ω− (B̂ − B) ≤
λ
Ω+ (B̂ − B).
(1 − δ)λ
We then obtain using the definition of Ω+ in Lemma 2.3
√
Ω+ (B̂ − B) ≤ kB̂ − BkF 3 s
√
≤ (kB̂ − B 0 kF + kB − B 0 kF )3 s
√
≤ G(kB̂ − B 0 kF ) + G(kB − B 0 kF ) + 2H(3 s).
Invoking the convex conjugate inequality and Assumption 2 we get
δλΩ+ (B̂ − B) + δλΩ− (B̂ − B)
√
≤ 2H(λ(1 + δ)3 s) + R(B) − R(B 0 ) + (R(B) − R(B 0 ))
+ λ∗ + 2λkB − knuclear
√
≤ 2H(λ(1 + δ)3 s) + 2(R(B) − R(B 0 ))
+ λ∗ + 2λkB − knuclear .
Combining the two cases we have for the estimation error
δλ(Ω+ + Ω− )(B̂ − B)
√
≤ 2H(λ(1 + δ)3 s) + 2λ∗
+ 2(R(B) − R(B 0 )) + 4λkB − knuclear .
and for the second claim we conclude that
R(B̂) − R(B) ≤ λΩ+ (B̂ − B) + λ∗ + 2λkB − knuclear
√
1
≤
2H(λ(1 + δ)3 s) + λ∗ + 2(R(B) − R(B 0 ))
δ
+ 2λkB − knuclear + λ∗ + 2λkB − knuclear .
Proof of Lemma 2.1. The theoretical risk function arising from the Huber loss
is given by
n
R(B) =
1X
EXi [E [ρH (Yi − trace(Xi B)) |Xi ]] .
n i=1
(A.11)
Elsener and van de Geer/Robust Low-Rank Matrix Estimation
26
Suppose that Xi has its only 1 at entry (k, j). Then XB = (B)jk . Define
r(x, B) := E [ρH (Yi − trace (Xi B)) |Xi = x]
= E [ρH (Yi − Bjk )]
h
i
dρH (Yi −Bjk )
=
E
. The derivative with respect
We notice that ṙ(x, B) = dr(x,B)
dBjk
dBjk
to Bjk of ρH (Yi − Bjk ) is given by
−2(Yi − Bjk ), if |Yi − Bjk | ≤ κ
−2κ,
if Yi − Bjk > κ
ρ̇H (Yi − Bjk ) :=
2κ,
if Yi − Bjk < −κ.
Then,
Z
Bjk +κ
ṙ(x, B) = −2
Z
∞
Z
(y − Bjk )dF (y) − 2κ
Bjk −κ
Z
Bjk +κ
Bjk +κ
= −2
Z
dF (y)
−∞
Bjk +κ
dF (y) − 2κ [1 − F (κ + Bjk )]
ydF (y) + 2Bjk
Bjk −κ
Bjk −κ
dF (y) + 2κ
Bjk −κ
+ 2κF (Bjk − κ)
= −2(Bjk + κ)F (Bjk + κ) + 2(Bjk − κ)F (Bjk − κ)
Z Bjk +κ
+2
F (y)dy + 2Bjk [F (Bjk + κ) − F (Bjk − κ)] − 2κ
Bjk −κ
+ 2κF (κ + Bjk ) + 2κF (Bjk − κ).
Z Bjk +κ
=2
F (y)dy − 2κ.
Bjk −κ
The second derivative of r(x, B) with respect to Bjk is then given by
r̈(x, B) = 2[F (Bjk + κ) − F (Bjk − κ)].
Therefore, the Taylor expansion around B 0 is given by
0
r(x, B) = r(x, B 0 ) + ṙ(x, B 0 )(Bjk − Bjk
)+
r̈(x, B̃)
0 2
(Bjk − Bjk
) ,
2
where B̃ ∈ B is an intermediate point.
We can see that Assumption 3 holds with G(u) = u2 /(2C12 pq).
Proof of Lemma 2.2. For the (theoretical) risk function R arising from the absolute value loss we have
R(B) = E [Rn (B)]
n
1X
=
E [|Yi − trace (Xi B)|] .
n i=1
(A.12)
(A.13)
(A.14)
Elsener and van de Geer/Robust Low-Rank Matrix Estimation
27
Using the tower property of the conditional expectation we obtain
n
R(B) =
1X
EXi [E [|Yi − trace(Xi B)| |Xi ]] .
n i=1
(A.15)
Suppose that Xi has its only 1 at entry (k, j). Then XB = (B)jk . Define
r(x, B) := E [|Yi − trace (Xi B)| |Xi = x]
= E [|Yi − Bjk |]
Z
Z
=
(y − Bjk ) dF (y) +
y≥Bjk
(Bjk − y) dF (y)
y<Bjk
Z ∞
Z
(Bjk − y) dF (y)
(y − Bjk ) dF (y) +
=
−∞
y≥Bjk
Z ∞
−
(Bjk − y) dF (y)
Bjk
Z ∞
Z
∞
(y − Bjk ) dF (y) +
=2
(Bjk − y) dF (y)
−∞
Bjk
Z ∞
Z
∞
dF (y) −
Bjk
−∞
{z
}
|
=1
Z ∞
Z ∞
=2
(1 − F (y)) dy + Bjk −
ydF (y).
Z
∞
(y − Bjk ) dF (y) + Bjk
=2
ydF (y)
−∞
−∞
Bjk
The Taylor expansion of r(x, B) around B 0 , assuming that B 0 minimizes r,
is given by
r̈(x, B̃)
0 2
0
r(x, B) = r(x, B 0 ) + ṙ(x, B 0 ) Bjk − Bjk
+
Bjk − Bjk
2
0 2
= r(x, B 0 ) + f (B̃jk ) Bjk − Bjk
,
where B̃ ∈ B is an intermediate point.
r(x, B) − r(x, B 0 ) ≥
1
0 2
Bjk − Bjk
2
C2
which means that the one point margin Condition 2 is satisfied with G(u) =
u2 /(2C22 pq).
Appendix B: Supplemental Material
This supplemental material contains an application to real data sets, the proofs
of the lemmas in Section 2 of the main text, and a section on the bound of the
empirical process part of the estimation problem.
Elsener and van de Geer/Robust Low-Rank Matrix Estimation
28
B.1. Example with real data
In Section 5 we have shown several synthetic data examples. The convex optimization problems there were solved using the semidefinite programming (SDP)
toolbox CVX Research (2012). These algorithms work very well for comparably
low-dimensional optimization problems. When real datasets are considered, due
to the much larger problem sizes different algorithms are needed. To solve the
optimization problem with real data a proximal gradient algorithm is used. The
algorithm is given in pseudocode.
We define F1 to be the empirical risk for the Huber loss
n
F1 (B) =
1X
ρH (Yi − trace(Xi B)).
n i=1
We define F2 to be the empirical risk for the quadratic loss
n
F2 (B) =
1X
(Yi − trace(Xi B))2 .
n i=1
The gradient of F1 is given by
n
OF1 (B) = −
1X
ρ̇H (Yi − trace(Xi B))XiT ,
n i=1
where ρ̇H (Yi −trace(Xi B)) is given in the proof of Lemma 2.1 in the main paper.
The gradient of F2 is given by
n
OF2 (B) = −
2X
(Yi − trace(Xi B))XiT .
n i=1
The proximity operator proxnuc for W ∈ Rp×q and γ > 0 is defined as
1
proxnucγ (W ) = arg min γkBknuclear + kB − W k2F .
2
p×q
B∈R
For the nuclear norm the proximity operator has a closed form: let W =
U diag(σ1 , . . . , σmin(p,q) )V 0 be the singular value decomposition of W , then
proxnucγ (W ) = U Sγ (σ1 , . . . , σmin(p,q) ) V 0 ,
where for x ∈ R the operator Sγ (x) applied elementwise is given by
x − γ, if x > γ
0,
if x = γ
Sγ (x) =
x + γ, if x < γ.
It is known that the solution of the optimization problem B̂H satisfies the
following fixed point equation
Elsener and van de Geer/Robust Low-Rank Matrix Estimation
29
B̂H = proxnucγ B̂H − γOF .
The same holds also for the quadratic loss function. To compute the proximal operator
in
Algorithm
1
the
function
prox nuclearnorm from the Matlab toolbox Unlocbox (2016) was used. The
algorithm is a Nesterov-type Accelerated Proximal Gradient algorithm. We refer to Section 4.3 in Parikh and Boyd (2014) and references therein. In particular,
the present form of Algorithm 1 goes back to Beck and Teboulle (2009). The
Huber constant is chosen to be κ = 2 in all the examples. Algorithm 1 is applied to the Huber loss (i = 1) and to the quadratic loss (i = 2). The tuning
parameter for the quadratic loss case is smaller than for the Huber loss.
Algorithm 1 HuberQuadProx
Start with an initial value B = Bobs containing the entries belonging
p to the training set.
Define the dummy variable v = B. Choose L = 0.1, β = 1.2, λ1 = 20 log(p + q)/(nq), and
p
λ2 = log(p + q)/(nq).
while Stopping criterion is not satisfied OR t ≤ maxiter do
obj ← Fi (B) + λi kBknuclear
Bprev ← B
δ=1
while δ > 0.001 do
B ← v − 1/LOFi (v)
B ← proxnucλ/L (B)
T
2
← FβL
δL ←
i (B) + λkBknuclear − obj − trace(OFi (Bprev ) (B − Bprev )) − L/2kB − Bprev kF
end while
L ← L/β
v ← B + t/(t + 3)(B − Bprev )
end while
This method was applied to the MovieLens data set which consists of p = 943
users, q = 1682 movies, and 1000 000 observed ratings, we call the set containing
the indices of the observed ratings Γ. The minimal and maximal ratings are
1 and 5 respectively. Every user has rated at least 20 movies. The data set is
available at MovieLens (1998). Training and testing sets of different sizes are
drawn randomly without replacement from the set of observed ratings. The
testing error is computed as follows: with 1000 000 = ntest + ntrain and using the
disjoint union Γ = Γtest ∪ Γtrain (where |Γtest | = ntest and |Γtrain | = ntrain ) we
have
X
1
0
test error =
(Bj,k
− B̂j,k )2 .
ntest
(j,k)∈Γtest
Elsener and van de Geer/Robust Low-Rank Matrix Estimation
30
The results are reported in Table 1:
training set size
250 000
500 000
750 000
test set size
750 000
500 000
250 000
Huber loss
test error
Iterations
1.43
60 000
1.14
60 000
1.01
60 000
Table 1
Quadratic loss
test error
Iterations
1.32
60 000
1.06
60 000
0.96
60 000
It can be observed that the test error of the quadratic loss estimator is slightly
smaller than the test error for the Huber loss estimator. This might be due to
the fact that there are no heavily corrupted entries in the data. On the other
hand, this indicates that the Huber estimator is able to “adapt” also to the
usual setting without heavy corruptions.
We have applied Algorithm 1 also to the MovieLens data set with 10 0000 209
observed ratings from 60 040 users on 30 952 movies. The minimal and maximal
ratings are 1 and 5 respectively. Every user has rated at least 20 movies. The
data set is available at MovieLens (2003).
The results are reported in Table 2:
training set size
2500 000
5000 000
7500 000
test set size
7500 209
5000 209
2500 209
Huber loss
test error
Iterations
1.05
100 000
0.92
100 000
0.85
100 000
Table 2
Quadratic loss
test error
Iterations
1.04
100 000
0.90
100 000
0.84
100 000
B.2. Proofs of Lemmas in Section 2
The following lemma shows “equivalence” of the nuclear norm and Frobenius
norm. This fact is useful since it is more common and meaningful to measure
the estimation error in the Frobenius norm. The rest of the lemma contains
technicalities that are used in the sequel to derive the triangle property.
Lemma B.1. Let A ∈ Rp×q . Then
kAkF ≤ kAknuclear ≤
√
rank A kAkF .
Let P ∈ Rp×s with P T P = I and s ≤ p. Then
√
P P T A F ≤ sΛmax (A)
and
PPTA
F
≤ kAkF ,
kAQQT kF ≤ kAkF ,
kAP P T kF ≤ kAkF ,
where kAkF is the Frobenius norm of A and Λmax (A) its largest singular value.
Elsener and van de Geer/Robust Low-Rank Matrix Estimation
31
Proof of Lemma B.1. Consider the singular value decomposition (SVD) of A
with s := rank(A)
A = PA ΛA QTA
(B.1)
with PAT PA = 1s×s , QTA QA = 1s×s and ΛA = diag (ΛA,1 , . . . , ΛA,s ). Then the
nuclear norm of A can be written as
kAknuclear =
s
X
ΛA,k .
(B.2)
k=1
The first claim follows from Hölder’s inequality applied in two different manners for the lower and upper bounds.
For the second claim consider the p−dimensional j−th unit vector ej . Define
u := P P T ej .
Then
T
T
u AA u =
2
2
AT u
2
kuk2
2
kuk2
≤
max
kuk2 =1
2
T
A u
= Λ2max (A) P P T ej
2
kuk2
2
2
2
.
Then, by the invariance of the trace under cyclic permutations we obtain the
claimed result.
In order to show that the triangle property holds we also need the dual norm
of the nuclear norm so that to apply the dual norm inequality (see Lemma
B.2). The subdifferential of the nuclear norm is then used to deduce the triangle
property.
The following lemma gives the dual norm and the subdifferential of the nuclear norm. It cites results of Lange (2013) and Watson (1992).
Lemma B.2. Let A
kAknuclear is given by
∈
Rp×q .The dual norm of the nuclear norm
Ω∗ (A) = Λmax (A),
where Λmax (A) is the largest singular value of A . Moreover, the subgradient of
the nuclear norm is given by
∂ kBknuclear = Z = P QT + 1p×p − P P T W 1q×q − QQT :
Λmax (W ) = 1} .
Proof. The derivation of the dual norm of the nuclear norm can be found in
Example 14.3.6. from Lange (2013). A justification for the second claim can be
found in Watson (1992).
Proof of Lemma 2.1 in the main text. Let Z ∈ ∂ kB + knuclear , i.e.
T
T
T
Z = P + Q+ + 1p×p − P + P + W 1q×q − Q+ Q+ ,
Elsener and van de Geer/Robust Low-Rank Matrix Estimation
32
where W ∈ Rp×q is such that Λmax (W ) = 1. Therefore, it is possible to write
Z = Z1 + Z2 ,
T
where Z1 := P + Q+
T
T
and Z2 := 1p×p − P + P + W 1q×q − Q+ Q+ .
Recall the definition of the subdifferential of the nuclear norm
∂ B + nuclear = Z| kB 0 knuclear − B + nuclear ≥ trace Z T B 0 − B + ,
∀B 0 ∈ Rp×q
For Z ∈ ∂ kB + knuclear we have
B + nuclear − kB 0 knuclear ≤ trace Z T B + − B 0
= trace (Z1 + Z2 )
T
B+ − B0
.
To prove the first assertion we need to bound the right
hand side of the above
inequality. For simplicity consider first trace Z1T B 0 . Using the invariance of
the trace under cyclic permutations we have
T
T
trace Z1T B 0 = trace Q+ P + B 0 = trace P + B 0 Q+
T
T
T
= trace |P +{zP +} P + B 0 Q+ Q+ Q+
| {z }
=1
=1
+ +T + +T 0 + +T
= trace Q P P P B Q Q
T
T
T
≤ Ω∗ Q+ P + kP + P + B 0 Q+ Q+ knuclear
T
T
= P + P + B 0 Q+ Q+
nuclear
,
T
since Λmax P + Q+ = 1.
On the other hand, consider trace Z2T B 0 . Again by the invariance of the
trace under cyclic permutations we have
T
T
trace Z2T B 0 = trace 1q×q − Q+ Q+ W T 1p×p − P + P + B 0
T
T
= trace W T 1p×p − P + P + B 0 1q×q − Q+ Q+
T
T
≤ Λmax (W ) 1p×p − P + P + B 0 1q×q − Q+ Q+
| {z }
nuclear
=1
T
T
=
sup
trace W T 1p×p − P + P + B 0 1q×q − Q+ Q+
.
W :Λmax (W )=1
Hence, it is possible to find a W such that Λmax (W ) = 1 and such that
T
T
trace W T B 0 = 1p×p − P + P + B 0 1q×q − Q+ Q+
. (B.3)
nuclear
Elsener and van de Geer/Robust Low-Rank Matrix Estimation
33
Substituting B 0 with B 0 − B + we have that
max
trace Z T B 0 − B +
+
Z∈∂kB k
nuclear
n
T
T
trace 1q×q − Q+ Q+ W T 1p×p − P + P +
B0 − B+
=
max
Λmax (W )=1
o
T
+ trace Q+ P + B 0 − B +
Hence,
T
T
T
T
− 1 − P + P + B 0 1 − Q+ Q+
P + P + B 0 − B + Q+ Q+
nuclear
nuclear
+ +T
√
−
+ +T
0
+
0
≤ s P P
− ΩB + (B )
B −B Q Q
F
0
+
≤ Ω+
.
B+ B − B
For the second assertion we have
T
T
T
T
kB 0 knuclear = P + P + B 0 + B 0 Q+ Q+ − P + P + B 0 Q+ Q+
T
T
+ 1 − P + P + B 0 1 − Q+ Q+
nuclear
T
≤ P +P + B0
+
+T
T
nuclear
0
+
+ B 0 Q+ Q+
nuclear
+T
+ P P BQ Q
nuclear T
+ +T
+ 1−P P
B 0 1 − Q+ Q+
nuclear
√ + +T 0
T
T
0 + +T
≤ s P P B
+ BQ Q
+ P + P + B 0 Q+ Q+
F
F
+ +T
0
+ +T
+ 1−P P
B 1−Q Q
.
nuclear
F
Since the dual of Ω may not be easy to deal with we bound it by the dual norm
of the nuclear norm. By doing so, it will be possible to apply results about the
tail of the maximum singular value of a sum of independent random matrices.
Lemma B.3. For the dual norm of Ω we have that
Ω∗ ≤ Λmax .
Proof. By Lemma 2.1 in the main text it is known that
−
k·knuclear ≤ Ω+
B + + ΩB + = Ω.
Therefore, using Lemma 2.2 in the main text we have for the dual norms that
Ω∗ ≤ k·knuclear∗ = Λmax .
Elsener and van de Geer/Robust Low-Rank Matrix Estimation
34
B.3. Probability bounds for the empirical process
Pn
To bound the expectation of Λmax n1 i=1 ε̃i Xi where ε̃1 , . . . , ε̃n are i.i.d.
Rademacher random variables independent of (Xi , Yi )ni=1 we use Theorem 2.1
in the main text together with the fact that the 2-Orlicz norm of a Rademacher
random variable ε̃ is equal to
r
1
kε̃kψ2 =
.
log 2
Now we check the assumptions of Theorem 2.1 in the main text.
For the case of matrix completion with uniform sampling we obtain that
EXi XiT =
1 T
ιι ,
pq
where ι is a q−vector consisting of only 1’s. It follows that
1
1 T
T
ιι
= .
Λmax EXi Xi = Λmax
pq
p
Analogously, we obtain that
1
Λmax EXiT Xi = .
q
Notice that by the independence of the Rademacher sequence and the random matrices Xi we have Eε̃i XiT = Eε̃i EXiT = 0, Eε̃2i Xi XiT = EXi XiT , and
Eε̃2i XiT Xi = EXiT Xi . Moreover, since Λmax (·) is a norm we have
Λmax (ε̃i Xi ) = |ε̃i | Λmax (Xi ) ≤ |ε̃i | , ∀i.
Hence,
r
1
, ∀i.
log 2
Lemma B.4.
1. Suppose that ρ is differentiable and Lipschitz continuous
with constant L. Suppose further that ρ̇(x)x is Lipschitz continuous with
Lipschitz constant L̃. Assume that for a constant KΩ
Λ2max (ε̃i Xi ) ψ
2
≤ kε̃kψ2 =
max Ω∗ (Xi ) ≤ KΩ .
1≤i≤n
Define for all M > 0
ZM :=
sup
trace
Ṙn (B 0 ) − Ṙ(B 0 )
T
(B − B 0 ) . (B.4)
B 0 ∈B:Ω(B−B 0 )≤M
Then we have for a constant C0 > 0 and λ̃1 = 8η L̃p log(p + q)/(3n) that
ZM
≤ L̃M
(8C0 +
√
s
2)
p
log(p + q)
log(p + q)
+ 8C0 log(1 + q)
nq
n
with probability at least 1 − exp(−p log(p + q)).
!
+ λ̃1
Elsener and van de Geer/Robust Low-Rank Matrix Estimation
35
2. Let ρ be a Lipschitz continuous function with Lipschitz constant L. Assume
that for a constant KΩ
max Ω∗ (Xi ) ≤ KΩ .
1≤i≤n
Define for all M > 0
ZM :=
|[Rn (B 0 ) − R(B 0 )] − [Rn (B) − R(B)]| .
sup
B 0 ∈B:Ω(B 0 −B)≤M
Then we have for a constant C0 > 0 and λ̃2 = 8ηLp log(p + q)/(3n) that
ZM
≤ LM
(8C0 +
√
s
2)
p
log(p + q)
log(p + q)
+ 8C0 log(1 + q)
nq
n
!
+ λ̃2
with probability at least 1 − exp(−p log(p + q)).
Proof. The proof of this lemma is based on the Symmetrization Theorem, on
the Contraction Theorem, on the dual norm inequality, and on Bousquet’s Concentration Theorem. We use the notation ρ(Xi B, Yi ) := ρ(Yi − trace(Xi B)) and
ρ̇(Xi B, Yi ) := ρ̇(Yi − trace(Xi B)).
We have
"
#
T
0
0
0
Ṙn (B ) − Ṙ(B ) (B − B )
trace
E
sup
B 0 ∈B:Ω(B−B 0 )≤M
"
n
1X
T
trace (B − B 0 ) XiT ρ̇(Xi B 0 , Yi )
B 0 ∈B:Ω(B−B 0 )≤M n i=1
#
T
0
− EXi ρ̇(Xi B , Yi )
=E
sup
"
n
1X
≤ 2E
sup
ε̃i trace (B − B 0 )T XiT ρ̇(Xi B 0 , Yi )
B 0 ∈B:Ω(B−B 0 )≤M n i=1
!
n
X
0
≤ 4L̃Ω(B − B )EΩ∗
ε̃i Xi /n,
(B.5)
#
i=1
≤ 4L̃M EΩ∗
n
X
!
ε̃i Xi
/n, since Ω ≤ Ω ⇒ Ω∗ ≤ Ω∗ .
(B.6)
i=1
The first inequality follows from Theorem B.1, the second
p from Theorem B.2
√
and the dual norm inequality. With S = 1/ q and K = 1/ log 2 in Theorem
2.1 in the main text together with the concavity of the logarithm
s r
s
s
q
1
1
1
q
1
log
=
log q + log
≤ log
+
log 2
2
2
log 2
2 2 log 2
p
≤ log (q + 1)
Elsener and van de Geer/Robust Low-Rank Matrix Estimation
36
we obtain integrating the tail that the expectation of the largest singular value
of the sum of masks is bounded by
s
!
!
n
X
log(p + q) p
log(p + q)
EΛmax
+ log (q + 1)
. (B.7)
ε̃i Xi /n ≤ C0
nq
n
i=1
Hence, defining
T
f (Xi B 0 ) := trace (B − B 0 ) XiT ρ̇(Xi B 0 , Yi )
(B.8)
we have using the Lipschitz continuity of the loss function and the fact that the
Xi ∈ χ are i.i.d.
sup
Var(f (X1 B 0 ))
B 0 ∈B:Ω(B−B 0 )≤M
=
sup
B 0 ∈B:Ω(B−B 0 )≤M
≤
sup
B 0 ∈B:Ω(B−B 0 )≤M
=
sup
B 0 ∈B:Ω(B−B 0 )≤M
q X
p
X
0
X1lj (Bjl − Bjl
Var
)ρ̇(X1 B 0 , Yi ))
l=1 j=1
2
q X
p
X
0
X1lj (Bjl − Bjl
)
L2 E
l=1 j=1
q X
p
X
0 2
L2 E
X12lj (Bjl − Bjl
)
l=1 j=1
2
=
L̃
0 2
kBjl − Bjl
kF
B 0 ∈B:Ω(B−B 0 )≤M pq
≤
L̃2 M 2
=: T 2 .
pq
sup
Therefore, using
kf k∞ ≤ 2η L̃ =: D
(B.9)
we obtain from Bousquet’s Concentration Theorem B.3 that for all t > 0
s
!
log(p + q) p
log(p + q)
P ZM ≥ 8L̃M C0
+ log (q + 1)
nq
n
!
r
M L̃ 2t 8tη L̃
+√
+
pq n
3n
≤ exp(−t).
Elsener and van de Geer/Robust Low-Rank Matrix Estimation
37
Replacing t by p log(p + q) we obtain
s
!
p
√
log(p + q)
log(p + q)
P ZM ≥ M L̃ (8C0 + 2)
+ 8C0 log(q + 1)
nq
n
!
8η L̃p log(p + q)
≤ exp(−p log(p + q)).
+
3n
For the second claim we proceed similarly. We have
"
E
0
#
0
|[Rn (B ) − R (B )] − [Rn (B) − R (B)]|
sup
B 0 ∈B:Ω(B 0 −B)≤M
"
=E
"
sup
B 0 ∈B:Ω(B 0 −B)≤M
"
n
#
n
1X
0
0
{ρ (Xi B , Yi ) − Eρ (Xi B , Yi )}
n i=1
##
1X
{ρ (Xi B, Yi ) − Eρ (Xi B, Yi )}
n i=1
"
#
" n
1X
=E
sup
{ρ (Xi B 0 , Yi ) − ρ (Xi B, Yi )}
n i=1
B 0 ∈B:Ω(B 0 −B)≤M
##
" n
1X
0
{Eρ (Xi B , Yi ) − Eρ (Xi B, Yi )}
−
n i=1
"
#
n
1X
≤ 2E
sup
ε̃i (ρ (Xi B 0 , Yi ) − ρ (Xi B, Yi ))
B 0 ∈B:Ω(B 0 −B)≤M n i=1
"
#
n
1X
0
≤ 4E
sup
Lε̃i trace (Xi (B − B))
B 0 ∈B:Ω(B 0 −B)≤M n i=1
( n
)
!#
"
1X
ε̃i Xi {B 0 − B}
trace
= 4LE
sup
n i=1
B 0 ∈B:Ω(B 0 −B)≤M
!
n
1X
≤ 4LEΩ∗
ε̃i Xi Ω (B 0 − B) , by the dual norm inequality
n i=1
!
n
1X
≤ 4LM EΛmax
ε̃i Xi
n i=1
s
!
log(p + q) p
log(p + q)
≤ 4C0 LM
+ log (q + 1)
, as before.
nq
n
−
The first inequality follows from Theorem B.1, the second from Theorem B.2.
Moreover, we have for all i = 1, . . . , n that
|ρ(Yi − trace(Xi B 0 )) − ρ(Yi − trace(Xi B))| ≤ L |trace (Xi (B 0 − B))|
≤ 2ηL =: D.
Elsener and van de Geer/Robust Low-Rank Matrix Estimation
38
In view of applying Bousquet’s Concentration Theorem B.3 we have with a
similar calculation as for the first claim
h
i
2
E (ρ(Yi − trace(Xi B 0 )) − ρ(Yi − trace(Xi B)))
≤ L2 E trace2 (Xi (B − B 0 ))
≤
M 2 L2
=: T 2
pq
Replacing t by p log(p + q) we obtain
s
!
log(p + q)
log(p + q) p
+ log (q + 1)
P ZM ≥ 8C0 LM
nq
n
!
r
M L 2p log(p + q)
+√
+ λ̃2
pq
n
≤ exp (−p log(p + q)) .
To obtain a uniform bound for all B ∈ B we use the peeling device given in
van de Geer (2000).
Lemma B.5.
λε,1
1. Let L̃ and λ̃1 be as in Lemma B.4. Define
s
!
p
√
log(p + q)
log(p + q)
+ 8C0 log(1 + q)
= L̃ (8C0 + 2)
nq
n
We have for any fixed B ∈ B
0
P ∃B ∈ B : trace
0
0
Ṙn (B ) − Ṙ(B )
T
0
(B − B)
!
0
> 2λε,1 (Ω(B − B) + 1) + λ̃1
≤ (j0 + 2) exp(−p log(p + q)).
2. Let L and λ̃2 be as in Lemma B.4. Define
s
!
p
√
log(p + q)
log(p + q)
+ 8C0 log(1 + q)
λε,2 = L (8C0 + 2)
nq
n
We have for any fixed B ∈ B
P(∃B 0 ∈ B : |(Rn (B 0 ) − R(B 0 )) − (Rn (B) − R(B))|
> 2λε,2 (Ω(B 0 − B) + 1) + λ̃2 )
≤ (j0 + 2) exp(−p log(p + q)).
Elsener and van de Geer/Robust Low-Rank Matrix Estimation
39
Proof. We subdivide the set B as follows for a fixed B ∈ B
B = {B 0 ∈ B : Ω(B − B 0 ) ≤ 1}
√
∪ {B 0 ∈ B : 1 < Ω(B − B 0 ) ≤ 14q pqη} .
Suppose we are on the first set, then
n
∃B 0 : Ω(B − B 0 ) ≤ 1, trace (Ṙn (B 0 ) − Ṙ(B 0 ))T (B − B 0 )
o
> 2λε,1 (Ω(B − B 0 ) + 1) + λ̃1
n
⊂ ∃B 0 : Ω(B − B 0 ) ≤ 1, trace (Ṙn (B 0 ) − Ṙ(B 0 ))T (B − B 0 )
o
> λε,1 + λ̃1 .
Therefore, by Lemma B.4 we conclude that
P ∃B 0 : Ω(B − B 0 ) ≤ 1, trace (Ṙn (B 0 ) − Ṙ(B 0 ))T (B − B 0 )
> 2λε,1 (Ω(B − B 0 ) + 1) + λ̃1
≤ exp(−p log(p + q)).
√
We then consider the set B 0 ∈ B : 1 < Ω(B − B 0 ) ≤ 14q pqη . We first refine
√
it by choosing j0 as the smallest integer such that j0 + 1 > log2 (14q pqη). This
leads us to
√
{B 0 ∈ B : 1 < Ω(B − B 0 ) ≤ 14q pqη}
⊂
j0
[
0 j
B : 2 < Ω(B − B 0 ) ≤ 2j+1 .
|
{z
}
j=0
=:Bj
For one j we have for the event
n
∃B 0 ∈ Bj : trace (Ṙn (B 0 ) − Ṙ(B 0 ))T (B 0 − B)
o
> 2λε,1 (Ω(B − B 0 ) + 1) + λ̃1
n
o
⊂ ∃B 0 ∈ Bj : trace (Ṙn (B 0 ) − Ṙ(B 0 ))T (B 0 − B) > 2j+1 λε,1 + λ̃1 .
By the first claim in Lemma B.4 we can conclude that
P ∃B 0 ∈ Bj : trace (Ṙn (B 0 ) − Ṙ(B 0 ))T (B 0 − B)
> 2λε,1 (Ω(B − B 0 ) + 1) + λ̃1
≤ exp(−p log(p + q)).
Elsener and van de Geer/Robust Low-Rank Matrix Estimation
40
To keep the notation clean we define the event
n
√
C = ∃B 0 ∈ B : 1 < Ω(B − B 0 ) ≤ 14q pqη,
o
trace (Ṙn (B 0 ) − Ṙ(B 0 ))T (B − B 0 ) > 2λε,1 (Ω(B − B 0 ) + 1) + λ̃1
and for all j = 0, . . . , j0 the events
n
Cj = ∃B 0 ∈ Bj : trace (Ṙn (B 0 ) − Ṙ(B 0 ))T (B 0 − B)
o
> 2j+1 λε,1 + λ̃1 .
Sj 0
Cj . Therefore, by the union bound
j0
j0
[
X
Cj ≤
P (Cj ) ≤ (j0 + 1) exp(−p log(p + q)).
P (C) ≤ P
Then C ⊂
j=0
j=0
j=0
The second claim follows by an analogous reasoning.
The proof of Lemma B.4 requires the Symmetrization Theorem and the Contraction Theorem (or Contraction Principle). The first theorem allows us to reduce the case of non-centred random variables to the case of mean zero random
variables. A main tool for this type of reduction is a sequence of Rademacher
random variables. The second theorem, the Contraction Principle, is used to
compare the limit behaviour of two series of Rademacher random variables with
different coefficients. We state these results for the sake of completeness.
Theorem B.1 (Symmetrization of Expectations, van der Vaart and Wellner
(1996)). Consider X1 , . . . , Xn independent matrices in χ and let F be a class of
real-valued functions on χ. Let ε̃1 , . . . , ε̃n be a Rademacher sequence independent
of X1 , . . . , Xn , then
"
#
"
#
n
n
X
X
E sup
(f (Xi ) − Ef (Xi )) ≤ 2E sup
ε̃i f (Xi ) .
(B.10)
f ∈F
i=1
f ∈F
i=1
Theorem B.2 (Contraction Theorem, Ledoux and Talagrand (1991)). Consider the non-random elements x1 , . . . , xn of χ. Let F be a class of real-valued
functions on χ. Consider the Lipschitz continuous functions ρi : R → R with
Lipschitz constant L, i.e.
|ρi (µ) − ρi (µ̃)| ≤ L |µ − µ̃| , for all µ, µ̃ ∈ R.
Let ε̃1 , . . . , ε̃n be a Rademacher sequence . Then for any function f ∗ : χ → R,
we have
"
#
"
#
n
n
X
X
E sup
ε̃i {ρi (f (xi )) − ρi (f ∗ (xi ))} ≤ 2E Lsup
ε̃i (f (xi ) − f ∗ (xi )) .
f ∈F i=1
f ∈F i=1
(B.11)
Elsener and van de Geer/Robust Low-Rank Matrix Estimation
41
Theorem B.3 (Bousquet’s Concentration Theorem, Bousquet (2002)). Suppose that for all i = 1, . . . , n and for all f ∈ F
n
E [f (Xi )] = 0,
1X
sup E f 2 (Xi ) ≤ T 2 .
n i=1 f ∈F
Assume further for a constant D > 0 and for all f ∈ F that
kf k∞ ≤ D.
Define
n
Z := sup
f ∈F
1X
(f (Xi ) − E [f (Xi )]) .
n i=1
Then we have for all t > 0
r
P Z ≥ 2EZ + T
2t 4tD
+
n
3n
!
≤ exp(−t).
B.4. On the distribution of the errors
In this section we discuss the consequences arising from the assumption that
the distribution of the errors is symmetric around 0. Let F be the distribution
function of the errors. For purposes of illustration, we discuss the location model.
The case of low-rank matrix estimation then follows easily. The location model
is as follows
X = µ∗ + ε,
(B.12)
where µ∗ is some fixed real number such that |µ∗ | ≤ η and ε is additive noise
with symmetric (around 0) distribution.
Lemma B.6. Assume that f is the density with respect to Lebesgue measure of
the errors. Suppose further that f (u) > 0 for all |u| ≤ 2η. Define
µ01 = arg min E [|X − µ|] .
(B.13)
µ∗ = µ01 .
(B.14)
|µ|≤η
Then,
Proof. We first notice that µ01 is the median of the distribution of X. Since the
distribution is continuous and the density positive everywhere the median is
unique.
Since the distribution of the errors is symmetric around 0 the distribution of X
is symmetric around µ∗ . This implies that µ∗ must be the median. But since
the median is unique we must have µ∗ = µ01 .
Elsener and van de Geer/Robust Low-Rank Matrix Estimation
42
Lemma B.7. Assume that the distribution function of the errors satisfies
F (u + κ) − F (u − κ) ≥ 1/C12 for all |u| ≤ 2η and κ ≤ 2η.
(B.15)
Define
µ02 = arg min E [ρH (X − µ)] ,
(B.16)
|µ|≤η
where ρH is the Huber loss as defined in the main text. Then,
µ∗ = µ02 .
(B.17)
Proof. We notice that the first derivative of the Huber loss evaluated in X − µ02
is given by
−2(X − µ02 ), if |X − µ02 | ≤ κ
0
−2κ,
if X − µ02 > κ
ρ̇H (X − µ2 ) =
(B.18)
2κ,
if X − µ02 < −κ.
By straightforward computation
E ρ̇H (X − µ02 ) = 2
Z
κ+µ02 −µ∗
F (ε)dε − 2κ = 0.
−κ+µ02 −µ∗
We now use the symmetry of the distribution of the errors. This translates to
F (ε) = 1 − F (−ε).
(B.19)
This implies
Z
κ+µ∗ −µ02
F (ε̃)dε̃ = κ.
−κ+µ∗ −µ02
On the other hand, we also have that
Z
κ+µ02 −µ∗
F (ε)dε = κ.
−κ+µ02 −µ∗
By changing the variables in the previous integrals we arrive at
Z κ
F (ε + µ02 − µ∗ ) − F (ε + µ∗ − µ02 )dε = 0.
−κ
The previous integral is always larger than 0 by assumption unless µ∗ = µ02 .
Elsener and van de Geer/Robust Low-Rank Matrix Estimation
43
B.5. Proof of Lemma 4.1 in the main text
Proof of Lemma 4.1 in the main text. The proof of this lemma is analogous to
the first part of the proof of Corollary 2 in Negahban and Wainwright (2012).
Take σ > 0 such that
s = max k ∈ {1, . . . , q} |Λ0k > σ .
Then, we have
B − nuclear =
q
X
Λk
k=s+1
q
X
Λk
σ
k=s+1
r
q
X
Λk
≤σ
σ
=σ
k=s+1
≤ σ 1−r ρrr .
Moreover, we have
sσ r = σ r
s
X
k=1
1{Λ0k >σ} = σr
s
X
k=1
1 Λ0k r
σ
>1
≤ σr
s 0 r
X
Λ
k
k=1
σ
≤ ρrr ,
where Λ0k denotes a singular value of the matrix B 0 . Therefore we obtain
s ≤ σ −r ρrr .
References
Beck, A. and Teboulle, M. (2009). Gradient-based algorithms with applications to signal recovery problems. Convex optimization in signal processing
and communications. 42–88. MR2767564
Bousquet, O. (2002). A Bennett concentration inequality and its application
to suprema of empirical processes. C. R. Math. Acad. Sci. Paris 334 495–500.
MR1890640
Bühlmann, P. and van de Geer, S. (2011). Statistics for High-Dimensional
Data: Methods, Theory and Applications. Springer Series in Statistics.
Springer, Heidelberg. MR2807761
Cambier, L., Absil, P-A.(2016). Robust low-rank matrix completion by Riemannian optimization SIAM J. Sci. Comput. 38 S440-S460. MR3565571
Candès, E. J., Li, X., Ma, Y. and Wright, J. (2011). Robust Principal
Component Analysis? J. ACM 58 11. MR2811000
Elsener and van de Geer/Robust Low-Rank Matrix Estimation
44
Candès, E. J. and Plan, Y. (2010). Matrix Completion with Noise Proceedings
of the IEEE 98 925–936.
Chandrasekaran, V., Sanghavi, S., Parrilo, P. A. and Willsky, A. S.
(2011). Rank-sparsity incoherence for matrix decomposition. SIAM J. Optim.
21 572–596. MR2817479
Chen, Y., Jalali, A., Sanghavi, S. and Caramanis, C. (2013). Low-rank
matrix recovery from errors and erasures. IEEE Trans. Inform. Theory 59
4324–4337.
CVX Research, I. (2012). CVX: Matlab Software for Disciplined Convex
Programming, version 2.0. http://cvxr.com/cvx.
Foygel, R., Shamir, O., Srebro, N. and Salakhutdinov, R. R. (2011).
Learning with the weighted trace-norm under arbitrary sampling distributions. Advances in Neural Information Processing Systems 2133–2141.
Klopp, O., Lounici, K. and Tsybakov, A. B. (2016). Robust matrix completion. Probab. Theory Related Fields. 1–42
Koltchinskii, V., Lounici, K. and Tsybakov, A. B. (2011). Nuclear-norm
penalization and optimal rates for noisy low-rank matrix completion. Ann.
Statist. 39 2302–2329. MR2906869
Lafond, J. (2015). Low rank matrix completion with exponential family noise.
J. Mach. Learn. Res.: Workshop and Conference Proceedings. COLT 2015
Proceedings 40 1–20.
Lange, K. (2013). Optimization. Springer Texts Statist. Springer, New York.
MR3052733
Ledoux, M. and Talagrand, M. (1991). Probability in Banach Spaces:
Isoperimetry and Processes. Ergeb. Math. Grenzgeb. 23. Springer-Verlag,
Berlin. MR1102015
Li, X. (2013). Compressed sensing and matrix completion with constant proportion of corruptions. Constr. Approx. 37 73–99. MR3010211
MovieLens 100k Dataset (1998). Available at http://grouplens.org/
datasets/movielens/100k/.
MovieLens 1m Dataset (2003). Available at http://grouplens.org/
datasets/movielens/1m/.
Negahban, S. and Wainwright, M. J. (2011). Estimation of (near) low-rank
matrices with noise and high-dimensional scaling. Ann. Statist. 39 1069–1097.
MR2816348
Negahban, S. and Wainwright, M. J. (2012). Restricted strong convexity
and weighted matrix completion: Optimal bounds with noise. J. Mach. Learn.
Res. 13 1665–1697. MR2930649
Parikh, N. and Boyd, S. (2014). Proximal Algorithms. Foundations and
Trends in optimization 1 123–231.
Rohde, A. and Tsybakov, A. B. (2011). Estimation of high-dimensional lowrank matrices. Ann. Statist. 39 887–930. MR2816342
Srebro, N., Rennie, J. and Jaakkola, T. S. (2005). Maximum-margin matrix factorization. Proceedings of the NIPS Conference 1329–1336. Vancouver,
Canada.
Srebro, N. and Shraibman, A. (2005). Rank, trace-norm and max-norm. In
Elsener and van de Geer/Robust Low-Rank Matrix Estimation
45
Learning Theory 545–560. Springer. MR2203286
Tibshirani, R. (1996). Regression shrinkage and selection via the lasso.
Journal of the Royal Statistical Society. Series B (Methodological) 267–288.
MR1379242
Unlocbox (2016). Unlocbox: (Matlab convex optimization toolbox version
1.7.3). https://lts2.epfl.ch/unlocbox/.
van de Geer, S. A. (2000). Applications of Empirical Process Theory 91.
Cambridge University Press, Cambridge. MR1739079
van de Geer, S. (2001). Least squares estimation with complexity penalties.
Math. Methods Statist. 10 355–374. MR1867165
van de Geer, S. (2016). Estimation and Testing under Sparsity: École d’Été
de Probabilités de Saint-Flour XLV-2015. Lecture Notes in Mathematics.
Springer. MR3526202
van der Vaart, A. W. and Wellner, J. A. (1996). Weak Convergence and
Empirical Processes. Springer Series in Statistics, Springer-Verlag, New York.
MR1385671
Watson, G. A. (1992). Characterization of the Subdifferential of some matrix
norms. Linear Algebra and its Applications 170 33–45. MR1160950
| 10math.ST
|
arXiv:1710.10710v2 [cs.CV] 16 Nov 2017
On Pre-Trained Image Features and Synthetic Images for Deep Learning
Stefan Hinterstoisser
X
Vincent Lepetit
Université de Bordeaux
Paul Wohlhart
X
[email protected]
[email protected]
[email protected]
Kurt Konolige
X
[email protected]
Abstract
Deep Learning methods usually require huge amounts of
training data to perform at their full potential, and often
require expensive manual labeling. Using synthetic images
is therefore very attractive to train object detectors, as the
labeling comes for free, and several approaches have been
proposed to combine synthetic and real images for training.
In this paper, we show that a simple trick is sufficient to
train very effectively modern object detectors with synthetic
images only: We ’freeze’ the layers responsible for feature
extraction to generic layers pre-trained on real images, and
train only the remaining layers with plain OpenGL rendering. Our experiments with very recent deep architectures for object recognition (Faster-RCNN, R-FCN, MaskRCNN) and image feature extractors (InceptionResnet and
Resnet) show this simple approach performs surprisingly
well.
Figure 1. We show that feature extractor layers from modern object detectors pre-trained on real images can be used on synthetic
images to learn to detect objects in real images. The top-left image
shows the CAD model we used to learn to detect the object in the
three other images.
1. Introduction
The capability of detecting objects in challenging environments is a key component for many computer vision
and robotics task. Current leading object detectors—FasterRCNNs [19], SSD [13], RFCN [4], Yolo9000 [18]—all rely
on convolutional neural networks. However, to perform at
their best, they require huge amounts of labeled training
data, which is usually time consuming and expensive to create.
Using synthetic images is therefore very attractive to
train object detectors, as the labeling comes for free. Unfortunately, synthetic rendering pipelines are usually unable to
reproduce the statistics produced by their real-world counterparts. This is often referred to as the ’domain gap’ between synthetic and real data and the transfer from one to
another usually results in deteriorated performance, as observed in [26] for example.
Several approaches have tried to overcome this domain
gap. For instance, [5, 7, 17] use synthetic images in addition to real ones to boost performance. While this usually
results in good performance, it is still dependent on real
world labeled data. Transfer learning approaches are also
possible [21, 3, 6], however they also require real images of
the objects to detect. [8, 1] create photo-realistic graphics
renderings and [8, 1, 7, 27] compose realistic scenes which
both shows to improve performance. Unfortunately, these
strategies are usually difficult to engineer, need domain specific expertise and require some additional data such as illumination information and scene labeling to create realistic scenes. [26] uses ’domain randomization’ to narrow
the gap. While this has shown very promising results, it
1
Proposal Generator
Proposal Generator
Objectness
Classification
Objectness
Classification
Box
Regression
(vgg, incep+on,
resnet, etc)
Box
Regression
(vgg, incep+on,
resnet, etc)
Box Classifier
Feature Extractor
Multiway
Classification
Box Classifier
Feature Extractor
Box
Refinement
Faster-RCNN
Multiway
Classification
Box
Refinement
R-FCN
Figure 2. The architectures of two recent object detectors with their feature extractors isolated as described in [12] (Figure taken from
[12]).
has mainly been demonstrated to work with simple objects
and scenarios. Other works [22, 2] use Generative Adversarial Networks (GANs) to remove the domain gap, however, GANs are still very brittle and hard to train, and to our
knowledge they have not been used for detection tasks yet.
In this paper we consider a simple alternative solution.
As shown by [12] and illustrated in Fig. 2, many of today’s
modern feature extractors can be split into a feature extractor and some remaining layers that depend on the metaarchitecture of the detector. Our claim is twofold: a) the
pre-trained feature extractors are already rich enough and
do not need to be retrained when considering new objects
to detect; b) when applied to an image synthetically generated using simple rendering techniques, the feature extractors work as a “projector” and output image features that are
close to real image features.
Therefore, by freezing the weights of feature extractor
pre-trained on real data and by only adapting the weights of
the remaining layers during training, we are able to train
state-of-the-art object detectors purely on synthetic data.
While using pre-trained layers for feature extraction is not
new (for example, VGG [23] has been used extensively
for this purpose), our contribution is therefore to show that
this approach is also advantageous when training with synthetic data. We also show in this paper that this observation is fairly general and we give both qualitative and
quantitative experiments for different detectors—FasterRCNN [19], RFCN [4] and Mask-RCNN [9]—and different feature extraction networks—InceptionResnet [25] and
Resnet101 [10].
Furthermore, we show that different cameras have different image statistics that allow different levels of performance when re-trained on synthetic data. We will demonstrate that performance is significantly boosted for these
cameras if our simple approach is applied.
In the remainder of the paper we first discuss related
work, describe how we generate synthetic data, demonstrate
the domain gap between synthetic and real data, and show
how to boost object detection by freezing the pre-trained
feature extractor during training. Finally, we show qualita-
tive and quantitative experiments.
2. Related Work
Mixing real and synthetic data to improve detection performance is a well established process. Many approaches
such as [5, 7, 24], to mention only very recent ones, have
shown the usefulness of adding synthetic data when real
data is limited. In contrast to [5, 7] which use real masked
image patches, [24] uses 3D CAD models and a structurepreserving deformation pipeline to generate new synthetic
models to prevent overfitting. However, while these approaches obtain better results compared to detectors trained
on real data only, they still require real data.
In order to avoid expensive labeling in terms of time
and money, some approaches learn object detectors purely
from synthetic data. For instance, a whole line of work uses
photo-realistic rendering [1, 8] and complex scene composition [8, 1, 7, 27] to achieve good results, and [15] stresses
the need for photo-realistic rendering. Some approaches
even use physics engines to enable realistic placing of objects [14]. This requires significant resources and highly
elaborate pipelines that are difficult to engineer and need
domain specific expertise [20]. Furthermore, additional effort is needed to collect environment information like illumination information [1] to produce photo-realistic scenes.
For real scene composition, one also needs to parse real
background images semantically in order to place the objects meaningful into the scene.
This usually needs manual post-processing or labeling
which is both expensive and time consuming. While these
graphics based rendering approaches already show some of
the advantages of learning from synthetic data, they usually
suffer from the domain gap between real and synthetic data.
To address this, a new line of work [5, 7, 17] moves away
from graphics based renderings to composing real images.
The underlying theme is to paste masked patches of objects into real images, and thus reducing the dependence on
graphics renderings. This approach has the advantage that
the images of the objects are already in the right domain—
the domain of real images—and thus, the domain gap be-
tween image compositions and real images is smaller than
the one of graphics based rendering and real images. While
this has shown quite some success, the amount of data is
still restricted to the number of images taken from the object in the data gathering step and therefore does not allow
to come up with new views of the object. Furthermore, it is
not possible to generate new illumination settings or proper
occlusions since shape and depth are usually not available.
In addition, this approach is dependent on segmenting out
the object from the background which is prone to segmentation errors when generating the object masks.
Recently, several approaches [22, 2] tried to overcome
the domain gap between real and synthetic data by using
generative adversarial networks (GANs). This way they
produced better results than training with real data. However, GANs are hard to train and up to now, they have
mainly shown their usefulness on regression tasks and not
on detection applications.
Yet another approach is to rely on transfer learning [21,
3, 6], to exploit a large amount of available data in a source
domain, here the domain of synthetic images, to correctly
classify data from the target domain, here the domain of
real images, for which the amount of training data is limited. This is typically done by tighting two predictors together, one trained on the source domain, the other on the
target domain or by training a single predictor on the two
domains. This is a general approach as the source and target domains can be very different, compared to synthetic
and real images, which are more related to each other. In
this paper, we exploit this relation by applying the same
feature extractor to the two domains. However, in contrast
to [21, 3, 6] we do not need any real images of the objects
of interest in our approach.
3. Method
In this section, we will present our simple synthetic data
generation pipeline and describe how we change existing
state-of-the-art object detectors to enable them to learn from
synthetic data. In this context, we will focus on object instance detection. Throughout this paper, we will mainly rely
on Faster-RCNN [19] since it demonstrated the best detection performance among a whole family of object detectors
as shown in [12]. However, in order to show the generability of our approach, we will also present additional quantitative and qualitative results of other detectors (RFCN [4]
and Mask-RCNN [9]) in Section 4.7.
3.1. Synthetic Data Generation Pipeline
Similar to [5], we believe that while global consistency
can be important, local appearance—so called patch-level
realism—is also important. The term patch-level realism
refers to the observation that the content of the bounding
box framing the rendered object looks realistic.
This principle is an important assumption for our synthetic data generation pipeline, shown in Fig. 3. For each
object, we start by generating a large set of poses uniformly
covering the pose space in which we want to be able to detect the corresponding object. As in [11], we generate rotations by recursively dividing an icosahedron, the largest
convex regular polyhedron. We substitute each triangle into
four almost equilateral triangles, and iterate several times.
The vertices of the resulting polyhedron give us then the
two out-of-plane rotation angles for the sampled pose with
respect to the coordinate center. In addition to these two
out-of-plane rotations, we also use equally sampled in-plane
rotations. Furthermore, we sample the scale logarithmically
to guarantee an approximate linear change in pixel coverage
of the reprojected object between consecutive scale levels.
The object is rendered at a random location in a randomly selected background image using a uniform distribution. The selected background image is part of a large
collection of highly cluttered real background images taken
with the camera of choice where the objects of interest are
not included. To increase the variability of the background
image set, we randomly swap the three background image channels and randomly flip and rotate the images (0◦ ,
90◦ , 180◦ and 270◦ ). We also tried to work without using real background images and experimented with backgrounds only exhibiting one randomly chosen color, however, that didn’t lead to good results.
We use plain OpenGL with simple Phong shading [16]
for rendering where we allow small random perturbations
of the ambient, the diffuse and the specular parameters. We
also allow small random perturbations of the light color.
We add random Gaussian noise to the rendered object and
blur it with a Gaussian kernel, including its boundaries with
the adjacent background image pixels to better integrate the
rendering with the background. We also experimented with
different strategies for integrating the rendered object in images as [5], however this did not result in significant performance improvements.
3.2. Freezing a Pre-Trained Feature Extractor
As shown in [12] and illustrated in Fig. 2, many state-ofthe-art object detectors including Faster-RCNN [19], MaskRCNN [9], and R-FCN [4] can be decoupled as a ’metaarchitecture’ and a feature extractor such as VGG [23],
Resnet [10], or InceptionResnet [25].
While the meta-architecture defines the different modules and how they work together, the feature extractor is
a deep network cut at some selected intermediate convolutional level. The remaining part can be used as part of the
multi-way classification+localization of the object detector.
As discussed in the introduction, for the feature extractor,
we use frozen weights pre-learned on real images, to enable
training the remaining part of the architecture on synthetic
3D Poses
Random Light Color
3D CAD Model
Large collection of highly
textured background images
(random channel swap and
rotation applied)
Random Illumination
Random Noise
Blurring
Figure 3. Our synthetic data generation pipeline. For each generated 3D pose and object, we render the object over a randomly selected
cluttered background image using OpenGL and the Phong illumination model [16]. We use randomly perturbed light color for rendering
and add image noise to the rendering. Finally, we blur the object with a Gaussian filter. We also compute a tightly fitting bounding box
using the object’s CAD model and the corresponding pose.
eras, as the quality of the camera influences the recognition
results as we will show. The rest of the section reports our
experiments and the conclusions we draw from them.
4.1. Objects and 3D CAD models
(a) Real Objects
(b) 3D CAD Models
Figure 4. (a) The real objects used in our experiments and (b) their
CAD models. We chose our objects carefully to represent different
colors and 3D shapes and to cover different fields of applications
(industrial objects, household objects, toys).
images only.
In practice, we use the Google’s public available OpenSource version [12] of Faster-RCNN and RFCN, and our
own implementation of Mask-RCNN. The ’frozen’ parts
are taken according to [12], by training InceptionResnet
and Resnet101 on a classification task on the ImageNetCLs dataset. We freeze InceptionResnet (v2) after the repeated use of block17 and right before layer Mixed 7a and
Resnet101 after block3. All other remaining parts of the
networks are not ’frozen’, meaning their weights are free to
adapt when we train the detector on synthetic images.
We evaluate this approach in the next section.
4. Experiments
In this section, we first describe the dataset we created
for these evaluations, made of synthetic and real images of
10 different objects. We also considered two different cam-
As shown in Fig. 4, we carefully selected the objects we
used in our experiments: We tried to represent different colors, textures (homogeneous color versus highly textured),
3D shapes and material properties (reflective versus nonreflective). Except for the mug and the bottle, the 3D shapes
of the objects we selected can look very different from different views. We also tried to consider objects from different application fields (industrial objects, household objects,
toys). For each real object we have a textured 3D CAD
model at hand which we generated using our in-house 3D
scanner.
4.2. Cameras
We consider two cameras, an AsusXtionPROLive and
a PtGreyBlackfly. For each camera, we generated a training dataset and an evaluation dataset. The training datasets
consist of approximatively 20K and the evaluation datasets
of approximatively 1K manually labeled real world images.
Each sample image contains one of the 10 objects shown
in Fig. 4 in challenging environments: heavy background
clutter, illumination changes, etc. In addition, we made sure
that each object is shown from various poses as this is very
important for object instance detection. Furthermore, for
each dataset all objects have the same amount of images.
4.3. Freezing the Feature Extractor
Fig. 5 shows that when Faster-RCNN is trained on synthetic images and tested on real images, it performs signif-
1.0
0.6
0.4
0.20
0.2
0.15
0.0
InceptionResnet
Resnet101
Prec [mAP]
0.8
synthetic re-trained
synthetic frozen features
real
0.10
0.00
0
20
40
60
80
100
120
140
Euclidean Distance of generated Image Features [from Real and Synthetic Images]
Figure 6. Top: Two examples of pairs of a real image and a synthetic one for the same object under the same pose. Bottom: Distributions of the Euclidean distances between image features generated for the real images and the corresponding synthetic images.
See text in Section 4.3 for details.
0.6
0.4
0.2
0.0
synthetic re-trained
synthetic frozen features
0.05
(a) PtGreyBlackfly
1.0
Relative Frequency
Prec [mAP]
0.8
synthetic re-trained
synthetic frozen features
real
InceptionResnet
Resnet101
(b) AsusXtionPROLive
Figure 5. The effect of freezing the pre-trained feature extractor,
for two different cameras. Training the feature extractors on synthetic images performs poorly, and totally fails in the case of the
AsusXtionPROLive camera. When using feature extractors pretrained on real images without retraining them, the performances
of detectors trained on synthetic data are almost as good as when
training them on real data, except when ResNet101 is used with
images from the AsusXtionPROLive camera.
icantly worse than when trained on real data. By contrast,
when we freeze the feature extractor’s weights during training to values pre-trained on real images, and only train the
remaining parts of the detector, we get a significant performance boost. We even come close to detectors trained
purely on real world data, as we typically obtained up to
95% of the performance when trained on synthetic data.
To get a better intuition why freezing the feature extractor gives significant better results than retraining it on synthetic data, we performed the following experiment: We
created 1000 image pairs with different objects under various poses. Each image pair consists of one image that
shows the real object and of another image where we superpose a rendering of the object’s CAD model on top of
the real image, under the same pose as the real object. The
top of Fig. 6 shows two examples.
We then compared the distributions of the Euclidean distances between image features generated for the real images
and the corresponding synthetic images. As we can see at
the bottom of Fig. 6, the distribution is much more clustered
around 0 when the features are computed using a frozen feature extractor pre-trained on real images (red) compared to
the distribution obtained when the pre-trained feature extractor is finetuned on synthetic images (blue).
4.4. Freezing the Feature Extractor at Different
Layers
We also performed experiments where we freeze the feature extractor at different intermediate layers i.e. layers lying between the input and the output layers of the feature extractor as specified in Section 3.2. As can be seen in Fig. 7,
freezing the full feature extractor always performs best. For
the AsusXtionPROLive camera, freezing the feature extractor on intermediate levels even results in a dramatic loss of
performance.
4.5. On Finetuning the Feature Extractor
One may wonder if the domain shift between synthetic
and real images still leads to decreased performance after
the detector was trained for some time with the pre-trained
feature extractor frozen. One could argue that all remaining
detector weights have already started to converge and therefore, the domain shift is far less influential. As a result,
1.0
0.8
0.6
Prec [mAP]
Prec [mAP]
0.8
1.0
synthetic re-trained
synthetic frozen features at layer Mixed_5b
synthetic frozen features at layer Mixed_6a
synthetic frozen features
0.4
0.2
0.0
synthetic re-trained
synthetic frozen features after block1
synthetic frozen features after block2
synthetic frozen features
0.6
0.4
0.2
PtGrey
0.0
Asus
Ptgrey
a) InceptionResnet [25]
Asus
b) Resnet101 [10]
Figure 7. We freeze features at different layers of InceptionResnet [25] and Resnet101 [10]. We can see that freezing the full feature
extractor performs best (yellow).
inc_1200K
inc_900K
inc_400K
inc_frozen_features
res_1200K
res_900K
res_400K
res_frozen_features
0.7
Precision [mAP]
0.6
0.5
0.4
0.3
inc_1200K
inc_900K
inc_400K
inc_frozen_features
res_1200K
res_900K
res_400K
res_frozen_features
0.6
0.5
0.4
0.3
0.2
0.2
0.1
0.1
0.0
0.7
Precision [mAP]
0.8
0
500000
1000000
Step
1500000
2000000
a) PtGreyBlackfly
0.0
0
500000
1000000
Step
1500000
2000000
b) AsusXtionPROLive
Figure 8. Finetuning the feature extractor after 400K, 900K and 1200K steps where the pre-trained feature extractor was frozen for the
PtGreyBlackfly and the AsusXtionPROLive cameras. We show results for the InceptionResnet [25] and Resnet101 [10] architectures.
the frozen feature extractor could be unfrozen to finetune
its weights to adapt to the learning task.
However, as we show in Fig. 8, this is not true. Even
after 1200K training steps where the feature extractor was
frozen and the detection performance starts to plateau the
detector’s performance degrades significantly if the frozen
feature extractor is unfrozen and its weights are finetuned.
Table 1 gives the corresponding numbers.
4.6. Ablation Experiments
In the following experiments, we investigated the influence of the single steps in the image generation pipeline.
For all these experiments we used InceptionResnet [25] as
feature extractor. The feature extractor itself was frozen.
We found out that blurring the rendered object and its adjacent image pixels gives a huge performance boost. Adding
noise to the rendered object or enabling random light color
did not give much improvement in performance and its influence depends on the camera used. As already mentioned,
we also experimented with different blending strategies as
in [5], that is using different blending options in the same
dataset: no blending, Gaussian blurring and Poisson blending, however we could not find significant performance improvements.
We also investigated what happens if we use the internal
camera parameter of our target camera but a background
dataset taken with another camera. As we can see in Fig. 9
results seem to stay approximately the same for the PtGreyBlackfly camera and seem to improve for the AsusXtionPROLive camera. The later seems reasonable since the
background images taken with the PtGreyBlackfly camera
are more cluttered and are showing more background variety than the background images taken with the AsusXtionPROLive camera. These results suggest that the camera
images can be taken from an arbitrary source and we only
have to make sure that a high amount of background variety
is provided.
4.7. RFCN and MASK-RCNN
To show the generality of our approach, we also performed several addition experiments. Fig. 10 shows the results for RFCN [4] trained only on synthetic data with the
synthetic
frozen
400K
900K
1200K
Asus
Prec [mAP]
Prec [[email protected]]
Prec [[email protected]]
Acc [@100]
.000/.000
.000/.000
.000/.000
.000/.010
.617/.171
.948/.385
.733/.130
.686/.256
.000/.000
.000/.000
.000/.000
.000/.000
.000/.000
.000/.000
.000/.000
.000/.000
.061/.006
.114/.016
.064/.004
.079/.007
.719/.681
.983/.988
.872/.844
.772/.742
PtGrey
real
Prec [mAP]
Prec [[email protected]]
Prec [[email protected]]
Acc [@100]
.374/.243
.537/.410
.431/.239
.461/.324
.725/.687
.971/.966
.886/.844
.771/.736
.426/.317
.606/.491
.495/.355
.483/.384
.514/.485
.717/.685
.593/.564
.577/.551
.709/.626
.936/.912
.835/.756
.768/.695
.764/.742
.987/.987
.908/.916
.808/.804
Table 1. Outcomes of all our experiments. We give numbers for InceptionResnet [25] / Resnet101 [10]. Except for the experiments with
real data (last column), all experiments were performed on synthetic data only. We emphasized the best results trained on synthetic data.
Prec [mAP]
0.8
0.6
w/o Postprocessing
Noise
Colored Light
Noise & Colored Light
Blur
Blur & Noise
Blur & Colored Light
Blur & Noise & Colored Light
Asus Background Images
1.0
0.8
Prec [mAP]
1.0
0.4
0.6
0.4
0.2
0.2
0.0
0.0
a) PtGreyBlackfly
w/o Postprocessing
Noise
Colored Light
Noise & Colored Light
Blur
Blur & Noise
Blur & Colored Light
Blur & Noise & Colored Light
PtGrey Background Images
b) AsusXtionPROLive
Figure 9. Influences of the different building blocks for synthetic rendering for the PtGreyBlackfly and the AsusXtionPROLive cameras.
Results were obtained with InceptionResnet [25] as the feature extractor. Blurring is clearly a useful yet simple operation to apply to the
synthetic images to improve the results.
1.0
Prec [mAP]
0.8
synthetic re-trained
synthetic frozen features
real
0.6
0.4
0.2
0.0
RFCN + InceptionResnet
Mask-RCNN + InceptionResnet
Figure 10. Results using RFCN [4] on the PtGreyBlackfly dataset.
Freezing the feature extractor boosts performance significantly on
this method as well. We observe the same results if we train MaskRCNN on the AsusXtionPROLive dataset.
feature extractor frozen and compares them with those using RFCN trained on real data and and those using RFCN
re-trained on synthetic data. Freezing the feature extractor
helps to unlock significant performance improvements also
here.
Fig. 10 also shows quantitative results of MaskRCNN [9] trained only on synthetic data with the feature
extractor frozen. Similar to what we observed with FasterRCNN and RFCN, freezing the feature extractor significantly boosts the performance when trained on synthetic
data. Fig. 12 shows that we are able to detect objects in
highly cluttered environments under various poses and get
reasonable masks. This result is especially important since
it shows that exhaustive (manual) pixel labeling is made redundant by training from synthetic data.
4.8. Qualitative Results
Fig. 11 shows some qualitative results on images exhibiting several of the 10 objects we considered in various poses with heavy background clutter and illumination
changes. We use Faster-RCNN [19] with the InceptionRes-
Figure 11. Results of Faster-RCNN trained on synthetic images only with the feature extractor frozen. The objects are detected in highly
cluttered scenes and many different instances are available in one image. Note that the different objects are seen in different arbitrary poses.
Figure 12. Results of Mask-RCNN [9] trained on synthetic images only with the feature extractor frozen. The images were taken with the
AsusXtionPROLive camera in a highly cluttered environment under various poses.
Figure 13. Objects with similar shapes and colors detected in challenging environments. The detector was trained on synthetic images
only.
net [25] as feature extractor and trained the rest of the network on synthetic images only. Fig. 12 shows results of
Mask-RCNN [9] trained on synthetic images only. Fig. 13
shows some other objects trained with the method presented
in this paper.
5. Conclusion
We have shown that by freezing a pre-trained feature extractor we are able to train state-of-the-art object detectors
on synthetic data only. The results are close to approaches
trained on real data only. While we have demonstrated that
object detectors re-trained on synthetic data lead to poor
performances and that images from different cameras lead
to different results, freezing the feature extractor always
gives a huge performance boost.
The results of our experiments suggest that simple rendering is sufficient to achieve good performances and that
complicated scene composition does not seem necessary.
Training from rendered 3D CAD models allows us to detect objects from all possible viewpoints which makes the
need for a real data generation and expensive manual labeling pipeline redundant.
Acknowledgments: The authors thank Google’s VALE
team for tremendous support using the Google Object Detection API, especially Jonathan Huang, Alireza Fathi,
Vivek Rathod, and Chen Sun. In addition, we thank Kevin
Murphy, Vincent Vanhoucke, Konstantinos Bousmalis and
Alexander Toshev for valuable discussions and feedback.
References
[1] H. A. Alhaija, S. K. Mustikovela, L. Mescheder, A. Geiger,
and C. Rother. Augmented Reality Meets Deep Learning
for Car Instance Segmentation in Urban Scenes. In British
Machine Vision Conference, 2017. 1, 2
[2] K. Bousmalis, N. Silberman, D. Dohan, D. Erhan, and D. Krishnan. Unsupervised Pixel-Level Domain Adaptation with
Generative Adversarial Networks. In Conference on Computer Vision and Pattern Recognition, 2017. 2, 3
[3] K. Bousmalis, G. Trigeorgis, N. Silberman, D. Krishnan, and
D. Erhan. Domain separation networks. In Advances in Neural Information Processing Systems, 2016. 1, 3
[4] J. Dai, Y. Li, K. He, and J. Sun. R-FCN: Object Detection via
Region-Based Fully Convolutional Networks. In Advances
in Neural Information Processing Systems, 2016. 1, 2, 3, 6,
7
[5] D. Dwibedi, I. Misra, and M. Hebert. Cut, Paste and Learn:
Surprisingly Easy Synthesis for Instance Detection. In arXiv
Preprint, 2017. 1, 2, 3, 6
[6] Y. Ganin, E. Ustinova, H. Ajakan, P. Germain, H. Larochelle,
F. Laviolette, M. Marchand, and V. Lempitsky. Domainadversarial training of neural networks. In Journal of Machine Learning Research, 2016. 1, 3
[7] G. Georgakis, A. Mousavian, A. C. Berg, and J. Kosecka.
Synthesizing Training Data for Object Detection in Indoor
Scenes. In Robotics: Science and Systems Conference, 2017.
1, 2
[8] A. Gupta, A. Vedaldi, and A. Zisserman. Synthetic Data
for Text Localisation in Natural Images. In Conference on
Computer Vision and Pattern Recognition, 2016. 1, 2
[9] K. He, G. Gkioxari, P. Dollár, and R. Girshick. Mask RCNN. In arXiv Preprint, 2017. 2, 3, 7, 8
[10] K. He, X. Zhang, S. Ren, and J. Sun. Deep Residual Learning
for Image Recognition. In Conference on Computer Vision
and Pattern Recognition, 2016. 2, 3, 6, 7
[11] S. Hinterstoisser, V. Lepetit, S. Ilic, S. Holzer, G. Bradski,
K. Konolige, and N. Navab. Model Based Training, Detection and Pose Estimation of Texture-Less 3D Objects in
Heavily Cluttered Scenes. In Asian Conference on Computer
Vision, 2012. 3
[12] J. Huang, V. Rathod, C. Sun, M. Zhu, A. Korattikara,
A. Fathi, I. Fischer, Z. Wojna, Y. Song, S. Guadarrama, and
K. Murphy. Speed and Accuracy Trade-Offs for Modern
Convolutional Object Detectors. In Conference on Computer
Vision and Pattern Recognition, 2017. 2, 3, 4
[13] W. Liu, D. Anguelov, D. Erhan, C. Szegedy, S. E. Reed,
C. Fu, and A. C. Berg. SSD: Single Shot Multibox Detector.
In European Conference on Computer Vision, 2016. 1
[14] C. Mitash, K. E. Bekris, and A. Boularias. A Self-Supervised
Learning System for Object Detection Using Physics Simulation and Multi-View Pose Estimation. In International
Conference on Intelligent Robots and Systems, 2017. 2
[15] Y. Movshovitz-attias, T. Kanade, and Y. Sheikh. How Useful
is Photo-Realistic Rendering for Visual Learning? In European Conference on Computer Vision, 2016. 2
[16] B. T. Phong. Illumination for Computer Generated Pictures.
In Communications of the ACM, 1975. 3, 4
[17] M. Rad and V. Lepetit. Bb8: A Scalable, Accurate, Robust
to Partial Occlusion Method for Predicting the 3D Poses of
Challenging Objects Without Using Depth. In International
Conference on Computer Vision, 2017. 1, 2
[18] J. Redmon and A. Farhadi. Yolo9000: Better, Faster,
Stronger. In Conference on Computer Vision and Pattern
Recognition, 2017. 1
[19] S. Ren, K. He, R. Girshick, and J. Sun. Faster R-CNN:
Towards Real-Time Object Detection with Region Proposal
Networks. In Advances in Neural Information Processing
Systems. 2015. 1, 2, 3, 7
[20] S. R. Richter, V. Vineet, S. Roth, and V. Koltun. Playing for
Data: Ground Truth from Computer Games. In European
Conference on Computer Vision, 2016. 2
[21] A. Rozantsev, M. Salzmann, and P. Fua. Beyond Sharing
Weights for Deep Domain Adaptation. In Conference on
Computer Vision and Pattern Recognition, 2017. 1, 3
[22] A. Shrivastava, T. Pfister, O. Tuzel, J. Susskind, W. Wang,
and R. Webb. Learning from Simulated and Unsupervised
Images through Adversarial Training. In Conference on
Computer Vision and Pattern Recognition, 2017. 2, 3
[23] K. Simonyan and A. Zisserman. Very Deep Convolutional
Networks for Large-Scale Image Recognition. In International Conference for Learning Representations, 2015. 2, 3
[24] H. Su, C. R. Qi, Y. Li, and L. J. Guibas. Render for CNN:
viewpoint estimation in images using cnns trained with rendered 3d model views. In ICCV, 2015. 2
[25] C. Szegedy, S. Ioffe, V. Vanhoucke, and A. A. Alemi.
Inception-V4, Inception-Resnet and the Impact of Residual
Connections on Learning. In American Association for Artificial Intelligence Conference, 2017. 2, 3, 6, 7
[26] J. Tobin, R. Fong, A. Ray, J. Schneider, W. Zaremba, and
P. Abbeel. Domain Randomization for Transferring Deep
Neural Networks from Simulation to the Real World. In International Conference on Intelligent Robots and Systems,
2017. 1
[27] G. Varol, J. Romero, X. Martin, N. Mahmood, M. J. Black,
I. Laptev, and C. Schmid. Learning from Synthetic Humans.
In Conference on Computer Vision and Pattern Recognition,
2017. 1, 2
| 1cs.CV
|
Estimation of the Hurst and the stability indices
of a H-self-similar stable process
Thi To Nhu DANG∗, Jacques ISTAS†
arXiv:1506.05593v3 [math.ST] 18 Oct 2017
October 19, 2017
Abstract
In this paper we estimate both the Hurst and the stability indices of a H-self-similar stable process.
More precisely, let X be a H-sssi (self-similar stationary increments) symmetric α-stable process. The
process X is observed at points nk , k = 0, . . . , n. Our estimate is based on β-negative power variations
with − 21 < β < 0. We obtain consistent estimators, with rate of convergence, for several classical
H-sssi α-stable processes (fractional Brownian motion, well-balanced linear fractional stable motion,
Takenaka’s process, Lévy motion). Moreover, we obtain asymptotic normality of our estimators for
fractional Brownian motion and Lévy motion.
Keywords: H-sssi processes; stable processes; self-similarity parameter estimator; stability parameter
estimator.
Contents
1 Introduction
2
2 Main results
3
3 Examples
3.1 Fractional Brownian motion . . . . .
3.2 SαS-stable Lévy motion . . . . . . .
3.3 Well-balanced linear fractional stable
3.4 Takenaka’s processes . . . . . . . . .
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
4
4
6
7
8
4 Proofs
4.1 Negative power expectation and auxiliary results
4.1.1 Auxiliary results . . . . . . . . . . . . . .
4.1.2 Proof of Lemma 4.1 . . . . . . . . . . . .
4.1.3 Proof of Theorem 4.1 . . . . . . . . . . .
4.1.4 Proof of Theorem 4.2 . . . . . . . . . . .
4.1.5 Proof of Lemma 4.2 . . . . . . . . . . . .
4.1.6 Proof of Lemma 4.3 . . . . . . . . . . . .
4.2 Proof of Theorem 2.1 . . . . . . . . . . . . . . . .
4.3 Proofs related to Section 3 . . . . . . . . . . . . .
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
8
8
9
10
10
13
19
19
19
23
. . . . .
. . . . .
motion
. . . . .
∗
.
.
.
.
The university of Danang, University of Economics, 71 Ngu Hanh Son, Danang, Vietnam. E-mail: [email protected]
Laboratoire Jean Kuntzmann, Université de Grenoble Alpes et CNRS, F-38000 Grenoble, France.
E-mail: [email protected]
†
1
5 Appendix
5.1 Auxiliary results related to Fractional Brownian motion . . . . . . . . . . . . . . . . . . .
5.2 Auxiliary results related to Takenaka’s process . . . . . . . . . . . . . . . . . . . . . . . .
5.3 Auxiliary results related to rate of convergence . . . . . . . . . . . . . . . . . . . . . . . .
31
31
35
36
References
37
1
Introduction
Self-similar processes play an important role in probability because of their connection to limit theorems
and they are widely used to model natural phenomena. For instance, persistent phenomena in internet
traffic, hydrology, geophysics or financial markets, e.g., [9], [17], [21], are known to be self-similar. Stable
processes have attracted growing interest in recent years: data with ”heavy tails” have been collected in
fields as diverse as economics, telecommunications, hydrology and physics of condensed matter, which
suggests using non-Gaussian stable processes as possible models, e.g., [21]. Self-similar α-stable processes
have been proposed to model some natural phenomena with heavy tails, as in [21] and references therein.
The estimation of various indices of H−sssi α−stable processes has been a problem studied since
several decades ago and, even nowadays, it continues to be a challenge. In the case of fractional Brownian
motion, the estimation of the self-similarity index H has attracted attention to many authors and many
methods have been proposed for solving this problem. Among these, one can mention the quadratic
variation method (see e.g. [6], [7], [9], [13]), the p-variation method (see e.g. [8], [18]), the wavelet
coefficients method (see e.g. [1], [5], [14]), the log-variation method (see e.g. [9], [12]). Other references,
like the works of J. Istas, recommend the use of complex variations for estimating the self-similarity index
H of H−sssi processes, but not for estimating α, (see e.g. [11]). For linear fractional stable motions,
strongly consistent estimators of the self-similarity index H, based on the discrete wavelet transform of
the processes, have been proposed without requirement that α to be known, as in [2], [20], [23], [24].
Thus, regarding the estimation of the stability index α, in [3], the authors presented a wavelet estimator
for linear fractional stable motions assuming that H is known. Recently, the corresponding estimation
problem of the stability function and the localisability function for a class of multistable processes was
considered in the discussion paper of R. Le Guével, see [15], based on some conditions that involve the
consistency of the estimators. For linear multifractional stable motions, in [4], the authors presented
strongly consistent estimators of the localisability function H(.) and the stability index α using wavelet
coefficients when α ∈ (1, 2) and H(.) is a Hölder function smooth enough, with values in a compact
subinterval [H, H] of (1/α, 1).
The aim of this work is to construct consistent estimators of the self-similar index H and the stable
index α of H-sssi, SαS-stable processes using a new framework. In the view of the fact that a stable
random variable has a density function, β− negative power variations have expectations and covariances
for −1/2 < β < 0. Our estimates are thus based on these variations. This new approach provides
estimators of H and α without assumptions on the existence moments of the underlying processes. It
also allows us to give an estimator for the self-similarity parameter H without assumption on α and vice
versa, we can estimate the stability index α without assumption on H. In other words, using β− negative
power variations (−1/2 < β < 0), one can obtain the estimators of H and α separately. We prove the
consistency and rates of convergence of the proposed estimators for H and α for the underlying processes
under an assumption on the series of covariances of β-negative power variations (−1/2 < β < 0). Then
obtained results were illustrated by some classical examples: fractional Brownian motions, SαS-stable
Lévy motions, well-balanced linear fractional stable motions and Takenaka’s processes. We then show
that the asymptotic normality of our estimates can be ascertained for the proposed estimators when the
underlying process is a fractional Brownian motion or an SαS-stable Lévy motion.
The remainder part of this article is organized as follows: in the next section, we present the setting,
the assumption and main results to construct the estimators of H and α. In Section 3, some classical
examples for the obtained results in Section 2 are given: fractional Brownian motions, SαS−stable Lévy
2
motions, well-balanced linear fractional stable motions, Takenaka’s processes. In this Section, we also
show the central limit theorem for the cases of the fractional Brownian motion and the SαS−stable Lévy
motion. Finally, in Section 4, we gather all the proofs of the main results and of the illustrated examples:
Subsection 4.1 contains auxiliary results on negative power variations which play an important role in
the proofs in Subsection 4.2 of the main results and in the proofs in Subsection 4.3 of the results of four
examples.
2
Main results
Let us recall the definition of a H−sssi process and an α− stable process (see e.g., [21]): A real-valued
process X
(d)
• is H-self-similar (H-ss) if for all a > 0, {X(at), t ∈ R} = aH {X(t), t ∈ R},
(d)
• has stationary increments (si) if, for all s ∈ R, {X(t + s) − X(s), t ∈ R} = {X(t) − X(0), t ∈ R}
(d)
where = stands for equality of finite dimensional distributions. A random variable X is said to have
a symmetric α-stable distribution (SαS) if there are parameters α ∈ (0, 2] and σ > 0 such that its
characteristic function has the form:
EeiθX = exp (−σ α | θ |α ) .
When σ = 1, a SαS is said to be standard. Let X be a H-sssi, SαS random process with 0 < α ≤ 2.
Let L ≥ 1, K ≥ 1 be fixed integers, a = (a0 , . . . , aK ) be a finite sequence with exactly L vanishing first
moments, that is for all q ∈ {0, . . . , L}, one has
K
X
kq ak = 0,
k=0
K
X
k=0
kL+1 ak 6= 0
(1)
with convention 00 = 1. For example, here we can choose K = L + 1 and
(L + 1)!
.
k!(L + 1 − k)!
ak = (−1)L+1−k
(2)
The increments of X with respect to the sequence a are defined by
△p,n X =
K
X
ak X(
k=0
k+p
).
n
(3)
We define now an estimator of H. Let β ∈ R, − 12 < β < 0, we set
Vn (β) =
n−K
X
1
|△p,n X|β ,
n − K + 1 p=0
Wn (β) = nβH Vn (β).
(4)
(5)
Notice that Vn (β) is the empirical mean of order β and Wn (β) is expected to converge to its mean. The
estimator of H is defined by
bn =
H
Vn/2 (β)
1
· log2
.
β
Vn (β)
(6)
We are now in position to define an estimator of α. We define first auxiliary functions ψu,v , hu,v , ϕu,v
before introducing the estimator of α, where u > v > 0.
Let ψu,v : R+ × R+ → R be the function defined by
ψu,v (x, y) = −v ln x + u ln y + C(u, v),
3
(7)
v
1−u
u
1−v
where C(u, v) = u−v
2 ln(π) + u ln Γ(1 + 2 ) + v ln Γ( 2 ) − v ln Γ(1 + 2 ) − u ln Γ( 2 ) .
Let hu,v : (0, +∞) → (−∞, 0) be the function defined by
u
v
hu,v (x) = u ln Γ(1 + ) − v ln Γ(1 + )
x
x
(8)
We will prove later that hu,v is bijective. Let ϕu,v : R → [0, +∞) be the function defined by
ϕu,v (x) =
(
0
h−1
u,v (x)
if x ≥ 0
if x < 0
(9)
where hu,v is defined as in (8).
Let β1 , β2 be in R such that −1/2 < β1 < β2 < 0. The estimator of α is defined by
α̂n = ϕ−β1 ,−β2 (ψ−β1 ,−β2 (Wn (β1 ), Wn (β2 ))) ,
(10)
where ψu,v , ϕu,v are defined as in (7) and (9), respectively.
With β ∈ (− 12 , 0) fixed, we will make the following assumption: There exist a sequence {bn , n ∈ N} and
a constant C such that lim bn = 0, bn/2 = O(bn ) and
n→+∞
lim sup
n→+∞
1
nb2n
X
k∈Z,|k|≤n
|cov(|△k,1 X|β , |△0,1 X|β )| ≤ C 2 .
(11)
Remark 2.1. The assumption (11) is important to prove the consistency of the estimators of the selfsimilarity and the stability indices. We will see its role in the main theorem below.
Now we are in position to present our main results for the estimation of H and α, based on the
assumption (11).
Theorem 2.1. Let X be a H-sssi, SαS random process that satisfies assumption (11). Also, let
b n , α̂n be defined as in (6) and (10), respectively.
β, β1 , β2 ∈ R, − 12 < β < 0, − 21 < β1 < β2 < 0 and H
Then as n → +∞, one has
P
P
bn −
H
→ H, α̂n −
→ α,
b n − H = OP (bn ), α̂n − α = OP (bn ), where OP is defined by:
moreover H
•Xn = OP (1) iff for all ǫ > 0, there exists M > 0 such that sup P(|Xn | > M ) < ǫ,
n
•Yn = OP (an ) means Yn = an Xn with Xn = OP (1).
See Subsection 4.2 for the proof of Theorem 2.1.
3
Examples
In this section, we study four classical examples: fractional Brownian motion, SαS-stable Lévy motion,
well-balanced linear fractional stable motion, Takenaka’s process. For these, we will show in Section 4
that (11) is valid, so that the conclusion of Theorem 2.1 holds. We precise this theorem by providing the
rate of convergence defined in (11) and a central limit theorem for the first two cases.
3.1
Fractional Brownian motion
Definition 3.1. Fractional Brownian motion
Fractional Brownian motion is a centered Gaussian process with covariance given by
EX(t)X(s) =
EX(1)2
{|s|2H + |t|2H − |s − t|2H }.
2
4
Fractional Brownian motion is a H-sssi 2-stable process (see, e.g., [9], p. 59). We will prove that the
condition (11) is satisfied with bn = n−1/2 , then the results in Theorem 2.1 are obtained. Moreover, we
can obtain the asymptotic normality of the estimators of the self-similarity index H and the stability
index α = 2.
Let X be a H fractional Brownian motion with H ∈ (0, 1). We first present the variances Ξ1 , Σ1 for the
limit distributions of the central limit theorems for the estimators of H and α.
We will mimic the Breuer-Major’s theorem (see e.g., Theorem 7.2.4 in [19]) to define these variances. For
β ∈ R, −1/2 < β < 0, let us introduce the following function
fβ (x) =
△ X
where Z0 = √ 0,1
var△0,1 X
q
β
var△0,1 X (|x|β − E|Z0 |β ),
(12)
.
Following Proposition 5.1 in Appendix, we can write fβ in terms of Hermite polynomials in a unique way
X
fβ (x) =
fβ,q Hq (x),
(13)
q≥d
where d is the Hermite rank of fβ and d ≥ 2,
ρ(r) =
P
2 < +∞. Let
q!fβ,q
q≥d
K
P
ap ap′ |r + p − p′ |2H
p,p′ =0
K
P
p,p′ =0
ρ1 (r) =
K
P
p,p′ =0
2H
ap ap′ |r + p − 2p′ |2H
K
P
ap a |p −
p′
p,p′ =0
P
q≥d
P
Γ1 =
q≥d
2
q!fβ,q
2
q!fβ,q
and φ : R+ × R+ → R be defined by
P
r∈Z
q≥d
ρq1 (r)
φ(x, y) =
Then Ξ1 is defined by
P
ρq (r)
r∈Z
P
,
(14)
ap ap′ |p − p′ |2H
2
P
2
q!fβ,q
q≥d
,
(15)
p′ |2H
P
r∈Z
2
q!fβ,q
P
ρq1 (r)
(16)
ρq (r).
r∈Z
x
1
log2 .
β
y
(17)
Ξ1 = φ′ (x0 , y0 )Γ1 φ′ (x0 , y0 )t ,
(18)
(x0 , y0 ) = (E|△0,1 X|β , E|△0,1 X|β ).
(19)
where
To define Σ1 , let −1/2 < β1 < β2 < 0, following Proposition 5.1 in Appendix, we can write fβ1 , fβ2 in
terms of Hermite polynomials in a unique way
fβ1 (x) =
X
fβ1,q Hq (x), fβ2 (x) =
X
fβ2 ,q Hq (x)
q≥d1
q≥d1
where d1 is the minimum of the Hermite ranks of fβ1 and fβ2 , d1 ≥ 2 and
X
q≥d
q!fβ21 ,q < +∞,
X
q≥d
5
q!fβ22 ,q < +∞.
(20)
Let
Σ1 = ∇ϕ−β1 ,−β2 ◦ψ−β1 ,−β2 (x1 , y1 )Γ2 ∇ϕ−β1 ,−β2 ◦ψ−β1 ,−β2 (x1 , y1 )t
(21)
where ψu,v , ϕu,v are defined by (7), (9) respectively, ∇ is the differential operator and
(x1 , y1 ) = (E|△0,1 X|β1 , E|△0,1 X|β2 ),
Γ2 =
σβ21 =
+∞
X
q=d1
q!fβ21 ,q
X
ρ(k)q , σβ22 =
k∈Z
+∞
X
(22)
!
σβ21 ρβ1 ,β2
,
ρβ1 ,β2 σβ22
q!fβ22,q
q=d1
X
ρ(k)q , ρβ1 ,β2 =
k∈Z
(23)
+∞
X
q!fβ1,q fβ2 ,q
q=d1
X
ρ(k)q .
(24)
k∈Z
We can now state the following theorem, which precises the results for the estimation of H and α in the
case of fractional Brownian motion.
Theorem 3.1. Let X be a fractional Brownian motion. Then
a)
b n − H = OP (n−1/2 ), α̂n − 2 = OP (n−1/2 ),
H
b)
√ b
√
(d)
(d)
n(Hn − H) −−→ N1 (0, Ξ1 )), n(α̂n − 2) −−→ N1 (0, Σ1 )
as n → +∞, where Ξ1 , Σ1 are defined by (18) and (21), respectively.
See Subsection 4.3 for the proof of Theorem 3.1.
3.2
SαS-stable Lévy motion
Definition 3.2. SαS-stable Lévy motion
A stochastic process {X(t), t ≥ 0} is called (standard) SαS-stable Lévy motion if X(0) = 0 (a.s.), X
has independent increments and, for all 0 ≤ s < t < ∞ and for some 0 < α ≤ 2, X(t) − X(s) is a SαS
random variable with characteristic function given by
Eeiθ(X(t)−X(s)) = exp (−(t − s)|θ|α ) .
The condition (11) is proved to be satisfied with bn = n−1/2 , then the results in Theorem 2.1 are
ascertained. Similar to the case of fractional Brownian motion, we obtain the asymptotic normality of
H and α.
The variances Ξ2 , Σ2 for the limit distributions of the central limit theorems for the estimators of H and
α are defined as follows.
Let X be a SαS−stable Lévy motion, we define the variance for the limit distribution of the central limit
theorem for the estimator of H by
Ξ2 = φ′ (x0 , y0 )Γ3 φ′ (x0 , y0 )t ,
(25)
where φ(x, y), (x0 , y0 ) are defined by (17), (19), respectively and
Γ3 =
!
σ12 σ1,2
,
σ1,2 σ22
6
(26)
σ12 = var|△0,1 X|β + cov(|△0,1 X|β , |△1,1 X|β )
+2
K−1
X
p=1
+2
cov(|△0,1 X|β , |△2p,1 X|β ) + cov(|△0,1 X|β , |△2p+1,1 X|β )
K−1
X
cov(|△1,1 X|β , |△2p,1 X|β ) + cov(|△1,1 X|β , |△2p+1,1 X|β ) ,
p=1
σ22 = 2 var|△0,1 X|β + 2
K−1
X
p=1
cov(|△0,1 X|β , |△p,1 X|β )
(27)
(28)
σ1,2 = 2βH cov(|△0,2 X|β , |△0,1 X|β ) + cov(|△1,2 X|β , |△0,1 X|β )
K−1
X
βH
+2
p=1
K−1
X
+ 2βH
p=1
cov(|△0,2 X|β , |△p,1 X|β ) + cov(|△1,2 X|β , |△p,1 X|β )
cov(|△0,1 X|β , |△2p,2 X|β ) + cov(|△0,1 X|β , |△2p+1,2 X|β )
(29)
The variance for the limit distribution of the central limit theorem for the estimator of α is defined by
Σ2 = ∇ϕ−β1 ,−β2 ◦ψ−β1 ,−β2 (x1 , y1 )Γ4 ∇ϕ−β1 ,−β2 ◦ψ−β1 ,−β2 (x1 , y1 )t
(30)
where ψu,v , ϕu,v , (x1 , y1 ) are defined as in (7), (9) and (22), respectively,
Γ4 =
σ12 = var|△0,1 X|β1 + 2
σ22 = var|△0,1 X|β2 + 2
K−1
X
k=1
K−1
X
k=1
σ1,2
!
σ12 σ1,2
,
σ1,2 σ22
(31)
cov(|△0,1 X|β1 , |△k,1 X|β1 ),
(32)
cov(|△0,1 X|β2 , |△k,1 X|β2 )
(33)
X
1 K−1
= cov(|△0,1 X| , |△0,1 X| ) +
cov(|△0,1 X|β1 , |△k,1 X|β2 ) + cov(|△0,1 X|β2 , |△k,1 X|β1 ) .
2 k=1
β1
β2
(34)
We now present the results on the asymptotic normality for the case of SαS-stable Lévy motion.
Theorem 3.2. Let X be a SαS-stable Lévy motion. Then
a)
b n − H = OP (n−1/2 ), α̂n − α = OP (n−1/2 )
H
b)
√
√
(d)
(d)
b n − H) −
n(H
−→ N1 (0, Ξ2 )), n(α̂n − α) −−→ N1 (0, Σ2 )
as n → +∞, where Ξ2 , Σ2 are defined by (25) and (30), respectively.
The proof of Theorem 3.2 is given in Subsection 4.3.
3.3
Well-balanced linear fractional stable motion
Definition 3.3. Well-balanced linear fractional stable motion
Let M be a SαS random measure, 0 < α ≤ 2, with Lebesgue control measure and consider
X(t) =
Z
+∞
−∞
(| t − x |H−1/α − | x |H−1/α )M (dx), −∞ < t < +∞
7
where 0 < H < 1, H 6= 1/α. The process X is called the well-balanced linear fractional stable motion.
Then X is a H-sssi process (Proposition 7.4.2, [21]).
Let
bn =
n−1/2
n
if H < L + 1 −
αH−(L+1)α
4
if H > L + 1 −
q
ln n
if H = L + 1 −
n
2
α
2
α
2
α.
(35)
It is clear that lim bn = 0 and bn/2 = O(bn ). We get the following results for the estimation of H and
n→+∞
α.
Theorem 3.3. Let {X(t)}t∈R be a well-balanced linear fractional stable motion with 0 < H < 1, H 6= 1/α
and 0 < α < 2. Then for every β ∈ (−1/2, 0), Theorem 2.1 is true with bn defined by (35).
See Subsection 4.3 for the proof of Theorem 3.3.
3.4
Takenaka’s processes
Definition 3.4. Takenaka’s process
Let M be a symmetric α− stable random measure (0 < α < 2) with control measure
m(dx, dr) = r ν−2 dxdr, (0 < ν < 1).
Let t ∈ R, set
Ct = {(x, r) ∈ R × R+ , |x − t| ≤ r}, St = Ct ∆C0
where ∆ denotes the symmetric difference between two sets.
Takenaka’s process is defined by
X(t) =
Z
1St (x, r)M (dx, dr).
(36)
R×R+
Following Theorem 4 in [25], the process X is ν/α−sssi. Let
bn = n
ν−1
2
.
(37)
We can now ascertain the following.
Theorem 3.4. Let {Xt , t ∈ R} be a Takenaka’s process defined by (36). Then for every β, β ∈ (−1/2, 0),
Theorem 2.1 is true with bn defined by (37).
The proof of Theorem 3.4 is given in Subsection 4.3.
4
Proofs
First, we give results on expectation of negative power variations of H-sssi, SαS random processes in
Subsection 4.1. Then we apply these results in Subsection 4.2 to the estimation of H and α, in order to
prove Theorem 2.1. Finally, we prove that Theorem 2.1 is true for four classical examples presented in
Section 3.
4.1
Negative power expectation and auxiliary results
Now we present some results on expectation of negative power variations of H-sssi, SαS random processes
proved by using theory of distribution. These results are the tools to prove assumptions (11) for four
examples in Section 3 and to prove the main result on the estimation for α.
8
4.1.1
Auxiliary results
We start with the following lemma which confirms the existence of the expectation of β-negative power
variation of a symmetric stable random variable when β ∈ C, Re(β) ∈ (−1, 0).
Lemma 4.1. Let X be a SαS random variable, β ∈ C, Re(β) ∈ (−1, 0), then |E|X|β | < +∞.
The proof of Lemma 4.1 is given in Subsection 4.1.2
The next two important results will be used to prove the condition (11) for our examples in next section.
Theorem 4.1 gives a way to determine the expectation of β-negative power variation of a symmetric stable
random variable whereas Theorem 4.2 helps to establish the inequality of (11) for illustrated examples.
Let (S, µ) be a measure space, h, g ∈ Lα (S, µ) and M be a symmetric α-stable random measure on S
with control measure µ, α ∈ (0, 2). Set
U=
Z
h(s)M (ds), V =
S
Z
g(s)M (ds).
(38)
S
Let
Cβ =
2β+1/2 Γ( β+1
2 )
(39)
Γ(− β2 )
where β ∈ C such that Re(β) ∈ (−1, 0).
Theorem 4.1. For β ∈ C, Re(β) ∈ (−1, 0), we have
1
E|U | = √
2π
β
Z
R
iU y
FT (y)Ee
Cβ
dy = √
2π
Z
R
EeiU y
dy
|y|β+1
(40)
in the sense of distributions, where U, V are defined by (38), T = |x|β and FT is Fourier transform of
T.
See Subsection 4.1.3 for the proof of Theorem 4.1.
Theorem 4.2. Assume that
||U ||αα =
[U, V ]2 =
Z
S
Z
S
|h(s)|α µ(ds) = 1, ||V ||αα =
Z
S
|g(s)|α µ(ds) = 1
|h(s)g(s)|α/2 ≤ η < 1,
where U, V are defined as in (38). Then for −1/2 < Re(β) < 0, we have
β
β
E|U | |V | =
Cβ Cβ Z
2π
R2
EeixU +iyV
|x|1+β |y|1+β
dxdy.
(41)
Moreover, there exists a constant C(η) such that
|cov(|U |β , |V |β )| ≤ C(η)
Z
S
|h(s)g(s)|α/2 ds.
(42)
The proof of Theorem 4.2 is given in Subsection 4.1.4.
The following two lemmas follow from Theorem 4.1 in which Lemma 4.2 provides an important formula
to construct the estimator for α.
9
Lemma 4.2. Let X be a standard SαS variable with 0 < α ≤ 2 and β ∈ C, −1 < Re(β) < 0, then
β
2β Γ( β+1
2 )Γ(1 − α )
.
√
πΓ(1 − β2 )
E|X|β =
(43)
See Subsection 4.1.5 for the proof of Lemma 4.2.
Lemma 4.3. Let X be a SαS process where 0 < α ≤ 2, β ∈ C, − 21 < Re(β) < 0, then
E|△0,1 X|β 6= 0.
See Subsection 4.1.6 for the proof of Lemma 4.3. Now we will give the proofs for the latter results.
4.1.2
Proof of Lemma 4.1
Since X is a SαS -stable random variable, X has a density function f (x) that is even and continuous on
R. We first consider the case β ∈ R and −1 < β < 0.
For −1 < β < 0, we can write:
E|X|β =
Z
R
We have
A=
Z
|x|≤1
|x|β f (x)dx =
Z
|x|β f (x)dx +
|x|≤1
β
|x| f (x)dx ≤ sup |f (x)|
|x|≤1
Z
Z
|x|≥1
|x|β f (x)dx := A + B.
β
|x|≤1
|x| dx < +∞, B = 2
Z∞
1
xβ f (x)dx ≤ 2.
It follows that E|X|β < +∞. For β ∈ C, −1 < Re(β) < 0, we have
E|X|β =
Z
R
|x|a+ib f (x)dx ≤
Z
R
|x|Re(β) f (x)dx < +∞.
Then we obtain the conclusion.
4.1.3
Proof of Theorem 4.1
To prove Theorem 4.1, we start with the following lemma.
Lemma 4.4. For all x ∈ R, β ∈ C, −1 < Re(β) < 0, let T (x) = |x|β . Then T has Fourier transform
defined by
Cβ
(44)
FT (y) = β+1
|y|
in the sense of distributions, where Cβ is defined as in (39).
Proof. For β ∈ C, −1 < Re(β) < 0, following example 5, §7, chapter VII of [22], then T is a distribution
and it has Fourier transform FT (y) = C|y|−(β+1) , where C is a constant. We will find C using function
2
k(x) = e−x /2 . Since T ∈ L1loc (R) and k ∈ S(R), in the sense of distributions, we have hFT, ki = hT, Fki.
On the other hand,
1
Fk(y) = √
2π
Z
e−ixy e−x
2 /2
dx = e−y
2 /2
R
then
Z
R
β −x2 /2
|x| e
dx =
Z
C|y|−(β+1) e−y
R
10
2 /2
dy.
By taking the change of variable, we obtain that
Z
R
β −x2 /2
|x| e
dx = 2
β+1
2
β+1
),
Γ(
2
It follows that
C=
From Lemma 4.4, we have FT (y) =
Z
|y|−(β+1) e−y
R
2β+1/2 Γ( β+1
2 )
Cβ
,
|y|1+β
Γ(− β2 )
2 /2
β
dy = 2−β/2 Γ(− ).
2
= Cβ .
Γ( u+1 )
where f is the density function of U and Cu = 2u+1/2 Γ(−2u ) .
2
R
Let ϕ be a non-negative, even function, ϕ ∈ C0∞ (R), suppϕ ⊂ [−1, 1], ϕ(y)dy = 1. Set ϕǫ (x) =
R
ϕ(x/ǫ)
ǫ ,
we will prove that gǫ = F −1 f ∗ ϕǫ ∈ S(R).
Indeed, let χ(x) be a function in C0∞ (R) such that χ(x) = 1 for |x| ≤ 1 and χ(x) = 0 for |x| ≥ 2.
We can write the characteristic function corresponding with the density function f as
√
α
α
e−σ |x| = 2πF −1 f (x) := g(x) = χ(x)g(x) + (1 − χ(x))g(x) := g1 (x) + g2 (x)
and
g ∗ ϕǫ (x) = g1 ∗ ϕǫ (x) + g2 ∗ ϕǫ (x).
It is clearly that g1 ∈ L1 (R), g1 has compact support, ϕǫ ∈ C0∞ (R), so g1 ∗ ϕǫ ∈ S(R).
We also have g2 ∗ ϕ ∈ S(R) since g2 and ϕǫ (x) are in S(R).
Then we get gǫ ∈ S(R).
We have
√
Fgǫ (x) = 2πf (x)Fϕǫ (x).
Since
hT, Fgǫ i = hFT, gǫ i,
we obtain
Z √
2π|x|β f (x)Fϕǫ (x)dx =
Z
R
R
=
Z
R
FT (y)F −1 f ∗ ϕǫ (y)dy
(45)
F −1 f (y)FT ∗ ϕǫ (y)dy,
(46)
Here we used Fubini’s theorem since FT, F −1 f ∈ L1loc (R), ϕǫ ∈ C0∞ (R) and ϕǫ is an even function.
Now we will find the limits of two sides of the equation (46) when ǫ → 0. We first consider the left hand
side of (46). One has
lim Fϕǫ (x) = lim
ǫ→0
ǫ→0
Z
R
1
√ e−itx ϕǫ (t)dt = lim
ǫ→0
2π
Z
R
ϕ(t/ǫ)
1
√ e−itx
dt = lim
ǫ→0
ǫ
2π
R
Z
R
1
√ e−iǫux ϕ(u)du.
2π
For x, u ∈ R, e−iǫux ϕ(u) → ϕ(u) when ǫ → 0, and |e−iǫux ϕ(u)| = ϕ(u), ϕ(u)du = 1, following Lebesgue
R
dominated convergence theorem, one gets
lim Fϕǫ (x) =
ǫ→0
Z
R
1
1
√ ϕ(u)du = √ .
2π
2π
11
Therefore, for x 6= 0,
√
2π|x|β f (x)Fϕǫ (x) →
1
|x|β f (x)Fϕǫ (x) = √ |x|Re(β) f (x)|
2π
1
≤ √ |x|Re(β) f (x)
2π
√
2π|x|β f (x) pointwise when ǫ → 0. We have
Z
e−itx
Z
1
ϕ(u)du = √ |x|Re(β) f (x).
2π
R
R
Moreover, applying Lemma 4.1, it follows that
R
R
ϕ(t/ǫ)
1
dt| = √ |x|e(β) f (x)
ǫ
2π
Z
e−iǫux ϕ(u)du
R
|x|Re(β) f (x)dx < ∞. Thus applying Lebesgue dominated
convergence theorem, the left hand side of (46) converges to
R
R
|x|Re(β) f (x)dx.
Turning back to the right hand side of (46), since FT is continuous at y 6= 0 and FT ∈ L1loc (R), we get
lim FT ∗ ϕǫ (y) = FT (y)
ǫ→0
for y ∈ R∗ . It follows that
lim F −1 f (y)FT ∗ ϕǫ (y) = F −1 f (y)FT (y)
ǫ→0
pointwise almost everywhere. We have the following inequality on FT and ϕǫ .
Lemma 4.5. There exists a constant C > 0 such that for all ǫ > 0, x 6= 0, we have
|FT | ∗ ϕǫ (x) ≤ C|FT |(x).
Proof. Since FT and ϕǫ are even functions, we just need to prove this lemma for x > 0. From the fact
that ϕ has compact support, then there exists a constant C such that for all x > 0, ϕ(x) ≤ C 1[−1,1] (x).
We consider first the case x > 2ǫ. One has
|FT | ∗ ϕǫ (x) =
Z
C
|FT |(y)ϕǫ (x − y)dy ≤
ǫ
R
x+ǫ
Z
x−ǫ
|FT |(y)dy
≤ 2C|FT |(x − ǫ) ≤ 2C|FT |(x/2) = C1 |FT |(x).
If x ≤ 2ǫ, then
C
|FT | ∗ ϕǫ (x) ≤
ǫ
=
x+ǫ
Z
x−ǫ
2C
ǫ
Z3ǫ
0
C
|FT |(y)dy ≤
ǫ
|Fβ |
dy
1+Re(β)
|y|
=
Z3ǫ
−3ǫ
|FT |(y)dy
1
C1
≤ C2 |FT |(x).
ǫ (3ǫ)Re(β)
Applying Lemma 4.5, then we deduce that
|F −1 f (y)FT ∗ ϕǫ (y)| ≤ C|F −1 f (y)||FT |(y)
almost everywhere. But
Z
R
|F −1 f (y)||FT |(y)dy =
Z
R
α
|Cβ |
|Cβ |
e−|y|
√
dy = √
1+Re(β)
2π |y|
2π
Z
R
α
e−|y|
dy < ∞
|y|1+Re(β)
since Re(β) ∈ (−1,R 0). Applying Lebesgue dominated convergence theorem again, the right hand side of
(46) converges to F −1 f (y)FT (y)dy. So we get (40).
R
12
4.1.4
Proof of Theorem 4.2
Let χ be in C0∞ (R), χ ≥ 0, χ(x) = 1 if x ∈ [−1, 1], suppχ ∈ [−2, 2]. For ǫ > 0, we define
φǫ (x) = (1 − χ(x/ǫ))χ(ǫx).
Let µ be the distribution of random vector (U, V ), then µ is a probability measure on R2 .
Let T1 (x) = |x|β , T2 (x) = |y|β . Following Lemma 4.4, T1 , T2 are distributions and have Fourier transforms
FT1 (y) =
Cβ
Cβ
,
,
FT
(x)
=
2
|y|1+β
|x|1+β
respectively, in the sense of distributions, where Cu = 2u+1/2
Γ( u+1
)
2
.
Γ( −u
)
2
Set F1ǫ (x) = T1 (x)φǫ (x), F2ǫ (y) = T2 (y)φǫ (y). It is clearly that F1ǫ (x) ∈ S(R), F2ǫ (y) ∈ S(R).
Then F1ǫ ⊗ F2ǫ (x, y) ∈ S(R2 ). It follows that
Z
R2
Z
F1ǫ ⊗ F2ǫ (x, y)dµ(x, y) =
R2
F −1 (dµ)(x, y)F(F1ǫ ⊗ F2ǫ )(x, y)dxdy.
(47)
Now we consider the right-hand side of (47). We have
F(F1ǫ ⊗ F2ǫ )(x, y) = FF1ǫ ⊗ FF2ǫ (x, y).
We can write
x
F1ǫ (x) = T1 (x)χ(ǫx) − T1 (x)χ(ǫx)χ( ).
ǫ
Set ψ(x) = Fχ(x). One has
1
1
1
1
1
FT1 ∗ ψǫ ∗ ψ1/ǫ = FT1 ∗ ( √ ψǫ ) − √ FT1 ∗ ( √ ψǫ ) ∗ ψ1/ǫ .
FF1ǫ (x) = √ FT1 ∗ ψǫ −
2π
2π
2π
2π
2π
We will use the following lemma.
Lemma 4.6. Let ψ be a function in the Schwartz class, T (t) =| t |β where Re(β) ∈ (−1, 0). Then for
all x 6= 0, there exists a constant C > 0 such that
| T ∗ ψ(x) |≤ C | T (x) | .
Proof. We denote C a running constant which may change from an occurrence to another occurrence.
For ǫ > 0, set
1 x
ψǫ (x) = ψ( ).
ǫ ǫ
We first prove that there exists a constant C > 0 such that for all ǫ > 0,
| T ∗ ψǫ (x) |≤ C sup
a>0
1
2a
Let k(y) =| T (x − y) |, Iǫ = T ∗ ψǫ (x).
Z
x+a
x−a
By taking the change of variable u = yǫ , we obtain that Iǫ =
has
F (x) =
1
ǫ
Z
0
ǫx
| T (t) | dt.
R
k(ǫu)ψ(u)du. Set F (x) =
R
k(t)dt, F ′ (x) = k(ǫx).
13
Rx
0
k(ǫu)du, one
Combining with the fact that lim F (x)ψ(x) = 0 and F (0) = 0, we deduce that
x→∞
+∞
Z
k(ǫu)ψ(u)du =
Z
+∞
′
F (u)ψ(u)du = −
0
0
=−
Since k(u) ≥ 0, it follows that
1
ǫu
Zǫu
0
+∞
Z
F (u)ψ ′ (u)du
0
1 Zǫu
k(t)dt uψ ′ (u)du.
ǫu
+∞
Z
0
0
k(t)dt ≤ sup
a>0
1
a
Z+a
k(t)dt.
−a
We also have ψ is a function in the Schwartz class, then
|
+∞
Z
0
k(ǫu)ψ(u)du| ≤ |
+∞
Z
0
1
uψ (u)du| sup
a>0 a
′
We can get a similar bound for the integral |
R0
Z
+a
−a
1
k(t)dt = C sup
a>0 2a
Z+a
k(t)dt.
−a
k(ǫu)ψ(u)du|. Therefore we obtain
−∞
1
|Iǫ | ≤ C sup
a>0 2a
x+a
Z
x−a
| T (t) | dt.
Taking ǫ = 1, it follows that
1
| T ∗ ψ(x) |≤ C sup
a>0 2a
x+a
Z
x−a
| T (t) | dt.
Now we will prove that there exists a constant C > 0 such that for all a > 0, then
1
2a
x+a
Z
x−a
| T (t) | dt ≤ C|T (x)|.
We first consider the case x > 0.
If x > 2a, then 0 < x2 < x − a < x + a and T (t) decreases over [x − a, x + a]. We get
1
2a
x+a
Z
x−a
| T (t) | dt ≤
≤
1
(x + a − (x − a)) (x − a)Re(β)
2a
(x/2)Re(β)
= C | T (x) | .
2
If 0 < x ≤ 2a < 3a, then
1
2a
x+a
Z
x−a
1
| T (t) | ≤
2a
=
1
a
Z3a
−3a
Z
3a
| T (t) | dt
tRe(β) dt =
0
(3a)1+Re(β)
a(1 + Re(β))
≤ C(3a)Re(β) ≤ CxRe(β) = C|T (x)|.
14
For the case x < 0, if x ≤ −2a, then x − a < x + a < x/2 < 0, we obtain
1
2a
x+a
Z
x−a
(x + a − (x − a))|x + a|Re(β)
2a
| T (t) | ≤
≤| x/2 |Re(β) = C | T (x) | .
If −2a < x < 0, then −3a < x − a < x + a < 3a, one gets
1
2a
x+a
Z
x−a
1
| T (t) | ≤
2a
=
1
a
Z3a
−3a
Z3a
0
| T (t) | dt
tRe(β) ≤ C | x |Re(β) = C | T (x) | .
One can therefore obtain the conclusion.
Since ψǫ ∗ ψ1/ǫ ∈ S(R), following Lemma 4.6, we have
|FT1 ∗ (ψǫ ∗ ψ1/ǫ )(x)| ≤ C|FT1 (x)|.
Then, there exists a constant C > 0 such that
|FF1ǫ (x)| ≤ C|FT1 (x)|.
In a similar way, we also get |FF2ǫ (y)| ≤ C|FT2 (y)|. It follows that
|F −1 (dµ)(x, y)F(F1ǫ ⊗ F2ǫ )(x, y)| ≤ C|F −1 (dµ)(x, y)||FT1 (x)FT2 (y)|.
Let us recall that
R
R
ψ(t)
√
dt
2π
= χ(0) = 1. We will use the two following lemmas to get
lim FF1ǫ (x) = FT1 (x), lim FF2ǫ (y) = FT2 (y).
ǫ→0
ǫ→0
(48)
Lemma 4.7. Let T (x) =| x |β where Re(β) ∈ (−1, 0), ψ be a function in Schwartz class. Then almost
everywhere,
lim T ∗ ψ1/ǫ (x) = 0.
ǫ→0
where ψ1/ǫ (x) = ǫψ(ǫx).
Proof. Let x ∈ R, x 6= 0, we have
T ∗ ψ1/ǫ (x) =
Z
R
T (y)ǫψ(ǫ(x − y))dy =
Z
R
| y |β ǫψ(ǫ(x − y))dy.
By taking the change of variable t = ǫy, one gets
| T ∗ ψ1/ǫ (x) |≤
Z
R
| t/ǫ |Re(β) | ψ(ǫx − t) | dt = ǫ−Re(β)
Z
R
| t |Re(β) | ψ(ǫx − t) | dt.
We write
Z
R
Re(β)
|t|
| ψ(ǫx − t) | dt =
Z1
−1
Re(β)
|t|
| ψ(ǫx − t) | dt +
15
Z
|t|≥1
| t |Re(β) | ψ(ǫx − t) | dt := I1 + I2 .
We consider I1 and I2 . Since ψ is a function in Schwartz class, one gets || ψ ||∞ < ∞, || ψ ||1 < ∞. Then
I1 ≤ 2 || ψ ||∞
Z
I2 =
|t|≥1
Z1
tRe(β) dt = C < +∞.
0
| t |Re(β) | ψ(ǫx − t) | dt ≤
Z
|t|≥1
| ψ(ǫx − t) | dt ≤|| ψ ||1 = C < +∞.
Since then | T ∗ ψ1/ǫ (x) |≤ Cǫ−Re(β) → 0 as ǫ → 0. It follows that T ∗ ψ1/ǫ (x) → 0 almost everywhere as
ǫ → 0.
Lemma 4.8. Let ψ be a function in Schwartz class such that
R
R
ψ(t)dt = 1, T (t) =| t |β where Re(β) ∈
(−1, 0). Then we have lim T ∗ ψǫ (x) = T (x) almost everywhere, where ψǫ (x) =
ǫ→0
ψ(x/ǫ)
ǫ .
Proof. Let x ∈ R, x 6= 0, we consider I = T ∗ ψǫ (x) − T (x). Let us recall that
Z
Z
ψǫ (t)dt =
R
Therefore I =
R
R
ψ(t/ǫ)
dt =
ǫ
R
Z
ψ(u)du = 1.
R
{T (x − y) − T (x)} 1ǫ ψ(y/ǫ)dy.
Let θ > 0 be a constant. There exists 0 < δ <| x | such that for | y |≤ δ, we have | T (x−y)−T (x) |≤
Then
| I |≤
Z
|y|≤δ
| T (x − y) − T (x) |
θ
I1 ≤
2ǫ || ψ ||1
Z
|y|≤δ
| ψ(y/ǫ) |
dy +
ǫ
θ
| ψ(y/ǫ) | dy =
2ǫ || ψ ||1
Z
| T (x − y) − T (x) |
|y|≥δ
Z
|u|≤ǫδ
θ
2||ψ||1 .
| ψ(y/ǫ) |
dy := I1 + I2 .
ǫ
θ
ǫ | ψ(u) | du ≤
2ǫ || ψ ||1
Z
R
ǫ | ψ(u) | du =
θ
.
2
Now we consider I2 . Since ψ is a function in Schwartz class, there exists a constant C > 0 such that for
| t |≥ 1 then | ψ(t) |≤ tC2 .
We choose ǫ > 0 such that δǫ ≥ 1. By taking the change of variable t = y/ǫ, we get
Z
I2 =
|t|≥ δǫ
=
| T (x − ǫt) − T (x) || ψ(t) | dt ≤
Z
|t|≥ δǫ ,|x−ǫt|≤1
| T (x − ǫt) || ψ(t) | dt +
Z
|t|≥ δǫ
| T (x − ǫt) || ψ(t) | dt +
Z
|t|≥ δǫ ,|x−ǫt|≥1
Z
|t|≥ δǫ
| T (x) || ψ(t) | dt
| T (x − ǫt) || ψ(t) | dt +
Z
|t|≥ δǫ
| T (x) || ψ(t) | dt
:= J1 + J2 + J3 .
We have
J1 =
Z
|t|≥ δǫ ,|x−ǫt|≤1
| T (x − ǫt) || ψ(t) | dt ≤ C
Z
|x−ǫt|≤1
| T (x − ǫt) | (ǫ/δ)2 dt.
By taking the change of variable u = ǫt − x, one gets
Cǫ2
J1 ≤ 2
δ
Z
|u|≤1
| T (u) |
du = C1 ǫ.
ǫ
16
Here C1 is a constant depending on δ.
Let us consider J2 . Since | T (t) |=| t |Re(β) and Re(β) ∈ (−1, 0), if | x − ǫt |≥ 1 we get | T (x − ǫt) |≤ 1.
Moreover δ/ǫ ≥ 1, it follows that
Z
C
ǫ
J2 ≤
dt = C2
2
t
δ
|t|≥δ/ǫ
where C2 is a constant depending on δ. Similarly, since δ/ǫ ≥ 1, we get
J3 ≤| T (x) |
Z
|t|≥δ/ǫ
C
dt = C3 ǫ
t2
where C3 is a constant depending on x, δ. So we get I2 ≤ Cǫ where C is a constant depending on x, δ.
We can choose ǫ small enough to get I2 ≤ 2θ .
Then for all θ > 0, there exists ǫ0 such that for all 0 < ǫ < ǫ0 , we have | I |≤ θ. Therefore we get the
conclusion.
From (48), one gets
lim F(F1ǫ ⊗ F2ǫ )(x, y) = FT1 (x)FT2 (y) =
ǫ→0
Cβ Cβ
| x |1+β | y |1+β
.
ixU +iyV
. We use Theorem 4.1, Lemma 4.1 and the following lemma to deduce
Moreover F −1 (dµ)(x, y) = Ee 2π
that
Z
Z
|F −1 (dµ)(x, y)|
EeixU +iyV
dxdy =
dxdy < +∞.
| x |1+Re(β) | y |1+Re(β)
2π | x |1+Re(β) | y |1+Re(β)
2
2
R
R
Lemma 4.9. Set
ixU +iyV
MU,V (x, y) = Ee
[U, V ]2 =
Z
S
ixU
− Ee
iyV
Ee
,I =
Z
| MU,V (x, y) |
dxdy.
| xy |1+Re(β)
R2
|h(s)g(s)|α/2 ≤ η < 1,
where U, V are defined as in Theorem 4.2. Then I ≤ C(η)[U, V ]2 < ∞, where the constant C(η) depends
on η.
Proof. We just need to consider the integral only over (0, +∞) × (0, +∞). We divide this domain into
four regions (0, 1) × (0, 1), (0, 1) × (1, +∞), (1, +∞) × (0, 1), (1, +∞) × (1, +∞) and let I1 , I2 , I3 , I4 be the
integrals over those domains, respectively.
Over (0, 1) × (0, 1), by using inequality (3.4) in [20] we get
I1 ≤
Z1 Z1
0 0
| MU,V (x, y) |
dxdy ≤ 2
| xy |1+Re(β)
Z1 Z1
(xy)α/2−1−Re(β) dxdy[U, V ]2 = C[U, V ]2 .
0 0
Over (1, +∞) × (1, +∞), by using inequality (3.6) in [20] and assumptions
|| U
||αα =
1, || V
||αα =
1, [U, V ]2 =
Z
S
| f (s)g(s) |α/2 ≤ η < 1,
we can bound the integral over this domain by
I2 ≤ 2
+∞
Z +∞
Z
α/2
(xy)α/2−1−Re(β) e−2(1−η)(xy)
1
1
17
dxdy[U, V ]2 .
α/2
up to a constant depending on η by (xy)−p for arbitrarily large p > 0.
Here we can bound e−2(1−η)(xy)
So I2 ≤ C(η)[U, V ]2 . Over (0, 1) × (1, +∞), by using inequality (3.5) in [20] we obtain a bound
I3 ≤ 2
Z1 +∞
Z
α/2 −y α/2 )2
≤2
Z
Z1 +∞
α/2 −1)2
(xy)α/2−1−Re(β) e−(x
0
1
(xy)α/2−1−Re(β) e−(y
0
1
+∞
Z
(xy)α/2−1−Re(β) e−(y
≤C
1
+∞
Z
y α/2−1−Re(β) e−(y
=C
dxdy[U, V ]2
α/2 −1)2
α/2 −1)2
dxdy[U, V ]2
dxdy[U, V ]2
dy[U, V ]2 .
1
Since
+∞
R
1
y α/2−1−Re(β) e−(y
α/2 −1)2
dy < +∞, we get I3 ≤ C[U, V ]2 .
A similar bound holds for I4 . Then we have the conclusion.
By Lebesgue dominated convergence theorem, as ǫ → 0, the right-hand side of (47) converges to
Cβ Cβ Z
2π
R2
EeixU +iyV
| x |1+β | y |1+β
dxdy.
Now we consider the left-hand side of (47). Since lim φǫ (x) = 1 for all x ∈ R, it follows
ǫ→0
lim F1ǫ (x)F2ǫ (y) =| x |β | y |β
ǫ→0
for all x 6= 0, y 6= 0.
It is clear that | F1ǫ (x) |≤ C | x |Re(β) , | F2ǫ (y) |≤ C | y |Re(β) . Moreover
Z
R2
Re(β)
|x|
Re(β)
|y|
β
β
dµ(x, y) = E|U | |V |
≤
q
E|U |2β E|V |2β < +∞
since Re(β) ∈ (−1/2, 0).
We can therefore apply Lebesgue dominated convergence theorem for the left-hand side of (47). As ǫ → 0,
it converges to
Z
R2
| x |β | y |β dµ(x, y) = E | U |β | V |β .
This proves the result (41). Now we prove (42).
Following Theorem 4.1 and (41), for −1/2 < Re(β) < 0, we get
Cβ
E|U | = √
2π
β
E | U |β | V |β =
Z
R
Cβ Cβ Z
2π
Cβ Z EeiyV
EeixU
β
√
dy
dx, E | V | =
| x |1+β
2π
| y |1+β
R2
R
EeixU +iyV
| x |1+β | y |1+β
18
dxdy.
Then
cov(| U |β , | V |β ) = E | U |β | V |β −E | U |β E | V |β = Cβ Cβ
≤| Cβ Cβ |
Z
EeixU +iyV
−
EeixU +iyV − EeixU EeiyV
R2
| x |1+β | y |1+β
dxdy
EeixU EeiyV
dxdy.
| xy |1+Re(β)
R2
Z
Applying Lemma 4.9, we obtain (42).
4.1.5
Proof of Lemma 4.2
For the case α = 2, let Y be a standard S2S variable. Then for −1 < Re(β) < 0, we have
1
E|Y | = √
2 π
β
+∞
Z
−∞
β
2
− x4
|x| e
+∞
Z
1
dx = √
π
xβ e−x
2 /4
0
2β
β+1
dx = √ · Γ(
).
π
2
Let us now consider the case α 6= 2. Following (44) and Theorem 4.1, we have
Cβ
E|X| = √
2π
β
Z
R
EeiXy
Cβ
dy = √
β+1
|y|
2π
Z
α
e−|y|
2Cβ
dy = √
β+1
|y|
2π
R
+∞
Z
0
α
e−y
dy.
y β+1
By making the change of variable y α = t, then
2Cβ
E|X| = √
2π
β
+∞
Z
t
0
2Cβ
e dt = √
2π
−β/α−1 −t
+∞
Z
t
√
2Cβ Γ(−β/α)
√
e dt =
.
α π
−β/α−1 −t
0
Since Γ(x + 1) = xΓ(x), one gets
E | X |β =
4.1.6
2β Γ( β+1
)Γ(1 − αβ )
.
√ 2
πΓ(1 − β2 )
Proof of Lemma 4.3
From the fact that X is a SαS process, one can write
△0,1 X =
K
X
(d)
ak X(k) = σY
k=0
where σ > 0 and Y is a standard SαS random variable. Then E | △0,1 X |β = σ β E | Y |β . Following
Theorem 4.1, since there doesn’t exist any x ∈ C such that Γ(x) = 0, we deduce that E | Y |β 6= 0.
Thus E | △0,1 X |β 6= 0.
4.2
Proof of Theorem 2.1
Proof of Theorem 2.1. In this proof, we shall denote by C a generic constant which may change from
occurrence to occurrence.
We will prove that Wn (β) − E|△0,1 X|β = OP (bn ) where bn is defined by (11).
Indeed, from Lemma 4.1, it follows that E|△0,1 X|β < +∞.
Because of H-self similarity and stationary increment properties of X, one has
△p,n X =
K
X
k=0
ak X(
K
K
K
X
k + p (d) X
△0,1 X
ak
ak
(d) X ak
)=
X(k
+
p)
=
(X(k
+
p)
−
X(p))
=
X(k) =
.
H
H
H
n
n
n
n
nH
k=0
k=0
k=0
19
We get E|△p,n X|β =
when n → ∞.
We have
(P)
E|△0,1 X|β
nβH
and EWn (β) = E|△0,1 X|β . Now we will prove that Wn (β) −−→ E|△0,1 X|β
E|Wn (β)|2 =
n−K
X
n2βH
E|△p,n X|β |△p′ ,n X|β .
(n − K + 1)2 p,p′ =0
Moreover
(d)
|△p,n X|β |△p′ ,n X|β = |
K
X
ak
n
k=0
1
=
n
|
2βH
X(k + p)|β |
H
K
X
K
X
ak
nH
k=0
X(k + p′ )|β
ap [X(k + p) − X(p′ )]|β |
k=0
K
X
K
X
K
X
k=0
ak [X(k + p′ ) − X(p′ )]|β
=
1
|
ak X(k + p − p′ )|β |
ak X(k)|β
n2βH k=0
k=0
=
1
|△p−p′ ,1 X|β |△0,1 X|β .
n2βH
(d)
It follows that
E|△p,nX|β |△p′ ,n X|β =
E|△k,1 X|β |△0,1 X|β
E|△p−p′ ,1 X|β |△0,1 X|β
=
n2βH
n2βH
with k = p − p′ . Thus
E|Wn (β)|2 =
X
1
|k|
)E|△k,1 X|β |△0,1 X|β .
(1 −
n − K + 1 |k|≤n−K
n−K +1
One has
E|Wn (β) − E|△0,1 X|β |2 = E|Wn (β)|2 − E|△0,1 X|β E|△0,1 X|β .
On the other hand, since E|△k,1 X|β = E|△0,1 X|β and
E|Wn (β) − E|△0,1 X|β |2 =
1
n−K+1
P
|k|≤n−K
(1 −
|k|
n−K+1 )
= 1, it follows that
X
1
|k|
)cov(|△k,1 X|β , |△0,1 X|β ).
(1 −
n − K + 1 |k|≤n−K
n−K +1
(49)
Using (49) and the assumption (11), one obtains
lim sup
n
1
E|Wn (β) − E|△0,1 X|β |2 ≤ Σ2 .
b2n
For all ǫ > 0, applying Markov’s inequality and using (49), we get
Σ
E|Wn (β) − E|△0,1 X|β |2
sup P (|Wn (β) − E|△0,1 X|β | > bn √ ) ≤ lim sup
≤ ǫ.
2
ǫ
n
n
b2n Σǫ
It follows that
Wn (β) − E|△0,1 X|β = OP (bn ).
(50)
In a similar way, combining with the fact that bn/2 = O(bn ), one also has
Wn/2 (β) − E|△0,1 X|β = OP (bn ).
20
(51)
b n − H = OP (bn ).
Now we will prove that H
+
+
Let φ : R × R → R be defined by
φ(x, y) = log2
x
.
y
(52)
b n − H = φ(W (β), Wn (β)).
Then H
n/2
We need the following lemma.
Lemma 4.10. Let f : D ⊂ R2 → R, be differentiable at a constant vector (a, b) ∈ D. Let (Xn , Yn ) be
P
P
random vectors whose ranges lie in D such that Xn −
→ a, Yn −
→ b and Xn − a = OP (bn ), Yn − b = OP (bn )
where {bn }n is a non-negative sequence and bn → 0 as n → +∞.
Then f (Xn , Yn ) − f (a, b) = OP (bn ).
Proof. Since f is differentiable at (a, b), we can write
f (a + h1 , b + h2 ) = f (a, b) + h1
∂f
∂f
(a, b) + h2 (a, b) + o(||(h1 , h2 ||).
∂x
∂y
as ||h|| = ||(h1 , h2 )|| → 0.
∂f
By applying Lemma 2.12 in [26] for R(x, y) = f (a + x, b + y) − f (a, b) − x ∂f
∂x (a, b) − y ∂y (a, b) and the
sequence random vector (Xn − a, Yn − b), we get
∂f
∂f
(a, b) + (Yn − a) (a, b) + oP (||(Xn − a, Yn − b)||)
∂x
∂y
= (Xn − a)OP (1) + (Yn − a)OP (1) + oP (||(Xn − a, Yn − b)||)
f (Xn , Yn ) − f (a, b) = (Xn − a)
= bn OP (1)OP (1) + bn OP (1)OP (1) + bn OP (1)oP (1)
= bn OP (1) + bn OP (1) + bn oP (1)
= bn OP (1)
Applying Lemma 4.10 with f = φ and vector (E | △0,1 X |β , E | △0,1 X |β ), combining with (50), (51)
b n − H = OP (bn ).
and the fact that φ(E | △0,1 X |β , E | △0,1 X |β ) = 0, it follows that H
(P)
b n = H.
Since lim bn = 0, it induces that lim H
n→+∞
n→+∞
b n − α = OP (bn ), we first prove that
To prove that α
h−β1 ,−β2 (α) = ψ−β1 ,−β2 (E|△0,1 X|β1 , E|△0,1 X|β2 )
(53)
where hu,v , ψu,v are defined by (8), (7), respectively.
Indeed, since {Xt , t ∈ R} is a H-sssi SαS−stable process, there exists a constant σ > 0 such that
△0,1 X = σY , where Y is the standard H-sssi, SαS random variable. For β1 , β2 ∈ R, −1/2 < β1 , β2 < 0,
from Lemma 4.2, we have
E|△0,1 X|β1 = σ β1 E|Y |β1 = σ β1
2β1 Γ( β12+1 )Γ(1 −
√
πΓ(1 − β21 )
β1
α)
.
Thus
β1
E|△0,1 X|
β2
=σ
β1 β2
=σ
β1 β2
2β1 Γ( β12+1 )Γ(1 −
√
πΓ(1 − β21 )
β1
α)
2β2 Γ( β22+1 )Γ(1 −
√
πΓ(1 − β22 )
β2
α)
!β 2
.
!β 1
.
Similarly, we also get
β2
E|△0,1 X|
β1
21
Moreover, from Lemma 4.3, E|△0,1 X|β 6= 0 for all −1/2 < β < 0, then it induces
E|△0,1 X|β1
β2
=
β1
(E|△0,1 X|β2 )
π
β1 −β2
2
β2
β1
β2 β1 +1
β2
2 )Γ ( 2 )Γ (1 − α )
.
β1
β2
β1 β2 +1
β1
2 )Γ ( 2 )Γ (1 − α )
Γβ1 (1 −
Γβ2 (1 −
Taking the natural logarithm, we have
β2 ln(E|△0,1 X|β1 ) − β1 ln(E|△0,1 X|β2 ) =
β1 − β2
β2
ln(π) + β1 ln Γ(1 − ) +
2
2
β1 + 1
β1
β2 ln Γ(
) + β2 ln Γ(1 − )
2
α
β2 + 1
β2
β1
) − β1 ln Γ(1 − ) .
− β2 ln Γ(1 − ) − β1 ln Γ(
2
2
α
It follows that
β2
β1
β2 ln Γ(1 − ) − β1 ln Γ(1 − ) =β2 ln(E|△0,1 X|β1 ) − β1 ln(E|△0,1 X|β2 )
α
α
β2
β1 + 1
β2 − β1
ln(π) − β1 ln Γ(1 − ) − β2 ln Γ(
)
+
2
2
2
β2 + 1
β1
)
+ β2 ln Γ(1 − ) + β1 ln Γ(
2
2
or h−β1 ,−β2 (α) = ψ−β1 ,−β2 (E|△0,1 X|β1 , E|△0,1 X|β2 ).
From the following lemma, we can deduce that hu,v is a strictly increasing function on (0, +∞) and
lim hu,v (x) = 0, lim hu,v (x) = −∞. Moreover, there exists an inverse function
x→+∞
x→0
h−1
u,v : (−∞, 0) → (0, +∞)
which is continuous and on (−∞, 0).
Lemma 4.11. Let 0 < v < u and gu,v : (0, +∞) → R be a function defined by
gu,v (x) = u ln (Γ(1 + vx)) − v ln (Γ(1 + ux)) .
Then gu,v is a strictly decreasing function on (0, +∞) and
lim gu,v (x) = 0, lim gu,v (x) = −∞.
x→0
x→+∞
Proof. We have
′
gu,v
(x) = uv
Γ′ (1 + vx) Γ′ (1 + ux)
Γ′ (1 + ux)
Γ′ (1 + vx)
− uv
= uv
−
.
Γ(1 + vx)
Γ(1 + ux)
Γ(1 + vx)
Γ(1 + ux)
Following Bohn-Mollerup’s theorem, Γ is a log-convex function. Let k(y) be defined by k(y) = ln Γ(y).
′ (y)
Then k′′ (y) ≥ 0 for all y > 0. It follows that ψ(y) := k′ (y) = ΓΓ(y)
is an increasing function.
′
′
Since Γ(x + 1) = xΓ(x), we have Γ (x + 1) = Γ(x) + xΓ (x). We obtain
ψ(x + 1) =
Γ(x) + xΓ′ (x)
1
Γ′ (x + 1)
=
= + ψ(x).
Γ(x + 1)
xΓ(x)
x
We will prove that ψ increases strictly.
Assume that there exist x0 , y0 such that 0 < x0 < y0 and ψ(x0 ) = ψ(y0 ), then
ψ(x0 + 1) − ψ(y0 + 1) =
1
1
y 0 − x0
−
=
> 0.
x0 y0
x0 y0
22
However, x0 + 1 < y0 + 1, then ψ(x0 + 1) ≤ ψ(y0 + 1) but this could not happen. Thus ψ is a strictly
increasing function.
′ (1+vx)
′ (1+ux)
We also have 1 < 1 + vx < 1 + ux, so ΓΓ(1+vx)
− ΓΓ(1+ux)
< 0.
′
It induces that gu,v (x) < 0 for all 0 < v < u and x > 0. This proves that gu,v (x) is a strictly decreasing
function.
It is clear that lim gu,v (x) = 0. Now we need to prove that lim gu,v (x) = −∞.
x→+∞
x→0
Applying Stirling’s formula, we have
1
1
ln Γ(1 + z) = ln(zΓ(z)) = ln z + ln Γ(z) = (z + ) ln z − z + ln(2π) + O(z −1 )
2
2
as z → +∞. Then
1
1
gu,v (x) = u (vx + ) ln(vx) − vx + ln(2π) + O((vx)−1 )
2
2
1
1
−1
− v (ux + ) ln(ux) − ux + ln(2π) + O((ux) )
2
2
u ln v − v ln u (u − v) ln(2π)
u
v
u−v
ln x +
+
+ O( ) − O( )
= uv(ln v − ln u)x +
2
2
2
vx
ux
as x → +∞.
Since 0 < v < u we deduce uv(ln v − ln u) < 0. Moreover lim
ln x
x→+∞ x
= 0, it follows that
lim gu,v (x) = −∞.
x→+∞
Then since α ∈ (0, 2], we obtain that h−β1 ,−β2 (α) = ψ−β1 ,−β2 (E|△0,1 X|β1 , E|△0,1 X|β2 ) < 0. On the
other hand
ψ−β1 ,−β2 (Wn (β1 ), Wn (β2 )) = ψ−β1 ,−β2 (Vn (β1 ), Vn (β2 )).
We deduce that
α̂n − α = ϕ−β1 ,−β2 (ψ−β1 ,−β2 (Wn (β1 ), Wn (β2 ))) − h−1
−β1 ,−β2 (h−β1 ,−β2 (α))
= ϕ−β1 ,−β2 (ψ−β1 ,−β2 (Wn (β1 ), Wn (β2 ))) − ϕ−β1 ,−β2 (h−β1 ,−β2 (α))
= ϕ−β1 ,−β2 (ψ−β1 ,−β2 (Wn (β1 ), Wn (β2 ))) − ϕ−β1 ,−β2 (ψ−β1 ,−β2 (E|△0,1 X|β1 , E|△0,1 X|β2 )).
Moreover, ϕ−β1 ,−β2 ◦ ψ−β1 ,−β2 is continuous and differentiable at
x0 = (E|△0,1 X|β1 , E|△0,1 X|β2 ).
(54)
Combining with (50), (51), we apply Lemma 4.10, and get
(P)
b n = α.
It also induces that lim α
b n − α = OP (bn ).
α
n→+∞
4.3
Proofs related to Section 3
Now we are in position to prove Theorems related to examples presented in Section 3.
Proof of Theorem 3.1. a) From Lemma 5.2 in Appendix, the assumption (11) is satisfied. Then
following Theorem 2.1, we have
Ĥn − H = OP (n−1/2 ), α̂n − 2 = OP (n−1/2 ).
23
b) We now prove the asymptotic normality for the estimators of H and α. To prove
converges to a normal distribution as n → +∞, we will first prove that
√
n(Ĥn − H)
(d)
√
n (Wn (β), Wn/2 (β)) − (E|△0,1 X|β , E|△0,1 X|β ) −−→ N2 (0, Γ1 )
(55)
as n → +∞, where Γ1 is defined by (16). Then, we need to prove that for all a, b ∈ R, ab 6= 0,
√
√
Vn := a n(Wn (β) − E|△0,1 X|β ) + b n(Wn/2 (β) − E|△0,1 X|β )
converges to G ∼ N1 (0, σ 2 ) as n → +∞, where
σ 2 = (a2 + 2b2 )
X
q!fq2
X
r∈Z
q≥d
X
ρq (r) + 2ab
q!fq2
q≥d
X
(56)
ρq1 (r),
(57)
r∈Z
fq s, ρ, ρ1 are defined by(13), (14), (15), respectively. Since {Xt }t≥0 is a H-sssi process, for all n ∈ N∗ , we
get
△0,n X, △1,n X, . . . , △n−K,nX, △0,n/2 X, △1,n/2 X, . . . , △n/2−K,n/2 X
(d)
=
1
△
X,
△
X,
.
.
.
,
△
X,
△
X,
△
X,
.
.
.
,
△
X
.
0,2
1,2
n−K,2
0,1
1,1
n/2−K,1
(n/2)H
Moreover var△k,2 =
√
var△0,1 X
, var△k,1 X
22H
= var△0,1 X. It follows that
n (Wn (β), Wn/2 (β)) − (E|△0,1 X|β , E|△0,1 X|β )
√
= n
(d)
=
√
n/2−K
X |△k,2 |β
X |△k,1 |β
(n/2)βH
(n)βH n−K
β
−
E|△
X|
,
− E|△0,1 X|β
0,1
n − K + 1 k=0 (n/2)βH
n/2 − K + 1 k=0 (n/2)βH
n/2−K
n−K
X
X
2βH
1
β
β
n
|△k,2 X| − E|△0,1 X| ,
|△k,1 X|β − E|△0,1 X|β
n − K + 1 k=0
n/2 − K + 1 k=0
√
√
n/2−K
n−K
X
X
n
n
(|Yk |β − E|Z0 |β ),
(|Zl |β − E|Z0 |β )
= (var△0,1 X)β/2
n − K + 1 k=0
n/2 − K + 1 l=0
√
√
n/2−K
n−K
X
X
n
n
=
fβ (Yk ),
fβ (Zl )
n − K + 1 k=0
n/2 − K + 1 l=0
△ X
where Yk = √ k,2
var△k,2 X
△ X
, Zl = √ l,1
var△l,1 X
and fβ is defined by (12).
We obtain that Yk ∼ N1 (0, 1), Zl ∼ N1 (0, 1), and
E(△k,2 X△k′ ,2 X)
EYk Yk′ =
var△0,1 X
22H
=
K
P
p,p′ =0
ap ap′ |k − k′ + p − p′ |2H
K
P
p,p′ =0
EYk Zl =
E(△k,2 X△l,1 X)
var△0,1 X
2H
=
K
P
p,p′ =0
ap ap′ |k − 2l + p − 2p′ |2H
2H
K
P
p,p′ =0
EZl Zl′ =
E(△l,1 X△
l′ ,1
var△0,1 X
2H
X)
=
K
P
p,p′ =0
p,p′ =0
24
ap ap′ |p −
p′ |2H
ap ap′ |k − l + p − p′ |2H
K
P
.
ap ap′ |p − p′ |2H
ap ap′ |p −
p′ |2H
,
,
Then EYk Yk′ = ρ(k − k′ ), EZl Zl′ = ρ(l − l′ ) and EYk Zl = ρ1 (k − 2l) where ρ, ρ1 are defined by (14), (15),
respectively. As in the proof of Lemma 5.2 in Appendix, we can prove that for r big enough
ρ(r)| ≤ C|r|2H−3 , |ρ1 (r)| ≤ C|r|2H−3 .
(58)
(d)
We then mimic the proof of Theorem 7.2.4 in [19] to get Vn −−→ N1 (0, σ 2 ) as n → +∞, it follows (55).
On the other hand, we have
√
n(Ĥn − H) =
√ 1
Wn/2 (β) √
= n φ(Wn (β), Wn/2 (β)) − φ(E|△0,1 X|β , E|△0,1 X|β ) .
n log2
β
Wn (β)
where φ is defined as in (17).
Since φ is differentiable at (x0 , y0 ) = (E|△0,1 X|β , E|△0,1 X|β ), we can apply Theorem 3.1 in [10] to get
√
(d)
n(Ĥn − H) −−→ N1 (0, Ξ1 )
as n → +∞, where Ξ1 is defined by (18).
Now we prove central limit theorem for the estimation of α. We will prove that
(d)
√
n(Wn (β1 ) − E|△0,1 X|β1 ), n(Wn (β2 ) − E|△0,1 X|β2 )) −−→ N2 (0, Γ2 )
√
(59)
as n → +∞, with Γ2 defined by (23).
Since {Xt }t∈R is a H-sssi process, we have
(d)
(△0,n X, . . . , △n−K,n X) =
1
(△0,1 X, . . . , △n−K,1 X) .
nH
(60)
On the other hand, var△k,1 X = var△0,1 X. Then we can write
√
√
n(Wn (β1 ) − E|△0,n X|β1 ), n(Wn (β2 ) − E|△0,n X|β2 ))
=
√
n
=
√
(d)
=
r
!
n−K
n−K
X |△k,1 X|β1
X |△k,1 X|β2
√
nβ1 H
nβ2 H
β1
n
−
E|△
X|
− E|△0,1 X|β2
,
0,1
β
H
n − K + 1 k=0
n 1
n − K + 1 k=0
nβ2 H
!
n−K
X
√
1
|△k,1 X|β1 − E|△0,1 X|β1 , n
n
n − K + 1 k=0
n
n−K +1
n−K
X
1
|△k,1 X|β2 − E|△0,1 X|β2
n − K + 1 k=0
!
!!
!!
n−K
n−K
X
X
1
1
√
fβ1 (Zk ), √
fβ (Zk ) .
n − K + 1 k=0
n − K + 1 k=0 2
△ X
where fβ1 and fβ2 are defined as in (12) and Zk = √ k,1
, Zk
var△k,1 X
+∞, Efβ22 (Z0 ) <
∼ N1 (0, 1).
We have Efβ1 (Z0 ) = Efβ2 (Z0 ) = 0, Efβ21 (Z0 ) <
+∞ and EZk Zl = ρ(k − l) where ρ is
defined by (14).
We mimic the proof of Theorem 7.2.4 of [19] to obtain that
√
√
a n Wn (β1 ) − E|△0,1 |β1 + b n Wn (β2 ) − E|△0,1 |β2
converges to N1 (0, σ 2 ) as n → +∞ for all a, b ∈ R, ab 6= 0, where σ 2 =
+∞
P
q=d
ρ, hq , gq are defined by (14), (20) respectively. This proves (59).
The function ϕ−β1 ,−β2 ◦ ψ−β1 ,−β2 : R+ × R+ → [0, +∞) is differentiable at
(x1 , y1 ) = (E|△0,1 X|β1 , E|△0,1 X|β2 ),
where ψu,v , ϕu,v are defined by (7), (9) respectively.
We can therefore apply Theorem 3.1 of [10] to get the conclusion.
25
q!(ahq + bgq )2
P
r∈Z
ρ(r)q . Here
Proof of Theorem 3.2. a) We will check the assumption (11).
Let 0 ≤ l < k, k −l ≥ K, then △k,1 X =
K
P
p=0
ap [X(k + p) − X(k)] and △l,1 X =
K
P
p′ =0
ap′ [X(l + p′ ) − X(l)] .
By the fact that {Xt }t≥0 has independent increments, we obtain that X(l + p′ ) − X(l), X(k + p) − X(k)
are independent for all p, p′ = 0, . . . , K since 0 ≤ l ≤ l + p′ ≤ k ≤ k + p.
It follows that △k,1 X and △0,1 X are independent for |k| ≥ K. Thus cov(| △k,1 X |β , | △0,1 X |β ) = 0. We
deduce that
X
1
C
1 X
| cov(| △k,1 X |β , | △0,1 X |β ) |=
| cov(| △k,1 X |β , | △0,1 X |β ) |=
n k∈Z,|k|≤n
n k∈Z,|k|≤K
n
where C is a positive constant. We thus get (11) with bn = n−1/2 and it follows that Ĥn − H =
OP (n−1/2 ), α̂n − α = OP (n−1/2 ).
b) To prove the asymptotic normality for the estimator of H, we first prove that for all n ∈ N, n > 2K,
√
n (Wn , Wn/2 ) − (E|△0,1 X|β , E|△0,1 X|β )
converges in distribution to a normal distribution as n → +∞.
Since {Xt }t≥0 is a H-sssi process, one has
△0,n X, △1,n X, . . . , △n−K,nX, △0,n/2 X, △1,n/2 X, . . . , △n/2−K,n/2 X
(d)
=
1
△
X,
△
X,
.
.
.
,
△
X,
△
X,
△
X,
.
.
.
,
△
X
.
0,2
1,2
n−K,2
0,1
1,1
n/2−K,1
(n/2)H
Moreover
E|△0,1 X|β
, E|△k,1 X|β = E|△0,1 X|β ,
2βH
var|△0,1 X|β
, var|△k,1 X|β = var|△0,1 X|β .
var|△k,2 |β =
22βH
E|△k,2 |β =
It follows that
√
= n
(d
√
= n
2βH
n−K+1
2βH
n−K+1
n−K
P
p=0
n−K
P
p=0
√
n Wn , Wn/2 − E|△0,1 X|β , E|△0,1 X|β
1
|△p,2 X)|β , n/2−K+1
△p,2
X)|β
−
n/2−K
P
p=0
△ X β
|
E| 20,1
H
,
1
n/2−K+1
Now we need to prove that for all a, b ∈ R, ab 6= 0,
√
Sn :=a n
|△p,1 X)|β
!
− E|△0,1 X|β , E|△0,1 X|β
n/2−K
P
p=0
|△p,1
X)|β
n−K
X
2βH
△0,1 X
|△p,2 X)|β − E| H |β
n − K + 1 p=0
2
− E|△0,1
X|β
!
!
.
n/2−K
X
1
|△p,1 X)|β − E|△0,1 X|β
+b n
n/2 − K + 1 p=0
√
converges to a normal distribution when n → +∞. Let
Zp :=
It follows that
Sn =
s
2βH a
|△2p,2 X|β + |△2p+1,2 X|β + b|△p,1 X|β .
2
n/2−K
X
n
1
p
(Zp − EZp + Un ,
n/2 − K + 1
n/2 − K + 1 p=0
26
(61)
where
√
n−2K+1
n−K
X
X
△0,1 X β
2βH a n 1 − K
△0,1 X β
β
β
|△p,2 X| − E| H | +
Un =
|△p,2 X| − E| H |
n − K + 1 n − 2K + 2 p=0
2
2
p=n−2K+2
=
Since
√
2βH a n
n−K +1
K
P
k=0
1−K
n − 2K + 2
n−2K+1
X
n−K
X
Yp +
p=0
p=n−2K+2
Yp , Yp = |△p,2 X|β − E|
△0,1 X β
| .
2H
ak = 0, one can write
Zp =
2βH a
2
+b
K
X
k=0
K
X
ak X(
k=0
k + 2p
) − X(p)
2
β
+
K
X
ak X(
k=0
β
k + 2p + 1
) − X(p)
2
β
ak (X(k + p) − X(p)) .
If p − p′ > K − 1, since X has independent increments and
k + 2p′
k + 2p k + 2p + 1
≤ p ≤ min{
,
, k + p},
2
2
2
k + 2p′ + 1
k + 2p k + 2p + 1
0 ≤ p′ ≤
≤ p ≤ min{
,
, k + p},
2
2
2
k + 2p k + 2p + 1
0 ≤ p′ ≤ k + p′ ≤ p ≤ min{
,
, k + p},
2
2
0 ≤ p′ ≤
for all k = 0, . . . , K, it follows that Zp , Zp′ are independent. It induces that {Zp }p∈N is a (K−1)-dependent
sequence of random variables. For l ∈ R fixed, we also have
Zp+l
K
k + 2(p + l)
2βH a X
ak X(
) − X(l)
=
2
2
k=0
+b
K
X
k=0
β
+
K
X
k=0
β
ak
k + 2(p + l) + 1
X(
) − X(l)
2
β
ak (X(k + p + l) − X(l)) .
On the other hand, since X has stationary increments and X(0) = 0 almost surely, we have
(d)
(X(t + l) − X(l))t∈R = (X(t))t∈R .
(d)
Then (Zp+l , p ∈ R) = (Zp , p ∈ R) or in another way, (Zp , p ∈ R) is stationary.
It follows that {Zp }p∈N is a stationary (K − 1)-dependent sequence of random variables. From Theorem
2.8.1 in [16], we get
s
n/2−K
X
1
n
p
(Zp − EZp )
n/2 − K + 1
n/2 − K + 1 p=0
converges in distribution to a centered normal distribution with variance
σ 2 = 2(varZ0 + 2
K−1
X
cov(Z0 , Zk )) = a2 σ12 + b2 σ22 + 2abσ1,2
k=1
27
(62)
(d)
where σ12 , σ22 , σ1,2 are defined as in (27), (28), (29). We also have EYp = 0, EUn = 0, Yp = Y0 and
EYp2 = EY02 for all p. Thus
EUn2 =
≤
22βH a2 n
(n − K + 1)2
22βH+1 a2 n
(n − K +
1)2
22βH+1 a2 n
≤
(n − K + 1)2
=
E
1−K
n − 2K + 2)
n−2K+1
X
Yp +
p=0
K)2
(1 −
E
(n − 2K + 2)2
n−2K+1
X
p=0
n−K
X
p=n−2K+2
2
2
Yp
Yp + E
n−K
X
p=n−2K+2
2
Yp
(1 − K)2
(n − 2K + 2)2 EY02 + (K − 1)2 EY02
(n − 2K + 2)2
!
22βH+2 a2 (K − 1)2 n
nC
EY02 =
.
2
(n − K + 1)
(n − K + 1)2
It follows that EUn2 converges to 0 as n → +∞. Moreover EUn = 0, using Chebyshev’s inequality, we
(P)
obtain that Un −−→ 0 as n → +∞.
Following Slutsky’s theorem, as n → +∞, Sn converges in distribution to a centered normal distribution
with variance σ as in (61).
(d)
√
We deduce that n (Wn , Wn/2 ) − (E|△0,1 X|β , E|△0,1 X|β ) → N2 (0, Γ3 ) where Γ3 is defined by (26).
Since
√
n(Ĥn − H) =
√ 1
Wn/2 √
n log2
= n φ(Wn (β), Wn/2 (β)) − φ(E|△0,1 X|β , E|△0,1 X|β )
β
Wn
(63)
√
(d)
where φ is defined by (17). Applying Theorem 3.1 of [10], we get n(Ĥ − H) → N1 (0, Ξ2 ) with Ξ2
defined by (25).
We now prove the central limit theorem for the estimation
of α in the case of SαS−stable Lévy motion.
√
We need to prove that for all n ∈ N, n > K, then n (Wn (β1 ), Wn (β2 )) − (E|△0,1 X|β1 , E|△0,1 X|β2 )
converges in distribution to a normal distribution as n → +∞.
We consider
√
√
Sn = a n Wn (β1 ) − E|△0,1 X|β1 + b n Wn (β2 ) − E|△0,1 X|β2
for all a, b ∈ R, ab 6= 0. Since {Xt , t ∈ R} is a H self-similar process, we have
√
n−K
X
n
Sn =
a(|△k,1 X|β1 − E|△k,1 X|β1 ) + b(|△k,1 X|β2 − E|△k,1 X|β2 ) .
n − K + 1 k=0
√
n−K
X
n
(Zk − EZk )
=
n − K + 1 k=0
(d)
where
Zk = a|△k,1 X|β1 + b|△k,1 X|β2 .
(64)
Since {Xt , t ∈ R} has stationary increments, {Zk , k ∈ N} is stationary.
Moreover, if k − k′ > K − 1, since {Xt , t ∈ R} has independent increments, then Zk , Zk′ are independent.
We obtain that {Zk , k ∈ N} is a stationary (K − 1)-dependent sequence of random variables. Then
applying Theorem 2.8.1 of [16], as n → +∞, Sn converges to a centered normal distribution with
variance:
2
σ = varZ0 + 2
K−1
X
k=0
28
cov(Z0 , Zk ).
(65)
We can write σ 2 in details
2
2
σ =a
+b
2
β1
var|△0,1 X|
β2
var|△0,1 X|
+2
+2
K−1
X
k=1
K−1
X
k=1
"
β1
β1
β2
β2
!
cov(|△0,1 X| , |△k,1 X| )
!
cov(|△0,1 X| , |△k,1 X| ) + 2ab
X
1 K−1
× cov(|△0,1 X| , |△0,1 X| ) +
cov(|△0,1 X|β1 , |△k,1 X|β2 ) + cov(|△0,1 X|β2 , |△k,1 X|β1 )
2 k=1
β1
β2
#
= a2 σ12 + b2 σ22 + 2abσ1,2 ,
where σ12 , σ22 , σ1,2 are defined by (32), (33), (34) respectively.
(d)
√
It follows that n (Wn (β1 ), Wn (β2 )) − (E|△0,1 X|β1 , E|△0,1 X|β2 ) → N2 (0, Γ4 ), where Γ4 defined by
(31).
The function ϕ−β1 ,−β2 ◦ ψ−β1 ,−β2 : R+ × R+ → [0, +∞) is differentiable at
(x1 , y1 ) = (E|△0,1 X|β1 , E|△0,1 X|β2 ),
where ψu,v , ϕu,v are defined by (7), (9) respectively. Then we apply Theorem 3.1 of [10] to get the
conclusion.
Proof of Theorem 3.3. Set f (t) =
△k,1 X =
and || △k,1 X ||αα =
R
R
Z X
K
R j=0
PK
k=0 ak |k
H−1/α
aj (| k + j − s |
△k,1 X
||△k,1 X||α ,
H−1/α
−|s|
=
Z
R
R
f (s − k)M (ds)
|f (u)|α du = ||△0,1 X||αα .
then || Uk ||αα = 1 and Uk =
(11) is satisfied. Therefore, we consider
Sn =
R
R
f (s−k)
||△k,1 X||α M (ds).
k∈Z,|k|≤n
We now prove that the assumption
1 X
| cov(|△k,1 X|β , |△0,1 X|β )|.
n k∈Z,|k|≤n
Since ||△k,1 X||α = ||△0,1 X||α , it follows that
X
)M (ds) =
Z
| f (s − k) |α ds. By taking the change of variable u = s − k, we get
||△k,1 X||αα
Let Uk =
− t|H−1/α . For all k ∈ Z, one has
|cov(|△k,1 X|β , |△0,1 X|β )| = ||△0,1 X||2β
α
X
k∈Z,|k|≤n
(66)
|cov(|Uk |β , |U0 |β )|.
Moreover
[Uk , U0 ]2 =
Z
R
f (s − k)f (s)
||△0,1 X||2α
α/2
ds
Together with Lemma 3.6 in [11], there exist k0 > 4K and 0 < η < 1 such that for all k ∈ Z, |k| > k0 ,one
has
[Uk , U0 ]2 ≤ η < 1.
29
Applying Theorem 4.2, there exists C(η) > 0 depending on η such that
|cov(|Uk |β , |U0 |β )| ≤ C(η)
Z
|f (s − k)f (s)|α/2 ds
R
for all |k| > k0 . Then for n > k0 , one obtains that
X
k∈Z,|k|≤n
X
|cov(|Uk |β , |U0 |β )| =
k∈Z,|k|≤k0
≤C
Because f (x) ∈ Lα (R, dx), one has
Then Sn =
C
n
P
k∈Z,k0 <|k|≤n
|k|
X
|cov(|Uk |β , |U0 |β )| +
Z
k∈Z,|k|≤k0 R
R
P
.
k∈Z,k0 <|k|≤n
−1
O(n )
αH−(L+1)α
2
O(n
ln n
O(
n
X
|f (s − k)f (s)|α/2 ds + C
Since αH − (L + 1)α < 0, using Lemma 5.4 in Appendix, we also get
Sn =
|cov(|Uk |β , |U0 |β )|
k∈Z,k0 <|k|≤n
|k|
αH−(L+1)α
2
.
| f (s − k)f (s) |α/2 ds < +∞.
k∈Z,|k|≤k0 R
αH−(L+1)α
2
X
)
)
if H < L + 1 −
if H > L + 1 −
if H = L + 1 −
2
α
2
α
2
α
,
where Sn is defined by (66). Applying Theorem 2.1, we have
where bn is defined by(35).
b n − H = OP (bn ),
Wn (β) − E|△0,1 X|β = OP (bn ), H
Proof of Theorem 3.4. We have
+∞
Z Z
△k,1 X =
||△k,1 X||αα
0 R
+∞
Z Z
=
0 R
+∞
Z Z
=
0
R
K
X
!
ai 1Sk+i (x, r) M (dx, dr)
i=0
|
K
X
|
K
X
ai 1Sk+i (x, r)|α (r ν−2 )α dxdr
i=0
i=0
ai 1Si (x − k, r)|α (r ν−2 )α d(x − k)dr.
By taking the change of variable u = x − k, one obtains that
||△k,1 X||αα
Set Uk =
Set
△k,1 X
||△k,1 X||α
In =
=
△k,1 X
||△0,1 X||α .
X
k∈Z,|k|≤n
=
+∞
Z Z
0
R
|
K
X
i=0
ai 1Si (u, r)|α (r ν−2 )α dudr = ||△0,1 X||αα .
Obviously, ||Uk ||αα = 1. We now prove that the condition (11) is satisfied.
|cov(|△k,1 X|β , |△0,1 X|β )| = ||△0,1 X||βα
X
k∈Z,|k|≤n
|cov(|Uk |β , |U0 |β )|.
For n > 2K, applying Lemma 5.3 in Appendix, one gets
In ≤ C
≤C
X
k∈Z,|k|≤2K
X
k∈Z,|k|≤2K
|cov(|Uk |β , |U0 |β )| +
|cov(|Uk |β , |U0 |β )| +
30
X
k∈Z,2K<|k|≤n
X
k∈Z,2K<|k|≤n
|cov(|Uk |β , |U0 |β )|
|k|ν−1 .
Since 0 < ν < 1, one gets −1 < ν − 1 < 0. Following Lemma 5.4 in Appendix, one obtains
X
1
|k|ν−1 = O(nν−1 ).
n k∈Z,2K<|k|≤n
Then we get the condition (11). Applying Theorem 2.1, we obtain that Wn (X) − E|△0,1 X|β = OP (bn )
and
b n − H = OP (bn ), α̂n − α = OP (bn )
H
where bn is defined as in (37).
5
Appendix
We present here some technical results related to examples introduced in Section 3.
5.1
Auxiliary results related to Fractional Brownian motion
We are in position to provide and prove some technical results related to fractional Brownian motion.
These results are used to present the variances for the limit distributions of the central limit theorems
for the estimators of H and α and to prove Theorem 3.1 in Subsection 4.3.
Proposition 5.1. Let X be a H fractional Brownian motion with H ∈ (0, 1). For β ∈ R, −1/2 < β < 0,
let fβ be defined as in (12),
fβ =
△ X
where Z0 = √ 0,1
var△0,1 X
q
β
var△0,1 X (|x|β − E|Z0 |β )
. Then fβ can be expanded in a unique way into series of Hermite polynomials
fβ (x) =
X
fβ,q Hq (x)
q≥d
and
P
q≥d
2 < +∞, where d is the Hermite rank of f , moreover d ≥ 2.
q!fβ,q
β
Proof. Since −1/2 < β < 0, one has
√1
2π
R
fβ (x)e−x
2 /2
dx = 0 and
R
√1
2π
R
R
fβ2 (x)e−x
2 /2
dx < +∞. Then
following Proposition 1.4.2-(iv) in [19], we can write fβ in terms of Hermite polynomials in a unique way
fβ (x) =
X
fβ,q Hq (x),
q≥d
where d ≥ 1 is the Hermite rank of fβ and Hq s are the Hermite polynomials.
Moreover, it is clear that Z0 ∼ N (0, 1). From Proposition 2.2.1 in [19], we get
E[Hp (Z0 )Hq (Z0 )] =
(
0 if p 6= q
p! if p = q
Then since H1 (x) = x, one has EH1 (Z0 )fβ (Z0 ) = EZ0 fβ (Z0 ) = fβ,1 EZ02 = fβ,1 . Combining with the fact
that
(var△0,1 X)β
√
EH1 (Z0 )fβ (Z0 ) =
2π
Z
R
x(|x|β − E|Z0 |β )e−x
we deduce that fβ,1 = 0. It follows that d ≥ 2.
Moreover,
Efβ2 (Z0 )
1
=√
2π
Z
fβ2 (x)e−x
R
31
2 /2
dx < +∞.
2 /2
dx = 0,
On the other hand,
Efβ2 (Z0 ) =
X
fβ,p fβ,q E[Hp (Z0 )Hq (Z0 )] =
P
2
q!fβ,q
.
q≥d
p,q≥d
It follows that
X
2 < +∞.
q!fβ,q
q≥d
!!
1 ρ
, | ρ |≤ 1. Then for each β ∈ C, Re(β) ∈ (−1/2, 0),
Lemma 5.1. Let (U, V ) = N2 (0, 0),
ρ 1
there exists a constant C > 0 such that ∀ | ρ |≤ 1, we have:
(d)
Proof. Let Σ =
!
1 ρ
ρ 1
| cov(| U |β , | V |β ) |≤ Cρ2
. We have det(Σ) = 1 −
ρ2
and
Σ−1
=
!
1 ρ
. The density function of
ρ 1
1
1−ρ2
(U, V ):
−1/2
f (x, y) =| 2πΣ |
"
!#
1
x
exp − (xy)Σ−1
y
2
= (2π)2 detΣ
−1/2
exp −
1
(x2 + y 2 − 2ρxy) .
2(1 − ρ2 )
We get
β
β
E |U | |V |
1
=
2π 1 − ρ2
E | U |β E | V |β =
1
2π
p
Z
R2
Z
1
(x2 + y 2 − 2ρxy) dxdy
| x | | y | exp −
2(1 − ρ2 )
β
β
R2
| x |β | y |β exp(−
x2 + y 2
)dxdy
2
and
cov(| U |β , | V |β ) = E(| U |β | V |β ) − E | U |β E | V |β
=
where
Since
1
2π
Z
R2
| x |β | y |β exp(−
x2 + y 2
)Aρ (x, y)dxdy
2
!
R
R
| x |β xe−x
1
ρ2
ρxy
Aρ (x, y) = p
exp −
(x2 + y 2 ) exp
2
1−ρ
1 − ρ2
1 − ρ2
2 /2
− 1.
dx = 0 we obtain that
1
cov(| U | , | V | ) =
2π
β
β
Z
R2
| x |β | y |β exp(−
x2 + y 2
)Bρ (x, y)dxdy
2
with
!
ρxy
1
ρ2
2
2
(x
exp
+
y
)
Bρ (x, y) = Aρ (x, y) − ρxy = p
exp
−
2
1−ρ
1 − ρ2
1 − ρ2
32
− 1 − ρxy.
Using L’Hôpital rule, we get:
′
Bρ
Bρ
lim 2 = lim
ρ→0 ρ
ρ→0 2ρ
′
′
Bρ = Aρ (x, y) − xy
=
h
ρ(1 − ρ2 )
−3
2
!
ρ2
ρxy
× exp −
(x2 + y 2 ) exp
2
1−ρ
1 − ρ2
′
lim
ρ→0
Bρ
= xy · lim
ρ→0
2ρ
h
exp
− (x2 + y 2 )[2ρ(1 − ρ2 )−1 + 2ρ3 (1 − ρ2 )−2 ] + xy (1 − ρ2 )−1 + 2ρ2 (1 − ρ2 )−2
−ρ2
2
1−ρ2 (x
+ y2) +
ρxy
1−ρ2
2ρ
− xy
i
(1 − ρ2 )−1 − 1
+
i
1
1
− (x2 + y 2 ) := A + − (x2 + y 2 ).
2
2
Then we continue using L’Hôpital rule for the remaining limit:
xy
A=
lim exp
2 ρ→0
"
=
−ρ2
ρxy
(x2 + y 2 ) +
2
1−ρ
1 − ρ2
!
×
2ρ
2ρ3
2ρ
2
2
−
(x
+
y
)
+
(1 − ρ2 )2
1 − ρ2 (1 − ρ2 )2
!
+ xy
x2 y 2
.
2
1
2ρ2
+
1 − ρ2 (1 − ρ)2
!#
One has
Bρ
x2 y 2 + 1
− (x2 + y 2 )
=
ρ→0 ρ2
2
!
ρxy
∂ 2 Bρ (x, y)
ρ2
2
2
= Pρ (x, y)exp −
(x + y ) exp
.
∂ρ2
1 − ρ2
1 − ρ2
lim
where Pρ (x, y) is a fourth degree polynomial that depends continuously on ρ. We also have
′
B0 (x, y) = 0, Bρ (x, y) |ρ=0 = 0.
A Taylor expansion up to order 2 leads to
!
ρ̃xy
ρ̃2
(x2 + y 2 ) exp
Bρ (x, y) = ρ Pρ̃ (x, y)exp −
1 − ρ̃2
1 − ρ̃2
2
with ρ̃ ∈ (0, ρ). On the compact set | ρ |≤ 1/2, the polynomial Pρ (x, y) can be bounded by a fourth
degree polynomial P (x, y), for all x, y ∈ R, | Pρ (x, y) |≤| P (| x |, | y |) |. Moreover
!
ρ̃xy
ρ̃2
(x2 + y 2 ) exp
exp −
1 − ρ̃2
1 − ρ̃2
But with | ρ̃ |≤ 1/2, we get |
ρ̃
1−ρ̃2
|≤ 2/3. So exp
ρ̃xy
1−ρ̃2
≤ exp
≤ exp
2|xy|
3
ρ̃xy
.
1 − ρ̃2
. Thus
2 | xy |
.
| Bρ (x, y) |≤ ρ | P (| x |, | y |) | exp
3
2
Because the power function grows faster than the polynomial function, we have
Z
R2
Re(β)
| xy |
x2 + y 2
2 | xy |
exp(−
)P (| x |, | y |) | exp
dxdy < ∞.
2
3
So we have the conclusion.
33
Lemma 5.2. Let X be a fractional Brownian motion, β ∈ C, Re(β) ∈ (−1/2, 0). Then
X
k∈Z
| cov(| △k,1 X |β , | △0,1 X |β ) |< +∞.
Proof. We have
cov(△k,1 X, △0,1 X) = −
K
K
k2H X
p − p′ 2H
1 X
ap ap′ | k + p − p′ |2H = −
ap ap′ | 1 +
| .
2 p,p′ =0
2 p,p′ =0
k
We just need to consider cov(△k,1 X, △0,1 X) when | k |≥ K. Since 1 +
cov(△k,1 X, △0,1 X) = −
Set
g(x) = −
If H =
1
2
p−p′
k
≥ 0, we get
K
k2H X
p − p′ 2H
ap ap′ (1 +
) .
2 p,p′ =0
k
K
1 X
ap ap′ (1 + (p − p′ )x)2H .
2 p,p′ =0
then g(x) = 0. If H 6= 12 , using Taylor expansion as x → 0, we get
g′ (x) = −H
K
X
p,p′ =0
ap ap′ (p − p′ )(1 + (p − p′ )x)2H−1
g′′ (x) = −H(2H − 1)
K
X
p,p′ =0
ap ap′ (p − p′ )2 (1 + (p − p′ )x)2H−2
g(3) (x) = −H(2H − 1)(2H − 2)
K
X
p,p′ =0
ap ap′ (p − p′ )3 (1 + (p − p′ )x)2H−3 .
Thus g(0) = 0, g′ (0) = 0, g′′ (0) = 0, g(3) (0) = 0 and we obtain that g(x) = o(x3 ) as x → 0. It follows that
cov(△k,1 X, △0,1 X) ∼ k2H · o(
1
) ∼ o(k2H−3 )
k3
as k → +∞. We can apply similarly as k → −∞. Then there exists a constant C such that for all
k, | k |≥ K and for all H ∈ (0, 1),
| cov(△k,1 X, △0,1 X) |≤ C | k |2H−3 .
For all k ∈ Z, we have
var△k,1 X = E
K
X
ap X(k + p)
p=0
We now apply the Lemma 5.1 with U = √
K
X
p′ =0
ap′ X(k + p′ ) = −
△k,1 X
,V
var(△k,1 X)
It follows that
=√
(67)
K
1 X
ap ap′ | p − p′ |2H .
2 p,p′ =0
△0,1 X
.
var(△0,1 X)
Then
△k,1 X
cov 2 (△k,1 X, △0,1 X)
△0,1 X
|β , | q
cov | q
.
|β ≤ C ·
var 2 △0,1 X
var(△k,1 X))
var(△0,1 X))
cov(| △k,1 X |β , | △0,1 X |β ) ≤ Ccov 2 (△k,1 X, △0,1 X), ∀k, k ∈ Z.
34
Since H ∈ (0, 1), we get
X
k∈Z
P
| k |4H−6 < +∞. Applying inequality (67), we obtain
k∈Z
cov(| △k,1 X |β , | △0,1 X |β ) =
X
cov(| △k,1 X |β , | △0,1 X |β )
k∈Z,|k|<K
X
+
cov(| △k,1 X |β , | △0,1 X |β )
k∈Z,|k|≥K
≤C
5.2
X
k∈Z,|k|<K
cov 2 (△k,1 X, △0,1 X) + C
X
k∈Z,|k|≥K
| k |4H−6 < +∞.
Auxiliary results related to Takenaka’s process
Lemma 5.3. Let {Xt , t ∈ R} be a Takenaka’s process defined by (36). Then for β ∈ R, β ∈ (−1/2, 0)
and |k| > 2K, we have
cov(| △k,1 X |β , | △0,1 X |β ) ≤ Ckν−1 .
Proof. One has
△k,1 X =
where fk =
K
P
i=0
ai 1Sk+i =
K
P
i=0
K
X
Z
ai X(k + i) =
i=0
K
X
ai 1Sk+i (x, r)M (dx, dr)
i=0
R×R+
K
P
ai (1Ck+i − 1C0 )2 . From the fact that
K
X
| fk | =| 1 − 21C0 ||
i=0
ai 1Ck+i |=|
i=0
K
X
i=0
ai = 0 and | 1 − 21C0 |= 1, it induces
ai 1Ck+i | .
Therefore we have to estimate, as |k| → +∞,
Ik = [△k,1 X, △0,1 X]2 =
+∞
Z
r ν−2
0
Z
| fk (x, r)f0 (x, r) | dxdr.
R
We will find an upper bound for Ik when |k| ≥ 2K.
If x > K + r then 1Ci (x, r) = 0 for all i = 0, . . . , K, thus f0 (x, r) = 0. If x < k − r, 1Ck+i (x, r) = 0 for
all i = 0, . . . , K, it follows that fk (x, r) = 0.
As a result, fk (x, r)f0 (x, r) = 0 for all x ∈ (−∞, k − r) ∪ (K + r, +∞).
Let k > 2K. If r < k−K
2 ⇔ k − r > K + r then fk (x, r)f0 (x, r) = 0 for all x. Thus one gets
Ik =
Z
R
=
Z
R
k−K
2
Z
r ν−2 | fk (x, r)f0 (x, r) | dr +
0
+∞
Z
k−K
2
r
ν−2
+∞
Z
k−K
2
| fk (x, r)f0 (x, r) | drdx =
r ν−2 | fk (x, r)f0 (x, r) | dr
dx
+∞
Z
r ν−2
k−K
2
Z
R
| fk (x, r)f0 (x, r) | dxdr.
Here we consider fk (x, r)f0 (x, r).
Since k > 2K, then k + K − r ≤ k + r. For k + K − r ≤ x ≤ k + r, | x − k − i |≤ r for all i = 0, . . . , K, it
follows that
1Ck+i (x, r) = 1 and fk (x, r) =
K
P
i=0
ai = 0.
Therefore fk (x, r)f0 (x, r) = 0 if x ∈ (−∞, k − r) ∪ (k + K − r, +∞). We also have
| fk (x, r) |=|
K
X
i=0
ai 1Ck+i (x, r) |≤
35
K
X
i=0
| ai |
for all k ∈ N. Thus
Ik ≤
+∞
Z
r
ν−2
k−r
k−K
2
≤C
k+K−r
Z
+∞
Z
r
ν−2
| fk (x, r)f0 (x, r) | dxdr
k+K−r
Z
dxdr = C(
k−r
k−K
2
k − K ν−1
)
≤ Ckν−1
2
k
since k−K
2 ≥ 4 and 0 < ν < 1.
Let k < −2K. If r < − k+K
2 ⇔ k + K + r < −r, then for all i = 1, . . . , K,
1Ck+i (x, r) = 0, ∀x ∈ (k + K + r, +∞), 1Ci (x, r) = 0, ∀x ∈ (−∞, −r).
It follows that fk (x, r)f0 (x, r) = 0, for all x ∈ (−∞, −r) ∪ (k + K + r, +∞) = R. Therefore
+∞
Z
Ik =
r ν−2
− k+K
2
Z
R
|fk (x, r)f0 (x, r)|dxdr
For r > − k+K
2 , r > K − K/2 = K/2. We have fk (x, r)f0 (x, r) = 0 for all x ∈ (−∞, k − r) ∪ (k − r +
K, k + r) ∪ (k + r + K, K + r) ∪ (K + r, +∞). It induces that
Ik =
+∞
Z
r ν−2
− k+K
2
=
+∞
Z
− k+K
2
≤C
Z
R
r ν−2
+∞
Z
− k+K
2
|fk (x, r)f0 (x, r)|dxdr
k−r+K
Z
k−r
|fk (x, r)f0 (x, r)|dxdr +
k+r+K
Z
k+r
|fk (x, r)f0 (x, r)|dxdr
r ν−2 dr ≤ C|k|ν−1 .
Putting together with Theorem 4.2, for |k| > 2K we obtain that
cov(| △k,1 X |β , | △0,1 X |β ) ≤ Ckν−1 .
5.3
Auxiliary results related to rate of convergence
We present here a lemma used to determine rate of convergence in the proofs of Theorems 3.3 and 3.4.
Lemma 5.4. For p < 0, let Sn =
1
n
P
p
|k|≤n |k| ,
Sn =
then lim Sn = 0. Moreover
n→+∞
−1
O(n )
if p < −1
if − 1 < p < 0
if p = −1.
O(np )
O( ln n )
n
Proof. Set
Sn =
n
1X
kp .
n k=1
36
If p < −1, since
R∞
xp dx < +∞, following the integral test for convergence, we get
∞
P
kp < +∞. Then
k=1
1
Sn = O(n−1 ).
If −1 < p < 0, we take a constant ǫ such that 0 < ǫ < −p, then
Sn =
Since p + ǫ < 0, we get
If p = −1, then Sn =
1
n
n
P
k=1
n
P
k=1
1
k 1+ǫ
1
k
n
n
n
X
k1+ǫ+p
n1+ǫ+p
1X
1
1X
p+ǫ
≤
=
n
.
1+ǫ
n k=1 k1+ǫ
n k=1 k1+ǫ
k
k=1
< +∞. Then Sn = O(np+ǫ ) for all 0 < ǫ < −p. Thus Sn = O(np ).
= O( lnnn ).
In all cases, we have lim Sn = 0.
n→+∞
Acknowledgements
The authors would like to thank Michel Zinsmeister for his helpful suggestions.
References
[1] P. Abry, L. Delbeke, and P. Flandrin. Wavelet-based estimators for the self-similarity parameter of
α-stable processes. In IEEE International Conference on Acoustics, Speech and Signal Processing
III, page 1581–1584. Az, USA: Phoenix, 1999.
[2] P. Abry, B. Pesquet-Popescu, and M.S. Taqqu. Estimation ondelette des paramètres de stabilité
et d’autosimilarité des processus α− stables autosimilaires. In 17ème Colloque sur le traitement du
signal et des images, FRA, 1999. GRETSI, Group d’Etudes du Traitement du Signal 1999.
[3] A. Ayache and J. Hamonier. Linear fractional stable motion: a wavelet estimator of the alpha
parameter. Statistics and Probability Letters, 82(8):1569–1575, 2013.
[4] A. Ayache and J. Hamonier. Linear multifractional stable motion: wavelet estimation of H(.) and α
parameter. Lithuanian Mathematical Journal, 55(2):159–192, 2015.
[5] J. M. Bardet and C. Tudor. A wavelet analysis of the Rosenblatt process: chaos expansion and estimation of the self-similarity parameter. Stochastic Processes and their Applications, 120(12):2331–
2362, 2010.
[6] A. Benassi, P. Bertrand, S. Cohen, and J. Istas. Identification of the Hurst index of a step fractional
Brownian motion. Stat. Inf. Stoc. Proc., 3(1):101–111, 2000.
[7] A. Benassi, S. Cohen, and J. Istas. Identifying the multifractional function of a Gaussian process.
Stat. and Proba. Letters, 39(4):337–345, 1998.
[8] J. F. Coeurjolly. Estimating the parameters of a fractional Brownian motion by discrete variations
of its sample paths. Stat. Inf. Stoc. Proc., 4(2):199–227, 2001.
[9] S. Cohen and J. Istas. Fractional fields and applications. Springer-Verlag, Berlin, 2013.
[10] D. Dacunha-Castelle and M. Duflo. Probabilités et Statistiques, volume 2. Masson, Paris, 1983.
[11] J. Istas. Estimating self-similarity through complex variations. Electronic Journal of Statistics,
6:1392–1408, 2012.
37
[12] J. Istas. Manifold indexed fractional fields. ESAIM: Probability and Statistics, 16:222–276, 2012.
[13] J. Istas and G. Lang. Quadratic variations and estimation of the Höder index of a Gaussian process.
Ann. Inst. H. Poincaré Probab. Statist, 33(4):407–436, 1997.
[14] C. Lacaux and J.-M. Loubes. Hurst exponent estimation of fractional Lévy motions. ALEA Lat.
Am. J. Probab. Math. Stat., 3:143–161, 2007.
[15] R. Le Guével. An estimation of the stability and the localisability functions of multistable processes.
Electronic Journal of Statistics, 7:1129–1166, 2013.
[16] E. L. Lehmann. Elements of large-sample theory. Springer-Verlag, New York, 1999.
[17] J. H. McCulloch. Financial applications of stable distributions, volume 14. North-Holland, New
York, in G. S. Maddala and C. R. Rao, Handbook of Statistics edition, 1996.
[18] I. Nourdin, D. Nualart, and C. Tudor. Central and non-central limit theorems for weighted power
variations of fractional Brownian motion. Ann. Inst. H. Poincaré Probab. Statist., 46(4):1055–1079,
2010.
[19] I. Nourdin and G. Peccati. Normal approximations with Malliavin calculus: from Stein’s method to
universality, volume 192. Cambridge University Press, Cambridge, 2012.
[20] V. Pipiras, M. S. Taqqu, and P. Abry. Bounds for the covariance of functions of infinite variance
stable random variables with applications to central limit theorems and wavelet-based estimation.
Bernoulli, 13(4):1091–1123, 2007.
[21] G. Samorodnitsky and M. S. Taqqu. Stable Non-Gaussian Random Processes. Chapmann and Hall,
New York, 1994.
[22] L. Schwartz. Théorie des distributions. Hermann, Paris, 1978.
[23] S. Stoev, V. Pipiras, and M. S Taqqu. Estimation of the self-similarity parameter in linear fractional
stable motion. Signal Processing, 82:1873–1901, 2002.
[24] S. Stoev and M. S Taqqu. Asymptotic self-similarity and wavelet estimation for long-range dependent
fractional autogressive intergrated moving average time series with stable inovations. Journal of Time
Series Analysis, 26(2):211–249, 2005.
[25] S. Takenaka. Integral-geometric construction of self-similar stable processes. Nagoya Math. J.,
123:1–12, 1991.
[26] A.W. Van Der Vaart. Asymptotic statistics. Cambrigde University Press, Cambridge, 2000.
38
| 10math.ST
|
arXiv:1408.3215v1 [cs.NE] 14 Aug 2014
Cortical Processing with Thermodynamic-RAM
M. Alexander Nugent
Timothy W. Molter
M. Alexander Nugent Consulting, Santa Fe, NM, USA
KnowmTech LLC, Albuquerque, NM, USA
Xeiam LLC, Santa Fe, NM, USA
Email: [email protected]
M. Alexander Nugent Consulting, Santa Fe, NM, USA
KnowmTech LLC, Albuquerque, NM, USA
Xeiam LLC, Santa Fe, NM, USA
Email: [email protected]
Abstract—AHaH computing forms a theoretical framework
from which a biologically-inspired type of computing architecture
can be built where, unlike von Neumann systems, memory and
processor are physically combined. In this paper we report
on an incremental step beyond the theoretical framework of
AHaH computing toward the development of a memristorbased physical neural processing unit (NPU), which we call
Thermodynamic-RAM (kT-RAM). While the power consumption
and speed dominance of such an NPU over von Neumann
architectures for machine learning applications is well appreciated, Thermodynamic-RAM offers several advantages over other
hardware approaches to adaptation and learning. Benefits include
general-purpose use, a simple yet flexible instruction set and easy
integration into existing digital platforms. We present a high level
design of kT-RAM and a formal definition of its instruction set.
We report the completion of a kT-RAM emulator and the successful port of all previous machine learning benchmark applications
including unsupervised clustering, supervised and unsupervised
classification, complex signal prediction, unsupervised robotic
actuation and combinatorial optimization. Lastly, we extend a
previous MNIST hand written digits benchmark application, to
show that an extra step of reading the synaptic states of AHaH
nodes during the train phase (healing) alone results in plasticity
that improves the classifier’s performance, bumping our best F1
score up to 99.5%.
I. I NTRODUCTION
Because of its superior power, size and speed advantages
over traditional von Neumann computing architectures, the
biological cortex is one of Nature’s structures that humans are
trying to understand and recreate in silicon form [1]–[4]. The
cognitive ability of causal understanding that animals possess
along with other higher functions such as motor control,
reasoning, perception and planning are credited to different
anatomical structures in the brain and cortex across animal
groups [5]. While the overall architecture and neural topology
of the cortex across the animal kingdom may be different, the
common denominator is the neuron and its connections via
synapses.
In this paper, we present a general-purpose neural processing unit (NPU) hardware circuit we call ThermodynamicRAM that provides a physical adaptive computing resource to
enable efficient emulation of functions associated with brains
and cortex. The motivation for this work is to satisfy the need
for a chip where processor and memory are the same physical
substrate so that computations and memory adaptations occur
together in the same location. For machine learning applications that intend to achieve the ultra-low power dissipation
of biological nervous systems, the shuttling of bits back and
forth between processor and memory grows with the size and
complexity of the model. The low-power solution to machine
learning occurs when the memory-processor distance goes to
zero and the voltage necessary for adaption is reduced. This
can only be achieved through intrinsically adaptive hardware.
Other experts in the field have also expressed the need
for this new type of computing. Traversa and Ventra recently
introduced the idea of ‘universal memcomputing machines’,
a general-purpose computing machine that has the same
computational power as a non-deterministic Universal Turing
Machine but also exhibiting intrinsic parallelization and information overhead [6]. Their proposed DCRAM and other
similar solutions employ memristors, elements that exhibit the
capacity for both memory and processing. They show that a
memristor or memcapacitor can be used as a subcomponent for
computation while at the same time storing a unit of data. A
previous study by Thomas et al. also argued that the memristor
can better be used to implement neuromorphic hardware than
traditional CMOS electronics [7]. In recent years there has
been an explosion of worldwide interest in memristive devices
[8], their connection to biological synapses [9], [10], and use
in alternative computing architectures [11].
Thermodynamic-RAM is adaptive hardware operating on
the principles of ‘AHaH computing’ [12], a new computing paradigm where processor and memory are united. The
fundamental building blocks it provides are ‘AHaH nodes’
and ‘AHaH synapses’ analogous to biological neurons and
synapses respectively. An AHaH node is built up from one
or more synapses, which are are implemented as differential memristor pairs. Spike streams asynchronously drive coactivation of synapses, and kT-RAM’s instruction set allows
for specification of supervised or unsupervised adaptive feedback. The co-active synaptic weights are summed on the
AHaH node’s output electrode as an analog sum of currents
providing both a state and a magnitude. While reverse engineering the biological brain was not the starting point for
developing AHaH computing, it is encouraging that the present
design shares similarities with the cortex such as sparse spike
data encoding and the capacity for continuous on-line learning.
Much like a graphical processing unit (GPU) accelerates
graphics, kT-RAM plugs into existing computer architectures
to accelerate machine learning operations. This opens up the
possibility to give computer hardware the ability to perceive
and act on information flows without being explicitly programmed. In neuromorphic systems, there are three main
specifications: the network topology, the synaptic values and
the plasticity of the interconnections or synapses. Any generalpurpose neural processor must contend with the problem that
hard-wired neural topology will restrict the available neural
algorithms that can be run on the processor. In kT-RAM, the
topology is defined in software, and this flexibility allows for
it to be configured for different machine learning applications
requiring different network topologies such as trees, forests,
meshes, and hierarchies. Much like a central processing unit
(CPU) carries out instructions of a computer program to
implement any arbitrary algorithm, kT-RAM is also generalpurpose, in that it does not enforce a specific network topology.
Additionally, a simple instruction set allows for various forms
of synaptic adaptation, each useful within specific contexts. So
far as we currently understand, the kT-RAM instruction set is
capable of supporting universal machine learning functions.
For an exhaustive introduction to AHaH computing, AHaH
nodes, the AHaH rule, and the AHaH node as well as circuit
simulation results, machine learning benchmark applications,
and accompanying open source code, please refer to a previous
paper entitled “AHaH computingFrom Metastable Switches to
Attractors to Machine Learning” [12].
II. M ETHODS
A. Architecture
The architecture of Thermodynamic-RAM presented in this
paper is a particular design that prioritizes flexibility and
general utility above anything else, much in the same way that
a CPU is designed for general purpose use. Different machine
learning applications require different network topologies (the
way in which neurons are connected to each other), and having
a chip that can be configured for any desired network topology
has the broadest general appeal across the field. Our kT-RAM
design borrows heavily from commodity RAM using its form
factor and the row and column address space mapping to
specific bit cells as a basis to build upon. Converting RAM
to kT-RAM requires the following steps: (1) the removal of
the RAM reading circuitry, (2) minor design modifications of
the RAM cells, (3) the addition of memristive synapses to the
RAM cells, (4) the addition of H-Tree circuitry connecting the
synapses (5) and the addition of driving and output sensing
circuitry.
Figure 1 shows an example of what kT-RAM would look
like with its H-Tree sensing node on the top metal layers of
the chip connecting all the underlying synapses located at each
cell in the RAM array. While at first glance it appears like this
architecture leads to one giant AHaH node per chip or core, the
core can be partitioned into smaller AHaH nodes of arbitrary
size by temporally partitioning sub portions of the tree (AHaH
nodes A and B for example). All other synapses that are
not included in the currently-focused AHaH node space are
unaffected by any read or write operations as they are floating.
In other words, so long as it is guaranteed that synapses
assigned to a particular AHaH node are never co-activated
Fig. 1. Thermodynamic-RAM utilizes standard RAM technology for synaptic
activation over a two-dimensional address space (light gray-bordered cells).
The fractal H-Tree wire (red) forms a common electrical node (y) for summing
the synaptic weights of an AHaH node (neuron) and also for providing a
learning/feedback signal. Here an 16 x 16 cell array is shown, but in practice
a much larger array containing many thousands of synapses can be fabricated.
AHaH node temporal partitioning is achieved by addressing different spaces
at different points in time (AHaH nodes A and B).
with other AHaH node partitions, these ‘virtual’ AHaH nodes
can co-exist on the same physical core. Any desired network
topology linking AHaH nodes together can be achieved by this
temporal partitioning concept by utilizing standard RAM to
store network topologies and core partitions. Software enforces
the constraints, while the hardware remains flexible. It is worth
mentioning that many real-world applications of machine
learning do not cleanly fit within the idealized structure of
a pure network. Rather, mixed computation and algorithms
are often used. By pairing kT-RAM with a standard CPU and
RAM, enormous flexibility in use is possible.
Figure 2 shows a model of an AHaH node as implemented
with switches and memristors. The AHaH node has one common node linking n synapses, with k active inputs and n − k
inactive (floating) inputs at any given time. The individual
gated switches for activating spike inputs are labeled S0 , S1 ,
· · · Sn . Each synapse is made up of two serially-connected
memristors labeled M0 , M1 , · · · Mn . The driving voltage
source for supervised and unsupervised learning is labeled F.
The subscript values a and b indicate the positive and negative
dissipative pathways, respectively.
B. Thermodynamic-RAM Emulator
The substantial cost of hardware, let alone the cost of
new hybrid memristor-CMOS hardware, provides tremendous
inhibitory pressure against the realization of an NPU. On the
Application developers can commence with building a market for kT-RAM on existing hardware platforms (smart phones,
servers, etc.) while hardware developers can work to build
more accurate and detailed emulators. Importantly, hardware
developers can look to the application developers to see where
the application spaces actually are and what is, and is not,
useful in the real world. Although memristive kT-RAM is the
long term goal, multiple generations of pure digital kT-RAM
emulators can be built, each offering various trade-offs in
terms of bandwidth, memory and power. Through adherence to
the kT-RAM instruction set, programs can be ported from one
technology generation to another and multiple industry participants can coordinate their actions across the whole technology
stack. Developments at the hardware level can be informed
by the application level, where utility is demonstrated, and
innovations at the hardware level can propagate back to the
application level.
C. Thermodynamic-RAM Instruction Set and Operation
Fig. 2. An AHaH node is the basic building block of AHaH computing.
Thermodynamic-RAM provides a physical substrate for forming AHaH nodes
(neurons) made up of serially-connected memristor pairs (synapses). The
individual neural inputs corresponding to an incoming sparse-encoded spike
stream are activated using standard RAM address mapping. The common
electrode labeled ‘y’ serves as both a node for summing the weights of the
activated synapses and also for delivering learning feedback to the synapses.
one hand an NPU design needs to justify its existence by
demonstrating utility across one or more application spaces.
On the other hand, one cannot develop applications without the
hardware. We believe we have found a solution to this chickenand-egg problem in the form of a digital kT-RAM emulator.
The AHaH node is very simple, and we have exploited this
simplicity to create an efficient digital emulator.
The digital emulator emulates simplified AHaH nodes
where the memristors are represented as having a set of incremental conductance states. The ‘nibble core’, for example,
treats each memristor as having 16 discrete states and can store
a synapse with one byte. The ‘byte core’ models memristors
as having 256 discrete states and can store a synapse with
two bytes. The ‘float core’ uses floating-point precision and
provides a point of correspondence to our analog emulator. The
analog emulator is more computationally expensive to run as it
emulates a circuit containing memristors which behave according to our ‘generalized metastable switch memristor model’
[12]. The memristors which we have previously applied our
model to include the Ag-chalcogenide [13], AIST [14], GST
[10], and WOx [15] devices. We have not changed the model’s
parameters during the integration of the memristors into the
analog emulator since our previous study [12]. Node voltages
are solved using Kirchoff’s Current Law (KCL). RC delays
and capacitive effects of switching have been ignored for now.
It is worth noting that all above memristor devices have proven
to produce similar performance scores on the benchmarks
amenable to using the analog emulator. In general, as long
as the memristor is incremental and its state change is voltage
dependent, it should be a viable candidate for kT-RAM.
As stated previously, Thermodynamic-RAM can be plugged
into existing digital computing architectures. The envisioned
hardware format is congruent with standard RAM chips and
RAM modules and would plug into a motherboard in a variety
of different ways. In general there are two main categories of
integration. First, kT-RAM is tightly coupled with the CPU,
on the CPU die itself or connected via the north bridge. In this
case, the instruction set of the CPU would have to be modified
to accommodate the new capabilities of kT-RAM. The CPU
would communicate with the kT-RAM module in a similar
manner as with normal RAM, over an address and data bus.
Whereas normal RAM returns a block of bits, kT-RAM would
return a digitized value of the AHaH node’s analog voltage
on electrode y, shown in Figure 2. Alternatively, kT-RAM is
loosely coupled as a peripheral device either connected via the
PCI bus, the LPC bus, or via cables or ports to the south bridge.
In these cases, no modification to the CPU’s instruction set
would be necessary, as the interfacing would be implemented
over the generic plug-in points of the south bus. As in the
case with other peripheral devices, a device driver would need
to be developed. Additional integration configurations are also
possible.
Given the above hardware integration, kT-RAM simply
becomes an additional resource that software developers have
access to via an API. Currently kT-RAM is implemented as an
emulator running on von Neumann architecture. Later, when
a chip is available, it will replace the emulator, and existing
programs written against the API will not need to be rewritten
to benefit from the accelerated capabilities offered by the
new hardware. In any case, kT-RAM or emulated kT-RAM
operates asynchronously. As new spikes arrive, the driver in
control of kT-RAM is responsible for activating the correct
synapses and providing an instruction pair for each AHaH
node. The returned activation value can then be passed back
to the program and used as needed.
The following steps are carried out by the system to
process spike streams: (1) sequentially load all active synapses
TABLE I
K T-RAM I NSTRUCTION
1:
S ET
Instruction
Synapse Driving Voltage
Feedback Voltage (F)
FF
Forward-Float
None/Floating
FH
Forward-High
−V
FL
Forward-Low
+V
FU
Forward-Unsupervised
−V if y ≥ 0 else +V
FA
Forward-Anti-Unsupervised
+V if y ≥ 0 else −V
FZ
Forward-Zero
0
RF
Reverse-Float
None/Floating
RH
Reverse-High
−V
RL
Reverse-Low
+V
RU
Reverse-Unsupervised
−V if y ≥ 0 else +V
RA
Reverse-Anti-Unsupervised
+V if y ≥ 0 else −V
RZ
Reverse-Zero
0
corresponding to the spike pattern, (2) provide an instruction
or instruction pair (read-write), (3) kT-RAM executes the supplied read instruction, (4) receive the AHaH node’s activation
value (confidence) from kT-RAM and (5) kT-RAM executes
the supplied write instruction (if provided).
One can theoretically exploit the kT-RAM instruction set
(Table I) however they wish. However, to prevent saturation
of the memristors in a maximally- or minimally-conductive
state, one must pair ‘forward’ instructions with ‘reverse’
instructions. For example, a forward-read operation F F should
be followed by a reverse operation (RF , RH, RL, RZ, RA or
RU ) and vise versa. The only way to extract state information
is to leave the feedback voltage floating, and thus there are
two possible read instructions: F F and RF . There is no
such thing as a ‘non-destructive read’ operation in kT-RAM.
Every memory access results in weight adaptation according to
AHaH plasticity. By understanding how the AHaH rule works
(AHaH computing), we can exploit the weight adaptations to
create, among other things, ‘self-healing hardware’.
D. Thermodynamic-RAM Classifier
Figure 3 contains pseudo code demonstrating how to construct a multi-label online classifier in software by loading
spikes and executing instructions in the kT-RAM instruction
set. The network topology of the classifier is simply N AHaH
nodes with M synapses, where N is the number of labels
being classified and M is the number of unique spikes in the
entire spike stream space. The active spikes S, a subset of M ,
is loaded onto each AHaH, and the execute method returns the
voltage on the AHaH node’s output electrode, y. Although all
the AHaH nodes may exist on the same physical chip and share
the same output electrode, temporal partitioning, as described
above, allows for a virtual separation of AHaH nodes.
E. MNIST Benchmark
The Mixed National Institute of Standards and Technology
(MNIST) database [16] is a classic dataset in the machine
learning community. It consists of 60,000 train and 10,000
test samples of handwritten digits, each containing a digit 0
2:
3:
4:
5:
6:
7:
8:
9:
10:
11:
12:
13:
14:
15:
16:
17:
procedure C LASSIFY(active spikes set S, truth labels set
L)
for Each AHaH Node N do
KTRAM.loadSpikes(N , S)
y ← KTRAM.execute(N , F F )
. forward read
if supervised then
if N ∈ L then
KTRAM.execute(N , RH)
else if y ≥ 0 then
. false-positive
KTRAM.execute(N , RL)
else
. true-negative
KTRAM.execute(N , RF )
end if
else
. unsupervised
KTRAM.execute(N , RF )
end if
end for
end procedure
Fig. 3. A multi-label online linear classifier with confidence estimation can
be easily constructed via calls to the kT-RAM instruction set.
to 9 (10 classes). The 28 x 28 pixel grayscale images have
been preprocessed to size-normalize and center the digits.
Our approach to the MNIST benchmark is to couple decision trees acting as spike encoders to a back-end classifier. We have ported multiple variants of adaptive and nonadaptive decision trees, as well as the classifier, to the kTRAM emulator. For the work presented here we use both a
simple non-adaptive and efficient decision tree to reduce the
computational load as well as the same adaptive decision tree
used in [12]. Our choice of the non-adaptive random decision
tree is also motivated by the desire to isolate adaptation only to
the classifier while studying the effects of healing. Previously,
AHaH attractor states have been shown to be computationally
complete [12], which means that a decision tree for spike
encoding formed of AHaH nodes in various configurations is
capable of performing a very large number of possible feature
encodings. We have only just begun exploring the space.
The classification procedure is to (1) spike-encode the data
and (2) perform a multi-label classification of the resulting
spike stream. The mechanisms of the spike encoding strongly
effect the resulting classification performance, both in terms of
error rate but also computational efficiency. The optimal spike
encoding method is determined by the problem constraints.
The encoding can be done automatically or manually. The
former entails feature learning or selection, and the latter
entails feature engineering.
We first create binary representations of each image by
thresholding pixels exceeding a value of ten. The output of
this thresholding over an image patch of 8 x 8 pixels produces
a spike stream with a space of 64 channels. This spike stream
is in turn fed into a simple random spike decision tree, where
each node in the tree is looking for the presence of a spike
on one spike channel. The leaves of the decision tree encode
features, and these ‘feature spike streams’ were joined with
pooling coordinates to produce a final output spike stream that
was fed to the classifier. Each spike in the final spike stream
represents the presence of a feature in a non-overlapping
pooling region (8 x 8 pixels). This architecture could be
compared roughly to a simple convolutional neural network.
The output of the classifier produces a list of AHaH node
‘activations’ in the form of voltages. The more positive the
voltage, the more likely that it is correct. By increasing
a confidence threshold, one can increase the classification
precision at the expense of reducing recall. Since the choice
of confidence threshold is arbitrary, we report here the peak
F1 score.
In the absence of any training labels, each AHaH node in
the classifier receives an F F −RF instruction sequence, which
amounts to a forward voltage to read the node state followed
by a reverse voltage without feedback. This instruction sequence is carried out during the train phase of the classification
and the RF operation is necessary to prevent the memristors
from saturating. Depending on the initial state of the synapses
and the evaluation voltage, the synapses will slightly adapt,
meaning there is no possible way to perform a non-destructive
read operation. While one might assume that a read operation
which adapts the memristors’ state is detrimental, this adaptation actually improves the classifier’s performance indicating
that a slight dose of online unsupervised learning occurs by
just reading.
This mechanism can be taken advantage of to further
improve the results of the MNIST classification benchmark by
performing a series of unsupervised classifications on subsets
of the original spike stream during the training phase (healing).
We extended the baseline classifier procedure as shown in Figure 3 so that during the train phase an extra classification was
performed. For each training exemplar, the base supervised
classification was performed followed by another classification
with a randomly-chosen subset of the original spike stream’s
activated spikes. We varied this amount from 0 to 100%. We
also tested two variations of this, where the spike pattern
subset was reclassified either via ‘unsupervised’ (F F −RF ) or
‘supervised’ (see Figure 3). Each configuration was repeated
10 times, and each experiment included 3 training epochs over
a reduced set of 10,000 training and 1,000 test images. The
reduced train and test set was motivated by the need to speed
up total experiment time. As stated above, a non-adaptive spike
encoder was used for this experiment in order to isolate the
adaptive healing effects to only the classifier.
As a final experiment, we repeated our original MNIST
classification benchmark with the full train and test image set,
an adaptive spike encoder, and the new healing method.
signal prediction challenge, Breast Cancer Wisconsin, Reuters21578 Distribution 1.0, and Census Income classification
benchmarks [17]. All benchmarks and challenges performed
roughly the same as before using the kT-RAM emulator with
the various core types. As expected, the low-resolution nibble
and byte cores, while extremely fast, cause slight degradation
of performance. This trade off was observed with the higher
resolution core types as well, where their performance was
better at the cost of longer processing times. The analog
core, which models a particular Ag-chalcogenide memristor
from Boise State University showed excellent congruence with
previous benchmarks, and also exhibited the expected speed
degradation given the extra complexity of the model.
III. R ESULTS AND D ISCUSSION
In this paper, we have proposed one possible hardware implementation of AHaH computing, which we call
Thermodynamic-RAM or kT-RAM for short. While a detailed
description of the chip design is beyond the scope of this
paper, an overview of how standard RAM can be modified to
create kT-RAM was given. Additionally, we described how kTRAM can be plugged into existing digital computing platforms
Here, we report that all the benchmarks and challenges from
the previous AHaH computing publication [12] were successfully re-evaluated using the kT-RAM emulator. The complete
list includes visual clustering challenges, unsupervised robotic
arm challenge, 64-city traveling-salesman problem, complex
A. MNIST Benchmark
As shown in Figure 4, the baseline classification F1 score
for the non-adaptive spike encoder, without healing, and with
10,000 training images and 1,000 test images was 0.967. The
degradation of F1 score is due to the reduced train-test set. It
can been seen that reclassifying a subset of each exemplar’s
spike set during the train phase (healing) improves the results
beyond the baseline. Furthermore, the unsupervised version
outperformed the supervised version. In general a percentage
of active spikes used for the reclassification of 30% to 70%
gave the best results. The kT-RAM classifier is therefore
capable of self-optimizing its performance in an unsupervised
way. Furthermore, this demonstrates that read operations do
not adversely affect the classification performance but rather
improve them. In other words, there is no need for a nondestructive read operation, which is commonly assumed by
circuit designers. This opens up the possibility of ultra-low
power adaptive circuitry, since adaptation can theoretically
occur at low voltages and the circuit can heal itself from the
damage that will occur due to thermal fluctuations.
Repeating the MNIST classification benchmark on the full
train and test dataset (60,000 and 10,000 images respectively)
using our new kT-RAM emulator, the original adaptive spike
encoder and healing yielded better performance than our
previous results of 0.98 - 0.99 F1 score [12]. Using the
adaptive encoder of [12] over 20 training epochs yielded an
F1 score of 0.995, taking approximately 2.5 hours to train
running on a 2012 MacBook Pro. These results show that
our port of AHaH computing functional code over to our
new kT-RAM emulator code did not degrade the outcome of
the classifier. Just adding the healing mechanism significantly
improved results. In addition, the experiment time actually
decreased, indicating that our emulator is very efficient.
IV. C ONCLUSION
nodes (healing) alone results in plasticity that improves the
classifier’s performance. Our previous best F1 score on the
MNIST classification benchmark has been improved to 0.995.
Note that it is not our intention to produce the absolute best
MNIST benchmark performance and compete with established
machine learning techniques. We are simply choosing a wide
range of machine learning benchmarks and challenges to show
utility of our proposed NPU. The advantage will be low-power
adaptive hardware with a consistent API that can be used to
solve numerous problems across the wide spectrum of the
machine learning application space. What we are attempting to
present is therefore not just a specific solution to a benchmark
problem but rather that kT-RAM offers a path toward a new
type of adaptive generic computing substrate.
V. F UTURE W ORK
Fig. 4. For each training exemplar in the MNIST classification benchmark,
a non-adaptive spike encoder was used to generate a spike encoded input
to the kT-RAM classifier. Random subsets of the original spike-encoded
input (from 0% to 95%) were generated from the original spike pattern
and reclassified (healing), either unsupervised (purple, dashed) or supervised
(orange). Average baseline F1 score for 3 epochs was 0.967 (blue, dotted).
The results demonstrate that the act of reading the synaptic states of AHaH
nodes results in plasticity that improves the classifier’s performance on the
test set.
in a similar way as RAM and RAM modules. Whether kTRAM is integrated directly into a CPU’s die, or it is off-board
as a peripheral device or anything in between, the kT-RAM
instruction set provides a minimal and complete interface
for integrating the chip into existing systems for accelerating
machine learning tasks where large scale adaptive networks
become a bottleneck in typical von Neumann architecture.
Writing software to utilize kT-RAM will only require the
addition of the kT-RAM API. Our kT-RAM emulator allows us
to develop applications, demonstrate utility, and justify a large
investment into chip development. When chips are available,
existing applications using the emulator API will not have
to be rewritten in order to take advantage of new hardware
acceleration capabilities.
In addition, we report that we have successfully ported all of
our previous examples of AHaH machine learning capabilities
to use our new kT-RAM emulator including classification,
prediction, clustering, robotic control, and combinatorial optimization. All of these capabilities are are associated with
biological cortex. The computing architecture in both cortex
and kT-RAM is one in which the memory and processor
are the same physical substrate. The temporal partitioning
feature of the kT-RAM design presented allows for a general
purpose chip where the network topology is configured in
software rather than being ‘hard coded’ in the chip’s circuitry.
The software procedure for implementing a classifier with
kT-RAM and the instruction set was given as pseudo code,
and it shows how the spike streams and truth labels (for
supervised learning) are processed. Extending our previous
MNIST hand written digit classification benchmark, it was
shown that the act of reading the synaptic states of AHaH
The classifier demonstration in this paper represents just one
network topology and one basic capability of the cortex, but as
mentioned, all the machine learning capabilities reported in a
previous publication [12] have been successfully ported over
to the kT-RAM emulator. Different topologies implemented
by different procedures and via the described AHaH node
temporal partitioning can be utilized for a wide range of
cortical capabilities. Because the network topology is defined
in software and not ‘hard-coded’ in kT-RAM circuitry, any
topology can be created, explored and tested. This flexibility
reminds one of the advantages and attractiveness of the CPU
- it is a jack of all trades and master of none. This is not to
say that kT-RAM could not be redesigned to an applicationspecific version, just as ASICs are optimized circuits that
are designed to do one thing well. Substantial work remains in hardware design and fabrication and more generally
in further developing AHaH computing. The simplicity of
Thermodynamic-RAM, combined with the fact that we have
demonstrated a number of machine learning capabilities leads
us to conclude that this work will be well worth the effort.
ACKNOWLEDGMENT
The authors would like to thank the Air Force Research
Labs in Rome, NY for their support under the SBIR/STTR
programs AF10-BT31, AF121-049.
R EFERENCES
[1] B. McCormick, “Applying cognitive memory to cybersecurity,” in Network Science and Cybersecurity. Springer, 2014, pp. 63–73.
[2] J. Navaridas, S. Furber, J. Garside, X. Jin, M. Khan, D. Lester, M. Luján,
J. Miguel-Alonso, E. Painkras, C. Patterson et al., “SpiNNaker: Fault
tolerance in a power-and area-constrained large-scale neuromimetic
architecture,” Parallel Computing, vol. 39, no. 11, pp. 693–708, 2013.
[3] B. V. Benjamin, P. Gao, E. McQuinn, S. Choudhary, A. R. Chandrasekaran, J. Bussat, R. Alvarez-Icaza, J. Arthur, P. Merolla, and
K. Boahen, “Neurogrid: A mixed-analog-digital multichip system for
large-scale neural simulations,” Proceedings of the IEEE, vol. 102, no. 5,
pp. 699–716, 2014.
[4] S. K. Esser, A. Andreopoulos, R. Appuswamy, P. Datta, D. Barch,
A. Amir, J. Arthur, A. Cassidy, M. Flickner, P. Merolla et al., “Cognitive
computing systems: Algorithms and applications for networks of neurosynaptic cores,” in Neural Networks (IJCNN), The 2013 International
Joint Conference on. IEEE, 2013, pp. 1–10.
[5] J. Dugas-Ford, J. J. Rowell, and C. W. Ragsdale, “Cell-type homologies
and the origins of the neocortex,” Proceedings of the National Academy
of Sciences, vol. 109, no. 42, pp. 16 974–16 979, 2012.
[6] F. L. Traversa and M. Di Ventra, “Universal memcomputing machines,”
arXiv preprint arXiv:1405.0931, 2014.
[7] A. Thomas, “Memristor-based neural networks,” Journal of Physics D:
Applied Physics, vol. 46, no. 9, p. 093001, 2013.
[8] T. Prodromakis and C. Toumazou, “A review on memristive devices and
applications,” in Electronics, Circuits, and Systems (ICECS), 2010 17th
IEEE International Conference on. IEEE, 2010, pp. 934–937.
[9] T. Chang, S.-H. Jo, and W. Lu, “Short-term memory to long-term
memory transition in a nanoscale memristor,” ACS Nano, vol. 5, no. 9,
pp. 7669–7676, 2011.
[10] Y. Li, Y. Zhong, L. Xu, J. Zhang, X. Xu, H. Sun, and X. Miao, “Ultrafast
synaptic events in a chalcogenide memristor,” Scientific Reports, vol. 3
doi:10.1038/srep01619, no. 1619, 2013.
[11] F. C. Morabito, A. G. Andreou, and E. Chicca, “Neuromorphic engineering: from neural systems to brain-like engineered systems.” Neural
Networks, vol. 45, pp. 1–3, 2013.
[12] M. A. Nugent and M. T. W, “Ahah computing-from metastable switches
to attractors to machine learning,” PLoS ONE, vol. 9, p. e85175, 02 2014.
[13] A. S. Oblea, A. Timilsina, D. Moore, and K. A. Campbell, “Silver
chalcogenide based memristor devices,” in Proc. 2010 IEEE International Joint Conference on Neural Networks (IJCNN), 2010, pp. 1–3.
[14] J. J. Zhang, H. J. Sun, Y. Li, Q. Wang, X. H. Xu, and X. S. Miao,
“AgInSbTe memristor with gradual resistance tuning,” Applied Physics
Letters, vol. 102, no. 18, p. 183513, 2013.
[15] T. Chang, S.-H. Jo, K.-H. Kim, P. Sheridan, S. Gaba, and W. Lu,
“Synaptic behaviors and modeling of a metal oxide memristive device,”
Applied Physics A, vol. 102, no. 4, pp. 857–863, 2011.
[16] Y. LeCun and C. Cortes, “The mnist database of handwritten digits,”
1998.
[17] K. Bache and M. Lichman, “UCI machine learning repository,” 2013.
[Online]. Available: http://archive.ics.uci.edu/ml
| 9cs.NE
|
arXiv:1604.08435v1 [math.AC] 28 Apr 2016
Dissertation
zur Erlangung des Doktorgrades (Dr. rer. nat.)
des Fachbereichs Mathematik/Informatik
der Universität Osnabrück
Hilbert-Kunz functions of surface rings of type
ADE
vorgelegt
von
Daniel Brinkmann
Osnabrück
eingereicht: Mai 2013
veröffentlicht: August 2013
Betreuer: Prof. Dr. Holger Brenner
Contents
Introduction
The goal of this thesis and a sketch of the most important ideas
Summary of the thesis
Acknowledgements
Notations and conventions
5
5
10
15
15
Chapter 1. A survey on Hilbert-Kunz theory
1.1. The algebraic viewpoint of Hilbert-Kunz theory
1.2. The geometric viewpoint of Hilbert-Kunz theory
1.2.1. Generalities on vector bundles
1.2.2. Frobenius
1.2.3. Syzygy bundles
1.2.4. Stability
1.2.5. Relation to Hilbert-Kunz theory
1.2.6. Curves of degree three
1.3. Related invariants and further readings
1.3.1. Limit Hilbert-Kunz multiplicity
1.3.2. F-Signature
1.3.3. Further readings
17
17
31
31
33
34
37
40
42
44
44
46
46
Chapter 2. Examples of Hilbert-Kunz functions and Han’s δ function
2.1. Examples of Hilbert-Kunz functions
2.2. Han’s δ function
2.3. The behaviour of the Hilbert-Kunz multiplicity in families
2.3.1. The family F1
2.3.2. The family FU
2.3.3. The family FV
2.3.4. The family FW
49
49
55
67
68
70
72
73
Chapter 3. Matrix factorizations and representations as first syzygy modules
3.1. Properties of maximal Cohen-Macaulay modules
3.1.1. Maximal Cohen-Macaulay modules and Hilbert-Kunz theory
3.2. Generalities on matrix factorizations
3.3. Constructing isomorphisms of maximal Cohen-Macaulay modules
3.4. The case An with n ≥ 0
3.5. The case Dn with n ≥ 4
3.6. The case E6
3.7. The case E7
77
77
79
83
88
91
92
95
99
3
4
CONTENTS
3.8. The case E8
3.9. An idea of a purely algebraic construction of the isomorphisms
3.10. A comment on the graded situation
102
106
108
Chapter 4. The Hilbert-series of first syzygy modules of monomial ideals
4.1. A short exact sequence
4.2. Examples
4.2.1. The case An
4.2.2. The case Dn
4.2.3. The case E6
4.2.4. The case E7
4.2.5. The case E8
111
111
115
116
116
118
119
120
Chapter 5. Relations between the maximal Cohen-Macaulay modules
5.1. The Case An−1 with n ≥ 1
5.2. The Case Dn+2 with n ≥ 2
5.2.1. Pull-backs to At−1 with t|2n
5.2.2. Pull-backs to Dn/2+2
5.3. The Case E6
5.4. The Case E7
5.5. The Case E8
5.6. Overview
123
125
126
127
128
129
133
135
136
Chapter 6. The Hilbert-Kunz functions of surface rings of type ADE
6.1. The case An−1 with n ≥ 1
6.2. The case Dn+2 with n ≥ 2
6.3. The case E6
6.4. The case E7
6.5. The case E8
139
139
140
142
143
147
Chapter 7. Extensions and further examples
7.1. Hilbert-Kunz functions of E8 with respect to various ideals I , m
7.2. The Hilbert-Kunz function of A∞
7.3. The Hilbert-Kunz function of D∞
7.4. On the Hilbert-Kunz functions of two-dimensional Fermat rings
7.5. Open questions
7.5.1. Questions concerning rings of dimension two
7.5.2. Questions concerning rings of dimension at least three
151
151
153
155
158
166
166
169
Appendix A.
171
An implementation of Corollary 2.31 for CoCoA
Appendix B. Computer computations to Chapter 5
B.1. Normal subgroups with GAP and rings of invariants with Singular
B.2. Macaulay2 computations in the E6 -case
B.3. Macaulay2 computations in the E8 -case
175
175
177
178
Bibliography
187
Introduction
The goal of this thesis and a sketch of the most important ideas
The starting point for Hilbert-Kunz theory was the paper “Characterizations of regular
local rings of characteristic p” and its sequel “On Noetherian rings of characteristic p”
by Kunz published 1969 resp. 1976 (cf. [Kun69] and [Kun76]). The central objects are
the Hilbert-Kunz function and the Hilbert-Kunz multiplicity, which are defined in the case
of affine k-algebras R := k[X1 , . . . , Xn ]/I in dependence on the classical Hilbert-Samuel
multiplicity and function as
pe
e
HK(R, pe ) = dimk R/ X1 , . . . , Xnp
and
HK(R, pe )
,
e→∞ pe·dim(R)
eHK (R) = lim
where k denotes a field of characteristic p > 0 and dim(R) denotes the Krull dimension of
R. For an (X1 , . . . , Xn )-primary ideal J := ( f1 , . . . , fm ), the Hilbert-Kunz function of R with
respect to J and the Hilbert-Kunz multiplicity of R with respect to J are defined as
pe
e
HK(J, R, pe ) = dimk R/ f1 , . . . , fmp
resp.
HK(J, R, pe )
.
e→∞ pe·dim(R)
eHK (R) = lim
Seven years later, Monsky proved the existence of Hilbert-Kunz multiplicities in a more
general situation (cf. [Mon83a, Theorem 1.8]). In the same paper the term “Hilbert-Kunz
function” firstly appeared. Hilbert-Kunz theory has applications in Iwasawa theory as
well as in tight closure theory (see for example [Mon83b] and [Hun96]). Initially, Kunz
expected a relation to the resolution of singularities in positive characteristics (cf. the
discussion after Example 4.3 of [Kun76]) but up to today there are no substantial results
in this direction. Maybe the most central question in Hilbert-Kunz theory was whether
Hilbert-Kunz multiplicities have to be rational or not. For long time the favourite answer
was that they have to be rational (cf. [Mon83a]) and there were many results in this
direction (e.g. [Bre06], [Eto02], [FT03], [HM93], [Sei97], [Tri05a] or [Wat00]). But
2008 Monsky came up with a conjecture whose correctness would imply the existence
of a four-dimensional ring of characteristic two with irrational Hilbert-Kunz multiplicity
(cf. [Mon08]). Extending the geometric approach to Hilbert-Kunz theory (cf. [Bre06] or
[Tri05a]), Brenner recently proved the existence of irrational Hilbert-Kunz multiplicities
in all characteristics (cf. [Bre13]).
The goal of this thesis is to compute the Hilbert-Kunz functions of the surface rings
• An := k[X, Y, Z]/(X n+1 − YZ) with n ≥ 0,
• Dn := k[X, Y, Z]/(X 2 + Y n−1 + YZ 2 ) with n ≥ 4
• E6 := k[X, Y, Z]/(X 2 + Y 3 + Z 4 ),
5
6
INTRODUCTION
• E7 := k[X, Y, Z]/(X 2 + Y 3 + YZ 3 ) and
• E8 := k[X, Y, Z]/(X 2 + Y 3 + Z 5 ),
where k denotes an algebraically closed field. These are called surface rings of type ADE.
In characteristic zero surface rings of type ADE show up as quotient singularities of C2
and are unique up to isomorphism as it was shown by Klein (cf. [Kle86]) in 1886. This
is also how they have share in string theory (cf. [Wit02]). Surface rings of type ADE
were studied for example by du Val (cf. [dV34a], [dV34b], [dV34c]) and Brieskorn
(cf. [Bri68]) and are sometimes called Kleinian or du Val singularities. Due to Artin
(cf. [Art66]) they are exactly the rational double points. In positive characteristics Artin
classified the rational double points of surfaces. In characteristics at least seven, these are
exactly the surface rings of type ADE. In characteristics two, three and five all rational
double points are given by surface rings of type ADE or certain deformations of them (cf.
[Art77]). Note that the rational double points are exactly the simple surface singularities
(cf. [Art77]).
Due to Watanabe and Yoshida (cf. [WY00]) the Hilbert-Kunz multiplicity of the surface
1
, where G ( SL2 (C) is the group corresponding to
rings of type ADE is known to be 2 − |G|
the singularity in characteristic zero and under the additional assumption that the corresponding ring is F-rational (cf. [WY00]). Note that this assumption is satisfied in almost
all characteristics. The Hilbert-Kunz function of the surface rings of type An is known
due to Kunz (cf. [Kun76]) and the Hilbert-Kunz functions of surface rings of type E6 or
E8 can be computed at least for explicitly given characteristic using the algorithm of Han
and Monsky (cf. [HM93]). In general the Hilbert-Kunz function of surface rings R of
type ADE has due to Huneke, McDermott and Monsky the shape
HK(R, pe ) = eHK (R) · p2e + γ(pe ),
where γ(pe ) is a bounded function (cf. [HMM04]).
By a geometric approach of Brenner and Trivedi for two-dimensional standard-graded
rings S one needs to control the Frobenius pull-backs of the cotangent bundle on C :=
Proj(S ), e.g. the sheaves
[pe ] ∼
,
SyzS S +
to compute the Hilbert-Kunz function of S (cf. [Bre06] or [Tri05a]). Unfortunately,
surface rings of type ADE are not standard-graded. But we can associate to them the
standard-graded rings S := k[U, V, W]/(F), where (F) is the image of the principal ideal
defining a surface ring of type ADE under the map
k[X, Y, Z] → k[U, V, W], X 7→ U deg(X) , Y 7→ V deg(Y) , Z 7→ W deg(Z) .
Since the map R → S is a flat extension of Noetherian domains, one obtains
HK((U deg(X) , V deg(Y) , W deg(Z) ), S , pe )
,
[Q(S ) : Q(R)]
where [Q(S ) : Q(R)] denotes the degree of the extension of the fields of fractions of R
and S . The benefit of this approach is that we can use tools from algebraic geometry to
compute
ÄÄ
ä
ä
HK U deg(X) , V deg(Y) , W deg(Z) , S , pe .
HK(R, pe ) =
THE GOAL OF THIS THESIS AND A SKETCH OF THE MOST IMPORTANT IDEAS
7
The disadvantage of this approach is that Proj(S ) is much more complicated then Proj(R),
which is just P1k . This is because the quotient map A2k → Spec(R) induces a finite map
P1k → Proj(R) of smooth projective curves. By Hurwitz’ formula the curve Proj(R) has
genus zero, hence Proj(R) P1k . To compute
ÄÄ
ä
ä
HK U deg(X) , V deg(Y) , W deg(Z) , S , pe ,
we need to control the OProj(S ) -modules corresponding to the S -modules
Ä
e
eä
e
Ne := SyzS U deg(X)·p , V deg(Y)·p , W deg(Z)·p .
To do so, we compute the R-modules
Ä
e
e
eä
Me := SyzR X p , Y p , Z p
and use our map η : R → S as follows.
(i) The map η induces a map ϕ : Spec(S ) \ {S + } → Spec(R) \ {R+ }.
pt
pt
pt
(ii) The pull-back of g
Me along ϕ is just ›
Ne , where g
Me denotes the restriction to
pt
Spec(R) \ {R+ } of the sheaf on Spec(R) associated to Me and analogue for ›
Ne .
pt
pt
(iii) The isomorphism ϕ∗ (g
Me ) ›
Ne extends to a global isomorphism η(Me ) Ne .
But how to compute the Me ?
Since over two-dimensional rings first syzygy modules of homogeneous ideals are graded
maximal Cohen-Macaulay modules, we will make use of the fact that the surface rings
of type ADE are of finite graded Cohen-Macaulay type, meaning that there are up to degree shifts only finitely many isomorphism classes of indecomposable, graded maximal
Cohen-Macaulay modules (cf. [Knö87] and [BGS87]). Representatives for the isomorphism classes are described in [KST07] using Eisenbud’s theory of matrix factorizations
(cf. [Eis80]). In our situation, a matrix factorization is a pair (ϕ, ψ) of n × n matrices with
coefficients in k[X, Y, Z] such that
ϕ · ψ = ψ · ϕ = f · Id,
where f is the generator of the principal ideal defining a surface ring R of type ADE.
Then coker(ϕ) is a graded maximal Cohen-Macaulay R-module. Following the common
strategy for classifications, we need an invariant that detects the isomorphism class of a
given (indecomposable) graded maximal Cohen-Macaulay module and then compute the
invariant for all R-modules Me . The first step is to find “good representatives” for the
finitely many isomorphism classes of indecomposable, graded maximal Cohen-Macaulay
modules. It is not surprising that the phrase “good representatives” means first syzygy
modules of homogeneous ideals in our case. We will use sheaf-theoretic tools to find
isomorphisms of the form
coker(ϕ) SyzR (Iϕ ),
where the Iϕ are homogeneous ideals of R and (ϕ, ψ) runs through a complete list of matrix
factorizations representing the isomorphism classes of indecomposable, graded maximal
Cohen-Macaulay R-modules. To be more precise, we will show (in our situation) that
for every matrix factorization (ϕ, ψ), where ϕ is a n × n-matrix and coker(ϕ) has rank r,
there are r + 1 columns of ψ (!) such that on U := Spec(R) \ {R+ } those r + 1 columns
8
INTRODUCTION
generate im(ψ) and on R they generate SyzR ( f1 , . . . , fr+1 ) for suitable homogeneous fi ∈ R.
Therefore, we have the short exact sequence
/
0
/
( f1 , . . . , fn )∼ |U
‡
/ im(ψ)|
U
Or+1
U
/
0,
showing that the dual module im(ψ)∨ and SyzR ( f1 , . . . , fr+1 ) correspond on U. Since all
rings in question are Gorenstein in codimension one, this isomorphism extends to a global
isomorphism. It remains to observe im(ψ) coker(ϕ).
Since we have only a finite list of isomorphism classes, one might hope that they can be
distinguished by their Hilbert-series. This is not true in general. But we will see that the
Hilbert-series of a given (indecomposable) graded maximal Cohen-Macaulay R-module
detects the isomorphism class of M up to dualizing, by which we mean that there is a
graded maximal Cohen-Macaulay module N, having the same Hilbert-series as M, such
that either M N or M N ∨ . We will prove a theorem which allows us to compute the
Hilbert-series of the modules Me by reducing the computation to the computation of the
Hilbert-series of first syzygy modules of homogeneous ideals in a polynomial ring in only
two variables. The idea to this theorem was inspired by work of Brenner (cf. [Bre05b]).
We will see that the Hilbert-series of Me is enough to detect its isomorphism class (up
to dualizing) if R is of type An or E8 . In the other cases we need to show that the modules Me are indecomposable. This is done by using invariant theory. To be more precise, we will use the facts that the surface rings of type ADE are rings of invariants of
k[x, y] under finite subgroups of SL2 (k) and that normal subgroups H ⊆ G induce maps
ι : Spec(k[x, y]H ) → Spec(k[x, y]G ) resp. on the punctured spectra. The claimed result will
follow by comparing (on the punctured spectra) the pull-backs of g
Me along ι with the
∗
g
Frobenius pull-backs of ι ( M0 ).
From the indecomposability of the modules Me we will obtain a periodicity up to degree
shift of the sequence
Ä
Ä
e
e
SyzR X p , Y p , Z p
e ää
e∈N
by which we mean that there exists an e0 such that
e+e0
SyzR X p
e+e0
,Yp
e+e0
,Zp
Ä
e
eä
e
SyzR X p , Y p , Z p (me )
holds for all e ∈ N and suitable me ∈ Z. This leads to a periodicity up to degree shift of the
sequence
Ä
Ä
e
e
e ää
SyzS U deg(X)·p , V deg(Y)·p , W deg(Z)·p
e∈N
.
Putting everything together, we will be able to compute the Hilbert-Kunz functions of
surface rings of type ADE.
Theorem (cf. Theorem 6.1). The Hilbert-Kunz function of Dn+2 is given by (with n ≥ 2,
q = pe and q ≡ r mod 2n, where r ∈ {0, . . . , 2n − 1})
2
2·q
e 7−→ Ç
2−
å
1
r + 1 r2
q2 −
+
4n
2
4n
if e ≥ 1, p = 2,
otherwise.
THE GOAL OF THIS THESIS AND A SKETCH OF THE MOST IMPORTANT IDEAS
9
Theorem (cf. Theorem 6.3). The Hilbert-Kunz function of E6 is given by
2
2·q
Ç
e 7−→
2−
if e ≥ 1 and p ∈ {2, 3},
å
1
23
q2 −
24
24
otherwise.
Theorem (cf. Theorem 6.5). The Hilbert-Kunz function of E7 is given by
2
2·q
Ç
if e ≥ 1 and p ∈ {2, 3},
å
1
2 47
e 7−→ 2 − 48 q − 48
Ç
å
1
71
q2 −
2−
48
48
if q mod 24 ∈ {±1, ±7},
otherwise.
Theorem (cf. Theorem 6.7). The Hilbert-Kunz function of E8 is given by
2 · q2
Ç
if e ≥ 1 and p ∈ {2, 3, 5},
å
1
2 119
e 7−→ 2 − 120 q − 120
Ç
å
1
191
q2 −
2−
120
120
if q mod 30 ∈ {±1, ±11},
otherwise.
Note that Seibert gave in [Sei97] an algorithm to compute Hilbert-Kunz functions of
maximal Cohen-Macaulay rings and modules. Given a ring R := k[ f1 , . . . , fn ] ⊆ k[x, y],
where k is a field of characteristic p > 0, the first step of this algorithm is to decompose R
into indecomposable R p -modules, which gets very hard if the fi are not monomial. This
is why we did not use this algorithm. For similar reasons it seems not practicable to count
the monomials in k[X, Y, Z]/(F, X q , Y q , Z q ) directly if F is not binomial.
Applying the developed methods in the case of Fermat rings R := k[X, Y, Z]/(X n + Y n + Z n )
and using work of Kaid, Kustin et al. and Trivedi (cf. [Kai09], [KRV12] resp. [Tri05b]),
we will obtain results on the Hilbert-Kunz function of R as well as on the behaviour of the
Frobenius pull-backs of the cotangent bundle on Proj(R). We say that a vector bundle S
admits a (s, t)-Frobenius periodicity if there are s < t ∈ N such that Ft∗ (S) F s∗ (S) holds
up to degree shift.
Theorem (cf. Theorem 7.17). Let C be the projective Fermat curve of degree n over
an algebraically closed field of characteristic p > 0 with gcd(p, n) = 1. The bundle
SyzC (X, Y, Z) admits a Frobenius periodicity if and only if
Ç
å
1 1 1
δ , ,
= 0.
n n n
Moreover, the length of this periodicity is bounded above by the order of p modulo 2n.
Theorem (cf. Theorem 7.18). Let R := k[X, Y, Z]/(X n + Y n + Z n ) with an algebraically
closed field k of characteristic p > 0 with gcd(p, n) = 1. Assume that SyzProj(R) (X, Y, Z)
is strongly semistable and is not trivialized by the Frobenius. Let q = pe = nθ + r with θ,
10
INTRODUCTION
r ∈ N and 0 ≤ r < n. Then
3n · q2 − 3n r2 + r3
if θ is even,
4
HK(R, q) = 4
3n 2 3n
· q − (n − r)2 + (n − r)3 if θ is odd.
4
4
These theorems generalize the result of Brenner and Kaid who obtained them under the
condition p ≡ −1 modulo 2n (cf. [BK13]).
Moreover, the last theorem recovers a result of Han and Monsky who proved that the
Hilbert-Kunz function of a Fermat ring R has the shape eHK (R) · p2e + ∆e , where the function e 7→ ∆e is eventually periodic with non-positive values (cf. [HM93]).
We will give examples showing that in every odd characteristic for every 1 ≤ l there exists
a n ∈ N such that the bundle Syz(X, Y, Z) on the projective Fermat curve of degree n admits a (0, l)-Frobenius periodicity. In characteristic two only (1, 2)-Frobenius periodicities
appear. Moreover, this happens only in the case n = 3.
Summary of the thesis
In the first chapter we will give a survey on Hilbert-Kunz theory. We divide this survey
into three sections. In the first one we give the basic definitions and state various results as
the existence of Hilbert-Kunz multiplicities, bounds for Hilbert-Kunz multiplicities, some
classification results and the connection to tight closure theory. In the second section we
explain the geometric approach to Hilbert-Kunz theory in graded dimension two by Brenner and Trivedi (cf. [Bre06] and [Tri05a]). This section will start with short discussions
of vector bundles, Frobenius maps, syzygy bundles and various stability conditions for
vector bundles. It will end with the classification of Hilbert-Kunz functions of the homogeneous coordinate rings of projective curves of degree three defined over algebraically
closed fields of positive characteristic. In the third section we define the invariants limit
Hilbert-Kunz multiplicity and F-signature which are strongly connected to Hilbert-Kunz
multiplicities. At the end of the first chapter we give a list of further references related to
Hilbert-Kunz theory.
In the first section of the second chapter we compute some explicit examples of HilbertKunz functions with the help of the geometric approach. For example we will compute
(cf. Example 2.6) the Hilbert-Kunz functions of the diagonal hypersurfaces
R := k[X, Y, Z]/(X 2 + Y l + Z m ),
Ä
ä
where l, m are positive integers such that the triple 12 , 1l , m1 does not satisfy the strict
triangle inequality.
The goal of the second section is to generalize a theorem of Monsky which connects the
Hilbert-Kunz multiplicities of standard-graded quotients R := k[X, Y, Z]/(F), where F is a
so-called regular, irreducible trinomial of degree d, with Han’s δ function. In this case
regular means that the trinomial F = MX + MY + MZ satisfies degX (MX ) > d2 , degY (MY ) > d2
and degZ (MZ ) > d2 . Explicitly, Monsky showed
Ç
eHK
3d
λ2
sα sβ sγ
(X , Y , Z ), R = s2 + δ
, ,
4
4d
λ λ λ
s
s
s
å2
,
SUMMARY OF THE THESIS
11
where s is a positive integer and α, β, γ, λ are natural numbers depending only on the
exponents appearing in F (cf. [Mon06a, Theorem 2.3] or Corollary 2.30). Leaning on
Monsky’s proof, we will show that a positively-graded analogue of Han’s δ function as
well as
µ(l1 , . . . , ln ) := eHK x1l1 , . . . , xnln , R
for standard-graded domains R and homogeneous R+ -primary ideals x1l1 , . . . , xnln extend
to continuous functions on [0, . . . , ∞)3 resp. [0, . . . , ∞)n .
Theorem (cf. Theorem 2.29 resp. Corollary 2.31). With the previous notations we have
eHK
ÄÄ
ä
ä
X a , Y b , Z c , k[X, Y, Z]/(F)
Ç Ç
åå
d
α(a, b, c) β(a, b, c) γ(a, b, c) 2
λ2
= · (2ab + 2ac + 2bc − a2 − b2 − c2 ) +
δ
,
,
,
4
4d
λ
λ
λ
where λ is a positive integer depending only on the exponents appearing in F and the
abbreviations α(a, b, c), β(a, b, c), γ(a, b, c) denote natural numbers depending on a, b, c
and the exponents in F.
An explicit implementation of this Theorem for CoCoA can be found in the Appendix.
Using a result of Watanabe and Yoshida (cf. [WY00, Theorem 2.7] or Theorem 1.21) on
the behaviour of Hilbert-Kunz multiplicities under module-finite extensions of Noetherian local domains, the above result enables us to compute Hilbert-Kunz multiplicities of
positively-graded quotients k[U, V, W]/(G) such that the image of G under the map
k[U, V, W] −→ k[X, Y, Z], U 7→ X deg(U) , V 7→ Y deg(V) , W 7→ Z deg(W)
is a regular, irreducible trinomial (cf. Remark 2.34). In particular, we recover Han’s result
on the Hilbert-Kunz multiplicities of two-dimensional diagonal hypersurfaces
Ä
ä
k[U, V, W]/ U d1 + V d2 + W d3 ,
with 2 ≤ di ∈ N (cf. [Han91, Theorem 2.30] or Corollary 2.36) and the result of Watanabe
and Yoshida on the Hilbert-Kunz multiplicities of the surface rings of type ADE (cf.
[WY00, Theorem 5.4], Theorem 1.26 or Example 2.38).
In the third section of the second chapter we use Theorem 2.29 resp. Remark 2.34 to
study the behaviour of the Hilbert-Kunz multiplicity in certain families of positivelygraded quotients k[U, V, W]/(G) of the form described above. Writing the trinomial G as
G = MU + MV + MW , the families are definied by the condition that one of the exponents
degU (MU ), degV (MV ), degW (MW ) is free and all other exponents in G are fixed. For
a fixed family we will study the Hilbert-Kunz multiplicity for large values of the free
exponent. We will see that in all cases the sequence of the Hilbert-Kunz multiplicities
in such a family converges to the Hilbert-Kunz multiplicity of a binomial quotient ring,
where the binomial is the deformation of the trinomial G obtained by leaving out the
monomial with the free exponent.
In the first section of Chapter 3 we give a brief overview of properties of maximal CohenMacaulay modules including a proof of the fact that the concepts maximal Cohen-Macaulay and reflexive coincide in normal domains of dimension two. We will demonstrate how
12
INTRODUCTION
one might compute Hilbert-Kunz functions of two-dimensional rings of finite CohenMacaulay type with the help of the geometric approach. The example will be the n-th
Veronese subring R of k[x, y], where k denotes an algebraically closed field of characteristic p > 0. We obtain (cf. Example 3.14)
n + 1 Ä 2e 2 ä
r(r + 1)
· p −r +n·
+ r − n,
2
2
where r is the smallest non-negative representative of the class of pe in Z/(n). During this
computation we will give a complete list of first syzygy modules of ideals representing
the isomorphism classes of maximal Cohen-Macaulay R-modules, show how SyzR (R+ )
decomposes in this choice of representatives and explain the action of the Frobenius morphism on the Picard group.
In Section 3.2 we will explain the notion of matrix factorizations developed by Eisenbud
in [Eis80] and their connection to maximal Cohen-Macaulay modules over rings of the
form S /( f ), where (S , m) is a local regular Noetherian ring and f ∈ m2 \ {0}.
In Section 3.3 we will develop a sheaf-theoretic approach to compute first syzygy modules
of ideals that are isomorphic to the maximal Cohen-Macaulay module defined by a fixed
matrix factorization. For this approach we have to restrict to two-dimensional local rings
which are Gorenstein in codimension one.
In Sections 3.4-3.8 we use the approach from Section 3.3 to obtain a system of first syzygy
modules of ideals representing the isomorphism classes of the indecomposable, maximal
Cohen-Macaulay modules over the completions of the surface rings of type ADE. Each
section ends with a theorem gathering the following informations.
HK(R, pe ) =
• A complete list of non-isomorphic first syzygy modules of ideals representing the isomorphism classes of non-free, indecomposable, maximal CohenMacaulay modules.
• Representatives of the isomorphism classes of the dual modules.
• The isomorphism classes of the corresponding determinant bundles on the punctured spectrum.
• Ideals representing the isomorphism classes of the non-free, indecomposable,
maximal Cohen-Macaulay modules of rank one.
For example if R is the (X, Y, Z)-adic completion of E6 and U its punctured spectrum, the
corresponding theorem is the following.
Theorem (cf. Theorem 3.40). The pairwise non-isomorphic modules
M1 = SyzR (X, Y, Z),
M2 = SyzR (X, Y 2 , YZ, Z 2 ),
M3 = SyzR (iX + Z 2 , Y 2 , YZ),
M4 = SyzR (−iX + Z 2 , Y 2 , YZ),
M5 = SyzR (−iX + Z 2 , Y),
M6 = SyzR (iX + Z 2 , Y)
SUMMARY OF THE THESIS
13
give a complete list of representatives of the isomorphism classes of indecomposable, nonfree, maximal Cohen-Macaulay modules. Moreover, M1 and M2 are selfdual, M3∨ M4
M3 |U resp. g
M4 |U are g
M5 |U resp. g
M6 |U . For the rank
and M5∨ M6 . The Determinants of g
2
one modules we have the isomorphisms M5 (iX + Z , Y) and M6 (−iX + Z 2 , Y).
In Section 3.9 we discuss an idea how to avoid sheaves in the approach of Section 3.3 and
give an example showing the disadvantages of this idea.
In Section 3.10 we comment on the correpondence between the local and the graded
situation. Especially, we will explain a theorem due to Auslander and Reiten saying that
for a positively-graded ring R, where R0 is a perfect field, the isomorphism classes of
indecomposable, graded maximal Cohen-Macaulay R-modules up to degree shift and the
isomorphism classes of indecomposable, maximal Cohen-Macaulay R̂-modules coincide
(cf. [AR89]).
In Chapter 4 we want to find an easy way to compute the Hilbert-series of syzygy modules
of the form
Ma := SyzR (X a · V1 , . . . , X a · Vm , Vm+1 , . . . , Vm+l ),
where R is a positively-graded ring of the form k[X, Y1 , . . . , Yn ]/(X d − F(Y1 , . . . , Yn )) and
the Vi are monomials in the variables Y j . Inspired by Brenner’s work in [Bre05b], we will
prove the following theorem.
Theorem (cf. Theorem 4.5). Let a = d · q + r with q, r ∈ N and 0 ≤ r ≤ d − 1. We have a
short exact sequence
0 −→ Ma+d−2r (s − α · r) −→ Mdq (s − α · r) ⊕ Mdq+d (s) −→ Ma (s) −→ 0
for all s ∈ Z, where α = deg(X).
Note that the modules Mdq and Mdq+d in the middle spot are already defined over the
polynomial ring k[Y1 , . . . , Yn ]. From this theorem we can compute the Hilbert-series of
the R-modules Ma .
Theorem (cf. Theorem 4.7). The Hilbert-series of Ma is given by
H Ma (t) =
(tα·r − tα·d ) · H Mdq (t) + (1 − tα·r ) H Mdq+d (t)
.
1 − tα·d
In the second section we compute for each isomorphism class of indecomposable, maximal Cohen-Macaulay modules over surface rings of type ADE the Hilbert-series of the
representing first syzygy modules of ideals found in Chapter 3. Moreover, we discuss
which splitting behaviours of a given maximal Cohen-Macaulay module M can be excluded by computing its Hilbert-series.
In Chapter 5 we use the fact that the surface rings of type ADE defined over C are the
rings of invariants of C[x, y] under actions defined by the finite subgroups of SL2 (C). If R
is of type ADE over an algebraically closed field of characteristic p > 0 and G ( SL2 (C)
is the corresponding group in characteristic zero with order invertible in k, it remains true
that R is the ring of invariants of k[x, y] under the action of a finite subgroup G0 ( SL2 (k)
with |G0 | = |G|. If H, G are finite subgroups of SL2 (k) such that H is a normal subgroup
of G, then k[x, y]G is a subring of k[x, y]H , inducing a morphism ι : V → U, where V and
14
INTRODUCTION
Ä
ä
Ä
ä
U are the punctured spectra of Spec k[x, y]H resp. Spec k[x, y]G . Fixing a surface
ring of type ADE with corresponding group G and non-isomorphic, indecomposable,
maximal Cohen-Macaulay modules M1 , . . . , Mn representing all isomorphism classes, we
will compute the pull-backs of all the corresponding OU -modules under all maps ι as
above. These explicit computations are done in Sections 5.1 to 5.5. In Section 5.6 we will
gather the results of the previous sections and obtain the following lemma.
Lemma (cf. Lemma 5.4). Let U be the punctured spectrum of a surface ring R of type D
or E. Assume that the order Äof the group corresponding
to the singularity is invertible in
e
e
eä
the ground field. Then SyzU X p , Y p , Z p is indecomposable for all e ∈ N.
In Chapter 6 we compute the Hilbert-Kunz functions of the surface rings of type DE (and
A). During these computations the whole strength of Theorem 4.7 becomes obvious.
In the last chapter we discuss possible extensions of the approach using matrix factorizations. More explicitly, we compute the Hilbert-Kunz functions of E8 with respect to the
ideals (X, Y, Z 2 ) and (X, Y 2 , YZ, Z 2 ) (cf. Examples 7.1 and 7.2) as well as the Hilbert-Kunz
functions of A∞ := k[X, Y, Z]/(XY) and D∞ := k[X, Y, Z]/(X 2 Y − Z 2 ) (cf. Sections 7.2 and
7.3). We also compute for every isomorphism class of non-free, indecomposable, maximal Cohen-Macaulay modules over A∞ and D∞ a representing first syzygy module of an
ideal and the corresponding Hilbert-series.
In the fourth section we focuss on projective Fermat curves C of degree n with C :=
Proj(R) and R := k[X, Y, Z]/(X n + Y n + Z n ), where we assume k to be an algebraically
closed field of characteristic p > 0 coprime to n. We start the section byÄa brief discussion
e
e
eä
of Kaid’s result on the strong Harder-Narasimhan filtration of SyzC X p , Y p , Z p for
e 0 and the Hilbert-Kunz function of R in the case, where SyzC (X, Y, Z) is not strongly
semistable. Since the vector bundle SyzC (X, Y, Z) is strongly semistable and is not trivialized by the Frobenius if and only if all quotients R/(X q , Y q , Z q ) have infinite projective
dimension (cf. Corollary 7.11), we can use a theoremÄ of Kustin, äRahmati and Vraciu
which gives the free resolution of the quotient Q := R/ X N , Y N , Z N if the projective dimension of Q is infinite (cf. [KRV12] or Theorem 7.15) to obtain
Corollary
(cf. äCorollary 7.16). Let N = θ · n + r with θ, r ∈ N and 0 ≤ r < n. Assume that
Ä
R/ X N , Y N , Z N has infinite projective dimension. Then
Ä
SyzR X , Y , Z
N
N
N
ä
(
(m)
SyzR (X r , Y r , Z r )
SyzR (X
n−r
,Y
n−r
if θ is even,
,Z
n−r
) if θ is odd.
holds for some m ∈ Z.
This corollary is used to prove that
Ç
å
1 1 1
δ , ,
=0
n n n
is sufficient for having a Frobenius periodicity of SyzProj(R) (X, Y, Z). Moreover, the corollary enables us to compute the Hilbert-Kunz function of R, provided SyzC (X, Y, Z) is
strongly semistable and is not trivialized by the Frobenius.
At the end we discuss the question which Frobenius periodicities the bundles SyzC (X, Y, Z)
might admit.
NOTATIONS AND CONVENTIONS
15
Finally, the last section of Chapter 7 is reserved for a discussion of open questions.
We should mention that we used CoCoA for a numerical treatment of all explicit results on
Hilbert-Kunz functions and multiplicities (of course only for small values of the various
parameters).
Acknowledgements
First of all, my deepest thanks go to my supervisor Holger Brenner. Without our countless
discussions it would not have been possible to write this thesis. I thank him for various
usefull comments and his extraordinary patience during the last years.
I would like to thank Tim Römer and Heinz Spindler for their great lectures on commutative algebra and algebraic geometry. Especially, because many of these courses were held
in addition to their teaching duties.
I thank Almar Kaid for several discussions on Hilbert-Kunz theory. These made the
beginning of my Ph.D. studies easier.
I am thankful to Axel Stäbler for very helpfull discussions on vector bundles and his
usefull comments to my thesis.
Special thanks go to Igor Burban who brought the theory of matrix factorizations to my
attention and to Kevin Tucker who is responsible for the corollaries about the F-signature
functions of the surface rings of type ADE.
I thank Alessio Caminata for his corrections to my thesis as well as the whole institute of
mathematics in Osnabrück for the great time I had in the last years. There was always a
warm and familar working atmosphere.
Finally my deep gratitude goes to my family and my friends for their support and patience.
Notations and conventions
During this thesis we will make use of the following notations and conventions.
• N := {0, 1, 2, . . .} and N≥a := {a, a + 1, a + 2, . . .}.
• All rings are assumed to be non-zero, commutative and Noetherian. Moreover,
they have a unit element.
• A standard-graded ring R is generated over R0 by finitely many elements of
degree one.
• A positively-graded ring R is generated over R0 by finitely many elements of
positive degree.
• A hypersurface denotes a quotient of a polynomial ring by a principal ideal.
• Prime and maximal ideals are always proper ideals.
• We denote by R̂ the m-adic completion of a local ring (R, m) or the R+ -adic
completion of a graded ring.
• If R → S is a homomorphism of rings and I ⊆ R an ideal, we denote by IS the
ideal of S which is generated by the image of I.
• Krull dimensions of rings and modules are denoted by dim(_).
• If V is a k-vector space, its dimension is denoted by dimk (V).
CHAPTER 1
A survey on Hilbert-Kunz theory
This chapter is devoted to give an overview of Hilbert-Kunz theory. In the first section
we will give the basic definitions and state some results. In the second section we will
give a brief introduction to vector bundles and explain how they appear in Hilbert-Kunz
theory. Finally, in section three we will define the related invariants limit Hilbert-Kunz
multiplicity and F-signature and list up some further references.
1.1. The algebraic viewpoint of Hilbert-Kunz theory
We start by gathering some properties of the length function λR .
Proposition 1.1. Let (R, m), (S , n) be local rings, let M be a finitely generated R-module
and N a finitely generated S -module. Then the following hold.
(i) If R is a k-algebra with dimk (R/m) < ∞, then λR (M) · dimk (R/m) = dimk (M).
(ii) If R → S is local and λR (M), λS (S /mS ) are finite, then
λR (M) · λS (S /mS ) ≥ λS (S ⊗R M).
Moreover, equality holds if the morphism is flat.
(iii) Let R → S be local with λS (N) < ∞ and [S /n : R/m] < ∞. Then we have
λR (N) = λS (N) · [S /n : R/m].
Proof. See Exercises 1.6 (d), 1.8 (b) and Lemma 1.36 (a) in Chapter 7 of [Liu06].
Definition 1.2. Let (R, m) be a local ring of characteristic p > 0. Let I = ( f1 , . . . , fm ) be an
m-primary ideal and M a finitely generated R-module. The function HK(I, M, pe ) : N → N,
pe
e
e 7−→ λR M/ f1 , . . . , fmp
M
is called the Hilbert-Kunz function of M with respect to I.
Remark 1.3.
(i) Denote by F : R → R the Frobenius morphism r 7→ r p . Since this
pe
pe
is a homomorphism and f1 , . . . , fm = Fe (I), it is immediately clear that the
Hilbert-Kunz function only depends on I and not on the generators. This is not
true for the so-called generalized Hilbert-Kunz function HK(I, M, n) : N → N,
n 7−→ λR M/( f1n , . . . , fmn )M .
To show this, we compute the values for n = 2 with respect to two different
representations of m in a ring of characteristic three. Let
R := Z/(3)~X, Y, Z/(X 2 + Y 3 + Z 4 ),
17
18
1. A SURVEY ON HILBERT-KUNZ THEORY
I = m and M = R. Choosing I = (X, Y, Z) we have
R/(X 2 , Y 2 , Z 2 ) = Z/(3)~X, Y, Z/(X 2 , Y 2 , Z 2 ),
which has length eight. But taking I = (X + YZ, Y, Z), we have
R/((X + YZ)2 , Y 2 , Z 2 ) = Z/(3)~X, Y, Z/(X 2 , Y 2 , Z 2 , XYZ),
which has length seven.
(ii) Using Proposition 1.1, one can show that the Hilbert-Kunz function is independent under completion of R resp. taking the algebraic closure of R/m (compare [Kun69, Lemma 3.1]). Note that this proof also works for the generalized
Hilbert-Kunz function.
(iii) In view of the previous remark, the Hilbert-Kunz function of a positively-graded
affine k-algebra, where k is a field and may assumed to be algebraically closed,
is always the Hilbert-Kunz function of the R+ -adic completion of R.
(iv) Assume that (R, m) is an integral domain of characteristic p > 0 with perfect
residue class field. Denote by R1/q the ring of q-th roots of elements in R in an
algebraic closure of the field of fractions of R. If R1/q is a finite R-module for
all q = pe then we have by [Wat00, Remark 1.3] for any m-primary ideal I the
isomorphism
R/Fe (I) R1/q /IR1/q .
Notation 1.4.
(i) If I = m, we will omit the ideal in our notation, hence
HK(M, q) := HK(m, M, q).
Moreover, we will call this function the Hilbert-Kunz function of M.
(ii) We will also use the notation I [q] := Fe (I) with q = pe .
Example 1.5. We will compute the generalized Hilbert-Kunz function of the surface
rings of type An , given by R := k[X, Y, Z]/(X n+1 − YZ), with respect to the graded maximal ideal m = (X, Y, Z). We need to compute the length of R/(X m , Y m , Z m ) for all m ∈ N,
which is the same as the dimension of the quotient as a k-vector space. For any m ∈ N, let
m = l(n + 1) + r with l ∈ N and 0 ≤ r ≤ n. Counting the monomials in
k[X, Y, Z]/(X m , Y m , Z m , X n+1 − YZ),
we can always reduce the power of X to a number in {0, . . . , n} using the defining equation
X n+1 = YZ. Since all monomials divisible by X m vanish, we obtain the equations
Xm
=
..
.
X m+n−r =
X m+n−r+1 =
..
.
X m+n
Xr Y lZl
XnY lZl
Y l+1 Z l+1
= 0
= 0
= 0
= X r−1 Y l+1 Z l+1 = 0
n + 1 − r equations
r equations
Since the powers of X are linearly independent, we see by the above equations that a
monomial X h Y i Z j ∈ R/(X m , Y m , Z m ) is non-zero if and only if the exponents satisfy the
following relations
1.1. THE ALGEBRAIC VIEWPOINT OF HILBERT-KUNZ THEORY
19
(i) If 0 ≤ h ≤ r − 1, then either 0 ≤ i ≤ l and 0 ≤ j ≤ m − 1 or l + 1 ≤ i ≤ m − 1 and
0 ≤ j ≤ l.
(ii) If r ≤ h ≤ n, then either 0 ≤ i ≤ l − 1 and 0 ≤ j ≤ m − 1 or l ≤ i ≤ m − 1 and
0 ≤ j ≤ l − 1.
Counting the number of triples (h, i, j) satisfying the above constraints, we get
HK(R, m) = dimk (k[X, Y, Z]/(X m , Y m , Z m , X n+1 − YZ))
l−1 m−1
l m−1
m−1
l−1
m−1
l
X
X
X
XX
X
X X
= (n + 1 − r)
1+
1 + r
1+
1
i=0 j=0
i=0 j=0
i=l j=0
i=l+1 j=0
= (n + 1 − r)(ml + (m − l)l) + r((l + 1)m + (m − (l + 1))(l + 1))
= (n + 1 − r)(2ml − l2 ) + r(2(l + 1)m − (l + 1)2 )
Ç
Å
ã å
Ç
ã
Å
ã2 å
Å
m−r 2
m−r
m−r
m−r
−
+1 −
+1
+ r 2m
= (n + 1 − r) 2m
n+1
n+1
n+1
n+1
!
2m2 2rm
m2
2rm
r2
= (n + 1 − r)
−
−
+
−
n + 1 n + 1 (n + 1)2 (n + 1)2 (n + 1)2
2m2 2n + 2 − 2r
m2 − 2rm + r2 2m − 2r
+r
+
m−
−1
−
n+1
n+1
(n + 1)2
n+1
2n + 1 2
2rn
r2
= (n + 1 − r)
m
−
m
−
(n + 1)2
(n + 1)2
(n + 1)2
Ç
!
!
å
2r
2r
2
r2
2r
2n + 1 2
m
+
2
−
+
−
m
−
+
−1
+r
(n + 1)2
n + 1 (n + 1)2 n + 1
(n + 1)2 n + 1
!
2n + 1 2
2rn
2r2 n
2r2
2r2
2r
=
m + −
+
+
m
+
2r
−
−
2
2
n+1
n + 1 (n + 1)
n + 1 (n + 1)
n+1
!
r2
r3
r3
2r2
+
−r
−
+
n + 1 (n + 1)2 (n + 1)2 n + 1
Ç
å
1
r2
= 2−
m2 +
− r.
n+1
n+1
Remark 1.6. Similarly to the last example Kunz computed in [Kun76, Example 4.3] the
Hilbert-Kunz functions of binomial hypersurfaces of the form
−
R := k[X, Y1 , . . . , Yn ]/(X d − Y1a1 · . . . · Ynan ),
where k is a field of characteristic p > 0 coprime to d ∈ N≥2 and a1 , . . . , an are natural
P
numbers with ai > 1. He obtains (compare with the introduction of [Mon83a])
Å
ãã
YÅ
ai
d · 1−
1−
if ai < d for all i ∈ {1, . . . , n},
d
eHK (R) =
d
otherwise.
The first deep theorem in Hilbert-Kunz theory was the following.
Theorem 1.7 (Kunz). Let (R, m) be a local ring of dimension d and positive characteristic
p. The following statements are equivalent.
20
1. A SURVEY ON HILBERT-KUNZ THEORY
(i)
(ii)
(iii)
(iv)
(v)
R is regular.
R is reduced and a flat R p -module.
HK(R, p) = pd .
e
R is reduced and a flat R p -module for all e ∈ N.
HK(R, pe ) = ped for all e ∈ N.
Proof. See [Kun69, Theorem 2.1, Theorem 3.3].
This theorem leads to the question how the quotient HK(I, M, pe )/ped behaves for large e.
This question was answered by Monsky.
Recall that for a set X and a function f : X → R, the set O( f ) consists of all functions
g : X → R with the property that |g/ f | is bounded.
Theorem 1.8 (Monsky). Let (R, m) be a local ring of positive characteristic p and dimension d. Let I be an m-primary ideal and M a finitely generated R-module.
(i) There is a positive real constant c(M) such that
Ä
ä
Ä
ä
λR M/I [q] M = c(M)qd + O qd−1 .
(ii) If 0 → L → M → N → 0 is a short exact sequence of finitely generated Rmodules, then
Ä
ä
Ä
ä
Ä
ä
Ä
ä
λR M/I [q] M = λR L/I [q] L + λR N/I [q] N + O qd−1 .
Proof. See [Mon83a, Theorem 1.8].
Note that part (i) of this theorem ensures that the limit in the following definition exists.
Definition 1.9. Let (R, m) be a local ring of positive characteristic p and dimension d.
Let I be an m-primary ideal and M a finitely generated R-module. We call the limit
HK(I, M, pe )
e→∞
ped
lim
the Hilbert-Kunz multiplicity of M with respect to I. We denote this limit by eHK (I, M)
and set eHK (M) := eHK (m, M) for short. We will also call eHK (M) the Hilbert-Kunz multiplicity of M.
Remark 1.10. We give a few remarks on results one obtains on the way to prove Theorem
1.8. The references are [Mon83a] and [Hun96].
(i) We have the inequalities (e(I) denotes the ordinary multiplicity of I)
e(I)
≤ eHK (I, R) ≤ e(I).
d!
Hence, the Hilbert-Kunz multiplicity of R with respect to I equals the multiplicity of I if dim(R) ≤ 1 and generalizes it in higher dimensions.
(ii) If I is generated by a system of parameters, one always has eHK (I, R) = e(I).
1.1. THE ALGEBRAIC VIEWPOINT OF HILBERT-KUNZ THEORY
21
(iii) The associativity formula holds for Hilbert-Kunz multiplicities, that is
eHK (I, M) =
X
eHK (I, R/p) · λRp (Mp ).
p∈Ass(M),
dim(R/p)=dim(R)
(iv) We have eHK (I, M) = 0 if and only if dim(M) < dim(R).
(v) Two finitely generated R-modules have the same Hilbert-Kunz multiplicity with
respect to the ideal I if their localizations at all minimal primes of maximal
dimension are isomorphic. Hence, the Hilbert-Kunz multiplicity does not detect
embedded primes.
We give an example to the last remark.
Example 1.11. Let p be a prime number, R := Z/(p)~X, Y, Z/(XZ, YZ) and M := R/(Z).
There is only one minimal prime of maximal dimension of R, namely (Z). The localizations of R and M at (Z) are both isomorphic to the field of fractions of the ring R/(Z).
Their Hilbert-Kunz functions are
HK(M, pe ) = p2e and
HK(R, pe ) = p2e + pe − 1
as one sees by counting monomials as in Example 1.5. This gives eHK (R) = eHK (M) = 1.
Treating R and M as Z/(p)~X, Y, Z-modules, their Hilbert-Kunz functions are the same
but their Hilbert-Kunz multiplicities are zero.
The next example shows that the limit
λR M/ f1n , . . . , fmn M
lim
,
n→∞
ndim(R)
sometimes called generalized Hilbert-Kunz multiplicity, does not exist even in easy cases.
Example 1.12. Let R := k[X, Y, Z]/(X + Y + Z), where k is any field of characteristic p > 0.
Consider the sequence
å
Ç
dimk (R/(X n , Y n , Z n ))
.
n2
n∈N
The subsequence to indices of the form pe converges to 1, but the subsequence to indices
of the form 2pe converges to 34 as one sees by counting monomials as in Example 1.5.
By Theorem 1.7 one has that every regular local ring of positive characteristic has HilbertKunz multiplicity one. In fact, one can show in this situation that
HK(I, R, pe ) = λR (R/I) · pe·dim(R)
holds for any m-primary ideal (see [WY00, 1.4] and [BH98, Exercise 8.2.10]). By Example 1.11 the converse does not hold in general. One can only expect it to be true if
there are no embedded primes. In fact this is - as for the Hilbert-Samuel multiplicity (see
[Nag75]) - already enough.
Theorem 1.13 (Watanabe, Yoshida). Let (R, m) be an unmixed local ring of positive characteristic. If R has Hilbert-Kunz multiplicity one, then R is regular.
22
1. A SURVEY ON HILBERT-KUNZ THEORY
Proof. See [WY00, Theorem 1.5] or [HY02, Theorem 3.1] for an alternative proof.
If the ring has dimension zero, we always have I [q] M = 0 for q large enough. Therefore,
all Hilbert-Kunz functions in these cases are eventually constant. Due to Monsky the
structure of Hilbert-Kunz functions over one-dimensional rings is known completely.
Theorem 1.14 (Monsky). Let (R, m) be a local ring of positive characteristic p and dimension one. Let I be an m-primary ideal and M a finitely generated R-module. There is
a periodic function ϕ : N → N such that
HK(I, M, pe ) = eHK (I, M) · pe + ϕ(e)
for all e 0.
Proof. See [Mon83a, Theorem 3.11].
Remark 1.15. In [Kre07] the author gives an algorithm in the case of an one-dimensional,
standard-graded, affine k-algebra R, where k is a finite field, that computes ϕ and a bound
e0 ∈ N such that the above theorem holds for all e ≥ e0 .
For the next theorems dealing with the structure of Hilbert-Kunz functions over normal
domains, we need the divisor class group C(R) of R, which is the quotient of the group
of Weil divisors by the normal subgroup of principal divisors. For a finitely generated Rmodule M, we choose a primary decomposition with quotients R/Pi . We attach to M the
P
divisor − P j , where P j runs through all Pi of height one (with repetitions) in the primary
P
decomposition. The image of − P j in C(R) is independent of the chosen decomposition
and is denoted by c(M).
Theorem 1.16 (Huneke, McDermott, Monsky). Let (R, m) be a local normal domain of
positive characteristic p and dimension d. Let M and N be finitely generated, torsion-free
R-modules of the same rank r. Then the following statements hold.
Ä
ä
(i) If c(M) = 0, then HK(M, q) = r · HK(R, q) + O qÄd−2 .ä
(ii) If c(M) = c(N), then HK(M, q) = HK(N, q) + O qd−2 .
Proof. See [HMM04, Theorem 1.4, Lemma 1.6].
Theorem 1.17 (Huneke, McDermott, Monsky). Let (R, m, k) be a local, excellent, normal
domain of positive characteristic p, dimension d and with k perfect. Let I be an m-primary
ideal and M a finitely generated R-module. Then there exists a constant β(I, M) ∈ R with
Ä
ä
HK(I, M, q) = eHK (I, M) · qd + β(I, M) · qd−1 + O qd−2 .
Proof. See [HMM04, Theorem 1.12].
Remark 1.18.
(i) The numbers β(I, M) in the previous theorem can be computed
using values of a homomorphism τ : C(R) → (R, +). If τ is constant zero (for
example if C(R) is torsion), all β(I, M) are zero. An important example, where
C(R) is torsion, is the case, where (R, m, k) is a complete, local, two-dimensional,
1.1. THE ALGEBRAIC VIEWPOINT OF HILBERT-KUNZ THEORY
23
normal domain and k is the algebraic closure of a finite field (cf. [HMM04,
Lemma 2.1]).
(ii) In the introduction of [HMM04] the authors
briefly (non-)possible exÄ discuss
ä
d−2
tensions of the previous theorem to the O q
-term. They point out that by
an example in [HM93] the Hilbert-Kunz function of the ring
R = Z/(5)[X, Y, Z, W]/(X 4 + Y 4 + Z 4 + W 4 )
is given by the formula
168 3e 107 e
5 −
3,
61
61
which shows that one cannot hope to have a constant factor in front of the qd−2 term, even for rings that are regular in codimension two.
(iii) In [HY09, Corollary 2.5] the authors weaken the assumptions in the above theorem to an excellent, equidimensional, reduced ring such that the ideal of the
singular locus has height at most two.
(iv) In [CK13, Theorem 3.2] the authors weaken the assumptions in the previous
theorem to an excellent, local ring with perfect residue class field such that all
localizations at primes p of dimension d resp. d − 1 are fields resp. DVRs.
Moreover, they argue with the Chow group instead of the divisor class group.
(v) In [Con95] Contessa uses Koszul homology and the fact
HK(R, 5e ) =
Ä
ä
Ä
Ä
λR M/I [q] M = λR H0 I [q] ; M
ää
to compute Hilbert-Kunz functions of finitely generated modules over regular
rings with algebraically closed residue class field with respect to ideals generated by a system of parameters. She shows that for a finitely generated torsionfree module M one has
Ä
ä
HK(I, M, q) = rank(M) · qd + const · qd−2 + O qd−3 ,
where the constant is a non-negative real number, depending only on (M ∨ )∨ /M
(see [Con95, Theorem 3.4]). In particular, the constant is zero if M is reflexive.
The next two statements give bounds for Hilbert-Kunz functions (of rings).
Lemma 1.19. Let (R, m, k) be a complete local ring of positive characteristic p and assume k to be algebraically closed. Then the inequalities
y1 ,...,yd
λR (R/(y1 , . . . , yd )) · q
system of parameters
min
d
≥ HK(R, q) ≥ qd
hold. Moreover, if R is Cohen-Macaulay, we have e(R) · qd ≥ HK(R, q) ≥ qd .
Proof. See [Kun69, Proposition 3.2].
Theorem 1.20 (Watanabe, Yoshida / Huneke, Yao). Let (R, m) be a local ring of characteristic p > 0. Let I ⊆ J be ideals, where I is m-primary (and J = R is allowed). Then
HK(I, R, q) ≤ HK(J, R, q) + λR (J/I) · HK(R, q).
In particular, for J = R one has HK(I, R, q) ≤ λR (R/I) · HK(R, q).
24
1. A SURVEY ON HILBERT-KUNZ THEORY
Proof. See [WY00, Lemma 4.2 (2)] for the analogue statement about the HilbertKunz multiplicities and [HY02, Lemma 2.1, Corollary 2.2 (1)] for the general case.
The next theorem shows how Hilbert-Kunz multiplicities behave under module-finite extensions of local domains.
Theorem 1.21 (Watanabe, Yoshida). Let (R, m) ⊆ (S , n) be an extension of local domains
of positive characteristic, where S is a finite R-module. Let I be an m-primary ideal of R.
Then we have
eHK (IS , S ) · [S /n : R/m]
,
eHK (I, R) =
[Q(S ) : Q(R)]
where Q(_) denotes the field of fractions.
Proof. See [WY00, Theorem 2.7] resp. [WY05, Theorem 1.6].
Remark 1.22. Using Proposition 1.1, it follows easily that Theorem 1.21 holds even for
the Hilbert-Kunz functions under the additional assumptions that R is an R/m-algebra and
that the extension R ⊆ S is flat. Explicitly, we have (the numbers refer to Proposition 1.1)
Ä
λR R/I [q]
Ä
ä
λS S /(IS )[q]
=
λS (S /mS )
Ä
ä
[q] · [S /n : R/m]
(iii) λS S /(IS )
=
λR (S /mS )
Ä
ä
[q] · [S /n : R/m]
(i) λS S /(IS )
.
=
dimk (S /mS )
ä (ii)
The denominator is the same as in the theorem, since
dimk (S /mS ) = dimQ(R) (Q(R) ⊗k k ⊗R S ) = dimQ(R) Q(S ).
Using Theorem 1.21 Brenner reproved a theorem in invariant theory of Smith.
Theorem 1.23 (Smith / Brenner). Let B := k[X1 , . . . , Xn ] with an algebraically closed
field k of positive characteristic p. Let G ( GL(n, k) be a finite subgroup and consider its
natural linear action on B. Let A := BG and m := A+ . We set h := dimk (B/mB), where mB
is the Hilbert ideal. Let R be one of the rings Am and d
Am . Then the following statements
hold.
(i) h = |G| · eHK (R) ≥ |G|.
(ii) Equality holds if and only if A is a polynomial ring.
Proof. The original proof was given in [Smi03] and the alternative proof can be found
in [Bre].
A second application of Theorem 1.21 is the computation of the Hilbert-Kunz multiplicities of surface rings with isolated singularities of type ADE.
1.1. THE ALGEBRAIC VIEWPOINT OF HILBERT-KUNZ THEORY
25
Definition 1.24. Let R := k[X, Y, Z]/(F), where k is a field. We say that R and Rb are
surface rings of type
An
Dn
E6
E7
E8
if F = X n+1 + YZ and n ≥ 0,
if F = X 2 + Y n−1 + YZ 2 and n ≥ 4,
if F = X 2 + Y 3 + Z 4 ,
if F = X 2 + Y 3 + YZ 3 ,
if F = X 2 + Y 3 + Z 5 .
In any case a surface ring of type ADE has in almost all characteristics an isolated singularity at the origin and we refer to these singularities as singularities of type ADE.
Remark 1.25. Recall that the surface rings C[X, Y, Z]/(F) of type ADE appear as rings
of invariants of C[x, y] by the actions of the finite subgroups of SL2 (C). The groups
corresponding to the singularities of type An , Dn , E6 , E7 resp. E8 are the cylic group with
n + 1 elements, the binary dihedral group of order 4n − 8, the binary tetrahedral group of
order 24, the binary octahedral group of order 48 resp. the binary icosahedral group of
order 120. If k is algebraically closed of characteristic p > 0 the groups above can be
viewed as finite subgroups of SL2 (k), provided their order is invertible in k. In these cases
k[X, Y, Z]/(F) is again the ring of invariants of k[x, y] under the action of the corresponding
group (cf. [LW12, Chapter 6, §2]).
Now, we show how one might use Theorem 1.21 to compute the Hilbert-Kunz multiplicities of surface rings of type ADE. See [WY00, Theorem 5.4] for a stronger formulation
of the next theorem.
Theorem 1.26 (Watanabe, Yoshida). Let R be a surface ring of type ADE. Assume that
k is algebraically closed and that the order of the corresponding group G is invertible
modulo char(k) > 0. Then
1
eHK (R) = 2 − .
|G|
Proof. We prove the graded version of the statement. The proof carries over to the
1
. Let
complete situation. Let R be of type An . Then Example 1.5 shows eHK (R) = 2 − n+1
R be of type Dn or E. Denote the corresponding group by G and let k[x, y] be the ring on
which G acts. Then the extension Q(k[x, y]) : Q(R) has degree |G| (cf. [LW12, Proposition
5.4]). The ring R = k[x, y]G is generated as k-algebra by three invariant homogeneous
polynomials X, Y, Z such that X 2 is a polynomial in Y and Z and such that deg(Y) ·
deg(Z) = 2 · |G| holds (cf. Chapter 5). Since k[x, y] is regular, we have eHK (I, k[x, y]) =
λ(k[x, y]/I) for every (x, y)-primary ideal I. This shows
eHK ((X, Y, Z), k[x, y]) = λ(k[x, y]/(X, Y, Z)) = λ(k[x, y]/(Y, Z)) − 1 = 2 · |G| − 1.
By Theorem 1.21 we have
eHK (R) =
eHK ((X, Y, Z), k[x, y])
1
= 2− .
|G|
|G|
26
1. A SURVEY ON HILBERT-KUNZ THEORY
Keeping the assumptions that R is a Cohen-Macaulay local ring of dimension two, one
can prove the inequality
r+2
· e(R),
eHK (R) ≥
2r + 2
where r denotes the number of generators of m/J for a minimal reduction J of m (cf.
[WY00, Lemma 5.5]). In [WY01b] the authors study the numbers eHK (I n , R) for mprimary ideals I. In particular, they show that for a local, two-dimensional, CohenMacaulay ring with algebraically closed residue class field one has
e(R) 2 n
eHK (mn , R) ≥
n +
2
2
(cf.
and that this minimal bound is achieved for all n if and only if eHK (R) = e(R)+1
2
[WY01b, Theorem 2.5]). Moreover, they show that a local, two-dimensional, CohenMacaulay ring R with algebraically closed residue class field has minimal Hilbert-Kunz
multiplicity (n = 1 in the above inequality) if and only if the associated graded ring of R
is isomorphic to the e(R)-th Veronese subring of k[X, Y] (cf. [WY01b, Theorem 3.1]).
One big class of rings in which Hilbert-Kunz functions and multiplicities are studied is
the class of rings that arise from combinatorial problems. For example monoid and toric
rings are studied in [Eto02] and [Wat00]. An explicit treatment of the two-dimensional
toric case is done in [CH03]. In [WY00] and [EW03] the authors focus on Rees algebras.
In [Eto00] (see also [EW03] and [BCP97]), Eto computes the Hilbert-Kunz multiplicity
of Segre products of polynomial rings. The (generalized) Hilbert-Kunz functions of 2 × 2
determinantal rings (in m · n many variables) are studied in [MS12]. They obtain in the
case m = 2 the following theorem (cf. [MS12, Theorem 4.4, Corollary 4.5]).
Theorem 1.27 (Miller, Swanson). Let k be a field and
R := k[Xi, j |i = 1, 2; j = 1, . . . , n]/I2 ,
where I2 is the ideal generated by the two-minors of the matrix (Xi j )i j . For any natural
number m ∈ N we obtain
(i)
nmn+1 − (n − 2)mn
n+m−1
HK(R, m) =
+n
2
n+1
!
(ii)
eHK (R) =
n
n
+
2 (n + 1)!
Remark 1.28. The above result was extended in [RS13] to the case
R := k[Xi, j |i = 1, . . . , m; j = 1, . . . , n]/I2 .
The structure of (generalized) Hilbert-Kunz functions for monomial rings and binomial
hypersurfaces is studied in [Con96].
Theorem 1.29 (Conca). Let I be a monomial ideal in S = k[X1 , . . . , Xn ], where k is a field
and let R = S /I. Then there is a polynomial PR (Y) ∈ Z[Y] of degree dim(R) with leading
coefficient e(R) such that HK(R, m) = PR (m) for all m ∈ N that are at least the highest
exponent appearing in the generators of I.
1.1. THE ALGEBRAIC VIEWPOINT OF HILBERT-KUNZ THEORY
27
Proof. See [Con96, Theorem 2.1].
Theorem 1.30 (Conca). Let R = k[X1 , . . . , Xn+m ]/(F), where k is a field and
bm+1
bm+n
· . . . · Xm+n
F = X1a1 · . . . · Xmam − Xm+1
is a homogeneous binomial with maximal appearing exponent u. Then there exists a
rational polynomial PR (Y, Z) ∈ Q[Y, Z] of degree n + m − 1 with leading coefficient C and
an integer α ≥ 0 with
HK(R, l) = PR (l, ε)
for all l ≥ α and ε ≡ l (u) with ε ∈ {0, . . . , u − 1}. Moreover, we have
C=
m X
n
X
(−1)i+ j sim (a)s jn (b)
i=1 j=1
ij
,
(i + j − 1)ui+ j−1
with a = (a1 , . . . , am ), b = (bm+1 , . . . , bm+n ) and sαβ denotes the elementary symmetric polynomial of degree α in β indeterminates.
Proof. See [Con96, Theorem 3.1].
We now turn to the hypersurface case. Studying properties of the function DF : Nn → N,
(a1 , . . . , an ) 7→ dimk (k[X1 , . . . , Xn ]/(X1a1 , . . . , X an , X1 + . . . + Xn )),
Han and Monsky developed in [HM93] an algorithm to compute Hilbert-Kunz functions
of diagonal hypersurfaces
Ä
ä
k[X1 , . . . , Xn ]/ X d1 + . . . + X dn with d1 , . . . , dn ∈ N≥2 .
In particular, they obtain the following result on the structure of Hilbert-Kunz functions
in this case.
Theorem 1.31 (Han, Monsky). Let R be a d-dimensional diagonal hypersurface over a
field k of positive characteristic p. Then the following holds.
(i) If char(k) = 2 or d = 2, we have
HK(R, pe ) = eHK (R) · pde + ∆e ,
where e 7→ ∆e is eventually periodic.
(ii) If char(k) ≥ 3 and d ≥ 3 there are integers µ ≥ 1 and 0 ≤ l ≤ p(d−2)µ such that
HK(R, pe ) = eHK (R) · pde + ∆e ,
where ∆e+µ = l · ∆e for e 0.
Moreover, ∆e ≤ 0 in any case.
Proof. See [HM93, Theorem 5.7] and [HM93, Corollary 5.9]. See [CH98a, Theorem
1.1] for a different proof using the same tools.
28
1. A SURVEY ON HILBERT-KUNZ THEORY
Remark 1.32.
(i) The ideas of [HM93] are generalized in [CH98b] to hypersurfaces R = k[Xi, j |i = 1, . . . , m, j = 1, . . . , ti ]/( f ) with
f=
ti
m Y
X
d
Xi,i,j j .
i=1 j=1
(ii) In her thesis [Cha93], Chang used the algorithm of Han and Monsky to compute
the Hilbert-Kunz function of the Fermat-quartic
R = Z/(p)[W, X, Y, Z]/(W 4 + X 4 + Y 4 + Z 4 )
for odd primes p. Explicitly, she obtains
2
3
HK(R, p ) =
e
8p2 + 8p + 12 3e 1 10p2 + 10p + 21
p −
2p2 + 2p + 1
3
2p2 + 2p + 1
!
8p3 + 4p + 12 3e 1 10p3 + 11p + 21
p −
3 2p2 − p + 1
3
2p3 − p + 1
!
2
!Ç
!Ç
p+1
2
p−1
2
åe
if p ≡ 1 (4),
åe
if p ≡ 3 (4).
(iii) In his thesis [Che98], Chen remarks that there cannot be a numerical function
of polynomial type having the Hilbert-Kunz multiplicity of the Fermat-quartic
as leading coefficient.
(iv) In the special case n = 2, the function DF is also known as Han’s δ-function,
which we will study in more detail in Chapter 2.
(v) For a hypersurface R, the sequence
Ç
HK(R, pe )
pe·dim(R)
å
e∈N
is non-decreasing by [HM93, Theorem 5.8].
The special case of the Hilbert-Kunz multiplicities of d-dimensional Fermat-quadrics
reappear as lower bounds for non-regular complete intersections.
Theorem 1.33 (Enescu, Shimomoto). Let (R, m, k) be a local, non-regular, complete intersection of dimension d ≥ 2 with char(k) , 2. Then
Ñ
eHK (R) ≥ eHK k~X0 , . . . , Xd /
Ñ
d
X
éé
Xi2
.
i=0
Proof. This bound was conjectured in [WY05] and the d-dimensional case was proven
in [ES05, Theorem 4.6].
Remark 1.34. In [GM10, Theorem 3.8] the authors showed that
Ñ
Ñ
eHK k~X0 , . . . , Xd /
Ñ
d
X
i=0
ééé
Xi2
char(k)=p
tends to 1+ the coefficient of td−1 in the power series of sec(t) + tan(t) around zero as p
goes to infinity.
1.1. THE ALGEBRAIC VIEWPOINT OF HILBERT-KUNZ THEORY
29
Recall that for regular rings one has HK(I, R, pe ) = λR (R/I) · pe·dim(R) (discussion before
Theorem 1.13). This is generalized to complete intersections by the following theorem.
Theorem 1.35 (Dutta, Miller). Let (R, m, k) be a local ring of dimension d which is a
complete intersection and let M be a finitely generated R-module of finite length. Then
the following statements hold.
(i) We have λR (Fe (M)) ≥ λR (M) · ped for all e ∈ N.
(ii) The sequence (λR (Fe (M)) · p−ed )e∈N is non-decreasing.
(iii) The following statements are equivalent.
(a) The module M has finite projective dimension.
(b) We have λR (Fe (M)) = λR (M) · ped for all e ∈ N.
(c) The limit lime→∞ λR (Fe (M)) · p−ed exists and equals λR (M).
In particular, if I is an m-primary ideal such that the quotient R/I has finite
projective dimension, one has
HK(I, R, pe ) = λR (R/I) · ped .
Proof. Part (i) and the implication (a) ⇒ (b) of (iii) can be found in [Dut83, Theorem
1.9]. For the other statements see [Mil00, Theorem 2.1].
An open question in Hilbert-Kunz theory is whether Hilbert-Kunz multiplicities are rational or not. Many experts thought for a long time Hilbert-Kunz multiplicities would
always be rational (cf. [Mon83a]), since this was true in all explicit examples and in all
cases, where the Hilbert-Kunz multiplicity was formulated in terms of a different theory
as combinatorial algebra or vector bundles. This opinion has changed due to a conjecture
of Monsky.
Theorem 1.36 (Monsky). Let S := Z/(2)[X, Y, Z] and h := X 3 + Y 3 + XYZ ∈ S . Suppose
that Monskys conjecture [Mon08, Conjecture 1.5] on the value of
Ä Ä
ä
ä 1 Ä
Ä Ä
ää
Ä Ä
äää
HK S / h2l+1 , 2e+1 − · HK S / h2l , 2e+1 + HK S / h2l+2 , 2e+1
2
e
holds for all 0 ≤ l < 2 . Then the following statements hold.
(i) Let 2e ≥ j. Then there are recursively defined integers v j and u j such that
7 j e j2
·4 −
· 2e + u j if j ≡ 1 (3),
3
3
HK(S /(h ), 2 ) =
7 j e j2 e
· 4 − · 2 + v j if j ≡ 2 (3).
3
3
j
e
(ii)
eHK (Z/(2)[X, Y, Z, U, V]/(UV + h)) =
(iii) The number
X
n≥0
2n
n
!2 Ç
1
· 16
2
ån
4
5
√ .
+
3 14 · 7
30
1. A SURVEY ON HILBERT-KUNZ THEORY
is transcendental and a Q-linear combination of Hilbert-Kunz multiplicities of
hypersurfaces in characteristic two.
Proof. For (i) and (ii) see [Mon08, Theorem 1.9, Corollary 2.7]. The proof that the
number in (iii) is transcendental (and exists) can be found in [Sch57] and the rest of the
statement in [Mon09, Theorem 2.2].
Very recently, Brenner came up in [Bre13] with a quite explicit example that Hilbert-Kunz
multiplicities might be irrational.
Theorem 1.37 (Brenner). There exists a local domain whose Hilbert-Kunz multiplicity is
irrational and there exists a three-dimensional hypersurface ring
R = k[X, Y, Z, W]/(F),
where F is homogeneous of degree 4 and k is an algebraically closed field of large characteristic, and an Artinian R-module M such that eHK (M) is irrational.
At the end of this section we state some results on the interaction of Hilbert-Kunz multiplicities with the theory of tight closure.
Definition 1.38. Let R be a ring of characteristic p > 0 and let I be an ideal of R. We say
that x ∈ R belongs to the tight closure of I if cxq ∈ I [q] for a c in the complement of the
minimal primes of R not depending on q and all large q. The tight closure of I is an ideal,
which we denote by I ∗ and we say that I is tightly closed in the case I = I ∗ .
The following theorem shows that the Hilbert-Kunz multiplicity measures the membership to the tight closure.
Theorem 1.39 (Hochster, Huneke). Let (R, m) be a local ring of characteristic p > 0 and
let I ⊆ J be m-primary ideals. Then
I ∗ = J ∗ =⇒ eHK (I, R) = eHK (J, R).
The converse holds, if R is quasi-unmixed and analytically unramified.
Proof. See [Hun96, Theorem 5.4].
The previous theorem tells us that Hilbert-Kunz multiplicities behave to tight closure as
Hilbert-Samuel multiplicities behave to integral closure.
Definition 1.40. We call a ring R of characteristic p > 0 an F-rational ring if every ideal
generated by parameters is tightly closed.
The following lemma gives a criterion for Cohen-Macaulay rings of multiplicity two to
be F-rational.
Lemma 1.41. Let (R, m) be Cohen-Macaulay of positive characteristic with multiplicity
two. Then R is F-rational if and only if eHK (R) < 2.
Proof. See [WY00, 5.3].
1.2. THE GEOMETRIC VIEWPOINT OF HILBERT-KUNZ THEORY
31
In view of Theorem 1.21 we obtain the following corollary.
Corollary 1.42. Let R be a surface ring of type ADE such that the order of the associated
group is invertible in the ground field. Then R is F-rational.
1.2. The geometric viewpoint of Hilbert-Kunz theory
In this section we introduce some basic notations concerning vector bundles, discuss
briefly the different Frobenius morphisms appearing in algebraic geometry, focus a little more on syzygy bundles and different stability properties of vector bundles. We will
often restrict ourselves to projective curves, since the definitions get much simpler in this
situation. Thereafter we will explain the connection of the theory of vector bundles and
Hilbert-Kunz theory. We will finish this section by a discussion of the Hilbert-Kunz functions of the homogeneous coordinate rings of elliptic curves and irreducible curves of
degree three.
1.2.1. Generalities on vector bundles.
Definition 1.43. Let X be a scheme. A (geometric) vector bundle of rank r over X is a
scheme f : E → X together with an open covering {Ui }i∈I of X and isomorphisms Ψi :
f −1 (Ui ) → ArUi such that for any i, j and for any open affine set V := Spec(A) ⊆ Ui ∩ U j
the automorphism
Ψ := Ψ j | f −1 (U j ) ◦ Ψ−1
i |ArV
is given by an A-linear automorphism of A[X1 , . . . , Xr ].
Remark 1.44.
(i) There is an equivalence of the categories of vector bundles over
X and the category of locally free sheaves of constant (finite) rank over X. This
equivalence attaches to a vector bundle of rank r its sheaf of sections, which carries a natural OX -module structure and is locally free of rank r. For the converse,
attach to a locally free sheaf of rank r the relative spectrum of its symmetric algebra. This gives a vector bundle of rank r, whose sheaf of sections is the dual
of the locally free sheaf we started with (cf. [Har87, Exercise II.5.18]).
(ii) Note that the quotient of a vector bundle by a subbundle is again a vector bundle
(cf. [Pot97, Proposition 1.7.1]). Hence an inclusion of locally free sheaves
S ⊆ T yields by the above equivalence a closed immersion of vector bundles if
and only if the quotient T /S is itself a locally free sheaf (cf. [GD61, Proposition
1.7.11]).
Definition 1.45. Let E be a locally free sheaf of rank r on X. Then
det(E) :=
r
^
E
is called the determinant-bundle or determinant of E.
Example 1.46. If L is an invertible sheaf, we have det(L) = L.
By [Har87, Exercise II.5.16] determinants satisfy the following properties.
32
1. A SURVEY ON HILBERT-KUNZ THEORY
Proposition 1.47. The situation is the same as in the previous definition.
(i) Determinants are invertible.
(ii) If r = 2, there is an isomorphism
E E∨ ⊗ det(E).
(iii) If f : Y → X is a morphism of schemes and E a locally free OX -module of rank
r, we have an isomorphism
f ∗ (det(E)) det( f ∗ (E)).
(iv) From a short exact sequence
0→F →E→G→0
of locally free sheaves, we obtain the isomorphism
det(F ) ⊗ det(G) det(E).
Definition 1.48. Let E be a locally free sheaf of rank r on a smooth projective curve C.
Then deg(E) := deg(det(E)) is called the degree of E, where the degree of the line bundle
det(E) is the degree of the corresponding Weil divisor. If C is a projective curve with an
ample invertible sheaf OC (1) the degree of C is defined as deg(C) := deg(OC (1)).
Example 1.49. We have deg(OC (m)) = m · deg(OC (1)) = m · deg(C).
Remark 1.50.
(i) Note that one could also define the degree of E as
χ(E) − r · χ(OX ) = χ(E) − r(1 − g),
where χ denotes the Euler characteristic. This definition is equivalent to our by
[Pot97, Theorem 2.6.9].
(ii) From the alternative definition it becomes clear that the degree is additive on
short exact sequences (see [Mar81, Lemma 1.5 (2)]).
(iii) By [Mar81, Lemma 1.5 (4)] we have the formula
deg(E ⊗ F ) = rank(E) deg(F ) + rank(F ) deg(E).
We close this subsection by stating two prominent theorems for locally free sheaves.
Theorem 1.51 (Serre duality). Let X be a smooth projective variety of positive dimension
V
d defined over an algebraically closed field and let ωX := d ΩX be the canonical bundle
on X. For every locally free sheaf E on X there are natural isomorphisms
Hi (X, E) Hd−i (X, E∨ ⊗ ωX )∗
for all i = 0, . . . , d, where ∗ denotes the dual vector space.
Proof. See [Har87, Corollary III.7.7].
Theorem 1.52 (Riemann-Roch). Let C be a smooth projective curve of genus g defined
over an algebraically closed field k and let E be a locally free sheaf on C. Then we have
h0 (C, E) − h1 (C, E) = deg(E) + (1 − g) · rank(E),
where hi (_) := dimk (Hi (_)).
1.2. THE GEOMETRIC VIEWPOINT OF HILBERT-KUNZ THEORY
33
Proof. See [Pot97, Theorem 2.6.9].
1.2.2. Frobenius. In this section we define the (absolute) Frobenius morphism.
Definition 1.53. Let X be a scheme over a field k of positive characteristic p. Then
the map FX : X → X given by the identity on the topological spaces and the Frobenius
homomorphism on the section rings, is called the (absolute) Frobenius. Note that we will
write F instead of FX if we deal only with one scheme.
Lemma 1.54.
(i) Let g : X → Y be a morphism of schemes over Z/(p). Then the
equality FY ◦ g = g ◦ FX holds.
(ii) For x ∈ X, we have FX (x) = x.
(iii) If all stalks of X are regular rings, then FX is flat.
(iv) If X is integral and geometrically reduced, we have deg(FX ) = pdim(X) .
Proof. See [Liu06, Lemma 3.2.22] for parts (i) and (ii). Part (iii) follows from Kunz’
Theorem 1.7. Part (iv) is Corollary 3.2.27 of [Liu06].
Since the absolute Frobenius acts by x 7→ x p on k, the morphism FX : X → X is k-linear
if and only if k = Z/(p). In the general case, one obtains a k-linear morphism by the
following construction. First, form the pull-back
pr2
X (p) := X ×Spec(k) Spec(k) → Spec(k)
with respect to the structure map f : X → Spec(k) and the absolute Frobenius FSpec(k) (see
the diagram below).
Secondly, one obtains by the universal property of pull-backs (and part (i) of the previous
lemma) a morphism Frel : X → X (p) of varieties over k, which we call relative Frobenius
or k- linear Frobenius.
The following diagram shows all involved maps and schemes.
X
FX
Frel
'
X (p)
f
$
/)
pr1
pr2
Spec(k)
FSpec(k)
/
X
f
Spec(k)
If X is an affine variety Spec(k[T 1 , . . . , T n ]/I) over k, then the absolute Frobenius acts
on OX by sending g ∈ k[T 1 , . . . , T n ]/I to g p , while the relative Frobenius acts on OX by
raising only the (images of the) variables T i to their p-th power (compare [Lie13, Section
2.1]). See [Liu06, Section 3.2.4] for a more general definition of the absolute and relative
Frobenius morphisms.
Note that if k is perfect, then X and X (p) are isomorphic as schemes over k, since their
structural morphisms differ only by an isomorphism.
34
1. A SURVEY ON HILBERT-KUNZ THEORY
Definition 1.55. Let F : X → X be the absolute Frobenius morphism and G an OX -module.
We call the pull-back F∗ (G) of G defined by the following diagram
/
F∗ (G)
X
F
/
G
X
the Frobenius pull-back of G.
For line bundles the Frobenius pull-back is easy to compute.
Lemma 1.56. If X is separable and k is algebraically closed, we have
F∗ (L) L p
for every invertible sheaf L on X.
Proof. See [BK05, Lemma 1.2.6].
We end up this section by defining the Hasse invariant of elliptic curves (compare with
[Har87, Chapter IV.4]). Let X be an elliptic curve over a perfect field k of positive characteristic p. Then the absolute Frobenius F : X → X induces a map
H1 (F) : H1 (X, OX ) → H1 (X, OX )
on cohomology. This map fulfills
H1 (F)(λ · a) = λ p · H1 (F)(a)
for all λ ∈ k and a ∈ H1 (X, OX ). Since X is elliptic, we have dimk (H1 (X, OX )) = 1. Thus,
the map H1 (F) has to be either 0 or bijective, since k is perfect.
Definition 1.57. The Hasse invariant of X is zero if H1 (F) is the zero map and one if
H1 (F) is bijective.
1.2.3. Syzygy bundles. In this subsection we will turn our attention to syzygy bundles, since they give a nice class of examples of vector bundles and because exactly these
are the vector bundles that will appear in the application of vector bundles to Hilbert-Kunz
theory.
Definition 1.58. Let k be a field and R an affine k-algebra. Let f1 , . . . , fn ∈ R and X :=
Spec(R).
(i) We call the kernel of the map
f1 ,..., fn
OnX −→ OX
the sheaf of syzygies for f1 , . . . , fn and denote it by SyzX ( f1 , . . . , fn ).
(ii) Let IZ be the ideal sheaf of the closed subscheme Z := V( f1 , . . . , fn ) ⊆ X. We
call the short exact sequence
f1 ,..., fn
0 −→ SyzX ( f1 , . . . , fn ) −→ OnX −→ IZ −→ 0
the presenting sequence of the sheaf of syzygies.
1.2. THE GEOMETRIC VIEWPOINT OF HILBERT-KUNZ THEORY
35
Definition 1.59. Let k be a field and R a standard-graded, affine k-algebra. Let f1 , . . . , fn ∈
R be homogeneous elements with deg( fi ) = di . Let Y := Proj(R).
(i) We call the kernel of the map
n
M
f1 ,..., fn
OY (−di ) −→ OY
i=1
the sheaf of syzygies for f1 , . . . , fn and denote it by SyzY ( f1 , . . . , fn ).
(ii) Let IZ be the ideal sheaf of the closed subscheme Z := V+ ( f1 , . . . , fn ) ⊆ Y. We
call the short exact sequence
0 −→ SyzY ( f1 , . . . , fn ) −→
n
M
f1 ,..., fn
OY (−di ) −→ IZ −→ 0
i=1
the presenting sequence of the sheaf of syzygies.
Note that in both situations the sheaf of syzygies SyzX ( f1 , . . . , fn ) resp. SyzY ( f1 , . . . , fn )
is nothing but the sheafification of the (graded) R-module SyzR ( f1 , . . . , fn ). Therefore
we may define in the graded case for any m ∈ Z the sheaf of syzygies of total degree
m of f1 , . . . , fn , denoted by SyzY ( f1 , . . . , fn )(m), as the sheafification of the R-module
SyzR ( f1 , . . . , fn )(m). Note that the equality
SyzY ( f1 , . . . , fn )(m) = SyzY ( f1 , . . . , fn ) ⊗ OY (m)
holds, since R is standard-graded (for counter-examples in the non standard-graded case
see for example [Dol82, Section 1.5]). Moreover, if R has dimension at least two and
satisfies S 2 the sheaf SyzY ( f1 , . . . , fn )(m) encodes the homogeneous component of degree
m of the R-module SyzR ( f1 , . . . , fn ). We get this homogeneous component back by taking
global sections
Γ Y, SyzY ( f1 , . . . , fn )(m) SyzR ( f1 , . . . , fn )m
(cf. [Har87, Exercise III.3.5]).
Example 1.60. Let R := k[X0 , . . . , Xn ] and Y := Proj(R) = Pnk . Then there is a short exact
sequence (cf. [Har87, Theorem II.8.13])
X0 ,...,Xn
0 −→ ΩY/k −→ On+1
Y (−1) −→ OY −→ 0,
which arises from the Euler sequence. This sequence is a presenting sequence of ΩY/k and
of SyzY (X0 , . . . , Xn ), hence they are isomorphic and SyzY (R+ ) is nothing but the cotangent
bundle on Y. If Y 0 ⊆ Y is a closed subvariety, the sheaf of syzygies SyzY 0 (X0 , . . . , Xn ) on
Y 0 is just the restriction of the cotangent bundle ΩY/k to Y 0 .
Recall some basic properties of syzygy modules.
Lemma 1.61. Let R be a standard-graded, affine k-domain and let f, f1 , . . . , fn ∈ R be
homogeneous and f non-zero. The following statements hold.
(i) Let g1 , . . . , gm ∈ ( f1 , . . . , fn ) be homogeneous elements with g j =
defines a graded R-module homomorphism
(hi j )i j
SyzR (g1 , . . . , gm ) −→ SyzR ( f1 , . . . , fn ).
Pn
i=1 hi j · fi .
This
36
1. A SURVEY ON HILBERT-KUNZ THEORY
(ii) The following isomorphism holds
SyzR ( f f1 , . . . , f fn ) SyzR ( f1 , . . . , fn )(− deg( f )).
(iii) If f, fn is an R-regular sequence, we have
SyzR ( f f1 , . . . , f fn−1 , fn ) SyzR ( f1 , . . . , fn )(− deg( f )).
(iv) If deg( f fi ) = deg( fn ) for some i ∈ {1, . . . , n − 1}, we have
SyzR ( f1 , . . . , fn ) SyzR ( f1 , . . . , fn−1 , fn + f fi ).
Proof. Since the parts (i), (ii) and (iv) are obvious, we only prove part (iii). Let
(s1 , . . . , sn ) ∈ SyzR ( f f1 , . . . , f fn−1 , fn )
be a syzygy of total degree m. We obtain
0=
n−1
X
si f fi + sn fn ≡ sn fn modulo f.
i=1
This forces sn ∈ ( f ), since fn is a non-zero divisor in R/( f ). But then (s1 , . . . , sn−1 , sn / f )
is a syzygy for f1 , . . . , fn of total degree m − deg( f ). On the other hand, if (t1 , . . . , tn ) is a
syzygy for f1 , . . . , fn of total degree m0 , then (t1 , . . . , tn−1 , f tn ) is a syzygy for the elements
f f1 , . . . , f fn−1 , fn of total degree m0 + deg( f ). These operations lead to two homomorphisms that are inverse to each other.
Proposition 1.62. Let k be an algebraically closed field and R a standard-graded kalgebra of dimension at least two and Y := Proj(R). Let f1 , . . . , fn ∈ R be homogeneous
elements with deg( fi ) = di . Assume that at least one of the fi is a non-zero divisor. Then
the following statements hold.
(i) The rank of SyzY ( f1 , . . . , fn ) is n − 1.
(ii) If the ideal generated by the fi is R+ -primary, then the morphism
n
M
f1 ,..., fn
OY (−di ) −→ OY
i=1
is surjective and SyzY ( f1 , . . . , fn ) is locally free.
(iii) The sheaf SyzY ( f1 , . . . , fn ) is locally free on U := D+ ( f1 , . . . , fn ) ⊆ Y.
(iv) If Y is an irreducible curve and the ideal ( f1 , . . . , fn ) is R+ -primary, then
Ñ
det(SyzY ( f1 , . . . , fn )(m)) OY
(n − 1)m −
n
X
é
di .
i=1
In particular, we have deg(SyzY ( f1 , . . . , fn )(m)) = (n − 1)m − ni=1 di · deg(Y).
(v) Let ϕ : R → S be a morphism of degree d of normal, standard-graded k-domains
of dimension two. Let g : Y 0 := Proj(S ) → Y be the corresponding morphism of
smooth projective curves. Then the following holds
P
g∗ (SyzY ( f1 , . . . , fn )(m)) SyzY 0 (ϕ( f1 ), . . . , ϕ( fn ))(dm).
1.2. THE GEOMETRIC VIEWPOINT OF HILBERT-KUNZ THEORY
37
(vi) Assume that char(k) = p > 0 and that Y is an irreducible curve. For the iterated
pull-backs under the Frobenius, we have
q
Fe∗ (SyzY ( f1 , . . . , fn )(m)) SyzY ( f1 , . . . , fnq )(mq),
where q = pe and e ∈ N.
Proof.
(i) Since f1 , . . . , fn contains a non-zero divisor, the rank of IZ (on every
irreducible component of Y) is one. Since the rank is additive on short exact sequences, the statement follows from the presenting sequence of SyzY ( f1 , . . . , fn ).
(ii) Since ( f1 , . . . , fn ) is R+ -primary, the ideal sheaf IZ is just OY . The supplement
follows, since kernels of surjective maps between locally free sheaves are locally free.
(iii) On the affine patches D+ ( fi ), the map
OY (−d j )|D+ ( fi ) −→ SyzY ( f1 , . . . , fn )|D+ ( fi ) ,
M
j∈{1,...,n}, j,i
Ç
(g1 , . . . , gi−1 , gi+1 , . . . , gn ) 7−→ g1 , . . . , gi−1 , −
P
å
j∈{1,...,n}, j,i f j g j
, gi+1 , . . . , gn
fi
is an isomorphism.
(iv) By part (ii) the presenting sequence of SyzY ( f1 , . . . , fn ) ends in OY . We obtain
det(SyzY ( f1 , . . . , fn )(m)) ⊗ OY (m) det
Ñ
n
M
Ñ
OY
é
OY (m − di )
i=1
nm −
n
X
é
di .
i=1
Tensoring with OY (−m) gives the result. The supplement is clear.
(v) The statement follows by taking the pull-back of the presenting sequence of
SyzY ( f1 , . . . , fn ) and by [Har87, Proposition II.6.9].
(vi) Similarly to (v), the statement follows by taking the Frobenius pull-back of the
presenting sequence of SyzY ( f1 , . . . , fn ) using Lemma 1.56.
Remark 1.63. In the affine situation the analogues of parts (i), (iii), (v) and (vi) of the
above proposition are still true.
1.2.4. Stability. An important property of vector bundles is the semistablility. It is
essential to construct meaningfull moduli spaces (see [HL97]).
Definition 1.64. Let C be a smooth projective curve defined over an algebraically closed
deg(E)
field and let E be a locally free OC -module. We call the quotient µ(E) := rank(E)
the slope
of the sheaf E.
Definition 1.65. Let C be a smooth projective curve defined over an algebraically closed
field and let E be a locally free OC -module. We call E semistable if for every locally free
subsheaf 0 , F ( E the inequality µ(F ) ≤ µ(E) holds. We say that E is stable if the above
inequality is always strict.
38
1. A SURVEY ON HILBERT-KUNZ THEORY
Note that our definition of stability is due to Mumford (another not equivalent definition
of stability using Euler characteristics is due to Gieseker).
Example 1.66. Let C ( P2k be a smooth plane curve of degree at least two defined over an
algebraically closed field k. Then ΩP2 /k |C is semistable as shown in [Bre05a, Proposition
k
6.2] and [Tri05a, Corollary 3.5].
Proposition 1.67. Let C be a smooth projective curve defined over an algebraically
closed field and let E, F be locally free OX -modules. The following statements hold.
(i) Line bundles are stable.
(ii) If E and F are semistable then E ⊕ F is semistable if and only if µ(F ) = µ(E).
(iii) The OX -module E is (semi-)stable if and only if for every line bundle L the sheaf
E ⊗ L is (semi-)stable.
Proof. To prove (i) assume that the line bundle L is not semistable. Then there exists
a line bundle E ( L with deg(E) > deg(L). The inclusion
0→E→L
leads to the inclusion
0 → O → L ⊗ E∨ .
Because O injects into L ⊗ E∨ its degree deg(L ⊗ E∨ ) = deg(L) − deg(E) has to be nonnegative contradicting deg(E) > deg(L).
To prove part (ii) assume that E ⊕ F is semistable. Then E ⊕ 0 is a locally free subsheaf of
E ⊕ F . The inequality
µ(E ⊕ 0) = µ(E) ≤ µ(E ⊕ F )
is equivalent to µ(E) ≤ µ(F ). Replacing E by F gives µ(F ) ≤ µ(E), hence µ(E) = µ(F ).
Now let µ(E) = µ(F ). It is easy to see that µ(E ⊕ F ) = µ(E) = µ(F ) holds. Assume
that E ⊕ F is not semistable. Then there exists a locally free subsheaf G ( E ⊕ F with
µ(G) > µ(E ⊕ F ). From the short exact sequence
0 → E → E⊕F → F → 0
we obtain locally free subsheaves G0 := ker(G → F ) and G00 := im(G → F ) of E resp. F
and the short exact sequence
(1.1)
0 → G0 → G → G00 → 0.
If G0 = 0, we obtain the contradiction
µ(F ) ≥ µ(G00 ) = µ(G) > µ(E ⊕ F ) = µ(F ).
Similarly, if G00 = 0, one gets µ(E) > µ(E). If G0 and G00 are both non-zero, one obtains
from (1.1) the contradiction
max(µ(G0 ), µ(G00 )) ≥ µ(G) > µ(E ⊕ F ) = µ(E) = µ(F )
since E and F are semistable.
To prove part (iii) fix a line bundle L. If 0 , F ( E contradicts (semi-)stability, then
0 , F ⊗ L ( E ⊗ L contradicts (semi-)stability. Similarly, if 0 , F 0 ( E ⊗ L contradicts
(semi-)stability, then 0 , F 0 ⊗ L∨ ( E contradicts (semi-)stability.
1.2. THE GEOMETRIC VIEWPOINT OF HILBERT-KUNZ THEORY
39
Theorem 1.68 (Grothendiek). Let E be a vector bundle of rank r on P1k , where k is an
algebraically closed field. Then there are uniquely determined integers a1 ≥ . . . ≥ ar with
E
r
M
i=1
OP1 (ai ).
k
Proof. See [HL97, Theorem 1.3.1].
Using this theorem and the last proposition, one gets immediately a classification of all
(semi-)stable vector bundles on the projective line:
Corollary 1.69. Let E be a vector bundle on P1k , where k is an algebraically closed field.
Then E is semistable if and only if all ai in Grothendiecks Theorem are equal. Moreover,
a vector bundle is stable if and only if it is a line bundle.
The next theorem shows that one can find for a given vector bundle a filtration with
semistable quotients.
Theorem 1.70 (Harder, Narasimhan). Let C be a smooth projective curve. Then every
locally free sheaf E over C has a unique filtration
0 = E0 ( E1 ( . . . ( Et = E
satisfying the following properties
(i) all quotients Ei /Ei−1 are semistable and
(ii) the slopes µi := µ(Ei /Ei−1 ) form a strictly decreasing chain µ1 > . . . > µt .
Proof. See [HN75, Lemma 1.3.7].
Definition 1.71. The filtration from the previous theorem is called Harder-Narasimhan
filtration. The sheaf E1 in the filtration is called the maximal destabilizing subsheaf of
E and the quotient E/Et−1 is called the minimal destabilizing quotient of E. The corresponding slopes µmax (E) := µ1 and µmin (E) := µt are called maximal slope resp. minimal
slope of E.
Working in positive characteristic, it is not true that Frobenius pull-backs of semistable
vector bundles are again semistable. A counter example to this can be found on the curve
C := Proj(k[X, Y, Z]/(X 4 + Y 4 + Z 4 )),
where k is an algebraically closed field with char(k) = 3. Then SyzC (X, Y, Z) is semistable,
but F∗ (SyzC (X, Y, Z)) is not as shown in [BK13, Example 3.1]. The following definition is
due to Miyaoka (cf. [Miy87, Section 5]).
Definition 1.72. Let C be a smooth projective curve over an algebraically closed field k
of positive characteristic. A locally free sheaf E over C is called strongly semistable, if all
its Frobenius pull-backs Fe∗ (E), e ≥ 0, are semistable.
Definition 1.73. We call a Harder-Narasimhan filtration a strong Harder-Narasimhan
filtration if all quotients are strongly semistable.
40
1. A SURVEY ON HILBERT-KUNZ THEORY
The next theorem shows that for every locally free sheaf there is a Frobenius pull-back
with a strong Harder-Narasimhan filtration. In fact the Frobenius pull-backs stabilize in a
certain way (see the next remark for the precise statement).
Theorem 1.74 (Langer). Let C be a smooth projective curve over an algebraically closed
field k with char(k) = p > 0. For every locally free sheaf E there exists a (minimal) natural
number e0 such that the Harder-Narasimhan filtration of Fe0 ∗ (E) is strong.
Proof. See [Lan04, Theorem 2.7].
Remark 1.75. Let 0 ( E1 ( . . . ( Et−1 ( Fe0 ∗ (E) be the strong Harder-Narasimhan filtration from Langers Theorem. Since Frobenius pull-backs respect inclusions and commute
with quotients, the filtration
0 ( Fe−e0 ∗ (E1 ) ( . . . ( Fe−e0 ∗ (Et−1 ) ( Fe∗ (E)
is a strong Harder-Narasimhan filtration for Fe∗ (E) for all e ≥ e0 .
1.2.5. Relation to Hilbert-Kunz theory. In this subsection we explain how the theory of vector bundles arises in Hilbert-Kunz theory. The next proposition will be crucial.
Proposition 1.76. Let k be an algebraically closed field of positive characteristic p, let
R be a standard-graded k-domain of dimension at least two satisfying S 2 and let Y =
Proj(R). Let I = ( f1 , . . . , fn ) be an R+ -primary ideal with deg( fi ) = di . Then for every
q = pe ≥ 1 and m ∈ Z the following formula holds
Ä
dimk R/I [q]
ä
= h0 (Y, OY (m)) −
m
n
X
q
h0 (Y, OY (m − q · di )) + h0 (Y, SyzY ( f1 , . . . , fnq )(m)).
i=1
Proof. Starting with the exact sequence
q
q
n
X
f1 ,..., fn
q
q
0 −→ SyzY ( f1 , . . . , fn )(m) −→
OY (m − q · di ) −→
i=1
OY (m) −→ 0,
we obtain by taking global sections the exact sequence
(1.2)
q
q
n
X
f1 ,..., fn
q
q
0 −→ Γ(Y, SyzY ( f1 , . . . , fn )(m)) −→
Rm−q·di −→
i=1
Rm ,
where Γ(Y, OY (l)) = Rl for any l ∈ Z, since
III.3.5]).
Ä
äR satisfies S 2 (cf. [Har87,
Ä Exercise
ä
Obviously, the image of the last map is I [q] m , hence we may add → R/I [q] m → 0 at the
right hand side
of (1.2).
Computing the alternating sum of the k-vector space dimensions
ä
P Ä
the sum m R/I [q] m is finite since I [q] is R+ -primary and it is exactly HK(I, R, q).
At this point we should mention that - although Proposition 1.76 holds for rings of higher
dimension than two - there are no examples in dimension at least three in which the above
proposition is helpfull. The reason for this is that second syzygy modules of quotients
R/I do not behave well if dim(R) ≥ 3. In fact, the dim(R)-th syzygy modules of R/I are
the correct modules to deal with. This is emphasized by a recent work of Brenner. Using
more general tools from algebraic geometry than we presented, he was able to prove the
following theorem (compare with [Bre13]).
1.2. THE GEOMETRIC VIEWPOINT OF HILBERT-KUNZ THEORY
41
Theorem 1.77 (Brenner). Let R be a standard-graded Cohen-Macaulay domain of dimension d + 1 ≥ 2 with an isolated singularity over an algebraically closed field k of
positive characteristic p. Let Y := Proj(R) and let I ( R be an R+ -primary ideal. Let
δ2
δ1
. . . → F2 → F1 → F0 → R/I → 0
(1.3)
be a graded complex with Fi := j∈Ji R(−βi j ). Let SyzdR := ker(δd ). If the complex on
Spec(R) \ {R+ } associated to (1.3) is exact, we have
L
HK(I, R, pe ) =
X
m∈N
hd (Fe∗ (SyzdY )(m)) +
d
X
Ñ
(−1)d−1−i
i=0
Ñ
éé
X
X
j∈Ji
m∈N
hd (OY (−βi j pe + m))
.
In the next chapter we will give some examples how one can compute Hilbert-Kunz functions with Proposition 1.76.
Starting with Proposition 1.76, using Langer’s Theorem 1.74 and properties of strongly
semistable vector bundles, Brenner and Trivedi computed independently the Hilbert-Kunz
multiplicity of two-dimensional, standard-graded, normal domains containing an algebraically closed field of positive characteristic. To state their result, we need some more
notation.
Notation 1.78. In the situation of Proposition 1.76 assume dim(R) = 2 and that Y is
smooth. Then there exists by Langer’s Theorem 1.74 an e0 such that for every e ≥ e0 the
e-th Frobenius pull-back of the sheaf SyzY ( f1 , . . . , fn ) has a strong Harder-Narasimhan
filtration
0 = S0 ( S1 ( . . . ( St−1 ( St = Fe∗ (SyzY ( f1 , . . . , fn )).
For i = 1, . . . , t, we define the numbers
µ(Si /Si−1 )
and ri := rank(Si /Si−1 ).
νi := −
deg(Y) · pe
Note that these numbers are independent of e by Remark 1.75.
Theorem 1.79 (Brenner/Trivedi). In the situation of Proposition 1.76 assume dim(R) = 2
and that Y is smooth. Then the Hilbert-Kunz multiplicity of R with respect to I is given by
Ñ
é
n
t
X
deg(Y) X
eHK (I, R) =
·
ri νi2 − d2j ,
2
i=1
j=1
where the ri and νi are computed with respect to SyzY (I).
Proof. See [Bre06, Theorem 3.6] for the general statement and [Tri05a, Theorem
4.12] for the case I = R+ .
We will state some generalizations of Theorem 1.79.
Theorem 1.80 (Trivedi). Drop the assumption in Theorem 1.79 that Y is smooth. Then
the inequality
Ñ P
é
n
n d 2 X
deg(Y)
i
2
i=1
eHK (I, R) ≥
− di
2
n−1
i=1
42
1. A SURVEY ON HILBERT-KUNZ THEORY
holds. Equality holds if and only if the pull-back of SyzY (I) to the normalization of Y
is strongly semistable. In particular, if Y is already smooth, equality holds if and only if
SyzY (I) is strongly semistable.
Proof. See [Tri05b, Theorem 2.2].
Corollary 1.81. Let k be an algebraically closed field of positive characteristic p. Let
R := k[X, Y, Z]/(F) be standard-graded with F irreducible and homogeneous of degree d
such that C := Proj(R) is a smooth projective curve. Let I = ( f1 , f2 , f3 ) ( R be a homogeneous and R+ -primary ideal with all generators having the same degree a ≥ 1. Then the
following statements hold.
(i) If SyzC (I) is not semistable, then there exists an integer 0 < l ≤ ad with l ≡ ad (2)
such that
3d 2 l2
·a + .
eHK (I, R) =
4
4d
(ii) If SyzC (I) is semistable but not strongly semistable, then there exists an integer
0 < l ≤ d(d − 3) with l ≡ pad (2) such that
eHK (I, R) =
3d 2
l2
·a +
,
4
4d · p2e
where e ≥ 1 is such that Fe−1∗ (SyzC (I)) is semistable, but Fe∗ (SyzC (I)) is not.
(iii) If SyzC (I) is strongly semistable, then
3d 2
·a .
eHK (I, R) =
4
Proof. For the case a = 1 see [Tri05a, Theorem 5.3] and [Bre06, Corollary 4.6]. For
the general case see [Kai09, Corollary 1.4.9 (2), Proposition 1.4.11].
Theorem 1.82 (Brenner). In the situation of Theorem 1.79, the Hilbert-Kunz function of
R with respect to I has the shape
HK(I, R, pe ) = eHK (I, R) · p2e + γ(p, e),
where γ(p, _) is a bounded function. Moreover, if k is the algebraic closure of a finite
field, the function γ(p, _) is eventually periodic.
Proof. See [Bre07, Theorem 6.1].
1.2.6. Curves of degree three. Using the just explained geometric approach, Brenner and Hein computed the Hilbert-Kunz function of the homogeneous coordinate ring R
associated to a very ample line bundle OC (1) of an elliptic curve C over an algebraically
closed field k. See [BH06, Corollary 4.6] for the case where SyzC ( f1 , . . . , fn ) is semistable
and [BH06, Theorem 4.4] for the general case. In the same paper they also look more
carefully at the case, where C is embedded by a complete linear system |OC (1)| into some
projective space PkN . In [BH06, Theorem 5.2] they give a full discription of the Hilbert2
Kunz function of R. In particular, the Hilbert-Kunz multiplicity of R is eHK (R) = (N+1)
2N .
We just state the explicit results of the Hilbert-Kunz functions in the cases N = 2, 3, 4,
1.2. THE GEOMETRIC VIEWPOINT OF HILBERT-KUNZ THEORY
43
given in [BH06, Corollaries 5.3-5.6]. Note that these results had already been proved in
[FT03, Theorem 3.17].
Proposition 1.83. Let C be an elliptic curve over an algebraically closed field k of characteristic p > 0, embedded by a complete linear system into PkN . Then the Hilbert-Kunz
function of the homogeneous coordinate ring R of C is given by the following formulas.
(i) Let N = 2 and let h be the Hasse invariant of C. Then
9 2 5
q −
4
4
HKR (R, q) = 8 − h
9
q2 − h
4
if q is odd,
if q = 2,
otherwise,
(ii)
8 q2 − 5
if N = 3 and p , 3,
3
HKR (R, q) = 3
25 2 17
q −
if N = 4 and p , 2.
8
8
The Hilbert-Kunz functions of the homogeneous coordinate rings of curves of degree
three have been studied by many authors using a whole bunch of different tools. In his
thesis (cf. [Par94]), Pardue made a conjecture on the Hilbert-Kunz functions of the homogeneous coordinate rings of plane curves of degree three. This conjecture was proven
case by case by different authors. We summarize the results and references.
If C is an irreducible nodal cubic, Pardue showed in his thesis (cf. [Par94] and also [BC97,
Theorem 3])
7
1
HK(R, q) = q2 − q − D(q)
3
3
with D(q) = 53 , if q ≡ 2 (3) and D(q) = 1 otherwise. Using Burbans results from [Bur12] on
the classification of Frobenius pull-backs of line bundles on plane rational nodal curves,
Monsky showed in [Mon11], that the Hilbert-Kunz function of R with respect to an mprimary ideal I is of the form HK(I, R, q) = µq2 + αq + D(q), where µ and α are constants
depending on the data of SyzC (I) and D(q) depends on these data and on the class of q
modulo three.
If C is cuspidal, we have
7
HK(R, 3e ) = 9e
3
and
7
4
HK(R, q) = q2 −
3
3
for p , 3 (see [Par94]). If C is smooth, the Hilbert-Kunz function of R was computed
for odd characteristics in [BC97, Theorem 4] based on Geronimus’ formula for the determinant of a matrix, whose entries are Legendre polynomials. In characteristic two,
one might use the algorithm from [HM93] if the Hasse invariant is zero. If the Hasse invariant is one, Monsky computed the Hilbert-Kunz function in [Mon97] using tools from
invariant theory.
44
1. A SURVEY ON HILBERT-KUNZ THEORY
1.3. Related invariants and further readings
In this section we want to discuss some invariants related to Hilbert-Kunz theory. We also
give a list of references related to Hilbert-Kunz theory that have not been mentioned in
this survey.
1.3.1. Limit Hilbert-Kunz multiplicity. The basic question in this subsection is
whether one can define a “Hilbert-Kunz multiplicity” in characteristic zero. We have
already defined the generalized Hilbert-Kunz function n 7→ λR (R/( f1n , . . . , fmn )R). From
this definition one might get the idea to study the limit
λR (R/( f1n , . . . , fmn )R)
lim
,
n→∞
ndim(R)
provided it exists. Consider for example R := Z[X, Y, Z]/(X + Y + Z). Then by [HM93,
Theorem 2.14] we obtain
3n2
.
dimQ (Q ⊗Z R/(X , Y , Z )) =
4
'
&
n
n
n
Dividing this number by n2 and taking the limit n → ∞, we get
dimQ (Q ⊗Z R/(X n , Y n , Z n ) 3
lim
= .
n→∞
n2
4
But since Q ⊗Z R Q[X, Y] is regular, we want the “Hilbert-Kunz multiplicity” to be one.
Hence the above limit cannot be a good candidate.
Definition 1.84. Let R be a positively-graded Z-domain and let I = ( f1 , . . . , fm ) be a homogeneous R+ -primary ideal. For a prime p let R p := R ⊗Z Z/(p) and denote by I p the
extended ideal IR p . We call the limit (provided it exists)
e∞
lim eHK (I p , R p )
HK (I, R) := p→∞
the limit Hilbert-Kunz multiplicity of R with respect to I.
The existence of e∞
HK (m, R), where R is a diagonal hypersurface over Z was shown by
Gessel and Monsky in [GM10]. To state their result, we need one more definition.
Definition 1.85. Let d1 , . . . , d s be positive integers. For any integer λ we define
Ç
å s−1
ε1
εs
(ε1 · . . . · ε s ) ·
+ . . . + − 2λ
d1
ds
,...,ε )∈{±1} s ,
X
Cλ :=
(ε1
.
s
ε1
εs
d1 +...+ d s >2λ
P
s X di with positive inteTheorem 1.86 (Gessel, Monsky). Let R := Z[X1 , . . . , X s ]/ i=1
i
Q
gers di and let d := di . The limit Hilbert-Kunz multiplicity of R is given by
e∞
HK (R) =
d(21−s )
(s − 1)!
Proof. See [GM10, Theorem 2.4].
Ñ
· C0 + 2
é
X
Cλ .
λ>0
1.3. RELATED INVARIANTS AND FURTHER READINGS
45
Another big class of rings for which the limit Hilbert-Kunz multiplicity is known to exist, is given by standard-graded, two-dimensional k-domains, where k is an algebraically
closed field of characteristic zero. To state the theorem, we need to define spreads.
Definition 1.87. Let R be a finitely generated, positively-graded, two-dimensional domain over an algebraically closed field k of characteristic 0 and let I be a homogeneous
R+ -primary ideal. There exist a finitely generated Z-algebra A ⊆ k, a finitely generated,
positively-graded algebra RA over A and a homogeneous ideal IA ( RA with RA ⊗A k = R
and such that the following properties hold for all closed points s ∈ Spec(A)
(i) the ring R s := RA ⊗A κ(s) is a finitely generated, positively-graded, two-dimensional domain over κ(s),
(ii) the rings R s are normal, if R is and
(iii) the ideal I s := im(IA ⊗A κ(s)) ( R s is homogeneous and (R s )+ -primary.
We call the data (A, RA , IA ) a spread of the pair (R, I). Note that the spread (A, RA , IA ) can
be chosen such that A contains a given finitely generated k-subalgebra A0 .
Theorem 1.88 (Trivedi). Let R be a standard-graded, two-dimensional domain over an
algebraically closed field k of characteristic 0. Let I ( R be a homogeneous R+ -primary
ideal and let (A, RA , IA ) be a spread of the pair (R, I). Denote by s0 the generic point of
Spec(A). Then the limit
e∞
HK (I, R) = lim eHK (I s , R s )
s→s0
over the closed points s ∈ Spec(A) exists. Moreover, if I = ( f1 , . . . , fn ) with deg( fi ) = di ,
this limit equals
deg(Proj(S ))
2
Ñ
t
X
i=1
Ç
ri
µi
deg(Proj(S ))
å2
−
n
X
é
di2
,
i=1
where S is the normalization of R and the numbers ri , µi are the ranks and slopes of the
consecutive quotients of the bundles in the Harder-Narasimhan filtration of SyzProj(S ) (IS ).
Proof. See [Tri07, Theorem 2.4].
Trivedi remarks (cf. [Tri07, Remark 2.6]) that all eHK (I s , R s ) are bounded below by
e∞
HK (I, R) and that equality holds for a closed point s ∈ Spec(A) if and only if the HarderNarasimhan filtration of SyzProj(Rs ) (I s ) is strong.
At least in this situation, the limit Hilbert-Kunz multiplicity behaves to solid closure like
the Hilbert-Kunz multiplicity to tight closure (cf. [Bre04, Theorem 3.3]).
The result of Trivedi was recovered in [BLM12] for standard-graded, flat Z-domains R,
where almost all fiber rings R p := R ⊗Z Z/(p) are geometrically normal, two-dimensional
domains. Moreover, they show that in this situation, one does not have to compute all
eHK (I p , R p ) to compute e∞
HK (I, R). In fact, one can skip the limit in the first number.
Theorem 1.89 (Brenner, Li, Miller). In the above situation, one has
[pe ]
e∞
HK (I, R) = lim
p→∞
λR p R p /I p
p2e
46
1. A SURVEY ON HILBERT-KUNZ THEORY
for every fixed integer e ≥ 1.
1.3.2. F-Signature. In this subsection we will define the F-signature (function) and
connect them to Hilbert-Kunz theory. For further properties of F-signature and explicit
examples we refer to the literature.
Definition 1.90. Let (R, m, k) be a reduced local ring of dimension d and characteristic
p > 0. Assume that k is perfect and that R is F-finite (meaning that F : R → R is a finite
map). For any e ∈ N denote by aq the maximal rank of a free submodule of R1/q . Then
the function
e 7→ Fsign (R, q) := aq
is called the F-signature function of R. The limit
Fsign (R, q)
s(R) := lim
e→∞
qd
(provided it exists) is called the F-signature of R.
The F-signature was firstly studied in [SdB97] and the existence was proven by Tucker
in [Tuc12, Theorem 4.9]. Yao has shown in [Yao06, Proposition 4.1] that the F-signature
of R is nothing but the minimal relative Hilbert-Kunz multiplicity of R, that is
√
inf{eHK (I, R) − eHK (J, R)|I ( J, I = m, λR (J/I) = 1}
and was defined in [WY04].
The following theorem gives a connection between the F-signature function and the
Hilbert-Kunz function.
Theorem 1.91 (Huneke, Leuschke). Let (R, m) be Gorenstein, let I be an m-primary ideal
generated by a system of parameters and J := (I, y), where y ∈ R is a representative for
the socle of R/I. We then get the equality
Fsign (R, q) = HK(I, R, q) − HK(J, R, q).
Proof. See [HL02, Theorem 11(2)].
1.3.3. Further readings. We give some more references on papers related to HilbertKunz theory that are not mentioned so far.
(i) In [Fie00] Fields investigates the theory of quasi-polynomial functions to length
functions of generalized Tor-modules. This theory has applications to HilbertKunz functions.
(ii) More on Hilbert-Kunz functions of hypersurfaces can be found in [Che98],
[Upa12a], [Upa12b], [Upa12c] and [Upa12d].
(iii) For more explicit examples of Hilbert-Kunz functions (in characteristic two) see
[Mon98a], [Mon98b] and [Mon05].
(iv) The geometric approach to Hilbert-Kunz theory is used in [FT03] and [Mon07].
(v) For a deeper analysis of Han’s δ-function and similar functions, especially of
their fractal structure, see [MT04], [MT06], [Tei02] and [Tei12].
(vi) More bounds for Hilbert-Kunz multiplicities can be found in [AE12], [AE08],
[BE04], [CDHZ12] and [Tuc12].
1.3. RELATED INVARIANTS AND FURTHER READINGS
47
(vii) For more on Frobenius functors and Frobenius pull-backs in higher dimensions
see [Dut83], [Dut89], [Dut96], [Sei89] and [Tri12].
(viii) For some generalizations of Hilbert-Kunz functions and multiplicities for example to ideals that are not m-primary see [DS13], [EY11] and [EY].
(ix) More properties and explicit examples of F-signatures can be found in [AE05],
[AL03], [Can12], [BST12a], [BST12b], [Kor11], [Sin05] and [SW11].
(x) More about tight and solid closure and their interactions with Hilbert-Kunz theory can be found for example in [HH90], [Hoc94], [Hun96], [Bre03], [Bre04]
and [Bre08].
(xi) Results on the Hilbert-Kunz function of maximal Cohen-Macaulay modules can
be found in [CK12], [Sei97] and [WY01a].
CHAPTER 2
Examples of Hilbert-Kunz functions and Han’s δ function
In the first section we will discuss some examples of Hilbert-Kunz functions. The second
section is devoted to a more detailed study of Han’s δ function and a generalization of
a theorem due to Monsky. We will end section two by computing some examples of
Hilbert-Kunz multiplicities with the help of this generalization. In the third section we
will use this generalization to study the behaviour of the Hilbert-Kunz multiplicity in
certain families of trinomial surface rings.
2.1. Examples of Hilbert-Kunz functions
We fix some notations for the next two lemmata in which we will compute formulas
for values of Hilbert-Kunz functions. These formulas will be used to compute HilbertKunz functions of explicit examples (cf. Example 2.6 as well as the computations in the
Chapters 6 and 7). Let R := k[X, Y, Z]/(F) be a normal, standard-graded domain with an
algebraically closed field k of prime characteristic p and deg(F) = d > 0. Moreover, we fix
the R+ -primary ideal I := (X α , Y β , Z γ ) with α, β, γ ∈ N≥1 . Let Q : N3 → Z be the quadratic
form defined as
(n1 , n2 , n3 ) 7−→ 2(n1 n2 + n1 n3 + n2 n3 ) − n21 − n22 − n23 .
Lemma 2.1. Let e ∈ N be such that
(2.1)
Ä
ä
SyzR I [q] = R(−n) ⊕ R(−l) for some n, l ∈ Z.
Then the value at e of the Hilbert-Kunz function of R with respect to I is given by
HK(I, R, pe ) =
d · Q(α, β, γ) · p2e + d · |n − l|2
.
4
Note that the assumption (2.1) means that the quotient R/I [q] has finite projective dimen0
sion and this carries over to all quotients R/I [q ] for q0 ≥ q.
Proof. Computing the degrees of the C := Proj(R)-bundles corresponding to the Rmodules in (2.1), one finds
l + n = q(α + β + γ).
Therefore, we can write l and n as
(2.2)
q(α + β + γ) ± |l − n|
.
2
49
2. EXAMPLES OF HILBERT-KUNZ FUNCTIONS AND HAN’S δ FUNCTION
50
Using Proposition 1.76 and Riemann-Roch 1.52, we obtain
dimk (R/(X αq , Y βq , Z γq ))
= x→∞
lim
+
x
X
0
h (OC (m)) −
m=0
x
X
h0 (OC (m − l)) +
x
X
x
X
0
h (OC (m)) −
m=0
x−l
X
h0 (OC (m)) +
m=0
x
X
h0 (OC (m − γq))
m=0
h0 (OC (m − n))
x−αq
X
0
h (OC (m)) −
x−n
X
x−βq
X
0
h (OC (m)) −
m=0
x−γq
X
h0 (OC (m))
m=0
h0 (OC (m))
m=0
= x→∞
lim
x
X
x
X
x−βq
X
h0 (OC (m)) −
m=x−α·q+1
−
h (OC (m − βq)) −
0
m=0
m=0
= x→∞
lim
x
X
m=0
+
h (OC (m − αq)) −
0
m=0
m=0
= x→∞
lim
x
X
x−βq
X
m=x−l+1
h0 (OC (m)) −
x
X
(dm + 1 − g) +
h1 (OC (m)) −
h (OC (m)) −
(dm + 1 − g) −
m=x−n+1
h0 (OC (m))
x−βq
X
1
m=x−α·q+1
x−γq
X
m=x−n+1
m=x−l+1
m=x−α·q+1
x−γq
X
(dm + 1 − g)
m=x−l+1
x−γq
X
h1 (OC (m)) ,
m=x−n+1
where g in the last equality denotes the genus of C. With the help of Equation (2.2) one
sees that the various sums over dm + 1 − g sum up to the constant
dQ(α, β, γ)q2 + d|l − n|2
.
4
Since both bounds of the sums over h1 (OC (m)) depend linearly on x and h1 (OC (m)) vanishes for large m by a theorem of Serre (cf. [Har87, Theorem III.5.2]) these sums have no
share in the limit over x. All in all, the claim follows.
Lemma 2.2. Let e ∈ N be such that
(2.3)
Ä
ä
SyzR I [q] SyzR (X a , Y b , Z c )(−n)
for some n ∈ Z and some positive integers a, b, c such that at least one of the inequalities a < αq, b < βq, c < γq holds. Computing the degrees of the C := Proj(R)-bundles
corresponding to the R-modules in (2.3), we find
q(α + β + γ) − a − b − c
.
n=
2
With D := dimk (R/(X a , Y b , Z c )) the value at e of the Hilbert-Kunz function of R with respect
to I is given by
HK(I, R, pe ) =
d · Q(α, β, γ) · p2e − d · Q(a, b, c)
+ D.
4
2.1. EXAMPLES OF HILBERT-KUNZ FUNCTIONS
51
Proof. From the presenting sequence of SyzC (X a , Y b , Z c )(m − n) we obtain
h0 (C, SyzC (X a , Y b , Z c )(m − n))
= h0 (C, OC (m − n − a)) + h0 (C, OC (m − n − b)) + h0 (C, OC (m − n − c))
(2.4)
− h0 (C, OC (m − n)) + dimk (R/(X a , Y b , Z c ))m−n .
AsÄin the previous
using Proposition 1.76, Riemann-Roch 1.52 and substituting
ä example,
ä
Ä
h0 C, SyzC I [q] (m) with the right hand side of (2.4), we get
dimk (R/(X αq , Y βq , Z γq ))
= lim
x→∞
−
x
X
h0 (OC (m)) −
m=0
x
X
+
h0 (OC (m − γq)) +
h0 (OC (m − n − c)) −
m=0
h0 (OC (m − n − a)) +
x
X
h0 (OC (m − n − b))
m=0
x
X
h0 (OC (m − n)) + D
x
X
0
h (OC (m)) −
x−βq
X
0
h (OC (m)) −
m=0
m=0
x−n−a
X
x−αq
X
h0 (OC (m)) +
x−n−b
X
h0 (OC (m)) +
m=0
2
d · Q(α, β, γ) · q − d · Q(a, b, c)
4
0
h (OC (m) −
x−γq
X
m=0
m=0
x−n−c
X
x−n
X
m=0
m=0
=
h0 (OC (m − βq))
m=0
+
x
X
m=0
m=0
= x→∞
lim
h0 (OC (m − αq)) −
m=0
x
X
m=0
x
X
x
X
h0 (OC (m)) −
h0 (OC (m))
h0 (OC (m)) + D
m=0
+ D.
An important application of Lemma 2.2 is the case, where SyzProj(R) (X, Y, Z) admits a
Frobenius periodicity.
Definition 2.3. Let X be a scheme over an algebraically closed field k of characteristic
p > 0. Let S be a vector bundle over X. Assume there are s < t ∈ N such that the Frobenius
pull-backs Fe∗ (S) of S are pairwise non-isomorphic for 0 ≤ e ≤ t − 1 and Ft∗ (S) F s∗ (S).
We say that S admits a (s, t)-Frobenius periodicity. Finally, the bundle S admits a Frobenius periodicity if there are s < t ∈ N such that S admits a (s, t)-Frobenius periodicity.
Remark 2.4.
(i) By a result of Lange and Stuhler the vector bundles S admitting
a (0, t)-Frobenius periodicity are exactly those which are étale trivializable (cf.
[LS77, Satz 1.4]), where S is étale trivializable if there exists a finite étale,
surjective morphism ϕ : D → C such that ϕ∗ (S) is trivial.
(ii) If X is a projective scheme only bundles of degree zero might admit a Frobenius
periodicity in the strong sence above. For bundles of non-zero degree we will
use the following weaker definition of Frobenius periodicity. We say that S
admits a (s, t)-Frobenius periodicity if
0
Fe∗ (S) Fe ∗ (S)(m)
and
Ft∗ (S) F s∗ (S)(n)
52
2. EXAMPLES OF HILBERT-KUNZ FUNCTIONS AND HAN’S δ FUNCTION
hold for all 0 ≤ e0 < e ≤ t − 1, all m ∈ Z and some n ∈ Z.
Example 2.5. Let d ≥ 2 be an integer and let k be a field of characteristic p with p ≡
−1 (2d). Let
Ä
Ä
ää
C := Proj k[X, Y, Z]/ X d + Y d − Z d
be the Fermat curve of degree d and S := SyzC (X, Y, Z). In this case Brenner and Kaid
showed that S is always strongly semistable and admits a (0, 1)-Frobenius periodicity (cf.
[BK13, Lemma 3.2, Theorem 3.4]). To be more precise, one has
å
Ç
3p − 3
∗
.
F (S) S −
2
From this periodicity one can compute the Hilbert-Kunz function of R as (cf. [BK13,
Corollary 4.1] or use Lemma 2.2 with α = β = γ = a = b = c = D = 1)
3d 2
3d
HK(R, q) =
·q +1− .
4
4
In Chapter 7 we recover
the
periodicity
by
a
different
approach
as in [BK13]. The isomorÄ 3p−3 ä
∗
phisms F (S) S − 2 as well as the trivializing étale covers are computed explicitly
in [Stä11].
Example 2.6. We are interested in the Hilbert-Kunz functions of the diagonal hypersurfaces of the form
R := k[X, Y, Z]/(X 2 + Y l + Z m ),
where k is an algebraically closed field of characteristic p > 0. With deg(X) = lm, deg(Y) =
2m and deg(Z) = 2l the ring R is homogeneous. Moreover, we assume l = 3 and m ≥ 6 or
4 ≤ l ≤ m. Note that l and m are chosen such that
1 1 1
(2.5)
+ + ≤ 1.
2 l m
Ä
ä
The above inequality is equivalent to the condition that the triple 12 , 1l , m1 does not satisfy
the strict triangle inequality. Since the map
Ä
R → S := k[U, V, W]/ U 2lm + V 2lm + W 2lm
ä
with X 7→ U lm , Y 7→ V 2m , Z 7→ W 2l is flat and the extension
Ä
ä
k[U, V, W]/ U lm , V 2m , W 2l : k
has degree 4l2 m2 (cf. [Mat05, Theorem 22.2]), we obtain
ÄÄ
ä
HK U lm , V 2m , W 2l , S , q
HK(R, q) =
4l2 m2
by Remark 1.22. Using the Lemmata 2.1 and 2.2, we obtain
ä
2lm · Q(lm, 2m, 2l) · q2 + 2lm · δ2
4 · 4l2 m2
if SyzS (U lmq , V 2mq , W 2lq ) S (−n) ⊕ S (−n − δ) for some n, δ ∈ Z and
(2.7)
Ä
ä
2lm · Q(lm, 2m, 2l) · q2 − 2lm · Q(lma, 2mb, 2lc) dimk S /(U a , V b , Z c )
HK(R, q) =
+
4 · 4l2 m2
4l2 m2
(2.6)
HK(R, q) =
2.1. EXAMPLES OF HILBERT-KUNZ FUNCTIONS
53
if SyzS (U lmq , V 2mq , W 2lq ) SyzS (U a , V b , W c )(−n) for some n ∈ Z and some a, b, c ∈ N
such that at least one of the inequalities a < lmq, b < 2mq, c < 2lq holds.
At first we present a geometric computation which will show that SyzS (U lmq , V 2mq , W 2lq )
gets free if q is large and the inequality (2.5) is strict. Afterwards, we turn to the algebraic
situation to compute the values of the Hilbert-Kunz functions that are not covered by the
geometric computation. The computation in the algebraic situation will recover the results
from the geometric one.
Denote by C the projective spectrum of S and assume that the inequality (2.5) is strict.
This means that (0, W 2l , −V 2m ) is a syzygy for U lm , V 2m , W 2l of minimal possible degree
2(l+m).ÄTherefore, theävector bundle OC (−2(l+m)) is the maximal destabilizing subsheaf
of SyzC U lm , V 2m , W 2l and we obtain the short exact sequence
Ä
ä
0 → OC (−2(l + m)) → SyzC U lm , V 2m , W 2l → OC (−lm) → 0.
The e-th Frobenius pull-back of this sequence splits if
Ext1 (OC (−lmpe ), OC (−2(l + m)pe )) H1 (C, OC (−(2l + 2m − lm)pe ))
H0 (C, OC ((2l + 2m − lm)pe + 2lm − 3))
vanishes, where the last isomorphism is due to Serre duality with ωC OC (2lm−3). Since
2m + 2l − lm < 0, the degree shift (2l + 2m − lm)pe + 2lm − 3 gets negative for large e such
that OC ((2l + 2m − lm)pe + 2lm − 3) has no global sections. This shows (with q = pe )
SyzC (U lmq , V 2mq , W 2lq ) OC (−2(l + m)q) ⊕ OC (−lmq)
for large e. We obtain the splitting
SyzS (U lmq , V 2mq , W 2lq ) S (−2(l + m)q) ⊕ S (−lmq)
of S -modules, showing
ÄÄ
ä
ä
2lm · Q(lm, 2m, 2l) · q2 + 2lm · δ2
= 8l2 m2 q2 and
4
HK(R, q) = 2q2
HK U lm , V 2m , W 2l , S , q =
2lm−3
for all q bigger than ml−2m−2l
.
We now show by an algebraic computation that we always have a splitting
SyzS (U lmq , V 2mq , W 2lq ) S (−2(l + m)q) ⊕ S (−lmq)
if 2l + 2m < lm or 2l + 2m = lm and p . 1 (l). In the case p = 2, we have (for q ≥ 2)
q
U lmq = (V 2lm + W 2lm ) 2 = V lmq + W lmq .
From this equality, we deduce that SyzS (U lmq , V 2mq , W 2lq ) is free. The generators are
Ä
ä
−1, V (l−2)mq , W (m−2)lq and (0, W 2lq , −V 2mq )
of total degrees mlq and 2mq + 2lq.
If q is divisible by l, we have
q
V 2mq = (−U 2ml − W 2ml ) l = −U 2mq − W 2mq .
54
2. EXAMPLES OF HILBERT-KUNZ FUNCTIONS AND HAN’S δ FUNCTION
From this equality we obtain with Lemma 1.61
SyzS (U lmq , V 2mq , W 2lq ) SyzS (U lmq , −U 2mq − W 2mq , W 2lq )
SyzS (U lmq , U 2mq , W 2lq )
SyzS (U (l−2)mq , 1, W 2lq )(−2mq)
(S (−(l − 2)mq) ⊕ S (−2lq))(−2mq)
S (−lmq) ⊕ S (−2mq − 2lq).
Now we can assume that p is odd and that we can write q = ln + r with 1 ≤ r ≤ l − 1. Using
U 2lm = −V 2lm − W 2lm , we get
q−1
(2.8)
U
lmq
= −U
lm
2
X
Ä
ä
q−1 !
q−1
q−1
2lmi 2lm 2 −i
2
2
.
(−1) V W
i
i=0
The summands for i ≥
†q£
l
= n + 1 are divisible by V 2mq and the summands with
ú
(m − 2)q − m
i≤
2m
ü
are divisible by W 2lq . We take a closer look at the bound for divisibility by W 2lq .
(2.9)
(m − 2)q − m (m − 2)(ln + r) − m
=
2m
2m
(ml − 2m − 2l)n + m(r − 1) − 2r
= n+
.
2m
If the numerator in the fraction in the last line is non-negative, all summands for indices
i ≤ n in Equation (2.8) are divisible by W 2lq . Since all summands with index i ≥ n + 1 are
multiples of V 2mq , we obtain U lmq ∈ (V 2mq , W 2lq ). As in the case p = 2, this implies
SyzS (U lmq , V 2mq , W 2lq ) S (−lmq) ⊕ S (−2mq − 2lq).
The number ml − 2m − 2l vanishes if and only if (l, m) ∈ {(3, 6), (4, 4)}, otherwise it is
positive. The cases (l, m) = (3, 6) and (l, m) = (4, 4) will be discussed seperatedly, hence we
assume ml − 2m − 2l ≥ 1. Starting with the case r = 1, we have to show (ml − 2m − 2l)n ≥ 2.
From r = 1, we obtain n ≥ 1. If l is even, ml − 2l − 2m is even and the result follows. If l is
odd, we get that n is even, since q is odd. Now let r > 1. To prove that the numerator in
(2.9) is non-negative, it suffices to show m(r − 1) − 2r ≥ 0. This is true, because otherwise
we obtain the contradiction
m(r − 1) < 2r ⇐⇒ r < 1 +
2
≤ 2,
m−2
since m ≥ 4 in all cases. This argument is still valid in the cases (l, m) ∈ {(3, 6), (4, 4)}.
Up to now, we found (except for the cases q odd with q ≡ 1 (l) and (l, m) ∈ {(3, 6), (4, 4)})
SyzS (U lmq , V 2mq , W 2lq ) S (−lmq) ⊕ S (−2mq − 2lq).
2.2. HAN’S δ FUNCTION
55
The Hilbert-Kunz function of R is given by Equation (2.6) as
2lm · Q(lm, 2l, 2m) · q2 + 2lm · (lm − 2m − 2l)2 · q2
16l2 m2
= 2q2
HK(R, q) =
We only have to treat the cases q = nl + 1 for (l, m) ∈ {(3, 6), (4, 4)} with q odd. If p . 1 (l)
we obtain a splitting
SyzS (U lmp , V 2mp , W 2lp ) S (−lmp) ⊕ S (−2mp − 2lp),
which carries over to all higher exponents of p. Hence, we may assume q = p with
p = nl + 1. We start with the case (l, m) = (3, 6) and work with the grading deg(X) = 18,
deg(Y) = 12 and deg(Z) = 6. With p = 3n + 1 only the summand for i = n in Equation
(2.8) does not belong to the ideal (V 12p , W 6p ). The binomial coefficient of this summand
is non-zero since all its factors are < p. This shows
Ä
SyzS (U 18p , V 12p , Z 6p ) SyzS U 18 V 12(p−1) W 6(p−1) , V 12p , W 6p
ä
SyzS (U 18 , V 12 , W 6 )(−18(p − 1)).
By Equation (2.7), we obtain
36 · Q(18, 12, 6) · q2 − 36 · Q(18, 12, 6) + 4 · 18 · 12 · 6
4 · 4 · 9 · 36
2
= 2q − 1.
HK(R, q) =
In the case (l, m) = (4, 4) we work with the grading deg(X) = 16, deg(Y) = deg(Z) = 8. Let
q = 4n + 1. Arguing as in the case (l, m) = (3, 6) we may assume q = p and obtain
SyzS (U 16p , V 8p , W 8p ) SyzS (U 16 , V 8 , W 8 )(−16(p − 1)).
Using Equation (2.7) in this situation, we find again HK(R, q) = 2q2 − 1.
All in all we have
1
HK(R, p ) =
e
2q2 − 1
2q2
if e = 0,
if e > 0, p ≡ 1 (l) and (l, m) ∈ {(3, 6), (4, 4)},
else.
Remark 2.7. The previous example covers all two-dimensional diagonal hypersurfaces
of multiplicity two which are not of type An , E6 or E8 with one exceptional family.
Namely the cases R := k[X, Y, Z]/(X 2 + Y 2 + Z l ) for l ≥ 2 and char(k) = 2 are not treated. In
this case an easy calculation shows HK(R, 2e ) = 2 · 4e .
2.2. Han’s δ function
In this section we introduce "Han’s δ function", which measures syzygy gaps in a polynomial ring in two variables. It was first studied by Han in her thesis [Han91]. We also
give a result of Monsky, that allows us to compute Hilbert-Kunz multiplicities of certain
curves in P2 with the help of this function. All statements are reproved in a slightly more
general context to make them work in non-standard graded situations.
56
2. EXAMPLES OF HILBERT-KUNZ FUNCTIONS AND HAN’S δ FUNCTION
Let k denote an algebraically closed field of char k = p > 0 and let R = k[U, V], where U
and V have the degrees a resp. b ∈ N≥1 . All polynomials are assumed to be (quasi-)homogeneous.
Definition 2.8. Let F1 , F2 , F3 ∈ R be homogeneous and coprime. By Hilberts Syzygy
Theorem we get a splitting SyzR (F1 , F2 , F3 ) R(−α) ⊕ R(−β) for some α ≥ β ∈ N. Let
syzgap(F1 , F2 , F3 ) := α − β.
Remark 2.9. Let I be the ideal generated by the Fi and di := deg(Fi ). Then by [KR05,
Proposition 5.8.9] the Hilbert-series of R/I is
HR/I (t) =
∞
X
dimk ((R/I)i ) · ti =
i=0
1 − td1 − td2 − td3 + tα + tβ
.
(1 − ta )(1 − tb )
Note that this is a polynomial, since we have dimk (R/I) < ∞ by the coprimeness of the
Fi , hence dimk ((R/I)i ) = 0 for i 0.
Proposition 2.10. With the previous notations we get the following formulas.
(i) α + β = d1 + d2 + d3 .
(ii) Let Q(r, s, t) := 2rs + 2rt + 2st − r2 − s2 − t2 . Then
Q(d1 , d2 , d3 ) + syzgap(F1 , F2 , F3 )2
.
4ab
Proof. Write P(t) for the numerator of H(R/I)(t). Differentiating both sides of P(t) =
HR/I (t)(1 − ta )(1 − tb ) one gets
dimk (R/I) =
− d1 td1 −1 − d2 td2 −1 − d3 td3 −1 + αtα−1 + βtβ−1
=(HR/I (t))0 (1 − ta )(1 − tb ) − a · HR/I (t)(1 − ta−1 )(1 − tb ) − b · HR/I (t)(1 − ta )(1 − tb−1 ).
The evaluation at 1 gives (i). Similarly, differentiating P(t) = HR/I (t)(1 − ta )(1 − tb ) twice
and evaluating at 1 gives (ii).
Note that (i) can be seen alternatively by the additivity of the degree of vector bundles.
The previous proposition gives
syzgap(F1 , F2 , F3 ) = α − β = d1 + d2 + d3 − 2β,
where β is the minimal degree of a non-trivial syzygy for I. This allows us to drop the
coprimeness condition in the definition of syzgap.
Definition 2.11. Let m(F1 , F2 , F3 ) be the minimal degree of a non-trivial syzygy for the
ideal generated by F1 , F2 and F3 . Then define
syzgap(F1 , F2 , F3 ) := d1 + d2 + d3 − 2 · m(F1 , F2 , F3 ).
Some basic properties of syzgap are given in the next proposition.
Proposition 2.12. With the previous notations and F ∈ R non-zero and homogeneous the
following (in-)equalities hold.
(i) syzgap(FF1 , FF2 , FF3 ) = deg(F) + syzgap(F1 , F2 , F3 ).
2.2. HAN’S δ FUNCTION
57
(ii) If F and F3 are coprime, we have syzgap(FF1 , FF2 , F3 ) = syzgap(F1 , F2 , F3 ).
(iii) | syzgap(FF1 , F2 , F3 ) − syzgap(F1 , F2 , F3 )| ≤ deg(F).
p
p
p
(iv) syzgap(F1 , F2 , F3 ) = p · syzgap(F1 , F2 , F3 ).
Proof. See Propositions 2.12, 2.17, 2.19 resp. 2.16 of [Tei02].
We now specify the situation:
Setup 2.13. Let k be an algebraically closed field of characteristic p > 0, let R := k[X, Y]
equipped with the standard grading. For a1 , a2 , a3 , e ∈ N and q := pe set
Ç
å
a1 a2 a3
δ
, ,
:= q−1 · syzgap(X a1 , Y a2 , (X + Y)a3 ).
q q q
Then Proposition 2.12 (iv) ensures, that this definition is independent of q and part (iii)
gives the inequality
Ç
å
Ç
å
a1 a2 a3
b1 b2 b3
δ
, ,
, ,
−δ
≤ q−1 · k(a1 − b1 , a2 − b2 , a3 − b3 )k1 ,
q q q
q q q
where k_k1 denotes the taxicap distance. Therefore
δ is Lipschitz and extends to a unique
t
3
−1
continuous function δ on [0, ∞) with δ q = q · δ(t) and |δ(t) − δ(s)| ≤ kt − sk1 .
Our next goal is to explain how δ can be computed in this situation. We follow [Mon06b],
which provides a simpler proof than [Han91].
Definition 2.14. Let
¶
Lodd := u ∈ N3 |u1 + u2 + u3 is odd
©
be the odd lattice.
We are interested in the taxicap distance of elements of the form pts to Lodd , where s is an
integer and t a three-tuple of non-negative real numbers. Note that for given pts there is at
most one u ∈ Lodd satisfying
t
− u < 1.
ps
1
Moreover, the only candidates for ui are the roundups and rounddowns of ptis .
Definition 2.15. For s ∈ Z and t ∈ [0, ∞)3 we define
0
E s (t) :=
1−
if there is no u ∈ Lodd with
t
−u
ps
if
1
t
−u
ps
t
−u
ps
< 1,
1
< 1.
1
Monsky has proved in [Mon06b, Lemma 21, Theorem 22] the following properties of the
functions E s (t).
Theorem 2.16 (Monsky). Let t = (t1 , t2 , t3 ) ∈ [0, ∞)3 .
(i) If the ti do not satisfy the strict triangle inequality, say t1 ≥ t2 + t3 , then
max s∈Z p s · E s (t) = t1 − t2 − t3 .
2. EXAMPLES OF HILBERT-KUNZ FUNCTIONS AND HAN’S δ FUNCTION
58
(ii) If the ti satisfy the strict triangle inequality, then either E s (t) = 0 for all s or
there exists a maximal s with E s (t) , 0. Moreover, for this maximal s we have
p s · E s (t) = maxr∈Z pr · Er (t) .
(iii)
δ(t) = max s∈Z p s · E s (t) .
From this theorem one gets Han’s Theorem [Han91, Theorem 2.25 and Theorem 2.29].
Theorem 2.17 (Han). Let t = (t1 , t2 , t3 ) ∈ [0, ∞)3 . If the ti do not satisfy the strict triangle
inequality (w.l.o.g. t1 ≥ t2 + t3 ), we have δ(t) = t1 − t2 − t3 . If the ti satisfy the strict triangle
inequality and there are s ∈ Z and u ∈ Lodd with kp s t − uk1 < 1, then there is such a pair
(s, u) with minimal s and with this pair (s, u) we get
δ(t) =
1
· 1 − pst − u 1 .
s
p
Otherwise, we have δ(t) = 0.
Proof. If the ti do not satisfy the strict triangle inequality, the statement is immediate
from Theorem 2.16 (i) and (iii). If the ti satisfy the strict triangle inequality, then by
Theorem 2.16 (ii) there is a minimal s such that E−s (t) , 0 or all E s (t) vanish. Part (iii)
gives the result.
Remark 2.18. At least if t is a triple of rational numbers, there is a finite range in which
the s in Theorem 2.17 has to lie. Namely, if the ti do not satisfy the strict triangle inequality, one has pl El (t) = 0 for all l ≥ 32 · max(ti ) by [Mon06b]. This gives a lower bound for
s. With ti := zi /ni , zi , ni ∈ N, the distance of t · pu for u ≥ 0 depends only on the residue
classes of zi · pu modulo 2 · ni . With n := 2 · lcm(ti ), the series
((z1 · pu , z2 · pu , z3 · pu ))u∈N ,
where x := x mod n, is eventually periodic. The number of different values for the elements of this series gives an upper bound for s.
Note that there is an explicit implementation of Han’s Theorem (for rational arguments)
in Appendix A.
Example 2.19. Let b ≤ a ∈ N≥2 with 2(a + b) ≥ ab and both prime to the odd prime p. For
e ∈ N we set q := pe for short. We want to show
Ç
å
1 1 1
δ , ,
= 0.
2 a b
The first step is to show that the s in Han’s Theorem 2.17 has to be positive. The nearest
1 1 1
elements in Lodd to ( 2q
, aq , bq ) are (1, 0, 0), (0, 1, 0), (0, 0, 1) and (1, 1, 1). The first three
2.2. HAN’S δ FUNCTION
59
points give a taxicap distance of at least 1. For the point (1, 1, 1) we get
1
1
1
1− +1− +1−
<1
2q
aq
bq
1
1
1
>2
⇐⇒ + +
2q aq bq
2 2
⇐⇒1 + +
> 4q.
a b
Since a, b ≥ 2, this gives the contradiction 4q < 3. Therefore s > 0 in Theorem 2.17.
Now consider the points ( q2 , qa , qb ) for q ≥ p. Since the first coordinate yields a summand
p−1
p+1
1
2 in the taxicap distance to Lodd and we may choose 2 or 2 , we get that the taxicap
distance of ( q2 , qa , qb ) to Lodd is exactly 12 plus the taxicap distance of ( qa , qb ) to N2 . The four
points in N2 that we have to consider have the roundup or rounddown of qa in the first and
of qb in the second component. Hence, the taxicap distance to N2 is of the form na + mb with
1 ≤ n < a and 1 ≤ m < b, since a and b are prime to p. If this taxicap distance is smaller
than 21 , we get the contradiction
ab
≤ a + b.
2
All in all, the taxicap distance of p s ( 12 , 1a , 1b ) to Lodd is greater than 1 for all s ∈ Z. Therefore
δ is zero.
nb + ma <
Before we connect the δ function with Hilbert-Kunz multiplicities, we prove a computation rule for a non-standard graded analogue.
Setup 2.20. Let k be an algebraically closed field of characteristic p > 0, let R := k[U, V]
equipped with the positive grading deg(U) = a, deg(V) = b. Fix c, d ∈ N≥1 such that
ac = bd. For a1 , a2 , a3 , c, d, e ∈ N with q := pe set
Ç
å
a1 a2 a3
τ
, ,
:= q−1 · syzgap(U a1 , V a2 , (U c + V d )a3 ).
q q q
Again Proposition 2.12 (iv) ensures, that this definition is independent of q and part (iii)
gives the inequality
Ç
å
Ç
a1 a2 a3
b1 b2 b3
τ
, ,
−τ
, ,
q q q
q q q
å
≤ q−1 · k(a1 − b1 , a2 − b2 , a3 − b3 )k1 .
Therefore τ is Lipschitz and extends to a unique continuous function τ on [0, ∞)3 with
τ( qt ) = q−1 · τ(t) and |τ(t) − τ(s)| ≤ kt − sk1 .
Lemma 2.21. With the notations from Setup 2.20 and t ∈ [0, ∞)3 we have
Å
ã
t1 t2
τ(t) = acδ
, , t3 .
c d
Proof. For t ∈ N3 both sides multiply by p when t is replaced by pt. Since rationals
of the form nc
, nd are dense in [0, ∞), it suffices to prove the statement for t = (ct1 , dt2 , t3 )
pj pj
with ti ∈ N. Let S be the standard-graded polynomial ring k[X, Y]. Consider the map
60
2. EXAMPLES OF HILBERT-KUNZ FUNCTIONS AND HAN’S δ FUNCTION
ϕ : S → R, X 7→ U c , Y 7→ V d , which is homogeneous of degree ac. Denote by ψ the
induced map on the projective spectra. We obtain
ψ∗ (SyzProj(S ) (X t1 , Y t2 , (X + Y)t3 ))
/
SyzProj(R) (U ct1 , V dt2 , (U c + V d )t3 )
/
ψ∗ (OProj(S ) (l) ⊕ OProj(S ) (l + δ))
OProj(R) (acl) ⊕ OProj(S ) (acl + acδ)
for some l ∈ Z and δ = δ(t1 , t2 , t3 ). Now the claim follows, since we have the isomorphism
SyzProj(R) (U ct1 , V dt2 , (U c + V d )t3 ) OProj(R) (m) ⊕ OProj(R) (m + τ)
for some m ∈ Z and τ = τ(ct1 , dt2 , t3 ).
We now prove a slightly more general version of [Mon06a, Theorem 2.3] (see Corollary
2.30 below). We will skip the proofs that are still valid in our situation.
Setup 2.22. Let R be a standard-graded domain of dimension d and J := (x1 , . . . , xn ) an
R+ -primary ideal with all xi non-zero and homogeneous. For l ∈ Nn let J(l) := x1l1 , . . . , xnln .
We denote eHK (J(l), R) by µ(l).
We will show that µ extends to a unique continuous function on [0, ∞)n with the property
µ(pt) = pd · µ(t) (this is Remark 5 in the appendix of [Tei12] and we will lean on the proof
given there in the case li = l1 for all i).
Lemma 2.23. In the situation of Setup 2.22 let l ∈ Nn . If li ≥ 1, let l(i) := l − ei , where ei
denotes the i-th canonical vector. Then
Ä
ä
Ä
µ(l) − µ l(i) = O (max(l j ))d−1
Ä
ä
ä
Proof. The R-module J l(i) /J(l) is annihilated by xi , which gives it a natural S :=
R/xi -module structure. Therefore we get the inequality
Ä Ä
ä
ä
Ä Ä
ä
ä
Ä
ä
λR J l(i) /J(l) = λS J l(i) S /J(l)S ≤ λS (S /J(l)S ) ≤ λS S /mn·max(l j ) .
But the last length is for large max(l j ) given by a polynomial in max(l j ) of degree d − 1.
By Theorem 1.20 we have
Ä Ä
µ(J(l)) − µ J l(i)
ää
Ä Ä
ä
ä
Ä
ä
≤ eHK (R) · λR J l(i) /J(l) = O (max(l j ))d−1 .
Lemma 2.24. The notations are the same as in Setup 2.22. Let I ( [0, 1]n be the set of all
elements of the form ( aq11 , . . . , qann ), where the qi ≥ 1 are powers of p and the ai are natural
numbers ≤ qi . Then µ : I → R is Lipschitz.
Proof. Given two elements from I, we may expand each component by a power of
p and end up with two elements of the form ql , mq , where l and m are n-tuples of natural
2.2. HAN’S δ FUNCTION
61
numbers ≤ q.
|µ(l) − µ(m)| = |µ(l) − µ(l1 , . . . , ln−1 , mn ) + µ(l1 , . . . , ln−1 , mn ) − µ(l1 , . . . , ln−2 , mn−1 , mn )
+ . . . + µ(l1 , m2 , . . . , mn ) − µ(m)|
≤ |µ(l) − µ(l1 , . . . , ln−1 , mn )| + |µ(l1 , . . . , ln−1 , mn ) − µ(l1 , . . . , ln−2 , mn−1 , mn )|
+ . . . + |µ(l1 , m2 , . . . , mn ) − µ(m)|
=
n
X
|µ(l1 , . . . , li , mi+1 , . . . , mn ) − µ(l1 , . . . , li−1 , mi , . . . , mn )|.
i=1
Expanding the i-th summand in the same way (w.l.o.g. li ≥ mi ), the previous lemma gives
the inequality
|µ(l1 , . . . , li−1 , li , mi+1 , . . . , mn ) − µ(l1 , . . . , li−1 , mi , mi+1 , . . . , mn )|
≤ |µ(l1 , . . . , li−1 , li , mi+1 , . . . , mn ) − µ(l1 , . . . , li−1 , li − 1, mi+1 , . . . , mn )|
+ . . . + |µ(l1 , . . . , li−1 , mi + 1, mi+1 , . . . , mn ) − µ(l1 , . . . , li−1 , mi , mi+1 , . . . , mn )|
≤
|ln+1−i −m
Xn+1−i |−1
Mi · max(l1 , . . . , li−1 , li − j, mi+1 , . . . , mn )d−1
j=0
≤ Mi · max(l1 , . . . , li , mi+1 , . . . , mn )d−1 · |li − mi |,
where Mi is the constant from the previous lemma. Summation over i yields
|µ(l) − µ(m)| ≤ n · max{Mi } · (max{l1 , m1 , . . . , ln , mn })d−1 kl − mk1 .
Dividing by qd gives
Ç å
Ç
l
m
µ
−µ
q
q
å
≤
l m
n · max{Mi }
kl − mk1 = n · max{Mi } −
q
q q
hence µ is Lipschitz on I with Lipschitz constant n · max{Mi }.
,
1
Lemma 2.25. The notations are the same as in Setup 2.22. For t ∈ [0, ∞)n choose e ∈ N
such that pe ≤ max(ti ) < pe+1 =: q. This gives qt ∈ [0, 1]n and we define µ(t) := qd · µ( qt ).
Then µ(t) is continous.
Proof. By the previous lemma we only have to take care of the "borders" pe . We have
Ç
å
1
1
,...,
=
lim e µ(t) = q · µ
max ti ↓p
p
p
d
Ç åd
q
p
· eHK (R).
For the "limit from below", we have to look for the limit t → (pe , . . . , pe ), t1 , . . . , tn < pe .
This limit is
Ç åd
Ç åd
q
q
lim
µ(t)
=
·
µ(1,
.
.
.
,
1)
=
· eHK (R).
t→(pe ,...,pe )
p
p
e
t1 ,...,tn <p
Next, we want to use the function µ to compute the Hilbert-Kunz multiplicities of certain hypersurfaces k[X, Y, Z]/(F) with respect to the ideals (X a , Y b , Z c ). This is a slight
62
2. EXAMPLES OF HILBERT-KUNZ FUNCTIONS AND HAN’S δ FUNCTION
generalization of Monsky’s result from [Mon06a], where he restricted himself to the case
a = b = c.
Definition 2.26. Let k be an algebraically closed field of characteristic p > 0. An irreducible, homogeneous (standard-graded) polynomial F ∈ k[X, Y, Z] of degree d is called
regular if each of the points (1 : 0 : 0), (0 : 1 : 0), (0 : 0 : 1) has multiplicity < d2 on the
projective curve defined by F. Otherwise we call F irregular.
Monsky has shown in [Mon06a, Lemma 2.2] that regular, irreducible trinomials
F = X a1 Y a2 Z a3 + X b1 Y b2 Z b3 + X c1 Y c2 Z c3
appear in two families:
a1 a2 a3 b1 b2 b3 c1 c2 c3
0 0 > d2
type I > d2
0 > d2
d
type II d
0 0
>2
0
> d2
An empty slot means, that there is no further restriction at the exponent (except making
F homogeneous).
The determinant of the matrix
Ö
è
a1 b1 c1
a2 b2 c2
(2.10)
a3 b3 c3
is dλ with 0 < λ ∈ N, because the rows sum up to (d, d, d).
Construction 2.27. Given a regular, irreducible trinomial F, compute λ as above. Let
A := k[r, s, t]/(rλ + sλ + tλ ). We denote by R, S , T the elements rλ , sλ , tλ of A. Let x, y, z
be the elements rb3 sb2 , sc1 tc3 , ra1 ta2 if F is of type I and rc3 −b3 sc2 tb3 , sd , rb1 td−b1 if F is of
type II.
Now we show where the ring k[X, Y, Z]/(F) appears in these data. Let B := k[x, y, z] ⊆ A
and consider the surjective map ϕ : k[X, Y, Z] → B, X 7→ x, Y 7→ y, Z 7→ z. Assume that
F is of type I. Replacing the third row in (2.10) by the sum of all three rows, one finds
λ = a1 b2 + a2 c1 − b2 c1 . Similarly, one obtains the equalities λ = b3 c1 + a1 c3 − a1 b3 and
λ = b2 c3 + a2 b3 − a2 c3 . Then
xa1 ya2 + yb2 zb3 + xc1 zc3 = (rb3 sb2 )a1 (sc1 tc3 )a2 + (sc1 tc3 )b2 (ra1 ta2 )b3 + (rb3 sb2 )c1 (ra1 ta2 )c3
= ra1 b3 sa1 b2 +a2 c1 ta2 c3 + ra1 b3 sb2 c1 tb2 c3 +a2 b3 + rb3 c1 +a1 c3 sb2 c1 ta2 c3
= ra1 b3 sb2 c1 ta2 c3 (sa1 b2 +a2 c1 −b2 c1 + tb2 c3 +a2 b3 −a2 c3 + rb3 c1 +a1 c3 −a1 b3 )
= ra1 b3 sb2 c1 ta2 c3 (rλ + sλ + tλ )
shows that ϕ factors through k[X, Y, Z]/(F). Since k[X, Y, Z]/(F) and B are two-dimensional
domains they have to be isomorphic. A similar computation shows that B and k[X, Y, Z]/(F)
are isomorphic if F is of type II.
We also define the subrings B0 := k[R, S , T ] and C := k[xλ , yλ , zλ ] of A. Note that B0 is
isomorphic to a polynomial ring in two variables.
Monsky has shown that B0 and B are finite over C of ranks d resp. λ2 .
2.2. HAN’S δ FUNCTION
63
Definition 2.28. Let F = X a1 Y a2 Z a3 + X b1 Y b2 Z b3 + X c1 Y c2 Z c3 be regular and irreducible.
For t = (t1 , t2 , t3 ) ∈ [0, ∞)3 set
α0 (t) := t1 d − t1 b3 − t1 c2 − t2 c1 − t3 b1 ,
β0 (t) := t2 d − t2 c1 − t3 a2 − t1 c2 − t2 a3 ,
γ0 (t) := t3 d − t3 a2 − t1 b3 − t3 b1 − t2 a3 .
Let α+ (t) := α0 (t) if α0 (t) is ≥ 0 and 0 otherwise. Similarly, α− (t) := −α0 (t) if α0 (t) < 0 and
0 otherwise. Do the same for β0 (t) and γ0 (t). After all these preparations set
type I
type II
α(t)
β(t)
γ(t)
α(t)
β(t)
γ(t)
α+ (t) + γ− (t) β+ (t) + α− (t) γ+ (t) + β− (t) α+ (t) + β− (t) + γ− (t) β+ (t) γ+ (t) + α− (t)
Theorem 2.29. For t = (t1 , t2 , t3 ) ∈ [0, ∞)3 we have
Ç Ç
λ2
α(t) β(t) γ(t)
d
,
,
δ
µ(t) = · Q(t1 , t2 , t3 ) +
4
4d
λ
λ
λ
åå2
.
Proof. Since both sides multiply by p2 , when t is replaced by pt and since rationals
of the form nλ
are dense in [0, ∞), it suffices to prove the statement for t = λ · (a, b, c) with
pi
natural numbers a, b, c. Therefore we have to show
Ç Ç
d · λ2
λ2
α(t) β(t) γ(t)
µ(λ · (a, b, c)) =
· Q(a, b, c) +
,
,
δ
4
4d
λ
λ
λ
åå2
.
Consider the ideal J := (xaλ , ybλ , zcλ ) of C. By Theorem 1.21 we have
eHK (J,C) =
eHK (JB, B)
λ2
and
eHK (J,C) =
eHK (JB0 , B0 )
.
d
This yields
λ2
λ2
· eHK (JB0 , B0 ) =
· dimk (B0 /JB0 ),
d
d
since B0 is regular (cf. [WY00, 1.4]).
If F is of type I, we have
eHK (JB, B) =
JB0 = ((Rb3 S b2 )a , (S c1 T c3 )b , (Ra1 T a2 )c ).
By Proposition 2.12 (ii) we have
syzgap((Rb3 S b2 )a , (S c1 T c3 )b , (Ra1 T a2 )c ) = syzgap(Rab3 S ab2 −bc1 , T bc3 , Rca1 T ca2 )
= syzgap(S ab2 −bc1 , T bc3 , Rca1 −ab3 T ca2 )
= syzgap(S ab2 −bc1 , T bc3 −ca2 , Rca1 −ab3 )
Ç
å
α(t) β(t) γ(t)
,
,
.
=δ
λ
λ
λ
Note that we made some assumptions in this computation, for example ab2 ≥ bc1 in the
first equality. We defined α, β and γ such that all cases that might appear in the above
2. EXAMPLES OF HILBERT-KUNZ FUNCTIONS AND HAN’S δ FUNCTION
64
computation end up in δ
Ä
α(t) β(t) γ(t)
λ , λ , λ
ä
. By Proposition 2.10 (ii) we have
ä
1 Ä
· Q(ad, bd, cd) + δ(ab2 − bc1 , bc3 − ca2 , ca1 − ab3 )2
4
Ç Ç
åå !
α(t) β(t) γ(t) 2
1
2
= · d · Q(a, b, c) + δ
,
,
.
4
λ
λ
λ
dimk (B0 /JB0 ) =
Multiplying by λd gives the result.
If F is of type II we have
2
JB0 = (Rac3 −ab3 S ac2 T ab3 , S bd , Rcb1 T cd−cb1 )
and
syzgap(JB0 ) = syzgap(Rac3 −ab3 S ac2 T ab3 , S bd , Rcb1 T cd−cb1 )
= syzgap(Rac3 −ab3 T ab3 , S bd−ac2 , Rcb1 T cd−cb1 )
= syzgap(Rac3 −ab3 , S bd−ac2 , Rcb1 T cd−cb1 −ab3 )
= syzgap(Rac3 −ab3 −cb1 , S bd−ac2 , T cd−cb1 −ab3 )
Ç
å
α(t) β(t) γ(t)
=δ
,
,
.
λ
λ
λ
As in the case where F is of type I we treated only one of the possible cases. Again all
other cases end up in the same result by definition of α, β and γ.
By Proposition 2.10 (ii) we have
ä
1 Ä
· Q(ad, bd, cd) + δ(ac3 − ab3 − cb1 , bd − ac2 , cd − cb1 − ab3 )2
4
Ç Ç
åå !
1
α(t) β(t) γ(t) 2
2
= · d · Q(a, b, c) + δ
,
,
.
4
λ
λ
λ
dimk (B0 /JB0 ) =
λ2
d
Multiplying by
gives the result.
We obtain [Mon06a, Theorem 2.3] as a corollary.
Corollary 2.30.
Ç
3d
λ2
(a1 + b2 − d)s (a1 + c3 − d)s (b2 + c3 − d)s
µ(s, s, s) = s2 + δ
,
,
4
4d
λ
λ
λ
å2
.
With t = (a, b, c) ∈ N3 in Theorem 2.29 one obtains
Corollary 2.31.
eHK
ÄÄ
ä
ä
Ç Ç
d
λ2
α(a, b, c) β(a, b, c) γ(a, b, c)
X , Y , Z , k[X, Y, Z]/(F) = · Q(a, b, c)+
δ
,
,
4
4d
λ
λ
λ
a
b
c
Remark 2.32. An explicit implementation of Corollary 2.31 is given in Appendix A.
åå2
.
2.2. HAN’S δ FUNCTION
65
Corollary 2.33. Let N be the normalization of the projective curve V+ (F). Then the
pull-back of SyzV+ (F) (X a , Y b , Z c ) to N is strongly semistable if and only if
å
Ç
α(a, b, c) β(a, b, c) γ(a, b, c)
,
,
= 0.
δ
λ
λ
λ
Proof. By Theorem 1.80 the pull-back of SyzV+ (F) (X a , Y b , Z c ) to the normalization N
is strongly semistable if and only if
ÄÄ
ä
ä
d (a + b + c)2
− a2 − b2 − c2
eHK X , Y , Z , k[X, Y, Z]/(F) = ·
2
2
d
= · Q(a, b, c).
4
The claim follows by comparing this Hilbert-Kunz multiplicity with Corollary 2.31.
a
b
c
!
Remark 2.34. Let R = k[U, V, W]/( f ) be positively-graded, such that the image F of f
under the map k[U, V, W] −→ k[X, Y, Z], where U, V, W are mapped to X deg(U) , Y deg(V) ,
Z deg(W) becomes of type I or II. In this situation we can compute
eHK
ÄÄ
ä
ä
X deg(U) , Y deg(V) , Z deg(W) , k[X, Y, Z]/(F)
with Corollary 2.31. By Theorem 1.21 we have to divide this number by the degree of
the extension Q(S ) : Q(R) to get eHK (R). This degree is nothing else but the product
deg(U) · deg(V) · deg(W).
In the next section we will use this remark to discuss the behaviour of the Hilbert-Kunz
multiplicity, when one exponent in f tends to infinity and the others are constant.
Definition 2.35. With the notations from the previous remark, we call the quasi-homogeneous trinomial f regular if the homogeneous trinomial F is regular.
In [Han91, Theorem 2.30], Han computed the Hilbert-Kunz multiplicity of diagonal hypersurfaces of dimension two. In view of the last remark, we can compute them using
Theorem 2.29.
Corollary 2.36. The Hilbert-Kunz multiplicity of k[U, V, W]/(U d1 + V d2 + W d3 ) equals
Ç
1
d1 d2 d1 d3 d2 d3
1 1 1
· 2d1 + 2d2 + 2d3 −
−
−
+ d1 d2 d3 · δ
, ,
4
d3
d2
d1
d1 d2 d3
Moreover, if the
1
di
å2 !
.
do not satisfy the strict triangle inequality, this is min {d1 , d2 , d3 }.
Proof. The second statement is clear from the first. Let deg(U) = d2 d3 , deg(V) = d1 d3
and deg(W) = d1 d2 . Then we get d = d1 d2 d3 , α = d1 d22 d32 , β = d12 d2 d32 , γ = d12 d22 d3 and
λ = d12 d22 d32 . Dividing
eHK
ÄÄ
ä
Ä
X d2 d3 , Y d1 d3 , Z d1 d2 , k[X, Y, Z]/ X d + Y d + Z d
by deg(U) · deg(V) · deg(W) = d12 d22 d32 , we get the result.
ää
Definition 2.37. For any P, Q ∈ N \ {0, 1}, we say that k[U, V, W]/(U P + V Q + UVW) has
a singularity of type T PQ∞ .
66
2. EXAMPLES OF HILBERT-KUNZ FUNCTIONS AND HAN’S δ FUNCTION
Example 2.38. We compute the Hilbert-Kunz multiplicities of R := k[U, V, W]/( f ) of type
ADE and T PQ∞ . We will denote the degrees of U, V, W and f by a, b, c and d.
An : Since we need a trinomial, we use f := U n+1 + V 2 + W 2 in this computation,
hence we have to restrict to odd characteristics. The arguments of δ in Corollary
1
. It is easy to see δ = 0 if p is odd. This gives
2.36 are 12 , 12 and n+1
Ç
å
1
4
1
eHK (R) =
2n + 10 − 2(n + 1) −
= 2−
.
4
n+1
n+1
Dn : Let f = U 2 + V n−1 + VW 2 with n ≥ 4 and a = n − 1, b = 2, c = n − 2. Then
d = 2n − 2, F = X 2n−2 + Y 2n−2 + Y 2 Z 2n−4 , λ = 4(n − 1)(n − 2), α = 2(n − 1)(n − 2),
β = 2(n − 1) and γ = 2(n − 1)(n − 2). The arguments of δ in Corollary 2.31 are
1
1
1
2 , 2(n−2) , 2 , hence δ = 0 as above (provided p is odd). We get
Ä
ä
2 · ((n − 1)(n − 2) + 2(2n − 3)) − ((n − 2)2 + 22 + (n − 1)2 )
eHK (R) =
2(n − 1)(n − 2)
1
.
= 2−
4(n − 2)
2n−2
4
1
1
If p = 2, we have δ 21 , 2(n−2)
, 12 = 2·(n−2)
and eHK (R) = 2.
2
3
4
E6 : Let f = U + V + W . By Corollary 2.36 and Example 2.19 we get
Ç
1
1 1 1
3 8
eHK (R) =
18 − − + 24 · δ , ,
4
2 3
2 3 4
å2 !
= 2−
1
,
24
1
if p ≥ 5. If p ≤ 3, we have δ = 12
and eHK (R) = 2.
2
3
3
E7 : Let f = U + V + VW , a = 9, b = 6, c = 4 and d = 18. Then F = X 18 + Y 18 +
Y 6 Z 12 , α = 6 · 18, β = 3 · 18, γ = 4 · 18 and λ = 12 · 18. The arguments of δ are
1 1 1
2 , 3 , 4 . By Example 2.19 we get δ = 0 if p ≥ 5 and
eHK (R) =
18
4
Ä
ä
2 · (9 · 6 + 9 · 4 + 6 · 4) − (92 + 62 + 42 )
1
= 2− .
9·6·4
48
1
If p ≤ 3, we have δ = 12
and eHK (R) = 2.
2
3
5
E8 : Let f = U + V + W . By Corollary 2.36 and Example 2.19 for p ≥ 7 we get
Ç
1
6 10 15
1 1 1
eHK (R) =
20 − − − + 30 · δ , ,
4
5 3
2
2 3 5
å2 !
= 2−
1
.
120
1
If p ≤ 5, one gets δ = 30
and eHK (R) = 2.
P
Q
T PQ∞ : Let f = U + V + UVW, a = Q, b = P, c = PQ − P − Q and d = PQ. To get c ≥ 1,
we assume P + Q > 4. Then F = X PQ + Y PQ + X Q Y P Z PQ−P−Q , α = (Q − 2)PQ,
β = (P − 2)PQ, γ = λ = (PQ − P − Q)PQ. Because
(P − 2)(Q − 2) ≥ 0 ⇐⇒
P−2
Q−2
+
≤ 1,
PQ − P − Q PQ − P − Q
2.3. THE BEHAVIOUR OF THE HILBERT-KUNZ MULTIPLICITY IN FAMILIES
67
we see that the arguments of δ do not satisfy the strict triangle inequality and
hence
eHK (R) =
P+Q−4
3P2 Q2 − 4P2 Q − 4PQ2 + 4PQ
= 3−
PQ(PQ − P − Q)
PQ − P − Q
in all characteristics. Since 4 < P + Q and
interval [2, 3) ∩ Q.
P+Q−4
PQ−P−Q
≤ 1, this number is in the
The computations in the ADE situation recover Theorem 1.26 with one exception: To
compute the Hilbert-Kunz multiplicity of a surface ring of type An we had to restrict to
odd characteristics, which is not necessary in Theorem 1.26 if n + 1 is odd.
We always assumed that F is regular, meaning that all singularities of V+ (F) have multiplicity strictly smaller than d2 . The Hilbert-Kunz multiplicity of an irregular trinomial was
computed by Trivedi in [Tri05b, Theorem 1.3].
Theorem 2.39 (Trivedi). Let C be a singular irreducible plane curve of degree d defined
over an algebraically closed field of characteristic p > 0. Assume that C has a singular
point of multiplicity r ≥ d2 . Let R be the homogeneous coordinate ring of C. Then
eHK (R) =
3d (2r − d)2
+
.
4
4d
2.3. The behaviour of the Hilbert-Kunz multiplicity in families of surface rings
given by certain quasi-homogeneous trinomials
In this section we want to apply Theorem 2.29 to study the behaviour of the Hilbert-Kunz
multiplicity in families of surface rings defined by irreducible, regular, quasi-homogeneous
trinomials (IRQT for short). To be more precise we study the Hilbert-Kunz multiplicity
of rings R := k[U, V, W]/(F), where F is a trinomial of the type
U a1 V a2 + V b2 W b3 + U c1 W c3
or
U a1 + U b1 V b2 W b3 + V c2 W c3 ,
homogeneous of degree d with respect to some positive grading deg(U) = a, deg(V) = b,
deg(W) = c and regular in the sense of Definition 2.37. Now we are interested in the
behaviour of the Hilbert-Kunz multiplicity of R, when a1 , b2 or c3 grows.
Definition 2.40. With the notations above let
F1 (a2 , b1 , b3 , c2 , c3 ) be the family of IRQT of the form U L V a2 + V b2 W b3 + U c1 W c3 ,
FU (b1 , b2 , b3 , c2 , c3 ) be the family of IRQT of the form U L + U b1 V b2 W b3 + V c2 W c3 ,
FV (a1 , b1 , b3 , c2 , c3 ) be the family of IRQT of the form U a1 + U b1 V L W b3 + V c2 W c3 ,
FW (a1 , b1 , b2 , b3 , c2 ) be the family of IRQT of the form U a1 + U b1 V b2 W b3 + V c2 W L .
We will write F1 , FU , FV , FW for short.
In what follows, we will always assume that the arguments of δ in Corollary 2.31 do not
satisfy the strict triangle inequality for L large enough.
2. EXAMPLES OF HILBERT-KUNZ FUNCTIONS AND HAN’S δ FUNCTION
68
2.3.1. The family F1 . The trinomials
U L V a2 + V b2 W b3 + U c1 W c3
are quasi-homogeneous in the grading
a = b2 c3 − a2 c3 + a2 b3
b = (c3 − b3 )L + b3 c1
c = b2 L − b2 c1 + a2 c1
of degree
d = b2 c3 L + a2 b3 c1 .
Note that b2 > a2 and c3 > b3 give a, b > 0. Assume L > c1 to get c > 0. The regularity
conditions aL, bb2 , cc3 > d/2 translate for large L to b2 c3 > 2a2 (c3 − b3 ) and c3 > 2b3 .
We have (recall Definition 2.28)
α0 = ad − acb3 − abc1
= ab(b2 − c1 ),
β = bd − bca2 − abc1
= bc(c3 − a2 ),
γ = cd − acb3 − bca2
= ac(L − b3 ).
0
0
Since γ0 is a degree two polynomial in L with leading coefficient (b2 −a2 )b2 c3 +a2 b2 b3 > 0
and α0 depends linearly on L, we get γ0 > 0 and α0 < γ0 for large L.
α0
+
+
−
−
β0
+
−
+
−
α
β
γ
α0
β0
γ0
α0
0
γ0 − β0
0
0
0 β −α
γ0
0
−α0 γ0 − β0
δ·λ
γ0 − α0 − β0 or β0 − α0 − γ0
γ0 − α0 − β0 or β0 − α0 − γ0
γ0 + α0 − β0 or β0 − α0 − γ0
γ0 + α0 − β0 or β0 − α0 − γ0 .
Table 2.1. The table shows the possible values of δ · λ depending the signs
of α0 , β0 .
According to Table 2.1 the value of δ · λ in Theorem 2.29 has the shape γ0 − β0 − α0 or
±(β0 − γ0 − α0 ). In the first case we have δ · λ = cd − bd − ad + 2abc1 , which gives the
Hilbert-Kunz multiplicity
d · (2ab + 2ac + 2bc − a2 − b2 − c2 ) (cd − bd − ad + 2abc1 )2
+
4abc
4abcd
2
d abc1 c1
= +
+ · (c − b − a)
c
cd
c
d − ac1 abc21 bc1
= c1 +
+
−
c
cd
c
b3 − c3 b3 c21 a
= c1 + c3 + c1
−
.
b2
b2 d
eHK (R) =
2.3. THE BEHAVIOUR OF THE HILBERT-KUNZ MULTIPLICITY IN FAMILIES
69
In this case L occurs only in d, hence the Hilbert-Kunz multiplicity is increasing in L and
3
.
tends to c1 + c3 + c1 b3b−c
2
Similarly, in the second case we have δ · λ = ±(bd − cd − ad + 2acb3 ) and
a
d + b3 (b − c) acb23 − adb3
+
= b2 + b3 − b2 b3 .
b
bd
d
Here the Hilbert-Kunz multiplicity is increasing in L and tends to b2 + b3 .
By the symmetry of the problem we get in the case b2 = L:
b
eHK (R) = c1 + c3 − c1 c3
in the cases λ · δ = ±(γ0 − β0 − α0 )
d
c1 − a1 a22 c1 b
−
in the case λ · δ = α0 − β0 − γ0
eHK (R) = a1 + a2 + a2
c3
c3 d
and in the case c3 = L:
c
eHK (R) = a1 + a2 − a1 a2
in the cases λ · δ = ±(α0 − β0 − γ0 )
d
a2 − b2 a2 b23 c
−
in the case λ · δ = β0 − α0 − γ0 .
eHK (R) = b2 + b3 + b3
a1
a1 d
eHK (R) =
Example 2.41. Consider the family (U L + V d2 + W d3 )L 0 of polynomials defining diagonal hypersurfaces with d2 , d3 . The above computations show Äthat their ÄHilbert-Kunz
ää
multiplicity tends to min(d2 , d3 ) for L → ∞. This is exactly eHK k[V, W]/ V d2 + W d3 .
We cannot apply the above computation in the case d2 = d3 , since then the triple L1 , d12 , d12
always satisfies the strict triangle inequality.
Next, we prove that the limits for L → ∞ above appear as the Hilbert-Kunz multiplicity
of a binomial hypersurface attached to the family F1 .
Definition 2.42. We set limL→∞ F1 := V b2 W b3 + U c1 W c3 = W b3 (V b2 + U c1 W c3 −b3 ).
Remark 2.43. By Theorem 1.8 (ii) and Remark 1.6 we have
c
b3 −c3
1 + c3 + c1 b2
eHK k[U, V, W]/(W b3 · (V b2 + U c1 W c3 −b3 ) =
b2 + b3
Ä
ä
if c1 , c3 − b3 < b2 ,
otherwise.
Lemma 2.44. We obtain the following limit behaviour of the Hilbert-Kunz multiplicity in
the family F1
Ä
ä
Å
ã
lim eHK k[U, V, W]/(U V + V W + U W ) = eHK lim F1 .
L a2
b2
b3
c1
c3
L→∞
L→∞
Proof. Since the two possible values for eHK (limL→∞ F1 ) coincide in the cases c1 = b2
and c3 − b3 = b2 , we may assume that both equations do not hold. Then we only have to
show the equivalence
δ · λ = γ0 − α0 − β0 ⇔ c1 , c3 − b3 < b2 .
By Table 2.1 the condition δ · λ = γ0 − α0 − β0 is equivalent to the conditions that α0 and
the leading coefficient of γ0 − β0 are positive. Now, the condition α0 > 0 gives b2 > c1 and
LC(γ0 − β0 ) = b2 c3 (b2 + b3 − c3 ) > 0 gives b2 > c3 − b3 .
2. EXAMPLES OF HILBERT-KUNZ FUNCTIONS AND HAN’S δ FUNCTION
70
2.3.2. The family FU . The trinomials
U L + U b1 V b2 W b3 + V c2 W c3
are quasi-homogeneous in the grading
a = b2 c3 − b3 c2
b = (c3 − b3 )L − c3 b1
c = (b2 − c2 )L + b1 c2
of degree
d = aL.
Since c3 > b3 and b2 > c2 the degrees are positive. In this case the regularity conditions
aL, bb2 , cc3 > d/2 translate to the condition
b2 c3 + b3 c2 > max(2b2 b3 , 2c2 c3 )
for large values of L. We have
α0 = ad − ac(b1 + b3 ) − abc2 ,
β0 = bd − abc2 ,
γ0 = cd − ac(b1 + b3 ).
In this case β0 and γ0 are polynomials in L of degree two with positive leading coefficients
and α0 depends linearly on L. Therefore we may assume L to be big enough such that β0 ,
γ0 ≥ 0 and α0 < β0 , γ0 .
α0
+
−
α β
γ
α0 β0
γ0
0 β0 γ0 − α0
δ·λ
γ0 − α0 − β0 or β0 − α0 − γ0
γ0 − α0 − β0 or − γ0 + α0 + β0 .
Table 2.2. The table shows the possible values of δ · λ depending the sign
of α0 .
Thus, δ · λ has the shape ±(γ0 − β0 − α0 ) or β0 − γ0 − α0 . In the first case we have δ · λ =
±(cd − bd − ad + 2abc2 ), which gives the Hilbert-Kunz multiplicity
d + c2 (c − b) abc22 − adc2
a
+
= c2 + c3 − c2 c3 .
c
cd
d
This is an increasing function in L, tending to c2 + c3 .
Similarly, in the second case we have δ · λ = bd − cd − ad + 2ac(b1 + b3 ), which gives the
Hilbert-Kunz multiplicity
eHK (R) =
d + (b1 + b3 )(b − c) ac(b1 + b3 )2 − ad(b1 + b3 )
+
b
bd
2
c2 − b2 b1 c3 c2 − b2 (b1 + b3 )2
b1 c2 ab3 + b21 c2
= b1 + b2 + b3 + b1
+
+
(b2 − c2 +
)−
.
c 3 − b3
b c3 − b3
b
L
b
eHK (R) =
2.3. THE BEHAVIOUR OF THE HILBERT-KUNZ MULTIPLICITY IN FAMILIES
71
−b2
.
This function tends to b1 + b2 + b3 + b1 cc32 −b
3
Example 2.45. Consider the family (U L + UVW + W Q )L 0 of T LQ∞ singularities. Since
α = (Q − 2)LQ, β = (Q2 − Q)L2 − Q2 L and γ = (L − 2)LQ, we are in the second case above
and see that the Hilbert-Kunz multiplicity is given by
Ç
1
1
3−
+ 3−Q−
Q−1
Q−1
å
1
.
LQ − L − Q
In particular, the Hilbert-Kunz multiplicity is constant for Q = 2 and strictly increasing in
the cases Q ≥ 3.
One question concerning the behaviour of the Hilbert-Kunz multiplicity as a function in
L is left.
Question 2.46. Assume that we are in the case δ · λ = bd − cd − ad + 2ac(b1 + b3 ). Are
there Hilbert-Kunz functions that are (strictly) decreasing as functions in L?
As in the case of the family F1 , the limit of the Hilbert-Kunz multiplicity for L → ∞
appears as Hilbert-Kunz multiplicity of an attached binomial hypersurface.
Definition 2.47. Let limL→∞ FU := U b1 V b2 W b3 + V c2 W c3 = V c2 W b3 (U b1 V b2 −c2 + W c3 −b3 ).
Remark 2.48. By Theorem 1.8 (ii) and Remark 1.6 we have
Ä
ä
eHK k[U, V, W]/(V c2 W b3 · (U b1 V b2 −c2 + W c3 −b3 ))
b
=
c2 −b2
1 + b2 + b3 + b1 c3 −b3
c2 + c3
if b1 , b2 − c2 < c3 − b3 ,
otherwise.
Lemma 2.49. The limit behaviour of the Hilbert-Kunz multiplicity in the family FU is
given by
Ä
ä
Å
ã
lim eHK k[U, V, W]/(U L + U b1 V b2 W b3 + V c2 W c3 ) = eHK lim FU .
L→∞
L→∞
Proof. Since the two possible values of eHK (limL→∞ FU ) coincide in the cases b1 =
c3 − b3 and b2 − c2 = c3 − b3 , we may assume that both equalities do not hold. According
to Table 2.2 we have
δ · λ = β0 − α0 − γ0 ⇔ α0 > 0 and max{β0 , γ0 } = β0 .
The condition α0 > 0 is equivalent to b1 < c3 − b3 and the condition max{β0 , γ0 } = β0 is
equivalent to the positivity of the leading coefficient of
β0 − γ0 = a · (c2 + c3 − b2 − b3 )L2 + O(L).
2. EXAMPLES OF HILBERT-KUNZ FUNCTIONS AND HAN’S δ FUNCTION
72
2.3.3. The family FV . The trinomials
U a1 + U b1 V L W b3 + V c2 W c3
are quasi-homogeneous in the grading
a = c3 L − b3 c2
b = a1 c3 − a1 b3 − b1 c3
c = a1 L − a1 c2 + b1 c2
of degree
d = aa1 .
For L > c2 we get positive values for a and c. The condition b > 0 is non-trivial. For large
L we get the condition
(∗)
a1 c3 > 2(a1 b3 + b1 c3 )
to get a regular polynomial. We have
α0 = ad − ac(b1 + b3 ) − abc2 ,
β0 = bd − abc2 ,
γ0 = cd − ac(b1 + b3 ).
In this case β0 depends only linearly on L and α0 , γ0 are O(L2 ).
α0
+
+
+
+
−
−
−
−
β0 γ0
+ +
+ −
− +
− −
+ +
+ −
− +
− −
α
α0
0
α − γ0
α0 − β 0
0
α − β0 − γ 0
0
−γ0
−β0
−β0 − γ0
β
γ
0
β
γ0
0
β
0
0
γ0
0
0
0
0
β γ − α0
β0 −α0
0 γ0 − α0
0
−α0
δ·λ
or γ0 − α0 − β0
0
0
0
α − β − γ or β0 − α0 + γ0
α0 − β0 − γ0 or γ0 − α0 + β0
α0 − β0 − γ0
α0 + β0 − γ0 or γ0 − α0 − β0
α0 − β0 − γ0 or γ0 − α0 − β0
α0 − β0 − γ0 or γ0 − α0 + β0
α0 − β0 − γ0 or γ0 − α0 + β0 .
α0 − β0 − γ0
Table 2.3. The table shows the possible values of δ · λ depending the signs
of α0 , β0 , γ0 .
As we can read off Table 2.3 the product δ · λ has always the shape ±(γ0 − β0 − α0 ) or
±(α0 − γ0 − β0 ). In the first case we have δ · λ = ±(cd − bd − ad + 2abc2 ). This gives the
constant Hilbert-Kunz multiplicity
d + c2 (c − a) abc22 − bdc2
c2 c3
+
= c2 + c3 −
.
c
cd
a1
Similarly, in the second case we have δ · λ = ±(ad − cd − bd), which gives the constant
Hilbert-Kunz multiplicity
d
eHK (R) = = a1 .
a
eHK (R) =
2.3. THE BEHAVIOUR OF THE HILBERT-KUNZ MULTIPLICITY IN FAMILIES
73
Definition 2.50. Let limL→∞ FV := U a1 + V c2 W c3 .
Remark 2.51. By Theorem 1.8 (ii) and Remark 1.6 we have
c + c − c2 c3
2
3
a1
eHK k[U, V, W]/(U a1 + V c2 W c3 ) =
a1
if c2 , c3 < a1 ,
otherwise.
Lemma 2.52. The limit behaviour of the Hilbert-Kunz multiplicity in the family FV is
given by
Ä
ä
ã
Å
lim eHK k[U, V, W]/(U + U V W + V W ) = eHK lim FV .
b1
a1
L
b3
c2
c3
L→∞
L→∞
Proof. We may assume c2 , c3 , a1 . Then Table 2.3 shows that δ · λ = ±(γ0 − α0 − β0 ) is
equivalent to β0 > 0 and LC(γ0 − α0 ) > 0. The condition β0 > 0 gives a1 > c2 and LC(γ0 −
α0 ) > 0 gives a1 > c3 .
2.3.4. The family FW . The trinomials
U a1 + U b1 V b2 W b3 + V c2 W L
are quasi-homogeneous in the grading
a = b2 L − b3 c2
b = (a1 − b1 )L − a1 b3
c = (b2 − c2 )a1 + b1 c2
of degree
d = aa1 .
The inequalities a1 > b1 and b2 > c2 give b, c > 0. Assume L > b3 to get a positive. The
regularity conditions give for large L the following conditions on the exponents
a1 > 2b1
and
a1 b2 > 2c2 (a1 − b1 ).
We have
α0 = ad − ac(b1 + b3 ) − abc2 ,
β0 = bd − abc2 ,
γ0 = cd − ac(b1 + b3 ).
At first we note that the equivalence
α0 < 0 ⇐⇒ c < 0
gives α0 ≥ 0. In this case γ0 depends linearly on L and α0 , β0 are polynomials in L of
degree two.
74
2. EXAMPLES OF HILBERT-KUNZ FUNCTIONS AND HAN’S δ FUNCTION
β0 γ0
+ +
+ −
− +
− −
α
β γ
0
α
β0 γ0
0
0
α −γ
β0 0
α0 − β 0
0 γ0
0
0
0
α −β −γ 0 0
δ·λ
or β0 − α0 − γ0
0
0
0
α − β − γ or β0 − α0 + γ0
α0 − β0 − γ0 or γ0 − α0 + β0
α0 − β0 − γ0 .
α0 − β0 − γ0
Table 2.4. The table shows the possible values of δ · λ depending the signs
of β0 , γ0 .
Therefore δ · λ has the shape β0 − α0 − γ0 or ±(α0 − β0 − γ0 ). In the first case we have
δ · λ = bd − cd − ad + 2ac(b1 + b3 ), giving
d + (b1 + b3 )(b − a) ac(b1 + b3 )2 − cd(b1 + b3 )
+
b
bd Ç
å
b2 b3
1
a1 b2 b3
b1 c + b3 c
= b1 + b2 + b3 −
+ (a1 − b1 − b3 )
− b3 c2 −
a1 − b1 b
a1 − b1
a1
Ç
å
b2 b3
1
b1 c2
b1 b2
= b1 + b2 + b3 −
+ (a1 − b1 − b3 )2
−
a1 − b1 b
a1
a1 − b1
eHK (R) =
b3
as Hilbert-Kunz multiplicity. It tends to b1 + b2 + b3 − ab12−b
. Treating eHK (R) as a function
1
in L, it is either constant or strictly increasing. It cannot be strictly decreasing, since
b2
would have to be negative, which is equivalent to −b1 c2 ≥
then the term ba1 1c2 − ab11−b
1
a1 · (b2 − c2 ) > 0. In the second case we have δ · λ = ±(ad − cd − bd) and get the constant
Hilbert-Kunz multiplicity
d
eHK (R) = = a1 .
a
We give two examples showing that both possibilities, eHK (R) constant or strictly inreasing in L, appear in the case δ · λ = β0 − α0 − γ0 .
Example 2.53. With F := U 5 + UV 3 + W L we have a = 3L, b = 4L, c = 15 and d = 15L.
Since α0 = 45L2 − 45L, β0 = 60L2 and γ0 = 180L, we have β0 − α0 − γ0 ≥ 0 for L ≥ 15, hence
we are in the first case of the above computation and the Hilbert-Kunz multiplicity of R is
given by
Ç
å
1 2
3
3
eHK (R) = 4 +
·4 · −
= 4− ,
4L
4
L
which is strictly increasing in L.
Example 2.54. Now consider F := U 5 + UV 3 W 4 + W L . In this case we have a = 3L,
b = 4L − 20, c = 15 and d = 15L. This gives α0 = 45L2 − 225L, β0 = 60L2 − 300L and
γ0 = 0. Therefore β0 − α0 − γ0 ≥ 0 for L ≥ 5 and we are again in the first case of the above
computation. Now, the Hilbert-Kunz multiplicity of R is given by
eHK (R) = 8 −
12
= 5.
4
Definition 2.55. Let limL→∞ FW := U a1 + U b1 V b2 W b3 = U b1 (U a1 −b1 + V b2 W b3 ).
2.3. THE BEHAVIOUR OF THE HILBERT-KUNZ MULTIPLICITY IN FAMILIES
75
Remark 2.56. By Theorem 1.8 (ii) and Remark 1.6 we have
Ä
ä
b
b2 b3
1 + b2 + b3 − a1 −b1
if b2 , b3 < a1 − b1 ,
otherwise.
eHK k[U, V, W]/(U b1 · (U a1 −b1 + V b2 W b3 )) =
a1
Lemma 2.57. The limit behaviour of the Hilbert-Kunz multiplicity in the family FW is
given by
ä
Ä
Å
ã
lim eHK k[U, V, W]/(U a1 + U b1 V b2 W b3 + V c2 W L ) = eHK lim FW .
L→∞
L→∞
Proof. We may assume b2 , b3 , a1 − b1 . According to Table 2.4 we have
δ · λ = β0 − α0 − γ0 ⇔ β0 , γ0 > 0 and max{α0 , β0 } = β0 .
The condition γ0 > 0 is equivalent to b3 < a1 − b1 and the condition max{α0 , β0 } = β0 is
equivalent to the positivity of the leading coefficient of
β0 − α0 = a1 · (b1 + b2 − a1 )L2 + O(L).
The results in this section lead to the following questions.
Question 2.58. What do the limits of the Hilbert-Kunz multiplicities tell us about the
families of singularities?
Question 2.59. Fix F ∈ k[X1 , . . . , Xn ] and a1 , . . . , an−1 ∈ N. Under which conditions does
the equality
Ä
Ä
an−1 L
Xn
lim eHK k[X1 , . . . , Xn ]/ F + X1a1 · . . . · Xn−1
ää
L→∞
= eHK (k[X1 , . . . , Xn ]/(F))
hold? Does it even hold for the Hilbert-Kunz functions?
Examples for which the limit behaviour holds for the Hilbert-Kunz functions are the diagonal hypersurfaces Rl,m := k[X, Y, Z]/(X 2 + Y l + Z m ) with 21 + 1l + m1 ≤ 1. For those we
have the equality
lim HK(Rl,m ) = HK(k[X, Y, Z]/(X 2 + Y l )).
m→∞
Further examples for which the limit behaviour from the question holds for the HilbertKunz functions are given in Sections two and three of Chapter seven.
CHAPTER 3
Matrix factorizations and representations as first syzygy modules of
ideals
In this chapter we want to introduce the theory of matrix factorizations, their connection
to maximal Cohen-Macaulay modules over hypersurfaces and explain how one gets with
their help isomorphisms of maximal Cohen-Macaulay modules with "nice" first syzygy
modules of ideals.
3.1. Properties of maximal Cohen-Macaulay modules
In this section we state some important properties of maximal Cohen-Macaulay modules.
Definition 3.1. Let R be a ring and M a finitely generated R-module. We call M a maximal Cohen-Macaulay R-module, if M = 0 or
depthRm (Mm ) = dim(Mm ) = dim(Rm )
holds for all maximal ideals m of R. We say that R is Cohen-Macaulay if R is maximal
Cohen-Macaulay as an R-module.
Remark 3.2. An equivalent definition of the maximal Cohen-Macaulay property using
Grothendieck’s vanishing theorem (cf. [BH98, Theorem 3.5.7]) is that
HimRm (Mm ) = 0
holds for all 0 ≤ i ≤ dim(Rm ) − 1 and all maximal ideals m of R.
Example 3.3. We give some examples.
(i) If dim(R) = 0, we have 0 ≤ depthRm (Mm ) ≤ dim(Rm ) = 0 for all m ∈ Spec(R),
giving that every finitely generated R-module is maximal Cohen-Macaulay.
(ii) Regular rings are Cohen-Macaulay and every maximal Cohen-Macaulay Rmodule M is free, since Mm has finite projective dimension and this is
proj. dimRm (Mm ) = depthRm (Rm ) − depthRm (Mm ) = 0
by the Auslander-Buchsbaum formula.
(iii) If (R, m) has dimension one and at least one non-zero divisor, then a finitely
generated R-module M is maximal Cohen-Macaulay if and only if it is torsionfree. That is because m and all its powers contain a non-zero divisor, giving the
inclusion H0m (M) ⊆ T(M), where T(M) denotes the torsion-submodule of M. For
the other inclusion, note that T(M) is Artinian because R is one-dimensional,
hence T(M) = H0m (T(M)) ⊆ H0m (M).
77
78
3. MATRIX FACTORIZATIONS AND REPRESENTATIONS AS FIRST SYZYGY MODULES
Definition 3.4. Let (R, m) be a Cohen-Macaulay local ring. We say that R has an isolated
singularity if Rp is regular for all p ∈ Spec(R) \ {m}.
We can characterize the property of having an isolated singularity by the structure of the
maximal Cohen-Macaulay modules.
Lemma 3.5. Let (R, m) be a Cohen-Macaulay local ring. The following are equivalent.
(i) The ring R has an isolated singularity.
(ii) All maximal Cohen-Macaulay R-modules are locally free on the punctured spectrum of R.
Proof. The implication (i) ⇒ (ii) is clear since the property maximal Cohen-Macaulay
localizes (cf. [BH98, Theorem 2.1.2]). See [Yos90, Lemma 3.3] for the converse.
Maximal Cohen-Macaulay modules appear in a natural way as kernels in free resolutions
as the following proposition shows.
Proposition 3.6. Let (R, m) be a Cohen-Macaulay ring and
Fn−1 → Fn−2 → . . . → F1 → F0
an exact sequence of free R-modules. If n ≥ dim(R), then ker(Fn−1 → Fn−2 ) is maximal
Cohen-Macaulay.
Proof. See [Yos90, Proposition 1.4].
Definition 3.7. Let R be a ring and M an R-module. Let M ∨ := HomR (M, R). The module
M ∨∨ is called the reflexive hull of M. We say that M is reflexive if the natural map
M −→ M ∨∨
m 7−→ evm
is an isomorphism.
Note that reflexive hulls are torsion-free, since HomR (M, N) is torsion-free if N is. They
appear for example as modules of sections of certain sheaves over the punctured spectrum.
We give the precise statement.
Lemma 3.8. Let (R, m) be a reduced, two-dimensional Cohen-Macaulay ring. Let M be a
f the O
finitely generated, torsion-free R-module. We denote by M
Spec(R) -module associated
to M and we set U := Spec(R) \ {m}. If R is Gorenstein in codimension one, we have the
isomorphism
f ) M ∨∨ .
Γ(U, M|
U
Proof. See Propositions 3.7 and 3.10 of [BD08].
Lemma 3.9. Let (R, m) be a reduced, two-dimensional Cohen-Macaulay ring which is
Gorenstein in codimension one. Then kernels of maps between finitely generated reflexive
modules are again reflexive.
3.1. PROPERTIES OF MAXIMAL COHEN-MACAULAY MODULES
79
Proof. Let 0 → K → M1 → M2 be an exact sequence of finitely generated R-modules
with M1 and M2 reflexive. From this sequence, we obtain the commutative diagram
/
0
/
K
0
/
‹ )
Γ(U, K
|U
/
M1
/ Γ(U, g
M1|U )
M2
/ Γ(U, g
M2|U ),
‹ ) K ∨∨ , since K - as a submodule of a
which induces an isomorphism K Γ(U, K
|U
torsion-free module - is itself torsion-free.
Our next goal is to show that for a two-dimensional normal domain maximal CohenMacaulay is equivalent to reflexive. For this we will need the following proposition.
Proposition 3.10. Let R be a ring and M a finitely generated R-module. Then M is
reflexive if and only if
(i) Mp is reflexive for all p ∈ Spec(R) with depthRp (Rp ) ≤ 1 and
(ii) depthRp (Mp ) ≥ 2 for all p ∈ Spec(R) with depthRp (Rp ) ≥ 2.
Proof. See [BH98, Proposition 1.4.1].
Theorem 3.11. Let R be a two-dimensional normal domain and M a finitely generated
R-module. Then M is maximal Cohen-Macaulay if and only if it is reflexive.
Proof. We give the proof of [Kar09, Proposition 1.28]. First assume that M is maximal Cohen-Macaulay. Then M satisfies condition (ii) of Proposition 3.10. Since R is
normal, we have depthRp (Rp ) ≤ 1, whenever dim(Rp ) ≤ 1 by Serre’s criterion. Since Rp is
regular for those primes, the localization Mp is free, hence reflexive. Note that this part of
the proof does not need the condition dim(R) = 2.
For the contrary, let M be reflexive. Then, since R is normal, by condition (ii) of Proposition 3.10, the modules Mp are maximal Cohen-Macaulay for all p ∈ Spec(R) such that
dim(Rp ) = 2. The modules Mp for dim(Rp ) = 0 are of course maximal Cohen-Macaulay.
For the primes p with dim(Rp ) = 1, the modules Mp are reflexive, hence torsion-free,
hence maximal Cohen-Macaulay.
Definition 3.12. Let R be a ring. We say that R is of finite Cohen-Macaulay type if
there are only finitely many isomorphism classes of indecomposable, maximal CohenMacaulay modules.
3.1.1. Maximal Cohen-Macaulay modules and Hilbert-Kunz theory.
Theorem 3.13 (Seibert). Let (R, m, k) be a local ring of dimension d and characteristic
p > 0. Assume that R is of finite Cohen-Macaulay type and that R p is a finite R-module.
(i) If R is Cohen-Macaulay, then eHK (R) ∈ Q.
(ii) If N is maximal Cohen-Macaulay, then eHK (N) ∈ Q.
Proof. See [Sei97, Theorem 4.1].
80
3. MATRIX FACTORIZATIONS AND REPRESENTATIONS AS FIRST SYZYGY MODULES
Moreover, Seibert gives an algorithm to compute the Hilbert-Kunz functions in these
cases. For example he obtains
HK(N, pe ) = 2pe , if e ≥ 1
for all maximal Cohen-Macaulay modules N over R := k~X, Y/(X n + Y 2 ), where n ≥ 3 is
odd (cf. [Sei97, Example 5.1]) and
HK(R, pe ) = 2p3e − 1,
HK(K, pe ) = 2p3e + 1,
HK(U, pe ) = 4p3e + 4
for all e ≥ 0, where R is the second Veronese subring of k~X, Y, Z (with char(k) > 2) and
K, U are representatives for the isomorphism classes of the non-free, indecomposable,
maximal Cohen-Macaulay modules with K = (XY, Y 2 , YZ)R (this is the canonical module)
and
Ö
è
XY
U = K 3 /R Y 2
YZ
(cf. [Sei97, Example 5.2]).
Let k be an algebraically closed field of prime characteristic such that n ∈ N≥2 is invertible
in k. Recall that the n-th Veronese subring R of k[X, Y] is the ring of invariants under the
action X 7→ ε · X, Y 7→ ε · Y, where ε denotes a primitive n-th root of unity. The group
acting on k[X, Y] is isomorphic to Z/(n). Since Z/(n) is commutative, all irreducible
representations are one-dimensional and there are exactly n different equivalence classes
(cf. [Dor71, Theorem 8.1]). By the correspondence between irreducible representations
of the acting group and the indecomposable, maximal Cohen-Macaulay modules over the
ring of invariants (cf. [LW12, Corollary 6.4]), we obtain that the n-th Veronese subring
of k[X, Y] has up to isomorphism exactly n different indecomposable, maximal CohenMacaulay modules of rank one and none of higher rank. This shows that the Picard group
of the punctured spectrum of R is isomorphic to Z/(n). In the next example we want to
demonstrate how one can make use of the finite Cohen-Macaulay type in the application
of Proposition 1.76.
Example 3.14. Let S := k[X, Y] and R its n-th Veronese subring, hence R is generated as
k-algebra by the monomials Zi := X i Y n−i . To compute the Hilbert-Kunz function of R,
we have to control the Frobenius pull-backs of Syz(Z0 , . . . , Zn ) on the punctured spectrum
of R. Since SyzR (Z0 , . . . , Zn ) is the second syzygy module of the quotient R/(Z0 , . . . , Zn ),
it is maximal Cohen-Macaulay. Since this module has rank n, it has to split into n indecomposable, maximal Cohen-Macaulay modules of rank one. Our first goal is to describe
these indecomposable modules as syzygy modules of two-generated monomial ideals.
To compute generators for syzygy modules of monomial ideals of R, we will work over
the factorial domain S as follows. Let P := SyzR (F1 , . . . , Fm ) be a syzygy module of a
P
monomial ideal in R and let (A1 , . . . , Am ) ∈ P. Then treat the relation A j · F j = 0 over S ,
meaning that we write the monomials F j in the variables X and Y. Then one can cancel
P
the common multiple of the F j , written in X, Y. The result is a relation A j ·G j (X, Y) = 0,
where the G j ∈ S are coprime. From this relation one can compute the A j (as elements in
3.1. PROPERTIES OF MAXIMAL COHEN-MACAULAY MODULES
81
R), using the fact that k[X, Y] is factorial. Let i < j ∈ {0, . . . , n} and let (A, B) ∈ SyzR (Zi , Z j ).
Then
AZi + BZ j = 0
⇔
AX i Y n−i = −BX j Y n− j
⇔
AY j−i = −BX j−i
⇔
(A, B) = C(X j−i , −Y j−i )
for some C ∈ k[X, Y]. Since A is an element in R, its degree as element in S is na for some
a ∈ N. Therefore, the degree of C is na + i − j. If a ≥ 2 we see that A and B have a common
factor in R, hence if (A, B) ∈ SyzR (Zi , Z j ) is of minimal degree, the degree of A ∈ S has
to be n. But then the possibilities for C of minimal degree are the monomials of degree
n − j + i. Hence, the R-modules SyzR (Zi , Z j ) are (minimally) generated by n − j + i + 1
elements. Explicitely, one has
ÆÇ
å Ç
å
Ç
å Ç
Z j−i
Z
Zn−1
Zn
, j−i+1 , . . . ,
,
−Z0
−Z1
−Zn− j+i−1
−Zn− j+i
SyzR (Zi , Z j ) =
å∏
.
This shows
SyzR (Zi , Z j ) SyzR (Zl , Zm ) ⇐⇒ |i − j| = |l − m|.
Therefore, it is possible to distinguish the different indecomposable, maximal CohenMacaulay modules by their number of generators. Moreover, this shows that the number
of generators of a maximal Cohen-Macaulay module of rank r is bounded above by n · r,
since it has to split into a direct sum of r indecomposable, maximal Cohen-Macaulay
modules of rank one and each of these has at most n generators. For i ∈ {0, . . . , n − 1} let
{(si,1 , si+1,1 ), . . . , (si,n , si+1,n )}
be a system of generators of SyzR (Zi , Zi+1 ). These give elements
(0, . . . , 0, si, j , si+1, j , 0, . . . , 0) ∈ SyzR (Z0 , . . . , Zn ).
We want to show that
(3.1)
{(0, . . . , 0, si, j , si+1, j , 0, . . . , 0)|1 ≤ j ≤ n, 0 ≤ i ≤ n − 1}
generates SyzR (Z0 , . . . , Zn ). Since all si, j have degree one we have to show that the elements in (3.1) are linear independent. Assuming the contrary, there are i0 ∈ {0, . . . , n − 1},
j0 ∈ {1, . . . , n} such that
(3.2)
0
..
.
0
.
.
.
0
.
.
.
0
0
0
s
s
i0 , j0 X
i0 , j X ti,i
=
+
,
λ
·
j
si0 +1, j0
si0 +1, j
j, j0
i,i0 ti,i+1
0
0
0
.
.
.
.
.
.
.
.
.
0
0
0
82
3. MATRIX FACTORIZATIONS AND REPRESENTATIONS AS FIRST SYZYGY MODULES
where λ j ∈ k and for each i the element (0, . . . , 0, ti,i , ti,i+1 , 0, . . . , 0) is a linear combination
of the elements (0, . . . , 0, si, j , si+1, j , 0, . . . , 0) for j ∈ {1, . . . , n}. Computing the first i0 components of the right hand side in Equation (3.2) (note that the summands for i ≥ i0 + 1 are
not necessary), we obtain the following system of linear equations
ti0 −1,i0 + j, j0 λ j · si0 , j = si0 , j0
ti0 −1,i0 −1 + ti0 −2,i0 −1 = 0
..
.
P
ti0 −l,i0 −l + ti0 −l−1,i0 −l = 0
..
.
t1,1 + t0,1 = 0
t0,0 = 0.
Since t0,0 = 0 and t0,0 Z0 + t0,1 Z1 = 0, we obtain t0,1 = 0. Solving the above linear equation
P
system from the bottom, we obtain the contradiction si0 , j0 = j, j0 λ j · si0 , j .
This shows that the n · n elements from (3.1) are part of a minimal system of generators of
SyzR (Z0 , . . . , Zn ). Since this module is maximal Cohen-Macaulay of rank n, it has at most
n · n generators. All in all, we obtain
SyzR (Z0 , . . . , Zn )
n−1
M
SyzR (Zi , Zi+1 ).
i=0
For i ∈ N write i = n · q + r with q ∈ N and r ∈ {0, . . . , n − 1}. Let (A, B) ∈ SyzR (Z0i , Z1i ). Then
AZ0i + BZ1i = 0
AY ni = −BX i Y (n−1)i
⇔
AY i = −BX i
⇔
⇔
(A, B) = C(X i , −Y i )
for some C ∈ k[X, Y]. An argumentation as above shows that the possibilities for C of
minimal degree are the monomials of degree n − r. This shows that SyzR (Z0i , Z1i ) has
n − r + 1 generators and hence
SyzR (Z0i , Z1i ) SyzR (Z0 , Zr )(m) SyzR (Z0r , Z1r )(l)
for some m, l ∈ Z. All in all, we obtain for every m ∈ Z
Ç
SyzR (Z0i , . . . , Zni )(m) SyzR (Z0r , . . . , Znr )
å
i−r
m − (n + 1)
.
n
If the characteristic of k is p and i = pe , we obtain from Proposition 1.76 as in Example
2.2 by using deg(Proj(R)) = n (cf. [Har87, Exercise I.7.1]) the formula
HK(R, q) =
with Dr := dimk R/ Z0r , . . . , Znr .
n+1 2 2
· (q − r ) + Dr ,
2
3.2. GENERALITIES ON MATRIX FACTORIZATIONS
83
Next, we give an explicit formula for Dr . Of course, we have
(3.3)
Dr =
r
X
dimk R/ Z0r , . . . , Znr
i=0
i+
X
dimk R/ Z0r , . . . , Znr
i.
i≥r+1
Since monomials in R of degree i correspond to monomials in k[X, Y] of degree ni, the
first sum above equals
r
X
(ni + 1) − (n + 1) = n ·
i=0
r(r + 1)
+ r − n.
2
We will show that the second summand in (3.3) vanishes by showing that a monomial in R
of degree r + 1 has a factor of the form Z rj for some j. Let M ∈ R be a monomial of degree
r + 1. This means that M = X i Y n(r+1)−i ∈ k[X, Y] for some i ∈ {0, . . . , n(r + 1)}. Assume that
i ∈ [ar, (a + 1)r) for some a ∈ N. We obtain
M = X i Y n(r+1)−i = X i−ar+ar Y nr+n−i+ar−ar = X ar Y nr−ar X i−ar Y ar+n−i = Zar X i−ar Y ar+n−i ,
showing that all monomials in Rr+1 vanish modulo (Z0r , . . . , Znr ).
Finally, under the assumption gcd(p, n) = 1 the Hilbert-Kunz function of R is given by
n + 1 Ä 2e 2 ä
r(r + 1)
HK(R, pe ) =
· p −r +n·
+ r − n,
2
2
where r is the smallest non-negative representative of the class of pe in Z/(n).
3.2. Generalities on matrix factorizations
Let (S , m, k) be a regular local ring, f ∈ m2 \ {0} and let R := S /( f ). The starting point for
the theory of matrix factorizations is the observation that by the Auslander-Buchsbaum
formula every maximal Cohen-Macaulay R-module M has - viewed as a S -module - a
free resolution of length one:
ϕ
0 → S n → S n → M → 0.
Because ϕ is injective and f · S n is contained in the image of ϕ, one can construct a linear
map ψ : S n → S n satisfying ϕ ◦ ψ = f · Id. Composing this equation from the right by ϕ
and using again that ϕ is injective, one gets ψ ◦ ϕ = f · Id. This motivates the following
definition.
Definition 3.15. For a given non-zero f ∈ m2 , a pair of n × n-matrices (ϕ, ψ) over S with
ϕ ◦ ψ = ψ ◦ ϕ = f · Idn
is called a matrix factorization for f of size n.
A matrix factorization (ϕ, ψ) is called reduced if all entries are non-units.
Example 3.16. The pairs ( f, 1) and (1, f ) are non-reduced matrix factorizations for f of
size one. If f = g · h with g, h ∈ m, then (g, h) and (h, g) are reduced matrix factorizations
for f of size one.
We denote the set of matrix factorizations for f by MFS ( f ). The following two definitions
will enrich its structure to an additive category.
84
3. MATRIX FACTORIZATIONS AND REPRESENTATIONS AS FIRST SYZYGY MODULES
Definition 3.17. Let (ϕ1 , ψ1 ) and (ϕ2 , ψ2 ) be two matrix factorizations for f of size n resp.
m. A morphism from (ϕ1 , ψ1 ) to (ϕ2 , ψ2 ) is a pair of m × n-matrices (α, β) over S such that
the following diagram commutes
ψ1
Sn
α
/
ψ2
Sm
/
Sn
ϕ1
β
/
Sm
ϕ2
/
Sn
α
S m.
We call (ϕ1 , ψ1 ) and (ϕ2 , ψ2 ) equivalent if there is a morphism (α, β) between them, where
α and β are isomorphisms. We will denote a morphism (α, β) from (ϕ1 , ψ1 ) to (ϕ2 , ψ2 ) by
(α, β) : (ϕ1 , ψ1 ) → (ϕ2 , ψ2 ).
Remark 3.18. Note that the commutativity of the second square is enough: Multiplying
ψ2 α by f · Id from the right gives
(ψ2 α)( f · Id) =(ψ2 α)(ϕ1 ψ1 )
=ψ2 (αϕ1 )ψ1
=ψ2 (ϕ2 β)ψ1
=(ψ2 ϕ2 )(βψ1 )
=( f · Id)(βψ1 ).
Definition 3.19. Let (ϕ1 , ψ1 ) and (ϕ2 , ψ2 ) be two matrix factorizations for f of size n resp.
m. We define their direct sum as
åå
å Ç
ÇÇ
ψ1 0
ϕ1 0
.
,
(ϕ1 , ψ1 ) ⊕ (ϕ2 , ψ2 ) :=
0 ψ2
0 ϕ2
Definition 3.20. A matrix factorization (ϕ, ψ) is called indecomposable if it is not equivalent to a direct sum of matrix factorizations.
Example 3.21. Let f = X 2 + Y 3 + YZ 2 . Assume that the ground field is not of characteristic two and contains i. Then (ϕ0 , ϕ0 ), (ϕ1 , ϕ1 ) and (ϕ2 , ϕ2 ) with
á
ϕ0 =
Ç
−X Y 2 YZ 0
Y
X
0 −Z
Z
0
X Y
0 −YZ Y 2 −X
å
ë
Ç
,
å
X
iY 2 + YZ
X
−iY 2 + YZ
ϕ1 =
and
ϕ2 =
Z − iY
−X
Z + iY
−X
are matrix factorizations for f . In this case (α, −α) : (ϕ0 , ϕ0 ) −→ (ϕ1 , ϕ1 ) ⊕ (ϕ2 , ϕ2 ) with
á
α=
i 0 0 −1
0 −1 −i 0
−i 0 0 −1
0 −1 i 0
is an equivalence of matrix factorizations for f .
ë
3.2. GENERALITIES ON MATRIX FACTORIZATIONS
85
Since f · S n ⊆ ϕ(S n ), the cokernel of ϕ is annihilated by f and therefore an R-module.
Taking the sequence of maps and modules
ϕ
ψ
ϕ
. . . −→ S n −→ S n −→ S n −→ coker(ϕ) −→ 0
modulo f , gives a two-periodic free resolution of coker(ϕ) as an R-module. Because of
this, coker(ϕ) is isomorphic to all of its 2i-th syzygy modules that are maximal CohenMacaulay by Proposition 3.6 if 2i is at least dim(R). All in all we get a functor
coker((ϕ, ψ)) := coker(ϕ) : MFS ( f ) −→ C(R),
where C(R) denotes the category of maximal Cohen-Macaulay R-modules. As described
at the beginning of this section, we have a functor Γ : C(R) −→ MFS ( f ) as well.
Theorem 3.22 (Eisenbud). The maps coker and Γ induce a bijection between the sets
of reduced matrix factorizations for f up to equivalence and non-free, maximal CohenMacaulay R-modules up to isomorphism. Moreover, this bijection respects direct sums.
Proof. See [Eis80, Corollary 6.3].
Remark 3.23. Note that the non-reduced matrix factorizations ( f, 1) and (1, f ) of f are
mapped under coker to R resp. 0. Using the above theorem one can show that any matrix factorization (ϕ, ψ) is equivalent to a direct sum of matrix factorizations of the form
L
L
(ϕ0 , ψ0 ) ⊕ai=1 ( f, 1) ⊕bi=1 (1, f ), where the natural numbers a, b are uniquely determined
and (ϕ0 , ψ0 ) is reduced and unique up to equivalence (cf. [Yos90, Remark 7.5]).
Theorem 3.22 generalizes to certain quotients of categories given by the following construction.
Construction 3.24. Let A be an Abelian category and B a subclass of objects of A. For
any A, B ∈ A, we define B(A, B) ⊆ HomA (A, B) as the set of morphisms that factor through
direct sums of objects in B. In fact, B(A, B) is a normal subgroup of HomA (A, B). We
define the category A/B, whose objects are the objects of A, but the morphisms from A
to B are given by the quotient
HomA/B (A, B) := HomA (A, B)/B(A, B).
Definition 3.25. Let MFS ( f ) := MFS ( f )/{(1, f ), ( f, 1)} be the category of equivalence
classes of reduced matrix factorizations and C(R) := C(R)/{R} the category of isomorphism classes of non-free maximal Cohen-Macaulay R-modules.
Now Theorem 3.22 generalizes as follows.
Theorem 3.26 (Eisenbud). The functors coker and Γ induce an equivalence of the categories MFS ( f ) and C(R).
Proof. See [Yos90, Theorem 7.4].
Remark 3.27. If f is prime and (ϕ, ψ) is a matrix factorization for f of size n, then there
is a unit u ∈ S and a number m ∈ {0, . . . , n} such that det(ϕ) = u · f m and det(ψ) = u−1 · f n−m .
In this case coker(ϕ) has rank m and coker(ψ) has rank n − m as R-modules. This can be
seen by localizing at the prime ideal ( f ). Since S ( f ) is a discrete valuation ring, the map
86
3. MATRIX FACTORIZATIONS AND REPRESENTATIONS AS FIRST SYZYGY MODULES
ϕ( f ) on the localization is equivalent to the map f · Idm ⊕ Idn−m . Obviously, coker(ϕ( f ) ) has
rank m over the field Q(R) S ( f ) / f S ( f ) . Note that m ∈ {1, . . . , n − 1}, if (ϕ, ψ) is reduced.
Compare with [Eis80, Proposition 5.6].
The next lemma describes the connection of a maximal Cohen-Macaulay R-module and
its dual in terms of matrix factorizations.
Ä
ä
Lemma 3.28. Given a matrix factorization (ϕ, ψ) for f , the pair ϕT , ψT is again a matrix
factorization for f . Moreover, we have
Ä
ä
coker ϕT (coker(ϕ))∨ .
Proof. The first statement is clear. For a matrix M over S , we denote the matrix over
R given by taking every entry from M modulo f by M. Applying Hom(_, R) to
ϕ
Rn −→ Rn −→ coker(ϕ) −→ 0,
produces the exact sequence
Ä
ϕT
ä
0 −→ (coker(ϕ))∨ −→ Rn −→ Rn −→ coker ϕT −→ 0.
This shows
Ä
ä
T
coker ϕT Rn / ker ψ
Ä
T
im ψ
Ä
ä
ker ϕT (coker(ϕ))∨ .
ä
Since coker(ϕ) and coker ϕT have
a natural R-module structure, they identify with the
Ä ä
T
R-modules coker(ϕ) resp. coker ϕ .
Due to Yoshino (cf. [Yos98]), there is the notion of the tensor product of matrix factorizations for power series in disjoint sets of variables.
Definition 3.29. Let (ϕ f , ψ f ) be a matrix factorization for f in k~X1 , . . . , Xr of size n and
(ϕg , ψg ) a matrix factorization for g in k~Y1 , . . . , Y s of size m. Then
ÇÇ
ˆ g , ψg ) :=
(ϕ f , ψ f )⊗(ϕ
å Ç
ϕ f ⊗ idm idn ⊗ϕg
ψ f ⊗ idm − idn ⊗ϕg
,
− idn ⊗ψg ψ f ⊗ idm
idn ⊗ψg ϕ f ⊗ idm
åå
is a matrix factorization for f + g in k~X1 , . . . , Xr , Y1 , . . . , Y s of size 2nm.
If (α, β) : (ϕ1 , ψ1 ) → (ϕ2 , ψ2 ) is a morphism in MF( f ), then
ÇÇ
ˆ g , ψg ) B
(α, β)⊗(ϕ
å Ç
α ⊗ idm
0
β ⊗ idm
0
,
0
β ⊗ idm
0
α ⊗ idm
åå
ˆ g , ψg ) to (ϕ2 , ψ2 )⊗(ϕ
ˆ g , ψg ) in MF( f +g). Similarly, if (γ, δ) :
is a morphism from (ϕ1 , ψ1 )⊗(ϕ
(ϕ3 , ψ3 ) → (ϕ4 , ψ4 ) is a morphism in MF(g), then
ÇÇ
ˆ δ) B
(ϕ f , ψ f )⊗(γ,
å Ç
idn ⊗γ
0
idn ⊗δ
0
,
0
idn ⊗δ
0
idn ⊗γ
åå
ˆ 3 , ψ3 ) to (ϕ f , ψ f )⊗(ϕ
ˆ 4 , ψ4 ) in MF( f + g).
is a morphism from (ϕ f , ψ f )⊗(ϕ
3.2. GENERALITIES ON MATRIX FACTORIZATIONS
87
Example 3.30. Let d1 , d2 , d3 ∈ N≥2 . Then for each 1 ≤ a ≤ d1 − 1 the pair (X a , X d1 −a ) is
a reduced matrix factorization for X d1 . Similarly, (Y b , Y d2 −b ) and (Z c , Z d3 −c ) with 1 ≤ b ≤
d2 − 1, 1 ≤ c ≤ d3 − 1 are reduced matrix factorizations for Y d2 resp. Z d3 . With these matrix
factorizations we can construct matrix factorizations for X d1 + Y d2 and for X d1 + Y d2 + Z d3 .
For given a, b, c as above, we get
ˆ b , Y d2 −b ))⊗(Z
ˆ c , Z d3 −c )
((X a , X d1 −a )⊗(Y
ÇÇ
=
áá
=
å Ç
X d1 −a −Y b
Xa
Yb
,
Y d2 −b X a
−Y d2 −b X d1 −a
åå
Xa
Yb
Zc
0
d
−b
d
−a
−Y 2
X 1
0
Zc
−Z d3 −c
0
X d1 −a −Y b
d
−c
0
−Z 3
Y d2 −b X a
ˆ c , Z d3 −c )
⊗(Z
ë á d1 −a
X
,
Ä
Y d2 −b
Z d3 −c
0
ä
−Y b
−Z c
0
a
X
0
−Z c
0
Xa
Yb
d
−c
d
−b
Z 3
−Y 2
X d1 −a
ëë
.
Denote the result by (ϕ, ψ). Then (ϕ, ψ) and ϕT , ψT are equivalent. Let
Ç
Ä
ä
å
0 −1
η :=
.
1 0
Then (α, α) : (ϕ, ψ) −→ ϕT , ψT given by
Ç
0 η
α=
η 0
å
is an equivalence of matrix factorizations. Note that the corresponding modules coker(ϕ)
have rank two by Remark 3.27, since det(ϕ) = (X d1 + Y d2 + Z d3 )2 .
This tensor product behaves much better than the tensor product of modules.
Theorem 3.31 (Yoshino). With the previous notations
ˆ g , ψg ) : MF( f ) −→ MF( f + g)
_⊗(ϕ
is an exact functor. Moreover, if (ϕg , ψg ) is reduced, this functor is faithful.
Proof. See [Yos98, Lemmata 2.8 and 2.11].
Remark 3.32. In Chapter 3 of [Yos98] Yoshino discusses how the indecomposability
of matrix factorizations behaves under tensor products. If char(k) , 2, then by [Yos98,
Example 3.8] the matrix factorizations computed in Example 3.30 are indecomposable if
and only if at most one of the following equalities holds: 2a = d1 , 2b = d2 , 2c = d3 .
Yoshino found also the following decomposition behaviour.
Lemma 3.33. Suppose char(k) , 2 and i ∈ k. For matrix factorizations (ϕ, ϕ) of size n and
ˆ ψ) decomposes as the direct sum (ξ, ζ) ⊕ (ζ, ξ)
(ψ, ψ) of size m the tensor product (ϕ, ϕ)⊗(ψ,
with
(ξ, ζ) = (ϕ ⊗ idm −i(idn ⊗ψ), ϕ ⊗ idm +i(idn ⊗ψ)).
Proof. See [Yos98, Lemma 3.2].
88
3. MATRIX FACTORIZATIONS AND REPRESENTATIONS AS FIRST SYZYGY MODULES
Ä
0ä
0
Ä
0
0
0ä
Example 3.34. Consider the matrix factorizations X d1 , X d1 for X 2d1 and Y d2 , Y d2 for
0
Y 2d2 . Then
Ä
0
0ä
Ä
0ä
0
Ä
0
0
0ä
0
Ä
0
0
0
0ä
X d1 , X d1 ⊗ˆ Y d2 , Y d2 X d1 − iY d2 , X d1 + iY d2 ⊕ X d1 + iY d2 , X d1 − iY d2 .
If we are in the case a = d10 = d1 /2 and b = d20 = d2 /2 of Example 3.30,
ÄÄ
0ä
0
Ä
0
0 ää
X d1 , X d1 ⊗ˆ Y d2 , Y d2
Ä
⊗ˆ Z c , Z d3 −c
ä
splits by Remark 3.32 (or Theorem 3.31) as the tensor products of the two direct summands above with (Z c , Z d3 −c ).
3.3. Constructing isomorphisms between maximal Cohen-Macaulay modules and
first syzygy modules of ideals
In this section we construct isomorphisms between maximal Cohen-Macaulay modules
and first syzygy modules of ideals over rings of dimension two. To be more precise,
given a full set M1 , . . . , Mn of representatives for the isomorphism classes of indecomposable, maximal Cohen-Macaulay R-modules, we want to find for each i non-zero elements
F1 , . . . , Fm ∈ R such that
Mi SyzR (F1 , . . . , Fm ).
This has to be possible, since first syzygy modules of ideals are second syzygy modules
of quotient rings.
From now on we assume that f is irreducible, that R = S /( f ) is a local Cohen-Macaulay
normal domain of dimension two and that R has no singularities outside the origin. In
this situation all maximal Cohen-Macaulay modules are locally free on the punctured
spectrum of R (cf. Lemma 3.5). We will denote the punctured spectrum of R by U.
Now we describe a way to find a representation of a maximal Cohen-Macaulay module,
given by a matrix factorization, as a syzygy bundle on U. At the end of the chapter we
discuss an algebraic approach to find an isomorphism between maximal Cohen-Macaulay
modules and first syzygy modules of ideals that reflects the main idea of restricting ϕ to a
submodule of lower rank better than the rather sheaf-theoretic approach. We will see that
this algebraic approach does not work very properly.
Construction 3.35. For a (reduced) matrix factorization (ϕS , ψS ) of size n, we have a two
periodic free resolution of M := coker(ϕS ) as R-module:
ψ
ϕ
ψ
ϕ
. . . −→ Rn −→ Rn −→ Rn −→ Rn −→ M −→ 0,
where ϕ and ψ arise from ϕS and ψS by taking their entries modulo f . Note that we can
identify M with coker(ϕ).
Therefore we get
M Rn / im(ϕ) Rn / ker(ψ) im(ψ).
Later we will see that in the case of rank one modules the isomorphism M im(ψ) is
already enough to describe M as a syzygy module.
For any subset J of {1, . . . , n} we denote by ψ J the matrix obtained from ψ by keeping all
columns whose index belongs to J (and deleting those with index in J C ).
3.3. CONSTRUCTING ISOMORPHISMS OF MAXIMAL COHEN-MACAULAY MODULES
89
‡
If coker(ϕ) im(ψ) has rank m as an R-module, the sheaf im(ψ)|
U gives a locally free
OU -module of rank m. Therefore a representing first syzygy module of an ideal has to be
a syzygy module of an (m + 1)-generated ideal.
‡
Suppose that we can choose a J ⊆ {1, . . . , n} of cardinality m + 1 such that im(ψ)|
U and
·
1
J
F B im(ψ )|U are isomorphic as sheavesÄof O
ä U -modules , where the isomorphism is the
J
map induced by the natural inclusion im ψ ⊆ im(ψ). A necessary condition for this is
that the matrix ψ J has full rank - namely m - in every non-zero point. Otherwise there
·
J ) has rank at most m − 1.
would be at least one point u ∈ U such that the stalk im(ψ
u
‡
Since im(ψ)|
U comes from a maximal Cohen-Macaulay R-module, it is locally free as
OU -module, giving that
ψJ
Om+1
U −→ F −→ 0
is a surjection of locally free OU -modules of ranks m + 1 and m. Therefore the kernel L
has to be locally free of rank one. This gives an isomorphism
L ⊗ det(F ) OU
of the determinants (cf. Proposition 1.47 (iv)). We distinguish whether the determinant
of F is trivial or not.
Case 1. Assume det(F ) OU . Then the line bundle L has to be trivial. Hence
(3.4)
ψJ
η
0 −→ OU −→ Om+1
U −→ F −→ 0
is a short exact sequence and η has to be the multiplication by a vector (G1 , . . . ,Gm+1 )T
with Gi ∈ R. But then Sequence (3.4) is exactly the dual of the presenting sequence of
SyzU (G1 , . . . ,Gm+1 ), hence there is an isomorphism
SyzU (G1 , . . . ,Gm+1 ) F ∨ .
Moreover, the ideal generated by the Gi has to be m-primary, since the determinant of the
syzygy bundle is trivial. Note that this case cannot appear if F has rank one.
Case 2. Now assume that L is a non-trivial line bundle, hence det(F ) L∨ . Dualizing
the sequence 0 → L → Om+1
U → F → 0 yields
(3.5)
(ψ J )T
∨
0 −→ F ∨ −→ Om+1
U → L −→ 0.
Since L∨ is a line bundle it corresponds to a Cartier divisor D. There is a finite cover
(Ui ) of U such that D|Ui is principal, say D|Ui = (hi ). The multiplication with the common
denominator of the hi defines an embedding of L∨ into OU . With this embedding, we can
extend Sequence (3.5) to
(3.6)
(ψ J )T
µ
0 −→ F ∨ −→ Om+1
U −→ OU −→ T −→ 0,
where T is a non-zero torsion sheaf and the map µ is the multiplication by a row vector
(G1 , . . . ,Gm+1 ). In this case, Sequence (3.6) is the presenting sequence of the syzygy
bundle SyzU (G1 , . . . ,Gm+1 ), where the ideal (G1 , . . . ,Gm+1 ) is not m-primary, since T is
non-zero.
1Note that we have no argument why this should be possible, but it works in all examples.
90
3. MATRIX FACTORIZATIONS AND REPRESENTATIONS AS FIRST SYZYGY MODULES
In both cases the column vector (G1 , . . . ,Gm+1 )T generates the kernel of ψÄJ . äIn order
to compute the Gi , we only need to find an element (F1 , . . . , Fm+1 )T ∈ ker ψ J . Then
(F1 , . . . , Fm+1 )T is just a mutiple of (G1 , . . . ,Gm+1 )T and equality (up to multiplication by
a unit) holds if we can prove that the Fi are coprime. For example this condition is
automatically fullfilled if the ideal generated by the Fi is m-primary. Anyway, the syzygy
modules of the two ideals are isomorphic.
All in all, our task is to find a vector (F1 , . . . , Fm+1 )T in the kernel of ψ J . This gives an
isomorphism
f∨ F ∨ Syz (F , . . . , F
M
m+1 ).
U 1
Since we have the isomorphisms
Ä
ä
Ä
ä
M ∨ coker(ϕ)∨ coker ϕT im ψT ,
we will apply this construction to ψ, if coker(ϕ) is selfdual, or to ψT , if coker(ϕ) is not
selfdual. In both cases we end up with an isomorphism
f Syz (F , . . . , F
M
m+1 ).
U 1
Remark 3.36. Since the underlying ring R is a two-dimensional normal domain the isomorphisms of OU -modules constructed above, extend to isomorphisms of R-modules,
since all involved modules are reflexive (cf. [BD08, Lemma 3.6]).
We will illustrate this construction by an explicit computation.
Example 3.37. In Example 3.30 we found the matrix factorizations
áá
(ϕ, ψ) :=
Xa
Yb
Zc
0
d
−b
d
−a
−Y 2
X 1
0
Zc
−Z d3 −c
0
X d1 −a −Y b
0
−Z d3 −c Y d2 −b X a
ë á d1 −a
X
,
−Y b
−Z c
0
a
X
0
−Z c
0
Xa
Yb
Z d3 −c −Y d2 −b X d1 −a
Y d2 −b
Z d3 −c
0
ëë
for X d1 + Y d2 + Z d3 with 1 ≤ a ≤ d1 − 1, 1 ≤ b ≤ d2 − 1 and 1 ≤ c ≤ d3 − 1.
The corresponding modules coker(ϕ) have rank two and are selfdual, hence we can apply
the above construction to ψ. The computations
á
Y d2 −b
·
Xa
−Y b
Xa
0
d
Z 3 −c
á
−
X d1 −a
·
Yb
ë
−Y b
Xa
0
d
Z 3 −c
á
+
Z d3 −c
·
Xa
ë
−Z c
0
Xa
−Y d2 −b
á
+
Z d3 −c
·
Yb
0
−Z c
Yb
X d1 −a
ë
á d −a ë
X 1
=
ë
Y d2 −b
Z d3 −c
0
,
á d −a ë
X 1
=
Y d2 −b
Z d3 −c
0
‡ ∨ is locally generated by the columns two, three and four. Now we have
show that im(ψ)|
U
to find three non-zero elements F1 , F2 and F3 such that (F1 , F2 , F3 )T belongs to the kernel
of ψ{2,3,4} .
The first row (−Y b , −Z c , 0) of ψ{2,3,4} gives the relation −Y b · F1 − Z c · F2 = 0 from which
we conclude (since Y b and Z c are coprime in R) that F1 = Z c ·G and F2 = −Y b ·G for some
3.4. THE CASE An WITH n ≥ 0
91
non-zero G ∈ R. From the second row (X a , 0, −Z c ) we get the relation Z c · X a ·G − Z c · F3 =
0. This gives F3 = X a · G. Looking at the relations given by the rows three and four of
ψ{2,3,4} , we see that G can be chosen as one. We get the isomorphism
Â{2,3,4} ∨
‡ ∨ im(ψ
im(ψ)|
)|U SyzU (Z c , −Y b , X a ),
U
extending to the global isomorphism
im(ψ)∨ SyzR (Z c , −Y b , X a ).
In fact, one can delete any column from ψ, giving the isomorphisms
im(ψ)∨ SyzR (Z c , −Y b , X a )
SyzR (Z c , X d1 −a , Y d2 −b )
SyzR (Y b , X d1 −a , −Z d3 −c )
SyzR (X a , −Y d2 −b , −Z d3 −c ).
In all four cases the ideal (F1 , F2 , F3 ) is m-primary, hence (F1 , F2 , F3 )T generates the
„
kernel of ψ J and det(coker(ϕ)|
U ) is always trivial.
In the next sections we give a complete list of representatives of reduced, indecomposable
matrix factorizations up to equivalence in the ADE case and compute first syzygy modules
of ideals isomorphic to the maximal Cohen-Macaulay modules given by the matrix factorizations. Proofs for the completeness of the lists of matrix factorizations can be found
in [KST07] and [LW12, Chapter 9, §4]. We use their enumeration of the matrix factorizations, which corresponds to the enumeration of the corresponding Dynkin-diagrams. In
all cases the matrices are various 2n × 2n- matrices and the corresponding modules have
rank n. Note that for an R-module M we will denote the corresponding OSpec(R) -module
again by M.
3.4. The case An with n ≥ 0
Let k be an algebraically closed field of odd characteristic not dividing n + 1 and let
R := k~X, Y, Z/(X n+1 + YZ)
with maximal ideal m. Denote by U := Spec(R) \ {m} the punctured spectrum of R. The
category of reduced, indecomposable matrix factorizations up to equivalence has the objects (ϕm , ψm ) for m ∈ {1, . . . , n} with
ÇÇ
(ϕm , ψm ) =
å Ç
Y X n+1−m
Z X n+1−m
, m
m
X
−Z
X
−Y
åå
.
We always have rank(coker(ϕm )) = 1. The modules
SyzR (Y, X n+1−m )
and
SyzR (X m , −Z)
are generated by the columns of ψm , hence they are isomorphic to coker(ϕm ). Note that
(ψm , ϕm ) (ϕn+1−m , ψn+1−m ) by the morphism
ÇÇ
å Ç
0 1
0 −1
,
−1 0
1 0
åå
92
3. MATRIX FACTORIZATIONS AND REPRESENTATIONS AS FIRST SYZYGY MODULES
and coker(ϕm )∨ = coker(ϕn+1−m ), because (ϕTm , ψTm ) = (ϕn+1−m , ψn+1−m ). Hence, the Picardgroup of R is isomorphic to Z/(n + 1).
Since the OU -modules SyzU (X m , Z) are line bundles, they have a description as the line
bundle of a reflexive ideal. From the short exact sequence of OU -modules
0 → SyzU (X m , Z) → O2U → (X m , Z)|U → 0,
we get by taking determinants
(X m , Z)|U SyzU (X m , Z)∨ SyzU (X m , Y).
Taking global sections Γ(U, _) gives the isomorphism
SyzR (X m , Y) (X m , Z)∨∨ .
It remains to compute the reflexive hulls of the ideals (X m , Z). The reflexive hull of a
finitely generated module over a normal ring is the intersection over all localizations of
that module at primes of height one (cf. the corollary to Theorem 1 in Chapter VII, §4,
no. 2 of [Bou89]). All (X m , Z) have exactly one minimal prime (of height one), namely
(X, Z), hence we get
(X m , Z)∨∨ = { f ∈ R| f ∈ (X m , Z)R(X,Z) } = { f ∈ R| f ∈ (X m )R(X,Z) },
where the last equality holds, since Y gets invertible and Z = −X n+1 /Y ∈ (X m )R(X,Z) .
We show by induction on m, that the ideals (X m , Z) are reflexive. This is clear for m = 1,
since (X, Z) is a prime ideal of height one. Let m > 1 and assume (X l , Z)∨∨ = (X l , Z) for all
l < m. The inclusion (X m , Z) ⊆ (X m , Z)∨∨ is clear. Let f ∈ (X, Z), such that f ∈ (X m )R(X,Z) .
Then there are a, b ∈ R with f = aX + bZ. Since bZ ∈ (X m )R(X,Z) , this gives f − bZ = aX ∈
(X m )R(X,Z) . From this we deduce a ∈ (X m−1 )R(X,Z) , which means a ∈ (X m−1 , Z) by the
induction hypothesis.
Theorem 3.38. The pairwise non-isomorphic modules Mm := SyzR (X m , Z) for m = 1, . . . , n
give a complete list of representatives of the isomorphism classes of indecomposable,
∨ M
non-free, maximal Cohen-Macaulay modules. Moreover, we have Mm
n+1−m and
m
Mm (X , Y).
3.5. The case Dn with n ≥ 4
Let k be an algebraically closed field of odd characteristic not dividing n − 2, let
R := k~X, Y, Z/(X 2 + Y n−1 + YZ 2 )
with maximal ideal m. Denote by U := Spec(R) \ {m} the punctured spectrum of R. In this
case the category of reduced, indecomposable matrix factorizations up to equivalence has
n objects (ϕi , ψi ), i ∈ {1, . . . , n}, given by
Ç
å
X Y n−2 + Z 2
ϕ1 = ψ1 =
,
Y
−X
3.5. THE CASE Dn WITH n ≥ 4
Ç
93
å
X
Y(iY (n−2)/2 + Z)
ϕn−1 = ψn−1 =
Z − iY (n−2)/2
−X
and
Ç
å
X
Y(−iY (n−2)/2 + Z)
ϕn = ψn =
Z + iY (n−2)/2
−X
Ç
å
Ç
å
if n is even,
X + iY (n−1)/2
YZ
,
ϕn−1 =
Z
−X + iY (n−1)/2
X − iY (n−1)/2
YZ
ψn−1 =
Z
−X − iY (n−1)/2
(ϕn , ψn ) =(ψn−1 , ϕn−1 ) if n is odd.
and
Independent of the parity of n we have
á
ϕm = ψ m =
0
YZ
Y m/2
−X Y n−1−m/2 −Z
Y m/2
X
0
−YZ
0
X
−X
0
Z
Y n−1−m/2
ë
if m ∈ {2, . . . , n − 2} is even, and
á
ϕm = ψ m =
Y n−1−(m−1)/2
YZ
0
X
0
−Z
0
X
Y n−2−(m−1)/2
−YZ
Y (m+1)/2
−X
−X
Y (m−1)/2
Z
0
ë
if m ∈ {2, . . . , n − 2} is odd.
The corresponding modules coker(ϕm ) have rank one for m ∈ {1, n − 1, n} and rank two
otherwise. Moreover, coker(ϕ1 ) is selfdual, but coker(ϕn−1 ) and coker(ϕn ) are selfdual
only when n is even. For n odd they are dual to each other. In any case the equivalences
are given by (α, −α) with
Ç
å
0 −1
α :=
.
1 0
The modules coker(ϕm ) are selfdual, since (η, ξ) : (ϕm , ψm ) −→ (ϕTm , ψTm ) with
Ç
å
−α 0
η=ξ=
0 α
Ç
α 0
η = −ξ =
0 α
if m is even and
å
if m is odd
are equivalences of matrix factorizations. This shows that the Picard-group of R is isomorphic to Z/(2) × Z/(2) if n is even and isomorphic to Z/(4) if n is odd.
We now compute representing first syzygy modules of ideals. In the rank one case we only
have to look at the columns of ψ j . The columns of ψ1 generate the modules SyzR (−Y, X)
and SyzR (X, Y n−2 + Z 2 ). Similarly, for n even we get
Ä
Ä
ää
Ä
Ä
ää
coker(ϕn−1 ) SyzR (−iY (n−2)/2 + Z, −X)
SyzR X, Y Z + iY (n−2)/2
coker(ϕn ) SyzR (iY (n−2)/2 + Z, −X)
SyzR X, Y Z − iY (n−2)/2
and
94
3. MATRIX FACTORIZATIONS AND REPRESENTATIONS AS FIRST SYZYGY MODULES
and for n odd we get
Ä
ä
Ä
ä
coker(ϕn−1 ) SyzR −Z, X + iY (n−1)/2
coker(ϕn ) SyzR −Z, X − iY (n−1)/2
SyzR (X − iY (n−1)/2 , YZ)
and
SyzR (X + iY (n−1)/2 , YZ).
The ideals (X, Y), (X, Z ± iY (n−2)/2 ) and (X ± iY (n−1)/2 , Z) are prime ideals of height one,
hence reflexive and we get the ideal representations
coker(ϕ1 ) (X, Y),
coker(ϕn−1 ) (X, Z − iY (n−2)/2 ) and
coker(ϕn ) (X, Z + iY (n−2)/2 ) if n is even,
coker(ϕn−1 ) (Z, X − iY (n−2)/2 ) and
coker(ϕn ) (Z, X + iY (n−2)/2 ) if n is odd.
Dealing with the rank two case, we have to omit one column from ψm (we can work
with ψm instead of ψTm , since coker(ϕm ) is selfdual) such that the rank of the OU -module
corresponding to the image of the reduced matrix remains two.
If we erase the first or fourth column in ψm for m even, the two-minors of the new matrix
vanish in the point (0, 0, 1) ∈ U. Hence, we cannot omit these columns, since we would get
a sheaf whose rank in (0, 0, 1) would be at most one. The following computation shows,
that we can omit the second column of ψm and still have a generating set for im(ψm )|U :
á
Y m/2
·
X
YZ
Y n−1−m/2
X
0
á
Y m/2
·
Z
ë
á m/2 ë
Y
−
YZ
·
X
+
Y n−1−m/2
=
á m/2 ë
Y
ë
−X
0
Z
−Z
0
X
X
·
Z
−Z
0
X
=
á
ë
á
ë
0
−X
Y m/2
−YZ
0
−X
Y m/2
−YZ
.
The rows of ψ{1,3,4}
generate the R-module SyzR (−X, Z, Y n−1−m/2 ), hence we get an isom
morphism of this syzygy module with coker(ϕm ). If we choose the coefficients
YZ Y n−1−m/2
− ,−
,0
X
X
!
and
X
Y n−1−m/2
, 0, −
Z
Z
!
on D(X) resp. D(Z) for the columns one, two, four, we see that we can leave out the third
column. Then the rows of ψ{1,2,4}
generate SyzR (Y m/2 , −Z, X), which gives an isomorm
phism of this syzygy module with coker(ϕm ). Both syzygy modules are syzygy modules
of m-primary ideals, hence the determinant of coker(ϕm )|U is trivial.
Now, only the case where m is odd is left. As in the case where m is even, erasing the
first or fourth column in ψm gives a matrix with all two-minors vanishing in the point
(0, 0, 1) ∈ U. Again we may omit the second or the third column, where the corresponding
coefficients are given in the table below (each row of the table is a syzygy for the columns
3.6. THE CASE E6
95
of ψm on the local piece D(X) or D(Z)).
1
D(X) − Y
n−1− m−1
2
X
2
3
4
−1
0
Y n−1−
Z
−1
YZ
X
− XZ
m+1
− Y X2
−1
0
D(Z)
0
−1
D(X)
− YZ
X
0
D(Z)
X
Z
−
m−1
Y 2
Z
m−1
2
The rows of ψ{1,3,4}
generate SyzR (YZ, X, Y (m+1)/2 ) and the rows of ψ{1,2,4}
generate the
m
m
n−1−(m−1)/2
module SyzR (Y
, X, −YZ). The ideals are not m-primary, hence the determinant
of D B coker(ϕm )|U is not trivial. Since the R-module coker(ϕm ) is selfdual, the determinant D has to be selfdual, too. Hence, if n is odd, we get directly
D coker(ϕ1 )|U .
If n is even, we need to compute the reflexive hull of (X, Y (m+1)/2 , YZ). This ideal has only
the minimal prime (X, Y). Since in the localization at (X, Y) the equality
(X, Y (m+1)/2 , YZ)(X,Y) = (X, Y)(X,Y)
holds, the reflexive hull of (X, Y (m+1)/2 , YZ) is (X, Y) and we obtain
D coker(ϕ1 )|U .
Theorem 3.39. The pairwise non-isomorphic modules
M1 = SyzR (X, Y),
Mm = SyzR (X, Y m/2 , Z)
if m ∈ {2, . . . , n − 2} is even,
Mm = SyzR (X, Y (m+1)/2 , YZ) if m ∈ {2, . . . , n − 2} is odd,
Mn−1 = SyzR (X, Z − iY (n−2)/2 ) and
Mn = SyzR (X, Z + iY (n−2)/2 ) if n is even, or
Mn−1 = SyzR (Z, X + iY (n−1)/2 ) and
Mn = SyzR (Z, X − iY (n−1)/2 ) if n is odd
give a complete list of representatives of the isomorphism classes of indecomposable, nonfree, maximal Cohen-Macaulay modules. Moreover, these modules are selfdual with the
∨ M if n is odd. The determinants of M
exception Mn−1
n
n−1 |U and Mn |U are M1 |U . For the
rank one modules we have the isomorphisms M1 (X, Y), Mn−1 (X, Z − iY (n−2)/2 ) and
Mn (X, Z + iY (n−2)/2 ) if n is even and Mn−1 (Z, X − iY (n−1)/2 ) and Mn (Z, X + iY (n−1)/2 )
if n is odd.
3.6. The case E6
Let k be an algebraically closed field of characteristic at least five, let
R := k~X, Y, Z/(X 2 + Y 3 + Z 4 )
96
3. MATRIX FACTORIZATIONS AND REPRESENTATIONS AS FIRST SYZYGY MODULES
with maximal ideal m. Denote by U := Spec(R) \ {m} the punctured spectrum of R. In this
case the category of reduced, indecomposable matrix factorizations up to equivalence has
six objects, which are given by
Ç
−Z 2 + iX
Y
ϕ5 = ψ6 =
2
2
Y
Z + iX
á
ϕ1 = ψ1 =
á
ϕ3 = ψ4 =
á
ϕ4 = ψ3 =
å
Ç
å
−Z 2 − iX
Y
ϕ6 = ψ5 =
,
2
2
Y
Z − iX
and
−X 0 Y 2 Z 3
0 −X Z −Y
Y
Z3 X 0
Z −Y 2 0 X
ë
,
−Z 2 + iX
0
YZ Y
2
−YZ
Z + iX Y 2 0
0
Y
iX Z
2
Y
−YZ
Z 3 iX
ë
−Z 2 − iX
0
YZ
Y
2
2
−YZ
Z − iX Y
0
0
Y
−iX Z
Y2
−YZ
Z 3 −iX
,
ë
,
−iX −Z 2 YZ
0
Y2
0
−Z 2 −iX
0
0
0
Y
0
0
−iX
−Y
0
Z
,
ϕ2 =
0
2
3
YZ −Y −iX Z
0
Y
0
0
Z −iX 0
0
Y2
Z3
0 YZ 2 −iX
iX −Z 2 YZ 0
Y2 0
−Z 2 iX
0
0
0
Y
0
0
iX
−Y
0
Z
,
ψ2 =
2
3
YZ −Y iX Z
0
0
Y
0
0
Z
iX 0
0
Y2
Z3
0 YZ 2 iX
where coker(ϕ j ) has rank one for j ∈ {5, 6}, rank two for j ∈ {1, 3, 4} and rank three for
j = 2. At first we clarify the dualities. The modules coker(ϕ5 ) and coker(ϕ6 ) are dual to
each other by the equivalence (α, −α) : (ϕ5 , ψ5 ) → (ϕT6 , ψT6 ) with
Ç
å
0 −1
α=
.
1 0
Since the Picard-group of R is given by the isomorphism classes of R, coker(ϕ5 ) and
coker(ϕ6 ), it is isomorphic to Z/(3). The modules coker(ϕ3 ) and coker(ϕ4 ) are dual to
each other by the equivalence (η, −η) : (ϕ3 , ψ3 ) → (ϕT4 , ψT4 ) with
Ç
å
−α 0
.
η=
0 −α
3.6. THE CASE E6
97
The modules coker(ϕ1 ) and coker(ϕ2 ) are the only modules of rank two resp. three left,
hence they are selfdual.
Now we compute representations as first syzygy modules of ideals.
By looking at the columns of ψ5 and ψ6 (since they give modules of rank one) we get
coker(ϕ5 ) SyzR (Y 2 , iX + Z 2 )
SyzR (Z 2 − iX, −Y)
coker(ϕ6 ) SyzR (Y 2 , Z 2 − iX)
SyzR (Z 2 + iX, −Y)
The two ideals (±iX + Z 2 , Y) are prime ideals of height one, hence reflexive. We obtain
the ideal representations
coker(ϕ5 ) (Z 2 + iX, Y)
coker(ϕ6 ) (Z 2 − iX, Y).
Dealing with the matrix factorizations (ϕ j , ψ j ) corresponding to modules of rank two, we
have to delete one column from ψ j .
Since (ϕ1 , ψ1 ) is equivalent to the matrix factorization from Example 3.30 with the values
a = b = c = 1 and di = 1 + i for the parameters, we get the following representations as first
syzygy modules of m-primary ideals:
coker(ϕ1 ) SyzR (X, Y 2 , Z 3 )
SyzR (X, Y, Z)
SyzR (X, Y, Z 3 )
SyzR (X, Y 2 , Z).
Fix a point of the form (x, 0, z) ∈ U, hence either ix + z2 = 0 or ix − z2 = 0. Deleting the
first column from ψ4 , all two-minors of the reduced matrix vanish iff z2 + ix = 0. If z2 = ix
generate SyzR (F1 , F2 , F3 ), we must have F3 = −Z · F2 by the first
and the rows of ψ{2,3,4}
4
row. The third row gives the condition 0 = Y · F1 + (iX − Z 2 ) · F2 = Y · F1 , hence F1 = 0.
This is a contradiction, since SyzR (0, F2 , F3 ) splits. Deleting the second column from ψ4
or the first column from ψ3 if ix = −z2 or the second column from ψ3 if ix = z2 , one gets
similar contradictions. By deleting the third columns we get isomorphisms
im(ψ3 ) SyzR (−iX + Z 2 , YZ, −Y 2 )
and
im(ψ4 ) SyzR (iX + Z 2 , YZ, −Y 2 ).
Deleting the fourth columns we get the isomorphisms
im(ψ3 ) SyzR (iXZ − Z 3 , iXY, Y 2 )
and
im(ψ4 ) SyzR (iXZ + Z 3 , iXY, −Y 2 ).
Since coker(ϕ3 ) and coker(ϕ4 ) are dual to each other, we get the isomorphisms
coker(ϕ3 ) SyzR (iX + Z 2 , YZ, −Y 2 )
SyzR (iXZ + Z 3 , iXY, −Y 2 ),
coker(ϕ4 ) SyzR (−iX + Z 2 , YZ, −Y 2 )
SyzR (iXZ − Z 3 , iXY, Y 2 ),
98
3. MATRIX FACTORIZATIONS AND REPRESENTATIONS AS FIRST SYZYGY MODULES
where the appearing ideals are not m-primary. The ideals (±iX + Z 2 , Y 2 , YZ) have only one
minimal prime, namely (±iX + Z 2 , Y). Since in the localization at (iX + Z 2 , Y) the equality
(iX + Z 2 , Y 2 , YZ)(iX+Z 2 ,Y) = (iX + Z 2 , Y)(iX+Z 2 ,Y)
holds (and similarly for the negative sign), the minimal primes have to be the reflexive
hulls. Hence, we obtain
det(coker(ϕ3 )|U ) coker(ϕ5 )|U
det(coker(ϕ4 )|U ) coker(ϕ6 )|U .
Considering ψ2 we have to erase two columns, since coker(ϕ2 ) has rank three. Not all
choices of two columns are possible, since there are choices {m, l} such that all threeminors of the matrix ψ{1,...,6}\{m,l} vanish. The non-possible choices {m, l} for columns of
ψ2 and points showing that these choices are not possible are given in the next table.
deleted columns
point
{1, 2} ∨ {3, 6} ∨ {4, 5} (i, 0, 1)
{1, 5} ∨ {2, 6} ∨ {3, 4} (i, 1, 0)
{1, 4} ∨ {1, 6} ∨ {4, 6} (0, −1, 1).
Deleting all other pairs of columns is possible and yields isomorphisms
im(ψ2 ) SyzR (F1 , . . . , F4 ),
where the polynomials are given as follows
deleted columns F1
F2
F3
F4
1, 3
Y 2 −iXZ Z 2 −iXY
2, 3
Y2
Z3
−iX YZ 2
3
iXY −Z
−Y 2 iXZ 2
2, 4
YZ −iX −Y 2
Z3
2, 5
3, 5
Z2
iX −YZ −Y 2
iXZ Z 3
−Y 2 −iXY.
5, 6
Note that all the ideals (F1 , F2 , F3 , F4 ) are m-primary.
Theorem 3.40. The pairwise non-isomorphic modules
M1 = SyzR (X, Y, Z),
M2 = SyzR (X, Y 2 , YZ, Z 2 ),
M3 = SyzR (iX + Z 2 , Y 2 , YZ),
M4 = SyzR (−iX + Z 2 , Y 2 , YZ),
M5 = SyzR (−iX + Z 2 , Y),
M6 = SyzR (iX + Z 2 , Y)
give a complete list of representatives of the isomorphism classes of indecomposable, nonfree, maximal Cohen-Macaulay modules. Moreover, M1 and M2 are selfdual, M3∨ M4
and M5∨ M6 . The Determinants of M3 |U resp. M4 |U are M5 |U resp. M6 |U . For the rank
one modules we have the isomorphisms M5 (iX + Z 2 , Y) and M6 (−iX + Z 2 , Y).
3.7. THE CASE E7
99
3.7. The case E7
Let k be an algebraically closed field of characteristic at least five, let
R := k~X, Y, Z/(X 2 + Y 3 + YZ 3 )
with maximal ideal m. Denote by U := Spec(R) \ {m} the punctured spectrum of R. In this
case the category of reduced, indecomposable matrix factorizations up to equivalence has
seven objects, which are given by
Ç
å
X
Y
ϕ7 = ψ7 =
,
2
3
Y + Z −X
á
ϕ1 = ψ1 =
á
ϕ4 = ψ4 =
á
ϕ6 = ψ6 =
X
0 −Y 2 Z
0
X YZ 2 Y
−Y Z −X 0
YZ 2 Y 2 0 −X
−X Z 2
0
YZ X −Y 2
0 −Y −X
Y 2 0 YZ 2
Y
0
Z
X
ë
,
ë
X
0 −YZ Y
0
X
Y 2 Z2
−Z 2 Y −X
0
Y 2 YZ
0
−X
,
ë
,
−X Z 2
YZ
0
Y2
0
YZ
X
0
0
0 −Y
0
0
X
−Y
0
Z
,
ϕ2 = ψ2 =
2
2
−X YZ
0
0 −YZ −Y
Y
0
0
Z
X
0
0 −Y 2 YZ 2 0 Y 2 Z −X
−X
0
YZ
0
0
Y
−YZ
2
2
X
0 −Z
−Y
0
2
Z
0
X
−Y
YZ
0
,
ϕ5 = ψ5 =
0
−YZ −Y 2 −X
0
0
0
−Y
0
0
−X −Z
Y2
0
0
YZ −YZ 2 X
−X 0 YZ −Z 2 0
0
Y2 0
0 −X
0
Z2
0
0
0
Y
2
2
Z
Z
X
0
0
−Y
0
0
0
2
YZ
0
X −Y
0
0
0
,
ϕ3 = ψ3 =
0
0
0
−Y
−X
0
0
Z
0
2
2
2
0 −Y
0
0 −X YZ Z
2
Y
0
0
0 −Z
Z
X
0
0 Y2
0
0 YZ 2 0
0
X
100
3. MATRIX FACTORIZATIONS AND REPRESENTATIONS AS FIRST SYZYGY MODULES
where coker(ϕ j ) has rank one for j = 7, rank two for j ∈ {1, 4, 6}, rank three for j ∈ {2, 5}
and rank four for j = 3. Since there is only one non-trivial isomorphism class of indecomposable, maximal Cohen-Macaulay modules of rank one, the Picard-group of R is
isomorphic to Z/(2).
Again we start by computing the dual modules. Clearly, coker(ϕ7 ) and coker(ϕ3 ) are
selfdual. We will see that all other modules are selfdual, too. Let
Ç
å
0 1
α :=
.
−1 0
The morphisms (αi , βi ) : (ϕi , ψi ) → (ϕTi , ψTi ) given by
å
Ç
α 0
,
α1 = β1 =
0 α
Ç
å
Ç
å
−α 0
α4 = −β4 =
,
0 α
α 0
α6 = β6 =
,
0 −α
Ö
α2 = −β2 =
Ö
α5 = −β5 =
α 0 0
0 α 0
0 0 α
è
,
è
α 0 0
0 α 0
0 0 −α
are equivalences of matrix factorizations.
Computing representations as first syzygy modules of ideals, we start again with the rank
one case. The columns of ψ7 generate SyzR (X, Y) and SyzR (Y 2 + Z 3 , −X). Therefore
coker(ϕ7 ) is isomorphic to both of these syzygy modules.
In the rank two cases, we have to delete one column from the matrices ψ j .
Deleting the second or fourth column from ψ1 , we get a matrix with vanishing two-minors
in (0, 0, 1). By deleting the first resp. the third column, we get isomorphisms
coker(ϕ1 ) SyzR (X, Z, Y 2 ) SyzR (Z, Y, −X).
Deleting the second or fourth column from ψ4 , we get a matrix with vanishing two-minors
in (0, 0, 1). By deleting the first resp. the third column, we get isomorphisms
coker(ϕ4 ) SyzR (Y 2 , X, −YZ 2 ) SyzR (−X, YZ, Y 2 ).
Note that both ideals are not m-primary (since {X, Y} is not a system of parameters).
Deleting the first or fourth column from ψ6 , we get a matrix with vanishing two-minors
in (0, 0, 1). By deleting the second resp. the third column, we get isomorphisms
coker(ϕ6 ) SyzR (X, −Z 2 , Y 2 ) SyzR (Y, Z 2 , −X).
We now deal with the rank three cases. To represent coker(ϕ2 ), we have to erase two
columns from ψ2 . If one of those two is an even column, we get a matrix with all threeminors vanishing in (0, 0, 1). In (1, −1, 0) all three-minors of the matrix obtained from ψ2
3.7. THE CASE E7
101
by deleting the columns one and five vanish. Deleting the columns one and three resp.
three and five, we obtain the isomorphisms
coker(ϕ2 ) SyzR (Y 2 , XZ, −Z 2 , XY) SyzR (Z 2 , X, −YZ, −Y 2 ).
Similarly, to represent coker(ϕ5 ), we have to erase two columns from ψ5 . If one of those
two is the column one, four or six, we get a matrix with all three-minors vanishing in
(0, 0, 1). Deleting the columns two and three, all three-minors vanish in (0, −1, −1) and all
three-minors vanish in (1, −1, 0), when erasing the columns two and five. In this case we
get only an isomorphism by deleting the columns three and five:
coker(ϕ5 ) SyzR (Y 2 , −XZ, YZ 2 , XY).
The ideal is not m-primary (since {XZ, Y} is not a system of parameters).
The last module to consider is coker(ϕ3 ) of rank four, hence we have to omit three columns
in ψ3 . The following choices of columns are not possible, since all four-minors of the
resulting matrix vanish in the given point (a question mark indicates an arbitrary choice)
deleted columns
point
4, ?, ?
(0, 0, 1)
8, ?, ?
(0, 0, 1)
1, 2, ?
(0, 0, 1)
5, 6, ?
(0, 0, 1)
1, 7, ?
(1, −1, 0)
3, 6, ?
(1, −1, 0)
2, 5, ?
(0, −1, −1).
The other four possibilities ((1,3,5), (2,3,7), (2,6,7) and (3,5,7)) give the isomorphisms
coker(ϕ3 ) SyzR (−XY, Y 2 Z, −XZ 2 , Z 3 , Y 3 )
SyzR (YZ 2 , −XY, Y 2 + Z 3 , Z 4 , XZ 2 )
SyzR (XZ 2 , −Z 4 , Y 3 , XY, −Y 2 Z 2 )
SyzR (Z 3 , −Y 2 − Z 3 , −XZ, −YZ 2 , −XY).
Note that the second ideal is not monomial. Multiplying all generators by Y yields a
monomial ideal and changes the syzygy module only by an isomorphism.
Theorem 3.41. The pairwise non-isomorphic modules
M1 = SyzR (X, Y, Z),
M2 = SyzR (X, Y 2 , YZ, Z 2 ),
M3 = SyzR (XY, XZ, Y 2 , YZ 2 , Z 3 ),
M4 = SyzR (X, Y 2 , YZ),
M5 = SyzR (XY, XZ, Y 2 , YZ 2 ),
M6 = SyzR (X, Y, Z 2 ),
M7 = SyzR (X, Y)
102
3. MATRIX FACTORIZATIONS AND REPRESENTATIONS AS FIRST SYZYGY MODULES
give a complete list of representatives of the isomorphism classes of indecomposable, nonfree, maximal Cohen-Macaulay modules. Moreover, all M j are selfdual and det(M j |U )
M7 |U for j ∈ {4, 5, 7}. The rank one module M7 is isomorphic to the ideal (X, Y).
3.8. The case E8
Let k be an algebraically closed field of characteristic at least seven, let
R := k~X, Y, Z/(X 2 + Y 3 + Z 5 )
with maximal ideal m. Denote by U := Spec(R) \ {m} the punctured spectrum of R. In this
case the category of reduced, indecomposable matrix factorizations up to equivalence has
eight objects, given by
á
X 0
Y
Z
4
0 X Z −Y 2
Y 2 Z −X 0
Z 4 −Y 0 −X
ϕ1 = ψ1 =
á
X 0
Y
Z2
3
0 X Z −Y 2
Y 2 Z 2 −X 0
Z 3 −Y 0 −X
ϕ8 = ψ8 =
ë
,
ë
,
X −Z 2 YZ 0 −Y 2
3
−X
0 0
0
−Z
0
0
−X Y
0
ϕ2 = ψ2 =
0
2 X
4
−YZ
Y
Z
−Y
0
0 Z −X
0
Y 2 Z 4 0 −YZ 3
−X 0
0
Z2
0
YZ
3
X
−Z
0 −Y 2
2
0 −Z
−X
Y
0
ϕ6 = ψ6 =
Z3
2
0
Y
X −YZ 2
0
−Y
0
0
−X
Y2
0 −YZ 2 0
Z4
0
Y
Z
,
0
0
X
Y
0
0
,
0
Z
X
−X 0 −YZ Z 2
0
0 Y2 0
0 −X
Z3
0
0
0
0 Y
2
0
Z
X
0
0
−Y
0
0
Z 3 YZ
2
0
X −Y
0
0 0
ϕ3 = ψ3 =
,
3 Z
0
0
0
−Y
−X
0
Z
0
0 −Y 2 0
0 −X 0 Z 2
2
Y
0
0
0
Z
−Z X 0
0 Y2
0
0
0
Z3 0 X
3.8. THE CASE E8
103
X
0
0
0 −Z 3 0
0 −Y
2
YZ −X
0
0
0 Z
Y2
0
2
3
0
0
−X
Z
0
Y
−Z
0
0
2
2
0
0
X −Y
0
0
Z
,
ϕ7 = ψ7 =
−Z 2
0
0 −Y −X 0
0
0
0
Z3 Y 2
0 YZ 2 X
0
0
0
Y −Z 2 0
0
0
X
Z
−Y 2 0
0
Z3
0
0
0 −X
X
0
YZ 0
0 −Z 2 Z 3
0 −Y 2 0
0
−X
0
0
0
0
0 −Z 2 0
Y
2
0
0
−X Z
0
0
0
Y
0
0
0
X
0
0 −Y 2 0
0
0
YZ Z 3
2
3
0
Z
0
0
X
−Y
0
0
Z
0
,
ϕ4 = ψ4 = 3
0
0
0 −Y 2 −X
0
0
0 Z2
−Z
0
0
0
−Y
0
0
−X
0
0
Z
0
3
2
2
−Z
Y
0
0
0 YZ
X
0
0
−Y
0
0
0
Z2
0
0
Z
−X 0
0
Y 2 YZ 2 0
0
0
Z4
0
0
X
−X 0
0
0
0
0
0
Z2
0
0
0
Y
0 −X −YZ
3
2
2
0
0
0
Z
−Z
0
0
Y
0
2
3
0
0
X
0
0
−Z
0
0
Z
−Y
0
0
YZ
3
2
0
0
X
−Z
0
0
0
−Y
0
0
0
2
0
0
0
−Z
−X
0
0
Y
0
0
0
0
0
3
2
2
2
0 −Z
0
0
−X −Y
0
0
0 YZ Z
,
ϕ5 = ψ5 =
Z2 Z2
0
0
0
−Y
X
0
0
0
0
0
Z3
2
2
0
0
0
Y
0
0
X −YZ
0
0
0
0
0
0
−Y
0
0
0
0
−X
0
0
Z
0
2
3
2
4
0 −Y −Z
0
0 YZ
0
0
−X −Z
0
2
0
Y
0
0
Z
0
0
0
0
−Z X
0
Y2 0
0
0 −YZ 2 0
0
0
Z4
0
0
X
where coker(ϕ j ) has rank two for j ∈ {1, 8}, rank three for j ∈ {2, 6}, rank four for j ∈ {3, 7},
rank five for j = 4 and rank six for j = 5. Since there are no non-trivial modules of rank
one, the Picard-group of R is trivial and all syzygy modules will be syzygy modules of
m-primary ideals.
Note that (ϕ1 , ψ1 ) and (ϕ8 , ψ8 ) are equivalent to the matrix factorizations from Example
3.30 with d1 = 2, d2 = 3, d3 = 5, a = b = 1 and c = 1 resp. c = 3. We get the isomorphisms
coker(ϕ1 ) SyzR (X, Y, Z)
SyzR (X, Y 2 , Z 4 )
SyzR (X, Y 2 , Z)
SyzR (X, Y, Z 4 ),
104
3. MATRIX FACTORIZATIONS AND REPRESENTATIONS AS FIRST SYZYGY MODULES
coker(ϕ8 ) SyzR (X, Y, Z 2 )
SyzR (X, Y 2 , Z 3 )
SyzR (X, Y 2 , Z 2 )
SyzR (X, Y, Z 3 ).
Now we Äshowäthat all other modules coker(ϕ j ) are selfdual. This is obvious for j = 4, 5.
0 1 . Then the morphisms (α , β ) : (ϕ , ψ ) −→ (ϕT , ψT ) given by
Let α := −1
j j
j
j
j
j
0
Ö
α2 = −β2 =
Ö
α6 = −β6 =
è
α 0 0
0 α 0
0 0 α
−α 0 0
0 α 0
0 0 α
á
α3 = β3 =
−α
0
0
0
á
α7 = −β7 =
,
è
,
ë
0 0
0
α 0
0
0 −α 0
0 0 −α
α 0 0 0
0 −α 0 0
0 0 α 0
0 0 0 −α
,
ë
are equivalences of matrix factorizations.
We continue with the computations of representations as syzygy modules.
Considering ψ2 and ψ6 we have to delete two columns. The following choices {l, m} are
not possible, since they give matrices with all three-minors vanishing in the given point:
point
deleted columns from ψ2 deleted columns from ψ6
(0, 1, −1)
{1, 4} ∨ {1, 6} ∨ {4, 6}
{2, 4} ∨ {2, 6} ∨ {4, 6}
(1, −1, 0)
{1, 5} ∨ {2, 6} ∨ {3, 4}
{1, 6} ∨ {2, 5} ∨ {3, 4}
(1, 0, −1)
{1, 2} ∨ {3, 6} ∨ {4, 5}
{1, 4} ∨ {2, 3} ∨ {5, 6}.
All other choices give isomorphisms coker(ϕ2 ) SyzR (F1 , F2 , F3 , F4 ) with
deleted columns
1, 3
2, 3
2, 4
2, 5
3, 5
5, 6
F1 F2
F3
F4
2
2
Y −XZ −Z
XY
2
4
Y
−Z
X YZ 3
4
XY Z
−Y 2 XZ 3
YZ −X Y 2
Z4
Z2
X
YZ −Y 2
4
XZ −Z
Y 2 XY
3.8. THE CASE E8
105
and coker(ϕ6 ) SyzR (F1 , F2 , F3 , F4 ) with
deleted columns F1
F2
F3
2
1, 2
−Y −XY Z 3
1, 3
Y 2 YZ 2
X
3
1, 5
Z
X
−Y 2
3, 5
−X YZ
Z3
4
3, 6
Z
XY XZ 2
4, 5
Y2
XZ −Z 3
F4
XZ 2
−Z 4
YZ 2
Y2
−Y 2
XY.
We get isomorphisms coker(ϕ3 ) SyzR (F1 , . . . , F5 ) with
deleted columns F1
F2
F3
F4
F5
1, 3, 5
XY −Y 2 Z XZ 2
Z3
X2
1, 5, 8
XZ 3 X 2 −YZ 4 −XY 2 −Y 2 Z
2, 4, 7
Y 2 Z −XY
Z4
−X 2
XZ 3
2
4
2, 5, 7
XZ
Y
−Z
XY
−YZ 3
3, 5, 7
Z3
Y2
XZ
YZ 2
XY
3
2
2
2
4, 6, 7
YZ
X
−XZ −Y Z −XY 2 .
We can represent coker(ϕ7 ) as SyzR (F1 , . . . , F5 ) with
deleted columns F1
F2
F3
F4
F5
2
2
2
4
1, 2, 3
−X
XY −Y Z
Z
−XZ 3
1, 3, 7
Z 4 −XY
Y2
XZ 2 −YZ 3
2
2
4
2, 3, 4
X
−XZ
YZ
Y 2 Z −XY 2
3, 5, 7
Y 2 −XZ −YZ 2
Z4
XY
2
4
2
5, 6, 7
XY Y Z
−Z
−XZ
X2
3
4
2
2
2
6, 7, 8
XZ
YZ
Y Z
XY
X2.
In all other cases the resulting matrix has vanishing four-minors in (at least) one of the
“test points” (0, 1, −1), (1, 0, −1) and (1, −1, 0).
Similarly, we have coker(ϕ4 ) SyzR (F1 , . . . , F6 ) with
deleted columns
1, 2, 3, 5
1, 2, 3, 6
1, 2, 4, 6
1, 2, 5, 7
1, 3, 5, 7
3, 5, 7, 9
3, 5, 9, 10
3, 6, 9, 10
6, 7, 8, 9
F1
F2
F3
F4
F5
F6
−XY 2
Z6
−X 2 XYZ 2 YZ 3
XZ 4
−Y 4
Z6
−XY 2 Y 3 Z 2 −XZ 3
Y 2Z4
Y4
XZ 4
Y 2Z3
XY 3 −X 2 Z XY 2 Z 2
2
3
4
−X
XZ
YZ
XY 2
Y 2Z
XYZ 2
X2
−XYZ −Y 2 Z 2 XZ 3
Z4
−XY 2
Z4
XY
−Y 2 Z XZ 2
YZ 3
−Y 3
3
4
2
6
3
−XZ
YZ
XY
Z
−Y
−XYZ 2
Y 2 Z 3 XZ 4
X2Y
−Z 6 −XY 2 −X 2 Z 2
4
2
3
−YZ
XY
XZ
X 2 Z −Y 2 Z 2 X 2 Y.
In all other cases the resulting matrix has vanishing five-minors in one of the test points.
106
3. MATRIX FACTORIZATIONS AND REPRESENTATIONS AS FIRST SYZYGY MODULES
Finally, we can represent coker(ϕ5 ) as SyzR (F1 , . . . , F7 ) with
deleted columns
F1
1, 3, 4, 6, 11
−X 2 Z 2
1, 3, 5, 6, 11
−XZ 4
1, 3, 5, 7, 11
−YZ 4
1, 3, 7, 9, 11
Y 2Z2
1, 7, 9, 10, 11
XYZ 2
2, 3, 5, 7, 9
−XY 2
2, 3, 5, 7, 12
Y 2Z4
3, 5, 7, 9, 11
X2
7, 8, 9, 10, 11
Y4
F2
F3
F4
F5
F6
F7
−Y 4
XZ 4
−XY 3 Y 2 Z 3
YZ 6
XY 2 Z 2
−Y 4
Z6
−Y 3 Z 2 −XY 2 −XYZ 3 Y 2 Z 4
XY 2
−Z 6
XYZ 2
X2
−Y 2 Z 3 −XZ 4
3
2
4
2
XZ
X
YZ
−XY
−Z 6
XYZ 2
Z 6 −Y 2 Z 3 −XY 2
XZ 4
Y4
−Y 3 Z 2
Y 3 Z −XYZ 2 Y 2 Z 3
XZ 4
Z5
Y4
3
6
2
2
4
2
3
XY
YZ
XY Z
−Y
X Z
XZ 4
Z 5 −XYZ −Y 2 Z 2 −XZ 3
YZ 4
−XY 2
YZ 5 −XZ 4 XY 2 Z −Y 2 Z 3 −X 2 Z 2 XY 3 .
In all other cases the resulting matrix has vanishing six-minors in one of test points.
Theorem 3.42. The pairwise non-isomorphic modules
M1 = SyzR (X, Y, Z),
M2 = SyzR (X, Y 2 , YZ, Z 2 ),
M3 = SyzR (XY, XZ, Y 2 , YZ 2 , Z 3 ),
M4 = SyzR (XY, XZ 2 , Y 3 , Y 2 Z, YZ 3 , Z 4 ),
M5 = SyzR (XY 2 , XYZ 2 , XZ 4 , Y 4 , Y 3 Z, Y 2 Z 3 , Z 5 ),
M6 = SyzR (X, Y 2 , YZ, Z 3 ),
M7 = SyzR (XY, XZ, Y 2 , YZ 2 , Z 4 ),
M8 = SyzR (X, Y, Z 2 )
give a complete list of representatives of the isomorphism classes of indecomposable,
non-free, maximal Cohen-Macaulay modules. Moreover, all M j are selfdual.
3.9. An idea of a purely algebraic construction of the isomorphisms
In this section we explain an idea how to find the isomorphisms from the previous sections
with purely algebraic methods.
Given a matrix factorization (ϕ, ψ) of size n and rank m, we want an isomorphism from
coker(ϕ) to SyzR (I), where I := (F1 , . . . , Fm+1 ) is an ideal (with unknown generators). Let
f : Rm+1 −→ R be the map h(F1 , . . . , Fm+1 )T , _i. Our goal is to find the Fi . Since we
want to restrict the map ϕ to Rm+1 , the idea is to factor f as f = p ◦ ϕ ◦ ι with linear
maps ι : Rm+1 −→ Rn and p : Rn −→ R. We restrict ourselves to inclusions ι that send
the standard base vectors of Rm+1 to (pairwise different) standard base vectors of Rn . To
simplify notation, lets assume im(ι) = Rm+1 × {0}n−m−1 . In this case ιT is right-inverse to ι,
giving pϕ = f ιT . But f ιT is f on the first m + 1 components and zero on the last n − 1 − m
components. Therefore p has to be a syzygy for the last n − 1 − m columns of ϕ and for
the first m + 1 columns of ϕ we get pϕc,i = fi , where ϕc,i denotes the i-th column of ϕ
3.9. AN IDEA OF A PURELY ALGEBRAIC CONSTRUCTION OF THE ISOMORPHISMS
107
(i ∈ {1, . . . , m + 1}). From the commutative diagram
/
0
/
0
/
ker(ϕ)
/
ker( f )
ϕ
Rn
/
Rn
p
ιT
f
Rm+1
/
R
we get an induced map ker(ϕ) −→ ker( f ), which is in fact an inclusion. Then one has to
show that this inclusion is surjective, which is equivalent to the condition that ker(ϕ) −→
ker( f ) splits, since ker(ϕ) is indecomposable and has the same rank as ker( f ).
Example 3.43. Let R := k[X, Y, Z]/(X d1 + Y d2 + Z d3 ) with di ∈ N≥2 and consider the matrix
á
ϕ=
Xa
Yb
Zc
0
d
−b
d
−a
2
1
−Y
X
0
Zc
−Z d3 −c
0
X d1 −a −Y b
d
−c
0
−Z 3
Y d2 −b X a
from Example 3.30. Choosing p := (1, 0, 0, 0) and
á
ι :=
1
0
0
0
0
1
0
0
ë
ë
0
0
1
0
,
we get pϕι = (X a , Y b , Z c ). Therefore we get the inclusion ker(ϕ) −→ SyzR (X a , Y b , Z c ). We
have to show that this map splits. That means we have to show that every (A, B,C) ∈
SyzR (X a , Y b , Z c ) extends uniquely to an element s ∈ ker(ϕ), which has to be of the form
(A, B,C, D), since ιT (s) = (A, B,C) (this already shows that there is at most one extension).
Multiplying AX a + BY b + CZ c = 0 by −Y d2 −b gives
− AX a Y d2 −b − BY d2 − CY d2 −b Z c
=0
⇐⇒ − AX a Y d2 −b + B(X d1 + Z d3 ) − CY d2 −b Z c
=0
⇐⇒X a (−AY d2 −b + BX d1 −a ) − Z c (−BZ d3 −c + CY d2 −b )
= 0.
Since X a and Z c are coprime in R this gives
−AY d2 −b + BX d1 −a = −Z c D
and
− BZ d3 −c + CY d2 −b = −X a D
for some D ∈ R.
Similarly, multiplying AX a + BY b + CZ c = 0 by −Z d3 −c gives
− AX a Z d3 −c − BY b Z d3 −c − CZ d3
=0
⇐⇒ − AX a Y d2 −b + BY b Z d3 −c + C(X d1 + Y d2 )
=0
⇐⇒X a (−AY d2 −b + CX d1 −a ) + Y b (−BZ d3 −c + CY d2 −b )
=0
⇐⇒X a (−AY d2 −b + CX d1 −a ) − X a Y b D
= 0.
The last line is equivalent to −AY d2 −b + CX d1 −a = Y b D, since X a is a non-zero divisor.
Now we can extend (A, B,C) ∈ SyzR (X a , Y b , Z c ) uniquely to (A, B,C, D) ∈ ker(ϕ).
108
3. MATRIX FACTORIZATIONS AND REPRESENTATIONS AS FIRST SYZYGY MODULES
Note that ϕ and ϕι had the same rank in this example. But this is not true in general.
Example 3.44. Let R := k[X, Y, Z]/(X 2 + Y n−1 + YZ 2 ) and consider
á
ϕm :=
−X
0
Z
Y n−1−m/2
0
YZ
Y m/2
−X Y n−1−m/2 −Z
Y m/2
X
0
−YZ
0
X
ë
for m even. In this case we may choose p to be either (1, 0, 0, 0) or (0, 0, 0, 1). In these cases
ι is the map (r1 , r2 , r3 ) 7→ (0, r1 , r2 , r3 ) resp. (r1 , r2 , r3 ) 7→ (r1 , r2 , r3 , 0). For the first pair (p, ι)
we get pϕm ι = (−X, Y n−1−m/2 , −Z) and for the second pair we get pϕm ι = (Z, Y m/2 , X). In
both cases we already saw that ker(ϕm ) coker(ϕm ) is isomorphic to the first syzygy
modules of these ideals. But the rank of ϕm ι in the point (0, 0, 1) is only one for both
choices. Choosing ι to embed R3 inside R4 outside the second resp. the third component
(in these cases the rank of ϕm ι would be two in every point), one gets f = (−X, YZ, Y m/2 )
resp. f = (Y n−1−m/2 , −YZ, X), whose first syzygy modules are not isomorphic to ker(ϕm ):
In the first case we have to find a B which extends (0, Y m/2−1 , −Z) to (0, B, Y m/2−1 , −Z) ∈
ker(ϕm ). The third row of ϕm gives the condition BY m/2 + Y m/2−1 X = 0, which is equivalent to X = −BY. This is a contradiction. Similarly, if the syzygy (Z, Y n−2−m/2 , 0) in
the second case would extend to (Z, Y n−2−m/2 ,C, 0) ∈ ker(ϕm ), the second row of ϕm gives
−XY n−2−m/2 + CY n−1−m/2 = 0, which is equivalent to X = CY.
The last example already showed that this approach causes a lot of troubles:
(i) There is no canonical choice for ι. Even the ranks of ϕ and ϕι might be different
in some points.
(ii) To find p one has to find a simultaneous syzygy for up to five column vectors of
length up to twelve.
(iii) Even if one finds possible maps ι and p, the induced inlusion ker(ϕ) −→ SyzR ( f )
might not be surjective.
All those disadvantages have analogues in the geometric approach from the third section.
But in the geometric approach these analogues are advantages (at least in all our explicit
examples):
(i) For every choice of columns such that the rank of the reduced matrix remains
the same, one gets a representation.
(ii) The generators of the ideal can be computed from the rows of the reduced matrices.
(iii) There are no “false friends“, meaning that everything that looks like a solution
is in fact a solution.
3.10. A comment on the graded situation
Since we want to use the results from this chapter to control the Frobenius pull-backs of
the sheaves SyzC (X, Y, Z), where C B Proj(R), we need a graded version of our results.
Therefore we summarize briefly Chapter 15 of [Yos90] (the original paper is [AR89]).
3.10. A COMMENT ON THE GRADED SITUATION
109
Let R be a positively-graded, Cohen-Macaulay ring with R0 = k a field. Denote by m the
graded maximal ideal and by R̂ the m-adic completion of R. We will need the following
categories.
symbol
objects
morphisms
grC(R) finitely generated graded R-modules degree-preserving homomorphisms
grM(R) graded maximal C-M R-modules degree-preserving homomorphisms
C(R̂)
finitely generated R̂-modules
homomorphisms
M(R̂)
maximal C-M R̂-modules
homomorphisms
We say that R is of finite graded Cohen-Macaulay type if there are - up to degree shift
- only finitely many isomorphism classes of graded, indecomposable, non-free, maximal
Cohen-Macaulay modules. The goal is to find a relation between the (graded) indecomposable, maximal Cohen-Macaulay modules over R and over R̂.
Recall that taking the m-adic completion −̂ is a functor from grM(R) to M(R̂) that restricts
to a functor from grC(R) to C(R̂).
We get the following lemma (cf. [Yos90, Lemma 15.2]).
Lemma 3.45.
(i) If M ∈ grM(R) is indecomposable, then M̂ ∈ M(R̂) is indecomposable.
(ii) If M, N ∈ grM(R) are indecomposable such that M̂ N̂ in M(R̂), then M N in
grM(R) up to degree shift.
An immediate consequence of the previous lemma is, that R is of finite graded CohenMacaulay type if R̂ is of finite Cohen-Macaulay type. Moreover, the graded analogous of
the Theorems 3.38 - 3.42 still give a full list of representatives of the isomorphism classes
of graded, indecomposable, non-free, maximal Cohen-Macaulay R-modules. But so far
we might have repetitions in the lists. At least if R0 = k is perfect, we can get rid of this.
First we need a definition.
Definition 3.46.
(i) We call a finitely generated R̂-module M graduable if there
exists a finitely generated graded R-module A such that  M.
(ii) An R̂-homomorphism f : M → N with M, N graduable is called a graduable
homomorphism if there is a graded homomorphism of finitely generated graded
R-modules A, B, such that the following diagram of R̂-modules and homomorphisms commutes.
M
N
/
f
/
Â
ĝ
B̂.
The next theorem, proven by Auslander and Reiten, gives the promised converse.
Theorem 3.47 (Auslander, Reiten). Suppose R0 = k is perfect. If R is of finite graded
Cohen-Macaulay type, then R̂ is of finite Cohen-Macaulay type. Moreover, in this case all
maximal Cohen-Macaulay R̂-modules are graduable.
Proof. See [Yos90, Theorem 15.14].
110
3. MATRIX FACTORIZATIONS AND REPRESENTATIONS AS FIRST SYZYGY MODULES
As a consequence of the last theorem, we obtain that the following definition makes sence.
Definition 3.48. Let S = k[X1 , . . . , Xm ] be positively-graded with k a perfect field. Let
f ∈ S +2 be homogeneous of degree d. Then a graded matrix factorization for f of size
m is a pair of m × m matrices ϕ, ψ with entries in S , such that ϕ ◦ ψ = ψ ◦ ϕ = f · idm and
the morphisms given by ϕ resp. ψ are homogeneous of degree zero resp. d. A morphism
from a matrix factorization into another is again a pair of matrices (α, β) such that the
corresponding diagram commutes and the morphism given by α and β are homogeneous
of degree zero.
By Theorem 3.47, the isomorphisms coker(ϕ) SyzR (F1 , . . . , Fm+1 ) of (ungraded) Rmodules constructed in the last sections, are graduable in the sence that they hold in the
graded situation up to a degree shift. But this shift cannot be computed with the definition
of a graded matrix factorization, since the modules coker(ϕ) have no “natural” grading.
We illustrate this by an example.
Example 3.49. Consider the hypersurface R := k[X, Y, Z]/(X n + YZ) of type An−1 with
n ≥ 1, graded by deg(X) = 2 and deg(Y) = deg(Z) = n. Let S := k[X, Y, Z]. From the matrix
factorizations
ÇÇ
å Ç
åå
Y X n−m
Z X n−m
(ϕm , ψm ) =
, m
X m −Z
X
−Y
we obtained the isomorphisms
coker(ϕm ) SyzR (X m , −Z)
SyzR (Y, X n−m ).
(3.7)
For all η ∈ Z the sequence
ψm
ϕm
S (−η − 2m + n) ⊕ S (−η) −→ S (−η − 2m + 2n) ⊕ S (−η + n) −→ S (−η − 2m + n) ⊕ S (−η)
shows that (ϕm , ψm ) are graded matrix factorizations for X n + YZ. Reducing the above
sequence by X n + YZ, we obtain the short exact sequence
0 → im(ψm ) → R(−η − 2m + n) ⊕ R(−η) → im(ϕm ) → 0.
The presenting sequences of the syzygy modules in (3.7) are
/
0
0
/
/
SyzR (X m , −Z)
SyzR (Y, X n−m )
/
/
R(−2m) ⊕ R(−n)
R(−n) ⊕ R(−2n + 2m)
/
(X m , Z)
/
0,
(Y, X n−m )
/
0.
Choosing η = n, we obtain coker(ϕm ) SyzR (X m , −Z) as graded modules and coker(ϕm )
SyzR (Y, X n−m ) is an isomorphism of graded modules, if we choose η = 2n − 2m.
Since every value of η is as good as any other, the ungraded isomorphisms coker(ϕ)
SyzR (F1 , . . . , Fm+1 ) have no natural graded analogues.
CHAPTER 4
The Hilbert-series of first syzygy modules of certain monomial ideals
in N-graded rings of the form k[X, Y1 , . . . , Yn ]/(X d − F(Y1 , . . . , Yn ))
The main results of this chapter are the Theorems 4.5 and 4.7, which allow us to compute the Hilbert-series of certain syzygy modules. In the second section we will compute the Hilbert-series of the syzygy modules representing the isomorphism classes of
the non-free, indecomposable, maximal Cohen-Macaulay modules over surface rings of
type ADE. We will see that in many cases these Hilbert-series carry enough information to
detect the isomorphism class of a given indecomposable, maximal Cohen-Macaulay module. It contains also enough information to exclude many possible splitting behaviours of
a given maximal Cohen-Macaulay module. Note that we will always work with the description of the Hilbert-series as rational functions und call these rational functions again
Hilbert-series.
4.1. A short exact sequence
The goal of this section is to provide an invariant for the modules SyzR (F1 , . . . , Fn ) from
Theorems 3.38-3.42. This invariant should be efficiently computable for the modules
q
q
SyzR (F1 , . . . , Fn ) at least in the case where n = 3 and the Fi are monomials. One idea was
to use the Hilbert-series as invariant. But how to compute it for various powers pe of various primes? The solution to this question was inspired by a work of Brenner who showed
that on a standard-graded projective curve C := Proj(R) with R := k[X, Y, Z]/(X d − F(Y, Z))
all generators of SyzC (X a , Y b , Z c ) (with a, b, c ∈ N≥1 ) come from (P1k )2 = (Proj(k[Y, Z]))2
(cf. [Bre05b, Lemma 1.1]). We give the precise statement.
Lemma 4.1. Let k be a field and let F(Y, Z) ∈ k[Y, Z] be a homogeneous polynomial of
degree d. Assume that the projective curve C := Proj(k[X, Y, Z]/(X d − F(Y, Z))) is smooth.
Let a, b, c ∈ N≥1 and let a = dq + r with 0 ≤ r < d and q ∈ N. Then there exists for every
m ∈ Z a surjective morphism
SyzC (F q , Y b , Z c )(m − r) ⊕ SyzC (F q+1 , Y b , Z c )(m) → SyzC (X a , Y b , Z c )(m).
Let k be a field and
R := k[X, Y1 , . . . , Yn ]/(X d − F(Y1 , . . . , Yn ))
with deg(X) = α, deg(Yi ) = βi and F homogeneous of degree dα (d, n, α, βi ∈ N≥1 ). Let
V1 , . . . , Vm+l be monomials in the Yi with m, l ≥ 1 and let a ∈ N≥1 . The goal of this section
is to compute the Hilbert-series of R-modules of the form
SyzR (X a · V1 , . . . , X a · Vm , Vm+1 , . . . , Vm+l ).
111
112
4. HILBERT-SERIES OF FIRST SYZYGY MODULES OF MONOMIAL IDEALS
Definition 4.2. With the previous notations we define for j ∈ N the module
S j := SyzR (F j · V1 , . . . , F j · Vm , Vm+1 , . . . , Vm+l ).
Lemma 4.3. Let a = d · q + r with 0 ≤ r ≤ d − 1 and q ∈ N. Then the homogeneous map
ϕ s : Sq (s − α · r) ⊕ Sq+1 (s) −→ SyzR (X a · V1 , . . . , X a · Vm , Vm+1 , . . . , Vm+l )(s),
which sends ( f1 , . . . , fm+l ), (g1 , . . . , gm+l ) to
( f1 + X d−r · g1 , . . . , fm + X d−r · gm , X r · fm+1 + gm+1 , . . . , X r · fm+l + gm+l )
is surjective for all s ∈ Z.
Proof. We check that the restriction of ϕ s to each direct summand is well-defined. Let
( f1 , . . . , fm+l ) ∈ Sq (s − α · r). The computation
f1 · X a · V1 + . . . + fm · X a · Vm + X r · fm+1 · Vm+1 + . . . + X r · fm+l · Vm+l
= X r · ( f1 · X dq · V1 + . . . + fm · X dq · Vm + fm+1 · Vm+1 + . . . + fm+l · Vm+l )
= X r · ( f1 · F q · V1 + . . . + fm · F q · Vm + fm+1 · Vm+1 + . . . + fm+l · Vm+l )
=0
shows that the restriction of ϕ s to Sq (s − α · r) is well-defined (it has the correct degree,
since every summand of the term in brackets in the last line is homogeneous of degree
s − α · r and X r has degree α · r).
Now let (g1 , . . . , gm+l ) ∈ Sq+1 (s). In this case we compute
X d−r · g1 · X a · V1 + . . . + X d−r · gm · X a · Vm + gm+1 · Vm+1 + . . . + gm+l · Vm+l
= g1 · F q+1 · V1 + . . . + gm · F q+1 · Vm + gm+1 · Vm+1 + . . . + gm+l · Vm+l
= 0,
which shows that the restriction of ϕ s to Sq+1 (s) is well-defined (it has the correct degree,
since every summand in the last line has degree s).
We now check the surjectivity. Let
h := (h1 , . . . , hm+l ) ∈ SyzR (X a · V1 , . . . , X a · Vm , Vm+1 , . . . , Vm+l )(s).
We write
hi := hi,0 + hi,1 X + . . . + hi,d−2 X d−2 + hi,d−1 X d−1
with hi, j ∈ k[Y1 , . . . , Yn ] for i = 1, . . . , m + l and j = 0, . . . , d − 1. From the equation
h1 · X a · V1 + . . . + hm · X a · Vm + hm+1 · Vm+1 + . . . + hm+l · Vm+l = 0
we get a system of equations by considering the k[Y1 , . . . , Yn ]-coefficients corresponding
to X j for j = 0, . . . , d − 1. Explicitly, these equations are
0 = h1,σ( j) · X σ( j) · X a · V1 + . . . + hm,σ( j) · X σ( j) · X a · Vm
(4.1)
+ hm+1, j · X j · Vm+1 + . . . + hm+l, j · X j · Vm+l ,
where σ( j) ∈ {0, . . . , d − 1} with σ( j) ≡ j − r modulo d. Since we have
h=
d−1
X
j=0
(h1,σ( j) · X σ( j) , . . . , hm,σ( j) · X σ( j) , hm+1, j · X j , . . . , hm+l, j · X j ),
4.1. A SHORT EXACT SEQUENCE
113
we may assume h to be one of those summands for a fixed j, hence
h := (h1,σ( j) · X σ( j) , . . . , hm,σ( j) · X σ( j) , hm+1, j · X j , . . . , hm+l, j · X j ).
We will show that h comes either from Sq (s − α · r) or from Sq+1 (s).
For j < r, we have σ( j) = j−r +d. Factoring out X j in equation (4.1) and using X a = F q X r ,
we get
0 = X j · (h1,σ( j) · X d−r · X r · F q · V1 + . . . + hm,σ( j) · X d−r · X r · F q · Vm
+ hm+1, j · Vm+1 + . . . + hm+l, j · Vm+l )
= X j · (h1,σ( j) · F q+1 · V1 + . . . + hm,σ( j) · F q+1 · Vm + hm+1, j · Vm+1 + . . . + hm+l, j · Vm+1 ).
This shows that
X j (h1,σ( j) , . . . , hm,σ( j) , hm+1, j , . . . , hm+l, j )
belongs to Sq+1 (s). Under ϕ s it is mapped to h.
For j ≥ r, we have σ( j) = j − r. Again by using X a = F q X r , we get from equation (4.1)
0 = h1,σ( j) · X σ( j) · X r · F q · V1 + . . . + hm,σ( j) · X σ( j) · X r · F q · Vm
+ hm+1, j · X σ( j)+r · Vm+1 + . . . + hm+l, j · X σ( j)+r · Vm+l
= X r · (h1,σ( j) · X σ( j) · F q · V1 + . . . + hm,σ( j) · X σ( j) · F q · Vm
+ hm+1, j · X σ( j) · Vm+1 + . . . + hm+l, j · X σ( j) · Vm+l ).
This shows that
X σ( j) · (h1,σ( j) , . . . , hm,σ( j) , hm+1, j , . . . , hm+l, j )
belongs to Sq (s − α · r). Under ϕ s it is mapped to h.
Lemma 4.4. The homogeneous map
ψ s : SyzR (X a+d−2r · V1 , . . . , X a+d−2r · Vm , Vm+1 , . . . , Vm+l )(s − α · r) −→ Sq (s − α · r) ⊕ Sq+1 (s),
which sends (h1 , . . . , hm+l ) to
(X d−r · h1 , . . . , X d−r · hm , hm+1 , . . . , hm+l ), (−h1 , . . . , −hm , −X r · hm+1 , . . . , −X r · hm+l )
is injective for all s ∈ Z.
Proof. Clearly, these maps are injective, provided they exist. The power X a+d−2r is
well-defined, since a + d − 2r = dq + r + d − 2r = dq + d − r is positive for r < d. Let
h := (h1 , . . . , hm+l ) ∈ SyzR (X a+d−2r · V1 , . . . , X a+d−2r · Vm , Vm+1 . . . , Vm+l )(s − α · r).
Then the first component of the image of h satisfies
X d−r · h1 · F q · V1 + . . . + X d−r · hm · F q · Vm + hm+1 · Vm+1 + . . . + hm+l · Vm+l
=h1 · X a+d−2r · V1 + . . . + hm · X a+d−2r · Vm + hm+1 · Vm+1 + . . . + hm+l · Vm+l
= 0.
The second component of the image of h satisfies (use F q = X a−r )
− h1 · F q+1 · V1 − . . . − hm · F q+1 · Vm − X r · hm+1 · Vm+1 − . . . − X r · hm+l · Vm+l
= − X r · (h1 · X a+d−2r · V1 + . . . + hm · X a+d−2r · Vm + hm+1 · Vm+1 + . . . + hm+l · Vm+l )
= 0.
114
4. HILBERT-SERIES OF FIRST SYZYGY MODULES OF MONOMIAL IDEALS
Theorem 4.5. For all s ∈ Z we have a short exact sequence
0 −→ SyzR (X a+d−2r · V1 , . . . , X a+d−2r · Vm , Vm+1 , . . . , Vm+l )(s − α · r)
ψs
−→ Sq (s − α · r) ⊕ Sq+1 (s)
ϕs
−→ SyzR (X a · V1 , . . . , X a · Vm , Vm+1 , . . . , Vm+l )(s) −→ 0.
Proof. By the two previous lemmatas we only have to check the exactness at the
middle spot. At first we show ϕ s ◦ ψ s = 0. With
h := (h1 , . . . , hm+l ) ∈ SyzR (X a+d−2r · V1 , . . . , X a+d−2r · Vm , Vm+1 , . . . , Vm+l )(s − α · r)
we have
ϕ s (ψ s (h)) = ϕ s ((X d−r · h1 , . . . , X d−r · hm , hm+1 , . . . , hm+l ),
(−h1 , . . . , −hm , −X r · hm+1 , . . . , −X r · hm+l ))
= (X d−r · h1 + X d−r · (−h1 ), . . . , X d−r · hm + X d−r · (−hm ),
X r · hm+1 + (−X r · hm+1 ), . . . , X r · hm+l + (−X r · hm+l ))
= (0, . . . , 0).
Let t := ( f1 , . . . , fm+l ), (g1 , . . . , gm+l ) ∈ ker ϕ s . This yields
fi + X d−r gi = 0 for all i = 1, . . . , m
X r fi + gi = 0 for all i = m + 1, . . . , m + l
and we get
t = (−X d−r g1 , . . . , −X d−r gm , fm+1 , . . . , fm+l ), (g1 , . . . , gm , −X r fm+1 , . . . , −X r fm+l )
= ψ s (−g1 , . . . , −gm , fm+1 , . . . , fm+l ).
Remark 4.6. Note that the syzygy modules Sq and Sq+1 appearing in the middle spot
are already defined over k[Y1 , . . . , Yn ]. If n = 2 they split by Hilberts syzygy theorem as a
direct sum of m + l − 1 (degree shifted) copies of R.
Theorem 4.7. Let M := SyzR (X a · V1 , . . . , X a · Vm , Vm+1 , . . . , Vm+l ). Then the Hilbert-series
of M is given by
H M (t) =
(tα·r − tα·d ) · HSq (t) + (1 − tα·r ) HSq+1 (t)
.
1 − tα·d
Proof. Let a0 := a + d − 2r = dq + d − r, r0 := d − r and
0
0
M 0 := SyzR (X a · V1 , . . . , X a · Vm , Vm+1 , . . . , Vm+l ).
Since a0 + d − 2r0 = a, Theorem 4.5 yields
H M (t) = tα·r HSq (t) + HSq+1 (t) − tα·r H M 0 (t)
H M 0 (t) = tα·r HSq (t) + HSq+1 (t) − tα·r H M (t).
0
0
Substituting H M 0 (t) in the first formula and solving for H M (t) gives the result.
4.2. EXAMPLES
115
4.2. Examples
Using the representation of the isomorphism classes of indecomposable, non-free, maximal Cohen-Macaulay modules over surface rings of type ADE as first syzygy modules of
ideals from Chapter 3, Theorem 4.7 enables us to compute their Hilbert-series if the ideal
in the representation is monomial. If the ideal in the representation is not monomial, we
compute the Hilbert-series of the syzygy module directly, using the short exact sequences
0 → SyzR (F1 , . . . , Fm+1 ) →
0 →
(F1 , . . . , Fm+1 )
→
Lm+1
i=1
R(− deg(Fi )) → (F1 , . . . , Fm+1 ) → 0,
R
→ R/(F1 , . . . , Fm+1 ) → 0.
From these sequences one obtains
Ñ
(4.2)
HSyzR (F1 ,...,Fm+1 ) (t) =
m+1
X
é
tdeg(Fi ) − 1 · HR (t) + HR/(F1 ,...,Fm+1 ) (t).
i=1
We start with a few remarks on the expected structure of the Hilbert-series. This structure
is given by the following lemma (cf. [BH98, Exercise 4.4.12]).
Lemma 4.8. Let R be a positively-graded algebra over a field k. Let M , 0 be a finitely
generated R-module and S a graded Noether normalization of R/ Ann(M) generated by
elements of degree a1 , . . . , ad with d = dim(M). Then the following holds.
(i) There is a polynomial Q(t) ∈ Z[t, t−1 ] such that
Q(t)
.
ai
i=1 (1 − t )
H M (t) = Qd
(ii) We have Q(1) = rankS (M) > 0.
(iii) If M is Cohen-Macaulay, the coefficients of Q(t) are non-negative.
Since the modules in question are maximal Cohen-Macaulay, the coefficients in Q(t) will
be non-negative and d = 2. Since Ann(M) = 0 in all cases, the Noether normalization S
of R/ Ann(M) is in fact a Noether normalization of R. This can be chosen to be generated
by two parameters from R. Since the equation of the An singularities is isomorphic to
X 2 + Y 2 + Z n+1 = 0 (if char(k) , 2), we may choose Y, Z as a parameter system in all
cases. Then the rank of R over S is two (since R S [X]/(X 2 + F(Y, Z))) and we get
Q(1) = rankS (M) = 2 · rankR (M).
We will always express the Hilbert-series as a rational function, since the explicit descriptions as power series are not very enlighting. Note that the isomorphisms M
SyzR (F1 , . . . , Fn ) from Chapter 3 are only isomorphisms of R-modules. By Theorem 3.47,
these isomorphisms extend after a degree shift to isomorphisms of graded R-modules, say
M(−l) SyzR (F1 , . . . , Fn ) as graded R-modules for some l ∈ Z. In this situation, we have
H M (t) = tl · HSyzR (F1 ,...,Fn ) (t).
Note that we will always compute the Hilbert-series of the syzygy module.
116
4. HILBERT-SERIES OF FIRST SYZYGY MODULES OF MONOMIAL IDEALS
4.2.1. The case An . Let R := k[X, Y, Z]/(X n+1 −YZ) with deg(X) = 2, deg(Y) = deg(Z) =
n + 1 and n ≥ 0. For any i ∈ {1, . . . , n}, let Mi := SyzR (X i , Z) (compare Theorem 3.38).
With the notations from Theorem 4.7, we have a = i and d = 2n + 2. In all cases for i, we
get q = 0 and r = i. Hence the syzygy modules S 0 and S 1 are SyzR (1, Z) resp. SyzR (YZ, Z)
for all i. They are free with basis (Z, −1) resp. (1, −Y) of total degrees n + 1 and 2n + 2.
By Theorem 4.7 we get
(t2i − t2(n+1) )tn+1 + (1 − t2i )t2n+2
(1 − t2 )(1 − tn+1 )2
t2i+n+1 + t2n+2
=
(1 − t2 )(1 − tn+1 )
Note that all these Hilbert-series are different. Since the Hilbert-series of a free R-module
of rank one is given by
H Mi (t) =
1 + tn+1
1 − t2n+2
l
,
=
t
·
(1 − t2 )(1 − tn+1 )2
(1 − t2 )(1 − tn+1 )
we see that the Hilbert-series of the Mi is different from the Hilbert-series of a free Rmodule, since the difference of the exponents in the numerator is |n + 1 − 2i| for Mi and
n + 1 for R(−l).
Therefore, the Hilbert-series of a maximal Cohen-Macaulay module of rank one tells
us not to which isomorphism class the module belongs, since in this situation the dual
modules SyzR (X i , Z) and SyzR (X i , Y) have the same Hilbert-series. In such a situation we
will simply say that the Hilbert-series detects the isomorphism class up to dualizing. Note
that we have a homogeneous isomorphism
HR(−l) (t) = tl ·
SyzR (X j , Y) ⊕ Syz(X j , Z) −→ SyzR (X j , Y, Z),
((A, B), (C, D)) 7−→ (A + C, B, D),
for all j ∈ N, which gives directly the Hilbert-series of SyzR (X j , Y, Z) as
HSyzR (X j ,Y,Z) (t) = HSyzR (X j ,Y) (t) + HSyzR (X j ,Z) (t)
t2n+2 + t2 j+n+1
2
·
(1 − t2 )(1 − tn+1 )
=
2j
2·t ·
if 1 ≤ j ≤ n and
1 + tn+1
otherwise.
(1 − t2 )(1 − tn+1 )
4.2.2. The case Dn . Let R := k[X, Y, Z]/(X 2 + Y n−1 + YZ 2 ) with n ≥ 4 and the degrees
of the variables are deg(X) = n − 1, deg(Y) = 2, deg(Z) = n − 2. According to Theorem
3.39 let
M1 := SyzR (X, Y),
M j := SyzR (X, Y j/2 , Z)
if j ∈ {2, . . . , n − 2} is even,
M j := SyzR (X, Y ( j+1)/2 , YZ)
if j ∈ {2, . . . , n − 2} is odd,
Mn−1 := SyzR (X, Z − iY (n−2)/2 ) and
Mn := SyzR (X, Z + iY (n−2)/2 ) if n is even, or
4.2. EXAMPLES
117
Mn−1 := SyzR (Z, X + iY (n−1)/2 ) and
Mn := SyzR (Z, X − iY (n−1)/2 ) if n is odd.
Using Theorem 4.7, we get
H M1 (t) =
=
H M j (t) =
=
H M j (t) =
=
(tn−1 − t2n−2 ) · t2 + (1 − tn−1 )t2(n−1)
(1 − tn−1 )(1 − tn−2 )(1 − t2 )
tn+1 + t2n−2
,
(1 − tn−2 )(1 − t2 )
(tn−1 − t2n−2 )(t j + tn−2 ) + (1 − tn−1 )(t2(n−1) + tn−2+ j )
(1 − tn−1 )(1 − tn−2 )(1 − t2 )
tn+ j−2 + tn+ j−1 + t2n−3 + t2n−2
if j is even,
(1 − tn−2 )(1 − t2 )
(tn−1 − t2n−2 )(t j+1 + tn ) + (1 − tn−1 )(t2(n−1) + tn+ j−1 )
(1 − tn−1 )(1 − tn−2 )(1 − t2 )
tn+ j−1 + tn+ j + t2n−2 + t2n−1
if j is odd.
(1 − tn−2 )(1 − t2 )
To compute the Hilbert-series of Mn−1 and Mn , we use Formula (4.2). Since the quotients
R/ X, Z ± iY
n−2
2
and R/ Z, X ± iY
n−1
2
are isomorphic to k[Y] as graded R-modules, we obtain for j ∈ {n − 1, n}
Ä
ä
H M j (t) = tn−1 + tn−2 − 1 ·
=
1 − tn−2
1 + tn−1
+
(1 − t2 )(1 − tn−2 ) (1 − t2 )(1 − tn−2 )
t2n−3 + t2n−2
.
(1 − tn−2 )(1 − t2 )
Moreover, we have
HR(−l) (t) = tl ·
1 + tn−1
.
(1 − t2 )(1 − tn−2 )
The Hilbert-series can distinguish the class of M1 from the classes of Mn−1 and Mn if
and only if n ≥ 5, since in these cases the difference of the exponents in the numerator
of H M1 (t) is at least two, while the difference in H Mn−1 (t) and H Mn (t) is one. Moreover,
the Hilbert-series can distinguish the classes of the Mm with m ∈ {2, . . . , n − 2}. Let M be
a maximal Cohen-Macaulay module of rank two, whose Hilbert-series is tl · H M j (t) with
2 ≤ j ≤ n−2 and l ∈ Z. If n = 4, we cannot say anything about the direct sum decomposition
of M - except that it has no free direct summands. For n ≥ 5 the module M might have
M1 as a direct summand only for j = 2. In this case we get M M1 (−l − 1) ⊕ M1 (−l).
We cannot exclude a splitting of M into degree shifted copies of Mn−1 and Mn only by
looking at the Hilbert-series. But we know at least that M has no free direct summands.
118
4. HILBERT-SERIES OF FIRST SYZYGY MODULES OF MONOMIAL IDEALS
4.2.3. The case E6 . Let R := k[X, Y, Z]/(X 2 + Y 3 + Z 4 ) with deg(X) = 6, deg(Y) = 4
and deg(Z) = 3. We define the modules
M1 := SyzR (X, Y, Z),
M2 := SyzR (X, Y 2 , YZ, Z 2 ),
M3 := SyzR (iX + Z 2 , Y 2 , YZ),
M4 := SyzR (−iX + Z 2 , Y 2 , YZ),
M5 := SyzR (−iX + Z 2 , Y),
M6 := SyzR (iX + Z 2 , Y).
Using Theorem 4.7, we get
t7 + t9 + t10 + t12
,
(1 − t4 )(1 − t3 )
t10 + t11 + 2t12 + t13 + t14
H M2 (t) =
.
(1 − t4 )(1 − t3 )
H M1 (t) =
To compute the Hilbert-series of the M j , j ∈ {3, 4, 5, 6}, we use Formula (4.2).
For j = 3, 4 the quotient
R/(±iX + Z 2 , Y 2 , YZ)
is isomorphic as a graded R-module to k[Z] ⊕ Y · k and for j = 5, 6 there is an isomorphism
R/(±iX + Z 2 , Y) k[Z]
as graded R-modules. Hence we obtain
Ä
ä
H M3 (t) = H M4 (t) = t8 + t7 + t6 − 1 ·
1 + t6
1 − t4
+
+ t4
(1 − t4 )(1 − t3 ) (1 − t4 )(1 − t3 )
t11 + t12 + t13 + t14
,
(1 − t4 )(1 − t3 )
Ä
ä
1 − t4
1 + t6
H M5 (t) = H M6 (t) = t6 + t4 − 1
+
(1 − t4 )(1 − t3 ) (1 − t4 )(1 − t3 )
t10 + t12
=
.
(1 − t4 )(1 − t3 )
=
For completeness, the Hilbert-series of a free R-module is given by
HR(−l) (t) = tl ·
1 + t6
.
(1 − t4 )(1 − t3 )
In this case the Hilbert-series can distinguish the class of M1 from the classes of M3 and
M4 . Let M be a maximal Cohen-Macaulay module, whose Hilbert-series is tl · H Mi (t) for
some l ∈ Z and some i ∈ {1, . . . , 6}. Then M has no free direct summands. Moreover, only
4.2. EXAMPLES
119
the following splittings are possible (with ε ∈ {3, 4} and δ, δ0 ∈ {5, 6})
i=1:
M Mδ (−l − 3) ⊕ Mδ0 (−l),
i=2:
M Mε (−l) ⊕ Mδ (−l) or
Mε (−l + 1) ⊕ Mδ (−l − 2),
i ∈ {3, 4} : M Mδ (−l − 1) ⊕ Mδ0 (−l − 2).
4.2.4. The case E7 . Let R := k[X, Y, Z]/(X 2 + Y 3 + YZ 3 ) with deg(X) = 9, deg(Y) = 6
and deg(Z) = 4. We want to compute the Hilbert-series of the modules
M1 := SyzR (X, Y, Z),
M2 := SyzR (X, Y 2 , YZ, Z 2 ),
M3 := SyzR (XY, XZ, Y 2 , YZ 2 , Z 3 ),
M4 := SyzR (X, Y 2 , YZ),
M5 := SyzR (XY, XZ, Y 2 , YZ 2 ),
M6 := SyzR (X, Y, Z 2 ),
M7 := SyzR (X, Y)
appearing in Theorem 3.41. These are
t10 + t13 + t15 + t18
,
(1 − t6 )(1 − t4 )
t14 + t16 + t17 + t18 + t19 + t21
H M2 (t) =
,
(1 − t6 )(1 − t4 )
t18 + t19 + t20 + 2t21 + t22 + t23 + t24
H M3 (t) =
,
(1 − t6 )(1 − t4 )
t16 + t18 + t19 + t21
H M4 (t) =
,
(1 − t6 )(1 − t4 )
t19 + t20 + t21 + t22 + t23 + t24
,
H M5 (t) =
(1 − t6 )(1 − t4 )
t14 + t15 + t17 + t18
H M6 (t) =
,
(1 − t6 )(1 − t4 )
t15 + t18
H M7 (t) =
,
(1 − t6 )(1 − t4 )
1 + t9
HR(−l) (t) = tl ·
.
(1 − t6 )(1 − t4 )
H M1 (t) =
The Hilbert-series can distinguish the indecomposable, maximal Cohen-Macaulay modules. Given a maximal Cohen-Macaulay module M, whose Hilbert-series is tl · H Mi (t)
for some l ∈ Z and some i ∈ {1, . . . , 6}, we cannot exclude a splitting of M into (various)
degree shifted copies of M7 only by looking at H M (t). But we see that M cannot have free
120
4. HILBERT-SERIES OF FIRST SYZYGY MODULES OF MONOMIAL IDEALS
direct summands and if rank(M) ≥ 3 there are a lot of possible splittings excluded by the
Hilbert-series. We list all possible splittings including at least one module of rank at least
two.
i = 2 : M M4 (−l + 2) ⊕ M7 (−l − 3) or
M4 (−l) ⊕ M7 (−l + 1),
i = 3 : M M4 (−l − 5) ⊕ M4 (−l − 6) or
M6 (−l − 7) ⊕ M6 (−l − 9) or
M5 (−l − 2) ⊕ M7 (−l − 9) or
M5 (−l − 3) ⊕ M7 (−l − 6),
i = 5 : M M4 (−l − 3) ⊕ M7 (−l − 5) or
M6 (−l − 5) ⊕ M7 (−l − 6) or
M6 (−l − 6) ⊕ M7 (−l − 4).
Negatively spoken, for i = 2 the module M cannot have M1 or M6 as a direct summand,
for i = 3 the modules M1 and M2 cannot appear as direct summands of M and for i = 5
again M1 is not a direct summand of M.
4.2.5. The case E8 . Let R := k[X, Y, Z]/(X 2 + Y 3 + Z 5 ) with deg(X) = 15, deg(Y) = 10
and deg(Z) = 6. As in the E7 case all modules from Theorem 3.42
M1 := SyzR (X, Y, Z),
M2 := SyzR (X, Y 2 , YZ, Z 2 ),
M3 := SyzR (XY, XZ, Y 2 , YZ 2 , Z 3 ),
M4 := SyzR (XY, XZ 2 , Y 3 , Y 2 Z, YZ 3 , Z 4 ),
M5 := SyzR (XY 2 , XYZ 2 , XZ 4 , Y 4 , Y 3 Z, Y 2 Z 3 , Z 5 ),
M6 := SyzR (X, Y 2 , YZ, Z 3 ),
M7 := SyzR (XY, XZ, Y 2 , YZ 2 , Z 4 ),
M8 := SyzR (X, Y, Z 2 )
are syzygy modules of monomial ideals. Using Theorem 4.7, we compute their Hilbertseries as
t16 + t21 + t25 + t30
,
(1 − t10 )(1 − t6 )
t27 + t31 + t32 + t35 + t36 + t40
H M2 (t) =
,
(1 − t10 )(1 − t6 )
t28 + t31 + t32 + t33 + t35 + t36 + t37 + t40
H M3 (t) =
,
(1 − t10 )(1 − t6 )
t34 + t36 + t37 + t38 + t39 + t40 + t41 + t42 + t43 + t45
H M4 (t) =
,
(1 − t10 )(1 − t6 )
H M1 (t) =
4.2. EXAMPLES
121
t46 + t47 + t48 + t49 + t50 + 2t51 + t52 + t53 + t54 + t55 + t56
,
(1 − t10 )(1 − t6 )
t26 + t28 + t30 + t31 + t33 + t35
,
H M6 (t) =
(1 − t10 )(1 − t6 )
t31 + t32 + t34 + t35 + t36 + t37 + t39 + t40
H M7 (t) =
,
(1 − t10 )(1 − t6 )
t22 + t25 + t27 + t30
,
H M8 (t) =
(1 − t10 )(1 − t6 )
1 + t15
l
HR(−l) (t) = t ·
.
(1 − t10 )(1 − t6 )
We see that the Hilbert-series of a given indecomposable, maximal Cohen-Macaulay
module detects its isomorphism class. Let M be a maximal Cohen-Macaulay module,
whose Hilbert-series is tl · H Mi (t) for some fixed i and l ∈ Z. Then M has no free direct summands. For i ∈ {1, 2, 4, 6, 8} the module M cannot split and we get M Mi (−l).
For i = 3 the only possible splitting is M M8 (−l − 6) ⊕ M8 (−l − 10), for i = 7 only
M M8 (−l − 9) ⊕ M8 (−l − 10) is possible and for i = 5 there is only the possible splitting M M6 (−l − 19) ⊕ M6 (−l − 20).
H M5 (t) =
CHAPTER 5
Relations between the maximal Cohen-Macaulay modules over
surface rings of type ADE
Considering the surface rings of type ADE as rings of invariants under finite subgroups of
SL2 (k), we get ringinclusions among them by the actions of normal subgroups. Especially,
we are interested in the pull-backs of the maximal Cohen-Macaulay modules along these
inclusions. From these pull-backs we will deduce that the Frobenius pull-backs of Syz(m)
over surface rings of type DE are indecomposable for almost all primes.
During this chapter the surface rings of type ADE are named by their type of singularity,
e.g. E8 := k[X, Y, Z]/(X 2 − Y 3 − 1728Z 5 ) (we use the hypersurface equations as they arise
directly from the computation of the rings of invariants) and the corresponding punctured
spectra get the same name but with curly letters, e.g. E8 := Spec(E8 ) \ {m}. For a module
M we will denote the restriction of its sheafification to the punctured spectrum by M.
We use the generators of the finite subgroups of SL2 (C), given in [BD08]. Note that
these representations are well-defined in positive characteristics, if the group order is
invertible. To get explicit generators of the normal subgroups, one might use [GAP12]
and [DGPS12] to compute the corresponding rings of invariants (see Appendix B.1 for
the computations).
Let us very briefly recall some facts from invariant theory. If G is a group, acting on a
ring R by ringautomorphisms and H is a normal subgroup of G, the ring of invariants RG
Ä ä
H and one has RG = RH G/H (cf. [Bre12,
is a subring of RH , the quotient G/H acts
on
R
Ä ä
Ä ä
Proposition 5.1 (3)]). Let X := Spec RG and Y = Spec RH (or the punctured resp.
projective spectra). The inclusion RG ,→ RH induces a morphism ι : Y → X. Given an
OX -module F , its pull-back ι∗ (F ) is an OY -module. If R = k[x, y] and G ( SL2 (k) finite,
we want to understand what happens to the sheaves associated to the indecomposable,
non-free, maximal Cohen-Macaulay modules under these pull-backs. For this purpose let
i : R := k[x, y]G = k[U, V, W]/H1 (U, V, W) ,→ S := k[x, y]H = k[X, Y, Z]/H2 (X, Y, Z)
be an inclusion between two (different) surface rings of type ADE, where H is a normal
subgroup of G. This inclusion is given by sending U, V, W to homogeneous polynomials
in the variables X, Y, Z and induces a morphism
ι : S := Spec(S ) \ {(X, Y, Z)} −→ R := Spec(R) \ {(U, V, W)}.
Let M := SyzR (F1 , . . . , Fm+1 ) be one of the indecomposable, non-free, maximal CohenMacaulay R-modules from Theorem 3.38-3.42. We obtain from Proposition 1.62 the
isomorphism
ι∗ (M) SyzS (i(F1 ), . . . , i(Fm+1 )).
123
124
5. RELATIONS BETWEEN THE MAXIMAL COHEN-MACAULAY MODULES
Now, the OS -module SyzS (i(F1 ), . . . , i(Fm+1 )) has to be the sheaf associated to a (not
necessarily indecomposable) maximal Cohen-Macaulay S -module N and we want to have
an explicit description of N. To get this description, we do the following manipulations
on P := SyzS (i(F1 ), . . . , i(Fm+1 )) (cf. Lemma 1.61).
Step 1: If the i(F j ) have a common factor A, replace all i(F j ) by i(F j )/A, which changes
P only by an isomorphism.
Step 2: If B is a common factor of i(F1 ), . . . , i(Fˇ j ), . . . , i(Fm+1 ) and the ideal (B, i(F j ))
is (X, Y, Z)-primary, replace all i(Fk ), k , j, by i(Fk )/B, which changes P again
only by an isomorphism.
In both steps, we allow also to replace i(F j ) by i(F j ) + G · i(Fl ) for G ∈ S and j , l or by a
constant multiple λ · i(F j ) with λ ∈ k× .
In surprisingly many cases, we end up with an syzygy module appearing explicitly in
Theorem 3.38-3.42 (or in a syzygy module that appeared in the computations in Chapter
3 as an alternative description).
If this is not the case, we compute generators for P and look for relations among these
generators. These relations will detect the module P∨ . To compute generators for P, we
will work over the factorial domain k[x, y] (the ring on which the group acts) as follows.
P
Let (A1 , . . . , Am+1 ) ∈ P. Then treat the relation A j ·i(F j ) = 0 over k[x, y], meaning that we
write the polynomials i(F j ) in the variables x, y. Then one can again cancel the common
P
multiple of the i(F j ), written in x, y. The result is a relation A j · G j (x, y) = 0, where the
G j are coprime. From this relation one can compute the A j (as elements in S !), using the
fact that k[x, y] is factorial. We give an example.
Example 5.1. Let S = k[X, Y, Z]/(X 5 − YZ) with X = xy, Y = x5 and Z = y5 . Consider
(A1 , A2 ) ∈ P := SyzS (X 6 , Y 2 ). Substituting X = xy and Y = x5 in the relation A1 · X 6 + A2 ·
Y 2 = 0, one obtains A1 · x6 y6 + A2 · x10 = 0, which is equivalent to
(5.1)
A1 · y6 + A2 · x4 = 0.
Since k[x, y] is factorial, there has to be a B ∈ k[x, y] such that
(5.2)
(A1 , A2 ) = (−x4 , y6 ) · B.
For all choices of B, the tuple (A1 , A2 ) in (5.2) solves (5.1). On the other hand, for each
(A1 , A2 ) ∈ P, there has to be a B ∈ k[x, y], such that (5.2) holds. All in all, to compute
S -module generators for P, we only have to compute the polynomials B ∈ k[x, y] with
the property that −x4 · B and y6 · B are polynomials in xy, x5 , y5 . Obviously, B = x and
B = y4 have this property and are minimal with respect to this property, in the sence that
every homogeneous monomial B0 < k having this property, has to be a multiple of either
x or y4 . This shows that P is generated by (−x4 , y6 ) · x = (−Y, XZ) and (−x4 , y6 ) · y4 =
(−X 4 , Z 2 ). Since P does not appear in Theorem 3.38, we compute relations among the
two generators. Since X · (−X 4 , Z 2 ) − Z · (−Y, XZ) = 0, we obtain
P SyzS (X, Z)∨ SyzS (X, Y).
Of course, the computation gets much more complicated for higher ranks or if the polynomials X = X(x, y), Y = Y(x, y), Z = Z(x, y) have high degrees or are not monomial. For this
5.1. THE CASE An−1 WITH n ≥ 1
125
reason, we stop in the exceptional cases E6 , E7 , E8 after Step 2 and compute generators
for the obtained syzygy modules with Macaulay2 [GS].
Z/(4n)
: O
t|4n
/
/
D2n+2
_
D2n
O
A4n−1
_ q
,
Z/(t)
O r
?
#
t|2n
t|2n
$ ?
Z/(2n)
: O
2|t
/
?
Dn+2
Dn
/
A _
; t−1
-
A2n−1
_ q
,
Z/(2)m r
t|4n
2|t
#
2|n
2|n
$ ?
Z/(4)
/
D2
/ T
_
EO 6
/
D4
/
;A
D 1
-
A3
2
?
I
O
E7
E8
Table 5.1. Inclusions of the finite subgroups (up to conjugation) of SL2 (C)
(left part) and of the corresponding rings of invariants (right part). All
direct inclusions in the left part are inclusions of normal subgroups.
5.1. The Case An−1 with n ≥ 1
Consider natural numbers t and n, where t is a divisor of n. LetÄ ε be äa primitive n-th root
of unity. On k[x, y] the group action of Z/(n), generated by 0ε ε0−1 , has the invariants
Ä
än/t
, has the
U := xy, V := xn , W := yn and the group action of Z/(t), generated by 0ε ε0−1
t
t
invariants X := xy, Y := x , Z := y . Since Z/(t) is a normal subgroup of Z/(n), invariant
theory tells us that An−1 is a subring of At−1 , where the inclusion is given by U 7→ X,
V 7→ Y n/t , W 7→ Z n/t , because the equalities U = X, V = Y n/t , W = Z n/t hold in k[x, y]. In
the grading (2, t, t) resp. (2, n, n) on At−1 resp. An−1 the inclusion is homogeneous. Denote
the inclusion by ϕt . The goal is to find an easier description of the module on the right
hand side in the isomorphism
ϕ∗t (SyzAn−1 (U l , V)) SyzAt−1 (X l , Y n/t ).
Lemma 5.2. Let l ∈ {1, . . . , n − 1} with l = at + r and 0 ≤ r ≤ t − 1. Let m ≥ a + 1. There is
an isomorphism
SyzAt−1 (X l , Y m ) SyzAt−1 (X r , Y).
Proof. Let (A, B) ∈ SyzAt−1 (X l , Y m ). Then over k[x, y] the relation A · X l + B · Y m = 0
becomes A · (xy)l + B · xtm = 0, which is equivalent to
A · yat+r + B · xt(m−a−1)+t−r = 0.
126
5. RELATIONS BETWEEN THE MAXIMAL COHEN-MACAULAY MODULES
We obtain the existence of a C ∈ k[x, y] such that
(A, B) = (xt(m−a−1) xt−r , −yat yr ) · C.
Since A and B belong to At−1 , all possibilities for C are At−1 -combinations of xr and yt−r .
We obtain
ÆÇ m−a å Ç t−r m−a−1 å∏
X ·Y
Y
l m
.
,
SyzAt−1 (X , Y ) =
r
a
−X · Z
−Z a+1
Since the generators fullfill the relation
å
Ç
å
Ç
X t−r · Y m−a−1
Y m−a
= 0,
+ Xr ·
−Z ·
r
a
−X · Z
−Z a+1
we get the isomorphisms
Ä
ä∨
SyzAt−1 (X l , Y m ) SyzAt−1 (X r , Z)
SyzAt−1 (X r , Y).
Since l ≤ n − 1 =
we obtain
Ä
n
t
ä
− 1 · t + t − 1, we have a ≤ nt − 1, hence by applying the above lemma,
with r ∈ {0, . . . , t − 1}.
SyzAt−1 (X l , Y n/t ) SyzAt−1 (X r , Y)
Remark 5.3. Using ideals to represent the indecomposable, maximal Cohen-Macaulay
An−1 -modules, we have by Theorem 3.38 the isomorphism ϕ∗t ((U l , W)) (X r , Z).
5.2. The Case Dn+2 with n ≥ 2
Ä
ä
The binary dihedral group Dn is generated by ζ := 0ε ε0−1 , where ε is a primitive 2n-th
Ä
ä
0 1 . The generating invariants are
root of unity and τ := −1
0
X := xy(x2n − y2n ), Y := x2 y2 and Z := x2n + y2n ,
fullfilling the equation
X 2 + 4Y n+1 − YZ 2 = 0.
If n is odd, the normal subgroups of Dn are exactly the normal subgroups of < ζ >. If n
is even there are two additional normal subgroups. The first one is generated by τ and ζ 2 .
The second normal subgroup is generated by ζτ and ζ 2 (cf. [DI09, Lemma 4.3]). These
two last groups lead to the same rings of invariances as a direct computation shows. This
ring is a Dn/2+2 -singularity. Recall from Theorem 3.39 that
M1 = SyzDn+2 (X, Y),
Å
ã
n
,
2
Ç
å
n+1
m
M2m−1 = SyzDn+2 (X, Y , YZ)
with 2 ≤ m ≤
,
2
Mn+1 = SyzDn+2 (X, Z − 2Y n/2 ) (with n even),
M2m = SyzDn+2 (X, Y m , Z)
with 1 ≤ m ≤
Mn+2 = SyzDn+2 (X, Z + 2Y n/2 ) (with n even),
5.2. THE CASE Dn+2 WITH n ≥ 2
Mn+1 = SyzDn+2 (X + 2iY
n+1
2
127
, Z) (with n odd),
n+1
2
Mn+2 = SyzDn+2 (X − 2iY , Z) (with n odd)
is a complete list of generators for the isomorphism classes of indecomposable, non-free,
maximal Cohen-Macaulay modules.
5.2.1. Pull-backs to At−1 with t|2n. It is enough to deal with the pull-backs to A2n−1 ,
because every pull-back from Dn+2 to At−1 (with t|2n) factors through A2n−1 . The pullback is induced by the map
ϕ : k[X, Y, Z]/(X 2 + 4Y n+1 − YZ 2 ) −→ k[R, S , T ]/(T 2n − RS ),
given by X 7→ T (R − S ), Y 7→ T 2 , Z 7→ R + S , where T = xy, R = x2n , S = y2n .
Except for the modules Mn+1 and Mn+2 , applying the steps 1 and 2 from the strategy
described at the beginning of this chapter is already enough to compute the pull-backs.
ϕ∗ (M1 ) = ϕ∗ (SyzDn+2 (X, Y))
SyzA2n−1 (T R − T S , T 2 )
SyzA2n−1 (R − S , T )
OA2n−1 ,
ϕ∗ (M2m ) = ϕ∗ (SyzDn+2 (X, Y m , Z))
SyzA2n−1 (T R − T S , T 2m , R + S )
SyzA2n−1 (R − S , T 2m−1 , R + S )
SyzA2n−1 (R, S , T 2m−1 )
SyzA2n−1 (R, T 2m−1 ) ⊕ SyzA2n−1 (S , T 2m−1 ),
ϕ∗ (M2m−1 ) = ϕ∗ (SyzDn+2 (X, Y m , YZ))
SyzA2n−1 (T R − T S , T 2m , T 2 R + T 2 S )
SyzA2n−1 (R − S , T 2m−1 , T R + T S )
SyzA2n−1 (R − S , T 2m−2 , R + S )
SyzA2n−1 (R, S , T 2m−2 )
SyzA2n−1 (R, T 2m−2 ) ⊕ SyzA2n−1 (S , T 2m−2 ).
For i = n+1 and i = n+2 we have ϕ∗ (Mi ) SyzA2n−1 (T n , R), independently of the cardinality of n. We only compute ϕ∗ (Mn+1 ) in the two cases corresponding to the the cardinality
of n. To compute ϕ∗ (Mn+2 ) one has only to change some signs. Let n be even. In this
case we get
n
ϕ∗ (Mn+1 ) = ϕ∗ SyzDn+2 X, Z − 2Y 2
SyzA2n−1 (T R − T S , R + S − 2T n )
SyzA2n−1 (R − S , R + S − 2T n )
SyzA2n−1 (R − S , R − T n ).
128
5. RELATIONS BETWEEN THE MAXIMAL COHEN-MACAULAY MODULES
For (A, B) ∈ SyzA2n−1 (R − S , R − T n ) the corresponding relation over k[x, y] is equivalent to
A · (xn + yn ) + B · xn = 0.
This gives (A, B) = (−xn , xn + yn ) ·C with C ∈ k[x, y]. The minimal possibilities for C are xn
and yn , giving the generators s1 := (−R, R + T n ) and s2 := (−T n , T n + S ) with the relation
T n · s1 − R · s2 = 0. The claim follows, since the syzygy module of (T n , R) is selfdual.
Now let n be odd. In this case we obtain
ϕ∗ (Mn+1 ) = ϕ∗ (SyzDn+2 (X + 2iY
n+1
2
, Z))
SyzA2n−1 (T R − T S + 2iT n+1 , R + S )
SyzA2n−1 (R − S + 2iT n , R + S )
SyzA2n−1 (R + iT n , R + S ).
Taking a syzygy (A, B) ∈ SyzA2n−1 (R + iT n , R + S ), the corresponding relation over k[x, y]
is equivalent to
A · xn + B · (xn − iyn ) = 0.
This gives (A, B) = (xn − iyn , −xn ) · C with C ∈ k[x, y]. The minimal possibilities for C are
again xn and yn , giving the generators s1 := (R − iT n , −R) and s2 := (T n − iS , −T n ) with the
relation T n · s1 − R · s2 = 0.
5.2.2. Pull-backs to Dn/2+2 . Consider the map
ψ : k[X, Y, Z]/(X 2 + 4Y n+1 − YZ 2 ) → k[R, S , T ]/(R2 + 4S 2 +1 − S T 2 ),
n
given by X 7→ RT , Y 7→ S , Z 7→ T 2 − 2S n/2 , where R = xy(xn − yn ), S = x2 y2 , T = xn + yn .
For all pull-backs the steps 1 and 2 from the beginning of this chapter are enough to detect
the isomorphism classes of the pull-backs.
ψ∗ (M1 ) SyzD n +2 (RT, S )
2
SyzD n +2 (R, S ),
2
n
ψ (M2m ) SyzD n +2 RT, S m , T 2 − 2S 2
∗
2
SyzD n +2 (RT, S m , T 2 )
2
SyzD n +2 (R, S m , T ),
2
ψ (M2m−1 ) SyzD n +2 RT, S m , S T 2 − 2S 2 +1
n
∗
2
SyzD n +2 (RT, S m , S T 2 )
2
SyzD n +2 (R, S m , S T ),
2
n
ψ (Mn+1 ) SyzD n +2 RT, T 2 − 4S 2
∗
2
n
SyzD n +2 R, T 2 − 4S 2
2
SyzD n +2 (R, S ),
2
5.3. THE CASE E6
129
ψ∗ (Mn+2 ) SyzD n +2 (RT, T 2 )
2
SyzD n +2 (R, T )
2
OD n +2 .
2
5.3. The Case E6
The binary tetrahedral group T of order 24 is generated by σ, τ and µ, where
Ç
å
i 0
,
σ=
0 −i
Ç
å
0 1
,
τ=
−1 0
Ç
1 ε7 ε7
µ= √
5
2 ε ε
å
with a primitive eighth root of unity ε. The generating invariants are
X := x12 − 33x8 y4 − 33x4 y8 + y12 , Y := x8 + 14x4 y4 + y8 , Z := x5 y − xy5 ,
hence the equation of the E6 -singularity is
X 2 − Y 3 + 108X 4 = 0.
There are two normal subgroups of T. The first one is generated by σ and τ and the ring of
invariants is a D4 -singularity. The second one is generated by τ2 and the ring of invariants
is an A1 -singularity. Recall from Theorem 3.40 that
M1 = SyzE6 (X, Y, Z),
M2 = SyzE6 (X, Y 2 , YZ, Z 2 ),
√
M3 = SyzR (iX + 6 3Z 2 , Y 2 , YZ),
√
M4 = SyzR (−iX + 6 3Z 2 , Y 2 , YZ),
√
M5 = SyzR (−iX + 6 3Z 2 , Y),
√
M6 = SyzR (iX + 6 3Z 2 , Y)
is a complete list of representatives of the isomorphism classes of indecomposable, nonfree, maximal Cohen-Macaulay modules.
We consider the map
ψ : k[X, Y, Z]/(X 2 − Y 3 + 108Z 4 ) −→ k[U, V, W]/(UV − W 2 ),
that is given by
X 7→ U 6 − 33U 2 W 4 − 33V 2 W 4 + V 6 ,
Y 7→ U 4 + V 4 + 14W 4 ,
A direct computation shows that ψ is the composition
(A3 → A1 ) ◦ (D4 → A3 ) ◦ (E6 → D4 ).
Therefore, it is enough to compute the pull-backs to D4 .
Z 7→ (U 2 − V 2 )W.
130
5. RELATIONS BETWEEN THE MAXIMAL COHEN-MACAULAY MODULES
To reduce the number of generators of the syzygy modules corresponding to the pullbacks of Mi , i ≥ 3, we first prove the isomorphisms
M5 ⊕ M6 L := SyzE6 (X, Y, Z 2 )
and
M3 ⊕ M4 N := SyzE6 (XY, XZ, Y 2 , YZ 2 , Z 3 )
SyzE6 (XY, XZ 2 , Y 3 , Y 2 Z, Z 3 )
(for later use).
The proofs use matrix factorizations. To avoid ugly constant coefficients in the matrices,
we turn back to the equation X 2 + Y 3 + Z 4 = 0 for these√proofs. Note that under this
isomorphism, we only have to change the coefficients 6 3 in M3 , . . . , M6 into 1. The
matrix factorization
ÇÇÇ
−Z 2 Y
Y 2 Z2
åå ÇÇ
−Z 2 Y
Y 2 Z2
,
ååå
ˆ X)
⊗(X,
splits by Lemma 3.33 as (ϕ5 , ψ5 ) ⊕ (ϕ6 , ψ6 ). Obviously, the module given by this matrix
factorization is selfdual. Computing the tensor product and deleting the first column (in
the second
matrix), one
obtains the
claimed isomorphism.
XE4 η
XE4 η
Let Ω =
with
η −XE4 ,
η −XE4
á
−Z 2
0
YZ Y
2
−YZ Z
Y2 0
0
Y
0 Z
Y 2 −YZ Z 3 0
η=
Because the map
Ä
E8 ,
Ä 0 E ää
4
E4 0
ë
.
ˆ X) → Ω
: (η, η)⊗(X,
is an isomorphism of matrix factorizations, we see by Lemma 3.33 that Ω splits as
(ϕ3 , ψ3 ) ⊕ (ϕ4 , ψ4 ). Again coker(Ω) is selfdual and deleting the columns one, three and
seven from the second matrix of Ω, the rows of the reduced matrix generate the module
SyzE6 (−XZ, XY, Y 2 , −Z 3 , YZ 2 ). For the other isomorphism delete the columns two, three
and five.
Now, we compute the pull-backs to D4 . The computations are done with Macaulay2 (with
coefficients in Z) and can be found in Appendix B.2. The map
ϕ : k[X, Y, Z]/(X 2 − Y 3 + 108Z 4 ) −→ k[U, V, W](U 2 + 4V 3 − VW 2 )
is given by
X 7→ W 3 − 36V 2 W,
Y 7→ W 2 + 12V 2 ,
Z 7→ U.
Therefore, the pull-back
ϕ∗ (SyzE6 (X, Y, Z)) SyzD4 (W 3 − 36V 2 W, W 2 + 12V 2 , U)
is generated by
Ö
s1 :=
0
−U
12V 2 + W 2
è
Ö
, s2 :=
−V
−2VW
3UW
è
Ö
, s3 :=
W
2
12V − W 2
36UV
è
Ö
, s4 :=
−U
−3UW
4W 3
è
.
5.3. THE CASE E6
131
The total degrees of these generators are seven, eight, eight and nine. Since there are no
elements of degree one and two and s2 and s3 are linearly independent, all generators are
necessary. We have the following global relations among these generators
−3U · s1 + W · s2 + V · s3 = 0,
−4UW · s1 + 4V 2 · s2 + VW · s3 + U · s4 = 0,
(12V 2 − 4W 2 ) · s1 + U · s3 + W · s4 = 0.
The second and third relation show that we can ommit the fourth generator locally on the
covering D(U) ∪ D(W) = D4 . Then the first relation gives the isomorphism
ϕ∗ (SyzE6 (M1 )) SyzD4 (U, V, W).
The pull-back
ϕ∗ (M5 ⊕ M6 ) ϕ∗ (L) ϕ∗ (SyzE6 (X, Y, Z 2 )) SyzD4 (W 3 − 36V 2 W, W 2 + 12V 2 , U 2 )
is generated by
Ö
s1 :=
V
2VW
−3W
è
Ö
, s2 :=
W
2
−W + 12V 2
36V
Ö
è
, s3 :=
0
U2
−12V 2 − W 2
Ö
è
, s4 :=
U2
0
2
36V W − W 3
è
Because s1 and s2 are linearly independent and of the same (minimal) total degree, both
are necessary as generators. The relations
1
1
s3 = (W · s1 − V · s2 ) and s4 = ((W 2 − 12V 2 ) · s1 + 2VW · s2 )
3
3
show that s1 and s2 generate the whole module if the characteristic is not three, hence M5
and M6 become trivial after the pull-back.
To get the pull-back of M2 we compute
ϕ∗ (SyzE6 (X, Y 2 , YZ, Z 2 ))
SyzD4 (W 3 − 36V 2 W, W 4 + 24V 2 W 2 + 144V 4 , UW 2 + 12UV 2 , U 2 ).
A set of generators for SyzD4 (W 3 − 36V 2 W, W 4 + 24V 2 W 2 + 144V 4 , UW 2 + 12UV 2 , U 2 ) is
given by
s1 := (0, 0, −U, 12V 2 + W 2 ),
s2 := (0, −U, 12V 2 + W 2 , 0),
s3 := (2VW, V, 3U, −6W 2 ),
s4 := (12V 2 − W 2 , W, 0, −72VW),
s5 := (UV, 0, 2VW, −3UW),
s6 := (UW, U, −2W 2 , 36UV),
s7 := (−U 2 , 0, −3UW, 4W 3 ).
Because we can write 3 · s7 = 6W · s1 − W · s2 + V · s3 , we can omit s7 as a generator (since
char(k) , 3). The remaining generators are linearly independent and have total degree ten
.
132
5. RELATIONS BETWEEN THE MAXIMAL COHEN-MACAULAY MODULES
or eleven. Since there are no elements of degree one, all of them are necessary to generate
the module. We have the relations
U · (−6s1 − s3 ) + V · (s2 + 2s6 ) = 0,
U · (6s1 − 2s3 + s4 ) + (W − 2V) · (−2s2 + 6s5 + s6 ) = 0,
U · (−6s1 + 2s3 + s4 ) + (W + 2V) · (2s2 − 6s5 + s6 ) = 0.
Since the characteristic of our base field is at least five, the module generated by s1 , . . . , s6
is isomorphic to the module generated by the elements in the brackets. We get
ϕ∗ (M2 ) SyzD4 (U, V) ⊕ SyzD4 (U, W − 2V) ⊕ SyzD4 (U, W + 2V).
Generators of
ϕ∗ (M3 ⊕ M4 ) ϕ∗ (N) = ϕ∗ (SyzE6 (XY, XZ, Y 2 , YZ 2 , Z 3 ))
are given by
s1 := (0, 0, 0, −U, W 2 + 12V 2 ),
s2 := (−U, 12V 2 + W 2 , 0, 0, 0),
s3 := (V, 0, 2VW, −3W, 0),
s4 := (W, 0, −W 2 + 12V 2 , 36V, 0),
s5 := (0, 0, U 2 , −12V 2 − W 2 , 0),
s6 := (0, 2VW, UV, 3U, −6W 2 ),
s7 := (U, −2W 2 , UW, 0, −72VW),
s8 := (0, UV, 0, 2VW, −3UW),
s9 := (0, UW, 0, 12V 2 − W 2 , 36UV),
s10 := (0, −U 2 , 0, −3UW, 4W 3 ),
s11 := (−U 2 W, 3UW 3 , 0, −2W 4 + 27U 2 V, 81U 3 ).
The equations
3 · s5 = −W · s3 + V · s4 ,
3 · s10 = 6W · s1 + V · (s2 + s7 ) − W · s7 ,
s11 = −27UV · s1 + UW · s2 − 12VW · s8 + 2W 2 · s9
show that s5 , s10 and s11 can be omitted (since char(k) , 3).
The remaining generators are linearly independent and have total degrees twelve, thirteen
or fourteen. The only possibility to have one generator to much, is an equation of the form
a14 = V · b12 + W · c12 , where a14 , b12 and c12 are linear combinations of the generators of
total degree given by the index. It’s easy to verify that there is no such equation.
5.4. THE CASE E7
133
Moreover, if char(k) ≥ 5, we can omit s8 and s9 on the covering D(U) ∪ D(W) as the
following relations show
3U · s8 = 3VW · s1 − V 2 · s2 − UV · s3 + 2VW · s6 ,
W · s8 = −3U · s1 + V · s9 ,
6U · s9 = (18W 2 − 72V 2 ) · s1 − UW · s3 + UV · s4 + (3W 2 − 12V 2 ) · s6 ,
6W · s9 = (2W 2 − 12V 2 ) · s3 + VW · s6 − 3U · s7 .
Finally, we have
W · (6s1 + s6 ) − U · s3 + V · s7 = 0,
W · (2S 2 + s6 ) + U · s4 − 12V · s6 = 0,
showing
ϕ∗ (M3 ⊕ M4 ) SyzD4 (U, V, W)2 .
5.4. The Case E7
The binary octahedral group O of order 48 is generated by σ, τ, µ and κ, where
Ç
å
i 0
σ=
,
0 −i
Ç
å
0 1
τ=
,
−1 0
Ç
å
1 ε7 ε7
,
µ= √
5
2 ε ε
Ç
å
ε 0
κ=
,
0 ε7
where ε is again a primitive eighth root of unity. The generating invariants are
X := x17 y − 34x13 y5 + 34x5 y13 − xy17 , Y := x10 y2 − 2x6 y6 + x2 y10 , Z := x8 + 14x4 y4 + y8 .
The equation of an E7 -singularity is given by
X 2 + 108Y 3 − YZ 3 = 0.
There are three normal subgroups of O. The first one is generated by σ and τ, which
is isomorphic to D2 and has D4 as ring of invariants. The second normal subgroup is
T =< σ, τ, µ >, since κ2 = σ. The corresponding ring of invariants is an E6 -singularity.
The third normal subgroup is generated by τ2 with ring of invariants A1 .
Recall from Theorem 3.41 that there are seven classes of non-free, indecomposable, maximal Cohen-Macaulay modules and represantatives are given by
M1 = SyzE7 (X, Y, Z),
M2 = SyzE7 (X, Y 2 , YZ, Z 2 ),
M3 = SyzE7 (XY, XZ, Y 2 , YZ 2 , Z 3 ),
M4 = SyzE7 (X, Y 2 , YZ),
M5 = SyzE7 (XY, XZ, Y 2 , YZ 2 ),
M6 = SyzE7 (X, Y, Z 2 ),
M7 = SyzE7 (X, Y).
134
5. RELATIONS BETWEEN THE MAXIMAL COHEN-MACAULAY MODULES
Note that we only need to compute their pull-backs to E6 , since the maps D4 → E7 and
A1 → E7 factor through E6 . The map
η : k[X, Y, Z]/(X 2 + 108Y 3 − YZ 3 ) −→ k[U, V, W]/(U 2 − V 3 + 108W 4 )
is given by X 7→ UW, Y 7→ W 2 , Z 7→ V. With l ∈ {1, 2} we get
η∗ (SyzE7 (X, Y, Z l )) SyzE6 (UW, W 2 , V l )
SyzE6 (U, W, V l )
SyzE6 (U, V, W),
showing η∗ (M1 ) η∗ (M6 ) SyzE6 (U, V, W). Similarly, one has
η∗ (SyzE7 (M4 )) SyzE6 (UW, W 4 , VW 2 )
SyzE6 (U, W 3 , VW)
SyzE6 (U, W 2 , V)
√
√
SyzE6 (iU + 6 3W 2 , V) ⊕ SyzE6 (−iU + 6 3W 2 , V),
where the last isomorphism is one of those, we have computed at the beginning of this
section.
The pull-back η∗ (SyzE7 (X, Y)) SyzE6 (U, W) is free.
We have the isomorphisms
η∗ (M2 ) = η∗ (SyzE7 (X, Y 2 , YZ, Z 2 ))
SyzE6 (UW, W 4 , VW 2 , V 2 )
SyzE6 (U, V 2 , VW, W 3 )
SyzE6 (U, V 2 , VW, W 2 ) and
η∗ (M5 ) = η∗ (SyzE7 (XY, XZ, Y 2 , YZ 2 ))
SyzE6 (UW 3 , UVW, W 4 , V 2 W 2 )
SyzE6 (UW 2 , UV, W 3 , V 2 W)
SyzE6 (UW, UV, W 2 , V 2 )
SyzE6 (U, V 2 , VW, W 2 ).
The last module to consider is M3 . The pull-back splits by the following computation
η∗ (SyzE7 (XY, XZ, Y 2 , YZ 2 , Z 3 ))
SyzE6 (UW 3 , UVW, W 4 , V 2 W 2 , V 3 )
SyzE6 (UW 2 , UV, W 3 , V 2 W, V 3 )
√
√
SyzE6 (iU + 6 3W 2 , V 2 , VW) ⊕ SyzE6 (−iU + 6 3W 2 , V 2 , VW),
where the last isomorphism is one of those, we have seen at the beginning of this section.
5.5. THE CASE E8
135
5.5. The Case E8
The binary icosahedral group I of order 120 has two generators ϕ and ψ with
Ç
ϕ=−
ε3 0
0 ε2
å
and
1
ψ= √
5
Ç
−ε + ε4 ε2 − ε3
ε2 − ε3 ε − ε4
å
,
where ε is a primitive fifth root of unity. The generating invariants are
X := x30 + 522x25 y5 − 10005x20 y10 − 10005x10 y20 − 522x5 y25 + y30 ,
Y := x20 − 228x15 y5 + 494x10 y10 + 228x5 y15 + y20 ,
Z := x11 y + 11x6 y6 − xy11 .
An E8 -singularity is given by the equation
X 2 − Y 3 − 1728Z 5 = 0.
The only normal subgroup of I is Z/2, generated by ϕ5 . The corresponding map
ι : k[X, Y, Z]/(X 2 − Y 3 − 1728Z 5 ) → k[R, S , T ]/(RS − T 2 )
is given by
X 7→ R15 + 522R10 T 5 − 10005R5 T 10 − 10005S 5 T 10 − 522S 10 T 5 + S 15 ,
Y 7→ R10 − 228R5 T 5 + 494T 10 + 228S 5 T 5 + S 10 ,
Z 7→ R5 T − S 5 T + 11T 6 .
There are eight isomorphism classes of non-free, indecomposable, maximal Cohen-Macaulay modules, represented by (cf. Theorem 3.42)
M1 = SyzE8 (X, Y, Z),
M2 = SyzE8 (X, Y 2 , YZ, Z 2 ),
M3 = SyzE8 (XY, XZ, Y 2 , YZ 2 , Z 3 ),
M4 = SyzE8 (XY, XZ 2 , Y 3 , Y 2 Z, YZ 3 , Z 4 ),
M5 = SyzE8 (XY 2 , XYZ 2 , XZ 4 , Y 4 , Y 3 Z, Y 2 Z 3 , Z 5 ),
M6 = SyzE8 (X, Y 2 , YZ, Z 3 ),
M7 = SyzE8 (XY, XZ, Y 2 , YZ 2 , Z 4 ),
M8 = SyzE8 (X, Y, Z 2 ).
Since there are only the indecomposable, maximal Cohen-Macaulay modules A1 and
SyzA1 (R, T ) on A1 , we only need the minimal numbers of generators of ι∗ (SyzE8 (Mi ))
to understand its splitting behaviour. Explicitely, the module A1 is one-generated, while
SyzA1 (R, T ) is two-generated. Now, a maximal Cohen-Macaulay A1 -module of rank r,
that is minimally generated by µ elements, has exactly µ − r direct summands isomorphic to SyzA1 (R, T ) and 2r − µ free direct summands. Therefore, we only explain the
Macaulay2 computation, given in Appendix B.3.
136
5. RELATIONS BETWEEN THE MAXIMAL COHEN-MACAULAY MODULES
First consider the ring A := Z[R, S , T ]/(RS − T 2 ). Since the syzygy modules, we are
interested in, have far to many generators, when working over A, we also work over
B := Q[R, S , T ]/(RS − T 2 ). The Macaulay2 command for the canonical inclusion A ,→ B
is map(B,A) and for a matrix M over A the command map(B,A)**M treats M as a
matrix over B. We denote the A-matrix, whose columns generate ι∗ (SyzE8 (Mi )) with
coefficients in Z by Si and the corresponding B-matrix by Ti (cf. B.3 up to i21).
Using mingens image Ti, we produce matrices M1, . . . , M8, such that the columns of
Mi and Ti generate the same module, but the Mi have less columns than the Ti (cf. B.3
i22-i29).
In B.3 i30-i37 we compute the canonical map from image Ti to image Mi. The
outputs are matrices, whose j − th column contains the coefficients necessary to write the
j − th column of Ti in terms of the columns of Mi. Working in positive characteristic,
we may use the Mi instead of the Ti, whenever all the coefficients are well-defined. We
always have to invert two and three, which is no problem, since our characteristic has
to be at least seven to write down the generators of the group I. In the case of T3 we
additionally need to invert five and seven and for T5 we have to invert five and 59. The
computations with coefficients in Z/(p), p ∈ {7, 59} are done in B.3 i38-i43.
In all cases we end up with a matrix Mi, whose columns generate ι∗ (SyzE8 (Mi )). For
each i these generators are linearly independent and have the same total degree, hence
they give a minimal system of generators.
The final results are gathered together in Table 5.5.
5.6. Overview
We end this section with an overview. The modules in the columns are modules over the
scheme in the head-spot of the column. The rows tell us to which module the module in
the first column pulls back.
Dn+2
M1
M2m
M2m−1
Mn+1 , n even
Mn+2 , n even
Mn+1 , n odd
Mn+2 , n odd
A2n−1
At−1 , t|2n
Dn/2+2 , (n even)
OA2n−1
OAt−1
M1
M2m−1 ⊕ M2(n−m)+1 Mr ⊕ Mt−r , r = 2m − 1 mod t
M2m
M2m−2 ⊕ M2(n−m)+2 Mr ⊕ Mt−r , r = 2m − 2 mod t
M2m−1
Mn
Mr , r = n mod t
M1
Mn
Mr , r = n mod t
ODn/2+2
Mn
Mr , r = n mod t
−
Mn
Mr , r = n mod t
−
Table 5.2. The table shows the pull-back behaviour of the Mi on Dn+2 .
In the third column of Table 5.2 the number r is the minimal non-negative integer satisfying the modulo condition.
5.6. OVERVIEW
E6
M1
M2
M3
M4
M5
M6
137
D4
A1
M2
M21
M1 ⊕ M3 ⊕ M4 O3A1
M2
M21
M2
M21
OD4
OA1
OD4
OA1
Table 5.3. The table shows the pull-back behaviour of the Mi on E6 .
E7
M1
M2
M3
M4
M5
M6
M7
E6
D4
M1
M2
M2
M1 ⊕ M3 ⊕ M4
M3 ⊕ M4
M22
M5 ⊕ M6
D24
M2
M1 ⊕ M3 ⊕ M4
M1
M2
E6
D4
A1
M21
O3A1
M41
O2A1
O3A1
M21
OA1
Table 5.4. The table shows the pull-back behaviour of the Mi on E7 .
E8
M1
M2
M3
M4
M5
M6
M7
M8
A1
M21
O3A1
M41
O5A1
M61
O3A1
O4A1
M21
Table 5.5. The table shows the pull-back behaviour of the Mi on E8 .
Using these direct computations, we can show that the Frobenius pull-backs of the syzygy
modules of the maximal ideals stay indecomposable in almost all characteristics if the
singularity is of type D or E.
Lemma 5.4. Let U be the punctured spectrum of a surface ring R of type D or E and let
F : U → U be the (restricted) Frobenius morphism. Assume that the order of the group
corresponding to the singularity is invertible in the ground field. Then Fe∗ (SyzU (m)) is
indecomposable for all e ∈ N.
Proof. Recall from Chapter 3 that SyzU (m) is isomorphic to M2 if R is of type D and
M1 if R is of type E. Assume first that R is of type D2n+2 or E. If Fe∗ (SyzU (m)) splits
138
5. RELATIONS BETWEEN THE MAXIMAL COHEN-MACAULAY MODULES
for some e ≥ 1, its pull-back to A1 (along the morphism ι, induced by the inclusion into
A1 ) gets trivial, since all line bundles pull back to the structure sheaf by Tables 5.2 - 5.5.
But the pull-back of SyzU (m) along ι is non-trivial (in fact, it is always M21 by the direct
computations) and its Frobenius pull-backs stay non-trivial in odd characteristics. This
gives a contradiction, since the two types of pull-backs commute.
Assume that R is of type Dn+2 with n odd. Since Fe∗ (SyzU (m)) might split as Mn+1 ⊕
Mn+2 and both pull back to M1 on A1 , the argument above gives no contradiction. In
this case, we have to compare the pull-backs to An−1 . Then all line bundles from Dn+2
pull back to the structure sheaf of An−1 , but ι∗ (M2 ) M1 ⊕ Mn−1 . The Frobenius pullbacks on the right hand side do not get trivial, since n is invertible modulo p.
Moreover, if R is of type D, the Frobenius pull-backs of M2 are again of the form M2m ,
since the M2m−1 get trivial after pulling them back to A1 .
By the previous lemma all Frobenius pull-backs of SyzU (m) are again indecomposable of
rank two. Since in all cases det(Fe∗ (SyzU (m))) OU , we obtain directly Fe∗ (SyzU (m))
SyzU (m) if R is of type E6 , since in this case there is only one rank two module, whose
representation is a syzygy module of an m-primary ideal. In the next chapter we will see
that Lemma 5.4 leads always to a (0, e0 )-Frobenius periodicity Fe0 ∗ (SyzU (m)) SyzU (m).
For example if R is of type E7 or E8 , the length e0 of this periodicity is - depending on
the characteristic - one or two.
CHAPTER 6
The Hilbert-Kunz functions of surface rings of type ADE
In this chapter we compute the Hilbert-Kunz functions of surface rings of type ADE. We
recall very briefly the situation. Let R := k[U, V, W]/( f ) be a surface ring of type ADE
with deg(U) = α, deg(V) = β, deg(W) = γ and let S := k[X, Y, Z]/(F) be the associated
standard-graded ring. Denote their projective spectra by CR resp. CS and their punctured
spectra by R resp. S. Then by Remark 1.22 the Hilbert-Kunz function of R is the HilbertKunz function of S with respect to the ideal (X α , Y β , Z γ ) divided by α · β · γ, since the map
R → S , sending U, V and W to X α , Y β resp. Z γ , is flat by [Mat05, Theorem 22.2] and the
extension Q(S ) : Q(R) has degree α · β · γ. In the standard-graded situation we may use the
geometric approach of Brenner/ Trivedi, which is possible, if we find for each q = pe an
isomorphism of the form
SyzS (X α·q , Y β·q , Z γ·q ) SyzS (X a , Y b , Z c )(n),
where the vector (a, b, c) and the degree shift n ∈ Z depend on q and at least one of the
inequalities α · q < a, β · q < b, γ · q < c holds. In each case we will show that there exists
a finite list of triples (a0 , b0 , c0 ) such that there are isomorphisms
Ä
0
0
0ä
SyzR (U q , V q , W q ) SyzR U a , V b , W c (m)
for some m ∈ Z and for all q. These isomorphisms induce isomorphisms in the standardgraded case of the desired form. After all we use the formula
HK(R, q)
(6.1)
0
0
0
d · Q(α, β, γ) · q2 − d · Q(a0 · α, b0 · β, c0 · γ) + 4 · dimk (S /(X a ·α , Y b ·β , Z c ·γ ))
=
,
4·α·β·γ
with d = deg( f ) and Q(r, s, t) = 2(rs + rt + st) − r2 − s2 − t2 from Lemma 2.2 (together with
Remark 1.22).
In small characteristics, it happens sometimes that the syzygy modules SyzR (U q , V q , W q )
split as R(−m + δ) ⊕ R(−m) for some m, δ ∈ Z. In those cases the Hilbert-Kunz function
can be computed by the formula (cf. Lemma 2.1 and Remark 1.22)
(6.2)
HK(R, q) =
d · Q(α, β, γ) · q2 + d · δ2
.
4·α·β·γ
6.1. The case An−1 with n ≥ 1
Let R := k[U, V, W]/(U n − VW) with deg(U) = 2, deg(V) = n, deg(W) = n and char(k) = p n. In Chapter 3 we saw that we may choose SyzR (U, V) as generator of the Picard group
of R. The e-th Frobenius pull-back is just the q-th tensor power, which is isomorphic to
139
140
6. THE HILBERT-KUNZ FUNCTIONS OF SURFACE RINGS OF TYPE ADE
the tensor power of order q modulo n, hence SyzR (U q , V q ) SyzR (U r , V) with r ≡ q mod
n, r ∈ {1, . . . , n − 1}, by Lemma 5.2. Using the isomorphism
SyzR (U r , V, W) SyzR (U r , V) ⊕ SyzR (U r , W) SyzR (U r , V) ⊕ SyzR (U n−r , V)∨ ,
we find
SyzR (U q , V q , W q ) SyzR (U r , V, W)(m)
for some m ∈ Z.
This yields with Formula (6.1)
2n · Q(2, n, n) · q2 − 2n · Q(2r, n, n) + 8rn2
8n2
(8n − 4)q2 − 8rn + 4r2
+r
=
4n
Ç
å
1 2
r2
= 2−
q −r+ .
n
n
HKR (R, q) =
Note that we can define all appearing syzygy modules in the case p|n as well. Since the
computations in Section 4.1 are correct in all characteristics, the above computation of
the Hilbert-Kunz function remains correct in the case p|n, if we would know that our list
of maximal Cohen-Macaulay modules is complete.
By a theorem of Herzog (cf. [LW12, Theorem 6.3]) we only need that R is a direct
summand of the polynomial ring k[u, v], on which the group Z/(n) acts, to have an one to
one correspondence between indecomposable, maximal Cohen-Macaulay R-modules and
direct summands of k[u, v] as an R-module.
Leuschke and Wiegand point out in [LW12, Remark 6.22] that R is a direct summand of
k[u, v] and that the direct summands of k[u, v] as an R-module are the same as in the nonmodular case (meaning that the algebraic definition is the same, although the modules in
the modular case might not come from representations of the group Z/n). Therefore, the
formula above is a full description of the Hilbert-Kunz function (compare with Example
1.5 or Remark 1.6).
By Theorem 1.91, the F-signature function of R is given by
1 2
r2
Fsign (R, q) = · q + r − .
n
n
6.2. The case Dn+2 with n ≥ 2
Let R := k[U, V, W]/(U 2 + V n+1 + VW 2 ) with deg(U) = n + 1, deg(V) = 2, deg(W) = n and
char(k) = p - 4n.
Recall from Chapter 5 that there is an embedding
ϕ : R −→ A2n−1 := k[X, Y, Z]/(X 2n − YZ)
6.2. THE CASE Dn+2 WITH n ≥ 2
141
with the following pull-back behaviour
R
M1
M2m
M2m−1
Mn+1
Mn+2
A2n−1
OA2n−1
N2m−1 ⊕ N2(n−m)+1
N2m−2 ⊕ N2(n−m)+2
Nn
Nn ,
where the Ni are the Mi on A2n−1 from Chapter 5. By Lemma 5.4 we have
Fe∗ (SyzR (m)) = Fe∗ (M2 ) M2m
(6.3)
for all e ∈ N and m = m(q) ∈ {1, . . . , n2 }. The parameter m = m(q) can be computed by considering the pull-back via ϕ: M2 pulls back to N1 ⊕ N2n−1 on A2n−1 . The e-th Frobenius
pull-back of this module is Nr ⊕ N2n−r with r ≡ q mod 2n (and 1 ≤ r ≤ 2n − 1). If r ≤ n − 1,
only Mr+1 pulls back to Nr ⊕ N2n−r via ϕ. If r ≥ n, only M2n−r−1 pulls back to Nr ⊕ N2n−r
via ϕ. But in Chapter 3 we saw that Mr+1 and M2n−r−1 are isomorphic (on R). We find
SyzR (U q , V q , W q ) SyzR (U, V (r+1)/2 , W),
where r ≡ q mod 2n with 1 ≤ r ≤ 2n − 1. This isomorphism induces a global isomorphism
SyzR (U q , V q , W q ) SyzR (U, V (r+1)/2 , W)(m),
for some m ∈ Z. Using Formula (6.1), we get
(2n + 2) · Q(n + 1, 2, n) · q2 − (2n + 2) · Q(n + 1, r + 1, n) + 4(r + 1)(n2 + n)
8(n2 + n)
(8n − 1)q2 − 4nr − 4n + r2 r + 1
=
+
4n
2
Ç
å
2
1
r+1 r
= 2−
q2 −
+ .
4n
2
4n
HK(R, q) =
Dealing with characteristics dividing 4n, the computations in Section 4.2 (and 4.1) still
hold if p is odd. By a similar argument as in the previous section (just change the groups,
the references are the same) the indecomposable, maximal Cohen-Macaulay modules (defined algebraically) are the same in the modular and the non-modular case. To get a full
description of the Hilbert-Kunz function, we only need to take a closer look at the case
p = 2. In this case the ring Dn+2 in Section 4.2 is not a Dn+2 -singularity, since it is a
binomial hypersurface.
142
6. THE HILBERT-KUNZ FUNCTIONS OF SURFACE RINGS OF TYPE ADE
Let q = 2e with e ≥ 1. Using Lemma 1.61, we obtain
q
SyzR (U q , V q , W q ) SyzR ((V n+1 + VW 2 ) 2 , V q , W q )
q
q
SyzR (V (n+1)· 2 + V 2 W q , V q , W q )
q
SyzR (V 2 W q , V q , W q )
q
SyzR (V 2 , V q , 1)(−nq)
q
SyzR (1, V 2 , 1)(−(n + 1)q)
R(−(n + 1)q) ⊕ R(−(n + 2)q).
Using Formula (6.2), we get
Q(n + 1, 2, n) · 4e + 4e
= 2 · 4e .
HK(R, 2 ) =
4n
e
Theorem 6.1. The Hilbert-Kunz function of Dn+2 is given by (with n ≥ 2, q = pe and r ≡ q
mod 2n for 0 ≤ r ≤ 2n − 1)
2
2·q
e 7−→ Ç
2−
å
if e ≥ 1, p = 2,
1
r + 1 r2
+
otherwise.
q2 −
4n
2
4n
Using Theorem 1.91, one obtains
Corollary 6.2. The F-signature function of R is given by
if e ≥ 1, p = 2,
0
Fsign (R, q) = 1 2 r + 1 r2
·q +
−
otherwise.
4n
2
4n
6.3. The case E6
Let R := k[U, V, W]/(U 2 +V 3 +W 4 ) with deg(U) = 6, deg(V) = 4, deg(W) = 3 and char(k) =
p ≥ 5. The Frobenius pull-backs of SyzR (U, V, W) are indecomposable by Lemma 5.4.
Since they have rank two and trivial determinant, we obtain a (0, 1)-Frobenius periodicity,
extending globally to
Ç
å
Ç
å
13
SyzR (U , V , W ) SyzR (U, V, W) − · (p − 1) .
2
p
p
p
Iterating yields the isomorphism
13
SyzR (U , V , W ) SyzR (U, V, W) − · (q − 1)
2
q
q
q
6.4. THE CASE E7
143
for all q and we get (with Formula (6.1))
12 · Q(6, 4, 3) · q2 − 12 · Q(6, 4, 3) + 12 · 24
12 · 24
2
47 · q − 47 + 24
=
24å
Ç
1
23
= 2−
q2 − .
24
24
HK(R, q) =
It remains to compute the Hilbert-Kunz function in characteristics two and three. Let
p = 2. For q ≥ 2 we have by using Lemma 1.61
q
SyzR (U q , V q , W q ) = SyzR V q , W q , (V 3 + W 4 ) 2
= SyzR (V q , W q , V
3q
2
+ W 2q )
SyzR (V q , W q , W 2q )
SyzR (V q , 1, W q )(−3q)
R(−6q) ⊕ R(−7q).
Using Formula (6.2), we get
HK(R, q) =
12 · Q(6, 4, 3) · q2 + 12 · q2
= 2 · q2 .
4·6·4·3
The analogue computation for p = 3 and q ≥ 3 gives HK(R, q) = 2 · q2 .
Theorem 6.3. The Hilbert-Kunz function of E6 is given by
2
2·q
Ç
e 7−→
2−
å
if e ≥ 1 and p ∈ {2, 3},
1
23
otherwise.
q2 −
24
24
From this we get by Theorem 1.91 the F-signature function.
Corollary 6.4. The F-signature function of R is given by
0
if e ≥ 1 and p ∈ {2, 3},
Fsign (R, q) = 1 2 23
·q +
otherwise.
24
24
6.4. The case E7
Let R := k[U, V, W]/(U 2 + V 3 + VW 3 ) with deg(U) = 9, deg(V) = 6, deg(W) = 4 and let
char(k) = p ≥ 5. We get from Lemma 5.4 that the Frobenius pull-backs of SyzR (U, V, W)
stay indecomposable and have to be isomorphic to either SyzR (U, V, W) or SyzR (U, V, W 2 ).
Recall from Chapter 4 that the Hilbert-series can distinguish these two classes. The
144
6. THE HILBERT-KUNZ FUNCTIONS OF SURFACE RINGS OF TYPE ADE
Hilbert-series of the corresponding R-modules are given by
t10 + t13 + t15 + t18
HSyzR (U,V,W) (t) =
(1 − t6 )(1 − t4 )
t14 + t15 + t17 + t18
HSyzR (U,V,W 2 ) (t) =
.
(1 − t6 )(1 − t4 )
With the help of Theorem 4.7 and Han’s Theorem 2.17 we compute the Hilbert-series
of SyzR (U p , V p , W p ) and SyzR (U p , V p , W 2p ). In the first case we need to compute the
degrees of the generators of
Å
SyzR V p , W p , (V 3 + VW 3 )
p−1
2
Å
ã
SyzR V p , W p , (V 3 + VW 3 )
and
p+1
2
ã
.
We have the isomorphisms
Å
SyzR V p , W p , (V 3 + VW 3 )
p−1
2
SyzR V p , W p , (V 3 + VW 3 )
p+1
2
Å
ã
ã
Å
SyzR V
Å
SyzR V
p+1
2
, W p , (V 2 + W 3 )
p−1
2
p−1
2
, W p , (V 2 + W 3 )
p+1
2
ã
(−3(p − 1)) and
ã
(−3(p + 1)).
Their syzygy gaps are given by
Ç
p+1 p p−1
12 · δ
, ,
4 3 2
(6.4)
å
Ç
resp.
å
p−1 p p+1
12 · δ
, ,
.
4 3 2
Similarly, in the second case the syzygy gaps of the corresponding S j from Theorem 4.7
are given by
Ç
p + 1 2p p − 1
12 · δ
, ,
4
3
2
(6.5)
å
Ç
resp.
å
p − 1 2p p + 1
, ,
12 · δ
.
4
3
2
The various arguments of δ do not satisfy the strict triangle inequality if and only if p ≤ 9
p p+1
and the arguments are p−1
4 , 3 , 2 . In this case we have 12 · δ = 9 − p. In all other cases
the s in Han’s Theorem has to be non-negative. For s = 0, p = 12l + r, r ∈ {1, 5, 7, 11} and
p
3 as second argument, the values of 12 · δ are given in Table 6.1.
r +, l even +, l odd −, l even −, l odd
1
2
2
8
4
5
2
2
4
8
7
8
4
2
2
11
4
8
2
2.
Table 6.1. The table shows the values of (6.4) for all primes p ≥ 5 depending on their decomposition as p = 12l + r. The sign in the headline is the
same as in the first argument of (6.4).
With
2p
3
as second argument, the values of 12 · δ are given in Table 6.2.
6.4. THE CASE E7
145
r +, l even +, l odd −, l even −, l odd
1
2
2
4
8
2
2
8
4
5
7
4
8
2
2
8
4
2
2.
11
Table 6.2. The table shows the values of (6.5) for all primes p ≥ 5 depending on their decomposition as p = 12l + r. The sign in the headline is the
same as in the first argument of (6.5).
Computing the Hilbert-series, we find
p−1
t19· 2
· HSyzR (U,V,W) (t)
HSyzR (U p ,V p ,W p ) (t) =
19· p−1
2 −2
HSyzR (U p ,V p ,W 2p ) (t) =
23· p−1
2 +2
t
p−1
t23· 2
t
if p mod 24 ∈ {±1, ±7},
· HSyzR (U,V,W 2 ) (t) if p mod 24 ∈ {±5, ±11},
· HSyzR (U,V,W 2 ) (t)
if p mod 24 ∈ {±1, ±7},
· HSyzR (U,V,W) (t) if p mod 24 ∈ {±5, ±11}.
For the higher pull-backs we get
Ç
å
19(q − 1)
SyzR (U, V, W) −
2
q q
q
å
Ç
SyzR (U , V , W )
19(q − 1)
2
Syz (U, V, W ) −
+2
R
2
if q mod 24 ∈ {±1, ±7},
if q mod 24 ∈ {±5, ±11}.
Therefore, the Hilbert-Kunz function is given by (use Formula (6.1))
18 · Q(9, 6, 4) · q2 − 18 · Q(9, 6, 4) + 18 · 48
18 · 48
2
95 · q − 95 + 48
=
48å
Ç
1
47
= 2−
q2 − ,
48
48
HK(R, q) =
if q mod 24 ∈ {±1, ±7} and (again with Formula (6.1))
18 · Q(9, 6, 4) · q2 − 18 · Q(9, 6, 8) + 18 · 96
18 · 48
2
95 · q − 167 + 96
=
48
Ç
å
1
71
= 2−
q2 − ,
48
48
HK(R, q) =
if q mod 24 ∈ {±5, ±11}.
146
6. THE HILBERT-KUNZ FUNCTIONS OF SURFACE RINGS OF TYPE ADE
Now let p = 2 and q ≥ 2. Then (by Lemma 1.61)
Å
Ä
SyzR (U , V , W ) SyzR V , W , V + VW
q
q
q
Å
q
q
SyzR V , W , V
q
q
Å
3
3q
2
q
2
Å
q
q
q
äq ã
+V W
SyzR V q , W q , V 2 W
SyzR V 2 , W q , W
3
3q
2
3q
2
2
3q
2
ã
ã
ã
(−3q)
q
SyzR V 2 , 1, W 2 (−7q)
R(−9q) ⊕ R(−10q).
The corresponding value of the Hilbert-Kunz function is 2 · q2 by Formula (6.2).
For p = 3 and q ≥ 3 we have (with Lemma 1.61)
q
SyzR (U q , V q , W q ) SyzR (U q , (U 2 + VW 3 ) 3 , W q )
SyzR (U q , U
2q
3
q
+ V 3 W q, W q)
2q
SyzR (U q , U 3 , W q )
q
SyzR (U 3 , 1, W q )(−6q)
R(−9q) ⊕ R(−10q).
Again we get HK(R, q) = 2 · q2 .
Theorem 6.5. The Hilbert-Kunz function of E7 is given by
2 · q2
Ç
å
if e ≥ 1 and p ∈ {2, 3},
1
2 47
q
−
if q mod 24 ∈ {±1, ±7},
e 7−→
48
48
Ç
å
71
1
q2 −
otherwise.
2−
48
48
2−
Reformulating the cases in the above theorem in terms of p, one obtains the function
2
2·q
Ç
å
if e ≥ 1 and p ∈ {2, 3},
1
2 71
e 7−→ 2 − 48 q − 48 if p mod 24 ∈ {±5, ±11} and e is odd,
Ç
å
47
1
q2 −
otherwise.
2−
48
48
By using Theorem 1.91, we can compute the F-signature function.
6.5. THE CASE E8
147
Corollary 6.6. The F-signature function of R is given by
0
1
if e ≥ 1 and p ∈ {2, 3},
2 47
Fsign (R, q) = 48 · q + 48 if q mod 24 ∈ {±1, ±7},
1 2 71
·q +
otherwise.
48
48
6.5. The case E8
Let R := k[U, V, W]/(U 2 + V 3 + W 5 ) with deg(U) = 15, deg(V) = 10, deg(W) = 6 and
let char(k) = p ≥ 7. By Lemma 5.4 the Frobenius pull-back of SyzR (U, V, W) is either
SyzR (U, V, W) or SyzR (U, V, W 2 ). Moreover, their classes can be distinguished by the
Hilbert-series. We have
t16 + t21 + t25 + t30
(1 − t10 )(1 − t6 )
t22 + t25 + t27 + t30
HSyzR (U,V,W 2 ) (t) =
.
(1 − t10 )(1 − t6 )
HSyzR (U,V,W) (t) =
We compute the Hilbert-series of SyzR (U p , V p , W p ) and SyzR (U p , V p , W 2p ), using Theorem 4.7 and Han’s Theorem 2.17.
In the first case we need to compute
Ç
p p p∓1
30 · δ
, ,
3 5 2
(6.6)
å
and in the second case we are interested in the numbers
å
Ç
p 2p p ∓ 1
, ,
.
(6.7)
30 · δ
3 5
2
Note that for primes at most 13 the numbers 3p , 5p and p+1
2 do not satisfy the strict triangle
inequality, in these cases we get 30 · δ = 15 − p. In all other cases the s in Han’s Theorem
is non-negative. With s = 0, p = 30l + r and r ∈ {1, 7, 11, 13, 17, 19, 23, 29} one gets for
30 · δ with 5p as second argument the values shown in Table 6.3 and with 2p
5 as second
argument the values are given in Table 6.4.
Computing the Hilbert-series, we find
p−1
t31· 2
HSyzR (U p ,V p ,W p ) (t) =
t
p−1
t37· 2
HSyzR (U p ,V p ,W 2p ) (t) =
· HSyzR (U,V,W) (t)
31· p−1
2 −3
if p mod 30 ∈ {±1, ±11},
· HSyzR (U,V,W 2 ) (t) if p mod 30 ∈ {±7, ±13},
· HSyzR (U,V,W 2 ) (t)
t37· p−1
2 +3 · H
SyzR (U,V,W) (t)
if p mod 30 ∈ {±1, ±11},
if p mod 30 ∈ {±7, ±13}.
148
6. THE HILBERT-KUNZ FUNCTIONS OF SURFACE RINGS OF TYPE ADE
r −, l odd −, l even +, l odd +, l even
1
14
4
4
14
8
2
2
8
7
11
4
14
14
4
2
8
8
2
13
17
2
8
8
2
4
14
14
4
19
23
8
2
2
8
14
4
4
14
29
Table 6.3. The table shows the values of (6.6) for all primes p ≥ 7 depending on their decomposition as p = 30l + r. The sign in the headline is the
same as in the last argument of (6.6).
r −, l odd −, l even +, l odd +, l even
1
8
2
2
8
7
14
4
4
14
11
2
8
8
2
13
4
14
14
4
.
17
4
14
14
4
19
2
8
8
2
23
14
4
4
14
29
8
2
2
8
Table 6.4. The table shows the values of (6.7) for all primes p ≥ 7 depending on their decomposition as p = 30l + r. The sign in the headline is the
same as in the last argument of (6.7).
For the higher pull-backs we get
å
Ç
31(q
−
1)
SyzR (U, V, W) −
2
q q
q
Ç
å
SyzR (U , V , W )
31(q − 1)
2
Syz (U, V, W ) −
+3
R
2
if q mod 30 ∈ {±1, ±11},
if q mod 30 ∈ {±7, ±13}.
Using Formula (6.1), the Hilbert-Kunz function is given by
30 · Q(15, 10, 6) · q2 − 30 · Q(15, 10, 6) + 30 · 120
30 · 120
239 · q2 − 239 + 120
=
120
Ç
å
1
119
= 2−
q2 −
,
120
120
HK(R, q) =
6.5. THE CASE E8
149
if q mod 30 ∈ {±1, ±11} and
30 · Q(15, 10, 6) · q2 − 30 · Q(15, 10, 12) + 30 · 240
30 · 120
2
239 · q − 431 + 240
=
120
å
Ç
191
1
q2 −
,
= 2−
24
120
if q mod 30 ∈ {±7, ±13}.
In characteristics two, three and five, all Frobenius pull-backs split and one obtains (similarly to the computations for E6 ) HKR (R, q) = 2 · q2 (with q > 1).
HK(R, q) =
Theorem 6.7. The Hilbert-Kunz function of E8 is given by
2 · q2
Ç
å
if e ≥ 1 and p ∈ {2, 3, 5},
1
2 119
e 7−→ 2 − 120 q − 120 if q mod 30 ∈ {±1, ±11},
Ç
å
1
191
otherwise.
q2 −
2−
120
120
Reformulating the cases in the above theorem in terms of p, one obtains the function
2 · q2
Ç
å
if e ≥ 1 and p ∈ {2, 3, 5},
1
2 191
q
−
if p mod 30 ∈ {±7, ±13} and e is odd,
e 7−→
120
120
Ç
å
1
119
otherwise.
q2 −
2−
120
120
With the help of Theorem 1.91 one gets the F-signature function.
2−
Corollary 6.8. The F-signature function of R is given by
0
if e ≥ 1 and p ∈ {2, 3, 5},
1
2 119
Fsign (R, q) = 120 · q + 120 if q mod 30 ∈ {±1, ±11},
1
191
· q2 +
otherwise.
120
120
In any case the Hilbert-Kunz functions of surface rings of type ADE have (at least for
1
p ≥ 7) a periodic O(1)-term. One possible value of this term is always 1 − |G|
. It is not
known whether the other possible values of the O(1)-term can be expressed in terms of
invariants of the corresponding groups. Another question is whether the length of the
Frobenius periodicity of Syz(m) is encoded in properties of the group G.
CHAPTER 7
Extensions and further examples
In this chapter we want to discuss further examples of the methods developed so far.
More explicitly, we will compute the Hilbert-Kunz function of a surface ring of type E8
with respect to different ideals I , m and the Hilbert-Kunz functions of the degenerations
A∞ := k[X, Y, Z]/(XY) and D∞ := k[X, Y, Z]/(X 2 Y − Z 2 ) of the singularities of type An and
Dn . In the fourth section we will combine our approach with a theorem of Kustin, Rahmati
and Vraciu to compute the Hilbert-Kunz functions of the homogeneous coordinate ring
of Fermat curves under the condition that Syz(m) is strongly semistable on the projective
spectrum. Finally, in section five we will state some open questions and discuss possible
generalizations of various theorems.
7.1. Hilbert-Kunz functions of E8 with respect to various ideals I , m
In this section we will demonstrate how the developed methods are helpfull to compute
the Hilbert-Kunz function of E8 with respect to the ideals (X, Y, Z 2 ) resp. (X, Y 2 , YZ, Z 2 ).
Example 7.1. Let k be an algebraically closed field of characteristic p ≥ 7,
R := k[X, Y, Z]/(X 2 + Y 3 + Z 5 ),
U := Spec(R) \ {(X, Y, Z)} and I := (X, Y, Z 2 ). By Lemma 5.4 the Frobenius pull-backs
of SyzU (I) are indecomposable, hence of the form SyzU (I) or SyzU (X, Y, Z) by Theorem 3.42. By the explicit computations of the Hilbert-series of SyzR (X p , Y p , Z p ) and
SyzR (X p , Y p , Z 2p ) in Section 6.5 we obtain
Ç
å
37(q − 1)
2
if q mod 30 ∈ {±1, ±11},
SyzR (X, Y, Z ) −
2
q q 2q
Ç
å
SyzR (X , Y , Z )
37(q − 1)
SyzR (X, Y, Z) −
−3
if q mod 30 ∈ {±7, ±13}.
2
Using Lemma 2.2 and Remark 1.22 one obtains that HK(I, R, q) is given by
30 · Q(15, 10, 12) · q2 − 30 · Q(15, 10, 6c)
+ c,
4 · 15 · 10 · 6
where c ∈ {1, 2} such that SyzR (X q , Y q , Z 2q ) SyzR (X, Y, Z c )(m) for some m ∈ Z. All in all,
the function
Ä
ä
71
2 119
3
+
120 ä q − 120
Ä
e 7−→ 3 +
Ä
3+
71
2
120 ä q −
71
2
120 q −
191
120
191
120
if p mod 30 ∈ {±7, ±13} and e is odd,
if p mod 30 ∈ {±7, ±13} and e is even,
if p mod 30 ∈ {±1, ±11}
is the Hilbert-Kunz function of R with respect to I.
151
152
7. EXTENSIONS AND FURTHER EXAMPLES
Example 7.2. Let k, R and U be as in the previous example and I := (X, Y 2 , YZ, Z 2 ). Then
SyzR (I) is the module M2 from Theorem 3.42. Since there are no non-trivial line bundles
on U the Frobenius pull-backs of SyzU (I) either stay indecomposable or become free. In
the case where SyzU (I [q] ) is indecomposable for some q it has to be isomorphic to either
SyzU (I) or SyzU (X, Y 2 , YZ, Z 3 ), since these are representatives for the two indecomposable vector bundles of rank three (cf. Theorem 3.42). Which of the possibilities is true
can be checked by the Hilbert-series. In view of Theorem 4.7 we need to compute the
degree of the generators of
S q−1 := SyzR (X q−1 , Y 2q , Y q Z q , Z 2q ) and S q+1 := SyzR (X q+1 , Y 2q , Y q Z q , Z 2q )
to compute the Hilbert-series of SyzR (X q , Y 2q , Y q Z q , Z 2q ). Since there is no effective way
to compute these we restrict to p = 7. An explicit computation using CoCoA shows that
S 7−1 is generated by elements of degree 140, 142 and 144, while S 7+1 is generated by
elements of degree 150, 152 and 154. This shows
HSyzR (X 7 ,Y 14 ,Y 7 Z 7 ,Z 14 ) (t) = t124 HSyzR (X,Y 2 ,YZ,Z 3 ) (t).
(7.1)
The module SyzR (X 6 , Y 14 , Y 7 Z 7 , Z 21 ) is generated by elements of degree 152, 156 and 160
and the module SyzR (X 8 , Y 14 , Y 7 Z 7 , Z 21 ) is generated by elements of degree 162, 166 and
170. This gives
HSyzR (X 7 ,Y 14 ,Y 7 Z 7 ,Z 21 ) (t) = t135 HSyzR (X,Y 2 ,YZ,Z 2 ) (t).
(7.2)
Combining (7.1) and (7.2) one finds
Ä
7e
SyzR X , Y
2·7e
7e
7e
,Y Z ,Z
ä
2·7e
2
3
R (X, Y , YZ, Z )(−21q + 23)
SyzR (Y, Y 2 , YZ, Z 2 )(−21q + 21)
Syz
if e is odd,
if e is even.
With the help of the standard-graded ring S := k[U, V, W]/(U 30 + V 30 + W 30 ) and the map
R → S , X 7→ U 15 , Y 7→ V 10 , Z 7→ W 6 , we can compute
ÄÄ
ä
HK U 15 , V 20 , V 10 W 6 , W 12 , S , 7e
ä
as in Lemma 2.2. Let q := 7e , a ∈ {2, 3}, n := 21q − 17 − 2a and
2700
D := dimk (S /(U 15 , V 20 , V 10 W 6 , W 6a ) =
3600
if a = 2,
if a = 3,
where a and n are chosen such that
SyzR (X q , Y 2q , Y q Z q , Z 2q ) SyzR (X, Y 2 , YZ, Z a )(−n).
7.2. THE HILBERT-KUNZ FUNCTION OF A∞
With these preparations we obtain
ÄÄ
ä
153
ä
HK U 15 , V 20 , V 10 W 6 , W 12 , S , q
= x→∞
lim
−
+
x
X
0
h (OC (m)) −
m=0
x
X
m=0
x
X
x
X
0
h (OC (m − 15q)) −
m=0
x
X
h0 (OC (m − 16q)) −
h0 (OC (m − n − 20)) +
x
X
h0 (OC (m − 20q))
m=0
x
X
h0 (OC (m − 12q)) +
m=0
m=0
−
x
X
x
X
h0 (OC (m − n − 15))
m=0
h0 (OC (m − n − 16)) +
m=0
x
X
h0 (OC (m − n − 6a))
m=0
h0 (OC (m − n)) + D
m=0
= lim
x→∞
−
x
X
h0 (OC (m)) −
m=0
x−12q
X
h0 (OC (m)) −
h0 (OC (m)) +
x−n−15
X
h0 (OC (m)) −
m=0
x−n
X
h0 (OC (m) −
x−16q
X
m=0
h0 (OC (m)) +
x−n−20
X
m=0
x−n−6a
X
x−20q
X
m=0
m=0
+
x−15q
X
h0 (OC (m))
m=0
h0 (OC (m)) +
m=0
x−n−16
X
h0 (OC (m))
m=0
h0 (OC (m)) + D
m=0
= 30 · (149q + 12a − 102a + 7) + D.
2
2
Since the degree of the extension
Ä
ä
k[U, V, W]/ U 15 , V 10 , W 6 : k
is 15 · 10 · 6 = 302 we obtain
30 · 149 · 49e − 2130
HK(IS , S , 7e ) =
30 · 149 · 49e − 1770
if e is odd,
if e is even
and
71
149
· 49e −
if e is odd,
30
30
HK(I, R, 7 ) =
149 · 49e − 59 if e is even.
30
30
By Theorem 1.80 we see that the vector bundle SyzProj(S ) (IS ) is strongly semistable.
e
7.2. The Hilbert-Kunz function of A∞
In this and the following section we will compute the Hilbert-Kunz functions of the rings
A∞ and D∞ . These examples will show, that we (at least in these cases) can drop the
assumptions
(i) isolated singularity,
(ii) normal domain,
(iii) finite Cohen-Macaulay type,
154
7. EXTENSIONS AND FURTHER EXAMPLES
(iv) there are finitely many isomorphism classes of indecomposable vector bundles
on the punctured spectrum.
Note that we discussed the idea to contsruct first syzygy modules of ideals representing
the indecomposable, maximal Cohen-Macaulay modules only under the assumption of a
normal ring. The condition normal was only required to ensure that there has to be such a
representation. If the ring is not normal, first syzygy modules of ideals are still reflexive,
but they might not be maximal Cohen-Macaulay.
We choose the standard-grading for the ring A∞ := k[X, Y, Z]/(XY).
Theorem 7.3 (Burban, Drozd). The indecomposable, non-free, maximal Cohen-Macaulay A∞ -modules are described up to isomorphism (and degree shift) by the following
matrix factorizations of XY
(i) (X, Y) and (Y, X),
(ii)
åå
å Ç
åå
ÇÇ
å Ç
ÇÇ
Y Zn
X −Z n
X −Z n
Y Zn
,
,
and
,
0 X
0 Y
0 Y
0 X
for all n ≥ 1.
Moreover, the modules corresponding to the matrix factorizations in (i) are the two components of the normalization k[X, Z] × k[Y, Z] of A∞ and are not locally free on the punctured spectrum of A∞ . The modules corresponding to the matrix factorizations in (ii) are
locally free of rank one on the punctured spectrum.
Proof. See [BD08, Theorem 5.3 and Remark 5.5].
The syzygy modules SyzA∞ (X) and SyzA∞ (Y) correspond to the matrix factorizations in
(i). Obviously, they are selfdual. The syzygy modules SyzA∞ (Y, Z n ) and SyzA∞ (X, −Z n )
correspond to the matrix factorizations in (ii). They are dual to each other, since
ÇÇ
(α, α) :
Ä
ä
å Ç
Y 0
X 0
,
n
Z X
−Z n Y
åå
ÇÇ
→
å Ç
X −Z n
Y Zn
,
0 Y
0 X
åå
,
with α := 01 −1
0 is an equivalence of matrix factorizations. Using Formula (4.2) on page
115, one obtains
1+t
,
HA∞ (t) =
(1 − t)2
t2
HSyzA∞ (X) (t) = HSyzA∞ (Y) (t) =
,
(1 − t)2
t2 + tn+1
HSyzA∞ (Y,Z n ) (t) = HSyzA∞ (X,−Z n ) (t) =
.
(1 − t)2
In this case the Hilbert-series of a maximal Cohen-Macaulay module that is locally free
on the punctured spectrum can detect its isomorphism class up to dualizing with one
exception. Namely, the modules SyzA∞ (X, Z 2 ), SyzA∞ (Y, Z 2 ) and A∞ (−2) have the same
Hilbert-series.
7.3. THE HILBERT-KUNZ FUNCTION OF D∞
155
From the splitting
SyzA∞ (X, Y, Z) = SyzA∞ (X, Z) ⊕ SyzA∞ (Y, Z),
we deduce the splittings
SyzA∞ (X q , Y q , Z q ) = SyzA∞ (X q , Z q ) ⊕ SyzA∞ (Y q , Z q ),
where by the symmetry of the problem SyzA∞ (X q , Y q , Z q ) is either free or of the form
SyzA∞ (X, Z n )(m) ⊕ SyzA∞ (Y, Z n )(m) for some n ∈ N and some m ∈ Z. All in all, we see that
it is enough to detect the isomorphism class of SyzA∞ (X q , Z q ) up to dualizing.
By Formula (4.2) on page 115, we have
(2tq − 1)(1 − t2 )
+ Hk[X,Y,Z]/(X q ,Z q ,XY) (t).
(1 − t)3
With M := k[X, Y, Z]/(X q , Z q ) one has by [KR05, Proposition 5.2.16]
(7.3)
HSyzA∞ (X q ,Z q ) (t) =
Hk[X,Y,Z]/(X q ,Z q ,XY) (t) = H M/(XY)M (t) = H M (t) − t2 H M/(0:M (XY)) (t)
= H M (t) − t2 H M/(X q−1 )M (t)
(1 − tq )2 2 (1 − tq−1 )(1 − tq )
−t ·
.
(1 − t)3
(1 − t)3
Substituting Equation (7.4) in Equation (7.3), one obtains
=
(7.4)
HSyzA∞ (X q ,Z q ) (t) =
2
q+1
tq+1 + t2q
q−1 t + t
=
t
·
.
(1 − t)2
(1 − t)2
If q , 2, this gives directly
SyzA∞ (X q , Z q ) SyzA∞ (X, Z q )(−q + 1) or SyzA∞ (X q , Z q ) SyzA∞ (Y, Z q )(−q + 1).
In the case q = 2, we might have SyzA∞ (X 2 , Z 2 ) A∞ (−3). But SyzA∞ (X 2 , Z 2 ) is obviously
not free.
All in all, we see
SyzA∞ (X q , Y q , Z q ) = SyzA∞ (X, Y, Z q )
and get by Lemma 2.2
2 · Q(1, 1, 1) · q2 − 2 · Q(1, 1, q) + 4q
= 2q2 − q.
4
Comparing this result with Example 1.5, we see
HK(A∞ , q) =
HK(A∞ , q) = lim HK(An , q).
n→∞
7.3. The Hilbert-Kunz function of D∞
For the ring D∞ := k[X, Y, Z]/(X 2 Y − Z 2 ) we choose the grading deg(X) = deg(Y) = 2 and
deg(Z) = 3.
Theorem 7.4 (Burban, Drozd). The indecomposable, non-free, maximal Cohen-Macaulay D∞ -modules are described up to isomorphism (and degree shift) by the following
matrix factorizations of X 2 Y − Z 2
156
7. EXTENSIONS AND FURTHER EXAMPLES
(i)
ÇÇ
å Ç
Z XY
−Z XY
,
X Z
X −Z
åå
,
(ii)
ÇÇ 2
X
Z
å Ç
Y −Z
Z
,
Y
−Z X 2
åå
,
(iii)
áá
Z XY 0 −Y n+1
X Z Yn
0
0 0 Z
XY
0 0 X
Z
ë á
−Z −XY 0 Y n+1
X
Z
Yn
0
0
0
−Z −XY
0
0
X
Z
,
ëë
,
for all n ≥ 1 and
(iv)
áá
Z XY −Y n 0
X Z
0
Yn
0 0
Z XY
0 0
X
Z
ë á
,
−Z −XY −Y n
0
X
Z
0
−Y n
0
0
−Z −XY
0
0
X
Z
ëë
,
for all n ≥ 1.
Moreover, all corresponding modules are selfdual, the module corresponding to the matrix factorization (i) has rank one, it is the normalization of D∞ and is not locally free on
the punctured spectrum. The modules corresponding to the matrix factorizations (ii), (iii)
and (iv) are locally free on the punctured spectrum and have rank one in case (ii) and
rank two in the cases (iii) and (iv).
Proof. See [BD08, Theorem 5.7 and Remark 5.9].
Corresponding first syzygy modules of ideals are given by
(i)
(ii)
(iii)
(iv)
SyzD∞ (X, Z) in case (i),
SyzD∞ (Z, Y) in case (ii),
SyzD∞ (Y n+1 , −XY, Z) in case (iii) and
SyzD∞ (Y n , −Z, X) in case (iv),
where one has to remove the second column in the second matrices to get the representations (iii) and (iv).
Although we do not need the Hilbert-series of the above modules to compute the HilbertKunz function of D∞ , we compute them for completeness, using Theorem 4.7. They are
7.3. THE HILBERT-KUNZ FUNCTION OF D∞
157
given by
1 + t3
,
(1 − t2 )2
t5 + t6
HSyzD∞ (X,Z) (t) =
,
(1 − t2 )2
t5 + t6
,
HSyzD∞ (Z,Y) (t) =
(1 − t2 )2
t6 + t7 + t2n+4 + t2n+5
HSyzD (Y n+1 ,−XY,Z) (t) =
and
∞
(1 − t2 )2
t5 + t6 + t2n+2 + t2n+3
n
HSyzD∞ (Y ,−Z,X) (t) =
.
(1 − t2 )2
One should mention the phenomenon that we have two indecomposable, maximal CohenMacaulay D∞ -modules with the same Hilbert-series, but one is locally free on the punctured spectrum, while the other one is not.
To compute the Hilbert-Kunz function of D∞ , we start with the case p = 2 and q ≥ 2.
Using Lemma 1.61, we obtain
HD∞ (t) =
q
SyzD∞ (X q , Y q , Z q ) SyzD∞ X q , Y q , X q Y 2
q
SyzD∞ 1, Y q , Y 2 (−2q)
q
SyzD∞ 1, Y 2 , 1 (−3q)
D∞ (−3q) ⊕ D∞ (−4q).
Using Formula (6.2) on page 139, one gets
6 · Q(2, 2, 3) · 4e + 6 · 4e
= 2 · 4e .
48
If p is odd and q > 1, we get by Lemma 1.61
HK(D∞ , 2e ) =
Å
SyzD∞ (X q , Y q , Z q ) SyzD∞ X q , Y q , ZX q−1 Y
Å
SyzD∞ X, Y , ZY
Å
SyzD∞ X, Y
q
q+1
2
q−1
2
q−1
2
ã
ã
(−2(q − 1))
ã
, Z (−3(q − 1)).
Using Formula (6.1) on page 139, one obtains
6 · Q(2, 2, 3) · q2 − 6 · Q(2, q + 1, 3) + 24 · (q + 1)
1
= 2 · q2 − · (q + 1).
48
2
To stay in our approach, we compute the Hilbert-series of SyzD∞ (X q , Y q , Z q ) to identify
its isomorphism class. By Theorem 4.7, we have
HK(D∞ , q) =
HSyzD∞ (X q ,Y q ,Z q ) (t) =
(t3 − t6 ) HSyzD
∞ (X
q ,Y q ,Z q−1 )
(t) + (1 − t3 ) HSyzD
∞ (X
1 − t6
q ,Y q ,Z q+1 )
(t)
.
158
7. EXTENSIONS AND FURTHER EXAMPLES
It is not hard to see that
±Ö
SyzD∞ (X q , Y q , Z q−1 ) =
±Ö
SyzD∞ (X q , Y q , Z q+1 ) =
Y
q−1 è
2
0
−X
Yq
−X q
0
Ö
X q−1
q+1
−Y 2
,
è Ö
,
èª
0
D∞ (−3q + 1) ⊕ D∞ (−4q + 2),
q+1 èª
XY 2
0
−1
D∞ (−4q) ⊕ D∞ (−3q − 3).
Therefore,
t3q+2 + t3q+3 + t4q + t4q+1
(1 − t2 )2
t5 + t6 + tq+3 + tq+4
= t3q−3 ·
.
(1 − t2 )2
HSyzD∞ (X q ,Y q ,Z q ) (t) =
Since the Hilbert-series of an indecomposable, maximal Cohen-Macaulay D∞ -module,
which is locally free on the punctured spectrum, detects its isomorphism class, we obtain
Å
SyzD∞ (X q , Y q , Z q ) SyzD∞ X, Y
q+1
2
ã
, Z (−3q + 3),
provided we can show that the module in question is indecomposable, since from the
Hilbert-series and the fact that SyzD∞ (X q , Y q , Z q ) is locally free on the punctured spectrum, the splitting
SyzD∞ (X q , Y q , Z q ) SyzD∞ (Z, Y)(−3q + 3) ⊕ SyzD∞ (Z, Y)(−4q + 5)
is still possible.
A comparison with Theorem 6.1 yields
HK(D∞ , q) = n→∞
lim HK(Dn , q).
7.4. On the Hilbert-Kunz functions of two-dimensional Fermat rings
In this section we want to compute the Hilbert-Kunz functions of the Fermat rings
R := k[X, Y, Z]/(X n + Y n + Z n ),
where k is algebraically closed of positive characteristic p with gcd(p, n) = 1, under the
condition that SyzC (X, Y, Z) is strongly semistable on C := Proj(R). For this computation
we need a result of [KRV12]. For completeness, we explain how Kaid computed in his
thesis (cf. [Kai09]) the values of the Hilbert-Kunz function of R for large q if SyzC (X, Y, Z)
is not strongly semistable.
We start with the case, where SyzC (X, Y, Z) is not strongly semistable. Recall that this is
equivalent to
å
Ç
1 1 1
,0
δ , ,
n n n
by Corollaries 2.33 and 2.36.
7.4. ON THE HILBERT-KUNZ FUNCTIONS OF TWO-DIMENSIONAL FERMAT RINGS
159
ä
Ä
Theorem 7.5 (Kaid). Assume gcd(p, n) = 1. If δ n1 , 1n , 1n , 0 and s is the integer of Han’s
Theorem 2.17, then F s∗ (SyzC (X, Y, Z)) is not semistable. Let p s = nl + r with l ∈ N and
0 ≤ r ≤ n − 1. Then the Frobenius pull-back F s∗ (SyzC (X, Y, Z)) has a strong HarderNarasimhan filtration given by
åå
Ç Ç
å
å
Ç Ç
l
l
s∗
s
→ F (SyzC (X, Y, Z)) → OC n l + 1 +
− 3p → 0,
0 → OC −n l + 1 +
2
2
if l is even and
Ç Ç ú üå
å
Ç Ç ú üå
å
l
l
s∗
s
0 → OC −n l +
− 3r → F (SyzC (X, Y, Z)) → OC n l +
+ 3r − 3p → 0,
2
2
if l is odd. Moreover, under the additional assumption gcd(p, n) = 1 this filtration is
minimal for p ≥ d − 3 and s ≥ 1 in the sense that F s−1∗ (SyzC (X, Y, Z)) is semistable.
Proof. At first one needs to prove that the inclusions in the above sequences are nontrivial. This is done in [Kai09, Lemma 4.2.8]. In the second step one has to show that the
kernels above are in fact the maximal destabilizing subbundles. This is done in [Kai09,
Theorem 4.2.12] by computing the degree of the maximal destabilizing subbundles. The
supplement is proven in [Kai09, Theorem 4.2.12] by using Corollary 1.81.
Remark 7.6. In the situation of Theorem 7.5, one has
Ç s s så
p p p
, ,
, Lodd < 1
n n n
1
by assumption. If l is even, this distance is achieved for
Ç¢ s • ¢ s • ¢ s •å
p
p
p
,
,
∈ Lodd ,
n
n
n
showing 3r > 2n. Similarly, if l is odd one obtains 3r < n.
From Theorem 7.5 one obtains the Hilbert-Kunz function of R for large q.
Theorem 7.7 (Kaid). In the situation of Theorem 7.5 one has
HK(R, pe ) = eHK (R)p2e for all e 0.
Proof. By Theorem 7.5 we have a short exact sequence
0 → OC (α) → F s∗ (SyzC (X, Y, Z)) → OC (β) → 0.
Since this is the strong Harder-Narasimhan filtration of F s∗ (SyzC (X, Y, Z)), we obtain by
Remark 1.75 the strong Harder-Narasimhan filtrations
0 → OC (α · pt ) → F s+t∗ (SyzC (X, Y, Z)) → OC (β · pt ) → 0.
These sequences split, whenever
Ä
ä
Ä
ä
Ext1 OC (β · pt ), OC (α · pt ) H1 C, OC ((α − β) · pt )
vanishes. Using Serre duality and ωC OC (n − 3) (cf. [Har87, Example II.8.20.3]), we
have to show that
OC ((2n − 3r) · pt + n − 3) if l is even or
OC ((3r − n) · pt + n − 3)
if l is odd
160
7. EXTENSIONS AND FURTHER EXAMPLES
have no global sections. This is true if (2n − 3r) · pt + n − 3 resp. (3r − n) · pt + n − 3 is
negative. In view of Remark 7.6 this happens for t big enough. In particular, we must
have t ≥ 1. Note that t = 1 is enough in the cases p > n − 3.
Then one can compute the Hilbert-Kunz function for large e with the help of Lemma 2.1.
See [Kai09, Corollary 4.2.17] for the complete proof.
Example 7.8. Let p = 5 and n = 6. Then s = 1 in Han’s Theorem 2.17 and by Theorem
7.5 the strong Harder-Narasimhan filtration of F∗ (SyzC (X, Y, Z)) is given by
0 → OC (−6) → SyzC (X 5 , Y 5 , Z 5 ) → OC (−9) → 0.
Taking the t-th Frobenius pull-back of this sequence, the resulting sequence splits if
−3 · 5t + 3 < 0 ⇔ t ≥ 1.
Explicitly, we have
Fe∗ (SyzC (X, Y, Z)) OC (−6 · 5e−1 ) ⊕ OC (−9 · 5e−1 )
for e ≥ 2. By Theorem 7.7 and Corollary 2.36 the Hilbert-Kunz function of R for e ≥ 2 is
given by
126 e
·5 .
HK(R, 5e ) = eHK (R) · 5e =
25
Ä
ä
Example 7.9. Let p = 3 and n = 7. Then δ 17 , 17 , 17 =
have the strong Harder-Narasimhan filtration
1
63
with s = 2. By Theorem 7.5 we
0 → OC (−13) → F2∗ (SyzC (X, Y, Z)) → OC (−14) → 0.
Since −3t + 4 < 0 ⇔ t ≥ 2, the e-th Frobenius pull-backs of SyzC (X, Y, Z) split as
Ä
ä
Ä
OC −13 · pe−2 ⊕ OC −14 · pe−2
ä
for e ≥ 4. By Theorem 7.7 and Corollary 2.36 we have
427 e
HK(R, 3e ) = eHK (R) · 3e =
·3
81
for e ≥ 4. The other values are HK(R, 30 ) = 1, HK(R, 31 ) = 27, HK(R, 32 ) = 419 and
HK(R, 33 ) = 3843 by an explicit computation with CoCoA. Moreover, another explicit
computation shows
SyzR (X 9 , Y 9 , Z 9 ) SyzR (X 5 , Y 5 , Z 5 )(−6) and SyzR (X 27 , Y 27 , Z 27 ) R(−39) ⊕ R(−42).
We now turn to the case, where SyzC (X, Y, Z) is strongly semistable.
Lemma 7.10. If p is odd, the bundle SyzC (X, Y, Z) is not strongly semistable if and only
if there is a q = pe such that SyzC (X q , Y q , Z q ) OC (l1 ) ⊕ OC (l2 ) holds for some l1 , l2 ∈ Z.
Proof. Let SyzC (X, Y, Z) be strongly semistable and assume that SyzC (X q , Y q , Z q )
OC (l1 ) ⊕ OC (l2 ) for some q. Since SyzC (X q , Y q , Z q ) is semistable, we must have l1 = l2 .
Computing the degrees of the vector bundles SyzC (X q , Y q , Z q ) and OC (l)2 , one obtains the
contradiction −3nq = 2nl.
If SyzC (X, Y, Z) is not strongly semistable, then by the proof of Theorem 7.7 the bundles
SyzC (X q , Y q , Z q ) split as OC (l1 ) ⊕ OC (l2 ) for all large q.
7.4. ON THE HILBERT-KUNZ FUNCTIONS OF TWO-DIMENSIONAL FERMAT RINGS
161
Corollary 7.11. If p is odd, the bundle SyzC (X, Y, Z) is strongly semistable if and only if
all quotients R/(X q , Y q , Z q ) have infinite projective dimension.
Proof. For a fixed q the quotient R/(X q , Y q , Z q ) has finite projective dimension if and
only SyzR (X q , Y q , Z q ) is free if and only if SyzC (X q , Y q , Z q ) OC (l1 ) ⊕ OC (l2 ) for some l1 ,
l2 ∈ Z.
Remark 7.12. Note that Lemma 7.10 and Corollary 7.11 hold in characteristic two under
the further assumption that SyzC (X, Y, Z) is not trivialized by the Frobenius.
In [KRV12] the authors study on Fermat rings R of degree n the minimal projective resolution of the quotients
Ä
ä
R/ X N , Y N , Z N
depending on N and n. In particular they have shown the following.
Theorem 7.13 (Kustin, Rahmati, Vraciu). Let R = k[X, Y, Z]/(X n + Y n + Z n ), where k denotes a field of characteristic p ≥ 0 and n ∈ N≥1 . For a given integer N ≥ 1 the quotient
R/(X N , Y N , Z N ) has finite projective dimension as R-module if and only if one of the following conditions is satisfied
(i) n|N,
(ii) p = 2 and n ≤ N,
(iii) p is odd and there exist positive integers J and e with J odd such that
e−1
3
e
N
Jp −
< p 3−1
n
pe +1
e
3
if p = 3,
if pe ≡ 1 mod 3,
if pe ≡ 2 mod 3
Proof. See [KRV12, Theorem 6.3].
Moreover, Kustin, Rahmati and Vraciu computed the minimal graded free resolution of
R/(X N , Y N , Z N ) if this quotient has infinite projective dimension.
Definition 7.14. Let r and s be positive integers with r + s = n. Then we define
á
ϕr,s :=
0
Z r −Y r X s
−Z r
0
Xr Y s
r
r
Y
−X
0 Zs
−X s −Y s −Z s 0
ë
.
Note that ϕr,s ϕ s,r = −(X n + Y n + Z n ) Id4 , hence (ϕr,s , ϕ s,r ) are matrix factorizations for
−(X n + Y n + Z n ). They are equivalent to the matrix factorizations from Example 3.30
with di = n and a = b = c = r. This gives the (ungraded) isomorphisms
coker(ϕr,s ) SyzR (X r , Y r , Z r ).
Theorem 7.15 (Kustin, Rahmati, Vraciu). Let N = θ · n + r with θ ∈ N and r ∈ {1, . . . , n − 1}.
Assume that
Ä
ä
Q := R/ X N , Y N , Z N
162
7. EXTENSIONS AND FURTHER EXAMPLES
has infinite projective dimension. If θ = 2 · η − 1, then the homogenous minimal free resolution of Q is given by
...
ϕr,n−r
/
F3
ϕn−r,r
/
F2
ϕr,n−r
/
/
/
R(−N)3
F1
R
with
F1 := R(−3ηn + n − r)3 ⊕ R(−3ηn + 2n − 3r),
F2 := R(−3ηn + n − 2r)3 ⊕ R(−3ηn),
F3 := R(−3ηn − r)3 ⊕ R(−3ηn + n − 3r).
If θ = 2 · η, then the homogenous minimal free resolution of Q is given by
...
ϕn−r,r
/
F3
ϕr,n−r
/
F2
ϕn−r,r
/
F1
/
/
R(−N)3
R
with
F1 := R(−3ηn − 2r)3 ⊕ R(−3ηn − n),
F2 := R(−3ηn − n − r)3 ⊕ R(−3ηn − 3r),
F3 := R(−3ηn − n − 2r)3 ⊕ R(−3ηn − 2n).
Proof. The statement follows from [KRV12, Theorem 3.5] combined with [KRV12,
Theorems 5.14 and 6.1].
Corollary 7.16. Under the hypothesis of Theorem 7.15, we have
Ä
ä
coker(ϕ
SyzR X N , Y N , Z N (m)
if θ is even,
coker(ϕn−r,r ) if θ is odd,
r,n−r )
Syz
R (X
SyzR
r, Y r, Zr)
(X n−r , Y n−r , Z n−r )
if θ is even,
if θ is odd.
for some m ∈ Z.
In [BK13] the authors proved that SyzC (X, Y, Z) admits a (0, 1)-Frobenius periodicity if
p ≡ −1 (2n) (cf. Example 7.20). The last corollary was the last ingredient necessary to
generalize this result.
Theorem 7.17. Let C be the projective Fermat curve of degree n over an algebraically
closed field of characteristic p > 0 with gcd(p, n) = 1. The bundle SyzC (X, Y, Z) admits a
Frobenius periodicity if and only if
Ç
å
1 1 1
= 0.
δ , ,
n n n
Moreover, the length of this periodicity is bounded above by the order of p modulo 2n.
Proof. If SyzC (X, Y, Z) admits a Frobenius periodicity, it has to be strongly semistable.
If it is strongly semistable we have to distinguish two cases. If SyzC (X, Y, Z) is trivialized
by the Frobenius (which might only happen if p = 2 by Lemma 7.10), we have of course
a Frobenius periodicity. Otherwise, we obtain a Frobenius periodicity by Corollary 7.16,
7.4. ON THE HILBERT-KUNZ FUNCTIONS OF TWO-DIMENSIONAL FERMAT RINGS
163
which can be applied by Corollary 7.11 (resp. Remark 7.12). The supplement follows
since the isomorphism class of SyzR (X q , Y q , Z q ) depends on r and the parity of q−r
n .
This result enables us to compute the Hilbert-Kunz function of R, if SyzC (X, Y, Z) is
strongly semistable. If SyzC (X, Y, Z) is trivialized by the Frobenius, we can use Lemma
2.1 to obtain HK(R, 2e ) = 3n · 4e−1 for e 0.
Theorem 7.18. Let R := k[X, Y, Z]/(X n + Y n + Z n ) with an algebraically closed field k of
characteristic p > 0 coprime to n. Assume that SyzC (X, Y, Z) is strongly semistable and
not trivialized by the Frobenius. Let q = pe = nθ + r with θ ∈ N and 0 ≤ r < n. Then
3n 2 3n 2 3
·q − r +r
if θ is even,
4
HK(R, q) = 4
3n 2 3n
· q − (n − r)2 + (n − r)3 if θ is odd.
4
4
Proof. If SyzC (X, Y, Z) is strongly semistable, the quotients R/(X q , Y q , Z q ) have infinite projective dimension and their resolutions are given by Theorem 7.15. Now combine
Corollary 7.16 with Lemma 2.2.
Example 7.19. Let pÄ ≡ 1 (2n).
We have always pe = nθ + 1 for some even θ. An easy
ä
1 1 1
computation shows δ n , n , n = 0. By Corollary 7.16, we have
Ç
3
SyzR (X , Y , Z ) Syz(X, Y, Z) − · (q − 1)
2
q
q
å
q
for all q, hence SyzC (X, Y, Z) admits a (0, 1)-Frobenius periodicity. With Theorem 7.18
we obtain
3n 2
3n
HK(R, q) =
·q +1− .
4
4
Example 7.20. Let p ≡ −1 (2n). We have pe = nθ + n − 1 for some oddÄ θ if e is
ä odd and
pe = nθ + 1 for some even θ if e is even. Again, it is easy to show δ 1n , 1n , 1n = 0. By
Corollary 7.16, we have
Ç
3
SyzR (X , Y , Z ) Syz(X, Y, Z) − · (q − 1)
2
q
q
å
q
for all q, hence SyzC (X, Y, Z) admits a (0, 1)-Frobenius periodicity. With Theorem 7.18
we obtain
3n 2
3n
·q +1− .
HK(R, q) =
4
4
This recovers Theorem 3.4 and Corollary 4.1 of [BK13].
Example 7.21. Let n = 5. If p ≡ ±1 (10), the Hilbert-Kunz function of R was computed
in the two previous examples and is given by
HK(R, q) =
15 2 11
·q − .
4
4
164
7. EXTENSIONS AND FURTHER EXAMPLES
In this case we can also compute the Hilbert-Kunz function for p ≡ ±3 (10). If p ≡ 3 (10),
we have pe = 5θ + r, where (θ, r) is of the form
(even, 3)
(odd, 4)
if e ≡ 1 (4),
if e ≡ 2 (4),
(θ, r) =
(odd, 2) if e ≡ 3 (4),
(even, 1) if e ≡ 0 (4).
If p ≡ 7 (10), we have pe = 5θ + r, where (θ, r) is of the form
(odd, 2)
(odd, 4)
if e ≡ 1 (4),
if e ≡ 2 (4),
(θ, r) =
(even, 3) if e ≡ 3 (4),
(even, 1) if e ≡ 0 (4).
It easy to check that δ
Ä
1 1 1
5, 5, 5
ä
= 0 in both cases. From Corollary 7.16 we obtain
Ç
å
3
3
3
3
SyzR (X , Y , Z ) − 2 · (q − 3)
e∗
Ç
å
F (SyzC (X, Y, Z))
3
Syz (X, Y, Z) − · (q − 1)
R
2
if e is odd,
if e is even.
Since the Hilbert-series
−3t6 + t5 − t3 + 3t2
HSyzR (X,Y,Z) (t) =
,
(1 − t)3
−t9 − 3t8 + 3t6 + t5
HSyzR (X 3 ,Y 3 ,Z 3 ) (t) =
(1 − t)3
do not differ by a factor tl , the R-modules SyzR (X, Y, Z) and SyzR (X 3 , Y 3 , Z 3 ) are not isomorphic, hence SyzC (X, Y, Z) admits a (0, 2)-Frobenius periodicity. The Hilbert-Kunz
function of R is given by
15 2 27
·q −
if e is odd,
4
HK(R, q) = 4
15 2 11
·q −
if e is even.
4
4
Example 7.22. Let p = 37 and n = 14. Then pe = 14 · θ + r with
(even, 9)
if e ≡ 1 (3),
(θ, r) = (odd, 11) if e ≡ 2 (3),
(even, 1) if e ≡ 0 (3).
7.4. ON THE HILBERT-KUNZ FUNCTIONS OF TWO-DIMENSIONAL FERMAT RINGS
165
This gives the Hilbert-Kunz function
243
21
· 372e −
2
2
21
135
HK(R, 37e ) =
· 372e −
2
2
21
19
· 372e −
if e ≡ 1 (3),
if e ≡ 2 (3),
if e ≡ 0 (3).
2
2
In this example, we see directly that we have a (0, 3)-Frobenius periodicity
ä
Ä
F3∗ (SyzC (X, Y, Z)) SyzC (X, Y, Z) − 32 · (q − 1) .
The next examples deal with the question which Frobenius periodicities of the bundles
SyzC (X, Y, Z) can be achieved. A sufficient condition for having a Frobenius periodicity is
(cf. Theorem 7.17)
Ç
å
1 1 1
δ , ,
= 0,
n n n
e e e
which is equivalent to the condition that the distances of all triples ve := pn , pn , pn to
Lodd are at least one. Let pe = θe · n + re with θe , re ∈ N and 0 ≤ re < n. The nearest element
to ve in Lodd is given by the component-wise round-up of ve if θe is even and by the
component-wise round-down of ve if θe is odd. This leads to the conditions
Å
ã
re
3· 1−
≥ 1 if θe is even,
2 · n ≥ 3 · re if θe is even,
n
⇔
re
3 · re ≥ n
if θe is odd.
3· ≥ 1
if θe is odd.
n
Example 7.23. Let p be odd, l ∈ N, l ≥ 0 and n =
(7.5)
0 · n + pe
Ä
pl+1 +1
2 .
For 0 ≤ e ≤ 2l + 2 we have
if 0 ≤ e ≤ l,
ä
pe = 2 · pe−l−1 − 1 · n + n − pe−l−1 if l + 1 ≤ e ≤ 2l + 1,
Ä
ä
2 · pl+1 − 2 · n + 1
if e = 2l + 2.
0
0
This shows that pe is of the form even · n + pe or odd ·Än + n −äpe for some 0 ≤ e0 ≤ l. Since
0
2n = pl+1 + 1 ≥ 3pe for all 0 ≤ e0 ≤ l, we see that δ 1n , 1n , 1n = 0. By Corollary 7.16 we
find the isomorphism
Ç
å
Å
0
0
0ã
3
e∗
pe
pe
pe
e
e0
− · (p − p )
F (SyzC (X, Y, Z)) SyzC X , Y , Z
2
0
0
with e ≡ e mod l + 1 and e ∈ {0, . . . , l}. Since the Hilbert-series
e
e
e
3 · t2p − 3 · tn+p + tn − t3p
HSyzR (X pe ,Y pe ,Z pe ) (t) =
(1 − t)3
for e ∈ {0, . . . , l} do not differ by a factor ta , a ∈ Z, the R-modules are non-isomorphic. All
in all, we found a (0, l + 1)-Frobenius periodicity. The Hilbert-Kunz function is given by
0ä
0
3n Ä 2e
HK(R, pe ) =
· p − p2e + p3e
4
with e ≡ e0 mod l + 1 and e0 ∈ {0, . . . , l}.
166
7. EXTENSIONS AND FURTHER EXAMPLES
Example 7.24. Let p = 2 and n = 3. Then 2e is of the form even
Ä · n + ä2 or odd · n + 1.
e
Since 2 · n = 3 · 2 ≥ 3 · 2 for 0 ≤ e ≤ 1 and 3 · 1 = n, we see that δ n1 , 1n , 1n = 0. Applying
Corollary 7.16, we find the isomorphism
(
SyzC (X, Y, Z)
if e = 0,
e∗
Ä
äÄ
ä
F (SyzC (X, Y, Z))
SyzC X 2 , Y 2 , Z 2 −3(pe−1 − 1) if e ≥ 1.
Since the Hilbert-series of SyzR (X, Y, Z) and SyzR (X 2 , Y 2 , Z 2 ) are given by
3t4 − 3t5 + t3 − t6
3t2 − 3t4 + t3 − t3
and
,
(1 − t)3
(1 − t)3
we see that the two modules are non-isomorphic. We obtain a (1, 2)-Frobenius periodicity.
By Theorem 7.18 the Hilbert-Kunz function is given by
1
HK(R, 2e ) =
9 · 22e−2 − 1
if e = 0,
if e ≥ 1.
Example 7.25. Let p = 2 and 2l ≤ n < 2l+1 , hence n = 2l + x with 1 ≤ x < 2l . We want to
show that SyzC (X, Y, Z) admits
a äFrobenius periodicity on the smooth curve C if and only
Ä
1 1 1
if n = 3. Since we need δ n , n , n = 0, we must have
2n = 2l+1 + 2x ≥ 3 · 2l ≥ 3 · 2e
for all 0 ≤ e ≤ l. This is equivalent to x ≥ 2l−1 . Since 2l+1 = n+2l − x with 0 < 2l − x < 2l < n,
we need that the inequality 3 · (2l − x) ≥ n holds. This is equivalent to x ≤ 2l−1 , hence
n = 2l + 2l−1 = 3 · 2l−1 . Because of the assumption gcd(2, n) = 1, only the case n = 3 is left.
From the Examples 7.24 - 7.25 we obtain the following corollary.
Corollary 7.26. Let p = 2. The bundle SyzC (X, Y, Z) admits a Frobenius periodicity on
the smooth curve C if and only if n = 3.
7.5. Open questions
At the end, we gather some open questions related to topics of this thesis.
7.5.1. Questions concerning rings of dimension two. In Chapter 3 we used some
sheaf-theoretic methods to compute from a given matrix factorization (ϕ, ψ) a first syzygy
module of an ideal such that this module is isomorphic to the cokernel of the matrix
factorization. At some point in Construction 3.35 we used the phrase “Suppose that we
can choose a J ⊆ {1, . . . , n}...”, where J was the set of indices of the columns of ψ we had
to keep. In general, it is not understood under which circumstances this J has to exist.
Question 7.27. Are there necessary (and sufficient) conditions (on the hypersurface R)
ensuring that the set J exists?
Another question is what happens to morphisms between matrix factorizations under this
construction.
Question 7.28. Is it possible to make the construction in Chapter 3 categorical?
7.5. OPEN QUESTIONS
167
Another related question is whether sheaf-theoretic tools are necessary or not.
Question 7.29. Is there a completely algebraic way to obtain first syzygy modules of
ideals from a given matrix factorization?
In Chapter 5 we computed the pull-backs of the indecomposable, maximal Cohen-Macaulay modules over surface rings of type ADE along the inclusions induced by normal subgroups. From these computations we obtained that the Frobenius pull-backs of SyzR (m),
where R is of type D or E, are indecomposable. This approach is quite complicated and
works only in this very special situation. This leads to the following question.
Question 7.30. Is there a (numerical) invariant of vector bundles, which measures the
splitting behaviour?
We turn back to the Fermat rings R = k[X, Y, Z]/(X n + Y n + Z n ) with an algebraically closed
field k of characteristic p > 0. In the case where S := SyzProj(R) (X, Y, Z) is not strongly
semistable, we saw in Theorem 7.5 that the strong Harder-Narasimhan filtration of F s∗ (S),
where s ≥ 0 comes from Han’s Theorem, is of the form
0 → OProj(R) (l1 ) → F s∗ (S) → OProj(R) (l2 ) → 0
for some l1 , l2 ∈ Z. In the proof of Theorem 7.7 we saw that there is a t ∈ N such that the
t-th Frobenius pull-back of the above sequence splits. In this situation we can compute
HK(R, pe ), when pe < n or e ≥ s + t.
Question 7.31. Let C be a projective Fermat curve of degree n, let s be the integer from
Han’s Theorem and t be minimal such that
F s+t∗ (SyzC (m)) OC (m1 ) ⊕ OC (m2 )
for some m1 , m2 ∈ Z. Find a way to compute HK(R, pe ) in the cases n < pe ≤ p s+t−1 .
Question 7.32. The situation is the same as in the previous question. Is it possible that
SyzR (m[q] ) splits as a direct sum of two non-free R-modules of rank one?
The next questions deal with general diagonal hypersurfaces of dimension two. We denote
by R the standard-graded Fermat ring of degree n and by C its projective spectrum. We
will denote by S a general diagonal hypersurface with exponent-triple (d1 , d2 , d3 ) ∈ N3≥2
and by D its projective spectrum.
Example 7.33. Han computed HK(S , 3e ) in her thesis [Han91], where S is the diagonal
hypersurface with exponent-triple (5, 8, 8). The Hilbert-Kunz multiplicity
is 24
5 and the
Ä
ä
19
81
1 1 1
e
tail of HK(S , 3 ) is − 5 if e is even and − 5 if e is odd. Since δ 5 , 8 , 8 = 0, we know by
Corollary 2.33 that SyzC (X 8 , Y 5 , Z 5 ) is strongly semistable on the projective Fermat curve
of degree 40. A direct computation using CoCoA shows
SyzD (X 3 , Y 3 , Z 3 )(−9q + 27)
7 7
if e 1 (4),
SyzD (X, Y , Z )(−9q + 39)
if e 2 (4),
SyzD (X 3 , Y 5 , Z 5 )(−9q + 37)
if e 3 (4),
if e 0 (4).
Fe∗ (SyzD (X, Y, Z))
SyzD (X, Y, Z)(−9q + 9)
168
7. EXTENSIONS AND FURTHER EXAMPLES
This shows that there is a (0, e0 )-Frobenius periodicity
F4∗ (SyzD (X, Y, Z)) SyzD (X, Y, Z)(m) for some m ∈ Z
of length e0 ≤ 4. Computing the Hilbert-series of the corresponding S -modules, one sees
that they do not differ by a factor tm , hence they cannot be isomorphic and the periodicity
has exactly length four.
Using Formula 6.1 on page 139, we recover Han’s result. With 3e = 5θ + r, θ ∈ N, r ∈
{0, . . . , 4}, the tuple (θ, r) is cyclic of the form (even, 1), (even, 3), (odd, 4), (even, 2) and
with 3e = 8θ0 + r0 , θ0 ∈ N, r0 ∈ {0, . . . , 7}, the tuple (θ0 , r0 ) is cyclic of the form (even, 1),
(even, 3), (odd, 1), (odd, 3) for e ≥ 0. Hence, we obtain in this case (for some m ∈ Z)
Syz
e∗
F (SyzD (X, Y, Z))(m)
D (X
r
0
0
, Y r , Zr )
0
if θ and θ0 are even,
0
SyzD (X 5−r , Y 8−r , Z 8−r ) if θ and θ0 are odd,
which has a similar shape as Corollary 7.15.
The surface rings of type An (in odd characteristics), E6 and E8 are other diagonal hypersurfaces and we saw that the corresponding δ vanishes in all characteristics coprime to the
degree of the defining polynomial. In these cases we saw that the modules Syz(X q , Y q , Z q )
satisfy isomorphisms similar to Corollary 7.15.
For 1 ≤ ai ≤ di − 1 we define the matrix factorizations (ϕa1 ,a2 ,a3 , ψa1 ,a2 ,a3 ) with
á
ϕa1 ,a2 ,a3 :=
X a1
Y a2
Z a3
0
d
−a
d
−a
2
2
1
1
−Y
X
0
Z a3
d
−a
d
−a
−Z 3 3
0
X 1 1 −Y a2
0
−Z d3 −a3 Y d2 −a2 X a1
á d1 −a1
X
ψa1 ,a2 ,a3 :=
Y d2 −a2
Z d3 −a3
0
−Y a2
−Z a3
0
a
X 1
0
−Z a3
0
X a1
Y a2
Z d3 −a3 −Y d2 −a2 X d1 −a1
ë
ë
.
Assume that Q := k[X, Y, Z]/(X N , Y N , Z N , X d1 +Y d2 +Z d3 ) has infinite projective dimension
as S := k[X, Y, Z]/(X d1 +Y d2 +Z d3 )-module. Let N := di θi +ri with θi ∈ N and 0 ≤ ri ≤ di −1.
Let Ai := ri if θi is even and Ai := di − ri if θi is odd.
Question 7.34. Is it true that the minimal S -free resolution of Q is given by
ϕA1 ,A2 ,A3
ψA1 ,A2 ,A3
ϕA1 ,A2 ,A3
. . . −→ S 4 −→ S 4 −→ S 4 −→ S 3 −→ S
up to the degree shifts necessary to make the resolution homogeneous?
Example 7.35. Consider R := k[X, Y, Z]/(X 12 + Y 12 + Z 12 ) and I := (X 4 , Y 4 , Z 3 ) in characteristic five. Then
å
Ç
1 1
4 4 3
, ,
=
· ,
δ
12 12 12
12 5
hence s = 1 in Han’s Theorem and SyzC (I) is not strongly semistable. A direct computation shows that the minimal degree of a generator of SyzR (X 20 , Y 20 , Z 15 ) is 27 and that we
have a splitting
SyzR (X 100 , Y 100 , Z 75 ) R(−135) ⊕ R(−140).
7.5. OPEN QUESTIONS
169
This gives for e ≥ 2
2928
· 25e and
25
61
HK(k[X, Y, Z]/(X 3 + Y 3 + Z 4 ), 5e ) =
· 25e
25
by Lemma 2.1. In the cases E6 and E8 we had in the characteristics two, three (and five)
a splitting for e ≥ 1. In these cases one has δ , 0 and s = 1.
HK(I, R, 5e ) =
Question 7.36. Assume δ
for large e?
Ä
a b c
n, n, n
ä
, 0. Do the modules SyzR (X aq , Y bq , Z cq ) become free
Equivalently, one could ask the following question.
Question 7.37. Assume δ
large e?
1 1 1
d1 , d2 , d3
, 0. Do the modules SyzS (m[q] ) become free for
7.5.2. Questions concerning rings of dimension at least three. As soon as the dimension of R grows, there are very few explicit examples of Hilbert-Kunz multiplicities
and Hilbert-Kunz functions. One might hope to use Knörrer’s periodicity (cf. [Yos90,
Theorem 12.10]) to obtain the Hilbert-Kunz function of
T := k[U1 , . . . , Un , X, Y, Z]/
Ñ
n
X
é
Ui2 + F(X, Y, Z) ,
i=1
where F is of type ADE. Recall that Knörrer’s periodicity says that the categories MF(R)
and MF(S ) with R = k[X1 , . . . , Xn ]/(G), S = k[U, V, X1 , . . . , Xn ]/(U 2 + V 2 +G) and char(k) ,
ˆ + iV, U − iV). In general first syzygy modules of
2 are equivalent via the functor _⊗(U
ideals are not (maximal) Cohen-Macaulay, but maybe they appear in the two periodic
free resolutions coming from the matrix factorizations in which case they would be (even
maximal) Cohen-Macaulay. This does not happen. We prove this for SyzT (m), where F
is of type E6 . We have
((n + 1)t6 + t4 + t3 − 1)(1 − t12 )
+1
(1 − t6 )n+1 (1 − t4 )(1 − t3 )
((n + 1)t6 + t4 + t3 − 1)(1 + t6 ) + (1 − t6 )n (1 − t4 )(1 − t3 )
=
.
(1 − t6 )n (1 − t4 )(1 − t3 )
HSyzT (m) (t) =
If SyzT (m) would be Cohen-Macaulay, all coefficients in the numerator of its Hilbertseries have to be positive (cf. Lemma 4.8). The numerator’s leading term is (−1)n t6n+7 ,
which is negative for n odd. If n is even, the second highest term of (1 − t6 )n (1 − t4 )(1 − t3 )
is −t6n+4 , which does not cancel with a term from ((n + 1)t6 + t4 + t3 − 1)(1 + t6 ). This
shows that one needs to work with syzygy modules of higher order.
Question 7.38. Is it possible to compute from a given matrix factorization a higher order
syzygy module of an ideal which is isomorphic to the cokernel of the matrix factorization?
APPENDIX A
An implementation of Corollary 2.31 for CoCoA
We give an implementation of Corollary 2.31 in CoCoA.
Function A.1. An implementation of the floor function for Q∈ Q.
Define Floor(Q)
M:=Num(Q);
N:=Den(Q);
If Mod(M,N)=0
Then Return M/N;
Else I:=0; While I<=M/N Do I:=I+1; EndWhile; Return I-1;
EndIf;
EndDefine;
Function A.2. An implementation of the ceil function for Q∈ Q.
Define Ceil(Q)
M:=Num(Q);
N:=Den(Q);
If Mod(M,N)=0
Then Return M/N;
Else I:=M; While I>=M/N Do I:=I-1; EndWhile; Return I+1;
EndIf;
EndDefine;
Function A.3. This function returns for a list L∈ ([0, . . . , ∞) ∩ Q)3 the minimal distance
to the lattice Lodd .
Define DistLodd(L)
Lodd:=[[Floor(L[1]),Floor(L[2]),Floor(L[3])],
[Floor(L[1]),Floor(L[2]),Ceil(L[3])],[Floor(L[1]),Ceil(L[2]),Floor(L[3])],
[Floor(L[1]),Ceil(L[2]),Ceil(L[3])],[Ceil(L[1]),Floor(L[2]),Floor(L[3])],
[Ceil(L[1]),Floor(L[2]),Ceil(L[3])],[Ceil(L[1]),Ceil(L[2]),Floor(L[3])],
[Ceil(L[1]),Ceil(L[2]),Ceil(L[3])]];
Dist:=[];
For I:=8 To 1 Step -1 Do
If IsEven(Sum(Lodd[I]))
Then Remove(Lodd,I);
Else
Append(Dist,Abs(L[1]-Lodd[I][1])+Abs(L[2]-Lodd[I][2])+Abs(L[3]-Lodd[I][3]));
EndIf;
EndFor;
Return Min(Dist);
EndDefine;
171
172
A. AN IMPLEMENTATION OF COROLLARY 2.31 FOR COCOA
Function A.4. This function computes a lower bound for the s in Han’s Theorem. The
argument P is a prime number and the argument T is a list of three non-negative rational
numbers.
Define Lower(P,T)
M:=Max(T);
L:=0;
If 3/2*M<=1
Then
While P^L>=3/2*M Do L:=L-1; EndWhile; L:=L+1;
Else
While P^L<3/2*M Do L:=L+1; EndWhile;
EndIf;
Return L;
EndDefine;
Function A.5. This function computes an upper bound for the s in Han’s Theorem. The
argument P is a prime number and the argument T is a list of three non-negative rational
numbers.
Define Upper(P,T)
U:=1;
Residues:=[];
M:=2*LCM(Den(T[1]),Den(T[2]),Den(T[3]));
A:=[Mod(P*Num(T[1]),M),Mod(P*Num(T[2]),M),Mod(P*Num(T[3]),M)];
While Not (A IsIn Residues) Do
Append(Residues,A);
U:=U+1;
A:=[Mod(P^U*Num(T[1]),M),Mod(P^U*Num(T[2]),M),Mod(P^U*Num(T[3]),M)];
EndWhile;
Return U-1;
EndDefine;
Function A.6. This function computes the characteristic P value of Han’s δ function of
the list T of three non-negative rational numbers.
Define Delta(P,T)
If 2*Max(T)>=Sum(T) Then Return 2*Max(T)-Sum(T); EndIf;
L:=Lower(P,T);
U:=Upper(P,T);
For S:=L To U Do
D:=DistLodd(P^S*T);
If D<1 Then Return (1-D)/P^S; EndIf;
EndFor;
Return 0;
EndDefine;
Function A.7. This function is the final implementation of Corollary 2.31. Within the
notation of this Corollary, the argument Exp is a 3 × 3 matrix whose i-th row consists of
A. AN IMPLEMENTATION OF COROLLARY 2.31 FOR COCOA
173
the exponents in the i-th summand of the trinomial F. The argument P is a prime number,
the argument T is the list [a, b, c] and Type denotes the type of F, hence 1 or 2.
Define HKMTri(Exp,P,T,Type)
D:=Sum(Exp[1]);
Alphap:=T[1]*(D-Exp[2][3]-Exp[3][2])-T[2]*Exp[3][1]-T[3]*Exp[2][1];
Betap:=T[2]*(D-Exp[3][1]-Exp[1][3])-T[1]*Exp[3][2]-T[3]*Exp[1][2];
Gammap:=T[3]*(D-Exp[1][2]-Exp[2][1])-T[1]*Exp[2][3]-T[2]*Exp[1][3];
If Alphap>=0 Then Alphau:=Alphap; Alphad:=0;
Else Alphau:=0; Alphad:=-Alphap; EndIf;
If Betap>=0 Then Betau:=Betap; Betad:=0;
Else Betau:=0; Betad:=-Betap; EndIf;
If Gammap>=0 Then Gammau:=Gammap; Gammad:=0;
Else Gammau:=0; Gammad:=-Gammap; EndIf;
If Type=1 Then Alpha:=Alphau+Gammad;
Else Alpha:=Alphau+Betad+Gammad; EndIf;
If Type=1 Then Beta:=Betau+Alphad;
Else Beta:=Betau; EndIf;
If Type=1 Then Gamma:=Gammau+Betad;
Else Gamma:=Gammau+Alphad; EndIf;
Lam:=Det(Mat(Exp))/D;
Q:=2*(T[1]*T[2]+T[1]*T[3]+T[2]*T[3])-T[1]^2-T[2]^2-T[3]^2;
Return D*Q/4+(Lam)^2/(4*D)*(Delta(P,[Alpha/Lam,Beta/Lam,Gamma/Lam]))^2;
EndDefine;
APPENDIX B
Computer computations to Chapter 5
We give the computer computations needed for the argumentation in Chapter 5. We print
only those output lines that are necessary for the arguments.
B.1. Normal subgroups with GAP and rings of invariants with Singular
In this section we compute explicit generators for the normal subgroups of T, O and I
(with their embeddings into SL2 (k) as given in Chapter 5, whenever the generators are
defined in k2×2 ). Moreover, we use Singular to compute the invariants of k[x, y] under
these normal subgroups, where we work over the smallest finite extensions of Q we need
to define the generators over k. For these computations one needs the library finvar.
We compute the normal subgroups of T with GAP.
gap> o:=[[E(4),0],[0,-E(4)]];
[ [ E(4), 0 ], [ 0, -E(4) ] ]
gap> t:=[[0,1],[-1,0]];
[ [ 0, 1 ], [ -1, 0 ] ]
gap> w:=Sqrt(2);
E(8)-E(8)^3
gap> m:=[[E(8)^7/w,E(8)^7/w],[E(8)^5/w,E(8)/w]];
[ [ 1/2-1/2*E(4), 1/2-1/2*E(4) ], [ -1/2-1/2*E(4), 1/2+1/2*E(4) ] ]
gap> NormalSubgroups(Group(o,t,m));
[ Group([ [ [ E(4), 0 ], [ 0, -E(4) ] ], [ [ 0, 1 ], [ -1, 0 ] ],
[ [ 1/2-1/2*E(4), 1/2-1/2*E(4) ], [ -1/2-1/2*E(4), 1/2+1/2*E(4) ] ] ]),
Group([ [ [ 0, E(4) ], [ E(4), 0 ] ], [ [ E(4), 0 ], [ 0, -E(4) ] ],
[ [ -1, 0 ], [ 0, -1 ] ] ]),
Group([ [ [ -1, 0 ], [ 0, -1 ] ] ]), Group([ ]) ]
Then Singular gives the corresponding rings of invariants. The three output lines give
the primary invariants (prim), the secondary invariants (sec) and the indecomposable,
secondary invariants (irrsec). Then k[x, y]G is a finitely generated module over k[prim]
with module generators sec. As a k-algebra k[x, y]G is generated by prim and irrsec.
> ring R=(0,Q),(x,y),lp;minpoly=rootofUnity(8);
> number w=Q-Q^3;
> w;
(-Q3+Q)
> matrix o[2][2]=Q2,0,0,-Q2;
> matrix t[2][2]=0,1,-1,0;
> matrix m[2][2]=Q7/w,Q7/w,Q5/w,Q/w;
175
176
B. COMPUTER COMPUTATIONS TO CHAPTER 5
> print(invariant_ring(o,t,m));
x5y-xy5,x8+14*x4y4+y8
1,x12-33*x8y4-33*x4y8+y12
x12-33*x8y4-33*x4y8+y12
> print(invariant_ring(o,o*t));
x2y2,x4+y4
1,x5y-xy5
x5y-xy5
The normal subgroups of O are given by
gap> k:=[[E(8),0],[0,E(8)^7]];
[ [ E(8), 0 ], [ 0, -E(8)^3 ] ]
gap> NormalSubgroups(Group(o,t,m,k));
[ Group([ [ [ E(4), 0 ], [ 0, -E(4) ] ], [ [ 0, 1 ], [ -1, 0 ] ],
[ [ 1/2-1/2*E(4), 1/2-1/2*E(4) ], [ -1/2-1/2*E(4), 1/2+1/2*E(4) ] ],
[ [ E(8), 0 ], [ 0, -E(8)^3 ] ] ]),
Group([ [ [ -1/2-1/2*E(4), 1/2-1/2*E(4) ],
[ -1/2-1/2*E(4), -1/2+1/2*E(4) ] ],
[ [ 0, E(4) ], [ E(4), 0 ] ],
[ [ E(4), 0 ], [ 0, -E(4) ] ], [ [ -1, 0 ], [ 0, -1 ] ] ]),
Group([ [ [ 0, E(4) ], [ E(4), 0 ] ], [ [ E(4), 0 ], [ 0, -E(4) ] ],
[ [ -1, 0 ], [ 0, -1 ] ] ]),
Group([ [ [ -1, 0 ], [ 0, -1 ] ] ]), Group([ ]) ]
The rings of invariants are
> matrix k[2][2]=Q,0,0,Q7;
> print(invariant_ring(o,t,m,k));
x8+14*x4y4+y8,x10y2-2*x6y6+x2y10
1,x17y-34*x13y5+34*x5y13-xy17
x17y-34*x13y5+34*x5y13-xy17
> print(invariant_ring(o,o*t,k*k*k*k*k*k*k*m*k*o*o*o));
x5y-xy5,x8+14*x4y4+y8
1,x12-33*x8y4-33*x4y8+y12
x12-33*x8y4-33*x4y8+y12
We compute the normal subgroups of I.
gap> w:=Sqrt(5);
E(5)-E(5)^2-E(5)^3+E(5)^4
gap> o:=[[-E(5)^3,0],[0,-E(5)^2]];
[ [ -E(5)^3, 0 ], [ 0, -E(5)^2 ] ]
gap> t:=[[(-E(5)+E(5)^4)/w,(E(5)^2-E(5)^3)/w],
[(E(5)^2-E(5)^3)/w,(E(5)-E(5)^4)/w]];
[ [ -1/5*E(5)-2/5*E(5)^2+2/5*E(5)^3+1/5*E(5)^4,
2/5*E(5)-1/5*E(5)^2+1/5*E(5)^3-2/5*E(5)^4 ],
[ 2/5*E(5)-1/5*E(5)^2+1/5*E(5)^3-2/5*E(5)^4,
1/5*E(5)+2/5*E(5)^2-2/5*E(5)^3-1/5*E(5)^4 ] ]
gap> NormalSubgroups(Group(o,t));
[ Group([ ]), <group of 2x2 matrices of size 2 in characteristic 0>,
<group of 2x2 matrices of size 120 in characteristic 0> ]
B.2. MACAULAY2 COMPUTATIONS IN THE E6 -CASE
177
The ring of invariants is given by
> ring R=(0,Q),(x,y),lp;minpoly=rootofUnity(5);
> matrix o[2][2]=-Q3,0,0,-Q2;
> number w=Q-Q2-Q3+Q4;
> matrix t[2][2]=(-Q+Q4)/w,(Q2-Q3)/w,(Q2-Q3)/w,(Q-Q4)/w;
> print(invariant_ring(o,t));
x11y+11*x6y6-xy11,x20-228*x15y5+494*x10y10+228*x5y15+y20
1,x30+522*x25y5-10005*x20y10-10005*x10y20-522*x5y25+y30
x30+522*x25y5-10005*x20y10-10005*x10y20-522*x5y25+y30
B.2. Macaulay2 computations in the E6 -case
+ M2.exe --no-readline --print-width 79
Macaulay2, version 1.4
with packages: ConwayPolynomials, Elimination, IntegralClosure, LLLBases,
PrimaryDecomposition, ReesAlgebra, TangentCone
i1 : A=ZZ[U,V,W]/(U^2+4*V^3-V*W^2)
i2 : X=W^3-36*V^2*W
i3 : Y=W^2+12*V^2
i4 : Z=U
i5 : S1=mingens image syz gens ideal(X,Y,Z)
o5 = {3} | 0
-V
W
-U
|
{2} | -U
-2VW 12V2-W2 -3UW |
{1} | 12V2+W2 3UW 36UV
4W3 |
3
o5 : Matrix A
4
<--- A
i6 : S2=mingens image syz gens ideal(X,Y^2,Y*Z,Z^2)
o6 = {3}
{4}
{3}
{2}
|
|
|
|
0
0
-U
12V2+W2
0
-U
12V2+W2
0
4
o6 : Matrix A
2VW
V
3U
-6W2
12V2-W2
W
0
-72VW
UV
0
2VW
-3UW
UW
U
-2W2
36UV
-U2
0
-3UW
4W3
|
|
|
|
7
<--- A
i7 : N=mingens image syz gens ideal(X*Y,X*Z,Y^2,Y*Z^2,Z^3)
o7 = {5} | 0
-U
V
W
0
0
U
0
0
0
{4} | 0
12V2+W2 0
0
0
2VW -2W2 UV
UW
-U2
{4} | 0
0
2VW 12V2-W2 U2
UV
UW
0
0
0
{4} | -U
0
-3W 36V
-12V2-W2 3U
0
2VW 12V2-W2 -3UW
{3} | 12V2+W2 0
0
0
0
-6W2 -72VW -3UW 36UV
4W3
--------------------------------------------------------------------------U2W
|
3UW3
|
0
|
-2W4+27U2V |
178
B. COMPUTER COMPUTATIONS TO CHAPTER 5
81U3
|
5
o7 : Matrix A
11
<--- A
i8 : L=mingens image syz gens ideal(X,Y,Z^2)
o8 = {3} | V
W
0
U2
|
{2} | 2VW 12V2-W2 U2
0
|
{2} | -3W 36V
-12V2-W2 36V2W-W3 |
3
o8 : Matrix A
4
<--- A
B.3. Macaulay2 computations in the E8 -case
+ M2.exe --no-readline --print-width 79
Macaulay2, version 1.4
with packages: ConwayPolynomials, Elimination, IntegralClosure, LLLBases,
PrimaryDecomposition, ReesAlgebra, TangentCone
i1 : A=ZZ[R,S,T]/(T^2-R*S)
i2 : X=R^15+522*R^10*T^5-10005*R^5*T^10-10005*S^5*T^10-522*S^10*T^5+S^15
i3 : Y=R^10-228*R^5*T^5+494*T^10+228*S^5*T^5+S^10
i4 : Z=R^5*T-S^5*T+11*T^6
i5 : S1=mingens image syz gens ideal(X,Y,Z)
i6 : S2=mingens image syz gens ideal(X,Y^2,Y*Z,Z^2)
i7 : S3=mingens image syz gens ideal(X*Y,X*Z,Y^2,Y*Z^2,Z^3)
i8 : S4=mingens image syz gens ideal(X*Y,X*Z^2,Y^3,Y^2*Z,Y*Z^3,Z^4)
i9 : S5=mingens image syz gens ideal(X*Y^2,X*Y*Z^2,X*Z^4,Y^4,Y^3*Z,Y^2*Z^3,Z^5)
i10 : S6=mingens image syz gens ideal(X,Y^2,Y*Z,Z^3)
i11 : S7=mingens image syz gens ideal(X*Y,X*Z,Y^2,Y*Z^2,Z^4)
i12 : S8=mingens image syz gens ideal(X,Y,Z^2)
i13 : B=QQ[R,S,T]/(T^2-R*S)
i14 : T1=map(B,A)**S1
i15 : T2=map(B,A)**S2
i16 : T3=map(B,A)**S3
i17 : T4=map(B,A)**S4
i18 : T5=map(B,A)**S5
B.3. MACAULAY2 COMPUTATIONS IN THE E8 -CASE
179
i19 : T6=map(B,A)**S6
i20 : T7=map(B,A)**S7
i21 : T8=map(B,A)**S8
i22 : M1=mingens image T1
o22 = {15} | R
-S
{10} | -R6+11S4T2-66RT5
S6-11R4T2-66ST5
{6} | 12S9T-684R6T4+2052S4T6+2964RT9 12R9T+684S6T4-2052R4T6+2964ST9
------------------------------------------------------------------------T
-T
|
-R5T+11S5T-66T6
-11R5T+S5T-66T6
|
12S10-684R5T5+2052S5T5+2964T10 12R10-2052R5T5+684S5T5+2964T10 |
3
o22 : Matrix B
4
<--- B
i23 : M2=mingens image T2
o23 = {15} | -5R5T-5S5T
R6-11S4T2+66RT5
{20} | -T
-R
{16} | 6R5-6S5+216T5
60S4T-180RT4
{12} | 3240R5T4-3240S5T4-9360T9 72S9-864R6T3+9072S4T5+8424RT8
------------------------------------------------------------------------S6-11R4T2-66ST5
|
-S
|
-60R4T-180ST4
|
72R9+864S6T3-9072R4T5+8424ST8 |
4
o23 : Matrix B
3
<--- B
i24 : M3=mingens image T3
o24 = {25} | -9T2
-9T2
{21} | -4R5T+104S5T+756T6
-104R5T+4S5T+756T6
{20} | -35R5T2-55S5T2+110T7 -55R5T2-35S5T2-110T7
{22} | 48R5+168S5-2160T5
168R5+48S5+2160T5
{18} | 64800S5T4+93600T9
64800R5T4-93600T9
------------------------------------------------------------------------ST
RT
4S6+16R4T2-84ST5
-4R6-16S4T2-84RT5
3S6T+7R4T3+22ST6
3R6T+7S4T3-22RT6
-24R4T-432ST4
-24S4T+432RT4
2880S6T3-10080R4T5+18720ST8 2880R6T3-10080S4T5-18720RT8
------------------------------------------------------------------------R2
-S2
-60S3T3+180R2T4
-60R3T3-180S2T4
-R7+11S3T4-66R2T5
S7-11R3T4-66S2T5
216S3T2-288R2T3
-216R3T2-288S2T3
288S8T-576R7T2+26208S3T6+14976R2T7 288R8T+576S7T2-26208R3T6+14976S2T7
------------------------------------------------------------------------3RT
-3ST
|
180
B. COMPUTER COMPUTATIONS TO CHAPTER 5
-2R6-158S4T2+408RT5
-R6T+31S4T3-176RT6
528S4T-504RT4
720S9+60480S4T5+28080RT8
5
o24 : Matrix B
-2S6-158R4T2-408ST5
S6T-31R4T3-176ST6
-528R4T-504ST4
720R9-60480R4T5+28080ST8
|
|
|
|
8
<--- B
i25 : M4=mingens image T4
o25 = {25} | 5R5T2+5S5T2
2S6T+8R4T3+33ST6
{27} | -12R5-12S5
36R4T-252ST4
{30} | T2
ST
{26} | 6R5T-6S5T-84T6
3S6+27R4T2-18ST5
{28} | 720T4
-72R4-576ST3
{24} | 7776R5T3-7776S5T3-22464T8 864S6T2-22464R4T4+33696ST7
-------------------------------------------------------------------------2R6T-8S4T3+33RT6
R7-11S3T4+66R2T5
-36S4T-252RT4
-216S3T2+288R2T3
-RT
-R2
3R6+27S4T2+18RT5
60S3T3-180R2T4
-72S4+576RT3
576S3T-288R2T2
864R6T2-22464S4T4-33696RT7 864S8-864R7T+56160S3T5+11232R2T6
------------------------------------------------------------------------S7-11R3T4-66S2T5
|
-216R3T2-288S2T3
|
-S2
|
-60R3T3-180S2T4
|
-576R3T-288S2T2
|
864R8+864S7T-56160R3T5+11232S2T6 |
6
o25 : Matrix B
5
<--- B
i26 : M5=mingens image T5
o26 = {35} | S8+3R7T-39S3T5-143R2T6 R8-3S7T+39R3T5-143S2T6 25R5T3+25S5T3
{37} | 0
0
0
{39} | 0
0
0
{40} | -S3+3R2T
-R3-3S2T
5T3
{36} | -6R7-294S3T4+84R2T5
-6S7-294R3T4-84S2T5
42R5T2-42S5T2-288T7
{38} | -720S3T2-1440R2T3
-720R3T2+1440S2T3
-72R5+72S5+1008T5
{30} | 0
0
0
------------------------------------------------------------------------9R6T2+41S4T4-176RT7
-6R6+306S4T2+684RT5
-720S4+1440RT3
5RT2
-8R6T-152S4T3+42RT6
-216S4T-1512RT4
509760S9T4-976320R6T7+5088960S4T9+2790720RT12
------------------------------------------------------------------------9S6T2+41R4T4+176ST7
-6S6+306R4T2-684ST5
-720R4-1440ST3
5ST2
B.3. MACAULAY2 COMPUTATIONS IN THE E8 -CASE
181
8S6T+152R4T3+42ST6
216R4T-1512ST4
509760R9T4+976320S6T7-5088960R4T9+2790720ST12
------------------------------------------------------------------------50S5T3-275T8
-150R5T+150S5T+2100T6
-18000T4
5T3
-83R5T2-167S5T2-288T7
228R5+372S5+1008T5
108000R10T3-108000S10T3-12312000R5T8-12312000S5T8
------------------------------------------------------------------------574S6T2+976R4T4+2211ST7
-1071S6+1521R4T2-15894ST5
-1080R4+125280ST3
155ST2
-342S6T+2352R4T3-8433ST6
-3924R4T+27468ST4
764640S11T2+88633440S6T7+79535520R4T9+4186080ST12
------------------------------------------------------------------------574R6T2+976S4T4-2211RT7
-1071R6+1521S4T2+15894RT5
-1080S4-125280RT3
155RT2
342R6T-2352S4T3-8433RT6
3924S4T+27468RT4
764640R11T2-88633440R6T7-79535520S4T9+4186080RT12
-------------------------------------------------------------------------7S7T-43R3T5-198S2T6
-360R3T3+720S2T4
864R3T+2592S2T2
-5S2T
-12S7-168R3T4-108S2T5
-288R3T2+2016S2T3
5184S12T-616896R8T5-580608S7T6+6697728R3T10-3348864S2T11
-------------------------------------------------------------------------7R7T-43S3T5+198R2T6
-360S3T3-720R2T4
864S3T-2592R2T2
-5R2T
12R7+168S3T4-108R2T5
288S3T2+2016R2T3
5184R12T-616896S8T5+580608R7T6-6697728S3T10-3348864R2T11
------------------------------------------------------------------------R7T+49S3T5-264R2T6
720S3T3-360R2T4
2592S3T+864R2T2
5R2T
-6R7-234S3T4+504R2T5
-1584S3T2-288R2T3
5184S13-580608S8T5-616896R7T6-3348864S3T10+6697728R2T11
-------------------------------------------------------------------------S7T-49R3T5-264S2T6
|
-720R3T3-360S2T4
|
-2592R3T+864S2T2
|
-5S2T
|
182
B. COMPUTER COMPUTATIONS TO CHAPTER 5
-6S7-234R3T4-504S2T5
|
-1584R3T2+288S2T3
|
5184R13+580608R8T5-616896S7T6-3348864R3T10-6697728S2T11 |
7
o26 : Matrix B
12
<--- B
i27 : M6=mingens image T6
o27 = {15} | S8+3R7T-39S3T5-143R2T6 R8-3S7T+39R3T5-143S2T6 25R5T3+25S5T3
{20} | -S3+3R2T
-R3-3S2T
5T3
{16} | -6R7-294S3T4+84R2T5
-6S7-294R3T4-84S2T5
42R5T2-42S5T2-288T7
{18} | -720S3T2-1440R2T3
-720R3T2+1440S2T3
-72R5+72S5+1008T5
------------------------------------------------------------------------|
|
|
|
4
o27 : Matrix B
3
<--- B
i28 : M7=mingens image T7
o28 = {25} | -S3-2R2T
-R3+2S2T
5RT2
{21} | 4R7-104S3T4-156R2T5
4S7-104R3T4+156S2T5
4R6T-104S4T3-156RT6
{20} | S8-2R7T-84S3T5+77R2T6 R8+2S7T+84R3T5+77S2T6 15R6T2+35S4T4-110RT7
{22} | -480S3T3-360R2T4
480R3T3-360S2T4
-24R6+144S4T2+576RT5
{24} | -288S3T-2016R2T2
288R3T-2016S2T2
-288S4-2016RT3
------------------------------------------------------------------------5ST2
|
-4S6T+104R4T3-156ST6 |
15S6T2+35R4T4+110ST7 |
-24S6+144R4T2-576ST5 |
-288R4+2016ST3
|
5
o28 : Matrix B
4
<--- B
i29 : M8=mingens image T8
o29 = {15} | -R4+7ST3
-S4-7RT3
{10} | R9+17S6T3+119R4T5+187ST8 S9-17R6T3-119S4T5+187RT8
{12} | -24S6T+624R4T3-936ST6
24R6T-624S4T3-936RT6
-------------------------------------------------------------------------R3T+7S2T2
-S3T-7R2T2
|
R8T+17S7T2+119R3T6+187S2T7 S8T-17R7T2-119S3T6+187R2T7 |
-24S7+624R3T4-936S2T5
24R7-624S3T4-936R2T5
|
3
o29 : Matrix B
4
<--- B
i30 : inducedMap(image M1,image T1)
o30 = {16} | 1 0 0 0
{16} | 0 1 0 0
|
|
B.3. MACAULAY2 COMPUTATIONS IN THE E8 -CASE
183
{16} | 0 0 1 1/12 |
{16} | 0 0 0 1/12 |
o30 : Matrix
i31 : inducedMap(image M2,image T2)
o31 = {21} | 1 0 0 -1/6R -1/6S -1/6T -1/36T |
{21} | 0 1 0 1/6T 0
1/6S 1/72S |
{21} | 0 0 1 0
1/6T 0
1/72R |
o31 : Matrix
i32 : inducedMap(image M3,image T3)
o32 = {27}
{27}
{27}
{27}
{27}
{27}
{27}
{27}
|
|
|
|
|
|
|
|
1
0
0
0
0
0
0
0
19/180
1/180
0
0
0
0
0
0
0
0
1
0
0
0
0
0
0
0
0
1
0
0
0
0
0
0
0
0
1
0
0
0
0
0
0
0
0
1
0
0
0
0
0
-3/10
0
0
1/10
0
0
0
3/10
0
0
0
0
1/10
17/48R
0
0
11/80T
0
0
1/60T
0
-11/168S
0
409/280T
0
0
0
0
1/60T
409/1800T
1/1800T
0
0
0
0
1/60S
0
97/900T
1/300T
0
0
0
0
1/720S
1/720R
|
|
|
|
|
|
|
|
o32 : Matrix
i33 : inducedMap(image M4,image T4)
o33 = {32} | 1 0 0 0 0 -5/3T 5/6T
-1/3S 1/3R 0
0
-1/12R -1/12S
{32} | 0 1 0 0 0 -2/3R 1/4R
-8/3T 0
0
-2/3S 0
-5/6T
{32} | 0 0 1 0 0 -7/3S -11/12S 0
-8/3T 2/3R 0
5/6T
0
{32} | 0 0 0 1 0 0
0
0
0
1/3T 0
1/12S 0
{32} | 0 0 0 0 1 0
0
0
0
0
1/3T 0
1/12R
------------------------------------------------------------------------23/24RT -29/24ST -3/8T2
167/144T2 RT2
9/2R3T-13/2S2T2 |
25/72R2 -3/8T2
-11/72RT 91/216RT 5/3S3+1/3R2T 3/2R4+227/6ST3
|
7/24T2 29/72S2 -13/24ST 125/216ST 14/3T3
13/6S3T+61/6R2T2 |
1/72ST 0
1/72S2
1/864S2
5/3ST2
-85/6RT3
|
0
1/72RT
0
1/864R2
-10/3S2T
-155/6T4
|
o33 : Matrix
i34 : inducedMap(image M5,image T5)
o34 = {43}
{43}
{43}
{43}
{43}
{43}
{43}
{43}
{43}
{43}
{43}
{43}
|
|
|
|
|
|
|
|
|
|
|
|
1
0
0
0
0
0
0
0
0
0
0
0
0
1
0
0
0
0
0
0
0
0
0
0
0
0
1
0
0
0
0
0
0
0
0
0
0
0
0
1
0
0
0
0
0
0
0
0
0
0
0
0
1
0
0
0
0
0
0
0
0
0
-1/25
0
0
1/25
0
0
0
0
0
0
0
0
0
0
-31/295
0
1/295
0
0
0
0
0
0
0
0
-31/295
0
0
0
1/295
0
0
0
0
0
0
0
0
0
0
0
0
1
0
0
0
0
0
0
0
0
0
0
0
0
1
0
0
0
0
0
0
0
0
0
0
0
1/6
1/6
0
0
0
0
0
0
0
0
0
-1/6
0
0
1/6
3/2T
1/2S
0
0
0
0
0
0
0
0
0
0
1/2R
-3/2T
0
0
0
0
0
0
0
0
0
0
0
0
59/375T
-1/30S
1/30R
-59/375T
0
0
0
0
0
0
184
B. COMPUTER COMPUTATIONS TO CHAPTER 5
------------------------------------------------------------------------0
0
0
0
-T
T
-1/3R 0
0
0
0
0
0
-1/3S 1/3S T
0
0
-1/2S -41/150T 1/18S
1/18R
0
0
0
-13/72R 13/72S
0
29/120S 0
-119/1770T 1/2R -1/3R 0
1/6T
0
1/2T -1/120R -119/1770T 0
0
0
1/3S
0
-1/6T
0
1/25T
0
0
0
0
0
0
0
0
0
1/2655T
0
0
0
0
0
0
0
0
0
1/2655T
0
0
0
0
0
0
0
0
0
0
0
-1/18T 0
0
0
0
0
0
1/6T 1/18T 0
0
0
0
0
0
0
0
1/18T 0
1/72S
0
0
0
0
0
0
0
1/18T 0
1/72R
------------------------------------------------------------------------ST
0
0
0
-20T3
1/3S2
0
0
0
35/6ST2
-47/720RT
19/144ST
-7841/54000T2 53/540T2
3/2S3+13/6R2T
-10139/42480T2 1/144S2
151/720ST
-181/1440ST 0
1/720R2
233/42480T2 13/720RT
-11/1440RT 0
0
0
-191/2250T2
19/540T2
0
0
-73/15930T2 0
0
0
-17/15930T2
0
0
0
0
0
0
0
0
0
0
0
0
0
0
1/432ST
0
1/432S2
1/5184S2
0
0
1/432RT
0
1/5184R2
0
-------------------------------------------------------------------------2R3T-7/2S2T2
893/72T4
-823/48RT3
|
10/3S3T-15/2R2T2 -823/48ST3
-893/72T4
|
109/15T4
-233/720S3T-773/720R2T2 233/720R3T-773/720S2T2 |
3/2R4-3ST3
11/80S4+1171/80RT3
-37/240S3T+1637/240R2T2 |
3/2S4+3RT3
-37/240R3T-1637/240S2T2 -11/80R4+1171/80ST3
|
-3/5T4
0
0
|
0
0
0
|
0
0
0
|
0
0
2183/144T4
|
0
2183/144T4
0
|
0
-649/48T4
0
|
0
0
649/48T4
|
o34 : Matrix
i35 : inducedMap(image M6,image T6)
o35 = {23} | 1 0 0 3/2T 1/2R -5/6R2 -5/2ST -5/6RT -5/2T2 -20T3
{23} | 0 1 0 1/2S -3/2T 5/2RT -5/6S2 5/2T2 -5/6ST 35/6ST2
{23} | 0 0 1 0
0
1/3ST 1/3RT 1/3S2 1/3R2 3/2S3+13/6R2T
-------------------------------------------------------------------------35/6RT2
7/72R3T2+23/36S2T3
|
-20T3
7/72S3T2-23/36R2T3
|
3/2R3-13/6S2T 1/72R5-1/72S5+53/36T5 |
o35 : Matrix
i36 : inducedMap(image M7,image T7)
B.3. MACAULAY2 COMPUTATIONS IN THE E8 -CASE
185
o36 = {28} | 1 0 0 0 -3/2T 3/4R 0
27/8T2 -39/8RT 1/4R2 5/24R2T
{28} | 0 1 0 0 3/4S 3/2T 0
39/8ST 27/8T2 1/2RT -5/24RT2
{28} | 0 0 1 0 -1/4R 0
7/4S 1/8RT -7/8R2 1/2S2 1/24R3-1/12S2T
{28} | 0 0 0 1 0
-1/4S -3/4R -7/8S2 -1/8ST -5/4T2 5/24T3
------------------------------------------------------------------------5/24ST2
5/24RT2
5/24T3
-5/8T4
5/24S2T
-5/24T3
5/24ST2
-5/6ST3
5/24T3
-1/12S3+1/24R2T 5/24RT2
-1/24S4-17/24RT3
-1/24S3-1/12R2T 5/24ST2
-1/12R3-1/24S2T 7/24R3T+1/4S2T2
------------------------------------------------------------------------5/6RT3
-5/72R3T3+5/96S2T4
|
-5/8T4
5/72S3T3+5/96R2T4
|
-7/24S3T+1/4R2T2 1/288S6-1/48R4T2+1/12ST5 |
-1/24R4+17/24ST3 1/288R6-1/48S4T2-1/12RT5 |
o36 : Matrix
i37 : inducedMap(image M8,image T8)
o37 = {19}
{19}
{19}
{19}
|
|
|
|
1
0
0
0
0
1
0
0
0
0
1
0
0
0
0
1
0
0
1/2S
-17/2T
0
0
17/2T
1/2R
-493/24RT2
-3757/24ST2
221/24S3
-29/24R3
|
|
|
|
o37 : Matrix
i38 : C=ZZ/(7)[R,S,T]/(T^2-R*S)
i39 : T3=map(C,A)**S3
i40 : M3=mingens image T3
o40 = {25} | -2T2
-2T2
-2ST
-2RT
R2
{21} | 3R5T-S5T R5T-3S5T -S6+3R4T2 R6-3S4T2 3S3T3-2R2T4
{20} | S5T2-2T7 R5T2+2T7 S6T-2ST6 R6T+2RT6 -R7-3S3T4-3R2T5
{22} | -R5+3T5 -S5-3T5 -R4T+3ST4 -S4T-3RT4 -S3T2-R2T3
{18} | S5T4+3T9 R5T4-3T9 S6T3+3ST8 R6T3-3RT8 S8T-2R7T2+3R2T7
-------------------------------------------------------------------------S2
-3RT
3ST
|
3R3T3+2S2T4
2R6-3S4T2-2RT5 2S6-3R4T2+2ST5 |
S7+3R3T4-3S2T5 R6T-3S4T3+RT6 -S6T+3R4T3+ST6 |
R3T2-S2T3
-3S4T
3R4T
|
R8T+2S7T2+3S2T7 S9-3RT8
R9-3ST8
|
5
o40 : Matrix C
8
<--- C
i41 : D=ZZ/(59)[R,S,T]/(T^2-R*S)
i42 : T5=map(D,A)**S5
i43 : M5=mingens image T5
o43 = {35} | S8+3R7T+20S3T5-25R2T6 R8-3S7T-20R3T5-25S2T6 R5T3+S5T3
{37} | 0
0
0
{39} | 0
0
0
186
B. COMPUTER COMPUTATIONS TO CHAPTER 5
{40} | -S3+3R2T
-R3-3S2T
12T3
{36} | -6R7+S3T4+25R2T5
-6S7+R3T4-25S2T5
-29R5T2+29S5T2+5T7
{38} | -12S3T2-24R2T3
-12R3T2+24S2T3
16R5-16S5+12T5
{30} | 0
0
0
------------------------------------------------------------------------14S6T2-28R4T4+5ST7 -14R6T2+28S4T4+5RT7 -18S5T3-19T8
-29S6+4R4T2-2ST5
29R6-4S4T2-2RT5
-5R5T+5S5T+11T6
R4+2ST3
-S4+2RT3
-10T4
-25ST2
25RT2
10T3
19S6T+7R4T3+26ST6 19R6T+7S4T3-26RT6
11R5T2+20S5T2+14T7
-18R4T+8ST4
-18S4T-8RT4
-16R5-23S5+10T5
S6T7-12R4T9+18ST12 R6T7-12S4T9-18RT12 R10T3-S10T3+4R5T8+4S5T8
-------------------------------------------------------------------------2S6T2-11R4T4-20ST7
-2R6T2-11S4T4+20RT7
-3S6+21R4T2-19ST5
-3R6+21S4T2+19RT5
4R4+18ST3
4S4-18RT3
-19ST2
-19RT2
-24S6T-21R4T3+5ST6
24R6T+21S4T3+5RT6
26R4T-5ST4
-26S4T-5RT4
S11T2-R9T4+7R4T9+13ST12 R11T2-S9T4-7S4T9+13RT12
------------------------------------------------------------------------23S7T-2R3T5+10S2T6
23R7T-2S3T5-10R2T6
-14R3T3+28S2T4
-14S3T3-28R2T4
10R3T-29S2T2
10S3T+29R2T2
8S2T
8R2T
-28S7+21R3T4-16S2T5
28R7-21S3T4-16R2T5
-23R3T2-16S2T3
23S3T2-16R2T3
S12T-R8T5+6S7T6-6R3T10+3S2T11 R12T-S8T5-6R7T6+6S3T10+3R2T11
------------------------------------------------------------------------22R7T+16S3T5-26R2T6
-22S7T-16R3T5-26S2T6
|
28S3T3-14R2T4
-28R3T3-14S2T4
|
-29S3T+10R2T2
29R3T+10S2T2
|
-8R2T
8S2T
|
-14R7-15S3T4-4R2T5
-14S7-15R3T4+4S2T5
|
21S3T2-23R2T3
21R3T2+23S2T3
|
S13+6S8T5-R7T6+3S3T10-6R2T11 R13-6R8T5-S7T6+3R3T10+6S2T11 |
7
o43 : Matrix D
12
<--- D
Bibliography
[AE05] I. M. Aberbach and F. Enescu. The structure of F-pure rings. Math. Z.,
240(4):791–806, 2005.
[AE08] I. M. Aberbach and F. Enescu. Lower Bounds for Hilbert-Kunz
Multiplicities in Local Rings of Fixed Dimension. Michigan Math. J.,
57:1–16, 2008.
[AE12] I. M. Aberbach and F. Enescu. New estimates of Hilbert-Kunz multiplicities
for local rings of fixed dimension. to appear in Nagoya Math. J.,
arXiv:1101.5078v4, 2012.
[AL03] I. M. Aberbach and G. J. Leuschke. The F-signature and strong F-regularity.
Math. Res. Lett., 10(1):51–56, 2003.
[AR89] M. Auslander and I. Reiten. Cohen-Macaulay modules for graded
Cohen-Macaulay rings and their completions. In Commutative Algebra,
Lecture Notes in Math., pages 21–31. Springer, 1989.
[Art66] M. Artin. On isolated rational singularities of surfaces. Amer. J. Math.,
88:129–136, 1966.
[Art77] M. Artin. Coverings of the Rational Double Points in Characteristic p. In
W. Baily and T. Shioda, editors, Complex analysis and algebraic geometry,
pages 11–22, Cambridge, 1977. Cambridge University Press.
[BC97] R.-O. Buchweitz and Q. Chen. Hilbert-Kunz functions of cubic curves and
surfaces. J. Algebra, 197(1):246–267, 1997.
[BCP97] R.-O. Buchweitz, Q. Chen, and K. Pardue. Hilbert-Kunz functions. preprint,
1997.
[BD08] I. Burban and Y. Drozd. Maximal Cohen-Macaulay modules over surface
singularities. In Trends in representation theory of algebras and related
topics, EMS Ser. Congr. Rep., pages 101–166, Zürich, 2008. EMS.
[BE04] M. Blickle and F. Enescu. On rings with small Hilbert-Kunz multiplicity.
Proc. AMS, 132(9):2505–2509, 2004.
[BGS87] R.-O. Buchweitz, G.-M. Greuel, and F.-O. Schreyer. Cohen-Macaulay
modules on hypersurface singularities II. Invent. Math., 88:165–182, 1987.
[BH98] W. Bruns and J. Herzog. Cohen-Macaulay rings, volume 39 of Cambridge
Studies in Advanced Mathematics. Cambridge University Press, Cambridge,
UK, 1998.
[BH06] H. Brenner and G. Hein. Restriction of the cotangent bundle to elliptic
curves and Hilbert-Kunz functions. Manuscripta Math., 119(1):17–36, 2006.
[BK05] M. Brion and S. Kumar. Frobenius splitting methods in geometry and
representation theory. Number 231 in Progress in Mathematics. Birkhäuser,
187
188
Bibliography
[BK13]
[BLM12]
[Bou89]
[Bre]
[Bre03]
[Bre04]
[Bre05a]
[Bre05b]
[Bre06]
[Bre07]
[Bre08]
[Bre12]
[Bre13]
[Bri68]
[BST12a]
[BST12b]
[Bur12]
[Can12]
Boston, 2005.
H. Brenner and A. Kaid. An explicit example of Frobenius periodicity. J.
Pure Appl. Algebra, 217(8):1412–1420, 2013.
H. Brenner, J. Li, and C. Miller. A direct limit for limit Hilbert-Kunz
multiplicity for smooth projective curves. J. Algebra, 372:488–504, 2012.
N. Bourbaki. Commutative algebra, Chapters 1-7. Springer, Berlin, 1989.
H. Brenner. A Hilbert-Kunz proof of a theorem in invariant theory. in
preparation.
H. Brenner. The theory of tight closure from the viewpoint of vector
bundles, 2003. Habilitationsschrift, Ruhr-Universität Bochum.
H. Brenner. A characteristic zero Hilbert-Kunz criterion for solid closure in
dimension two. Math. Res. Lett., 11:563–574, 2004.
H. Brenner. Computing the tight closure in dimension two. Math. Comp.,
74(251):1495–1518, 2005.
H. Brenner. On a problem of Miyaoka. In B. Moonen, R. Shoof, and
G. Geer, editors, Number fields and function fields - two parallel worlds,
number 239 in Progress in Mathematics, pages 51–59. Birkhäuser Boston,
2005.
H. Brenner. The rationality of the Hilbert-Kunz multiplicity in graded
dimension two. Math. Ann., 334(1):91–110, 2006.
H. Brenner. The Hilbert-Kunz function in graded dimension two. Comm.
Algebra, 35(10):3199–3213, 2007.
H. Brenner. Tight closure and vector bundles. In Three lectures on
commutative algebra, volume 42 of Univ. Lecture Ser., pages 1–71. AMS,
Providence, RI, 2008.
H. Brenner. Invariantentheorie, 2012. Lecture held at the Universität
Osnabrück in the winter term 2012. Available under
http://de.wikiversity.org/wiki/Kurs:
Invariantentheorie_(Osnabrück_2012-2013).
H. Brenner. Irrational Hilbert-Kunz multiplicities. preprint,
arXiv:1305.5873, 2013.
E. Brieskorn. Rationale Singularitäten komplexer Flächen. Invent. Math.,
4:336–358, 1968.
M. Blickle, K. Schwede, and K. Tucker. F-Signature of pairs and the
asymptotic behavior of Frobenius splittings. Adv. Math., 231(6):3232–3258,
2012.
M. Blickle, K. Schwede, and K. Tucker. F-Signature of pairs: Continuity,
p-fractals and minimal log discrepancies. to appear in J. LMS,
arXiv:1111.2762v2, 2012.
I. Burban. Frobenius morphism and vector bundles on cycles of projective
lines. Comm. Algebra, 40(8):2983–2988, 2012.
E. Canton. Relating F-signature and F-splitting ratio of pairs using
left-derivatives. preprint, arXiv:1206.4293v4, 2012.
Bibliography
189
[CDHZ12] O. Celikbas, H. Dao, C. Huneke, and Y. Zhang. Bounds on the Hilbert-Kunz
multiplicity. Nagoya Math. J., 205:149–165, 2012.
[CH98a] L. Chiang and Y. C. Hung. On Hilbert-Kunz function and representation
ring. Bull. Inst. Math. Acad. Sinica, 26(1):1–32, 1998.
[CH98b] L. Chiang and Y. C. Hung. On Hilbert-Kunz functions of some
hypersurfaces. J. Algebra, 199:499–527, 1998.
[CH03] S. Choi and S. Hong. The Hilbert-Kunz multiplicity of two-dimensional
toric rings. J. Korean Math. Soc., 40(2):169–177, 2003.
[Cha93] S. T. Chang. The asymptotic behaviour of Hilbert-Kunz functions and their
generalizations. PhD thesis, University of Michigan, 1993.
[Che98] Q. Chen. Hilbert-Kunz multiplicity of plane curves and a conjecture of K.
Pardue. PhD thesis, University of Toronto, 1998.
[CK12] C-Y. Jean Chan and K. Kurano. The cone spanned by maximal
Cohen-Macaulay modules and an application. preprint, arXiv:1211.4016v1,
2012.
[CK13] C-Y. Jean Chan and K. Kurano. Hilbert-Kunz functions over rings regular in
codimension one. preprint, arXiv:1301.5278v1, 2013.
[CoC] CoCoATeam. CoCoA: a system for doing Computations in Commutative
Algebra. Available at http://cocoa.dima.unige.it.
[Con95] M. Contessa. On the Hilbert-Kunz function and Koszul homology. J.
Algebra, 175:757–766, 1995.
[Con96] A. Conca. Hilbert-Kunz function of monomial ideals and binomial
hypersurfaces. Manuscripta Math., 90:287–300, 1996.
[DGPS12] W. Decker, G.-M. Greuel, G. Pfister, and H. Schönemann. Singular 3-1-5 —
A computer algebra system for polynomial computations. 2012.
http://www.singular.uni-kl.de.
[DI09] I. V. Dolgachev and V. A. Iskovskikh. Finite subgroups of the plane Cremona
group. In Algebra, arithmetic, and geometry: in honor of Yu. I. Manin, Vol.
I, volume 269 of Progr. Math., pages 443–548. Birkhäuser Boston, 2009.
[Dol82] I. Dolgachev. Weighted projective varieties. In Group actions and vector
fields, volume 956 of Lecture Notes in Math., pages 34–71, Berlin, 1982.
Springer.
[Dor71] L. Dornhoff. Group representation theory. Part A: Ordinary representation
theory, volume 7 of Pure and Applied Mathematics. Dekker, New York,
1971.
[DS13] H. Dao and I. Smirnov. On generalized Hilbert-Kunz function and
multiplicity. preprint, arXiv:1305.1833, 2013.
[Dut83] S. P. Dutta. Frobenius and multiplicities. J. Algebra, 85:424–448, 1983.
[Dut89] S. P. Dutta. Ext and Frobenius. J. Algebra, 127:163–177, 1989.
[Dut96] S. P. Dutta. Ext and Frobenius, II. J. Algebra, 186:724–735, 1996.
[dV34a] P. du Val. On isolated singularities of surfaces which do not affect the
conditions of adjunction. I. Proc. Camb. Philos. Soc., 30(4):453–459, 1934.
[dV34b] P. du Val. On isolated singularities of surfaces which do not affect the
conditions of adjunction. II. Proc. Camb. Philos. Soc., 30(4):460–465, 1934.
190
Bibliography
[dV34c] P. du Val. On isolated singularities of surfaces which do not affect the
conditions of adjunction. III. Proc. Camb. Philos. Soc., 30(4):483–491,
1934.
[Eis80] D. Eisenbud. Homological algebra on a complete intersection, with an
application to group representations. Trans. Amer. Math. Soc., 260(1):35–64,
1980.
[ES05] F. Enescu and K. Shimomoto. On the upper semi-continuity of the
Hilbert-Kunz multiplicity. J. Algebra, 285(1):222–237, 2005.
[Eto00] K. Eto. Hilbert-Kunz multiplicity of the Segre product of polynomial rings.
Rep. Res. Nippon Inst. Tech., 30:317–327, 2000.
[Eto02] K. Eto. Multiplicity and Hilbert-Kunz multiplicity of monoid rings. Tokyo J.
Math., 25(2):241–245, 2002.
[EW03] K. Eto and K.-i. Watanabe. Notes on Hilbert-Kunz multiplicity of Rees
algebras. Comm. Algebra, 31(12):5943–5976, 2003.
[EY] N. Epstein and Y. Yao. Unmixed Hilbert-Kunz multiplicity. in preparation.
[EY11] N. Epstein and Y. Yao. Some extensions of Hilbert-Kunz multiplicity.
preprint, arXiv:1103.4730v1, 2011.
[Fie00] J. Bruce Fields. Length functions determined by killing powers of several
ideals in a local ring. PhD thesis, University of Michigan, 2000.
[FT03] N. Fakhruddin and V. Trivedi. Hilbert-Kunz functions and multiplicities for
full flag varieties and elliptic curves. J. Pure Appl. Algebra, 181(1):23–52,
2003.
[GAP12] The GAP Group. GAP – Groups, Algorithms, and Programming, Version
4.5.6, 2012.
[GD61] A. Grothendieck and J. Dieudonné. Eléments de géométrie algébrique II.
Inst. Hautes Études Sci. Publ. Math., 8:5–222, 1961.
[GM10] I. M. Gessel and P. Monsky. The limit as p -> infinity of the Hilbert-Kunz
P
multiplicity of xidi . preprint, arXiv:1007.2004v1, 2010.
[GS] Daniel R. Grayson and Michael E. Stillman. Macaulay2, a software system
for research in algebraic geometry. Available at
http://www.math.uiuc.edu/Macaulay2/.
[Han91] C. Han. The Hilbert-Kunz function of a diagonal hypersurface. PhD thesis,
Brandeis University, 1991.
[Har87] R. Hartshorne. Algebraic geometry, volume 52 of Graduate Texts in
Mathematics. Springer, New York, 1987.
[HH90] M. Hochster and C. Huneke. Tight closure, invariant theory and the
Briançon-Skoda theorem. J. Amer. Math. Soc., 3:31–116, 1990.
[HL97] D. Huybrechts and M. Lehn. The geometry of moduli spaces of sheaves,
volume E31 of Aspects of Mathematics. Vieweg, Braunschweig, 1997.
[HL02] C. Huneke and G. J. Leuschke. Two theorems about maximal
Cohen-Macaulay modules. Math. Ann., 324(2):391–404, 2002.
[HM93] C. Han and P. Monsky. Some surprising Hilbert-Kunz functions. Math. Z.,
214:119–135, 1993.
Bibliography
191
[HMM04] C. Huneke, M. A. McDermott, and P. Monsky. Hilbert-Kunz functions for
normal rings. Math. Res. Lett., 11:539–546, 2004.
[HN75] G. Harder and M. S. Narasimham. On the cohomology groups of moduli
spaces of vector bundles on curves. Math. Ann., 212:215–248, 1975.
[Hoc94] M. Hochster. Solid closure. Contemp. Math., 159:103–172, 1994.
[Hun96] C. Huneke. Tight closure and its applications. AMS, Providence, RI, 1996.
[HY02] C. Huneke and Y. Yao. Unmixed local rings with minimal Hilbert-Kunz
multiplicity are regular. Proc. AMS, 130:661–665, 2002.
[HY09] M. Hochster and Y. Yao. Second coefficients of Hilbert-Kunz functions for
domains. preprint, available on the second author’s homepage, 2009.
[Kai09] A. Kaid. On semistable and strongly semistable syzygy bundles. PhD thesis,
University of Sheffield, 2009.
[Kar09] N. Karroum. MCM-einfache Moduln. PhD thesis, Ruhr-Universität
Bochum, 2009.
[Kle86] F. Klein. Vorlesungen über das Ikosaeder und die Auflösung der
Gleichungen vom 5ten Grade. Teubner, 1886.
[Knö87] H. Knörrer. Cohen-Macaulay modules on hypersurface singularities I.
Invent. Math., 88:153–164, 1987.
[Kor11] M. Von Korff. F-Signature of affine toric varieties. preprint,
arXiv:1110.0552v2, 2011.
[KR05] M. Kreuzer and L. Robbiano. Computational Commutative Algebra 2.
Springer, Berlin, 2005.
[Kre07] M. Kreuzer. Computing Hilbert-Kunz functions of 1-dimensional graded
rings. Univ. Iagellonicae Acta Math., 45:81–95, 2007.
[KRV12] A. R. Kustin, H. Rahmati, and A. Vraciu. The resolution of the bracket
powers of the maximal ideal in a diagonal hypersurface ring. J. Algebra,
369:256–321, 2012.
[KST07] H. Kajiura, K. Saito, and A. Takahashi. Matrix factorization and
representations of quivers. II: Type ADE case. Adv. Math., 211(1):327–362,
2007.
[Kun69] E. Kunz. Characterizations of regular local rings of characteristic p. Amer. J.
Math., 91(3):772–784, 1969.
[Kun76] E. Kunz. On Noetherian rings of characteristic p. Amer. J. Math.,
98(4):999–1013, 1976.
[Lan04] A. Langer. Semistable sheaves in positive characteristic. Ann. of Math. (2),
159:251–276, 2004.
[Lie13] C. Liedtke. Algebraic surfaces in positive characteristic. In F. Bogomolov,
B. Hassett, and Y. Tschinkel, editors, Birational Geometry, Rational Curves,
and Arithmetic. Springer, 2013.
[Liu06] Q. Liu. Algebraic geometry and arithmetic curves, volume 6 of Oxford
Graduate Texts in Mathematics. Oxford University Press, Oxford, UK,
2006.
[LS77] H. Lange and U. Stuhler. Vektorbündel auf Kurven und Darstellungen der
algebraischen Fundamentalgruppe. Math. Zeitschrift, 156:77–83, 1977.
192
Bibliography
[LW12] G. J. Leuschke and R. Wiegand. Cohen-Macaulay Representations, volume
181 of Mathematical Surveys and Monographs. AMS, Providence, RI, 2012.
[Mar81] M. Maruyama. The theorem of Grauert-Mülich-Spindler. Math. Ann.,
255:317–333, 1981.
[Mat05] H. Matsumura. Commutative Ring Theory. Number 8 in Cambridge Studies
in Advanced Mathematics. Cambridge University Press, Cambridge, UK,
2005.
[Mil00] C. Miller. A Frobenius characterization of finite projective dimension over
complete intersections. Math. Z., 233:127–136, 2000.
[Miy87] Y. Miyaoka. The Chern classes and Kodaira dimension of a minimal variety.
In Algebraic Geometry, Sendai 1985, volume 10 of Adv. Stud. Pure Math.,
pages 449–476, Amsterdam, 1987. North-Holland.
[Mon83a] P. Monsky. The Hilbert-Kunz function. Math. Ann., 263:43–49, 1983.
[Mon83b] P. Monsky. p-Ranks of Class Groups in Zdp -Extensions. Math. Ann.,
263:509–514, 1983.
[Mon97] P. Monsky. The Hilbert-Kunz function of a characteristic 2 cubic. J.
Algebra, 197:268–277, 1997.
[Mon98a] P. Monsky. The Hilbert-Kunz functions in a family: Line-s4 quartics. J.
Algebra, 208:359–371, 1998.
[Mon98b] P. Monsky. The Hilbert-Kunz functions in a family: Point-s4 quartics. J.
Algebra, 208:343–358, 1998.
[Mon05] P. Monsky. On the Hilbert-Kunz function of zd − p4 (x, y). J. Algebra,
291:350–372, 2005.
[Mon06a] P. Monsky. The Hilbert-Kunz multiplicity of an irreducible trinomial. J.
Algebra, 304:1101–1107, 2006.
[Mon06b] P. Monsky. Mason’s theorem and syzygy gaps. J. Algebra, 303:373–381,
2006.
[Mon07] P. Monsky. Hilbert-Kunz functions for irreducible plane curves. J. Algebra,
316:326–345, 2007.
[Mon08] P. Monsky. Rationality of Hilbert-Kunz multiplicities: A likely
counterexample. Michigan Math. J., 57:605–613, 2008.
[Mon09] P. Monsky. Transcendence of some Hilbert-Kunz multiplicities (modulo a
conjecture). preprint, arXiv:0908.0971v1, 2009.
[Mon11] P. Monsky. Hilbert-Kunz theory for nodal cubics, via sheaves. J. Algebra,
346(1):180–188, 2011.
[MS12] L. E. Miller and I. Swanson. Hilbert-Kunz functions of 2 × 2 determinantal
rings. to appear in Illinois J. Math., arXiv:1206.1015v2, 2012.
[MT04] P. Monsky and P. Teixeira. p-Fractals and power series I - Some 2 variable
results. J. Algebra, 280(2):505–536, 2004.
[MT06] P. Monsky and P. Teixeira. p-Fractals and power series II - Some
applications to Hilbert-Kunz theory. J. Algebra, 304:237–255, 2006.
[Nag75] M. Nagata. Local rings. Krieger, Huntington, NY, 1975.
[Par94] K. Pardue. Nonstandard Borel-fixed ideals. PhD thesis, Brandeis University,
1994.
Bibliography
193
[Pot97] J. Le Potier. Lectures on vector bundles, volume 54 of Cambridge Studies in
Advanced Mathematics. Cambridge University Press, Cambridge, UK, 1997.
[RS13] M. Robinson and I. Swanson. Explicit Hilbert-Kunz functions of 2 × 2
determinantal rings. preprint, arXiv:1304.7274, 2013.
[Sch57] T. Schneider. Einführung in die transzendenten Zahlen. Springer, Berlin,
1957.
[SdB97] K. E. Smith and M. Van den Bergh. Simplicity of rings of differential
operators in prime characteristic. Proc. London Math. Soc. (3), 75(1):32–62,
1997.
[Sei89] G. Seibert. Complexes with Homology of Finite Length and Frobenius
Functors. J. Algebra, 125:278–287, 1989.
[Sei97] G. Seibert. The Hilbert-Kunz function of rings of finite Cohen-Macaulay
type. Arch. Math. (Basel), 69:286–296, 1997.
[Sin05] A. K. Singh. The F-Signature of an affine semigroup ring. J. Pure Appl.
Algebra, 196(2):313–321, 2005.
[Smi03] L. Smith. A modular analog of a theorem of R. Steinberg on coinvariants of
complex pseudoreflection groups. Glasgow Math. J., 45:69–71, 2003.
[Stä11] A. Stäbler. An explicit example of a family of trivialising étale covers.
preprint, arXiv:1111.2131v1, 2011.
[SW11] A. Sannai and K.-i. Watanabe. F-Signature of graded Gorenstein rings. J.
Pure Appl. Algebra, 215(9):2190–2195, 2011.
[Tei02] P. Teixeira. p-Fractals and Hilbert-Kunz series. PhD thesis, Brandeis
University, 2002.
[Tei12] P. Teixeira. Syzygy gap fractals I. some structural results and an upper
bound. J. Algebra, 350(1):132–162, 2012.
[Tri05a] V. Trivedi. Semistability and Hilbert-Kunz multiplicities for curves. J.
Algebra, 284:627–644, 2005.
[Tri05b] V. Trivedi. Strong semistability and Hilbert-Kunz multiplicity for singular
plane curves. Contemp. Math., 390:165–173, 2005.
[Tri07] V. Trivedi. Hilbert-Kunz multiplicity and reduction mod p. Nagoya Math. J.,
185:123–141, 2007.
[Tri12] V. Trivedi. Frobenius pull-backs of vector bundles in higher dimensions.
Proc. Indian Acad. Sci., 122(4):615–628, 2012.
[Tuc12] K. Tucker. F-Signature exists. Invent. Math., 190(3):743–765, 2012.
[Upa12a] S. Upadhyay. An algorithm for the HK function for disjoint-term trinomial
hypersurfaces. preprint, arXiv:1204.5417v1, 2012.
[Upa12b] S. Upadhyay. The Hilbert-Kunz function for binomial hypersurfaces.
preprint, arXiv:1101.5936v2, 2012.
[Upa12c] S. Upadhyay. Irrational HK multiplicities are possible for trinomial
hypersurfaces. preprint, arXiv:1211.7244v1, 2012.
[Upa12d] S. Upadhyay. A rank computation problem related to the HK function of
trinomial hypersurfaces. preprint, arXiv:1210.2885v1, 2012.
[Wat00] K.-i. Watanabe. Hilbert-Kunz multiplicity of toric rings. Proc. Inst. Nat.
Sci., 35:173–177, 2000.
[Wit02] E. Witten. Singularities in String Theory. Proceedings of the ICM,
1:495–504, 2002.
[WY00] K.-i. Watanabe and K.-i. Yoshida. Hilbert-Kunz multiplicity and an
inequality between multiplicity and colength. J. Algebra, 230:295–317,
2000.
[WY01a] K.-i. Watanabe and K.-i. Yoshida. Hilbert-Kunz multiplicity, McKay
correspondence and good ideals in two-dimensional rational singularities.
Manuscripta Math., 104:275–294, 2001.
[WY01b] K.-i. Watanabe and K.-i. Yoshida. Hilbert-Kunz multiplicity of
two-dimensional local rings. Nagoya Math. J., 162:87–110, 2001.
[WY04] K.-i. Watanabe and K.-i. Yoshida. Minimal relative Hilbert-Kunz
multiplicity. Illinois J. Math., 48(1):273–294, 2004.
[WY05] K.-i. Watanabe and K.-i. Yoshida. Hilbert-Kunz multiplicity of
three-dimensional local rings. Nagoya Math. J., 177:47–75, 2005.
[Yao06] Y. Yao. Observations on the F-signature of local rings of characteristic p. J.
Algebra, 299:198–218, 2006.
[Yos90] Y. Yoshino. Cohen-Macaulay Modules over Cohen-Macaulay Rings, volume
146 of Lon. Mat. Soc. LNS. Cambridge University Press, Cambridge, UK,
1990.
[Yos98] Y. Yoshino. Tensor products of matrix factorizations. Nagoya Math. J.,
152:39–56, 1998.
| 0math.AC
|
Cross-validation estimation of covariance parameters
under fixed-domain asymptotics
F. Bachoc∗1 , A. Lagnoux1 , and T.M.N. Nguyen1
arXiv:1610.02872v2 [math.ST] 25 Jul 2017
1
Institut de Mathématiques de Toulouse, Université Paul Sabatier, France
118 route de Narbonne, 31062 TOULOUSE Cedex 9
{francois.bachoc,agnes.lagnoux,thi_mong_ngoc.nguyen}@math.univtoulouse.fr
July 26, 2017
Abstract
We consider a one-dimensional Gaussian process having exponential covariance
function. Under fixed-domain asymptotics, we prove the strong consistency and
asymptotic normality of a cross validation estimator of the microergodic covariance
parameter. In this setting, Ying [40] proved the same asymptotic properties for the
maximum likelihood estimator. Our proof includes several original or more involved
components, compared to that of Ying. Also, while the asymptotic variance of maximum likelihood does not depend on the triangular array of observation points under
consideration, that of cross validation does, and is shown to be lower and upper
bounded. The lower bound coincides with the asymptotic variance of maximum likelihood. We provide examples of triangular arrays of observation points achieving the
lower and upper bounds. We illustrate our asymptotic results with simulations, and
provide extensions to the case of an unknown mean function. To our knowledge, this
work constitutes the first fixed-domain asymptotic analysis of cross validation.
Keywords: Kriging, cross validation, strong consistency, asymptotic normality, spatial
sampling, fixed-domain asymptotics
1
Introduction
Kriging [35, 28] consists in inferring the values of a Gaussian random field given observations
at a finite set of observation points. It has become a popular method for a large range
∗
Corresponding author.
[email protected], Institut de Mathématiques de
Toulouse, Université Paul Sabatier, 118 route de Narbonne, 31062 TOULOUSE Cedex 9
1
of applications, such as geostatistics [25], numerical code approximation [29, 30, 8] and
calibration [27, 9] or global optimization [20].
Before Kriging can be applied, a covariance function must be chosen. The most common
practice is to statistically estimate the covariance function, from a set of observations of the
Gaussian process, and to plug [35, Ch.6.8] the estimate in the Kriging equations. Usually,
it is assumed that the covariance function belongs to a given parametric family (see [1]
for a review of classical families). In this case, the estimation boils down to estimating
the corresponding covariance parameters. For covariance parameter estimation, maximum
likelihood (ML) is the most studied and used method, while cross validation (CV) [36, 43, 5]
is an alternative technique. CV has been shown to have attractive properties, compared
to ML, when the parametric family of covariance functions is misspecified [5, 7].
There is a fair amount of literature on the asymptotic properties of ML. In this regard, the
two main frameworks are increasing-domain and fixed-domain asymptotics [35, p.62]. Under increasing-domain asymptotics, the average density of observation points is bounded,
so that the infinite sequence of observation points is unbounded. Under fixed-domain
asymptotics, this sequence is dense in a bounded domain.
Consider first increasing-domain asymptotics. Generally speaking, for all (identifiable)
covariance parameters, the ML estimator is consistent and asymptotically normal under
some mild regularity conditions. The asymptotic covariance matrix is equal to the inverse
of the (asymptotic) Fisher information matrix. This result was first shown in [24], and
then extended in different directions in [12, 13, 31, 6, 16].
The situation is significantly different under fixed-domain asymptotics. Indeed, two types
of covariance parameters can be distinguished: microergodic and non-microergodic parameters [18, 35]. A covariance parameter is microergodic if, for two different values of it, the
two corresponding Gaussian measures are orthogonal, see [18, 35]. It is non-microergodic if,
even for two different values of it, the two corresponding Gaussian measures are equivalent.
Non-microergodic parameters cannot be estimated consistently, but have an asymptotically
negligible impact on prediction [32, 33, 34, 42]. On the other hand, it is at least possible
to consistently estimate microergodic covariance parameters, and misspecifying them can
have a strong negative impact on prediction.
Under fixed-domain asymptotics, there exist results indicating which covariance parameters are microergodic, and providing the asymptotic properties of the corresponding ML
estimator. Most of these available results are specific to particular covariance models. In
dimension d = 1 when the covariance model is exponential, only a reparameterized quantity
obtained from the variance and scale parameters is microergodic. It is shown in [40] that
the ML estimator of this microergodic parameter is strongly consistent and asymptotically
normal. These results are extended in [11], by taking into account measurement errors,
and in [10], by taking into account both measurement errors and an unknown mean function. When d > 1 and for a separable exponential covariance function, all the covariance
parameters are microergodic, and the asymptotic normality of the ML estimator is proved
in [41]. Other results in this case are also given in [37, 2]. Consistency of ML is shown as
well in [23] for the scale parameters of the Gaussian covariance function and in [22] for all
the covariance parameters of the separable Matérn 3/2 covariance function. Finally, for
2
the entire isotropic Matérn class of covariance functions, all parameters are microergodic
for d > 4 [3], and only reparameterized parameters obtained from the scale and variance
are microergodic for d 6 3 [42]. In [21], the asymptotic normality of the ML estimators for
these microergodic parameters is proved, from previous results in [14] and [39]. Finally we
remark that, beyond ML, quadratic variation-based estimators have also been extensively
studied, under fixed-domain asymptotics (see for instance [19]).
In contrast to ML, CV has received less theoretical attention. Under increasing-domain
asymptotics, the consistency and asymptotic normality of a CV estimator is proved in
[6]. Also, under increasing-domain asymptotics, it is shown in [7] that this CV estimator asymptotically minimizes the integrated square prediction error. To the best of our
knowledge, no fixed-domain asymptotic analysis of CV exists in the literature.
In this paper, we provide a first fixed-domain asymptotic analysis of the CV estimator
minimizing the CV logarithmic score, see [28] Equation (5.11) and [43]. We focus on
the case of the one-dimensional exponential covariance function, which was historically
the first covariance function for which the asymptotic properties of ML were derived [40].
This covariance function is particularly amenable to theoretical analysis, as its Markovian
property yields an explicit (matrix-free) expression of the likelihood function. It turns out
that the CV logarithmic score can also be expressed in a matrix-free form, which enables
us to prove the strong consistency and asymptotic normality of the corresponding CV
estimator. We follow the same general proof architecture as in [40] for ML, but our proof,
and the nature of our results, contain several new elements.
In terms of proofs, the random CV logarithmic score, and its derivatives, have more complicated expressions than for ML. [This is because the CV logarithm score is based on
the conditional distributions of the observations, from both their nearest left and right
neighbors, while the likelihood function is solely based on the nearest left neighbors. See
Lemma 3.1 and Lemma 1 in [40] for details.] As a consequence, the computations are more
involved, and some other tools than in [40] are needed. In particular, many of our asymptotic approximations rely on Taylor expansions of functions of several variables (where each
variable is an interpoint distance going to zero, see the proofs for details). In contrast, only
Taylor approximations with one variable are needed in [40]. In addition, we use central
limit theorems for dependent random variables, while only independent variables need to
be considered in [40].
The nature of our asymptotic normality result also differs from that in [40]. In this reference, the asymptotic variance does not depend on the triangular array of observation
points. On the contrary, in our case, different triangular arrays of observation points can
yield different asymptotic variances. We exhibit a lower and an upper bound for these
asymptotic variances, and provide examples of triangular arrays reaching them. The lower
bound is in fact equal to the asymptotic variance of ML in [40]. Interestingly, the triangular array given by equispaced observation points attains neither the lower nor the upper
bound. It is also pointed out in [6] that equispaced observation points need not provide
the smallest asymptotic variance for covariance parameter estimation.
Finally, the fact that the asymptotic variance is larger for CV than for ML is a standard finding in the well-specified case considered here, where the covariance function of
3
the Gaussian process does belong to the parametric family of covariance functions under
consideration. In contrasts, as mentioned above, CV has attractive properties compared
to ML when this well-specified case does not hold [5, 7].
The rest of the paper is organized as follows. In Section 2, we present in more details
the setting and the CV estimator under consideration. In Section 3, we give our strong
consistency result for this estimator. In Section 4, we provide the asymptotic normality
result, together with the analysis of the asymptotic variance. In Section 5, we present
numerical experiments, illustrating our theoretical findings. In Section 6, we extend the
results of Sections 3 and 4 to the case of an unknown mean function. In Section 7, we give
a few concluding remarks. All the proofs are postponed to Section 8.
2
The context and the cross-validation estimators
We consider a centered Gaussian process Y on [0, 1] with covariance function
K0 (t1 , t2 ) = σ02 exp{−θ0 |t1 − t2 |}
for some fixed and unknown parameters θ0 > 0 and σ02 > 0. This process is commonly
known as the Ornstein-Uhlenbeck process. It satisfies the following stochastic differential
equation, called the Langevin’s equation,
p
dY (t) = −θ0 Y (t)dt + 2θ0 σ0 dB(t),
where (B(t))t denotes a standard Brownian motion process. The Ornstein-Uhlenbeck process has been widely used to model physical, biological, social, and many other phenomena.
It also possesses many useful mathematical properties that simplify the analysis.
We introduce the parametric set of covariance functions {Kθ,σ2 , a 6 θ 6 A, b 6 σ 2 6 B}
for some fixed 0 < a 6 A < +∞ and 0 < b 6 B < +∞ where
Kθ,σ2 (t1 , t2 ) = σ 2 exp{−θ|t1 − t2 |}.
For any n ∈ N, we consider a design of observation points {s1 , ..., sn }. Without loss of
generality, we may assume that 0 = s1 < ... < sn = 1. Similarly as in [40], there is no need
to assume that the sequences of observation points are nested. We consider the vector
of observations at locations s1 , ..., sn , (Y (s1 ), . . . , Y (sn ))0 . Now let ∆i := si − si−1 , for
i = 2, ..., n, and yi := Y (si ), for i = 1, ..., n. For ease of redaction, we do not mention in si
and ∆i the dependency in n. We define Rθ as the variance-covariance matrix of (y1 , ..., yn )0
under covariance function Kθ,1 ,
n
P
−θ
∆i
e−θ∆2 · · · e i=2
1
n
P
−θ
∆i
−θ∆
i=3
e 2
1
··· e
Rθ :=
.
..
..
..
..
.
.
.
.
n
n
P
P
−θ
∆i
−θ
∆i
e i=2
e i=3
···
1
4
From [4], we have
1
1−e−2θ∆2
R−1
θ
−e−θ∆2
−2θ∆
2
1−e
= 0
..
.
−e−θ∆2
1−e−2θ∆2
1
e−2θ∆3
+ 1−e
−2θ∆3
1−e−2θ∆2
...
..
.
···
0
···
...
0
...
...
0
0
..
.
.
0
−e−θ∆n
1−e−2θ∆n
e−2θ∆n
1
+ 1−e
−2θ∆n
1−e−2θ∆n−1
−e−θ∆n
1−e−2θ∆n
(1)
1
1−e−2θ∆n
We now address the CV estimators of θ0 and σ02 considered in [28, 43]. Let
Ŷθ,−i (si ) = Eθ,σ2 (Y (si )|Y (s1 ), ..., Y (si−1 ), Y (si+1 ), ..., Y (sn )),
where the conditional expectation Eθ,σ2 is calculated assuming that Y is centered and has
covariance function Kθ,σ2 . We remark that Ŷθ,−i (si ) does not depend on σ 2 . We define
similarly
2
σ̂θ,σ
2 ,−i (si ) = Varθ,σ 2 (Y (si )|Y (s1 ), ..., Y (si−1 ), Y (si+1 ), ..., Y (sn )).
Then, the CV estimators are given by
(θ̂, σ̂ 2 ) ∈
argmin
Sn (θ, σ 2 ),
a6θ6A,b6σ 2 6B
where
Sn (θ, σ 2 ) =
n
X
i=1
"
(yi − Ŷθ,−i (si ))2
2
log(σ̂θ,σ
2 ,−i (si )) +
2
σ̂θ,σ
2 ,−i (si )
#
(2)
is the logarithmic score. The rationale for minimizing the logarithmic score is that log(2π)+
(yi −Ŷθ,−i (si ))2
2
log(σ̂θ,σ
is equal to −2 times the conditional log-likelihood of yi ,
2 ,−i (si )) +
σ̂ 2
(si )
θ,σ 2 ,−i
given (y1 , ..., yi−1 , yi+1 , ..., yn )0 , with covariance parameters θ, σ 2 . The term cross-validation
underlines the fact that we consider leave-one-out quantities.
As already known [18, 40, 42], it is not possible to consistently estimate simultaneously θ0
and σ02 (the ML estimator of θ0 is a non-degenerate random variable, even if (Y (t))t∈[0,1]
is observed continuously [44]), but it is possible to consistently estimate θ0 σ02 . As a consequence, we have considered three different cases, as in [40]. (i) Set σ 2 = σ12 in (2) with
σ12 > 0 being a predetermined constant and consider the CV estimator θ̂1 of θ1 = θ0 σ02 /σ12
that minimizes (2) with σ 2 = σ12 . (ii) Set θ = θ2 in (2) with θ2 > 0 being a predetermined
constant and consider the CV estimator σ̂22 of σ22 = θ0 σ02 /θ2 that minimizes (2) with θ = θ2 .
(iii) Consider the estimator θ̂σ̂ 2 of θ0 σ02 , where θ̂ and σ̂ 2 are the CV estimators of θ0 and
σ02 .
Ying [40] considers the ML estimators of θ and σ 2 and establishes their consistency and
asymptotic normality. We carry out a similar asymptotic analysis for the above CV estimators. More precisely, we prove that θ̂σ̂ 2 (resp. θ̂1 and σ̂22 ) converges almost surely
5
to θ0 σ02 (resp. θ1 and σ22 ) in the next section. √In section 4, we establish
that, for a se√
2
2
2
quence
τ
which
is
lower
and
upper-bounded,
(
n/[θ
τ
])(
θ̂
−
θ
),
(
n/[σ
1 n
1
1
2 τn ])(σ̂2 − σ2 )
√ n
and ( n/[θ0 σ02 τn ])(θ̂σ̂ 2 − θ0 σ02 ) all converge in distribution to a standard Gaussian random
variable. We remark that the asymptotic variance τn2 depends on how the underlying design points {s1 , ..., sn } are chosen. On the contrary, considering the ML estimators [40],
the asymptotic variance is the same for any triangular array of design points.
3
Consistency
In this section, we establish the strong consistency of the CV estimator θ̂σ̂ 2 of θ0 σ02 described
in the previous section. In that view, we consider Sn (θ, σ 2 ) defined by (2). As done in
[40], we base our analysis on the Markovian property of the Ornstein-Uhlenbeck process
in order to handle the fact that, as n increases, the observed sample (y1 , . . . , yn )0 becomes
more and more correlated. We have
X (R−1 )ij
θ
Ŷθ,−i (si ) = −
yj
(3)
(R−1
θ )ii
j=1,...,n;
j6=i
and
σ2
,
(R−1
θ )ii
from [43, 5, 15]. Then, using Equation (1), we get the following lemma after some tedious
computations.
2
σ̂θ,σ
2 ,−i (si ) =
Lemma 3.1 (Logarithmic score). With Sn (θ, σ 2 ) as in (2), we have
Sn (θ, σ 2 ) = n log(σ 2 ) + log(1 − e−2θ∆2 ) + log(1 − e−2θ∆n )
n−1
(y1 − e−θ∆2 y2 )2 (yn − e−θ∆n yn−1 )2 X
1
e−2θ∆i+1
+ 2
+
−
log
+
−2θ∆i
σ (1 − e−2θ∆2 )
σ 2 (1 − e−2θ∆n )
1
−
e
1 − e−2θ∆i+1
i=2
2
e−θ∆i
e−θ∆i+1
n−1
−2θ∆i+1
X
y
+
y
i−1
i+1
−2θ∆
−2θ∆
1
1
e
i
i+1
1−e
yi − 1−e
.
+ 2
+
−2θ∆i+1
−2θ∆
−2θ∆
i
i+1
1
e
σ i=2 1 − e
1−e
+
−2θ∆
−2θ∆
1−e
i
1−e
i+1
Based on Lemma 3.1, we prove the following theorem in Section 8.2.
Theorem 3.2 (Consistency). Assume that
lim sup max ∆i = 0.
n→+∞ i=2,...,n
(4)
Let J = [a, A] × [b, B], where a, A, b and B are fixed and have been defined in the previous
section. Assume that there exists (θ̃, σ̃ 2 ) in J so that θ̃σ̃ 2 = θ0 σ02 . Define (θ̂, σ̂ 2 ) ∈ J as a
solution of
Sn (θ̂, σ̂ 2 ) = min
Sn (θ, σ 2 ).
2
(θ,σ )∈J
6
(5)
Then (θ̂, σ̂ 2 ) exists and
a.s.
θ̂σ̂ 2 → θ0 σ02 .
(6)
In particular, let σ12 > 0 and θ2 > 0 be predetermined constants satisfying σ02 θ0 /σ12 ∈ [a, A]
and σ02 θ0 /θ2 ∈ [b, B]. Define θ̂1 ∈ [a, A] and σ̂22 ∈ [b, B] as solutions of
Sn (θ̂1 , σ12 ) = min Sn (θ, σ12 )
(7)
Sn (θ2 , σ̂22 ) = 2min Sn (θ2 , σ 2 ).
(8)
θ∈[a,A]
and
σ ∈[b,B]
a.s.
a.s.
Then θ̂1 → θ1 := σ02 θ0 /σ12 and σ̂22 → σ22 := σ02 θ0 /θ2 .
Remark 3.3. It is worth remarking that the asymptotically preponderant terms in Lemma
3.1 are the same as those obtained in the context of ML estimation (see [40] and Section
8.2 for more details).
4
Asymptotic normality
Once the consistency has been established, the natural question of the convergence speed
arises. We address this point in this section. We first provide a central limit result in the
following theorem.
Theorem 4.1 (Central Limit Theorem). Consider the same notation and assumptions
as in Theorem 3.2. Assume further that either aB < θ0 σ02 ; Ab > θ0 σ02 or aB > θ0 σ02 ;
Ab < θ0 σ02 hold. Then the estimators are asymptotically normal. More precisely, we have
√
n
D
(θ̂σ̂ 2 − θ0 σ02 ) −−−→ N (0, 1).
(9)
2
n→∞
θ0 σ0 τn
Also, when (σ02 θ0 )/σ12 ∈ (a, A) we have
√
n
D
(θ̂1 − θ1 ) −−−→ N (0, 1).
n→∞
θ1 τn
Finally, when (σ02 θ0 )/θ2 ∈ (b, B) we have
√
n 2
D
(σ̂ − σ22 ) −−−→ N (0, 1).
2
n→∞
σ2 τn 2
(10)
(11)
The quantity τn2 depends on how the underlying design points {s1 , . . . , sn } have been chosen.
More precisely,
"
#
2
n−1
X
2
∆
∆
∆
∆
i+1
i−1
i i+1
τn2 =
+
+2
.
(12)
n i=3
∆i + ∆i+1 ∆i + ∆i−1
(∆i + ∆i+1 )2
7
Remark 4.2. The condition aB < θ0 σ02 ; Ab > θ0 σ02 or aB > θ0 σ02 ; Ab < θ0 σ02 ensures that
(∂/∂θ)Sn (θ̂, σ̂ 2 ) or (∂/∂σ 2 )Sn (θ̂, σ̂ 2 ) will be equal to zero for n large enough almost surely,
by applying Theorem 3.2. This is used in the proof of Theorem 4.1. A similar assumption
is made in [40], where the parameter domain for (θ, σ 2 ) is (0, ∞) × [b, B] or [a, A] × (0, ∞).
In the following proposition, we show that the quantity τ√n2 in Theorem 4.1 is lower and
upper bounded, so that the rate of convergence is always n in this theorem.
Proposition 4.3. We have, for any choice of the triangular array of design points {s1 , ..., sn }
satisfying (4),
2 6 lim inf τn2 6 lim supτn2 6 4.
n→∞
(13)
n→∞
Remark 4.4.
1. The asymptotic variance of the limiting distribution of θ̂σ̂ 2 − θ0 σ02
can be easily estimated. By the previous proposition, this asymptotic variance is
2
always larger than the one of the ML estimator. Indeed, with θ̂M L and σ̂M
L the ML
√
D
2
2
2
2
estimators of θ and σ we have ( n/[θ0 σ0 ])(θ̂M L σ̂M L − θ0 σ0 ) −−−→ N (0, 2), see [40].
n→∞
This fact is quite expected as ML estimates usually perform best when the covariance
model is well-specified, as is the case here.
1
2. As one can check easily, the regular design ∆i ≡ n−1
for all i = 2, . . . , n, does not
yield the limiting variance of the ML estimator. Instead, we have τn2 →n→∞ 3 for
this design. However, in Proposition 4.5, we exhibit a particular design realizing the
limiting variance of the ML estimator: lim τn2 = 2.
n→∞
In fact, the bounds in (13) are sharp as shown in the following proposition.
Proposition 4.5. (i) Let {s1 , . . . , sn } be such that s1 = 0, for i = 2, ..., n − 1,
(
(1 − γn ) n2 if i is even,
∆i = 2γn
if i is odd,
n
where γn ∈ (0, 1), and ∆n = 1 −
Pn−1
i=2
∆i . Then, taking γn = 1/n, we get τn2 → 4.
n→∞
(ii) Let {s1 , . . . , sn } and 0 < α < 1 be such that s1 = 0, ∆i = 1/(i !) for i = bnα c + 1, . . . , n
n
n
X
X
and ∆2 = · · · = ∆bnα c ≡ (1 − rn )/(bnα c − 1) with rn :=
∆i . Then
∆i = 1 and
i=bnα c+1
i=2
τn2 → 2.
n→∞
Remark 4.6. Intuitively, in Proposition 4.5 (ii), ∆i+1 will be much smaller than ∆i for
i+1
i ∆i+1
most of the indices i, so that the quantities ∆i∆+∆
and (∆∆i +∆
2 in (12) will be negligible.
i+1
i+1 )
We refer to the proof of Proposition 4.5 for further details.
8
5
Numerical experiments
We illustrate Theorem 4.1 by a Monte Carlo simulation. We set θ0 = 3 and σ02 = 1 and we
consider three sample size values, n = 12, 50, 200. For the sample size n = 12, we address
three designs {s1 , ..., sn }. The first one is the ‘minimal’ design given by Proposition 4.5
(ii) with α = 0.5, which asymptotically achieves the minimal estimation variance. The
second one is the ‘regular’ design given by {s1 , ..., sn } = {0, 1/(n − 1), ..., 1}. The third one
is the ‘maximal’ design given by Proposition 4.5 (i) with γn = 1/n, which asymptotically
achieves the maximal estimation variance. These three designs are show in Figure 1. For
the sample sizes n = 50 and n = 200, the ‘minimal’ design is not amenable to numerical
computation anymore, as the values of ∆i become too small; so that we only address the
‘regular’ and ‘maximal’ designs.
●
0.0
●
0.2
0.4
●
0.6
0.8
●●
●
1.0
0.0
●
●
0.2
●
●
●
0.4
●
●
0.6
●
●
0.8
●
●
1.0
●●
0.0
●●
0.2
●●
0.4
●●
●●
0.6
●
0.8
●
1.0
Figure 1: Plot of the points {s1 , ..., sn } for the ‘minimal’, ‘regular’ and ‘maximal’ designs
of the numerical experiments for n = 12 (from left to right). For the ‘minimal’ design, nine
points form a dense cluster around one and the asymptotic variance of the CV estimator
is 2 (Proposition 4.5 (i)), for the ‘regular’ design, the asymptotic variance is 3, and for the
‘maximal’ design, the asymptotic variance is 4 (Proposition 4.5 (ii)).
For a given configuration of n and a given design {s1 , ..., sn }, we repeat N = 2.000 data
generations and estimations. That is, we independently sample N Gaussian vectors of
size n with zero mean vector and covariance matrix [σ02 e−θ0 |si −sj | ]16i,j6n . For each of
these Gaussian vectors, we compute the CV estimators θ̂ and σ̂ 2 , with parameter space
[0.1, 10] × [0.3, 30], so that we consider case (9) of Theorem 4.1. The computation of θ̂
and σ̂ 2 is not challenging since, from Lemma 3.1, the logarithmic score Sn (θ, σ 2 ) can be
computed quickly, with a O(n) complexity. [For more general covariance functions, the
computation of CV or ML criteria is more costly, with a O(n3 ) complexity.] The criterion
Sn is minimized over (θ, σ 2 ) by repeating the Broyden–Fletcher–Goldfarb–Shanno (BFGS)
algorithm, with several starting points for (θ, σ 2 ), and by keeping the value of (θ, σ 2 ) with
smallest logarithmic score, over all the repetitions. The R software was used, with the
optim function. [We remark that, for fixed θ, one could find an explicit expression of
σ̂ 2 (θ) ∈ argminσ2 >0 Sn (θ, σ 2 ) (see also [5] for a different CV criterion). Hence, it would
be possible to minimize the profile logarithmic score minσ2 >0 Sn (θ, σ 2 ) over θ only. As
9
mentioned earlier, this improvement is not needed here, since the criterion Sn (θ, σ 2 ) can
be computed quickly.]
√
For the N values of (θ̂, σ̂ 2 ), we compute the N values of ( n/[θ0 σ02 ])(θ̂σ̂ 2 − θ0 σ02 ). In
Figure 2, we report the histograms of these latter N values, for the seven configurations
under consideration. In addition, we report the probability density functions of the seven
corresponding Gaussian distributions with mean 0 and variance τn2 , to which the histograms
converge when n → ∞, in view of Theorem 4.1.
In Figure 2, we observe that, for n = 12, the asymptotic Gaussian distributions are already
reasonable approximations of the empirical histograms. For n = 50, the asymptotic distributions become very close to the histograms, and for n = 200 the asymptotic distributions
are almost identical to the histograms. Hence, the convergence in distribution of Theorem
4.1 provides a good approximation of the finite sample situation already for small to moderate n. The case n = 12 illustrates the benefit of the ‘minimal’ design for estimation, as
2
the histogram is most concentrated around zero for this design. Similarly, the value of τ12
is the smallest for this design, compared to the ‘regular’ and ‘maximal’ designs. For n = 50
and 200, we also observe that the estimation is more accurate for the ‘regular’ design than
for the ‘maximal’ design, which also confirms Remark 4.4 and Proposition 4.5.
Finally, we have obtained similar conclusions for the case where either θ0 or σ02 is known
in the computation of θ̂, σ̂ 2 (cases of (10) and (11)). We do not report the corresponding
results for concision.
6
Extension to regression models
In this section, we extend Theorems 3.2 and 4.1 to the case of regression models. We
assume that, instead of Y , we observe the Gaussian process Z defined by Z(t) = β01 f1 (t) +
... + β0p fp (t) + Y (t). In the definition of Z, β 0 = (β01 , ..., β0p )0 ∈ Rp is fixed and unknown
and, for k = 1, ..., p, fk : [0, 1] → R is a known function. Hence, we estimate jointly
(θ, σ 2 , β 00 ) from the observation vector z = (z1 , ..., zn )0 , with zi = Z(si ).
−1 0 −1
Let F be the n × p matrix [fj (si )]i=1,...,n,j=1,...,p . Then β̂ = (F 0 R−1
θ F ) F Rθ z is the best
linear unbiased predictor of β 0 given z, under covariance function Kθ,σ2 for all σ 2 , see e.g.
[30].
We now address CV estimation. Let f i = (f1 (si ), ..., fp (si ))0 , let z −i = (z1 , ..., zi−1 , zi+1 , ..., zn )0 ,
let F −i be the matrix obtained by removing line i of F , and let Rθ,−i be the (n −
1) × (n − 1) matrix obtained by removing line and column i of Rθ . Then, for all σ 2 ,
−1
−1 0
β̂ −i = (F 0−i R−1
θ,−i F −i ) F −i Rθ,−i z −i is the best linear unbiased predictor of β 0 given z −i ,
under covariance function Kθ,σ2 .
We also let r θ,−i = (Kθ,1 (si , s1 ), ..., Kθ,1 (si , si−1 ), Kθ,1 (si , si+1 ), ..., Kθ,1 (si , sn ))0 . Then, from
e.g. [30],
Ẑθ,−i (si ) = f 0i β̂ −i + r 0θ,−i R−1
(14)
θ,−i (z −i − F −i β̂ −i )
is the best linear unbiased predictor of zi given z −i . We let
h
i2
2
σ̌θ,σ2 ,−i (si ) = Eθ,σ2 Ẑθ,−i (si ) − Z(si )
.
10
−5
0
5
0.3
0.0
0.1
0.2
0.3
0.2
0.1
0.0
0.0
0.1
0.2
0.3
0.4
n=12, maximal
0.4
n=12, regular
0.4
n=12, minimal
10
−5
0
5
−5
0
5
10
0.3
0.2
0.1
0.0
0.0
0.1
0.2
0.3
0.4
n=50, maximal
0.4
n=50, regular
10
−5
0
5
10
−5
5
10
0.3
0.2
0.1
0.0
0.0
0.1
0.2
0.3
0.4
n=200, maximal
0.4
n=200, regular
0
−5
0
5
10
−5
0
5
10
Figure
2: Illustration of Theorem 4.1. Histograms of N = 2.000 independent realizations of
√
( n/[θ0 σ02 ])(θ̂σ̂ 2 − θ0 σ02 ), together with the corresponding asymptotic Gaussian probability
density functions with mean 0 and variances τn2 (red lines). The sample size is n = 12 (top
row), n = 50 (middle row) and n = 200 (bottom row). For the top row, the designs are
the ‘minimal’ design (left), achieving the smallest asymptotic variance; the ‘regular’ design
(middle), with equispaced observation points; and the ‘maximal’ design (right), achieving
the largest asymptotic variance. For the middle and bottom rows, the designs are the
‘regular’ design (left) and the ‘maximal’ design (right).
Then, the CV estimator of (θ, σ 2 ) we shall study in this section is
(θ̌, σ̌ 2 ) ∈ argmina6θ6A,b6σ2 6B S̄n (θ, σ 2 )
11
with
S̄n (θ, σ 2 ) =
n
X
i=1
"
#
2
(z
−
Ẑ
(s
))
i
θ,−i i
2
log(σ̌θ,σ
.
2 ,−i (si )) +
2
σ̌θ,σ
2 ,−i (si )
We remark that [43] suggests to use a similar CV criterion, with the notable difference
that β̂ −i is replaced by β̂ in (14). The benefit of the CV predictor (14), compared to that
considered in [43], is that, in (14), no use of zi is made at all for predicting zi . In [15], the
following relations are shown, extending those of Section 3. We have
Ẑθ,−i (si ) = −
X (Q− )ij
θ
zj
−
(Q
)
ii
θ
j=1,...,n;
j6=i
and
2
σ̌θ,σ
2 ,−i (si ) =
σ2
,
(Q−
θ )ii
(15)
0 −1
−1
−1
−1 0 −1
with Q−
θ = Rθ − Rθ F (F Rθ F ) F Rθ .
Based on the two displays above, and again using the explicit matrix inverse in (1), we
are able to prove the consistency and asymptotic normality of θ̌σ̌ 2 where the asymptotic
distribution is identical to that of Section 4.
Theorem 6.1 (Consistency). Assume that lim sup max ∆i = 0 and that there exists (θ̃, σ̃ 2 )
n→+∞ i=2,...,n
2
θ0 σ02 .
in J so that θ̃σ̃ =
Assume also that f1 , ..., fp are twice continuously differentiable
and are linearly independent on [0, 1]. Then (θ̌, σ̌ 2 ) exists and
a.s.
θ̌σ̌ 2 → θ0 σ02 .
(16)
Theorem 6.2 (Central Limit Theorem). Assume that the conditions of Theorem 6.1 hold
and that aB > θ0 σ02 and Ab < θ0 σ02 . Then, with τn as in (12), we have
√
n
D
(θ̌σ̌ 2 − θ0 σ02 ) −−−→ N (0, 1).
(17)
2
n→∞
θ0 σ0 τn
In Theorems 6.1 and 6.2, the twice differentiability condition for f1 , ..., fp is mostly technical, and could be replaced by a continuous differentiability condition, at the price of more
technical proofs. [We remark that Theorems 6.1 and 6.2 apply in particular to polynomial
functions f1 , ..., fp which are widely considered, for instance in the framework of computer
experiments [30].] As remarked in [40], when f1 , ..., fp are continuously differentiable, the
parameter β 0 is non-microergodic and can not be consistently estimated.
Finally, assume now that f1 , ..., fp satisfy the conditions given in Theorem 3 (ii) in [40].
Then, one can show from the proof of this theorem that, for any sequence or random
variables (θ̄, σ̄ 2 ) ∈ J (and in particular for ( θ̌, σ̌ 2 )), the estimator β̂ = β̂(θ̄) given above
is consistent and asymptotically normal, with asymptotic distribution given in (4.5) in
[40]. In this setting, it would be interesting to study the joint asymptotic distribution of
(θ̌, σ̌ 2 , β̂).
12
7
Concluding remarks
We have proved the consistency and asymptotic normality of the CV estimator of the
microergodic√parameter θ0 σ02 , based on the logarithmic score. While the ML asymptotic
2
2
variance of ( n/[θ0 σ02 ])(θ̂M L σ̂M
L − θ0 σ0 ) is 2 for any triangular array of observation points,
the corresponding CV asymptotic variance is simply bounded between 2 and 4, those
bounds being tight. The triangular array we exhibit, achieving the asymptotic variance 2
for CV, is based on some ratios between interpoint distances (of the form (si+1 −si )/(sj+1 −
sj )) going to zero as n → ∞, which makes it too challenging to simulate numerically for
large n. It would be interesting to find the smallest possible asymptotic variance for CV,
when the ratios between interpoint distances are bounded away from zero.
One interesting agenda for future research would be to extend this asymptotic analysis of
CV in the other settings where such an analysis was possible for ML. These settings include the case of measurement errors for the exponential covariance function in dimension
one [11, 10], the case of the separable exponential covariance function in higher dimension [41] (consistency and asymptotic normality), of the separable Matérn 3/2 covariance
function [22] (consistency) and of the Gaussian covariance function [23] (consistency). In
these references, tractable approximations of the inverse covariance matrices are provided,
which could also be exploited in the case of CV. Finally, using techniques which are more
spectral in nature, [14, 21, 39] prove central limit theorems for the ML estimation of the
microergodic parameter in the isotropic Matérn model. An extension to CV would also be
valuable.
8
8.1
Proofs
Notation and auxiliary results
Remind that ∆i = si − si−1 and introduce Wi := [yi − e−θ∆i yi−1 ] and its normalized
version W i := [yi − e−θ∆i yi−1 ]/[σ 2 (1 − e−2θ∆i )]1/2 for i = 2, . . . , n (the dependency in n is
not mentioned in Wi and W i to lighten notation). When (θ, σ 2 ) = (θ0 , σ02 ), the random
variables will be denoted Wi,0 and W i,0 . By the Markovian and Gaussian properties of
Y , it follows that for each i > 2, W i,0 is independent of {yj , j 6 i − 1}. Therefore,
{W i,0 , 2 6 i 6 n} is an i.i.d. sequence of random variables having the standard normal
distribution N (0, 1).
It is convenient to have short expressions for terms that converge in probability to zero.
We follow [38]. The notation oP (1) (respectively OP (1)) stands for a sequence of random
variables that converges to zero in probability (resp. is bounded in probability) as n → ∞.
More generally, for a sequence of random variables Rn ,
P
Xn = oP (Rn ) means Xn = Yn Rn with Yn → 0
Xn = OP (Rn ) means Xn = Yn Rn with Yn = OP (1).
13
For deterministic sequences Xn and Rn , the stochastic notation reduce to the usual o and
O. Throughout the paper, by K, we denote a generic constant (i.e. K may or may not be
the same at different occurrences) that does not depend on (Y, θ, σ 2 , n).
We also denote by δn and δ̄n two sequences of random variables satisfying
sup |δn | = OP (1)
(θ,σ 2 )∈J
and
sup |δ̄n | = o(n), a.s.
(θ,σ 2 )∈J
The definition of δn and δ̄n may change from one line to the other. Similarly, we denote
by δi,n a triangular array of deterministic scalars satisfying
|δi,n | 6 K.
sup
n∈N,i=1,...,n
(θ,σ 2 )∈J
The definition of δi,n may change from one line to the other, and possibly also in different
occurrences within the same line. We also use several times that,
n−1
X
∆i = 1,
(18)
i=2
sup |Y (t)| < +∞ a.s.
(19)
t∈[0,1]
Before turning to the proof of Theorems 3.2 and 4.1, we state five auxiliary lemmas that
will be required in the sequel.
Lemma 8.1.
(i) Let λi > 0, i = 1, 2 be fixed. Then as x → 0,
λ1
1 − e−λ1 x
=
+ O(x).
−λ
x
1−e 2
λ2
(ii) Let λ > 0 be fixed. Then as x, y → 0,
x+y
1 − e−λx
=
(1 + O(x + y)).
−λy
1−e
y
Lemma 8.2. For any constant δ > 0, there exists a constant η > 0 such that
1 + e−λy
inf (x − 1 − log x) > η.
|x−1|>δ
x>0
(20)
Lemma 8.3. Suppose that for each n, the random variables Z, Z1,n , . . . , Zn,n are independent and identically distributed and centered. Suppose also that for some t > 0 and
p > 0, E[exp{t|Z|p }] < ∞. Then for all α > 0, a.s
sup |Zk,n | = o(nα )
(21)
16k6n
n
X
Zk,n = o(n(1/2)+α ).
k=2
14
(22)
The proof of Lemma 8.1 is direct and these of Lemmas 8.2 and 8.3 can be found in [40].
Lemma 8.4. Let for any i ∈ {2, ..., n − 1},
ηi =
∆i+1
(1 + δi,n (∆i + ∆i+1 ))
∆i + ∆i+1
τi =
∆i
(1 + δi,n (∆i + ∆i+1 )) .
∆i + ∆i+1
and
Then
sup
(i)
(θ,σ 2 )∈J
sup
(ii)
(θ,σ 2 )∈J
sup
(iii)
(θ,σ 2 )∈J
sup
(iv)
(θ,σ 2 )∈J
Pn−1
ηi Wi,0 yi = OP (1);
Pn−1
ηi Wi yi = OP (1);
Pn−1
τi Wi+1,0 yi−1 = OP (1);
Pn−1
τi Wi+1 yi−1 = OP (1).
i=2
i=2
i=2
i=2
Proof of Lemma 8.4. We only prove (iii) and (iv), the proof of (i) and (ii) being similar.
(iii) We have
n−1
X
τi Wi+1,0 yi−1 6
i=2
n−1
X
i=2
Using the fact that
∆i
∆i +∆i+1
n−1
X
∆i
Wi+1,0 yi−1 + K
∆i |Wi+1,0 yi−1 | =: L1 + L2 .
∆i + ∆i+1
i=2
6 1 and that for i 6= j,
E[Wi+1,0 yi−1 Wj+1,0 yj−1 ] = 0,
we get
E[L21 ]
6
n−1
X
2
2
E[Wi+1,0
yi−1
]
i=2
n−1
X
6K
=
n−1
X
2
2
E[yi−1
]E[Wi+1,0
]
i=2
(1 − e
−2θ0 ∆i+1
)6K
i=2
n−1
X
∆i+1 = O(1),
i=2
using the fact that Wi+1,0 and yi−1 are independent and a Taylor expansion of ∆i+1 7→
1 − e−2θ0 ∆i+1 . Now, by Cauchy-Schwarz,
E(|L2 |) 6K
n−1
X
∆i E(|Wi+1,0 ||yi−1 |) 6 K
i=2
n−1
X
i=2
15
q
q
2
2
∆i E(Wi+1,0 ) E(yi−1
) = O(1).
Hence, since L1 and L2 do not depend on θ and σ 2 , we have sup
Pn−1
(θ,σ 2 )∈J
i=2
τi Wi+1,0 yi−1 =
OP (1).
(iv) Now, we use the decomposition
Wi+1 = Wi+1,0 + (e−θ0 ∆i+1 − e−θ∆i+1 )yi
(23)
to get
n−1
X
τi Wi+1 yi−1 =
n−1
X
τi Wi+1,0 yi−1 +
n−1
X
i=2
i=2
6
n−1
X
τi Wi+1,0 yi−1 +
i=2
τi (e−θ0 ∆i+1 − e−θ∆i+1 )yi yi−1
i=2
n−1
X
τi |e−θ0 ∆i+1 − e−θ∆i+1 ||yi yi−1 |.
i=2
The first sum is OP (1) by (iii). The second sum is OP (1) by using the fact that supt∈[0,1] Y 2 (t) <
∞ a.s. and
|τi ||e−θ0 ∆i+1 − e−θ∆i+1 |
sup
6 K.
∆i+1
n∈N,i=2,...,n−1,(θ,σ 2 )∈J
Pn−1
Hence sup
i=2 τi Wi+1 yi−1 = OP (1).
(θ,σ 2 )∈J
We can show, after some tedious but straightforward calculations, the following Taylor
expansions.
Lemma 8.5. Let
αi := (σ 2 (1 − e−2θ∆i ))−1 ,
αi,0 := (σ02 (1 − e−2θ0 ∆i ))−1 ,
∆i+1
∆i−1
qi :=
+
,
∆i + ∆i+1 ∆i + ∆i−1
Ai := αi + αi+1 e−2θ∆i+1 ,
1
e−2θ∆i
+
,
Ai
Ai−1
Ci := (αi αi+1 e−θ∆i+1 )/Ai ,
Bi :=
for i = 2, ..., n,
for i = 2, ..., n,
(25)
for i = 3, ..., n − 1,
(26)
for i = 2, ..., n − 1,
(27)
for i = 3, ..., n − 1,
(28)
for i = 2, ..., n − 1.
(29)
Let αi0 = ∂αi /∂θ, A0i = ∂Ai /∂θ, Bi0 = ∂Bi /∂θ and Ci0 = ∂Ci /∂θ. Note that we have
αi αi+1 A0i
e−θ∆i+1
0
0
0
Ci =
αi αi+1 + αi αi+1 − αi αi+1 ∆i+1 −
.
Ai
Ai
16
(24)
(30)
We have,
αi −
sup
n∈N,i=2,...,n,
(θ,σ 2 )∈J
sup
n∈N,i=2,...,n,
(θ,σ 2 )∈J
αi2 =
sup
n∈N,i=2,...,n−1,
(θ,σ 2 )∈J
sup
1
4σ 4 θ2 ∆2i
1
Ai − 2
2σ θ
1
− 2 2
2σ θ
A0i
n∈N,i=2,...,n−1,
(θ,σ 2 )∈J
αi0 −
1
2σ 2 θ∆i
1
2σ 2 θ2 ∆i
+ δi,n
1
1
+
∆i ∆i+1
1
1
+
∆i ∆i+1
6K
(31)
6K
(32)
1
∆i
(33)
6K
(34)
6K
(35)
1
∆i ∆i+1
A0i
= + δi,n
Ai
θ
∆i + ∆i+1
1
αi Bi0 = − qi + δi,n (∆i−1 + ∆i + ∆i+1 )
θ
1
αi0 Bi = qi + δi,n (∆i−1 + ∆i + ∆i+1 )
θ
1
1
6 K.
sup
Ci0 − 2 2
2σ θ ∆i + ∆i+1
n∈N,i=2,...,n−1,
(36)
(37)
(38)
(39)
(θ,σ 2 )∈J
8.2
Proof of Theorem 3.2
The existence of θ̂ and σ̂ 2 is a consequence of the fact that Sn is a continuous function
defined on a compact set. Now, it suffices to show (16) since θ̂1 → θ1 (resp. σ̂22 → σ22 ) is a
particular case of (16) with b = B = σ12 (resp. a = A = θ2 ). Moreover, in view of (5), the
result (16) holds if we can show that for every ε > 0, a.s.
inf
eσ
{Sn (θ, σ 2 ) − Sn (θ,
e2 )} → ∞
(θ,σ 2 )∈J,
eσ 2 |>ε
|θσ 2 −θe
(40)
eσ
eσ 2 = θ0 σ 2 .
where (θ,
e2 ) ∈ J can be any non-random vector such that θe
0
eσ
Let us compute the difference Sn (θ, σ 2 ) − Sn (θ,
e2 ) and determine the preponderant terms.
After some computations, we naturally have
yi −
αi e−θ∆i yi−1 + αi+1 e−θ∆i+1 yi+1
1
=
αi Wi − αi+1 e−θ∆i+1 Wi+1 ,
Ai
Ai
17
(41)
where αi and Ai have already been defined in Equations (24) and (27). Hence, from Lemma
3.1,
Sn (θ, σ 2 ) =n log(σ 2 ) + log(1 − e−2θ∆2 ) + log(1 − e−2θ∆n ) −
n−1
X
log σ 2 Ai
(42)
i=2
−θ∆2
2
Wn2
− e−2θ∆n )
y2 )
(y1 − e
+ 2
2
−2θ∆
2)
σ (1 − e
σ (1
n−1
X 1
2
+
αi Wi − αi+1 e−θ∆i+1 Wi+1 .
Ai
i=2
+
(43)
(44)
In the following, we prove that the terms in (43) and those obtained by developing (44),
except one, are o(n) uniformly in (θ, σ 2 ) ∈ J, a.s. More precisely, we establish the following
lemma (see the proof in Section 8.3).
Lemma 8.6. One has
2
2
Sn (θ, σ ) =n log(σ ) + log(1 − e
−2θ∆2
−2θ∆n
) + log(1 − e
)−
n−1
X
log σ 2 Ai
i=2
+
n−1
X
2
αi2 Bi Wi,0
+ δ̄n .
i=3
As a consequence, we find that,
n−1
1 − e−2θ∆2
1 − e−2θ∆n X
σ2
2
e
e
−
log Ai /Ai
+ log
+ log
Sn (θ, σ ) − Sn (θ, σ
e ) =n log
e 2
e n
1 − e−2θ∆
1 − e−2θ∆
σe2
2
i=2
+
n−1
X
ei W 2 + δ̄n
αi2 Bi − α
ei2 B
i,0
(45)
i=3
ei and A
ei are the analogs of αi , Bi and Ai defined in Equations (24), (27)
where α
ei , B
and (28) with θ = θe and σ 2 = σ
e2 . More precisely,h they are naturally
defined by α
ei =
i
e
e
e i+1
1
e−2θ∆i
2
−2θ∆i −1 e
−θ∆
e
(e
σ (1 − e
)) , Ai = α
ei + α
ei+1 e
and Bi = Ae + Ae
.
i
i−1
eσ
eσ 2 = θ0 σ 2 and making some more
Using the fact that (θ,
e2 ) has been chosen such as θe
0
computations, we get the following lemma (see the proof in Section 8.3).
Lemma 8.7. Uniformly in (θ, σ 2 ) ∈ J, a.s.
2
2
θ
σ
θ
σ
0
0
0
0
2
2
eσ
+ o(n).
Sn (θ, σ ) − Sn (θ,
e )=n
− 1 − log
θσ 2
θσ 2
18
Hence by Lemma 8.7, a.s.
2
eσ
{Sn (θ, σ ) − Sn (θ,
e )} >
inf
θ0 σ02
θ0 σ02
n
+ o(n)
− 1 − log
θσ 2
θσ 2
2
(θ,σ 2 )∈J,
eσ 2 |>ε
|θσ 2 −θe
inf
(θ,σ 2 )∈J,
eσ 2 |>ε
|θσ 2 −θe
(46)
which by Lemma 8.2, for every ε > 0, is strictly positive, for n large enough, a.s. Then the
proof of Theorem 3.2 is now complete.
8.3
Proofs of the lemmas of Section 8.2
Proof of Lemma 8.6. (i) First, we study the terms in (43). We have, from (31)
2
2
y1 − e−θ∆2 y2
y1 − e−θ∆2 y2
6K 2
σ 2 (1 − e−2θ∆2 )
σ0 (1 − e−2θ0 ∆2 )
2
2
y1 − e−θ0 ∆2 y2
e−θ0 ∆2 − e−θ∆2
6K 2
+K 2
y2
σ0 (1 − e−2θ0 ∆2 )
σ0 (1 − e−2θ0 ∆2 ) 2
2
6 K sup W̄i,0
+ K∆2 sup Y 2 (t)
26i6n
(47)
t∈[0,1]
= o(n) a.s.
(48)
from Lemma 8.3. The random variable Wn2 /(σ 2 (1 − e−2θ∆n )) can be treated in the same
manner leading to the same result.
(ii) Second, we turn to the term in (44) that we aim at approximating by a sum of independent random variables. In this goal, we first show the relation
2
2
+ 2[e−θ0 ∆i − e−θ∆i ]yi−1 Wi,0 .
Wi2 = Wi,0
+ [e−θ0 ∆i − e−θ∆i ]2 yi−1
(49)
Hence, by (49), one has
n−1
X
2
1
αi Wi − αi+1 e−θ∆i+1 Wi+1
Ai
i=2
=
=
=
n−1 2
X
α
i
i=2
n−1
X
i=3
n−1
X
i=3
+
Ai
αi2
Wi2
+
n−1 2
X
αi+1 e−2θ∆i+1
i=2
Ai
2
Wi+1
−2
n−1
X
αi αi+1 e−θ∆i+1
i=2
Ai
Wi Wi+1
n−1
X
1
e−2θ∆i
α22 2 αn2 e−2θ∆n 2
2
+
Wi +
W +
Wn − 2
Ci Wi Wi+1
Ai
Ai−1
A2 2
An−1
i=2
2
αi2 Bi Wi,0
+
n−1
X
αi2 Bi (e−θ0 ∆i
−
2
e−θ∆i )2 yi−1
i=3
α22 2 αn2 e−2θ∆n 2
W +
Wn − 2
A2 2
An−1
+2
n−1
X
αi2 Bi (e−θ0 ∆i − e−θ∆i )Wi,0 yi−1
i=3
n−1
X
Ci Wi Wi+1
i=2
=:Σ1 + Σ2 + 2Σ3 + T2 + Tn − 2Σ4 ,
(50)
19
where αi , Ai , Bi and Ci have been defined in (24), (27), (28) and (29). We prove that all
the previous terms are o(n), uniformly in θ and σ 2 a.s, except Σ1 that still appears in the
expression of Sn (θ, σ 2 ) in Lemma 8.6.
α2
A2
• Term T2 : For n large enough, since
6 1, we get
α22 2
W 6 α2 W22 .
A2 2
|T2 | =
Hence, we can show sup |T2 | = o(n) a.s. in the same way as for (47).
(θ,σ 2 )∈J
αn e−2θ∆n
An−1
• Term Tn : For n large enough, since
|Tn | =
6 1, we get
αn2 e−2θ∆n 2
Wn 6 αn Wn2 .
An−1
Hence, we can show sup |Tn | = o(n) a.s. in the same way as for (47).
(θ,σ 2 )∈J
• Term Σ2 : The deterministic quantity
αi
(e−θ0 ∆i
∆i
− e−θ∆i )2 is bounded for n large enough,
−2θ∆
uniformly in (θ, σ 2 ) ∈ J (trivial inequalities and (31)) while Aαii 6 1 and αiAe i−1 i 6 1. Then,
we are led to
n−1
n−1
X
X
αi
2
−θ0 ∆i
−θ∆i 2 2
2
αi Bi (e
−e
) yi−1 =
∆i αi Bi (e−θ0 ∆i − e−θ∆i )2 yi−1
∆i
i=3
i=3
2
6 K sup Y (t)
t∈[0,1]
n−1
X
∆i = K sup Y (t)2
i=3
t∈[0,1]
from which, by (19), we deduce sup Σ2 = o(n) a.s.
(θ,σ 2 )∈J
• Term Σ3 : By the Cauchy-Schwarz inequality,
|Σ3 | 6
n−1
X
αi4 2 −θ0 ∆i
Bi (e
− e−θ∆i )2
α
i,0
i=3
!1/2
n−1
X
!1/2
2
2
W i,0
yi−1
.
(51)
i=3
As already mentioned, the deterministic term
αi −θ0 ∆i
(e
− e−θ∆i )2
∆i
is bounded uniformly in (θ, σ 2 ) ∈ J. Furthermore, αi /αi,0 is bounded uniformly in (θ, σ 2 ) ∈
J from (31). Finally, αi2 Bi2 6 Kαi2 (1/αi )2 = K. Hence the first term on the right-hand side
P
2
of (51) is bounded uniformly in (θ, σ 2 ) ∈ J. Now Lemma 8.3 yields that n−1
i=3 (W i,0 − 1) =
Pn−1 2
o(n(1/2)+α ) a.s. for any α > 0 leading to i=3
W i,0 = O(n) a.s and
!1/2
!1/2
n−1
n−1
X
X
2
2
2
yi−1
W i,0
6 sup Y 2 (t)
W i,0
= O(n1/2 ) a.s.
i=3
t∈[0,1]
i=3
20
As a consequence, sup |Σ3 | = O(n1/2 ) a.s. and naturally sup |Σ3 | = o(n) a.s.
(θ,σ 2 )∈J
(θ,σ 2 )∈J
• Term Σ4 : Using the trivial equality ab = ab − a0 b0 + a0 b0 = a0 b0 + a0 (b − b0 ) + b(a − a0 ),
one gets
Wi Wi+1 = [yi − e−θ∆i yi−1 ][yi+1 − e−θ∆i+1 yi ] = [yi − e−θ0 ∆i yi−1 ][yi+1 − e−θ0 ∆i+1 yi ]
+ [yi − e−θ0 ∆i yi−1 ](e−θ0 ∆i+1 − e−θ∆i+1 )yi + [yi+1 − e−θ∆i+1 yi ](e−θ0 ∆i − e−θ∆i )yi−1
= Wi,0 Wi+1,0 + (e−θ0 ∆i+1 − e−θ∆i+1 )Wi,0 yi + (e−θ0 ∆i − e−θ∆i )Wi+1 yi−1 .
(52)
Thus Σ4 rewrites
n−1
X
Ci Wi,0 Wi+1,0 +
i=2
n−1
X
Ci (e−θ0 ∆i+1 − e−θ∆i+1 )Wi,0 yi +
n−1
X
i=2
Ci (e−θ0 ∆i − e−θ∆i )Wi+1 yi−1 . (53)
i=2
We can show that
Ci
1/2
1/2
αi,0 αi+1,0
√
σ02 θ0 ∆i ∆i+1
= 2 2
+ δi,n (∆i + ∆i+1 ).
σ θ ∆i + ∆i+1
Hence the first random variable of (53) rewrites
sup
(θ,σ 2 )∈J
n−1
X
Ci Wi,0 Wi+1,0 = sup
(θ,σ 2 )∈J
i=2
n−1
X
Ci
W i,0 W i+1,0
1/2 1/2
i=2 αi,0 αi+1,0
n−1
X
√
n−1 2
X
σ0 θ0 ∆i ∆i+1
6 sup
W i,0 W i+1,0 +
δi,n (∆i + ∆i+1 )|W i,0 W i+1,0 |
2 2
(θ,σ 2 )∈J i=2 σ θ ∆i + ∆i+1
i=2
n−1
n−1 √
X
X
∆i ∆i+1
6K
W i,0 W i+1,0 +
δi,n (∆i + ∆i+1 )|W i,0 W i+1,0 |
∆i + ∆i+1
i=2
i=2
n−1 √
X
∆i ∆i+1
=K
W i,0 W i+1,0 + δ̄n ,
∆i + ∆i+1
i=2
from (21). Now, we have
n−1 √
X
∆i ∆i+1
W i,0 W i+1,0
∆i + ∆i+1
i=2
√
∆i ∆i+1
6
W i,0 W i+1,0 +
∆
+
∆
i
i+1
i=2,...,n−1
i odd
X
21
√
∆i ∆i+1
W i,0 W i+1,0 .
∆
+
∆
i
i+1
i=2,...,n−1
i even
X
In each of the two sums above, the summands constitute two triangular arrays of independent random variables. Thus, applying Theorem
Pn−1 2.1 in [17] with an = n, each of the two
sums is a o(n) a.s. Hence finally sup
i=2 Ci Wi,0 Wi+1,0 = o(n) a.s.
(θ,σ 2 )∈J
Let us now address the second term in (53) that by the Cauchy-Schwarz inequality is
bounded from above by
n−1
X
Ci2 −θ0 ∆i+1
− e−θ∆i+1 )2
(e
α
i,0
i=2
!1/2
n−1
X
!1/2
2
yi2 W i,0
= o(n) a.s.
i=2
where the last equality comes from similar computations as from the term Σ3 above, and
from the fact that
sup sup
sup
n∈N 26i6n−1 (θ,σ 2 )∈J
Ci2
(e−θ0 ∆i+1 − e−θ∆i+1 )2 6 K.
αi,0 ∆i+1
The third term in the right-hand side of (53) yields
sup
(θ,σ 2 )∈J
sup
(θ,σ 2 )∈J
n−1
X
i=2
n−1
X
Ci (e−θ0 ∆i − e−θ∆i )Wi+1 yi−1 6
−θ0 ∆i
Ci (e
−θ∆i
−e
n−1
X
)Wi+1,0 yi−1 + sup
(θ,σ 2 )∈J
i=2
Ci (e−θ0 ∆i − e−θ∆i )2 yi yi−1 . (54)
i=2
Since trivially Ci (e−θ0 ∆i −e−θ∆i ) = δi,n , the second term in (54) is bounded by K
O(1) a.s.
The first term in (54) is bounded by
K
n−1
X
|Wi+1,0 yi−1 | 6 K sup |Y (t)|
t∈[0,1]
i=2
n−1
X
Pn−1
i=2
∆i |yi yi−1 | =
−1/2
|W̄i+1,0 |αi,0
i=2
v
v
u n−1
u n−1
uX
uX
−1
t
2
W̄i+1,0 t
αi,0
6 K sup |Y (t)|
t∈[0,1]
√
= O( n) a.s.
i=2
i=2
P
2
since n−1
i=2 W̄i+1,0 = O(n) a.s. has been shown when handling Σ3 after (51).
One may conclude that sup Σ4 = o(n) a.s. The proof of Lemma 8.6 is then complete.
(θ,σ 2 )∈J
Proof of Lemma 8.7. We address each of the terms in (45).
• Using Lemma 8.1(i), we get that
sup log
θ∈[a,A]
1 − e−2θ∆2
e 2
1 − e−2θ∆
22
= O(1).
−2θ∆n
In the same way, supθ∈[a,A] log 1−e
= O(1).
e n
1−e−2θ∆
• We have, using Lemmas 8.1 (i) and (ii),
"
!#
e i+1
−2θ∆
−2θ∆i+1
1
e
e
1
ei ) = log
log(Ai /A
+
+
/
e i
e i+1
1 − e−2θ∆i 1 − e−2θ∆i+1
1 − e−2θ∆
1 − e−2θ∆
!#
"
e
e i
−2θ∆i
−2θ∆
1 − e−2θ∆i
1
−
e
1
−
e
e
= log
+ log 1 + e−2θ∆i+1
/ 1 + e−2θ∆i+1
e i+1
1 − e−2θ∆i
1 − e−2θ∆i+1
1 − e−2θ∆
= log
θe
+ δi,n (∆i + ∆i+1 ).
θ
Thus, by summation we have,
sup
n−1
X
log
θ∈[a,A] i=2
Ai
θe
− n log = O(1) = o(n).
θ
Ãi
• We want to show that
sup
(θ,σ 2 )∈J
n−1
X
2
αi2 Bi Wi,0
−n
i=3
θ0 σ02
= o(n) a.s.
θσ 2
By (28), one has
n−1
X
n−1
n−1
X
X
αi2 1 2
αi2 e−2θ∆i 2
=
W i,0 +
W i,0
αi,0 Ai
αi,0 Ai−1
i=3
i=3
i=3
n−2 2
2
X
αi+1
α2
αi 1 2
e−2θ∆i+1 2
α2 e−2θ∆3 2
2
=
W i,0 +
W i+1,0 + 3
W3,0 + n−1 Wn−1,0
.
α
A
α
A
A
A
i,0
i
i+1,0
i
2
n−1
i=3
2
αi2 Bi Wi,0
Then we use (31) to develop αi /αi,0 (respectively αi+1 /αi+1,0 ) and Lemma 8.1 (ii) to develop
αi /Ai (respectively αi+1 e−2θ∆i+1 /Ai ). We get
θ0 σ02 ∆i+1
αi2 1
=
+ δi,n (∆i + ∆i+1 )
αi,0 Ai
θσ 2 ∆i + ∆i+1
(55)
2
αi+1
e−2θ∆i+1
θ0 σ02
∆i
=
+ δi,n (∆i + ∆i+1 ).
2
αi+1,0 Ai
θσ ∆i + ∆i+1
(56)
In addition, we easily show, as in (47), that
sup
(θ,σ 2 )∈J
23
α23 e−2θ∆3
2
W3,0
A2
= o(n) a.s. and that
α2n−1
2
Wn−1,0
An−1
sup
(θ,σ 2 )∈J
n−1
X
= o(n) a.s. Then,
X
n−2 2
n−2
2
X
αi+1
θ0 σ02
αi 1 2
e−2θ∆i+1 2
θ0 σ02
−n 2 =
W i,0 +
W i+1,0 −
+ δ̄n
2
θσ
α
A
α
A
θσ
i,0
i
i+1,0
i
i=3
i=3
i=3
n−2 2
n−2
2
−2θ∆
2
X
X
i+1
αi+1 e
αi 1 2
θ0 σ0
∆i
∆i+1
2
=
W i,0 +
W i+1,0 −
+
+ δ̄n
2
α
A
α
A
θσ
∆
+
∆
∆
+
∆
i,0
i
i+1,0
i
i
i+1
i
i+1
i=3
i=3
6
2
αi2 Bi Wi,0
n−2
n−2
2
X
X
αi+1
αi2 1
θ0 σ 2 ∆i+1
e−2θ∆i+1
θ0 σ 2
∆i
2
2
− 2 02
− 2 02
W i,0 +
W i+1,0
αi,0 Ai θ σ ∆i + ∆i+1
αi+1,0 Ai
θ σ ∆i + ∆i+1
i=3
i=3
+
6K
n−2
n−2
θ0 σ02 X ∆i+1
θ0 σ02 X
∆i
2
2
(W
−
1)
+
(W
i,0
i+1,0 − 1) + |δ̄n |
θσ 2 i=3 ∆i + ∆i+1
θσ 2 i=3 ∆i + ∆i+1
n−2
n−2
X
X
2
2
(∆i + ∆i+1 )W i+1,0 + |δ̄n |
(∆i + ∆i+1 )W i,0 + K
i=3
i=3
+
θ0 σ02
θσ 2
(57)
n−2
X
i=3
n−2
θ0 σ02 X
∆i+1
∆i
2
2
(W i,0 − 1) +
(W i+1,0 − 1) .
2
∆i + ∆i+1
θσ i=3 ∆i + ∆i+1
Let us show that the terms in the right-hand side of (57) are a.s. o(n). We have
n−2
X
2
(∆i + ∆i+1 )W i,0 6 K
i=3
sup
i=3,...,n−2
2
W i,0 6 1 +
sup
i=3,...,n−2
a.s. from Lemma 8.3. Similarly
n−2
X
2
(∆i + ∆i+1 )W i+1,0 = o(n)
i=3
a.s. Also, using theorem 2.1 in [17] with an = n, we have
n−2
X
i=3
a.s. and
n−2
X
i=3
∆i+1
2
(W i,0 − 1) = o(n)
∆i + ∆i+1
∆i
2
(W i+1,0 − 1) = o(n)
∆i + ∆i+1
a.s. Hence finally
sup
(θ,σ 2 )∈J
n−1
X
2
αi2 Bi Wi,0
−n
i=3
24
2
|W i,0 − 1| = o(n),
θ0 σ02
= o(n) a.s.
θσ 2
• We can now conclude the proof. We have
θ0 σ02
σ2
θ0 σ02
θ̃
θ0 σ02
σ2θ
Sn (θ, σ ) − Sn (θ̃, σ̃ ) = n log 2 − n log + n 2 − n
+ δn = n log 2 +
− 1 + δn ,
σ̃
θ
θσ
σ0 θ0
θσ 2
θ̃σ˜2
2
2
by reminding that θ̃σ˜2 = θ0 σ02 . The proof of Lemma 8.7 is thus complete.
8.4
Proof of Theorem 4.1
Let us first prove (9) in the case aB < θ0 σ02 ; Ab > θ0 σ02 . We shall then discuss the other
cases at the end. In that view, let
ψ(θ, σ 2 ) =
∂
Sn (θ, σ 2 ).
∂θ
(58)
Then from Theorem 3.2, a.s. for n large enough, θ̂ ∈ (a, A). Thus a.s. for n large enough
(θ̂, σ̂ 2 ) satisfies ψ(θ̂, σ̂ 2 ) = 0. We shall approximate ψ(θ, σ 2 ) uniformly in (θ, σ 2 ) ∈ J.
Starting from (42), (43), (44) and (50) we can write
ψ(θ, σ 2 ) =
where
◦
Σ01
=
∂
Σ
∂θ 1
=
2∆2 e−2θ∆2 2∆n e−2θ∆n 2∆2 e−θ∆2 y2 (y1 − e−θ∆2 y2 )
+
+
1 − e−2θ∆2
1 − e−2θ∆n
σ 2 (1 − e−2θ∆2 )
2∆2 e−2θ∆2 (y1 − e−θ∆2 y2 )2 2∆n e−θ∆n yn−1 Wn
2∆n e−2θ∆n Wn2
−
+
−
σ 2 (1 − e−2θ∆2 )2
σ 2 (1 − e−2θ∆n )
σ 2 (1 − e−2θ∆n )2
n−1
X
A0i
−
+ Σ01 + Σ02 + 2Σ03 + T20 + Tn0 − 2Σ04
A
i
i=2
n−1
X
2
αi Di Wi,0
;
i=3
◦ Σ02 =
∂
Σ
∂θ 2
=
n−1
X
2
αi Di (e−θ0 ∆i − e−θ∆i ) + 2αi Bi ∆i e−θ∆i yi−1
(e−θ0 ∆i − e−θ∆i );
i=3
n−1
X
αi Di (e−θ0 ∆i − e−θ∆i ) + αi Bi ∆i e−θ∆i Wi,0 yi−1 ;
i=3
0
2∆2 e−θ∆2 y1 W2
W22
2
2 A2
∂
0
0
+α
− W2 2 ;
◦ T2 = ∂θ T2 = 2α2 α2
A2 2
A2
A2
0
−2θ∆n
A
α
e
n
n−1
0
−θ∆n
∂
0
◦ Tn = ∂θ Tn =
(2αn − 2αn ∆n −
αn )Wn + 2αn ∆n e
yn−1 Wn ;
An−1
An−1
n−1
n−1
n−1
X
X
X
0
−θ∆i
∂
0
◦ Σ4 = ∂θ Σ4 =
Ci Wi Wi+1 +
Ci ∆i e
Wi+1 yi−1 +
Ci ∆i+1 e−θ∆i+1 Wi yi
◦ Σ03 =
∂
Σ
∂θ 3
=
i=2
i=2
i=2
where Ci0 is the derivative of Ci w.r.t. θ defined in (30) and
Di := 2αi0 Bi + αi Bi0 ,
for i = 3, . . . , n − 1.
(59)
First, we consider the terms Σ01 and Σ04 in the following lemma (proved in Section 8.6).
25
Lemma 8.8. We have
Σ01 =
n−1
θ0 σ02 X
2
qi W i,0 + δn
2
2
θ σ i=3
and
Σ04
n−1 √
θ0 σ02 X ∆i ∆i+1
= 2 2
W i,0 W i+1,0 + δn ,
θ σ i=3 ∆i + ∆i+1
where qi and Ci0 have been defined in (26) and (30).
Now we prove that
the remaining terms in ψ(σ 2 , θ) are OP (1), uniformly in (θ, σ 2 ) ∈ J, at
Pn−1
the exception of i=2 A0i /Ai , leading to the following lemma (proved in Section 8.6).
Lemma 8.9. We obtain
√
n−1
θ0 σ02 X
∆i ∆i+1
n
2
ψ(θ, σ ) = 2 2
qi W i,0 − 2
W i,0 W i+1,0 − + δn .
θ σ i=3
∆i + ∆i+1
θ
2
Since θ̂2 σ̂ 2 ψ(θ̂, σ̂ 2 ) = 0 with probability going to 1, and since we can show that
n + O(1), we have
2
n(θ̂σ̂ −
θ0 σ02 )
=
θ0 σ02
n−1
X
2
qi (W i,0
i=3
√
∆i ∆i+1
− 1) − 2
W i,0 W i+1,0 + OP (1).
∆i + ∆i+1
We want to establish a Central
Limit Theorem for
√
2
Pn−1
√
i=3
qi =
(60)
n(θ̂σ̂ 2 − θ0 σ02 ). In that view, we define
∆i ∆i+1
Xi := qi (W i,0 − 1) − 2 ∆i +∆i+1 W i,0 W i+1,0 and we apply Theorem 2.1 in [26] for weakly
dependent variables (since Xi is not necessarily independent with Xi−1 and Xi+1 but is
independent with Xk for |i − k| > 2).
Note that we can show easily that τn2 =
1
n
Var(
Pn−1
i=3
Xi ), and assume
√ (θ̂σ̂ 2 − θ0 σ02 ) D
n
6
→
N (0, 1).
θ0 σ02 τn
n→∞
By Proposition 4.3, we can extract a subsequence n so that τ2n → τ 2 with τ 2 ∈ [2, 4]
n→∞
and so that
√ (θ̂σ̂ 2 − θ0 σ02 ) D
n
6
→
N (0, 1).
θ0 σ02 τεn
n→∞
26
√
The triangular array Xi / εn i=3,...,εn −1 satisfies the conditions of [26, Theorem 2.1], thus
we obtain
n −1
1 X
D
Xi → N (0, τ 2 ).
√
n→∞
n i=3
Now, from (60),
εn −1
εn −1
εn −1
√ (θ̂σ̂ 2 − θ0 σ02 )
1 X
1 X
1
1
1 X
Xi
Xi
=
+
o
(1)
=
+
−
n
Xi + oP (1).
√
√
√
P
θ0 σ02 τεn
n i=3 τεn
n i=3 τ
τn
τ
n i=3
Since
√1
n
Pεn −1
i=3
Xi = OP (1) and
1
τn
−
1
τ
= o(1), we get by Slutsky’s lemma
√ (θ̂σ̂ 2 − θ0 σ02 ) D
→ N (0, 1),
n
n→∞
θ0 σ02 τεn
which is contradictory and ends the proof of (9).
Now (10) is under consideration only when b = B = σ12 and so when aB < θ0 σ02 ; Ab > θ0 σ02 .
Thus (10) is a special case of (9). Now, when aB > θ0 σ02 ; Ab < θ0 σ02 , we have almost
surely for n large enough (∂/∂σ 2 )Sn (θ̂, σ̂ 2 ) = 0, so that the estimator σ̂22 can be expressed
explicitly, by differentiating the terms in (42), (43) and (44) w.r.t. σ 2 . Hence, (9) can
be proved in the case aB > θ0 σ02 ; Ab < θ0 σ02 by using identical techniques as in the
case aB < θ0 σ02 ; Ab > θ0 σ02 . We omit the details to save space. Finally, (11) is under
consideration only when a = A = θ2 and so when aB > θ0 σ02 ; Ab < θ0 σ02 . Thus (11) is a
special case of (9).
8.5
Proof of Propositions 4.3 and 4.5
Proof of Proposition 4.3. We have
τn2
n−1
2X 2
∆i ∆i+1
=
q +2
.
n i=3 i
(∆i + ∆i+1 )2
∆i−1
∆i+1
, note that
= 1 − ai−1 and qi =
∆i + ∆i+1
∆i + ∆i−1
ai + 1 − ai−1 . First, we have after some trivial computations,
(i) Upper bound for τn2 . Let ai =
n−1
τn2
2X
=
[(ai + 1 − ai−1 )2 + 2ai (1 − ai )]
n i=3
n−1
2X
=
(1 + 2ai − 2ai ai−1 ) + o(1)
n i=3
62 + 4m + o(1),
27
(61)
n−1
1 X
where m :=
ai .
n − 3 i=3
Also, since for k = 2, ..., n − 1, 0 6 ak 6 1, we have 1 + 2(1 − ai−1 )ai 6 3 − 2ai−1 . Thus,
from (61),
n−1
τn2 6
2X
(3 − 2ai−1 ) + o(1) = 6 − 4m + o(1).
n i=3
Finally, τn2 6 min (2 + 4m, 6 − 4m) + o(1). Since sup min (2 + 4m, 6 − 4m) = 4, τn2 6
m∈[0,1]
4 + o(1).
n−1
1X
∆i ∆i+1
Note that
qi = 1 + o(1). Since
> 0, we get
n i=3
(∆i + ∆i+1 )2
τn2 .
(ii) Lower bound for
n−1
2X 2
q >2
τn2 >
n i=3 i
n−1
1X
qi
n i=3
!2
+ o(1) = 2 + o(1).
Proof of Proposition 4.5. (i) After some computation, we have
τn2 =4γn2 − 4γn + 4 + o(1).
Since γn = 1/n, then τn2 → 4.
n→∞
(ii) We have
2
τn2 6
n
n−1
X
i=bnα c+2
i
1
+
i+2 i+1
2
2(i + 1)
+
(i + 2)2
!
+ o(1) 6 2 + o(1).
As a consequence, this particular design realizes τn2 = 2 + o(1) by (62).
8.6
Proofs of the lemmas of Section 8.4
Proof of Lemma 8.8. (i) From (37) and (38),
n−1
Σ01 =
n−1
X αi
1 X αi
2
2
qi W i,0 +
δi,n (∆i−1 + ∆i + ∆i+1 )W i,0
θ i=3 αi,0
αi,0
i=3
where qi has been defined in (26). By (31), we have
αi
θ0 σ 2
= 2 02 + δi,n ∆i .
αi,0
σ θ
28
(62)
2
Moreover, since E[W i,0 ] = 1, one clearly has
n−1
X
αi
2
δi,n (∆i−1 + ∆i + ∆i+1 )W i,0 = OP (1)
αi,0
i=3
that leads to the desired result.
P
0
(ii) Now we study the first sum of Σ04 that rewrites n−1
i=2 Ci (Mi,1 + Mi,2 + Mi,3 ) using (52)
and where Ci0 has been defined in (30) and
W i,0 W i+1,0
Mi,1 := α1/2 α1/2
i,0
i+1,0
Mi,2 := (e−θ0 ∆i+1 − e−θ∆i+1 )Wi,0 yi
Mi,3 := (e−θ0 ∆i − e−θ∆i )Wi+1 yi−1 .
• First, we consider
n−1
X
Ci0 Mi,1 . By (39) and (31) we can show
i=2
Ci0
1/2
1/2
αi,0 αi+1,0
√
θ0 σ02 ∆i ∆i+1
= 2 2
+ δi,n (∆i + ∆i+1 ).
θ σ ∆i + ∆i+1
Furthermore we have
"
E
n−1
X
#
δi,n (∆i + ∆i+1 )W i,0 W i+1,0
6K
i=2
=K
n−1
X
i=2
n−1
X
q
2
2
(∆i + ∆i+1 ) E[W i,0 ]E[W i+1,0 ]
(∆i + ∆i+1 ) 6 K.
i=2
Thus
n−1
X
Ci0
i=2
αi,0 αi+1,0
1/2
Hence
1/2
n−1
X
W i,0 W i+1,0
Ci0 Mi,1
i=2
n−1 √
θ0 σ02 X ∆i ∆i+1
= 2 2
W i,0 W i+1,0 + δn .
θ σ i=2 ∆i + ∆i+1
n−1 √
θ0 σ02 X ∆i ∆i+1
= 2 2
W i,0 W i+1,0 + δn .
θ σ i=2 ∆i + ∆i+1
• Second, one clearly has
n−1
X
Ci0 Mi,2
i=2
Hence sup
(θ,σ 2 )∈J
Pn−1
i=2
n−1
θ − θ0 X ∆i+1
= 2 2
(1 + δi,n (∆i + ∆i+1 ))Wi,0 yi .
2θ σ i=2 ∆i + ∆i+1
Ci0 Mi,2 = OP (1) by Lemma 8.4 (i).
29
• Third, we get
n−1
X
Ci0 Mi,3
i=2
and sup
(θ,σ 2 )∈J
Pn−1
i=2
n−1
θ − θ0 X
∆i
= 2 2
(1 + δi,n (∆i + ∆i+1 )) Wi+1 yi−1
2θ σ i=2 ∆i + ∆i+1
Ci0 Mi,3 = OP (1) by Lemma 8.4 (iv).
(iii) We now consider the second and third sums in Σ04 .
Using (31) and (34), we can show
Ci ∆i e−θ∆i =
1
∆i
(1 + δi,n (∆i + ∆i+1 ))
2θσ 2 ∆i + ∆i+1
and
Ci ∆i+1 e−θ∆i+1 =
∆i+1
1
(1 + δi,n (∆i + ∆i+1 )).
2
2θσ ∆i + ∆i+1
Hence by Lemma 8.4 (iv) and (ii), sup
(θ,σ 2 )∈J
n−1
X
Ci (∆i e−θ∆i Wi+1 yi−1 + ∆i+1 e−θ∆i+1 Wi yi ) =
i=2
OP (1).
Proof of Lemma 8.9. • We have
sup
a6θ6A
2∆2 e−2θ∆2 2∆n e−2θ∆n
= O(1).
+
1 − e−2θ∆2
1 − e−2θ∆n
• For n large enough,
sup
(θ,σ 2 )∈J
2∆2 e−θ∆2 y2 [y1 − e−θ∆2 y2 ]
6 K sup |y2 ||y1 − e−θ∆2 y2 | 6 K sup Y (t)2 = OP (1).
2
−2θ∆
2
σ (1 − e
)
a6θ6A
t∈[0,1]
2
+ (e−θ0 ∆i − e−θ∆i )yi−1 Wi + Wi,0 (e−θ0 ∆i − e−θ∆i )yi−1 we can eas• Using Wi2 = Wi,0
−2θ∆2
−θ∆2 y )2
2∆n e−θ∆n yn−1 Wn
1 −e
2
ily show sup 2∆2 e σ2 (1−e(y−2θ∆
=
O
(1),
sup
= OP (1) and
P
2
2)
σ 2 (1 − e−2θ∆n )
(θ,σ 2 )∈J
(θ,σ 2 )∈J
sup
(θ,σ 2 )∈J
2∆n e−2θ∆n Wn2
σ 2 (1−e−2θ∆n )2
= OP (1).
• Term Σ02 : First, using (37), (38) and the definition (26) of qi , the deterministic quantity
Di is bounded uniformly in (θ, σ 2 ) ∈ J, and so is (e−2θ0 ∆i − e−2θ∆i )2 αi /∆i from (31). By
(19), we are led to
sup
(θ,σ 2 )∈J
n−1
X
i=3
∆i
αi −θ0 ∆i
2
(e
− e−θ∆i )2 yi−1
Di = OP (1).
∆i
30
Similarly,
n−1
X
sup
(θ,σ 2 )∈J
2
= OP (1) and thus
2αi2 Bi ∆i e−θ∆i (e−θ0 ∆i − e−θ∆i )yi−1
i=3
sup |Σ02 | =
(θ,σ 2 )∈J
OP (1).
• Term Σ03 : First, from (37), (38) and (31), we have
αi Di (e−θ0 ∆i − e−θ∆i ) −
sup
n∈N,i=2,...,n−1,(θ,σ 2 )∈J
θ − θ0
qi 6 K.
2σ 2 θ2
Hence, proceeding as in the proof of Lemma 8.4, we can show
sup
(θ,σ 2 )∈J
n−1
X
αi Di (e−θ0 ∆i − e−θ∆i )Wi yi−1 = OP (1).
i=3
Second, we can show
sup
n∈N,i=3,...,n−1,(θ,σ 2 )∈J
αi2 Bi ∆i e−θ∆i
1
− 2
2σ θ
∆i+1
∆i−1
+
∆i + ∆i+1 ∆i + ∆i−1
6 K.
Hence we can show
n−1
X
sup
(θ,σ 2 )∈J
αi2 Bi ∆i e−θ∆i Wi,0 yi−1 = OP (1),
i=3
as in the proof of Lemma 8.4. Hence finally, sup Σ03 = OP (1).
(θ,σ 2 )∈J
• Term T20 : From (27), we have
|T20 | 6 2|α20 |W22 + |α2 |∆2 e−θ∆2 |y1 | + W22 |A02 |.
We can show
sup (|α20 |W22 + W22 |A02 |) = OP (1)
(θ,σ 2 )∈J
2
by using W22 = W2,0
+ (e−θ0 ∆2 − e−θ∆2 )y1 W2 + W2,0 (e−θ0 ∆2 − e−θ∆2 )y1 .
Finally, sup |α2 |∆2 e−θ∆2 = OP (1), which finally shows sup |T20 | = OP (1).
(θ,σ 2 )∈J
(θ,σ 2 )∈J
• Term Tn0 : Using (31), (34) and (35), we get
sup |2αn ∆n | 6 K
sup |2αn ∆n e−2θ∆n | 6 K.
and
(θ,σ 2 )∈J
(θ,σ 2 )∈J
−2θ∆n
Moreover, one has | αnAe n−1 | 6 1. Finally, we have sup Wn2 = OP (1) and sup |yn−1 Wn | =
(θ,σ 2 )∈J
OP (1). Hence, in order to show
sup
(θ,σ 2 )∈J
|Tn0 |
= OP (1) it remains to show
31
(θ,σ 2 )∈J
sup |αn0 Wn2 | =
(θ,σ 2 )∈J
sup |(A0n−1 /An−1 )αn Wn2 | =
(θ,σ 2 )∈J
− e−θ∆n )yn−1 Wn + Wn,0 (e−θ0 ∆n −
n−1
X
A0i
OP (1) and
2
OP (1). This is shown by using Wn2 = Wn,0
+
(e−θ0 ∆n
e−θ∆n )yn−1 .
• Term
i=2
Ai
: By (36),
n−1
X
A0
i
i=2
Ai
=
n
+ δn .
θ
Finally,
n−1
n−1 √
X
n
∆i ∆i+1
θ0 σ02 X
2
qi W i,0 − 2
W i,0 W i+1,0 − + δn
ψ(σ , θ) = 2 2
θ σ i=3
∆i + ∆i+1
θ
i=3
2
using Lemma 8.8.
8.7
Proof of Theorems 6.1 and 6.2
In this section and the next one, we let ||A|| denote the largest singular value of a matrix
A and ||v|| denote the Euclidean norm of a vector ||v||. Finally, for k ∈ {1, ..., p}, we let
f (k) = (fk (s1 ), ..., fk (sn ))0 . We first provide a decomposition of S̄n in the following lemma.
Lemma 8.10. We have
S̄n (θ, σ 2 ) = Sn (θ, σ 2 ) − r1 (θ) +
r2 (θ) + 2r3 (θ) − r4 (θ)
,
σ2
with
r1 (θ) =
n
X
¯−i − log (R−1
,
log (R−1
θ )ii −
θ )ii
i=1
n
X
2
r2 (θ) =
(R−1
θ )ii −i ,
r3 (θ) =
i=1
n
X
(Rθ−1 )ii −i (yi − Ŷθ,−i (si )),
i=1
r4 (θ) =
−i =
¯−i =
n
X
¯−i (yi − Ŷθ,−i (si ) + −i )2 ,
i=1
[f 0i − r 0θ,−i R−1
θ,−i F −i ](β 0 − β̂ −i ),
−1
−1 0 −1
0
ei,n Rθ F (F 0 R−1
θ F ) F Rθ ei,n ,
where ei,n is the i-th base column vector of Rn . We remark that rk (θ) does not depend on
σ 2 for k = 1, ..., 4.
32
We now show that, in Lemma 8.10, the term Sn only, corresponding to the zero-mean case,
is preponderant.
Lemma 8.11. With the notation of Lemma 8.10, we have for k = 1, ..., 4
sup |rk (θ)| = O(1) a.s.
θ∈[a,A]
Because of Lemmas 8.10 and 8.11, we have a.s.
sup
S̄n (θ, σ 2 ) − Sn (θ, σ 2 ) = O(1).
(θ,σ 2 )∈J
Hence Theorem 6.1 follows from (46).
Also, from Lemmas 3.1, 8.10 and 8.11, we have
S̄n (θ, σ 2 ) = n log(σ 2 ) + log(1 − e−2θ∆2 ) + log(1 − e−2θ∆n )
n−1
X
1
e−2θ∆i+1
−
log
+
− r1 (θ)
−2θ∆i
−2θ∆i+1
1
−
e
1
−
e
i=2
(
1 (y1 − e−θ∆2 y2 )2 (yn − e−θ∆n yn−1 )2
+
+ 2
σ
(1 − e−2θ∆2 )
(1 − e−2θ∆n )
e−θ∆i
n−1
−2θ∆
X
i+1
−2θ∆i yi−1 +
e
1
1−e
yi −
+
+
1
1 − e−2θ∆i 1 − e−2θ∆i+1
−2θ∆ +
i=2
1−e
i
2
e−θ∆i+1
y
1−e−2θ∆i+1 i+1
−2θ∆i+1
e
1−e−2θ∆i+1
)
+ δn (θ) ,
where δn (θ) does not depend on σ 2 and satisfies supa6θ6A |δn (θ)| = OP (1). Since θ̌σ̌ 2 → θ0 σ02
a.s. from Theorem 6.1, and since aB > θ0 σ02 ; Ab < θ0 σ02 , we have [∂/∂σ 2 ]S¯n (θ̌, σ̌ 2 ) = 0 for
n large enough almost surely. Hence we obtain for n large enough almost surely,
(
(y1 − e−θ̌∆2 y2 )2 (yn − e−θ̌∆n yn−1 )2
+
(1 − e−2θ̌∆2 )
(1 − e−2θ̌∆n )
"
#
e−θ̌∆i
n−1
−2θ̌∆i+1
y +
X
−2θ̌∆i i−1
1
e
yi − 1−e
+
+
1
1 − e−2θ̌∆i 1 − e−2θ̌∆i+1
+
i=2
1
n
− 4
0 =
2
σ̌
σ̌
1−e−2θ̌∆i
2
e−θ̌∆i+1
y
i+1
1−e−2θ̌∆i+1
e−2θ̌∆i+1
1−e−2θ̌∆i+1
)
+ δn (θ̌) .
As noted at the end of the proof of Theorem 4.1, using identical techniques as for proving
Theorem 4.1, one can finish the proof of Theorem 6.2.
8.8
Proofs of the lemmas in Section 8.7
Proof of Lemma 8.10. We have, with y −i = (y1 , ..., yi−1 , yi+1 , ..., yn )0 ,
Z(si ) − Ẑθ,−i (si ) = yi + f 0i β 0 − f 0i β̂ −i − r 0θ,−i R−1
θ,−i (y −i + F −i β 0 − F −i β̂ −i )
0
−1
0
= yi − r 0θ,−i R−1
θ,−i y −i + (f i − r θ,−i Rθ,−i F −i )(β 0 − β̂ −i )
= yi − Ŷθ,−i (si ) + −i .
33
Also
−1
−1
0 −1
−1
0
0
−1 0 −1
(Q−
¯−i .
θ )ii = ei,n Rθ ei,n − ei,n Rθ F (F Rθ F ) F Rθ ei,n = (Rθ )ii −
Then, from (15),
"
#
2
(z
−
Ẑ
(s
))
i
θ,−i
i
2
S̄n (θ, σ 2 ) =
log(σ̌θ,σ
2 ,−i (si )) +
2
σ̌θ,σ
2 ,−i (si )
i=1
n
n
X
1
1 X −
2
= n log(σ ) +
log
+ 2
(Qθ )ii [yi − Ŷθ,−i (si ) + −i ]2
−
σ i=1
(Qθ )ii
i=1
n
X
= Sn (θ, σ 2 ) −
n
n
X
1 X −1 2
−1
(Rθ )ii −i
log (R−1
)
−
¯
−
log((R
)
)
+
−i
θ ii
θ ii
2
σ
i=1
i=1
n
n
2 X −1
1 X
+ 2
(R )ii −i (yi − Ŷθ,−i (si )) − 2
¯−i (yi − Ŷθ,−i (si ) + −i )2
σ i=1 θ
σ i=1
= Sn (θ, σ 2 ) − r1 (θ) +
r2 (θ) + 2r3 (θ) − r4 (θ)
.
σ2
Before proving Lemma 8.11, we state and prove some intermediary results.
Lemma 8.12. We have
||F 0 R−1
θ y|| 6 K sup |Y (t)|.
t∈[0,1]
Proof of Lemma 8.12. From (1) we have, with k = 1, ..., p and with f = (f1 , ..., fn )0 = f (k) ,
0 −1
[F 0 R−1
θ y]k = f Rθ y
n
n
X
X
1
1
e−θ∆i
e−θ∆i
=
f
y
+
f
y
−
f
y
−
fi−1 yi
1
1
n
n
i
i−1
1 − e−2θ∆2
1 − e−2θ∆n
1 − e−2θ∆i
1 − e−2θ∆i
i=2
i=2
n−1
X
1
e−2θ∆i+1
+
+
fi y i
1 − e−2θ∆i 1 − e−2θ∆i+1
i=2
=
f1 y1 − f2 y1 e−θ∆2 − f1 y2 e−θ∆2 + f2 y2
1 − e−2θ∆2
n
X
fi yi − fi yi−1 e−θ∆i − fi−1 yi e−θ∆i + e−2θ∆i fi−1 yi−1
+
1 − e−2θ∆i
i=3
= a1 +
n
X
mi ,
i=3
34
say. We have, using (31), and that fk is continuously differentiable,
1
a1 =
+ δ1,n f1 y1 (1 − e−θ∆2 ) + (f1 − f2 )y1 e−θ∆2 + y2 f2 (1 − e−θ∆2 ) + e−θ∆2 y2 (f2 − f1 )
2θ∆2
1
f1 − f2
(f1 y1 + y2 f2 ) +
(y1 − y2 ) + δ1,n ∆2 sup |Y (t)|
=
2
2θ∆2
t∈[0,1]
and
−2θ∆i
fi − e−θ∆i fi−1
e
fi−1 − e−θ∆i fi
= yi
+ yi−1
1 − e−2θ∆i
1 − e−2θ∆i
fi − fi−1 + fi−1 (1 − e−θ∆i )
fi−1 (e−2θ∆i − e−θ∆i ) + (fi−1 − fi )e−θ∆i
= yi
+ yi−1
1 − e−2θ∆i
1 − e−2θ∆i
fi − fi−1 1
+ (yi − yi−1 )fi−1 + δi,n ∆i sup |Y (t)|.
= (yi − yi−1 )
2θ∆i
2
t∈[0,1]
mi
Hence, using that fk is continuously differentiable,
f 0 R−1
θ y =
f1 − f2
1
(f1 y1 + y2 f2 ) +
(y1 − y2 ) + δ1,n ∆2 sup |Y (t)|
2
2θ∆2
t∈[0,1]
!
fi − fi−1 1
+
(yi − yi−1 )
+ (yi − yi−1 )fi−1 + δi,n ∆i sup |Y (t)|
2θ∆i
2
t∈[0,1]
i=3
n−1
n−1
X
1X
fi − fi−1
(yi − yi−1 ) +
= δ1,n sup |Y (t)| +
(yi − yi−1 ) fi−1 .(63)
2θ∆i
2 i=3
t∈[0,1]
i=3
n
X
The absolute value of first sum in (63) is equal to, after a summation by part, and using
that fk is twice continuously differentiable,
n−1
X
fi+1 − fi fi − fi−1
−
yi
−
+ δn,n sup |Y (t)|
2θ∆i+1
2θ∆i
t∈[0,1]
i=3
n−1
1 X
6
yi (f 0 (si ) + δi,n ∆i+1 − f 0 (si ) + δi,n ∆i ) + δn,n sup |Y (t)|
2θ i=3
t∈[0,1]
= δn,n sup |Y (t)|.
t∈[0,1]
The second sum in (63) is equal to, using that fk is continuously differentiable,
n−1
X
(yi fi−1 − yi−1 fi−2 + yi−1 (fi−2 − fi−1 )) = δn,n sup |Y (t)|.
t∈[0,1]
i=3
Hence
|f 0 R−1
θ y| 6 K sup |Y (t)|.
t∈[0,1]
35
Lemma 8.13. We have, for k = 1, ..., p,
(
K(∆i + ∆i+1 ) when i 6∈ {1, n}
(k)
−1
|e0i,n Rθ f | 6
.
K
when i ∈ {1, n}
Proof of Lemma 8.13. Let f = (f1 , ..., fn )0 = f (k) .
i) When i ∈
/ {1, n}, from (1),
e0i,n R−1
θ f
=
=
=
=
e−θ∆i
1
e−2θ∆i+1
e−θ∆i+1
fi−1 +
fi +
fi −
fi+1
−
1 − e−2θ∆i
1 − e−2θ∆i
1 − e−2θ∆i+1
1 − e−2θ∆i+1
fi (1 − e−θ∆i ) + e−θ∆i (fi − fi−1 ) fi (e−2θ∆i+1 − e−θ∆i+1 ) + (fi − fi+1 )e−θ∆i+1
+
1 − e−2θ∆i
1 − e−2θ∆i+1
0
0
fi f (si )
fi f (si )
+
+ δi,n ∆i − −
+ δi,n ∆i+1
2
2θ
2
2θ
δi,n (∆i + ∆i+1 ),
where we have used (31) and that f is twice continuously differentiable.
ii) Similarly,
e01,n R−1
θ f =
f (0) f 0 (0)
−
+ δ1,n ∆2 .
2
2θ
and
e0n,n R−1
θ f
f (1) f 0 (1)
+
+ δn,n ∆n .
=
2
2θ
Lemma 8.14. We have
F 0 R−1
θ F = I f + W (θ)
with supθ∈[a,A] ||W (θ)|| →n→∞ 0 and with I f the p × p matrix defined by
1
[I f ]kl = fk (0)fl (0)+
2θ
Z
1
θ
fk0 (t)fl0 (t)dt+
0
2
Z
1
0
1
fk (t)fl (t)dt+
2
Z
1
(fk0 (t)fl (t)+fk (t)fl0 (t))dt.
0
Furthermore, I f is invertible.
Proof of Lemma 8.14. Let k, l ∈ {1, ..., p} and let g = fk , h = fl , g = (g1 , ..., gn )0 = f (k)
36
and h = (h1 , ..., hn )0 = f (l) . From (1), we have,
0 −1
[F 0 R−1
θ F ]kl = g Rθ h
n
n
X
X
g1 h1
gn hn
e−θ∆i
e−θ∆i
=
+
−
g
h
−
gi−1 hi
i
i−1
1 − e−2θ∆2 1 − e−2θ∆n
1 − e−2θ∆i
1 − e−2θ∆i
i=2
i=2
n−1
X
e−2θ∆i+1
1
+
+
gi hi
1 − e−2θ∆i 1 − e−2θ∆i+1
i=2
!
n
gi − e−θ∆i gi−1 hi − e−θ∆i hi−1
g1 h1 − e−θ∆2 (g2 h1 + g1 h2 ) + g2 h2 X
+
=
1 − e−2θ∆2
1 − e−2θ∆i
i=3
=
(g1 − g2 )(h1 − h2 ) + θ∆2 (g2 h1 + g1 h2 ) + δ1,n ∆22
2θ∆2 + δ1,n ∆22
n
X
1
+
(gi − gi−1 )(hi − hi−1 ) + gi−1 (1 − e−θ∆i )hi−1 (1 − e−θ∆i ) (64)
−2θ∆i
1
−
e
i=3
+gi−1 (1 − e−θ∆i )(hi − hi−1 ) + (gi − gi−1 )hi−1 (1 − e−θ∆i ) ,
where we have used (31). Since fk and fl are continuously differentiable we have
(g1 − g2 )(h1 − h2 ) + θ∆2 (g2 h1 + g1 h2 )
= g(0)h(0) + δ1,n ∆2 .
2θ∆2 + δ1,n ∆22
Also, the element i of the sum in (64) is equal to
1
1
1 (gi − gi−1 )
θ
(hi − hi−1 ) + gi−1 hi−1 ∆i + gi−1 (hi − hi−1 ) + hi−1 (gi − gi−1 ) + δi,n ∆2i
2θ
∆i
2
2
2
1 0
θ
1
1
=
g (si )h0 (si )∆i + g(si−1 )h(si−1 )∆i + g(si )h0 (si )∆i + h(si )g 0 (si )∆i + δi,n ∆2i ,
2θ
2
2
2
since f is twice continuously differentiable. Hence we have
g 0 R−1
θ h = g(0)h(0) + δ1,n ∆2
n
X
1 0
θ
1 0
1
0
0
2
+
∆i
g (si )h (si ) + g(si−1 )h(si−1 ) + g (si )h(si ) + g(si )h (si ) + δi,n ∆i
2θ
2
2
2
i=3
= [I f ]kl + w(θ),
where supθ∈[a,A] |w(θ)| →n→∞ 0, by dominated convergence. Finally, we have, for λ1 , ..., λp ∈
R,
!2
!2
Z 1
p
p
p
p
X
X
X
X
√
1
1
√
λk λl [If ]kl =
λk fk (0) +
λk fk0 (t) + θ
λk fk (t) dt.
2 0
θ k=1
k,l=1
k=1
k=1
Pp
Pp
Pp
0
Hence,
if
λ
λ
[I
]
=
0,
then
λ
f
(0)
=
0
and
for
all
t
∈
[0,
1],
k
l
f
kl
k
k
k,l=1
k=1
k=1 λk fk (t) =
Pp
−θ k=1 λk fk (t) so that, by linear independence, λ1 = ... = λp = 0. Hence I f is invertible.
37
Lemma 8.15. We have
max sup |yi − Ŷθ,−i (si )| 6 K sup |Y (t)|.
i=1,...,n a6θ6A
t∈[0,1]
Proof of Lemma 8.15. For i 6∈ {1, ..., n}, using (41),
1
(αi Wi − αi+1 e−θ∆i+1 Wi+1 )
Ai
= δi,n Wi + δi,n Wi+1
= δi,n sup |Y (t)|.
yi − Ŷθ,−i (si ) =
t∈[0,1]
Also
y1 − Ŷθ,−1 (s1 ) = y1 − e−θ∆2 y2 = δ1,n sup |Y (t)|
t∈[0,1]
and similarly yn − Ŷθ,−n (sn ) = δn,n supt∈[0,1] |Y (t)|.
We now prove Lemma 8.11.
Proof of Lemma 8.11. • Term r1 (θ):
n
n
X
X
−1
−1
log 1 −
r1 (θ) =
log (Rθ )ii − ¯−i − log((Rθ )ii ) =
=
=
i=1
n
X
i=1
n
X
i=1
log 1 −
i=1
!
−1
0 −1
0 −1
0
−1
(ei,n Rθ F )(F Rθ F ) (F Rθ ei,n )
(R−1
θ )ii
δi,n
log 1 −
(R−1
θ )ii
¯−i
(R−1
θ )ii
,
from Lemmas 8.13 and 8.14.
1
1
Now, from (1), for i = 2, . . . , n, we have supθ (R−1
= δi,n ∆i and supθ (R−1
= δ1,n ∆2 .
)ii
)11
θ
θ
Pn
Hence r1 (θ) = log (1 − δ1,n ∆2 ) + i=2 log (1 − δi,n ∆i ) so that supθ∈[a,A] |r1 (θ)| is bounded
as n → ∞.
• Term r2 (θ):
(k)
For k ∈ N, let f −i be obtained by removing component i of f (k) . We observe that fk (si ) −
(k)
r 0θ,−i R−1
θ,−i f −i can be interpreted as a leave-one-out prediction error for a n-dimensional
observation vector equal to f (k) . Hence from (3),
n
X
2
r2 (θ) =
(R−1
θ )ii −i
i=1
p
=
" n
X X
k,l=1
i=1
1
−1 (β 0 − β̂ −i )k (β 0 − β̂ −i )l
(Rθ )ii
38
(k)
e0i,n R−1
θ f
#
−1 (l)
0
ei,n Rθ f
.
For i ∈ {1, ..., n}, one can show that Lemmas 8.12 and 8.14 remain true with F , Rθ , y
replaced by F −i , Rθ,−i , y −i . In addition, these (modified) lemmas can be shown to be
uniform over i = 1, ..., n. As a consequence, we have, for i = 1, . . . , n,
−1
−1 0
||β̂ −i − β 0 || = ||(F 0−i R−1
θ,−i F −i ) F −i Rθ,−i y −i || 6 K sup |Y (t)|.
(65)
t∈[0,1]
Also, since from (1) maxi=1,...,n
n
X
1
(R−1
θ )ii
i=1
1
(R−1
θ )ii
6 K, we have
(k)
(l)
e0i,n R−1
6K+
e0i,n R−1
θ f
θ f
n−1
X
δi,n (∆i + ∆i+1 )2 ,
i=2
from Lemma 8.13. Hence, we have
r2 (θ) 6 K sup Y (t)2
K+
t∈[0,1]
n−1
X
!
(∆i + ∆i+1 )2 .
i=2
Hence finally, sup |r2 (θ)| is a.s. bounded as n → ∞.
θ∈[a,A]
• Term r3 (θ):
r3 (θ) =
n
X
(R−1
θ )ii −i (yi − Ŷθ,−i (si ))
i=1
p
n
XX
(k)
−1
0
(R−1
=
θ )ii (β 0 − β̂ −i )k [fk (si ) − r θ,−i Rθ,−i f −i ](yi − Ŷθ,−i (si )).
k=1 i=1
(k)
We make the same observation on fk (si ) − r 0θ,−i R−1
θ,−i f −i as for r2 (θ). Hence we have
r3 (θ) =
p
n
X
X
(k)
(β 0 − β̂ −i )k (e0i,n R−1
)(yi − Ŷθ,−i (si )).
θ f
k=1 i=1
Hence, from (65) and Lemmas 8.13 and 8.15, we have
|r3 (θ)| 6
p
n−1
X
X
k=1
!
δi,n (∆i + ∆i+1 ) sup Y (t)2
!
+ δ1,n sup Y (t)2 .
t∈[0,1]
i=2
t∈[0,1]
Hence, r3 (θ) is almost surely bounded as n → ∞.
• Term r4 (θ):
From Lemmas 8.13 and 8.14, we have for i ∈ {2, ..., n − 1}, ¯−i 6 K(∆i + ∆i+1 )2 and for
i ∈ {1, n}, ¯−i 6 K. Hence
|r4 (θ)| 6 K(y1 − Ŷθ,−1 (s1 ) + −1 )2 + K(yn − Ŷθ,−n (sn ) + −n )2
n−1
X
+
(∆i + ∆i+1 )2 (yi − Ŷθ,−i (si ) + −i )2
i=2
n−1
X
6 K sup Y (t) + K
(∆i + ∆i+1 )2 sup Y (t)2 ,
2
t∈[0,1]
i=2
39
t∈[0,1]
using also Lemma 8.15, and remarking that the previous treatment of r2 (θ) implies that
2−i 6 K supt∈[0,1] Y (t)2 . Hence r4 (θ) is a.s. bounded as n → ∞.
Acknowledgements
We are grateful to Jean-Marc Azais, for constructive discussions on the topic of this
manuscript. This work was partly funded by the ANR project PEPITO. We thank the associate editor and two referees, whose suggestions lead to a broader scope and an improved
presentation of the manuscript.
References
[1] P. Abrahamsen. A review of Gaussian random fields and correlation functions. Technical report, Norwegian computing center, 1997.
[2] M. Abt and W. J. Welch. Fisher information and maximum-likelihood estimation
of covariance parameters in Gaussian stochastic processes. The Canadian Journal of
Statistics, 26:127–137, 1998.
[3] E. Anderes. On the consistent separation of scale and variance for Gaussian random
fields. The Annals of Statistics, 38:870–893, 2010.
[4] A. B. Antognini and M. Zagoraiou. Exact optimal designs for computer experiments
via Kriging metamodelling. Journal of Statistical Planning and Inference, 140:2607–
2617, 2010.
[5] F. Bachoc. Cross validation and maximum likelihood estimations of hyper-parameters
of Gaussian processes with model mispecification. Computational Statistics and Data
Analysis, 66:55–69, 2013.
[6] F. Bachoc. Asymptotic analysis of the role of spatial sampling for covariance parameter
estimation of Gaussian processes. Journal of Multivariate Analysis, 125:1–35, 2014.
[7] F. Bachoc. Asymptotic analysis of covariance parameter estimation for Gaussian
processes in the misspecified case. To appear in Bernoulli, 2017.
[8] F. Bachoc, K. Ammar, and J.M. Martinez. Improvement of code behavior in a design
of experiments by metamodeling. Nuclear science and engineering, 183(3):387–406,
1016.
[9] F. Bachoc, G. Bois, J. Garnier, and J.M Martinez. Calibration and improved prediction of computer models by universal Kriging. Nuclear Science and Engineering,
176(1):81–97, 2014.
40
[10] C-H. Chang, H-C. Huang, and C-K. Ing. Mixed domain asymptotics for a stochastic
process model with time trend and measurement error. Bernoulli, 23(1):159–190,
2017.
[11] H-S. Chen, D. G. Simpson, and Z. Ying. Infill asymptotics for a stochastic process
model with measurement error. Statistica Sinica, 10:141–156, 2000.
[12] N. Cressie and S.N. Lahiri. The asymptotic distribution of REML estimators. Journal
of Multivariate Analysis, 45:217–233, 1993.
[13] N. Cressie and S.N. Lahiri. Asymptotics for REML estimation of spatial covariance
parameters. Journal of Statistical Planning and Inference, 50:327–341, 1996.
[14] J. Du, H. Zhang, and V. S. Mandrekar. Fixed-domain asymptotic properties of tapered
maximum likelihood estimators. The Annals of Statistics, 37:3330–3361, 2009.
[15] O. Dubrule. Cross validation of Kriging in a unique neighborhood. Mathematical
Geology, 15:687–699, 1983.
[16] R. Furrer, F. Bachoc, and J. Du. Asymptotic properties of multivariate tapering for
estimation and prediction. Journal of Multivariate Analysis, 149:177–191, 2016.
[17] T-C. Hu and R.L. Taylor. On the strong law for arrays and for the bootstrap
mean and variance. International Journal of Mathematics and Mathematical Sciences,
20(2):375–382, 1997.
[18] I.A. Ibragimov and Y.A. Rozanov. Gaussian Random Processes. Springer-Verlag, New
York, 1978.
[19] J. Istas and G. Lang. Quadratic variations and estimation of the local Hölder index
of a Gaussian process. Annales de l’Institut Henri Poincaré, 33:407–436, 1997.
[20] D.R. Jones, M. Schonlau, and W.J. Welch. Efficient global optimization of expensive
black box functions. Journal of Global Optimization, 13:455–492, 1998.
[21] C. G. Kaufman and B. A. Shaby. The role of the range parameter for estimation and
prediction in geostatistics. Biometrika, 100:473–484, 2013.
[22] W-L. Loh. Fixed-domain asymptotics for a subclass of Matérn-type Gaussian random
fields. The Annals of Statistics, 33:2344–2394, 2005.
[23] W-L. Loh and T-K. Lam. Estimating structured correlation matrices in smooth Gaussian random field models. The Annals of Statistics, 28:880–904, 2000.
[24] K. V. Mardia and R. J. Marshall. Maximum likelihood estimation of models for
residual covariance in spatial regression. Biometrika, 71:135–146, 1984.
41
[25] G. Matheron. La Théorie des Variables Régionalisées et ses Applications. Fasicule
5 in Les Cahiers du Centre de Morphologie Mathématique de Fontainebleau. Ecole
Nationale Supérieure des Mines de Paris, 1970.
[26] M. H. Neumann. A central limit theorem for triangular arrays of weakly dependent
random variables, with applications in statistics. ESAIM: Probability and Statistics,
17:120–134, 2013.
[27] R. Paulo, G. Garcia-Donato, and J. Palomo. Calibration of computer models with
multivariate output. Computational Statistics and Data Analysis, 56:3959–3974, 2012.
[28] C.E. Rasmussen and C.K.I. Williams. Gaussian Processes for Machine Learning. The
MIT Press, Cambridge, 2006.
[29] J. Sacks, W.J. Welch, T.J. Mitchell, and H.P. Wynn. Design and analysis of computer
experiments. Statistical Science, 4:409–423, 1989.
[30] T.J. Santner, B.J. Williams, and W.I. Notz. The Design and Analysis of Computer
Experiments. Springer, New York, 2003.
[31] B. A. Shaby and D. Ruppert. Tapered covariance: Bayesian estimation and asymptotics. Journal of Computational and Graphical Statistics, 21(2):433–452, 2012.
[32] M.L. Stein. Asymptotically efficient prediction of a random field with a misspecified
covariance function. The Annals of Statistics, 16:55–63, 1988.
[33] M.L. Stein. Bounds on the efficiency of linear predictions using an incorrect covariance
function. The Annals of Statistics, 18:1116–1138, 1990.
[34] M.L. Stein. Uniform asymptotic optimality of linear predictions of a random field
using an incorrect second-order structure. The Annals of Statistics, 18:850–872, 1990.
[35] M.L. Stein. Interpolation of Spatial Data: Some Theory for Kriging. Springer, New
York, 1999.
[36] S. Sundararajan and S.S. Keerthi. Predictive approaches for choosing hyperparameters
in Gaussian processes. Neural Computation, 13:1103–1118, 2001.
[37] A. W. van der Vaart. Maximum likelihood estimation under a spatial sampling scheme.
The Annals of Statistics, 24(5):2049–2057, 1996.
[38] A. W. van der Vaart. Asymptotic statistics, volume 3 of Cambridge Series in Statistical
and Probabilistic Mathematics. Cambridge University Press, Cambridge, 1998.
[39] D. Wang and W-L. Loh. On fixed-domain asymptotics and covariance tapering in
Gaussian random field models. Electronic Journal of Statistics, 5:238–269, 2011.
42
[40] Z. Ying. Asymptotic properties of a maximum likelihood estimator with data from a
Gaussian process. Journal of Multivariate Analysis, 36:280–296, 1991.
[41] Z. Ying. Maximum likelihood estimation of parameters under a spatial sampling
scheme. The Annals of Statistics, 21:1567–1590, 1993.
[42] H. Zhang. Inconsistent estimation and asymptotically equivalent interpolations in
model-based geostatistics. Journal of the American Statistical Association, 99:250–
261, 2004.
[43] H. Zhang and Y. Wang. Kriging and cross validation for massive spatial data. Environmetrics, 21:290–304, 2010.
[44] H. Zhang and D. L. Zimmerman. Towards reconciling two asymptotic frameworks in
spatial statistics. Biometrika, 92:921–936, 2005.
43
| 10math.ST
|
arXiv:1704.00026v3 [cs.NE] 13 Jun 2017
Upper Bounds on the Runtime of the Univariate Marginal
Distribution Algorithm on OneMax∗
Carsten Witt
DTU Compute
Technical University of Denmark
2800 Kgs. Lyngby
Denmark
June 15, 2017
Abstract
A runtime analysis of the Univariate Marginal Distribution Algorithm (UMDA)
is presented on the OneMax function for wide ranges of the parameters µ and λ.
If µ ≥ c log n for some constant c > 0 and λ = (1 + Θ(1))µ, a general bound
O(µn) on the expected runtime is obtained. This bound crucially assumes that all
marginal √
probabilities of the algorithm are confined to the interval [1/n, 1 − 1/n].
If µ ≥ c′ n log n for a constant c′ > 0 and λ = (1 + Θ(1))µ, the behavior
of the
√
algorithm changes and the bound on the expected runtime becomes O(µ n), which
typically even holds if the borders on the marginal probabilities are
√ omitted.
The results supplement the recently derived lower bound Ω(µ n + n log n) by
Krejca and Witt (FOGA
√ 2017) and turn out as tight for the two very different choices
µ = c log n and µ = c′ n log n. They also improve the previously best known upper
bound O(n log n log log n) by Dang and Lehre (GECCO 2015).
1
Introduction
Estimation-of-distribution algorithms (EDAs, [15]) are randomized search heuristics that
have emerged as a popular alternative to classical evolutionary algorithms like Genetic
Algorithms. In contrast to the classical approaches, EDAs do not store explicit populations of search points but develop a probabilistic model of the fitness function to be
optimized. Roughly, this model is built by sampling a number of search points from the
current model and updating it based on the structure of the best samples.
Although many different variants of EDAs (cf. [12]) and many different domains are
possible, theoretical analysis of EDAs in discrete search spaces often considers runtime
∗
An extended abstract of this report will appear in the proceedings of the 2017 Genetic and Evolutionary Computation Conference (GECCO 2017).
1
analysis over {0, 1}n . The simplest of these EDAs have no mechanism to learn correlations between bits. Instead, they store a Poisson binomial distribution, i. e., a probability
vector p of n independent probabilities, each component pi denoting the probability that
a sampled bit string will have a 1 at position i.
The first theoretical analysis in this setting was conducted by Droste [6], who analyzed the compact Genetic Algorithm (cGA), an EDA that only samples two solutions in
each iteration, on linear functions. Papers considering other EDAs, like, e. g., analysis
of an iteration-best Ant Colony Optimization (ACO) algorithm by Neumann et al. [21]
followed.
Recently, the interest in the theoretical runtime analysis of EDAs has increased [5,
10, 9, 26, 14]. Most of these works derive bounds for a specific EDA on the popular
OneMax function, which counts the number of 1s in a bit string and is considered to
be one of the easiest functions with a unique optimum [25, 28]. In this paper, we follow
up on recent work on the Univariate Marginal Distribution Algorithm (UMDA [20]) on
OneMax.
The UMDA is an EDA that samples λ solutions in each iteration, selects µ < λ best
solutions, and then sets the probability pi (hereinafter called frequency) to the relative
occurrence of 1s among these µ individuals. The algorithm has already been analyzed
some years ago for several artificially designed example functions [2, 1, 3, 4]. However,
none these papers considered the most fundamental benchmark function in theory, the
OneMax function. In fact, the runtime analysis of the UMDA on the simple OneMax
function has turned out to be rather challenging; the first such result, showing the upper
bound O(n log n log log n) on its expected runtime for certain settings of µ and λ, was
not published until 2015 [5]. Very recently, this result was supplemented by a general
√
lower bound of the kind Ω(µ n + n log n) [14], proving that the UMDA cannot be more
efficient than simple evolutionary algorithms on this function, at least if λ = (1+Θ(1))µ.
As upper and lower bounds were apart by a factor of Θ(log log n), it was an open problem
to determine the asymptotically best possible runtime of the UMDA on OneMax.
In this paper, we close this gap and show that the UMDA can optimize OneMax
in expected time O(n log n) for two very different, carefully chosen values of µ, always
assuming that λ = (1 + Θ(1))µ. In fact, we obtain two general upper bounds depending
√
on µ. If µ ≥ c n log n, where c is a sufficiently large constant, the first upper bound
√
is O(µ n). This bound exploits that all pi move more or less steadily to the largest
possible value and that there are no frequencies that ever drop below 1/4. Around µ =
√
Θ( n log n), there is a phase transition in the behavior of the algorithm. With smaller µ,
the stochastic movement of the frequencies is more chaotic and many frequencies will
hit the lowest possible value during the optimization. Still, the expected optimization
time is O(µn) for µ ≥ c′ log n and a sufficiently large constant c′ > 0 if all frequencies
are confined to the interval [1/n, 1 − 1/n], as typically done in EDAs. If frequencies are
allowed to drop to 0, the algorithm will typically have infinite optimization time below
√
the phase transition bound µ ∼ n log n, whereas it typically will be efficient above.
Interestingly, Dang and Lehre [5] used µ = Θ(ln n), i. e., a value below the phase
transition to obtain their O(n log n log log n) bound. This region turns out to be harder
2
to analyze than the region above the phase transition, at least with our techniques.
However, our proof also follows an approach being widely different from [5]. There the
so-called level based theorem, a very general upper bound technique, is applied to track
the stochastic behavior of the best-so-far OneMax-value. While this gives a rather
short and elegant proof of the upper bound O(n log n log log n), the generality of the
technique does not give much insight into how the probabilities pi of the individuals bits
develop over time. We think that it is crucial to understand the working principles of
the algorithm thoroughly and present a detailed analysis of the stochastic process at bit
level, as also done in many other runtime analyses of EDAs [10, 9, 26, 14].
This paper is structured as follows: in Section 2, we introduce the setting we are going
to analyze and summarize some tools from probability theory that are used throughout
the paper. In particular, a new negative drift theorem is presented that generalizes
previous formulations by making milder assumptions on steps in the direction of the
drift than on steps against the drift. In the section, we also give a detailed analysis of
the update rule of the UMDA, which results in a bias of the frequencies pi towards higher
values. These techniques are presented for the OneMax-case, but contain some general
insights that may be useful in analyses of different fitness functions. In Section 3, we
√
prove the upper bound for the case of µ above the phase transition point Θ( n log n).
The case of µ below this point is dealt with in Section 4. We finish with some conclusions.
The appendix gives a self-contained proof of the new drift theorem.
Independent, related work. Very recently, Lehre and Nguyen [16] independently
√
obtained the upper bound O(µn) for c log n ≤ µ = O( n) using a refined application of
the level-based method. Our approach also covers larger µ (but requires λ = µ(1+Θ(1)))
and is technically different.
2
Preliminaries
We consider the so-called Univariate Marginal Distribution Algorithm (UMDA [20]) in
Algorithm 1 that maximizes the pseudo-Boolean function f . Throughout this paper, we
have f := OneMax, where, for all x = (x1 , . . . , xn ) ∈ {0, 1}n ,
OneMax(x) =
n
X
xi .
i=1
Note that the unique maximum is the all-ones bit string. However, a more general
version can be defined by choosing an arbitrary optimum a ∈ {0, 1}n and defining, for all
x ∈ {0, 1}n , OneMaxa (x) = n − dH (x, a), where dH (x, a) denotes the Hamming distance
of the bit strings x and a. Note that OneMax1n is equivalent to the original definition
of OneMax. Our analyses hold true for any function OneMaxa , with a ∈ {0, 1}n , due
to symmetry of the UMDA’s update rule.
We call bit strings individuals and their respective OneMax-values fitness.
The UMDA does not store an explicit population but does so implicitly, as usual in
EDAs. For each of the n different bit positions, it stores a rational number pi , which
3
Algorithm 1: Univariate Marginal Distribution
Algorithm (UMDA); algorithm
UMDA∗ is obtained if the line indexed R is omitted.
t ← 0, pt,1 ← pt,2 ← · · · ← pt,n ← 12 ;
while termination criterion not met do
Pt ← ∅;
for j ∈ {1, . . . , λ} do
for i ∈ {1, . . . , n} do
(j)
xt,i ← 1 with prob. pt,i ,
(j)
xt,i ← 0 with prob. 1 − pt,i ;
(j)
Pt ← Pt ∪ {xt };
(1)
(µ)
Sort individuals in P descending by fitness (such that f (xt ) ≥ · · · ≥ f (xt )),
breaking ties uniformly at random;
for i ∈ {1, . . . , n} do
Pµ
x
(j)
pt+1,i ← j=1µ t,i ;
R Restrict pt+1,i to be within [ n1 , 1 − n1 ];
t ← t + 1;
we call frequency, determining how likely it is that a hypothetical individual would have
a 1 at this position. In other words, the UMDA stores a probability distribution over
{0, 1}n . The starting distribution is the uniform distribution.
In each so-called generation t, the UMDA samples λ individuals such that each
individual has a 1 at position i, where i ∈ {1, . . . , n} with probability pt,i , independent of
all the other frequencies. Thus, individuals are sampled according to a Poisson binomial
distribution with probability vector (pt,i )i∈{1,...,n} .
After sampling λ individuals, µ of them with highest fitness are chosen, breaking
ties uniformly at random (so-called selection). Then, for each position, the respective
frequency is set to the relative occurrence of 1s in this position. That is, if the chosen µ
best individuals have x 1s among them, the frequency pi will be updated to x/µ for the
next iteration. Note that such an update allows large jumps like, e. g., from (µ − 1)/µ
to 1/µ.
If a frequency is either 0 or 1, it cannot change anymore since then all values at this
position will be either 0 or 1. To prevent the UMDA from getting stuck in this way,
we narrow the interval of possible frequencies down to [1/n, 1 − 1/n] and call 1/n and
1 − 1/n the borders for the frequencies. Hence, there is always a chance of sampling
0s and 1s for each position. This is a common approach used by other EDAs as well,
such as the cGA or ACO algorithms (cf. the references given in the introduction). We
also consider a variant of the UMDA called UMDA∗ where the borders are not used.
That algorithm will typically not have finite expected runtime; however, it might still
4
be efficient with high probability if it is sufficiently unlikely that frequencies get stuck
at bad values.
Overall, we are interested in upper bounds on the UMDA’s expected number of
function evaluations on OneMax until the optimum is sampled; this number is typically
called runtime. Note that this equals the expected number of generations until the
optimum is sampled times λ.
In all of our analyses, we assume that λ = (1 + β)µ for some constant β > 0 and
use µ and λ interchangeably in asymptotic notation. Of course, we could also choose
λ = ω(µ) but then each generation would be more expensive. Choosing λ = Θ(µ) lets
us basically focus on the minimal number of function evaluations per generation, as µ
of them are at least needed to make an update.
2.1
Useful Tools from Probability Theory
In our analyses, we often approximate sums of random variables, which are independent
but not necessarily identically distributed, by the normal distribution through the central limit theorem (CLT). The following lemma describes sufficient conditions for such
approximations to be sufficiently exact, along with a bound on the approximation error.
Lemma 1 (CLT with Lyapunov condition, Berry-Esseen inequality [8, p. 544]). Let
X1 , . . . , Xm be a sequence of independent random variables, each with finite expected
value µi and variance σi2 . Define
s2m
:=
m
X
σi2
and
Cm
m
1 X
:=
(Xi − µi ) .
sm
i=1
i=1
If there exists a δ > 0 such that
lim
m→∞
m
1 X
s2+δ
m
i=1
E |Xi − µi |2+δ = 0
(assuming all the 2 + δth moments to be defined), then Cm converges in distribution to
a standard normally distributed random variable.
Moreover, the approximation error is bounded as follows: for all x ∈ R,
Pm
3
i=1 E |Xi − µi |
|Pr(Cm ≤ x) − Φ(x)|≤ C ·
s3m
where C is an absolute constant and Φ(x) denotes the cumulative distribution function
of the standard normal distribution.
We will use the following estimates for Φ(x). More precise formulas exist, but they
do not yield any benefit in our analysis.
5
Lemma 2 ([7, p. 175], [27]). For any x < 0,
x2
x2
1
1
1
1 1
−
− 3 √ e− 2 ≤ Φ(x) ≤ − √ e− 2 .
x x
x 2π
2π
If −1 ≤ x < 0,
1
1
Φ(x) ≥ + √
2
2π
x3
x−
3
.
An approximation of the binomial distribution by the normal distribution will often
be used to bound the probability of a tail from below. To bound it from above, many
classical techniques such as Chernoff-Hoeffding bounds exist. The following version,
which includes knowledge of the variance, is particularly handy to use.
Lemma 3 ([18]). If X1 , . . . , Xn are independent, and Xi − E(Xi ) ≤ b for i ∈ {1, . . . , n},
then for X := X1 + · · · + Xn and any d ≥ 0 it holds
−
Pr(X − E(X) ≥ d) ≤ e
d2
2σ 2 (1+δ/3)
,
where σ 2 := Var(X) and δ := bd/σ 2 .
The following lemma describes a result regarding the Poisson binomial distribution
which we find very intuitive. However, as we did not find a sufficiency related result in
the literature, we give a self-contained proof here.
Lemma 4. Let X1 , .P
. . , Xn be independent P
Poisson trials. Denote pi = P
Pr(Xi = 1) for
n
n
2
i ∈ {1, . . . , n}, X := i=1 Xi , µ := E(X) = i=1 pi and σ := Var(X) = ni=1 pi (1−pi ).
Given two constants ℓ, u ∈ (0, 1] such that ℓ + u < 1, let kℓ := min{i | Pr(X ≤ i) ≥ ℓ}
and ku := max{i | Pr(X ≥ i) ≥ u}. Then it holds Pr(X = k) = Ω(min{1, 1/σ}) for all
k ∈ {kℓ , . . . , ku },
Proof. We first note that kℓ ≤ ku . This holds since by assumption Pr(X < kℓ ) < ℓ and
Pr(X > ku ) < u, hence Pr(kℓ ≤ X ≤ ku ) ≥ 1 − ℓ − u > 0, using ℓ + u < 1. If kℓ > ku
happened, we would obtain a contradiction.
We first handle the case σ = o(1) separately. Since X takes only integral values,
this implies Pr(X < [E(X)]) = o(1) and Pr(X > [E(X)]) = o(1), where [r] denotes the
integer closest to r. Since ℓ and u are positive constants, it immediately follows that
kℓ = ku = [E(X)] and Pr(X = kℓ ) = Ω(1).
In the following, we assume σ = Ω(1) and distinguish between the cases kℓ ≤ ⌊E(X)⌋
and kℓ > ⌊E(X)⌋. Assume first kℓ ≤ ⌊E(X)⌋. We start by proving Pr(X = kℓ ) = Ω(1).
To this end, we recall the unimodality of the Poisson-binomial distribution function, more
precisely Pr(X = i) ≤ Pr(X = i + 1) for i ≤ ⌊E(X)⌋ − 1 and Pr(X = i) ≥ Pr(X = i + 1)
for i ≥ ⌈E(X)⌉ [24]. Hence, denoting α := Pr(X = kℓ ), we have Pr(X = i) ≤ α for all
i ≤ kℓ . It follows Pr(X ≤ kℓ − ℓ/(2α)) ≥ ℓ/2 = Ω(1) since Pr(X ≤ kℓ ) ≥ ℓ by definition.
We remark (but do not use) that this also implies a lower bound on kℓ .
p
Var(X) =
If α = o(1/σ) held, the fact Pr(X ≤ kℓ − ℓ/(2α))
=
Ω(1)
would
imply
p
Ω(1/α) = ω(σ), contradicting our assumption Var(X) = σ. Hence, Pr(X = kℓ ) =
6
Ω(1/σ). Again using the monotonicity of the Poisson-binomial distribution, we have
Pr(X = i) = Ω(1/σ) for all i ∈ {kℓ , . . . , ⌊E(X)⌋}. To show Pr(X = i) = Ω(1/σ) for
all i ∈ {⌈E(X)⌉, . . . , ku }, we apply the argumentation in a symmetrical way, using that
Pr(X ≥ ku ) = Ω(1) and that Pr(X = i) ≥ Pr(X = i + 1) for i ≥ ⌈E(X)⌉.
We are left with the case kℓ > ⌊E(X)⌋. Recall that ku ≥ kℓ . Moreover, Pr(X ≥
ku ) = Ω(1) by assumption. Similarly to the previous paragraph, if Pr(X = ku ) = α
for some α = o(1/σ) then we would obtain a contradiction to the given σ. So, Pr(X =
ku ) = Ω(1/σ). Now, again using the monotonicity, Pr(X = i) ≥ Pr(X = ku ) for all
i ∈ {kℓ , . . . , ku − 1}, hence Pr(X = i) = Ω(1/σ) for all i ∈ {kℓ , . . . , ku }.
As mentioned, we will study how the frequencies associated with single bits evolve
over time. To analyze the underlying stochastic process, the following theorem will be
used. It generalizes the so-called simplified drift theorem with scaling from [22]. The
crucial relaxation is that the original version demanded an exponential decay w. r. t.
jumps in both directions, more precisely the second condition below was on Pr(|Xt+1 −
Xt | ≥ jr). We now only have sharp demands on jumps in the undesired direction while
there is a milder assumption (included in the first item) on jumps in the desired direction.
The proof of the theorem is given in the appendix.
Theorem 1 (Generalizing [22]). Let Xt , t ≥ 0, be real-valued random variables describing a stochastic process over some state space, adapted to a filtration Ft . Suppose
there exist an interval [a, b] ⊆ R and, possibly depending on ℓ := b − a, a drift bound
ǫ := ǫ(ℓ) > 0, a typical forward jump factor κ := κ(ǫ) ≥ 1 as well as a scaling factor
r := r(ℓ) > 0 such that for all t ≥ 0 the following three conditions hold:
1. E((Xt+1 − Xt ) · 1{Xt+1 − Xt ≤ κǫ}) | Ft ; a < Xt < b) ≥ ǫ,
2. Pr(Xt+1 − Xt ≤ −jr | Ft ; a < Xt ) ≤ e−j for all j ∈ N,
3. λℓ ≥ 2 ln(4/(λǫ)), where λ := min{1/(2r), ǫ/(17r 2 ), 1/(κ2 ǫ)}.
Then for T ∗ := min{t ≥ 0 : Xt ≤ a | F0 ; X0 ≥ b} it holds that Pr(T ∗ ≤ eλℓ/4 ) =
O(e−λℓ/4 ).
To derive upper bounds on hitting times for optimal state, drift analysis is used, in
particular in scenarios where the drift towards the optimum is not homogeneous with
respect to place. Such a drift is called variable in the literature. A clean form of a variable
drift theorem, generalizing previous formulations from [13] and [19], was presented in
[23]. The following formulation has been proposed in [17].
Theorem 2 (Variable Drift, Upper Bound). Let (Xt )t∈N0 , be a stochastic process,
adapted to a filtration Ft , over some state space S ⊆ {0} ∪ [xmin , xmax ], where xmin > 0,
Let h(x) : [xmin , xmax ] → R+ be a monotone increasing function such that 1/h(x) is integrable on [xmin , xmax ] and E(Xt − Xt+1 | Ft ) ≥ h(Xt ) if Xt ≥ xmin . Then it holds for
the first hitting time T := min{t | Xt = 0} that
Z X0
1
xmin
E(T | F0 ) ≤
+
dx.
h(xmin )
xmin h(x)
7
Finally, we need the following lemma in our analysis of the impact of the so-called
2nd-class individuals in Section 2.2. Its statement is very specific and tailored to our
applications. Roughly, the intuition is to show that E(min{C, X}) is not much less than
min{C, E(X)} for X ∼ Bin(D, p) and D ≥ C.
Lemma 5. Let X be binomially distributed with parameters D and p. Let C ∈ {1, . . . , D}.
Then
E(min{C, X}) ≥ min{Cp + γp(1 − p)(D − C), C(1/2 + p/2)}
for some constant γ > 0.
Proof. First we note that Var(X) = Dp(1 − p). We denote f (X) = min{C, X} and note
that our aim is to compute E(f (X)). Obviously, E(f (X)) ≤ C.
We distinguish two cases with respect to p. If p ≤ (C/2+D/2)/D, then E := E(X) =
Dp ≤ C/2 + D/2, Var(X) ≤ D(1 − E/D) = D − E, so by Chebyshev’s inequality
√
√
4
D−E D−E
.
≤
Pr(X ≥ E + (D − E)/2) = Pr X ≥ E +
2
D−E
Note that D − E ≥ D − (C/2 + D/2) = D/2 − C/2. We get
Pr(X ≥ E + (D − E)/2) ≤
8
16
=
.
D/2 − C/2
D−C
To relate E(X) to E(f (X)), we note that all outcomes X ≤ E + (D − E)/2 are weighted
by a factor at most (E + (D − E)/2)/C ≤ (C/4 + 3D/4)/C larger in the first than in
the second expectation. The remaining outcomes are weighted by a factor at most D/C
higher. Noting that E(X) = Cp + (D − C)p, we obtain
16
C
C 16
· 1−
E(f (X)) ≥ E(X)
+ E(X)
C/4 + 3D/4
D−C
DD−C
16
C
· 1−
≥ Cp + (D − C)p
C + 3(D − C)/4
D−C
≥ Cp + γp(D − C)
for some sufficiently small constant γ > 0 if D−C ≥ κ for some sufficiently large constant
κ. Hence, we have E(f (X)) ≥ Cp + γp(D − C) in this case.
If p ≥ (C/2 + D/2)/D then, using E := E(X) = pD, we obtain Var(X) ≤ D(1 −
E/D) = D − E. By Chebyshev’s inequality, we have
Pr(X ≤ C) ≤ Pr(X ≤ E − (E − C))
!
p
Var(X)(E − C)
D−E
√
≤
≤ Pr X ≤ E(X) −
.
(E − C)2
D−E
8
Under the condition X ≤ C, the expected value is at least the same as if we only had C
trials. Hence,
1
D−E
Cp
D−E
+
C
1
−
+
C
1
−
≥
E(f (X)) ≥ Cp
(E − C)2
(E − C)2
2
2
p 1
=C
+
,
2 2
using E − C ≥ D − E and bounding (D − E)/(E − C)2 ≤ 1/2 if E − C is not too small.
We have E − C ≥ (D − C)/2 ≥ κ/2, so it suffices to choose the constant κ sufficiently
large to obtain E(f (X)) ≥ C(1/2 + p/2) in this case.
If D − C is not sufficiently large for the above estimations to hold, we argue more
directly. Let D − C ≤ κ for some constant κ > 0. We study the first C trials, denoted
by Y ∼ Bin(C, p), and apply the law of total probability w. r. t. the events Y = C and
Y < C. If the first C trials yields less than C successes then at least one more success
of probability p among the remaining D − C trials will be taken into account before the
minimum takes effect. Hence,
E(f (X)) = CpC + (1 − pC )(E(Y | Y < C) + p)
E(Y · 1{Y < C})
C
C
+p
= Cp + (1 − p )
Pr(Y < C)
= Cp + p(1 − pC ) ≥ Cp + p(1 − p)
since p ≤ 1 and C ≥ 1. Hence, E(f (X)) ≥ Cp + p(1 − p)(D − C)/(2κ) ≥ Cp + γp(1 −
p)(D −C) for some sufficiently small constant γ > 0, so the statement holds also if D −C
is bounded from above by a constant.
2.2
On the Stochastic Behavior of Frequencies
To bound the expected runtime of UMDA and UMDA∗ , it is crucial to understand
how the n frequencies associated with the bits evolve over time. The symmetry of the
fitness function OneMax implies that each frequency evolves in the same way, but not
necessarily independently of the others. Intuitively, many frequencies should be close
to their upper border for making it sufficiently likely to sample the optimum, i. e., the
all-ones string.
To understand the stochastic process on a frequency, it is useful to consider the
UMDA without selection for a moment. More precisely, assume that each of the λ offspring has the same probability of being selected as one of the µ individuals determining
the frequency update. Then the frequency describes a random walk that is a martingale,
i. e., in expectation it does not change. With OneMax, individuals with higher value
are more likely to be among the µ updating individuals. However, since only the accumulated number of one-bits matters for selection, a single frequency may still decrease
even if the step leads to an increase of the best-so-far seen OneMax value. We will spell
out the bias due to selection in the remainder of this section.
9
We consider an arbitrary but fixed bit position j and denote by pt := pt,j its frequency at time t. Moreover, let Xt , where 0 ≤ Xt ≤ µ, be the number of ones
1−1/n
among the µ offspring selected to compute pj . Then pt = cap1/n (Xt /µ), where
caphℓ (a) := max{min{a, h}, l} caps frequencies at their borders.
Consider the fitness of all individuals sampled during one generation of the UMDA
w. r. t. n − 1 bits, i. e., all bits but bit j. Assume that the individuals are sorted in
levels decreasingly by their fitness; each individual having a unique index. Level n − 1
is called the topmost, and level 0 the lowermost. Let w+ be the level of the individual
with rank µ, and let w− be the level of the individual with rank µ + 1. Since bit j has
not been considered so far, its OneMax-value can potentially increase each individual’s
level by 1. Now assume that w+ = w− + 1. Then, individuals from level w− can end up
with the same fitness as individuals from level w+ , once bit j has been sampled. Thus,
individuals from level w+ were still prone to selection.
Among the µ individuals chosen during selection, we distinguish between two different
types: 1st-class and 2nd-class individuals. 1st-class individuals are those which have
some many 1s at the n − 1 other bits such that they had to be chosen during selection
no matter which value bit j has. The remaining of the µ individuals are the 2nd-class
individuals; they had to compete with other individuals for selection. Therefore, their
bit value j is biased toward 1 compared to 1st-class individuals. Note that 2nd-class
individuals can only exist if w+ ≤ w− + 1, since in this case, individuals from level w−
can still be as good as individuals from level w+ after sampling bit j.
∗
Given Xt , let Ct+1
denote the number of 2nd-class individuals in generation t + 1.
Note that the number of 1s of 1st-class individuals during generation t + 1 follows a
∗
1st-class
binomial distribution with success probability Xt /µ. Since we have µ − Ct+1
∗
individuals, the distribution of the number of 1s of these follows Bin(µ − Ct+1 , Xt /µ).
We proceed by analyzing the number of 2nd-class individuals and how they bias
the number of 1-bits, leading to Lemma 6 below. The underlying idea is that both the
number of 2nd-class individuals is sufficiently large and that at the same time, these
2nd-class individuals were selected from an even larger set to allow many one-bits to be
gained at the considered position j. This requires a careful analysis of the level where
the rank-µ individual ends up.
For i ∈ {0, . . . , n − 1}, let Ci denote the cardinality of level i, i. e., the number
of
individuals
in level i during an arbitrary generation of the UMDA, and let C≥i =
Pn−1
C
.
Let
M denote the index of the first level from the top such that the number
a
a=i
of sampled individuals is greater than µ when including the following level, i.e.,
M := max{i | C≥i−1 > µ}.
Note that M can never be 0, and only if M = n − 1, CM can be greater than µ. Due to
the definition of M , if M 6= n − 1, level M − 1 contains the individual of rank µ + 1, so
level M − 1 contains the cut where the best µ out of λ offspring are selected. Individuals
in levels at least M + 1 are definitely 1st-class individuals since they still will have rank
at least µ even if the bit j sampled last turns out as 0. 2nd-class individuals, if any,
have to come from levels M , M − 1 and M − 2. Individuals from level M may still
10
be selected (but may also not) for the µ updating individuals even if bit j turns out
as 0. To obtain a pessimistic bound the bias introduced by 2nd-class individuals, we
concentrate on level M − 1. Note that all individuals from level M − 1 sampling bit j
as 1 will certainly be selected unless the µ − C≥M remaining slots for the µ best are filled
up. We call the individuals from levels M − 1 2nd-class candidates and denote their
∗
∗
:= CM −1 . By definition, Dt+1
≥ µ − C≥M , and we want to show that
number by Dt+1
∗
Dt+1 is expected to be larger than µ − C≥M . That is, we expect to have more 2nd-class
candidates in level M − 1 than can actually be selected as 2nd-class individuals.
We now present the lemma that summarizes crucial properties of the distribution
of 1st-class and 2nd-class individuals and candidates along with the number of one-bits
sampled in them. We write X ≻ Y to denote that X is stochastically at least as large
as Y .
P
Lemma 6. Let σt2 := ni=1 pt,i (1 − pt,i ) be the sampling variance of the UMDA. Then
∗
∗
1. Ct+1
≻ Bin(Θ(µ), Θ(1/σt )), implying E(Ct+1
| σt ) = Ω(µ/σt ).
∗
∗
∗
∗
| σt ) = Ω(µ/σt ).
− Ct+1
≻ Bin(Θ(µ), Θ(1/σt )), implying E(Dt+1
− Ct+1
2. Dt+1
∗ , X /µ) + min{C ∗ , Bin(D ∗ , X /µ)}.
3. Xt+1 ∼ Bin(µ − Ct+1
t
t
t+1
t+1
4. E(Xt+1 | Xt , σt ) = Xt + Ω((µ/σt )(Xt /µ)(1 − Xt /µ)).
5. If pt ≤ 1 − c/n, where c > 0 is a sufficiently large constant, then E(pt+1 | pt , σt ) =
pt + Ω(pt (1 − pt )/σt ).
Proof. We first analyze the distribution of C ∗ := µ − C≥M , which as outlined above,
∗ . To this end, we carefully investigate and then reformulate
is a lower bound on Ct+1
the stochastic process generating the λ individuals (before selection), restricted to n − 1
bits. Each individual is sampled by a Poisson binomial distribution for a vector of
probabilities p′t = (p′t,1 , . . . , p′t,n−1 ) obtained from the frequency vector of the UMDA
by leaving the entry belonging to bit j out (i. e., p′t = (pt,1 , . . . , pt,j−1 , pt,j+1 , . . . , pt,n ).
Counting its number of 1s, each of the λ individuals then falls into some level i, where
0 ≤ i ≤ n − 1, with some probability qi depending on the vector p′t . Since the individuals
are created independently, the number of individuals in level i is binomially distributed
with parameters n − 1 and qi .
Next, we take an alternative view on the process putting individuals into levels,
using the principle of deferred decisions. We imagine that the process first samples all
individuals in level 0 (through λ trials, all of which either hit the level or not), then
(using the trials which did not hit level 0) all individuals in level 1, . . . , up to level n − 1.
The number of individuals C0 in level 0 is still binomially distributed with parameters
λ and q0 . However, after all individuals in level 0 have been sampled, the distribution
changes. We have λ − C0 trials left, each of which can hit one of the levels 1 to n − 1.
In particular, such a trial will hit level 1 with probability q1 /(1 − q0 ), by the definition
of conditional probability since level 0 is excluded. This holds independently for all of
the λ − C0 trials so that C1 follows a binomial distribution with parameters λ − C0 and
11
q1 /(1 − q0 ). Inductively, also all Ci for i > 1 are binomially distributed; e. g., Cn−1 is
distributed with parameters λ − Cn−2 − · · · − C0 and 1. Note that this model of the
sampling process can also be applied for any other permutation of the levels; we will
make use of this fact.
Recall that our aim is to analyze C ∗ . Formally, by applying the law of total probability, its distribution looks as follows for k ∈ {0, . . . , λ}:
Pr(C ∗ ≥ k) =
n−1
X
i=1
Pr(M = i) · Pr(µ − C≥i ≥ k | M = i) .
(1)
We will bound the last expression from below by omitting all terms where Pr(M = i)
is small. Let X be the number of 1s in a single individual sampled in the process of
creating the λ offspring (without conditioning on certain levels being hit). The expected
number of one-bits follows a Poisson binomial distribution with vector p′t . We define
1
L := min i | Pr(X ≤ i) ≥
2 + 2β
and
U := max i | Pr(X ≥ i) ≥
β
2 + 2β
.
1
1+β λ = µ
β
λ=
than 1+β
By Chernoff bounds, the number of individuals sampled above U is less than
with probability 1 − e−Ω(λ) , and the number sampled below L is also less
λ − µ with probability 1 − e−Ω(λ) . Then the µ-th ranked individual will be within
Z := [U, L] with constant probability, which means that Pr(M ∈ Z) = Ω(1). Without
further mention, such constant probabilities will be incorporated in the lower bounds on
success probabilities derived below, using the law of total probability. Coming back to
Equation (1), it is now sufficient to bound Pr(µ − C≥i ≥ k | M = i) by the distribution
function belonging a binomial distribution for all i ∈ Z.
We reformulate the underlying event appropriately. Here we note that
(µ − C≥i ≥ k) | (M = i)
is equivalent to (C≤i−1 ≥ λ − µ + k) | (M = i), where C≤i =
definition of M , this is also equivalent to
Pi
j=0 Cj ,
and, using the
(C≤i−1 ≥ λ − µ + k) | (C≤i−2 < λ − µ).
We now use the above-mentioned view on the stochastic process and assume that levels
0 to i − 2 have been sampled and a number of experiments in a binomial distribution is
carried out to determine the individuals from level i − 1. Hence, given some C≤i−2 =
a < λ − µ, our event is equivalent to that the event
E ∗ := Ci−1 ≥ (λ − µ − a) + k | (Ci−2 = a) ∩ (Ci−1 ≥ λ − µ − a)
12
happens.
Recall from our model that Ci−1 follows a binomial distribution with λ − a trials
and with a certain success probability s. The number of trials left after having sampled
levels 0, . . . , i − 2 is at least µ since a < λ − µ. Regarding s, note that it denotes
the probability to sample an individual with i − 1 1s, given that it cannot have less
than i − 1 1s. By omitting this condition, we clearly do not increase the probability.
Hence, we pessimistically assume that s = Pr(X = i − 1). We now apply Lemma 4,
using ℓ := 1/(2 + 2β) and u := β/(2 + 2β), in accordance with the above definition
of L and U . Hence, every level in Z is hit with probability Ω(1/σt ). Therefore, since
there are at least Ω(µ) trials left, the event Ci−1 ≥ (λ − µ − a) + k, conditioned on
(Ci−2 = a) ∩ (Ci−1 ≥ λ − µ − a) is at least as likely as obtaining at least k successes
in Ω(µ) trials with success probability Ω(1/σt ) each. Here we handle the condition by
assuming that Ci−1 already contains λ−µ−a samples and that afterwards the remaining
∗
≥ C ∗ ≻ Bin(Θ(µ), Θ(1/σt )). Using the
individuals in Ci−1 are sampled. Hence, Ct+1
properties of the binomial distribution we obtain E(C ∗ ) = Ω(µ/σt ), which proves the
first item of the lemma.
We now use a dual line of argumentation to analyze CM −1 and prove the second item
of the lemma. The aim is to show that C≥M −1 ≻ µ + Bin(Θ(µ), Θ(1/σt )). We take the
same view on the stochastic process as above but imagine now that the levels are sampled
in the order from n − 1 down to 0. Conditioning on that levels n − 1, . . . , M have been
sampled, level M − 1 is still hit by λ − µ = Ω(µ) samples of success probability Ω(1/σt )
each since by definition less than µ samples fall into levels n − 1, . . . , M . Using the same
∗
∗
arguments as above, C≥M −1 ≻ µ + Bin(Θ(µ), Θ(1/σt )). Observing that Dt+1
− Ct+1
=
∗
∗
∗
C≥M −1 − µ since trivially Ct+1 ≤ µ, we obtain Dt+1 − Ct+1 ≻ Bin(Θ(µ), Θ(1/σt )).
∗ ) = Ω(µ/σ ) 2nd-class individuals, and an additional number
Altogether, we have E(Ct+1
t
∗
∗
of E(Dt+1 − Ct+1 ) = Ω(µ/σt ) samples that are 2nd-class candidates but not individuals.
This proves the second item of the lemma.
To prove the third and fourth item, we essentially show that the expected overhang
in candidates allows a bias in the frequency, as detailed in the following. In each of the
∗
2nd-class candidates, bit j is sampled as 1 with probability Xt /µ. Only an expected
Dt+1
∗
2nd-class individuals, are selected
constant fraction of the candidates, namely the Ct+1
for the best µ offspring determining the next frequency. As observed above, the number
∗
of ones in the 1st-class individuals is binomially distributed with parameters µ − Ct+1
and Xt /µ. Hence, we obtain for the number of one-bits in the µ selected offspring that
∗
∗
∗
Xt+1 ∼ Bin(µ − Ct+1
, Xt /µ) + min{Ct+1
, Bin(Dt+1
, Xt /µ)},
which represents the third item of the lemma. To determine the expected value, we use
Lemma 5. Hence, the expected value of the minimum is at least the minimum of
∗
Ct+1
Xt
Xt
∗
∗
+ γ(Dt+1
− Ct+1
)
µ
µ
and
∗
Ct+1
1 Xt
+
2 2µ
13
.
If the minimum is taken on the first expression, we have an expected value E(Xt+1 |
∗ , D ∗ ) of at least
Xt , Ct+1
t+1
Xt
Xt
∗
∗
Xt + γ(Dt+1 − Ct+1 )
1−
,
µ
µ
∗
∗
are still random. If it is on the second term, we obtain an expected
and Dt+1
where Ct+1
value of
1 Xt
1 Xt
∗
∗
∗ Xt
+ Ct+1
+
−
= Xt + Ct+1
.
Xt − Ct+1
µ
2 2µ
2 2µ
Altogether,
∗
∗
∗
E(Xt+1 | Xt ) = Xt + Ω(min{E(Ct+1
), E(Dt+1
− Ct+1
)}(Xt /µ)(1 − Xt /µ)),
∗
∗ ) = Ω(µ/σ ) and E(D ∗
which, as both E(Ct+1
t
t+1 − Ct+1 ) = Ω(µ/σt ), implies
E(Xt+1 | Xt ) = Xt + Ω((µ/σt )(Xt /µ)(1 − Xt /µ)),
which proves the fourth item of the lemma.
To conclude on the expected value of pt+1 , we recall from Algorithm 1 that pt+1 :=
1−1/n
cap1/n Xt+1 /µ. Using our assumption pt ≤ 1 − c/n we get 1 − Xt /µ ≥ c/n. Hence,
as hitting the upper border changes the frequency by only at most 1/n and the lower
border can be ignored here, we also obtain E(pt+1 | pt ) ≥ pt + Ω((1/σt )pt (1 − pt )) if c is
large enough to balance the implicit constant in the Ω. This proves the fifth item.
We note that parts of the previous proof are strongly inspired by [14]; in particular,
the modeling of the stochastic process and the definition of M are almost literally taken
from that paper. However, as [14] is concerned with lower bounds on the runtime,
it bounds the number of 2nd-class individuals from above and needs a very different
argumentation in the core of its proofs.
3
Above the phase transition
We now prove our main result for the case of large λ. It implies an O(n log n) runtime
√
behavior if µ = c n log n.
√
Theorem 3. Let λ = (1 + β)µ for an arbitrary constant β > 0 and let µ ≥ c n log n
for some sufficiently large constant c > 0. Then with probability Ω(1), the optimization
√
time of both UMDA and UMDA∗ on OneMax is bounded from above by O(λ n). For
UMDA, also the expected optimization time is bounded in this way.
The proof of Theorem 3 follows a well-known approach that is similar to techniques partially independently proposed in several previous analyses of EDAs and of
ant colony optimizers [21, 9, 26]. Here we show that the approach also works for the
UMDA. Roughly, a drift analysis is performed with respect to the sum of frequencies. In
14
Lemma 6, we have already established a drift of frequencies towards higher values. Still,
there are random fluctuations (referred to as genetic drift in [26]) of frequencies that
may lead to undesired decreases towards 0. The proof of Theorem 3 uses that under the
condition on µ, typically all frequencies stay sufficiently far away from the lower border;
more precisely, no frequency drops below 1/4. Then the drift is especially beneficial.
The following lemma formally shows that, if µ is not too small, the positive drift
along with the fine-grained scale imply that the frequencies will generally move to higher
values and are unlikely to decrease by a large distance. Using the lemma, we will
′
′
obtain a failure probability of O(n−cc ) within ncc generations, which can subsume any
polynomial number of steps by choosing c large enough.
Lemma 7. Consider an arbitrary bit and let pt be its frequency at time t. Suppose that
√
µ ≥ c n log n for a sufficiently large constant c > 0. For T := min{t ≥ 0 | pt ≤ 1/4} it
′
′
then holds Pr(T ≤ ecc log n ) = O(e−cc log n ), where c′ is another positive constant.
Proof. The aim is to apply Theorem 1. We consider the frequency pt,i associated with
the considered bit and its distance Xt := µpt,i from the lower border. By initialization
of the UMDA, X0 = µ/2. Note that Xt is a process on {µ/n, 1, 2, . . . , µ − 1, µ(1 − 1/n)}.
In the notation of the drift theorem, we set [a, b] := [µ/4, µ/2], hence ℓ = µ/4.
Next we establish the three conditions. First, we observe that E(Xt+1 − Xt | Xt ) =
√
√
Ω(Xt (1 − Xt /µ)/ n) (Lemma 6 along with the trivial bound E(σt ) = O( n)) for Xt ∈
√
√
{1, . . . , µ}. The bound is Ω(µ/ n) for Xt ∈ [a, b], hence we will set ǫ′ := cµ/ n for
some constant c > 0. Hereinafter, we will omit the conditions Xt ; a < Xt < b from the
expected values. To establish the first condition of the drift theorem, we need to show
that the expected value is “typical”; formally, we will find a not too large κ such that
ǫ′
E (Xt+1 − Xt ) · 1{Xt+1 − Xt ≤ κǫ′ /2} ≥ .
2
Then the first condition is established with ǫ := ǫ′ /2.
We show the equivalent claim that E((Xt+1 − Xt ) · 1{Xt+1 − Xt > κǫ}) ≤ ǫ′ /2 when
κ is chosen as a sufficiently large constant. To see this, we recall from the third item of
Lemma 6 that Xt+1 is the sum of two random variables Z1 ∼ Bin(µ − C ∗ , Xt /µ) and
Z2 ∼ min{C ∗ , Bin(D ∗ , Xt /µ)} for some random variables C ∗ and D ∗ which, by the first
two items of the lemma, themselves dominate binomial distributions with parameters
√
Θ(µ) and Θ(1/ n). Since the lower bound E(Xt+1 − Xt ) ≥ ǫ′ derived from Lemma 6
assumes the case of a binomial distribution for both C ∗ and D ∗ as a worst case, we ignore
the stochastic dominance and are altogether confronted with binomial distributions only.
In fact, for the range Xt ∈ [µ/4, µ/2] considered here, we have in the notation of Lemma 5
for p = Xt /µ both p ∈ [Ω(1), 1 − Ω(1)] and 1 − p ∈ [Ω(1), 1 − Ω(1)]. Thus, applying
Lemma 5 in the same way as in the proof of Lemma 6 we get
E(Xt+1 | Xt ) ≥ Xt + γ ′ p(1 − p) min{D ∗ − C ∗ , C ∗ }
for a sufficiently small constant γ ′ > 0. Hence, using p(1 − p) = Θ(1), the drift is proportional to either the cardinality of D ∗ − C ∗ or of C ∗ , both of which (pessimisticaly) follow
15
√
a binomial distribution with parameters Θ(µ) and Θ(1/ n). Hence, w. l. o. g. we assume
hereinafter that Xt+1 = Z3 + Z4 , where Z3 ∼ Bin(µ, Xt /µ) and Z4 ∼ Bin(C ∗ , Xt /µ).
To show the claim, we exploit that Z3 + Z4 with constant probability is at least the
expected value E(Z3 + Z4 ). First, we note that E(Z3 ) = Xt . Moreover, it is easy to see
that Pr(Z3 ≥ E(Z3 )) = Ω(1). This holds since Xt ∈ [µ/4, µ/2] so the distribution of Z3
can be approximated by a normal distribution. Similarly, we obtain Pr(Z4 ≥ E(Z4 )) =
√
Ω(1) as also Z4 can be approximated by a Normal distribution due to µ = Ω( n log n)
√
and 1/σt = Ω(1/ n). If both Z3 ≥ E(Z3 ) and Z4 ≥ E(Z4 ), then one the one hand,
Xt+1 − Xt is clearly non-negative, and on the other hand, the outcome of Z4 is at least
C ∗ Xt /µ. Both events together happen with probability Ω(1) each, which just reduces
the original bound E(Xt+1 −Xt | Xt ) by at most a constant factor. Hence, now it suffices
to show that
E(C ∗ )
E C ∗ 1{C ∗ ≥ κE(C ∗ )} ≤
2
∗
(analogously w. r. t. D ) to establish our claim that
ǫ′
E (Xt+1 − Xt ) · 1{Xt+1 − Xt > κǫ} ≤ .
2
Note that E(C ∗ ) = Ω(log n) due to our assumption on µ. Hence, by Chernoff bounds,
Pr(C ∗ ≥ P
kE(C ∗ )) = e−Ω(k) for all k being bigger than some constant. Using the identity
E(X) = x≥0 Pr(X ≥ x) for random variables X with non-negative integer values, we
conclude E(C ∗ · 1{C ∗ ≥ κE(C ∗ )}) ≤ E(C ∗ )/2 as desired, given that κ is a sufficiently
large constant; analogously w. r. t. D ∗ . This proves the claim and establishes the first
condition of the drift theorem.
To show the second condition, recall from Section 2.2 that Xt+1 stochastically dominates Bin(µ, Xt /µ). Hence, to analyze steps where Xt+1 < Xt , we may pessimistically
assume the martingale case, where Xt+1 follows this binomial distribution, and obtain
Xt
µ
Xt
2
1−
≤ ,
ς := Var(Xt − Xt+1 | Xt ) = µ
µ
µ
4
√
so ς = Θ( µ). Using Lemma 3 with d = jς and b = 1, we get Pr(Xt+1 − Xt ≤ −jς) ≤
2
√
e−Ω(min{j ,j}). Hence, we can work with some r = c′′′ µ for some sufficiently large
constant c′′′ > 0 and satisfy the second condition on jumps that decrease the state.
The third condition is also easily verified but requires a distinction according to the
√
√
mutual relation of ℓ, ǫ, r and κ. We recall that ℓ = Θ(µ), ǫ = Θ(µ/ n), r = Θ( µ) and
√
√
√
κ = Θ(1). We note that ℓ/r 2 = Θ(1/ n), 1/r = Θ(1/ µ) and 1/ǫ = Θ( n/µ). Hence,
if µ = o(n), then by definition, λ equals ǫ/(17r 2 ) and λℓ = Ω(c log n) by our assumption
√
µ ≥ c n log n from the lemma. By choosing c large enough, we obtain
ǫℓ
≥ 2 ln(4r 2 /(17ǫ2 ))
17r 2
√
since the right-hand side is at most 2 ln(Θ( n)) and satisfy the third condition. If
√
√
√ √
µ = Ω(n), then λ = Ω(min{1/ µ, n/µ}) = Ω( n/µ) and already λℓ = Ω( n). Hence,
16
in any case, the drift theorem implies that the first hitting time of states less than a,
starting from above b is at least ec log n with probability at least 1 − e−c log n , if n is large
enough.
Proof of Theorem 3. We use a similar approach and partially also similar presentation
of the ideas as in [26]. Following [21, Theorem 3] we show that, starting with a setting
√
where all frequencies are at least 1/2 simultaneously, with probability Ω(1) after O( n)
generations either the global optimum has been found or at least one frequency has
dropped below 1/4. In the first case we speak of a success and in the latter of a failure.
√
The expected number of generations until either a success or a failure happens is O( n).
With respect to UMDA, we can use the success probability Ω(1) to bound the expected optimization time. We choose a constant γ > 3. According to Lemma 7, the probability of a failure in altogether nγ generations is at most n−γ , provided the constant c in
√
the condition µ ≥ c n log n is large enough. In case of a failure we wait until all frequencies simultaneously reach values at least 1/2 again and then repeat the arguments from
the preceding paragraph. It is easy to show via additive drift analysis for the UMDA (not
the UMDA∗ ) that the expected time for one frequency to reach the upper border is always
bounded by O(n3/2 ), regardless of the initial probabilities. This holds since by Lemma 6
there is always an additive drift of Ω(pt,i (1 − pt,i )/σt ) = Ω(1/(nσt )) = Ω(1/n3/2 ). By
standard arguments on independent phases, the expected time until all frequencies have
reached their upper border at least once is O(n3/2 log n). Once a frequency reaches the
upper border, we apply a straightforward modification of Lemma 7 to show that the
probability of a frequency decreasing below 1/2 in time nγ is at most n−γ (for large
enough c). The probability that there is a frequency for which this happens is at most
n−γ+1 by the union bound. If this does not happen, all frequencies attain value at least
1/2 simultaneously, and we apply our above arguments again. As the probability of a
failure is at most n−γ+1 , the expected number of restarts is O(n−γ+1 ) and the expected
time until all bits recover to values at least 1/2 only leads to an additional term of
n−γ+1 · O(n3/2 log n) ≤ o(1) (as n−γ ≤ n−3 ) in the expectation. We now only need to
√
show that after O( n) generations without failure the probability of having found the
all-ones string is Ω(1).
P
In the rest of this proof, we consider the potential function φt := n − 1 − ni=1 pt,i ,
which denotes the total distance of the frequencies from the upper border 1 − 1/n. For
simplicity, for the moment we assume that no frequency is greater than 1 − c/n, where
c is the constant from Lemma 6. Using Lemma 6 and the linearity of expectation, we
obtain for some constant γ > 0 the drift
E(φt − φt+1 | φt ) =
n
X
(pt+1,i − pt,i )
i=1
n
X
(pt,i + γpt,i (1 − pt,i )/σt − pt,i ) = γσt ,
=
i=1
17
P
since ni=1 pt,i (1 − pt,i ) = σt2 . Using our assumption pt,i ≥ 1/4, we obtain the lower
bound
v
v
u n
u n
uX
uX
t
E(φt − φt+1 | φt ) ≥ γ
pt,i (1 − pt,i ) ≥ γ t (1 − pt,i )/4,
i=1
i=1
√
which equals γ φt /2.
Now consider the frequencies that are greater than 1 − c/n and fix an arbitrary one
of them. The underlying bit is set to 0 in a selected offspring with probability at most
c/n since conditioning on being selected cannot make the probability bigger. Hence, we
expect no more than µc/n selected offspring sampling a 0, which implies an expected
decrease of the frequency by at most (µc/n)/µ = c/n. As there are at most n frequencies
greater than 1 − c/n, this amounts to a total drift of no less than −c causes by these
frequencies. In total, we bound the drift by
√
√
p
γ φt
γ φt
−c≥
=: h(φt ).
E(φt − φt+1 | φt ; φt ≥ 8/(cγ)) ≥
4
8
By the variable drift theorem (Theorem 2) with drift function h(φt ), maximum n and
minimum xmin = 8/(cγ), the expected number of generations until the φ-value is at most
8/(cγ) is at most
Z n
Z n
√
dx
xmin
dx
8/(cγ)
√
+
+
≤ p
= O( n)
h(xmin )
γ 8/(cγ)/8
8/(cγ) γ x/8
xmin h(x)
√
since both c and γ are constant. Hence, by Markov’s inequality, O( n) generations,
√
amounting to O(λ n) function evaluations, suffice with probability Ω(1) to reach φt ≤
8/(cγ) = O(1). It is easy to see that φt = O(1) implies an at least constant probability
of sampling the all-ones string (assuming that all pt,i are at least 1/4). Hence, the
√
optimum is sampled in O( n) generations with probability Ω(1), which, as outlined
above, proves the first statement of the lemma and also the statement on UMDA’s
expected runtime.
4
Below the phase transition
√
Theorem 3 crucially assumes that µ ≥ c n log n for a large constant c > 0. As described
above, the UMDA shows a phase transition between unstable and stable behavior at the
√
threshold Θ( n log n). While above the threshold, the frequencies typically stay well
focused on their drift towards the upper border and do not drop much below 1/2, the
√
opposite is the case if µ < c′ n log n for a sufficiently small constant c′ > 0. Krejca and
Witt [14] have shown for this regime that with high probability nΩ(1) frequencies will
walk to the lower border before the optimum is found, resulting in a coupon collector
effect and therefore the lower bound Ω(n log n) on the runtime. It also follows directly
from their results (although this is not made explicit) that UMDA∗ will in this regime
18
with high probability have infinite optimization time since nΩ(1) frequencies will get
√
stuck at 0. Hence, in the regime µ = Θ( n log n), the UMDA∗ turns from efficient with
at least constant probability to inefficient with overwhelming probability.
√
Interestingly, the value Θ( n log n) has also been derived in [26] as an important
parameter setting w. r. t. the update strengths called K and 1/ρ in the simple EDAs
cGA and 2-MMASib , respectively. Below the threshold value, lower bounds are obtained through a coupon collector argument, whereas above the threshold, the runtime
√
√
is O(K n) (and O((1/ρ) n) since frequencies evolve smoothly towards the upper border. The UMDA and UMDA∗ describe the same threshold behavior, even at the same
threshold points.
The EDAs considered in [26] use borders 1/n and 1 − 1/n for the frequencies in
the same way as the UMDA. The only upper bounds on the runtime are obtained for
√
update strengths greater than c n log n. Below the threshold, no conjectures on upper
bounds on the runtime are stated; however, it seems that the authors do not see any
benefit in smaller settings of the parameter since they recommend always to choose
values above the threshold. Surprisingly, this does not seem to be necessary if the
borders [1/n, 1 − 1/n] are used. With respect to the UMDA, we will show that even
for logarithmic µ it has polynomial expected runtime, thanks to the borders, while we
already know that UMDA∗ will fail. We also think that a similar effect can be shown
for the EDAs in [26].
√
We now give our theorem for the UMDA with small µ. If µ = Ω( n log n), it is
weaker than Theorem 3, again underlining the phase transition. The proof is more
involved since it carefully has to bound the number of times frequencies leave a border
state.
Theorem 4. Let λ = (1 + β)µ for an arbitrary constant β > 0 and µ ≥ c log n for a
sufficiently large constant c > 0 as well as µ = o(n). Then the expected optimization
time of UMDA on OneMax is O(λn). For UMDA∗ , it is infinite with high probability
√
if µ < c′ n log n for a sufficiently small constant c′ > 0.
Proof. The second statement can be derived from [14], as discussedPabove. We now
focus on the first claim, reusing the potential function φt = n − 1 − ni=1 pt,i from the
proof of Theorem 3. Let k denote the number of frequencies below 1 − c/n for the c
from Lemma 6, w. l. o. g., these are the frequencies associated with bits 1, . . . , k. The
last n − k bits are actually at 1 − 1/n since 1/µ = ω(n) by assumption. They are set
to 0 with probability at most 1/n in each of the selected offspring, amounting to a total
expected loss of at most 1. Similarly as in the proof of Theorem 3, we compute the drift
E(φt −φt+1
≥
k
X
1
(pt+1,i − pt,i ) − (n − k)
| φt ) ≥
n
i=1
k
X
i=1
(pt,i + γpt,i (1 − pt,i )/σt − pt,i ) − 1
19
=q
Pk
− pt,i )
−1
Pk
((n − k)/n)(1 − 1/n) + i=1 pt,i (1 − pt,i )
γ
i=1 pt,i (1
(2)
where the last equality
Pkjust used the definition of σt . We now distinguish two cases
depending on V ∗ :=
i=1 pt,i (1 − pt,i ), the total variance w. r. t. the bits not at the
∗
′
upper border. If V ≥ c for some sufficiently large constant c′ > 0, we obtain
√
and therefore
γV ∗
≥ 2,
1+V∗
γV ∗
γV ∗
√
E(φt − φt+1 | φt ) ≥ √
−
1
≥
≥1
1+V∗
2 1+V∗
from (2). If V ∗ < c′ , we will show by advanced arguments that the bits that have
reached the upper border can almost√be ignored and that the drift with respect to the
other bits is still in the order Ω(V ∗ / 1 + V ∗ ). Using this (so far unproven) statement,
we apply variable drift (Theorem 2) with xmin = 1/µ (since each pi,t = i/µ for some
i ∈ {1, . . . , µ − 1} if it is not at a border) and
√
h(x) := min{1, c′′ x/ 1 + x}
√
for some constant c′′ . Let x∗ be the point where 1 = c′′ x∗ / 1 + x∗ and note that
x∗ = Θ(1). We obtain the upper bound
p
Z n
Z x∗ √
Z n
(1/µ) 1 + 1/µ
1
dx
xmin
1+x
+
dx =
+
dx +
′′
′′
h(xmin )
c /µ
c x
1/µ h(x)
x∗ 1
1/µ
on √
the expected number
√ of generations. Since the anti-derivative of
ln( 1 + x − 1) − ln( 1 + x + 1), the bound can be simplified to
√
√
1 + x/x is 2 1 + x+
O(1) + O(log µ) + O(n) = O(n)
using µ = o(n). This corresponds to an expected runtime √
of O(λn).
∗
We still have to show that we have a drift of Ω(V / 1 + V ∗ ) if V ∗ ≤ c′ . Actually, we will consider a phase of n2 generations and show that the claim holds with
high probability throughout the phase. We then show that under this assumption the
optimum is sampled with probability Ω(1) in the phase. In case of a failure, we repeat the argumentation and get an expected number of O(1) repetitions, altogether
an expected runtime of O(λn). Another important fact we will use concerns the drift
statement E(pt+1,i − pt,i | pi , σt ) = Ω(pi (1 − pi )/σt ). Above we bounded σt in some
deterministic way, but in fact we only need to bound σt in expectation. Since the function σt 7→ 1/σt is convex, we have by Jensen’s inequality E(1/σt ) ≥ 1/E(σt ), implying
E(pt+1,i − pt,i | pi ) = Ω(pi (1 − pi )/E(σt )).
We have seen above that frequencies at the upper border may contribute negative to
the drift of the φt -value. Hence, to show the claim that the potential also decreases when
20
V ∗ ≤ c′ , we will analyze an arbitrary but fixed frequency and prove that it is likely to
stay in the vicinity of the upper border once having been there. Let us for the moment
assume that we always have E(σt ) = O(1) in a phase of n2 generations. We claim that
this actually happens with probability at least 1/2, a proof for this will be given below.
In case of a failure, we repeat the argumentation.
We consider a frequency i and its distance Xt := µpt,i from the lower border from
the first time on where pt,i = 1 − 1/n. The aim is to show the following: when the
Xt -value drops below 3/4µ for the first time, it will with probability 1 − e−Ω(µ) increase
to at least 3/4µ again before it drops below µ/2. We do not apply Theorem 1 to show
this (although this would be possible) but use a more direct and somewhat simpler
argumentation based on the analysis of the sampling process.
Recalling from Section 2.2 that Xt+1 is obtained by summing up the number of
ones sampled in both the 1st-class individuals and the 2nd-class individuals at generation t + 1. Denote by D ∗ the number of 2nd-class candidates and by C ∗ the number of
2nd-class individuals as in Lemma 6. From the lemma we obtain that D ∗ − C ∗ follows at
least a binomial distribution with parameters Ω(µ) and Ω(1/σt ) = Ω(1). Hence, by Chernoff bounds the probability that D ∗ − C ∗ ≥ c1 µ is at least 1 − e−Ω(µ) . Assuming this to
happen, the number of ones-bits sampled in the D ∗ candidates is at least D ∗ (1 − δ)Xt /µ
with probability 1 − e−Ω(µ) by Chernoff bounds (using that Xt /µ = Θ(1)), where δ is a
constant that can be chosen small enough. Likewise, the number of one-bits sampled in
the µ−C ∗ 1st-class individuals is at least (µ−C ∗ )(1−δ)Xt /µ with probability 1−e−Ω(µ) .
Assuming all this to happen, we have Xt+1 ≥ Xt (1 − δ) + (D ∗ − C ∗ )(1 − δ)Xt /µ (or Xt+1
takes its maximum µ anyway) with probability 1 − e−Ω(µ) . If δ is chosen sufficiently
small compared to c1 , we obtain Xt+1 ≥ Xt + µδ/2 with probability 1 − e−Ω(µ) . Note
that the constants in the exponent of e−Ω(µ) may be different; we use the smallest ones
from these estimations and union bounds to obtain the bound 1 − e−Ω(µ) on the final
probability.
If Xt+1 ≥ Xt + µδ/2 for a certain constant number of iterations, then the frequency
has raised from 1/2 to at least 3/4 in this number of steps. By a union bound over
these many iterations, the probability that a frequency over 3/4 drops to less than 1/2
before returning to at least 3/4 is at least 1 − e−Ω(µ) = 1 − e−Ω(c log n) as claimed above.
Choosing c large enough, we obtain that always pt,i ≥ 1/2 within n2 steps with high
probability, provided that the frequency was at least 3/4 at the start of the phase. The
latter is fulfilled since our consideration start at the first time where Xt = 1 − 1/n.
Using the bound pt,i ≥ 1/2 just derived, we apply Lemma 6 again and obtain for
qt,i = 1 − pt,i that E(qt,i − qt+1,i | qt,i ) = Ω(qt,i ). Hence, using additive drift, the expected
return time from state j/µ (with respect to the q-values), where j ∈ {1, . . . , µ/2}, to
state 0 is O(k). Note that this bound is linear in k. Since the expected successor state
of 0, when left, is O(1/µ), the expected return time to 0 after having left 0 is O(1) as
well. Finally, note that the probability of leaving state 0 is O(µ/n). Hence, looking into
an arbitrary time after reaching the upper border for the first time, the frequency has an
expected value of at least 1 − O(1/n) (see Lemma 8 below for a detailed argumentation).
Consequently, if there are k bits that have been at the upper border at least once, the
21
probability of sampling only 1s at all these bits is at least
k
Y
1
1
1−O
1−
= Ω(1).
n
n
i=1
This still allows the optimum to be sampled with probability Ω(1) after the potential on
the bits that never have hit the border so far has decreased below c′ .
Finally, we have to justify why E(σt′ ) = O(1) after for all t′ ≥ t within n2 steps
after the first time t where φt = O(1). The frequencies that never have been at the
upper border are at most 1 − 1/µ = 1 − ω(1/n), contribute O(1) to φt by assumption
and, as already analyzed, will in expectation not decrease towards time t + 1. The
remaining frequencies have an expected value of 1 − O(1/n) (by the analysis above), so
their expected contribution to the φt′ -value is O(1) for all t′ ≥ t by induction. The claim
follows since always σt ≤ φt .
Finally, we give a proof of a property that was only sketched in the proof of Theorem 4
above. A frequency is considered after the first point of time where it has hit the upper
border and proved to be very likely to stay there in subsequent steps. Due to the Markov
property, the time where the order is hit can w. l. o. g. be assume to be 0.
Lemma 8. Let µ = o(n). Consider the frequencies pt , t ≥ 0, belonging to an arbitrary
but fixed bit and suppose that
• p0 = 1 − 1/n
• E(σt ) = O(1) for all t ≥ 0 and
• pt ≥ 1/2 for all pt ≥ 0.
Then for all t ≥ 0 the probability of sampling a one at the bit is 1 − O(1/n).
Proof. Consider an arbitrary point in time t where pt = 1 − 1/n. We analyze the
distribution of pt+1 (so we consider the original state space of the frequencies and not
the complementary probabilities used in the main theorem). Since the µ best individuals
are biased towards one-entries, the number of zero-bits sampled at the bit among the µ
best is stochastically smaller than Z ∼ Bin(µ, 1/n). Since µ = o(n), even the expected
number of zero-bits sampled is o(1). Hence Pr(Z = k) is exponentially decreasing in
k. Therefore, under the condition that pt+1 < 1 − 1/n (which means that at least one
zero-bit is sampled), the expected number is 0-bits sampled is O(1) and the expected
value of pt+1 is (µ − O(1))/µ. As the process is a submartingale (unless it has hit the
upper border) also the expected value of pt+i , i ≥ 2, is bounded in the same way. We
will show in the remainder of the proof that for all t ≥ 0, we have Pr(pt = 1 − 1/n) =
1 − O(µ/n) ≥ 1 − cµ/n for some constant c > 0. Using the law of total probability to
combine the cases pt = 1 − 1/n and pt < 1 − 1/n, we obtain that a 1-bit is sampled at
time t with probability at least
1
cµ µ − O(1)
cµ
= 1 − O(1/n).
1−
+
1−
n
n
n
µ
22
We are left with the claim Pr(pt = 1 − 1/n) = 1 − O(µ/n). Recall that on pt+1 <
1 − 1/n we have E(pt+1 ) = (µ − O(1))/µ. It also follows from the second and third
assumption of the lemma that we have additive drift Ω(1) the whole time. The expected
time to raise the frequency from i/µ to 1 − 1/n is therefore O(µ − i), which is linear in i.
Hence, after state 1 − 1/n has been left, the total expected time to return to 1 − 1/n is
O(1). By Markov’s inequality, there is a constant c such that after c steps the process
has returned to 1 − 1/n with probability at least 1/2. Again using the submartingale
property, the expected state after c steps is no less than the state before these steps.
Hence, we can repeat the argumentation and obtain that after every block of c steps the
process has returned to 1 − 1/n with probability at least 1/2.
We take now a simpler view by means of a two-state Markov chain, where state 0
corresponds to frequency 1− 1/n and state 1 to the rest of the states. Time is considered
in blocks of c steps, which will be justified in the final paragraph of this proof. The
transition probability from 0 to 1 is O(µ/n) = o(1) and the one from 1 to 0 is 1/2;
the remaining probabilities are self-loops. Now, it is easy to analyze the steady state
probabilities, which are 1 − O(µ/n) for state 0 and O(1/n) for state 1. Moreover, since
the chain starts in state 0, simple calculations of occupation probabilities over time yield
for state 0 a probability of 1 − O(µ/n) for all points of time t ≥ 0. More precisely, at the
transition from time t to time t+1 the occupation probability of state 0 can only decrease
by O(µ/n). When state 1 exceeds an occupation probability of c′ µ/n for a sufficiently
large constant c′ > 0, the process goes to state 0 with probability at least c′ µ/(2n), which
is less than the decrease of the occupation probability for state 0 for c′ large enough.
Hence, the occupation probability of state 0 cannot drop below 1 − O(µ/n).
Finally, we argue why we may consider phases of length c in the Markov chain
analysis. Note that only every c-th step a transition from state 1 to 0 is possible (in our
pessimistic model), however, in fact every step can transit from state 0 to 1. Formally,
we have to work in these additional steps in our two-state model. We do so by increasing
the probability of leaving state 0 by a factor of c, which vanishes in the O(µ/n) bound
used above.
We have now concluded the proof of Theorem 4. As mentioned before, we can from
this theorem extract a second value of µ that gives the O(n log n) runtime bound, namely
µ = c′ log n. We also believe that values µ = o(log n) will lead to a too coarse-grained
frequency scale and exponential lower bounds on the runtime, which can be regarded as
another phase transition in the behavior. We do not give a proof here but only mention
that such a phase transition from polynomial to exponential runtime is known from ACO
algorithms and non-elitist (1, λ) EAs when a parameter crosses log n [21, 23].
Conclusions
√
We have analyzed the UMDA on OneMax and obtained the upper bounds O(µ n) and
√
O(µn) on its expected runtime in different domains for µ, more precisely if µ ≥ c n log n
and µ ≥ c′ log n, respectively, where c, c′ are positive constants. This implies an expected
23
runtime of O(n log n) for two asymptotic values of µ, closing the previous gap between
√
the lower bound Ω(µ n + n log n) and the upper bound O(n log n log log n). In our
proofs, we provide detailed tools for the analysis of the stochastic processes at single
frequencies in the UMDA. We hope that these tools will be fruitful in future analyses of
EDAs.
We note that all our results assume λ = O(µ). However, we do not think that larger λ
can be beneficial; if λ = αµ, for α = ω(1), the progress due to 2nd-class individuals can
be by a factor of at most α bigger; however, also the computational effort per generation
would grow by this factor. A formal analysis of other ratios between µ and λ is open,
as is the case of sublogarithmic µ. Moreover, we do not have lower bounds matching
the upper bounds from Theorem 3 if µ is in the regime where both µ = ω(log n) and
√
µ = o( n log n).
Acknowledgments Financial support by the Danish Council for Independent Research (DFF-FNU 4002–00542) is gratefully acknowledged.
A
A.1
Appendix
Proof of Theorem 1
We will use Hajek’s drift theorem to prove Lemma 1. As we are dealing with a stochastic
process, we implicitly assume that the random variables Xt , t ≥ 0, are adapted to some
filtration Ft such as the natural filtration X0 , . . . , Xt , t ≥ 0.
We do not formulate the theorem using a potential/Lyapunov function g mapping
from some state space to the reals either. Instead, we w. l. o. g. assume the random
variables Xt as already obtained by the mapping.
The following theorem follows immediately from taking Conditions D1 and D2 in
[11] and applying Inequality (2.8) in a union bound over L(ℓ) time steps.
Theorem 5 ([11]). Let Xt , t ≥ 0, be real-valued random variables describing a stochastic
process over some state space, adapted to a filtration Ft . Pick two real numbers a(ℓ) and
b(ℓ) depending on a parameter ℓ such that a(ℓ) < b(ℓ) holds. Let T (ℓ) be the random
variable denoting the earliest point in time t ≥ 0 such that Xt ≤ a(ℓ) holds. If there are
λ(ℓ) > 0 and p(ℓ) > 0 such that the condition
1
E e−λ(ℓ)·(Xt+1 −Xt ) | Ft ; a(ℓ) < Xt < b(ℓ) ≤ 1 −
p(ℓ)
holds for all t ≥ 0 then for all time bounds L(ℓ) ≥ 0
Pr T (ℓ) ≤ L(ℓ) | X0 ≥ b(ℓ) ≤ e−λ(ℓ)·(b(ℓ)−a(ℓ)) · L(ℓ) · D(ℓ) · p(ℓ),
where D(ℓ) = max 1, E e−λ(ℓ)·(Xt+1 −b(ℓ)) | Ft ; Xt ≥ b(ℓ) .
24
(∗)
Proof of Theorem 1. We will apply Theorem 5 for suitable choices of its variables, some
of which might depend on the parameter ℓ = b−a denoting the length of the interval [a, b].
The following argumentation is also inspired by Hajek’s work [11].
Fix t ≥ 0. For notational convenience, we let ∆ := (Xt+1 − Xt | Ft ; a < Xt < b) and
omit to state the filtration Ft hereinafter. The aim is to bound the moment-generating
function (mgf.) from Condition (∗). First we observe that it is sufficient to bound the
mgf. of ∆ · 1{∆ ≤ κǫ} since
E(e−λ∆ ) = E(e−λ∆1{∆≤κǫ}−λ∆1{∆>κǫ} )
= E(e−λ∆1{∆≤κǫ} e−λ∆1{∆>κǫ} ) ≤ E(e−λ∆1{∆≤κǫ} ),
using ∆1{∆ > κǫ} > 0 and hence e−λ∆1{∆>κǫ} ≤ 1. In the following, we omit the factor
1{∆ ≤ κǫ} but implicitly multiply ∆ with it all the time.
To establish Condition (∗), it is sufficient to identify values λ := λ(ℓ) > 0 and p(ℓ) > 0
such that
1
.
E(e−λ∆ ) ≤ 1 −
p(ℓ)
Using the series expansion of the exponential function, we get
E(e−λ∆ ) = 1 − λE(∆) +
= 1 − λǫ +
∞
X
(−λ)k
k=2
k!
∞
X
(−λ)k
k=2
k!
E(∆k )
E(∆k 1{∆ ≥ 0}) + E(∆k 1{∆ < 0}) .
We first concentrate on the positive steps in the direction of the expected value, more
precisely, we consider for any odd k ≥ 3
Mk :=
λk+1
λk
E(∆k 1{∆ ≥ 0}) −
E(∆k+1 1{∆ ≥ 0}).
k!
(k + 1)!
Since we implicitly multiply with 1{∆ ≤ κǫ}, we have ∆k 1{∆ ≥ 0} ≤ (κǫ)k and hence
|E(∆k+1 1{∆ ≥ 0})/E(∆k 1{∆ ≥ 0})| ≤ κǫ. By choosing λ ≤ 1/(κ2 ǫ), we have
Mk ≥
λk
λk
E(∆k 1{∆ ≥ 0}) − 2
κǫE(∆k 1{∆ ≥ 0}) ≥ 0,
k!
κ ǫ(k + 1)!
for k ≥ 3, using κ ≥ 1. Hence,
λ2
E(∆2 1{∆ ≥ 0})
2
1
≤ 1 − λǫ + λ 2 (κǫ)2 ≤ 1 − λǫ/2
2κ ǫ
E(e−λ∆ ) ≤ 1 − λǫ +
using again λ ≤ 1/(κ2 ǫ). So, we have estimated the contribution of all the positive steps
by 1 − λE(∆)/2.
25
We proceed with the remaining terms. We overestimate the sum by using ∆′ :=
|∆·1{∆ < 0})| and bounding (−λk ) ≤ λk in all terms starting from k = 2. Incorporating
the contribution of the positive steps, we obtain for all γ ≥ λ
E(e−λ∆ ) ≤ 1 −
∞
λ
λ2 X γ k
E(∆) + 2
E(∆′k )
2
γ
k!
k=2
∞
2
X
λ
λ
≤ 1 − E(∆) + 2
2
γ
′
γk
E(eγ∆ )
λ
E(∆′k ) ≤ 1 − ǫ + λ2
,
k!
2
γ2
k=0
| {z }
=:C(γ)
where the last inequality uses the first condition of the theorem, i. e., the bound on the
drift.
Given any γ > 0, choosing λ := min{1/(κ2 ǫ), γ, ǫ/(4C(γ))} results in
E(e−λ∆ ) ≤ 1 −
ǫ
λǫ
1
λ
ǫ+λ·
· C(γ) = 1 −
= 1−
2
4C(γ)
4
p(ℓ)
with p(ℓ) := 4/(λǫ).
′
The aim is now to choose γ in such a way that E(eγ∆ ) is bounded from above by a
constant. We get
γ∆′
E(e
) ≤
∞
X
γ(j+1)r
e
j=0
Pr(∆ ≤ −jr) ≤
∞
X
eγ(j+1)r e−j
j=0
where the inequality uses the second condition of the theorem.
Choosing γ := 1/(2r) yields
′
E(eγ∆ ) ≤
∞
X
e(j+1)/2−j = e1/2
∞
X
j=0
j=0
e−j/2 = e1/2
1
≤ 4.2.
1 − e−1/2
Hence, λ ≤ ǫ/(4 · 4.2r 2 ) ≤ ǫ/17r 2 . From the definition of λ, we altogether have λ =
min{1/(2r), ǫ/(17r 2 ), 1/(κ2 ǫ)}. Since p(ℓ) = 4/(λǫ), we know p(ℓ) = O(r/ǫ + r 2 /ǫ2 + κ2 ).
Condition (∗) of Theorem 5 has been established along with these bounds on p(ℓ) and
λ = λ(ℓ).
To bound the probability of a success within L(ℓ) steps, we still need a bound on
D(ℓ) = max{1, E(e−λ(Xt+1 −b) | Xt ≥ b)}. If 1 does not maximize the expression then
D(ℓ) = E(e−λ(Xt+1 −b) | Xt ≥ b) ≤ E(e−λ|∆| | Xt ≥ b)
′
≤ 1 + E(eγ∆ | Xt ≥ b),
where the first inequality follows from Xt ≥ b and the second one from γ ≥ λ along with
the bound +1 for the positive terms as argued above. The last term can be bounded as
′
in the above calculation leading to E(eγ∆ ) = O(1) since that estimation uses only the
26
second condition, which holds conditional on Xt > a. Hence, in any case D(ℓ) = O(1).
Altogether, we have
e−λ(ℓ)·ℓ · D(ℓ) · p(ℓ) ≤ e−λℓ ·
4
λǫ
= e−λℓǫ+ln(4/(λǫ))
By the third condition, we have λℓ ≥ 2 ln(4/(λǫ)), which finally means that
e−λ(ℓ)·ℓ · D(ℓ) · p(ℓ) ≤ O(e−λℓǫ/2))
Choosing L(ℓ) = eλℓ/4 , Theorem 5 yields
Pr(T (ℓ) ≤ L(ℓ)) ≤ L(ℓ) · O(e−λℓ/4 ) = O(e−λℓ/4 ),
which proves the theorem.
References
[1] Tianshi Chen, Ke Tang, Guoliang Chen, and Xin Yao. On the analysis of average
time complexity of estimation of distribution algorithms. In Proc. of CEC ’07, pages
453–460, 2007.
[2] Tianshi Chen, Per Kristian Lehre, Ke Tang, and Xin Yao. When is an estimation of
distribution algorithm better than an evolutionary algorithm? In Proc. of CEC ’09,
pages 1470–1477, 2009.
[3] Tianshi Chen, Ke Tang, Guoliang Chen, and Xin Yao. Rigorous time complexity
analysis of univariate marginal distribution algorithm with margins. In Proc. of
CEC ’09, pages 2157–2164, 2009.
[4] Tianshi Chen, Ke Tang, Guoliang Chen, and Xin Yao. Analysis of computational
time of simple estimation of distribution algorithms. IEEE Transactions on Evolutionary Computation, 14(1):1–22, 2010.
[5] Duc-Cuong Dang and Per Kristian Lehre. Simplified runtime analysis of estimation
of distribution algorithms. In Proc. of GECCO ’15, pages 513–518. ACM Press,
2015.
[6] Stefan Droste. A rigorous analysis of the compact genetic algorithm for linear
functions. Natural Computing, 5(3):257–283, 2006.
[7] William Feller. An Introduction to Probability Theory and Its Applications, volume 1. Wiley, 1968.
[8] William Feller. An Introduction to Probability Theory and Its Applications, volume 2. Wiley, 1971.
27
[9] Tobias Friedrich, Timo Kötzing, Martin S. Krejca, and Andrew M. Sutton. The
benefit of recombination in noisy evolutionary search. In Proc. of ISAAC ’15, pages
140–150. Springer, 2015.
[10] Tobias Friedrich, Timo Kötzing, and Martin S. Krejca. EDAs cannot be balanced
and stable. In Proc. of GECCO ’16, pages 1139–1146. ACM Press, 2016.
[11] Bruce Hajek. Hitting-time and occupation-time bounds implied by drift analysis
with applications. Advanced Applied Probability, 14:502–525, 1982.
[12] Mark Hauschild and Martin Pelikan. An introduction and survey of estimation of
distribution algorithms. Swarm and Evolutionary Computation, 1(3):111–128, 2011.
[13] Daniel Johannsen. Random combinatorial structures and randomized search
heuristics. PhD thesis, Universität des Saarlandes, Germany, 2010. URL
http://scidok.sulb.uni-saarland.de/volltexte/2011/3529.
[14] Martin S. Krejca and Carsten Witt. Lower bounds on the run time of the univariate
marginal distribution algorithm on OneMax. In Proc. of FOGA 2017, pages 65–79.
ACM Press, 2017.
[15] Pedro Larrañaga and Jose A. Lozano, editors. Estimation of Distribution Algorithms: A New Tool for Evolutionary Computation, volume 2 of Genetic Algorithms
and Evolutionary Computation. Springer, 2002.
[16] Per Kristian Lehre and Phan Trung Hai Nguyen. Improved runtime bounds for
the univariate marginal distribution algorithm via anti-concentration. In Proc. of
GECCO ’17. ACM Press, 2017. To appear.
[17] Per Kristian Lehre and Carsten Witt. Concentrated hitting times of randomized
search heuristics with variable drift. In Proc. of ISAAC ’14, pages 686–697. Springer,
2014. Full technical report at http://arxiv.org/abs/1307.2559.
[18] Colin McDiarmid. Concentration. In M. Habib, C. McDiarmid, J. Ramirez-Alfonsin,
and B. Reed, editors, Probabilistic Methods for Algorithmic Discrete Mathematics,
page 195–247. Springer, 1998.
[19] Boris Mitavskiy, Jonathan E. Rowe, and Chris Cannings. Theoretical analysis of
local search strategies to optimize network communication subject to preserving the
total number of links. International Journal of Intelligent Computing and Cybernetics, 2(2):243–284, 2009.
[20] Heinz Mühlenbein and Gerhard Paass. From Recombination of Genes to the Estimation of Distributions I. Binary Parameters. In Proc. of PPSN IV, pages 178–187.
Springer, 1996.
[21] Frank Neumann, Dirk Sudholt, and Carsten Witt. A few ants are enough: ACO
with iteration-best update. In Proc. of GECCO ’10, pages 63–70. ACM Press, 2010.
28
[22] Pietro S. Oliveto and Carsten Witt. Improved time complexity analysis of the
simple genetic algorithm. Theoretical Computer Science, 605:21–41, 2015.
[23] Jonathan E. Rowe and Dirk Sudholt. The choice of the offspring population size in
the (1, λ) evolutionary algorithm. Theoretical Computer Science, 545:20–38, 2014.
[24] Stephen M. Samuels. On the number of successes in independent trials. The Annals
of Mathematical Statistics, 36(4):1272–1278, 08 1965.
[25] Dirk Sudholt. A new method for lower bounds on the running time of evolutionary
algorithms. IEEE Transactions on Evolutionary Computation, 17(3):418–435, 2013.
[26] Dirk Sudholt and Carsten Witt. Update strength in EDAs and ACO: How to avoid
genetic drift. In Proc. of GECCO ’16, pages 61–68. ACM Press, 2016.
[27] Eric W. Weisstein.
Normal distribution function.
In
MathWorld–A Wolfram Web Resource. Accessed February 1,
http://mathworld.wolfram.com/NormalDistributionFunction.html.
From
2017.
[28] Carsten Witt. Tight bounds on the optimization time of a randomized search
heuristic on linear functions. Combinatorics, Probability and Computing, 22(2):
294–318, 2013.
29
| 9cs.NE
|
Submitted to Bernoulli
arXiv: arXiv:1705.09898
arXiv:1705.09898v2 [cs.IT] 16 Jun 2017
Projection Theorems of Divergences and
Likelihood Maximization Methods
ATIN GAYEN and M. ASHOK KUMAR
Discipline of Mathematics, Indian Institute of Technology Indore, Indore 453 553.
Email: [email protected] and [email protected]
Abstract:
Projection theorems of divergences enable us to find reverse projection of a divergence on a
specific statistical model as a forward projection of the divergence on a different but rather
“simpler” statistical model, which, in turn, results in solving a system of linear equations. Reverse projection of divergences are closely related to various estimation methods such as the
maximum likelihood estimation or its variants in robust statistics. We consider projection theorems of three parametric families of divergences that are widely used in robust statistics, namely
the Rényi divergences (or the Cressie-Reed power divergences), density power divergences, and
the relative α-entropy (or the logarithmic density power divergences). We explore these projection theorems from the usual likelihood maximization approach and from the principle of
sufficiency. In particular, we show the equivalence of solving the estimation problems by the
projection theorems of the respective divergences and by directly solving the corresponding
estimating equations. We also derive the projection theorem for the density power divergences.
Keywords: Estimating equations, power divergence, power-law family, projection theorem, relative α-entropy, Rényi divergence, robust estimation, sufficient statistics.
1. Introduction
Minimum divergence1 (or minimum distance) method has a unique place in statistical
inference because of its ability to trade-off between efficiency and robustness [3, 28].
Minimization of Kullback-Leibler divergence (I-divergence) or relative entropy is closely
related to the maximum likelihood estimation (MLE) [17, Lemma 3.1]. MLE is not a
preferred method when the data set is contaminated by some unexpected sample points
called outliers. However, I-divergence can be generalized, replacing the logarithmic function by some power function, to produce divergences that are more robust with respect
to outliers [2, 21]. In this paper we consider three such families of divergences that are
well-known in the context of robust statistics. They are defined as follows.
1 By a divergence, we mean a non-negative extended real valued function defined for every pair of
probability measures (P, Q) satisfying D(P, Q) = 0 if and only if P = Q.
1
imsart-bj ver. 2014/10/16 file: ProjectionLikelihood.tex date: June 19, 2017
2
ATIN GAYEN AND ASHOK KUMAR
Let X be a finite set. Let P := P(X ) be the space of all strictly positive probability
measures2 on X . Let α > 0, α 6= 1, and let P, Q ∈ P.
1. The Dα -divergence3 (also known as Rényi divergence [30] or power divergence [29]):
X
1
log
P (x)α Q(x)1−α .
(1.1)
Dα (P, Q) :=
α−1
x∈X
2. The Bα -divergence (also known as density power divergence [2]):
X
α X
1 X
Bα (P, Q) :=
P (x)Q(x)α−1 −
P (x)α +
Q(x)α .
1−α
1−α
x∈X
x∈X
(1.2)
x∈X
3. The Iα -divergence [31], [32], [26], [19] (also known as relative α-entropy [23], [24],
logarithmic density power divergence [27]):
Iα (P, Q)
X
X
X
1
α
log
log
P (x)Q(x)α−1 −
P (x)α + log
Q(x)α .(1.3)
:=
1−α
1−α
x∈X
x∈X
x∈X
Throughout the paper log stands for the natural logarithm. It should be noted that,
although the above divergences are not defined for α = 1, they all coincide with the
I-divergence as α → 1 [7]. That is,
X
P (x)
. (1.4)
lim Bα (P, Q) = lim Iα (P, Q) = lim Dα (P, Q) = I(P, Q) :=
P (x) log
α→1
α→1
α→1
Q(x)
x∈X
In this sense each of these three classes of divergences can be regarded as a generalization
of the I-divergence.
Dα -divergences also arise as generalized cut-off rates in information theory [13]. The
Bα -divergences belong to the Bregman class which is characterized by transitive projection rules (see [12, Eq. (3.2) and Th. 3]). The Iα -divergence (for α < 1) arises in
information theory as redundancy in mismatched guessing moments [31], in mismatched
compression [24], and in mismatched encoding of tasks [5]. The three classes of divergences are associated with robust inference for α > 1 in case of Bα and Iα , and α < 1
in case of Dα , as we shall see now.
Suppose that X1 , . . . , Xn are an independent and identically distributed (i.i.d.) random
sample drawn according to a particular member of a parametric family of probability
measures, Π = {Pθ : θ ∈ Θ} ⊂ P, where Θ is an open subset of Rk . To find the MLE of
the parameter θ, one needs to solve the so-called score equation or estimating equation
for θ, given by
n
X
s(Xj ; θ) = 0,
(1.5)
j=1
2 P can be just thought of the set of all ordered |X |-tuples with strictly positive components that sum
to one.
3 Upto a monotone function, same as the Cressie-Read power divergence [8].
imsart-bj ver. 2014/10/16 file: ProjectionLikelihood.tex date: June 19, 2017
Projection Theorems of Divergences and Likelihood Maximization Methods
3
where s(x; θ) := ∇ log Pθ (x), called the score function and ∇ stands for gradient with
respect to θ. Observe that (1.5) can be re-written as
X
Pb(x)s(x; θ) = 0,
(1.6)
x∈X
where Pb (·) is the empirical probability measure of the sample X1 , . . . , Xn .
In the presence of outliers in the observed sample, one modifies the score equation by
scaling the score function in (1.6) by weights that down-weights the effect of outliers.
The following estimating equation, referred as generalized Hellinger estimating equation,
was proposed where the score function was weighted by Pb (·)α Pθ (·)1−α instead of Pb(·):
X
Pb (x)α Pθ (x)1−α s(x; θ) = 0,
(1.7)
x∈X
where α ∈ (0, 1). The above estimating equation was proposed based on the following
intuition. If x is an outlier in the sample, Pb(x)α is larger while Pθ (x)1−α is smaller for
sufficiently smaller value of α. Hence the terms in (1.7) corresponding to outliers are
down-weighted. (See [3, Sec. 4.3] and the references therein.) The following estimating
equation, where the score function is weighted by power of model probability measure
and equated to its hypothetical one, was proposed by Basu et. al. [2]:
X
X
Pb(x)Pθ (x)α−1 s(x, θ) =
Pθ (x)α s(x, θ),
(1.8)
x∈X
x∈X
where α > 1. Motivated by the works of Field and Smith [18] and Windham [35], further
an alternative estimating equation was proposed by Jones et. al. [21] where the weights
in (1.8) were normalized:
P
P
Pb(x)Pθ (x)α−1 s(x; θ)
Pθ (x)α s(x; θ)
x∈X
x∈X
P
,
(1.9)
=
P b
Pθ (x)α
P (x)Pθ (x)α−1
x∈X
x∈X
where α > 1. It should be noted that all the three
Pestimating equations coincide with
the usual score equation (1.6) when α = 1, since x Pθ (x)s(x; θ) = 0. The estimating
equations (1.6), (1.7), (1.8), and (1.9) are, respectively, associated with the divergences
in (1.4), (1.1), (1.2), and (1.3) in a sense that will be made clear in the following.
Observe that the estimating equations (1.6), (1.7), (1.8), and (1.9) are implications of
the first order optimality condition of, respectively, the usual log-likelihood function and
the following modified likelihood functions,
hX
i
1
(α)
log
Pb (x)α Pθ (x)1−α ,
(1.10)
L1 (θ) :=
1−α
x∈X
(α)
L2 (θ)
1
:=
n
n
X
j=1
X
αPθ (Xj )α−1 − 1
−
Pθ (x)α ,
α−1
(1.11)
x∈X
imsart-bj ver. 2014/10/16 file: ProjectionLikelihood.tex date: June 19, 2017
4
ATIN GAYEN AND ASHOK KUMAR
and
(α)
L3 (θ)
X
n
hX
i
1
α
α−1
− log
Pθ (Xj )
Pθ (x)α .
log
:=
α−1
n j=1
(1.12)
x∈X
Although the above likelihood functions (1.10), (1.11) and (1.12) are not defined for
α = 1, it can be shown that they all coincide with the following usual log-likelihood
function as α → 1:
n
L(θ) :=
1X
log Pθ (Xj ).
n j=1
(1.13)
Moreover, it is easy to see that the Pθ that maximizes the likelihood function in (1.10),
(1.11), (1.12) or (1.13) is same as the one that minimizes Dα (Pb, Pθ ), Bα (Pb , Pθ ), Iα (Pb, Pθ )
or I(Pb , Pθ ) respectively. Thus, for MLE or “robustified MLE”, one needs to solve
min D(Pb , Pθ ),
(1.14)
min D(P, Q)
(1.15)
Pθ ∈Π
where D is either I, Dα , Bα or Iα . The minimizing probability measure Pθ∗ (if exists
and unique) is known as the reverse D-projection of Pb on Π.
A “dual” minimization problem is the so-called forward projection problem, where the
minimization is over the first argument of the divergence function. Given C ⊂ P, and
Q ∈ P, any P ∗ ∈ C that attains
P ∈C
is called a forward D-projection of Q on C. Forward projection is usually on a convex
set or on an α-convex set of probability measures. Forward projection on a convex set
is motivated by the well-known maximum entropy principle of statistical physics [20].
Motivation for forward projection on α-convex set comes from the so-called non-extensive
statistical physics [33, 34, 23]. Forward I-projection on convex sets was extensively studied
by Csiszár [9, 10, 14], Csiszár and Matúš [16], [15], Csiszár and Shields [17], and Csiszár
and Tusnády [11].
Csiszár and Shields showed that the reverse I-projection on an exponential family is
same as the forward I-projection on a linear family4 , which, in turn, is a solution to a
system of linear equations [17, Th. 3.3]. We call such a result a projection theorem for the
associated divergence. Projection theorem for I-divergence was due to an “orthogonal”
relationship between the exponential family and the linear family. Projection theorem
for Iα -divergence was established by Kumar and Sundaresan where the so-called αpower-law family plays the role of exponential family [24, Th. 18 and Th. 21]. Projection
theorem for Dα -divergence was established by Kumar and Sason where a variant of the
α-power-law family, called an α-exponential family, plays the role of exponential family
and the so-called α-linear family plays the role of linear family [25, Th. 6]. Projection
4 See
Definition 2.
imsart-bj ver. 2014/10/16 file: ProjectionLikelihood.tex date: June 19, 2017
Projection Theorems of Divergences and Likelihood Maximization Methods
5
theorem for the more general class of Bregman divergences, where Bα is a subclass, was
established by Csiszár and Matúš [16] using techniques from convex analysis. One of
our goals in this paper is to derive the projection theorem for the Bα -divergence using
elementary tools. We also identify the parametric family of probability measures associated with the projection theorem of Bα -divergence, which is yet another generalization
of the exponential family. In all these projection theorems, the Pythagorean theorem of
the respective divergence plays a key role. Thus projection theorems enable us to find
the estimator (whether MLE or robustified MLE) as a forward projection if the estimation is done on a specific parametric family. While for MLE the required family is the
exponential family, for robustified MLE, it is one of the generalized exponential families.
Our contributions in this paper are the following.
(a) We show that the generalized Hellinger estimating equation (1.7) and the estimating
equation of Jones et al. (1.9) are equivalent under a transformation.
(b) We show the equivalence of the two methods of solving the estimation problems: the
one by the projection theorem and the other by directly solving the corresponding
estimating equation.
(c) We show that the statistics of the data that have bearing on the projection theorem
(that is, projection equation) are the sufficient statistics of the underlying statistical
model with respect to the associated likelihood function.
(d) We derive the projection theorem for the Bα -divergence.
Rest of the paper is organised as follows. The contributions mentioned in (a) and (b)
are covered in section 2. The contribution mentioned in (c) is covered in section 3. The
contribution in (d) is given in Appendix A. The paper ends with a summary in section
4.
2. Relationship Between Projection Theorems and
Estimating Equations
We now show that the solutions of the estimation problems obtained using the projection
theorems and those obtained by directly solving the estimating equations are the same.
Recall that, projection theorem of a divergence enables us to find the reverse projection
in terms of a forward projection if the former is done on the statistical model associated with the divergence. The statistical models associated with the I, Bα , Iα and Dα
divergences are, respectively, exponential family, non-normalized α-power-law family, αpower-law family, and α-exponential family which we shall define now. Indeed, for I, Bα ,
Iα and Dα divergences, the reverse projection (if exists) must satisfy an equation that
we call a projection equation. We shall first list down the statistical models along with
the projection equations associated with these divergences.
imsart-bj ver. 2014/10/16 file: ProjectionLikelihood.tex date: June 19, 2017
6
ATIN GAYEN AND ASHOK KUMAR
1. Projection equation for I-divergence on exponential family:
The exponential family, E := E(Q, f, Θ), characterized by a probability measure
Q ∈ P, k real valued functions fi , i = 1, . . . , k on X , and parameter space Θ ⊂ Rk ,
is given by E = {Pθ : θ ∈ Θ} ⊂ P, where
Pθ (x) = Z(θ)−1 exp log(Q(x)) + θT f (x) for x ∈ X ,
(2.1)
Z(θ) is a normalizing constant that makes Pθ a probability measure, and f =
(f1 , . . . , fk )T .
The reverse I-projection of Pb on E, if exists, is a unique solution of
Eθ [f (X)] = f¯,
(2.2)
P
where f¯ := (f¯1 , . . . , f¯k )T , f¯i := n1 nj=1 fi (Xj ) for i = 1, . . . , k, and Eθ denotes
expected value with respect to Pθ . The above result was due to [17, Th. 3.3]. We
call (2.2) the projection equation for I-divergence on exponential family.
2. Projection equation for Bα on non-normalized α-power-law family:
The non-normalized α-power-law family, B(α) := B(α) (Q, f, Θ) ⊂ P, characterized
by a Q ∈ P, k real valued functions fi , i = 1, . . . , k on X , and parameter space
Θ ⊂ Rk , is defined as in Definition 3.
The reverse Bα -projection of Pb on B(α) , if exists, is a unique solution of
Eθ [f (X)] = f¯,
by Theorem 7. (Notice that the projection equations for I and Bα -divergences are
the same.)
3. Projection equation for Iα on α-power-law family:
The α-power-law family, M(α) := M(α) (Q, f, Θ)5 , characterized by a Q ∈ P and k
real valued functions fi , i = 1, . . . , k on X , and parameter space Θ ⊂ Rk , is given
by M(α) = {Pθ : θ ∈ Θ} ⊂ P, where
1
Pθ (x) = Z(θ)−1 Q(x)α−1 + (1 − α)θT f (x) α−1 for x ∈ X ,
(2.3)
and Z(θ) is the normalizing constant (c.f. [24, Defn. 8]).
The reverse Iα -projection of Pb on M(α) , if exists, is a unique solution of
Eθ Q(X)α−1
Eθ fi (X) =
f¯i ,
(2.4)
Qα−1
Pn
for i = 1, . . . , k, where Qα−1 := n1 j=1 Q(Xj )α−1 [24, Th. 18 and Th. 21]6 .
5 In the continuous case, each Student-t distribution can be seen as point-wise limit of a sequence of
distributions from M(α) . (c.f. [23, Rem. 13].)
6 Notice that this result in [24] is true under the assumption that the associated linear family L
determined by the functions f1 , . . . , fk is non-empty.
imsart-bj ver. 2014/10/16 file: ProjectionLikelihood.tex date: June 19, 2017
Projection Theorems of Divergences and Likelihood Maximization Methods
7
4. Projection equation for Dα on α-exponential family:
The α-exponential family, Eα := Eα (Q, f, Θ), characterized by a Q ∈ P and k real
valued functions fi , i = 1, . . . , k on X , and parameter space Θ ⊂ Rk , is given by
Eα = {Pθ : θ ∈ Θ} ⊂ P, where
1
Pθ (x) = Z(θ)−1 Q(x)1−α + (1 − α)θT f (x) 1−α
for x ∈ X ,
(2.5)
and Z(θ) is the normalizing constant (c.f. [25, Eq. (62)].
The reverse Dα -projection of Pb on Eα , if exists, is a unique solution of
P
Pθ (x)α Q(x)1−α X
X
x∈X
Pb (x)α fi (x),
Pθ (x)α fi (x) = P
b(x)α Q(x)1−α
P
x∈X
x∈X
(2.6)
x∈X
for i = 1, . . . , k ([25, Th. 6]).
(2.6) can be re-written as
Eθ(α)
Eθ(α) Q(X)1−α (α)
fi ,
fi (X) =
(α)
Q1−α
(α)
(α)
(2.7)
(α)
where Eθ(α) denotes expectation with respect to Pθ ; Q1−α and fi are, respectively,
P
averages of Q(·)1−α and fi (·), with respect to Pb(α) , where P (α) (x) := P (x)α / y P (y)α
is the α-scaled measure7 associated with P .
Notice that while the projection equations for I and Bα depend on the sample only
through f¯, the projection equations for Iα and Dα , respectively, depend on f¯/Qα−1 and
(α)
(α)
f /Q1−α . This fact will be further explored in terms of sufficiency principle in section
3. Also notice that (2.4) and (2.7) are related by the transformations P ↔ P (α) and α ↔
1/α. The following lemmas explore this connection further. While lemma 1 establishes
the connection between M(α) and Eα , lemma 2 establishes the connection between the
estimation problem based on Dα -divergence on Eα and the estimation problem based on
Iα -divergence on M(α) . Lemma 1 is due to Karthik and Sundaresan [22, Th. 2], where
only the reverse implication was proved though.
Lemma 1. The map P 7→ P (α) establishes a one-to-one correspondence between the
parametric families Eα (Q, f, Θ) and M(1/α) (Q(α) , f, Θ′ ), where
n
o
(−α)θi
Θ′ = θ′ = (θ1′ , . . . , θk′ ) : θi′ = kQk
(2.8)
1−α , i = 1, . . . , k, θ ∈ Θ ,
P
α 1/α
. (We suppress the dependence of kQk on α for notational
and kQk :=
x Q(x)
convenience.)
7 Sometimes
referred as escort measure among the physicists.
imsart-bj ver. 2014/10/16 file: ProjectionLikelihood.tex date: June 19, 2017
8
ATIN GAYEN AND ASHOK KUMAR
Proof. For any Pθ ∈ Eα (Q, f, Θ), we have from (2.5), for x ∈ X ,
α
h
i 1−α
−α
(α)
1−α
T
Q(x)
+
(1
−
α)θ
f
(x)
Pθ (x) = Z(θ)
kPθ kα
hn
o 1−α
i 11
−α
α
(−α) T
α
−1
Q(x)α
1
α
)
= Z(θ)kPθ kkQk
+
(1
−
θ
f
(x)
α
α
kQk
α kQk1−α
h
i 11
1
−α
α
−1
T
−1
= Z(θ)kPθ kkQk
Q(α) (x) α + (1 − α1 )θ′ f (x) α .
α
(2.9)
(α)
Hence Pθ ∈ M(1/α) (Q(α) , f, Θ′ ). So, the mapping is well-defined. The map is clearly
(α)
(α)
one-one, since it is easy to see that, if Pθ = Pη for some θ, η ∈ Θ, then Pθ = Pη . To
verify it is onto, let P ∈ M(1/α) (Q(α) , f, Θ′ ) be arbitrary. Then
h
i 11
1
−1
T
−1
P (x) = Z(θ′ )−1 Q(α) (x) α + (1 − α1 )θ′ f (x) α
α
hn
o1−α
i 1−α
(1−α) T
= Z(θ′ )−1 Q(x)
θ
f
(x)
+
1−α
kQk
kQk
α
Z(θ ′ )−1
1−α
+ (1 − α)θT f (x) 1−α .
= kQkα Q(x)
This implies that
P (x)1/α =
and hence
P (1/α) (x) =
Z(θ ′ )−1/α
Q(x)1−α
kQk
1
+ (1 − α)θT f (x) 1−α ,
′ −1/α
Z(θ
P)
Q(x)1−α
kQk y P (y)1/α
1
+ (1 − α)θT f (x) 1−α .
Hence P (1/α) ∈ Eα (Q, f, Θ) and so P (1/α) = Pθ for some θ ∈ Θ. It is now easy to show
(α)
that Pθ = P . Thus, for any P ∈ M(1/α) (Q(α) , f, Θ′ ), there exists Pθ ∈ Eα (Q, f, Θ) such
(α)
that Pθ = P . Hence the mapping is onto.
Lemma 2.
The following statements hold.
(a) Solving the estimating equation (1.7) for Pθ ∈ Eα (Q, f, Θ) is equivalent to solving
(α)
the estimating equation (1.9) for the α-scaled measure Pθ ∈ M(1/α) (Q(α) , f, Θ′ ),
′
where Θ is as in (2.8).
(b) Solving the projection equation (2.7) for Pθ ∈ Eα (Q, f, Θ) is equivalent to solving
(α)
the projection equation (2.4) for the α-scaled measure Pθ ∈ M(1/α) (Q(α) , f, Θ′ ),
where Θ′ is as in (2.8).
Proof.
(a) The estimating equation in (1.7) can be re-written as
P
Pb(x)α Pθ (x)1−α s(x; θ)
X
x∈X
=
Pθ (x)s(x; θ),
P b α
P (x) Pθ (x)1−α
x∈X
(2.10)
x∈X
imsart-bj ver. 2014/10/16 file: ProjectionLikelihood.tex date: June 19, 2017
Projection Theorems of Divergences and Likelihood Maximization Methods
since
P
9
Pθ (x)s(x; θ) = 0. This can further be re-written as
x∈X
P b (α) (α) α1 −1
P (x) Pθ (x)
s(x; θ)
x∈X
P b (α) (α) α1 −1
P (x) Pθ (x)
=
x∈X
Observe that
P
x∈X
1
(α)
Pθ (x) α s(x; θ)
P (α) α1
Pθ (x)
.
(2.11)
x∈X
α
(α)
θ (x)
s(α) (x; θ) := ∇ log Pθ (x) = ∇ log PkP
α
θk
= ∇ log Pθ (x)α − log kPθ kα
= α s(x; θ) − ∇ log kPθ k .
Hence
s(x; θ) =
1 (α)
(x; θ)
αs
+ A(θ),
(2.12)
where A(θ) = ∇ log kPθ k. Plugging (2.12) in (2.11), we get
P b(α) (α) α1 −1 (α)
s (x; θ)
P (x) Pθ (x)
x∈X
P
x∈X
(α) 1 −1
Pb(α) (x) Pθ (x) α
=
P
x∈X
1
(α)
Pθ (x) α s(α) (x; θ)
P
x∈X
1
(α)
Pθ (x) α
.
(2.13)
(α)
This is same as (1.9) with Pb, Pθ , and α, respectively, replaced by Pb (α) , Pθ , and
1/α. Thus, solving the estimating equation (1.7) for Pθ ∈ Eα (Q, f, Θ) is equivalent
(α)
to solving (1.9) for Pθ ∈ M(1/α) (Q(α) , f, Θ′ ), by lemma 1.
(b) The projection equation (2.7) can further be re-written as
i
h
1
−1
Eθ(α) Q(α) (X) α
X
Eθ(α) fi (X) =
Pb (α) (x)fi (x).
1
P (α) (α) α −1
x∈X
Pb (x) Q (x)
(2.14)
x∈X
(α)
This is same as (2.4) with Pθ , Pb , Q, and α, respectively, replaced by Pθ , Pb (α) ,
(α)
Q , and 1/α. Thus solving (2.6) for Pθ ∈ Eα (Q, f, Θ) is equivalent to solving
(α)
(2.14) for Pθ ∈ M(1/α) (Q(α) , f, Θ′ ), by lemma 1.
We are now ready to state the main result of this section.
Theorem 1. The solutions of the estimation problems based on the estimating equations
(1.6), (1.7), (1.8), and (1.9) on E, Eα , B(α) , and M(α) respectively, is same as the
solutions obtained from their corresponding projection equations.
imsart-bj ver. 2014/10/16 file: ProjectionLikelihood.tex date: June 19, 2017
10
ATIN GAYEN AND ASHOK KUMAR
Proof.
(a) MLE on exponential family E:
For Pθ ∈ E,
log Pθ (x) = − log Z(θ) + log Q(x) + θT f (x).
(2.15)
Hence (1.5) implies that
n
∇ log Z(θ) =
1X
f (Xj ).
n j=1
(2.16)
Since Eθ [∇ log Pθ (X)] = 0, (2.15) implies that
∇ log Z(θ) = Eθ [f (X)].
(2.17)
Thus, from (2.16) and (2.17), we see that the MLE must satisfy
Eθ [f (X)] = f¯,
which is same as the projection equation for I-divergence.
(b) Robust estimation based on (1.8) on B(α) :
The estimating equation (1.8) can be re-written as
X
X
Pb (x)Pθ (x)α−2 ∇Pθ (x) =
Pθ (x)α−1 ∇Pθ (x).
x∈X
(2.18)
x∈X
For Pθ ∈ B(α) , from (A.22) we have
∇Pθ (x)
1 −1
= − Q(x)α−1 + (1 − α) Z(θ) + θT f (x) α−1 [∇Z(θ) + f (x)]
= −Pθ (x)2−α [∇Z(θ) + f (x)].
Substituting this in (2.18), we get
Eθ [f (X)] = f¯,
which is same as the projection equation for Bα -divergence.
(c) Robust estimation based on (1.9) on M(α) :
The estimating equation (1.9) can be re-written as
1
n
n
P
Pθ (Xi )α−2 ∇Pθ (Xi )
i=1
1
n
n
P
i=1
=
Pθ (Xi )α−1
P
x∈X
Pθ (x)α−1 ∇Pθ (x)
P
.
Pθ (x)α
(2.19)
x∈X
Now if Pθ ∈ M(α) , then from (2.3) we have
Pθ (x)α−1 = Z(θ)1−α Q(x)α−1 + (1 − α)θT f (x)
for x ∈ X .
(2.20)
imsart-bj ver. 2014/10/16 file: ProjectionLikelihood.tex date: June 19, 2017
Projection Theorems of Divergences and Likelihood Maximization Methods
11
Differentiating (2.20) with respect to θ, we get
Pθ (x)α−2 ∇Pθ (x)
= −Z(θ)−α ∇Z(θ) Q(x)α−1 + (1 − α)θT f (x) − Z(θ)1−α f (x)
= −Z(θ)1−α Z(θ)−1 ∇Z(θ) Q(x)α−1 + (1 − α)θT f (x) + f (x) . (2.21)
Using (2.20) and (2.21), left-hand side of (2.19) becomes
1
n
−
n n
P
j=1
o
Z(θ)−1 ∇Z(θ) Q(Xj )α−1 + (1 − α)θT f (Xj ) + f (Xj )
n
P
Q(Xj )α−1 + (1 − α)θT f (Xj )
1
n
j=1
= −Z(θ)−1 ∇Z(θ) −
f¯
Qα−1 + (1 − α)θT f¯
.
Similarly, using (2.20) and (2.21), one can show that the right-hand side of (2.19)
is
1
P
f (x) Q(x)α−1 + (1 − α)θT f (x) α−1
x∈X
−Z(θ)−1 ∇Z(θ) −
.
α
P
Q(x)α−1 + (1 − α)θT f (x) α−1
x∈X
Hence (2.19) is same as
f¯
Qα−1
+ (1 −
α)θT f¯
P
x∈X
=
1
f (x) Q(x)α−1 + (1 − α)θT f (x) α−1
P
x∈X
α
Q(x)α−1 + (1 − α)θT f (x) α−1
.
Using (2.20) the above can be re-written as
f¯
Qα−1 + (1 − α)θT f¯
=
Eθ [f (X)]
.
Eθ [Q(X)α−1 + (1 − α)θT f (X)]
(2.22)
Hence the estimator Pθ must satisfy the above equation.
On the other hand, the projection equation (2.4) can be re-written as
This implies
Eθ fi (X) =
f¯i
Qα−1
Eθ Q(X)α−1 ,
for i = 1, . . . , k.
θT f¯
Eθ Q(X)α−1 .
θT Eθ f (X) =
Qα−1
imsart-bj ver. 2014/10/16 file: ProjectionLikelihood.tex date: June 19, 2017
12
ATIN GAYEN AND ASHOK KUMAR
The above two equations can be combined to
(1 − α)θT f¯
Qα−1
f¯i
=
=
,
Eθ fi (X)
Eθ Q(X)α−1
(1 − α)θT Eθ f (X)
for i = 1, . . . , k.
This implies that the reverse projection Pθ must satisfy
f¯
Qα−1 + (1 − α)θT f¯
i =
,
Eθ fi (X)
Eθ Q(X)α−1 + (1 − α)θT f (X)
for i = 1, . . . , k.
(2.23)
This is same as (2.22) derived from the estimating equation. Hence any solution of
the projection equation (2.4) is a solution of the estimating equation (2.22). Hence,
it suffices to show that, if (2.22) has a solution, it is unique. To this end, we define
Φ := (φ1 , . . . , φk ), where
Eθ fi (X) Qα−1 + (1 − α)θT f¯
E [f (X)]Pθα−1
= θ i
φi (θ) :=
, for i = 1, . . . , k,
kPθ kα
Eθ Q(X)α−1 + (1 − α)θT f (X)
(2.24)
where the second equality follows from (2.20). Then (2.22) reduces to solving
Φ(θ) = f¯
(2.25)
for θ. Thus it remains to show that the mapping θ 7→ Φ(θ) is one-one. This is indeed
the case. The proof is given in Appendix B.
(d) Robust estimation based on (1.7) on Eα :
This clearly follows from (c) in view of lemmas 1 and 2.
3. Projection Theorems and the Principle of
Sufficiency
In the previous section we studied four estimation problems respectively on four statistical models from the perspectives of projection theorems and estimating equations. We
observed that the estimating equations or the projection equations arising from these
estimation problems depended on the given sample only through some specific statistics
of the sample. In this section we explore this observation further from the principle of
sufficiency. The following result, known as factorization theorem [6, Th. 6.2.6], tells us a
way to identify the sufficient statistics from the log-likelihood function.
Factorization Theorem: Let X1 , . . . , Xn be an i.i.d. random sample drawn according to
some model Pθ , θ ∈ Θ. A statistic T (X1 , . . . , Xn ) is a sufficient statistic for θ if and only if
there exists functions g(θ, T (X1 , . . . , Xn )) and h(X1 , . . . , Xn ) such that the log-likelihood
function L(θ) in (1.13) can be written as
L(θ) = g(θ, T (X1 , . . . , Xn )) + h(X1 , . . . , Xn ),
(3.1)
imsart-bj ver. 2014/10/16 file: ProjectionLikelihood.tex date: June 19, 2017
Projection Theorems of Divergences and Likelihood Maximization Methods
13
for all sample points (X1 , . . . , Xn ) and for all θ ∈ Θ.
However, when the data set is contaminated, we saw in section 1 that one should use
some modified likelihood function for inference instead of the usual log-likelihood function. The quantities in (1.10), (1.11) and (1.12) are some examples of modified likelihood
functions. Suppose that LG is one of the modified likelihood functions in (1.10) - (1.12).
The robust estimator of θ is given by
θE := arg max LG (θ).
θ
(3.2)
Now, if we can analogously write the modified likelihood function LG as in (3.1), that is,
if
LG (θ) = g(θ, T (X1 , . . . , Xn )) + h(X1 , . . . , Xn ),
(3.3)
for some functions g and h, then we have
θE = arg max LG (θ)
θ
= arg max[g(θ, T (X1 , . . . , Xn )) + h(X1 , . . . , Xn )]
θ
= arg max g(θ, T (X1 , . . . , Xn )).
θ
This means that the robustified MLE depends on the sample only through the function
T (·). Thus it is reasonable to call such T (·) a sufficient statistics for θ with respect to
the modified likelihood function. In view of this, in the following theorem, we find the
sufficient statistics for the parameters of each of the families E, B(α) , M(α) and Eα when
the appropriate likelihood function is used.
Theorem 2. Let X1 , . . . , Xn be an i.i.d. random sample drawn according to one of the
statistical models E, B(α) , M(α) or Eα . Then the following statements hold.
(a) For E, T1 (X1 , . . . , Xn ) = f¯ is a sufficient statistic for θ with respect to the usual
log-likelihood function L(θ) in (1.13).
(b) For B(α) , T2 (X1 . . . , Xn ) = f¯ is a sufficient statistic for θ with respect to the like(α)
lihood function L2 (θ) in (1.11).
(α)
(c) For M , T3 (X1 . . . , Xn ) = f¯/Qα−1 is a sufficient statistic for θ with respect to
(α)
the likelihood function L3 (θ) in (1.12).
(α)
(α)
(d) For Eα , T4 (X1 . . . , Xn ) = f /Q1−α
is a sufficient statistic for θ with respect
(α)
to the likelihood function L1 (θ) in (1.10).
Proof.
(a) See [6, Th. 6.2.10].
imsart-bj ver. 2014/10/16 file: ProjectionLikelihood.tex date: June 19, 2017
14
ATIN GAYEN AND ASHOK KUMAR
(b) For Pθ ∈ B(α) , using (A.22), the likelihood function in (1.11) can be re-written as
X
n
1 X α Q(Xj )α−1 + (1 − α)[Z(θ) + θT f (Xj )] − 1
(α)
L2 (θ) =
−
Pθ (x)α
n j=1
α−1
x∈X
i
h
X
1
α
+ αθT f¯ +
Pθ (x)α
Qα−1 − αZ(θ) +
=
α−1
α−1
x∈X
= h(X1 , . . . , Xn ) + g(θ, T2 (X1 , . . . , Xn )),
where
α
h(X1 , . . . , Xn ) := α−1
Qα−1 ,
T2 (X1 , . . . , Xn ) := f¯,
h
and g(θ, T2 (X1 , . . . , Xn )) := − αZ(θ) +
1
α−1
i
P
Pθ (x)α .
+ αθT f¯ +
x∈X
Hence T2 (X1 , . . . , Xn ) = f¯ is a sufficient statistics for θ.
(c) For Pθ ∈ M(α) , using (2.20), the likelihood function in (1.12) can be re-written as
(α)
L3 (θ)
n
i
h1 X
X
α
=
Pθ (x)α
Z(θ)1−α Q(Xj )α−1 + (1 − α)θT f (Xj ) − log
log
α−1
n j=1
x∈X
X
α
T ¯
α
α−1
log Q
= −α log Z(θ) +
+ (1 − α)θ f − log
Pθ (x)
α−1
x∈X
α
α
log Qα−1 +
log 1 + (1 − α) θT f¯/Qα−1
= −α log Z(θ) +
α−1
α−1
X
− log
Pθ (x)α
= h(X1 , . . . , Xn ) + g θ, T3 (X1 , . . . , Xn ) ,
where T3 (X1 , . . . , Xn ) := f¯/Qα−1 , h(X1 , . . . , Xn ) :=
g(θ, T3 (X1 , . . . , Xn )) := −α log Z(θ) +
x∈X
α
α−1
log Qα−1 , and
α
log 1 + (1 − α) θT f¯/Qα−1
α−1
X
− log
Pθ (x)α .
x∈X
Hence T3 (X1 , . . . , Xn ) = f¯/Qα−1 is a sufficient statistics for θ.
imsart-bj ver. 2014/10/16 file: ProjectionLikelihood.tex date: June 19, 2017
Projection Theorems of Divergences and Likelihood Maximization Methods
15
(d) For Pθ ∈ Eα , using (2.5), the likelihood function in (1.10) can be re-written as
(α)
L1 (θ) =
=
hX
i
1
log
Pb(x)α Z(θ)α−1 Q(x)1−α + (1 − α)θT f (x)
1−α
x∈X
i
h
(α)
1
(α)
log 1 + (1 − α)θT f /Q1−α
− log Z(θ) +
1−α
X
1
+
log
Pb (x)α Q(x)1−α
1−α
x∈X
=
h(X1 , . . . , Xn ) + g(θ, T4 (X1 , . . . , Xn )),
where
(α)
(α)
(α)
(α)
with f
being, respectively, averand Q1−α
T4 (X1 , . . . , Xn ) := f /Q1−α
ages of f (·) and Q(·)1−α with respect to the measure Pb (α) ,
oi
h
n
(α)
1
(α)
g(θ, T4 (X1 , . . . , Xn )) := − log Z(θ) +
,
log 1 + (1 − α) θT f /Q1−α
1−α
and
P
1
h(X1 , . . . , Xn ) := 1−α
Pb (x)α Q(x)1−α .
log
x∈X
(α)
Hence T4 (X1 , . . . , Xn ) = f
/Q1−α
(α)
is a sufficient statistic for θ.
(α)
(α)
Thus the sufficient statistics f¯, f¯/Qα−1 , and f /Q1−α
are precisely the statistics
of the sample that influence the projection equations (2.2), (2.4), and (2.7) respectively.
4. Summary
In this paper we studied four estimation problems on four statistical models distinctively
associated with them. Each of these estimation problems can be regarded as a reverse
projection problem of one of the divergences I, Bα , Iα and Dα , on one of these statistical
models. Projection theorems tell us that, the reverse projection, if exists, is the unique
solution of an equation, called projection equation. While the projection equations are
simple and easy to interpret in the case of I and Bα -divergences, it is not so in the
case of Iα and Dα -divergences. Our objective in this paper was to understand these
projection theorems (that is, projection equations) from the usual statistical method of
solving by estimating equations. In the case of I and Bα -divergences, we saw that the
projection equations and the estimating equations are the same and both depended only
on the statistics of f which is one of the defining entity of the underlying statistical
model. However, in the case of Iα and Dα -divergences, the estimating equations and
the projection equations are different and both depended not only on the statistics of f
but also on Q which is other defining entity of the model. Nevertheless, we showed that,
in the case of Iα and Dα also, the solutions of the estimation problems obtained both
imsart-bj ver. 2014/10/16 file: ProjectionLikelihood.tex date: June 19, 2017
16
ATIN GAYEN AND ASHOK KUMAR
from the projection equations and from the estimating equations are the same. We then
tried to understand the projection theorems via sufficient statistics in the sense of generalized likelihood functions and observed that the estimating equations (or the projection
equations) depended on the given sample only through these sufficient statistics.
Appendix A: Projection Theorem for Density Power
Divergence
The Projection theorem and the Pythagorean property of the more general class of
Bregman divergences were established by Csiszár and Matúš [16] in generality using tools
from convex analysis. The density power divergences Bα is a subclass of the Bregman
divergences. However, it is hard to extract the results for the Bα -divergence from [16].
Hence we derive those results for the Bα -divergence using some elementary tools as in
[17] for the KL-divergence. We must point out that the geometry of Bα -divergence is
quite a natural extension of that of KL-divergence. Here we assume that P := P(X ) is
the space of all probability measures on X . Let Bα be as defined in (1.2)8 . Let us also
recall the definitions of reverse and forward projections given in (1.14) and (1.15).
Definition 1. For P ∈ P, the support of P is defined as Supp(P ) = {x ∈ X : P (x) >
0}. For C ⊂ P, we will denote Supp(C) for the union of support of members of C.
First we show that, while the Pythagorean inequality is always satisfied when the forward projection is on a closed convex set, equality is satisfied when the forward projection
is on a linear family and α < 1. Here, in the sequel, we assume that Supp(Q) = X .
Theorem 3. Let α > 0, α 6= 1. Let P ∗ be the forward Bα -projection of Q on a closed
and convex set C. Then
Bα (P, Q) ≥ Bα (P, P ∗ ) + Bα (P ∗ , Q)
∀P ∈ C.
(A.1)
∗
Further, if α < 1, Supp(C) = Supp(P ).
Proof. Let P ∈ C and define
Pt (·) = (1 − t)P ∗ (·) + tP (·), for t ∈ [0, 1].
Since C is convex, Pt ∈ C. By mean-value theorem, we have for each t ∈ (0, 1),
0
≤
=
=
8 We
1
Bα (Pt , Q) − Bα (P ∗ , Q)
t
1
Bα (Pt , Q) − Bα (P0 , Q)
t
d
Bα (Pt , Q) t=t̃ , for some t̃ ∈ (0, t).
dt
(A.2)
assume the usual convention that Bα (P, Q) = ∞ when P 6≪ Q and α < 1.
imsart-bj ver. 2014/10/16 file: ProjectionLikelihood.tex date: June 19, 2017
Projection Theorems of Divergences and Likelihood Maximization Methods
17
Using definition of Bα , we have
α X
d
Bα (Pt , Q) =
P (x) − P ∗ (x) Pt (x)α−1 − Q(x)α−1 .
dt
α−1
x∈X
Therefore, using (A.2),
α X
P (x) − P ∗ (x) Pt̃ (x)α−1 − Q(x)α−1 ≥ 0.
α−1
(A.3)
x∈X
Hence, as t ↓ 0, we have
α X
P (x) − P ∗ (x) P ∗ (x)α−1 − Q(x)α−1 ≥ 0,
α−1
(A.4)
x∈X
which is equivalent to (A.1).
If Supp(P ∗ ) 6= Supp(C), that is, if P ∗ (x) = 0 for some x ∈ X and some P ∈ C such
that x ∈ Supp(P ), and if α < 1, then the left-hand side of (A.3) goes to −∞ as t ↓ 0,
which contradicts (A.3). This proves the claim.
If α > 1, in general, Supp(P ∗ ) 6= Supp(C). [24, Example 2] serves as a counterexample
here as well.
We will now show that equality holds in (A.1) for α < 1 when C is a linear family,
which we shall define now.
Definition 2. The linear family, determined by k real valued functions fi , i = 1, . . . , k
on X and k real numbers ai , i = 1, . . . , k, is defined as
o
n
X
L :=
P (x)fi (x) = ai ,
P ∈P:
i = 1, . . . , k .
(A.5)
x∈X
Theorem 4.
Let P ∗ be the forward Bα -projection of Q on L. The following hold.
(a) If α < 1, then the Pythagorean equality holds, that is,
Bα (P, Q) = Bα (P, P ∗ ) + Bα (P ∗ , Q)
∀P ∈ L.
(A.6)
(b) If α > 1 and if Supp(P ∗ ) = Supp(L), then the Pythagorean equality (A.6) holds.
Proof. (a) Let Pt be as in Theorem 3. Since Supp(P ∗ ) = Supp(L), there exists t′ < 0
such that Pt = (1 − t)P ∗ + tP ∈ L for t ∈ (t′ , 0). Hence, for every t ∈ (t′ , 0), there exists
t̃ ∈ (t, 0) such that
α X
P (x) − P ∗ (x) Pt̃ (x)α−1 − Q(x)α−1 ≤ 0.
α−1
x∈X
Proceeding as in Theorem 3, we get (A.4) with a reversed inequality. Thus we have
equality in (A.4). Hence (A.6) holds.
(b) Similar to (a).
imsart-bj ver. 2014/10/16 file: ProjectionLikelihood.tex date: June 19, 2017
18
ATIN GAYEN AND ASHOK KUMAR
Remark 1. When α > 1, equality in (A.6) does not hold in general. [24, Example 2]
serves as a counterexample here as well.
We will now find an explicit expression of the forward Bα -projection in both the cases
α < 1 and α > 1 separately.
Theorem 5.
(A.5).
Let Q ∈ P and consider a linear family L of probability measures as in
(a) If α < 1, the forward Bα -projection P ∗ of Q on L satisfies
k
X
1
θi fi (x) α−1
P ∗ (x) = Q(x)α−1 + (1 − α) Z +
∀x ∈ Supp(L),
(A.7)
i=1
where θ1 , . . . , θk are scalars and Z is a constant.
(b) If α > 1, the forward Bα -projection P ∗ of Q on L satisfies
k
X
1
θi fi (x) +α−1
P ∗ (x) = Q(x)α−1 + (1 − α) Z +
∀x ∈ X ,
(A.8)
i=1
where θ1 , . . . , θk are scalars, Z is a constant, and for any real number r, [r]+ :=
max{r, 0}.
Proof. (a) The linear family in (A.5) can be re-written as
o
n
X
P (x)fi (x) = ai , i = 1, . . . , k .
L := P ∈ P :
(A.9)
x∈Supp(L)
Let F be the subspace of R|Supp(L)| spanned by the k vectors f1 (·) − a1 , . . . , fk (·) − ak .
Then every P ∈ L can be thought of a |Supp(L)|-dimensional vector in F ⊥ . Hence F ⊥
is a subspace of R|Supp(L)| that contains a vector whose components are strictly positive
as P ∗ ∈ L. It follows that F ⊥ is spanned by its probability vectors. From (A.4), we see
that (A.6) is equivalent to
X
P (x) − P ∗ (x) P ∗ (x)α−1 − Q(x)α−1 = 0 ∀P ∈ L.
(A.10)
x∈X
This implies that the vector
P ∗ (·)α−1 − Q(·)α−1 −
X
x
Hence
P ∗ (x)α−1 − Q(x)α−1 −
X
x
P ∗ (x) P ∗ (x)α−1 − Q(x)α−1 ∈ (F ⊥ )⊥ = F .
k
X
P ∗ (x) P ∗ (x)α−1 − Q(x)α−1 =
ci fi (x) − ai
i=1
∀x ∈ Supp(L),
imsart-bj ver. 2014/10/16 file: ProjectionLikelihood.tex date: June 19, 2017
Projection Theorems of Divergences and Likelihood Maximization Methods
19
for some scalars c1 , . . . , ck . This implies (A.7) for appropriate choices of Z and θ1 , . . . , θk .
(b) The proof of this case is similar to the proof for relative α-entropy [24, Th. 14 (b)].
The optimization problem involved in the forward Bα -projection is
min Bα (P, Q)
P
X
subject to
P (x)fi (x) = ai ,
(A.11)
i = 1, . . . , k,
(A.12)
x
X
P (x) = 1,
(A.13)
x
P (x) ≥ 0 ∀x ∈ X .
(A.14)
Hence, by [4, Prop. 3.3.7], there exists Lagrange multipliers λ1 , . . . , λk , ν, and (µ(x), x ∈
X ) respectively associated with the above constraints such that, for x ∈ X ,
∂
Bα (P, Q)
∂P (x)
=
P =P ∗
λi [fi (x) − ai ] + µ(x) − ν,
(A.15)
i=1
µ(x)
µ(x)P ∗ (x)
Since
k
X
≥
=
0,
0.
(A.16)
(A.17)
∂
α
P (x)α−1 − Q(x)α−1 ,
Bα (P, Q) =
∂P (x)
α−1
(A.18)
(A.15) can be re-written as
k
X
α ∗ α−1
λi fi (x) − ai + µ(x) − ν
P (x)
− Q(x)α−1 =
α−1
i=1
for x ∈ X .
(A.19)
Multiplying both sides by P ∗ (x) and summing over all x ∈ X , we get
α X ∗
ν=
P (x) Q(x)α−1 − P ∗ (x)α−1 .
α−1
x∈X
∗
For x ∈ Supp(P ), from (A.17), we must have µ(x) = 0. Then, from (A.19), we have
k
P ∗ (x)α−1 = Q(x)α−1 +
If P ∗ (x) = 0, from (A.19), we get
α−1
α−1X
λi fi (x) − ai −
ν.
α i=1
α
(A.20)
k
Q(x)α−1 +
α−1
α−1
α−1X
λi fi (x) − ai −
ν=−
µ(x) ≤ 0.
α i=1
α
α
(A.21)
Combining (A.20) and (A.21) we get (A.8).
imsart-bj ver. 2014/10/16 file: ProjectionLikelihood.tex date: June 19, 2017
20
ATIN GAYEN AND ASHOK KUMAR
Remark 2. The expression for P ∗ in (A.7) (or in (A.8)) is given with the multiplicative factor 1 − α in the square brackets so that the formula in (A.7) can be rewritten, with the help of α-exponential and α-logarithmic functions [24, Defn. 7], as
k
P
θi fi (x) , analogous to the exponential probability
P ∗ (x)−1 = eα lnα (Q(x)−1 ) + Z +
i=1
measure (see (2.1)).
Theorem 5 suggests us to define a parametric family of probability measures that
extends the usual exponential family. We first formally define this family and then show
an orthogonality relationship between this family and the linear family. As a consequence
we will also show that the reverse Bα -projection on this generalized exponential family
is same as a forward projection on a linear family.
Definition 3. The non-normalized α-power-law family9 , B(α) := B(α) (Q, f, Θ), characterized by a Q ∈ P, k real valued functions fi , i = 1, . . . , k on X , and a parameter
space Θ ⊂ Rk , is given by B(α) = {Pθ : θ ∈ Θ} ⊂ P, where
1
Pθ (x) = Q(x)α−1 + (1 − α) Z(θ) + θT f (x) α−1
for x ∈ X .
(A.22)
Remark 3.
(a) Observe that B(α) is a special case of the family F[βh] in [16, Eq. (28)]
1
[tα − αt + α − t].
when h(·) = Q(·) and β(·, t) = α−1
(b) As in the case of E, M(α) , and Eα families, the family B(α) depends on the reference
measure Q only in a loose manner in the sense that any other member of B(α) can
play the role of Q. The change of reference measure only corresponds to a translation
of the parameter space.
The following theorem and its corollary together establish an “orthogonality” relationship between the non-normalized α-power-law family and its associated linear family.
Theorem 6. Let α < 1. Let B(α) be the non-normalized α-power-law family as in
Definition 3 and L be the corresponding linear family determined by the same functions
fi , i = 1, . . . , k and some constants ai , i = 1, . . . , k as in (A.5). If P ∗ is the forward
Bα -projection of Q on L, then we have the following:
(a) L ∩ cl(B(α) ) = {P ∗ } and
Bα (P, Q) = Bα (P, P ∗ ) + Bα (P ∗ , Q)
∀P ∈ L.
(A.23)
(b) Further, if Supp(L) = X , then L ∩ B(α) = {P ∗ }.
Proof. By Theorem 5, the forward Bα -projection P ∗ of Q on L is in B(α) . This implies
that P ∗ ∈ L ∩ B(α) . In general, it would suffice if we prove the following:
9 This term is coined due to the fact that this family differs from the α-power-law family only in the
normalizing constant.
imsart-bj ver. 2014/10/16 file: ProjectionLikelihood.tex date: June 19, 2017
Projection Theorems of Divergences and Likelihood Maximization Methods
21
(i) Every P̃ ∈ L ∩ cl(B(α) ) satisfies (A.6) with P̃ in place of P ∗ .
(ii) L ∩ cl(B(α) ) is non-empty.
We now proceed to prove both (i) and (ii).
(i) Let P̃ ∈ L ∩ cl(B(α) ). As P̃ ∈ cl(B(α) ), this implies there exists a sequence {Pn } ⊂
B
such that Pn → P̃ as n → ∞. Since Pn ∈ B(α) , we can write
(α)
k
X
(n)
Pn (x)α−1 = Q(x)α−1 + (1 − α) Zn +
θi fi (x)
∀x ∈ X
(A.24)
i=1
(n)
(n)
for some constants θ(n) = (θ1 , . . . , θk ) ∈ Rk and Zn . Now, for any P ∈ L we have,
P
P (x)fi (x) = ai , i = 1, . . . , k. Since P̃ ∈ L, we
from the definition of linear family,
x∈X
P
P̃ (x)fi (x) = ai , i = 1, . . . , k. Multiplying both sides of (A.24) by P (·) and
also have
x∈X
P̃ (·) separately, we get
X
P (x)Pn (x)α−1 =
X
P̃ (x)Pn (x)α−1 =
x∈X
and
k
i
h
X
(n)
θi a i
P (x)Q(x)α−1 + (1 − α) Zn +
X
k
h
i
X
(n)
P̃ (x)Q(x)α−1 + (1 − α) Zn +
θi a i ,
x∈X
x∈X
which imply
X
x∈X
X
x∈X
i=1
i=1
P (x) − P̃ (x) Pn (x)α−1 − Q(x)α−1 = 0.
As n → ∞, the above becomes
X
P (x) − P̃ (x) P̃ (x)α−1 − Q(x)α−1 = 0,
x∈X
which is equivalent to (A.6).
(ii) Let Pn∗ be the forward Bα -projection of Q onto the linear family
n
X
1 X
1
Ln := P :
ai +
P (x)fi (x) = 1 −
Q(x)fi (x),
n
n
x∈X
x∈X
1
o
i = 1, . . . , k .
By construction, 1 − n P + n1 Q ∈ Ln for any P ∈ L. Hence, since Supp(Q) = X , we
have Supp(Ln ) = X . Since Ln is also characterized by the same functions fi , i = 1, . . . , k,
we have Pn∗ ∈ B(α) for every n ∈ N. Hence limit of any convergent subsequence of {Pn∗ }
belongs to cl(B(α) ) ∩ L. Thus cl(B(α) ) ∩ L is non-empty. This completes the proof.
imsart-bj ver. 2014/10/16 file: ProjectionLikelihood.tex date: June 19, 2017
22
ATIN GAYEN AND ASHOK KUMAR
Corollary 1. Let L and B(α) be characterized by the same functions fi , i = 1, . . . , k.
Then L ∩ cl(B(α) ) = {P ∗ } and
Bα (P, Q) = Bα (P, P ∗ ) + Bα (P ∗ , Q)
∀P ∈ L,
∀Q ∈ cl(B(α) ).
(A.25)
Proof. By Theorem 6, we have L ∩ cl(B(α) ) = {P ∗ }. In view of Remark 3(b), notice that
every member of B(α) has the same projection on L, namely P ∗ . Hence (A.25) holds for
every Q ∈ B(α) . Thus we only need to prove (A.25) for every Q ∈ cl(B(α) ) \ B(α) . Let
Q ∈ cl(B(α) ) \ B(α) . There exists {Qn } ⊂ B(α) such that Qn → Q. Hence, for any P ∈ L,
Bα (P, Qn ) = Bα (P, P ∗ ) + Bα (P ∗ , Qn ) ∀n ∈ N.
(A.26)
10
Taking limit as n → ∞, we have (A.25) . This completes the proof.
The following theorem tells us that a reverse Bα -projection on a non-normalized αpower-law family can be turned into a forward Bα -projection on the associated linear
family. We shall refer this as the projection theorem for the Bα -divergence.
Theorem 7. Let α < 1. Let B(α) be as in Definition 3 and let X1 , . . . , Xn be an i.i.d.
sample drawn according to a particular member of B(α) . Let Pb be the empirical probability
measure of X1 , . . . , Xn and let
n
o
X
b := P ∈ P :
L
P (x)fi (x) = f¯i , i = 1, . . . , k ,
(A.27)
x∈X
Pn
∗
b
where f¯i =
j=1 fi (Xj ), i = 1, . . . , k. Let P be the forward Bα -projection of Q on L.
Then the following hold.
(i) If P ∗ ∈ B(α) , then P ∗ is the reverse Bα -projection of Pb on B(α) .
(ii) If P ∗ ∈
/ B(α) , then Pb does not have a reverse Bα -projection on B(α) . However, P ∗
is the reverse Bα -projection of Pb on cl(B(α) ).
1
n
b is so constructed that Pb ∈ L.
b Since the families L
b
Proof. Let us first observe that L
(α)
and B
are defined by the same functions fi , i = 1, . . . , k, by Corollary 1, we have
b ∩ cl(B(α) ) = {P ∗ } and
L
Bα (Pb , Q) = Bα (Pb , P ∗ ) + Bα (P ∗ , Q) ∀Q ∈ cl(B(α) ).
(A.28)
Hence it is clear that the minimizer of Bα (Pb , Q) over Q ∈ cl(B(α) ) is same as the minimizer of Bα (P ∗ , Q) over Q ∈ cl(B(α) ) (Notice that this statement is also true with
cl(B(α) ) replaced by B(α) ). But Bα (P ∗ , Q) over Q ∈ cl(B(α) ) is uniquely minimized by
Q = P ∗ . Hence, if P ∗ ∈
/ B(α) , since minimum of Bα (Pb, Q) over Q ∈ cl(B(α) ) is same as
b
the minimum of Bα (P , Q) over Q ∈ B(α) , the later is not attained on B(α) .
Remark 4. Theorems 6, 7, and Corollary 1 continue to hold for α > 1 as well if
attention is restricted to probability measures with strictly positive components and the
existence of P ∗ is guaranteed.
10 For
a fixed P , Q 7→ Bα (P, Q) is continuous as a function from P to [0, ∞].
imsart-bj ver. 2014/10/16 file: ProjectionLikelihood.tex date: June 19, 2017
Projection Theorems of Divergences and Likelihood Maximization Methods
23
Appendix B: The Map θ 7→ Φ(θ) Defined in (2.24) is
one-one:
Let us first recall that Φ = (φ1 , . . . , φk ), where
φi (θ) =
Eθ [fi (X)]Pθα−1
,
kPθ kα
for i = 1, . . . , k.
We will show that the determinant of the Jacobian matrix of Φ is non-zero. Then the
claim follows from [1, Th. 13.6]. Let ∂j denote ∂/∂θj . Then the (i, j)th entry of the
Jacobian matrix of Φ is given by
∂j φi
=
=
=
Pθα−1
kPθ kα ∂j (Eθ [fi (X)]) − Eθ [fi (X)]∂j (kPθ kα )
kPθ k2α
+
Eθ [fi (X)]
∂j (Pθα−1 )
kPθ kα
Pθ α−1
Eθ [fi (X)]∂j (kPθ kα ) Eθ [fi (X)]
∂j (Eθ [fi (X)]) − Pθα−1
+
∂j (Pθα−1 )
α
kPθ k
kPθ k2α
kPθ kα
i
Pθ α−1
Eθ [fi (X)] h
α−1
α−1
α
+
∂
(P
log
kP
k
∂
)
.
∂
(E
[f
(X)])
+
−P
j
θ
j
j
θ
i
θ
θ
kPθ kα
kPθ kα
(B.1)
We now calculate each of the terms in (B.1).
∂j Eθ [fi (X)]
=
X
x∈X
=
X
x∈X
=
−
∂j [Pθ (x)]fi (x)
1
∂j Z(θ)−1 Q(x)α−1 + (1 − α)θT f (x) α−1 fi (x)
X
Z(θ)−1 Pθ (x)fi (x)∂j [Z(θ)]
x∈X
−Z(θ)−1
X
1 −1
Q(x)α−1 + (1 − α)θT f (x) α−1 fi (x)fj (x)
x∈X
=
−Z(θ)−1 Eθ [fi (X)]∂j [Z(θ)]
X
1 −1
−Z(θ)−1
Z(θ)α−1 Pθ (x)α−1 α−1 fi (x)fj (x)
x∈X
=
−1
−Z(θ)
Eθ [fi (X)]∂j [Z(θ)]
X
Pθ (x)α [Pθ (x)1−α fi (x)][Pθ (x)1−α fj (x)].
1−α
−Z(θ)
x∈X
imsart-bj ver. 2014/10/16 file: ProjectionLikelihood.tex date: June 19, 2017
24
ATIN GAYEN AND ASHOK KUMAR
Hence
Pθ α−1
∂j (Eθ [fi (X)]) = −Pθ α−1 Z(θ)−1 ∂j [Z(θ)]Eθ(α) [Pθ1−α (X)fi (X)]
kPθ kα
−Z(θ)1−α Pθ α−1 Eθ(α) [Pθ1−α (X)fi (X)Pθ1−α (X)fj (X)],
(B.2)
where we used the fact that
Eθ [fi (X)]
= Eθ(α) [Pθ1−α (X)fi (X)].
kPθ kα
Similarly
∂j log kPθ kα
i
h
X
Pθ (x)α
= ∂j log
h
x∈X
= ∂j log Z(θ)−α
X
α i
Q(x)α−1 + (1 − α)θT f (x) α−1
x∈X
−α
= ∂j log Z(θ)
X
α
Q(x)α−1 + (1 − α)θT f (x) α−1
+ ∂j log
= −αZ(θ)−1 ∂j [Z(θ)] −
x∈X
1
P
α
Q(x)α−1 + (1 − α)θT f (x) α−1 fj (x)
x∈X
α
P
Q(x)α−1 + (1 − α)θT f (x) α−1
x∈X
−1
= −αZ(θ)
∂j [Z(θ)] −
α
Z(θ)α kPθ kα
= −αZ(θ)−1 ∂j [Z(θ)] − αZ(θ)1−α
Z(θ)
X
Pθ (x)fj (x)
x∈X
Eθ [fj (X)]
.
kPθ kα
(B.3)
Also
∂j (Pθα−1 ) = ∂j Z(θ)1−α Qα−1 + (1 − α)θT f¯
= (1 − α)Z(θ)−α ∂j [Z(θ)] Qα−1 + (1 − α)θT f¯ + (1 − α)Z(θ)1−α f¯j
= (1 − α)Z(θ)−1 Pθ α−1 ∂j [Z(θ)] + (1 − α)Z(θ)1−α
Eθ [fj (X)]Pθα−1
,
kPθ kα
(B.4)
where in (B.4) we used the fact that
Eθ [fj (X)]Pθα−1
f¯j =
kPθ kα
by (2.23) and (2.24).
imsart-bj ver. 2014/10/16 file: ProjectionLikelihood.tex date: June 19, 2017
Projection Theorems of Divergences and Likelihood Maximization Methods
Using (B.3) and (B.4), the last two terms in (B.1) together yield
i
Eθ [fi (X)] h
α−1
α−1
α
−P
+
∂
log
kP
k
∂
P
j
θ
j
θ
θ
kPθ kα
"
Eθ [fi (X)]
Eθ [fj (X)]
αPθα−1 Z(θ)−1 ∂j [Z(θ)] + αPθα−1 Z(θ)1−α
=
kPθ kα
kPθ kα
−1
+(1 − α)Z(θ)
Pθ
= Z(θ)−1 Pθ α−1 ∂j [Z(θ)]
α−1
∂j [Z(θ)] + (1 −
Eθ [fj (X)]Pθα−1
α)Z(θ)1−α
kPθ kα
25
#
Eθ [fi (X)]
Eθ [fi (X)] Eθ [fj (X)]
+ Z(θ)1−α Pθα−1
kPθ kα
kPθ kα
kPθ kα
= Z(θ)−1 Pθ α−1 ∂j [Z(θ)]Eθ(α) [Pθ1−α (X)fi (X)]
+Z(θ)1−α Pθα−1 Eθ(α) [Pθ1−α (X)fi (X)]Eθ(α) [Pθ1−α (X)fj (X)].
(B.5)
Substituting (B.2) and (B.5) in (B.1), we get
∂j φi
=
−Pθ α−1 Z(θ)−1 ∂j [Z(θ)]Eθ(α) [Pθ1−α (X)fi (X)]
−Z(θ)1−α Pθ α−1 Eθ(α) [Pθ1−α (X)fi (X)Pθ1−α (X)fj (X)]
+Z(θ)−1 Pθ α−1 ∂j [Z(θ)]Eθ(α) [Pθ1−α (X)fi (X)]
+Z(θ)1−α Pθα−1 Eθ(α) [Pθ1−α (X)fi (X)]Eθ(α) [Pθ1−α (X)fj (X)]
=
−Z(θ)1−α Pθα−1 Eθ(α) [Pθ1−α (X)fi (X)Pθ1−α (X)fj (X)]
+Z(θ)1−α Pθα−1 Eθ(α) [Pθ1−α (X)fi (X)]Eθ(α) [Pθ1−α (X)fj (X)]
=
−Z(θ)1−α Pθα−1 Covθ(α) [Pθ1−α (X)fi (X), Pθ1−α (X)fj (X)].
(B.6)
Let A denote the above covariance matrix, that is, the (i, j)th entry of A is
Covθ(α) [Pθ1−α (X)fi (X), Pθ1−α (X)fj (X)].
Then A is symmetric and positive semi-definite. Let B denote the Jacobian matrix of Φ.
Then, from (B.6), we have B = −Z(θ)1−α Pθα−1 A. So, B is also symmetric but negative
semi-definite. If we can show that B is negative definite then then the determinant of B
is non-zero. Thus it suffices to show that A is positive definite. Suppose that A is not
positive definite. Then there exists a non-zero vector c = (c1 , . . . , ck )T such that
cT Ac = 0.
(B.7)
Varθ(α) [c1 Pθ (X)1−α f1 (X) + · · · + ck Pθ (X)1−α fk (X)] = 0.
(B.8)
This implies
Therefore,
c1 Pθ (X)1−α f1 (X) + · · · + ck Pθ (X)1−α fk (X) = t,
imsart-bj ver. 2014/10/16 file: ProjectionLikelihood.tex date: June 19, 2017
26
ATIN GAYEN AND ASHOK KUMAR
where t = Eθ(α) [c1 Pθ (X)1−α f1 (X) + · · · + ck Pθ (X)1−α fk (X)].
c1 f1 (·) + · · · + ck fk (·) = tPθ (·)α−1 .
(B.9)
As Pθ ∈ M(α) , we have
Pθ (·)α−1 = Z(θ)1−α [Q(·)α−1 + (1 − α)θT f (·)],
for some θ = (θ1 , . . . , θk ) ∈ Θ. Substituting this in (B.9), we get
[c1 − (1 − α)tθ1 Z(θ)1−α ]f1 (·) + · · · + [ck − (1 − α)tθk Z(θ)1−α ]fk (·)
−tZ(θ)1−α Q(·)α−1 = 0.
(B.10)
Observe that, with out loss of generality, we can assume that the vectors f1 (·), . . . , fk (·)
in the α-power-law family are linearly independent. Further, in view of footnote 6, we
claim that the vector Q(·)α−1 is linearly independent of the vectors f1 (·), . . . , fk (·). If
T
not, let Q(·)α−1 = c1 f1P
(·) + · · · + ck fk (·) for
P some non-zero vector c = (c1 , . . . , ck ) .
α−1
P (x)[c1 f1 (x) + · · · + ck fk (x)] = 0. This
P (x)Q(x)
=
Then, for any P ∈ L,
x∈X
x∈X
is not possible since Q has full support. Then, from (B.10), we have t = 0 and hence
c1 = · · · = ck = 0, which is a contradiction. Hence A is positive definite. This completes
the proof.
Acknowledgements
Atin Gayen is supported by an INSPIRE fellowship of the Department of Science and
Technology, Government of India.
References
[1] Apostol, T. M. (2002). “Mathematical Analysis,” Indian student 2nd ed., Narosa
Publishing House.
[2] Basu, A., Harris, I. R., Hjort, N. L., and Jones, M. C. (1998). “Robust and efficient
estimation by minimising a density power divergence,” Biometrika, 85(3), pp. 549–
559.
[3] Basu, A., Shioya, H., and Park, C. (2011). “Statistical Inference: The Minimum
Distance Approach,” Chapman & Hall/ CRC Monographs on Statistics and Applied
Probability 120.
[4] Bertsekas, D. P. (2003). “Nonlinear Programming,” 2nd ed. Belmont, MA: Athena
Scientific.
[5] Bunte, C. and Lapidoth, A. (2014). “Encoding tasks and Rényi entropy,” IEEE
Trans. Info. Theory, 60(9), pp. 5065-5076.
[6] Casella, G. and Berger, R. L. (2002). “Statistical Inference,” 2nd ed. Duxbury Press.
imsart-bj ver. 2014/10/16 file: ProjectionLikelihood.tex date: June 19, 2017
Projection Theorems of Divergences and Likelihood Maximization Methods
27
[7] Cichocki, A. and Amari, S. (2010). “Families of Alpha-Beta-and GammaDivergences: Flexible and Robust Measure of Similarities,” Entropy, 12, pp. 1532–
1568.
[8] Cressie, N. and Read, T. R. C. (1984). “Multinomial Goodness-of-Fit Tests,” Journal
of the Royal Statistical Society. Series B (Methodological), 46(3), pp. 440–464.
[9] Csiszár, I. (1975). “I-divergence geometry of probability distributions and minimization problems,” Ann. Probab., 3, pp. 146–158.
[10] Csiszár, I. (1984). “Sanov property, generalized I-projection, and a conditional limit
theorem,” Ann. Probab., 12(3), pp. 768–793.
[11] Csiszár, I. and Tusnády, G. (1984). “Information geometry and alternating minimization procedures,” Statistics and Decisions, Supp. 1, pp. 205–237.
[12] Csiszár, I. (1991). “Why least squares and maximum entropy? An axiomatic approach to inference for linear inverse problems,” Ann. Statist., 19(4), pp. 2032–2066.
[13] Csiszár, I. (1995). “Generalized Cutoff Rates and Rényi’s Information Measures,”
IEEE Trans. Info. Theory, 41(1), pp. 26–34.
[14] Csiszár, I. (1995). “Generalized projections for non-negative functions”, Acta Math.
Hungar., 68(1–2), pp. 161–185.
[15] Csiszár, I. and Matúš, F. (2003). “Information projections revisited,” IEEE Trans.
Info. Theory, 49(6), pp. 1474–1490.
[16] Csiszár, I. and Matúš, F. (2012). “Generalized minimizers of convex integral functionals, Bergman distance, Pythagorean identities.” Kybernetika, 48(4), pp. 637–
689.
[17] Csiszár, I. and Shields, P. C. (2004). “Information Theory and Statistics: A Tutorial,” Foundations and Trends in Communications and Information Theory, 1(4),
pp. 417–528.
[18] Field, C. and Smith, B. (1994). “Robust estimation: A weighted maximum likelihood
approach,” International Statistical Review, 62(3), pp. 405–424.
[19] Fujisawa, H. and Eguchi, S. (2008). “ Robust parameter estimation with a small
bias against heavy contamination”, J. Multivariate Anal., 99, pp. 2053–2081.
[20] Jaynes, E. T. (1983). Papers on Probability, Statistics and Statistical Physics, R.
D. Rosenkrantz, Ed. P.O. Box 17,3300 AA Dordrecht, The Netherlands.: Kluwer
Academic Publishers.
[21] Jones, M. C., Hjort, N. L., Harris, I. R., and Basu, A. (2001). “A comparison of
related density based minimum divergence estimators,” Biometrika, 88(3), pp. 865–
873.
[22] Karthik, P. N. and Sundaresan, R. (2017). “ On the equivalence of projections in relative α-entropy and Rényi divergence,”preprint, [Online]. Available at
https://arxiv.org/abs/1701.06347.
[23] Kumar, M. A. and Sundaresan, R. (2015). “Minimization problems based on relative
α-entropy I: Forward Projection,” IEEE Trans. Info. Theory, 61(9), pp. 5063–5080.
[24] Kumar, M. A. and Sundaresan, R. (2015). “Minimization problems based on relative
α-entropy II: Reverse Projection,” IEEE Trans. Info. Theory, 61(9), pp. 5081–5095.
[25] Kumar, M. A. and Sason, I. (2016). “Projection theorems for the Renyi divergence
on alpha-convex sets, IEEE Trans. Info. Theory, 62(9), pp. 4924–4935.
imsart-bj ver. 2014/10/16 file: ProjectionLikelihood.tex date: June 19, 2017
28
ATIN GAYEN AND ASHOK KUMAR
[26] Lutwak, E., Yang, D., and Zhang, G. (2005) “Cramér-Rao and moment-entropy
inequalities for Rényi entropy and generalized Fisher information,” IEEE Trans.
Info. Theory, 51(2), pp. 473–478.
[27] Maji, A., Ghosh, A., and Basu, A. (2016). “The logarithmic super divergence and
asymptotic inference properties,” Advances in Statistical Analysis, 100, pp. 99–131.
[28] Pardo, L. (2006). “Statistical Inference Based on Divergence Measures,” Chapman
& Hall/CRC, Taylor and Francis group, Boca Raton, Florida, USA.
[29] Patra, S., Maji, A., Pardo, L., and Basu, A. (2013). “The power divergence and the
density power divergence families: the mathematical connection,” Sankhya, Series
B, 75, pp. 16–28.
[30] Rényi, A. (1961). “On measures of entropy and information,” Proceedings of
4th Berekely Symposium on Mathematical Statistics and Probability, pp. 547–561,
Berekeley, California, USA.
[31] Sundaresan, R. (2002). “A measure of discrimination and its geometric properties,”
in Proc. 2002 IEEE Int. Symp. Inf. Theory, Lausanne, Switzerland, pp. 264.
[32] Sundaresan, R. (2007). “Guessing under source uncertainty,” IEEE Trans. Info.
Theory, 53(1), pp. 269–287.
[33] Tsallis, C. (1988). “Possible generalization of Bolzmann-Gibbs statistics,” Journal
of Statistical Physics, 52(1-2), pp. 479–487.
[34] Tsallis, C., Mendes, R. S., and Plastino, A. R. (1998). “The role of constraints within
generalized non-extensive statistics,” Physica A, 261(34), pp. 534–554.
[35] Windham, M. P. (1995). “Robustifying model fitting,” Journal of the Royal Statistical Society, Series B (Methodological), 57(3), pp. 599–609.
imsart-bj ver. 2014/10/16 file: ProjectionLikelihood.tex date: June 19, 2017
| 7cs.IT
|
arXiv:1605.01746v1 [cs.NE] 5 May 2016
Biobjective Performance Assessment with the COCO
Platform
Dimo Brockhoff1 , Tea Tušar1 , Dejan Tušar1 , Tobias Wagner2 ,
Nikolaus Hansen3,4 , Anne Auger3,4
1
Inria, research centre Lille, 2 Technical University Dortmund,
3
Inria, research centre Saclay, 4 Université Paris-Saclay, LRI
Abstract
This document details the rationales behind assessing the performance of numerical blackbox optimizers on multi-objective problems within the COCO platform and in particular on
the biobjective test suite bbob-biobj. The evaluation is based on a hypervolume of all nondominated solutions in the archive of candidate solutions and measures the runtime until the
hypervolume value succeeds prescribed target values.
Contents
1
Introduction
1.1 Definitions and Terminology . . . . . . . . . . . . . . . . . . . . . . . . . . . . .
2
2
2
Performance Assessment with a Quality Indicator
2.1 Definition of the Quality Indicator . . . . . . . . . . . . . . . . . . . . . . . . . .
2.2 Rationales Behind the Performance Measure . . . . . . . . . . . . . . . . . . . . .
2.3 Specificities and Properties . . . . . . . . . . . . . . . . . . . . . . . . . . . . . .
3
3
4
5
3
Definition of Target Values
3.1 Target Precision Values . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . .
3.2 The Reference Hypervolume Indicator Value . . . . . . . . . . . . . . . . . . . .
6
6
7
4
Instances and Generalization Experiment
7
5
Data Storage and Future Recalculations of Indicator Values
7
1
1 Introduction
The performance assessment of (numerical) optimization algorithms with the COCO platform
[HAN2016co] is invariably based on the measurement of the runtime1 until a quality indicator
reaches a predefined target value. On each problem instance, several target values are defined and
for each target value a runtime is measured (or no runtime value is available if the indicator does
not reach the target value) [HAN2016perf]. In the single-objective, noise-free case, the assessed
quality indicator is, at each given time step, the function value of the best solution the algorithm
has obtained (evaluated or recommended, see [HAN2016ex]) before or at this time step.
In the bi- and multi-objective case, e.g. on the biobjective bbob-biobj test suite [TUS2016],
the assessed quality indicator at the given time step is a hypervolume indicator computed from all
solutions obtained (evaluated or recommended) before or at this time step.
1.1 Definitions and Terminology
In this section, we introduce the definitions of some basic terms and concepts.
function instance, problem In the case of the bi-objective performance assessment within COCO,
a problem is a 5-tuple of
• a parameterized function 𝑓𝜃 : R𝑛 → R2 , mapping the decision variables of a solution
𝑥 ∈ R𝑛 to its objective vector 𝑓𝜃 (𝑥) = (𝑓𝛼 (𝑥), 𝑓𝛽 (𝑥)) with 𝑓𝛼 : R𝑛 ↦→ R and 𝑓𝛽 :
R𝑛 ↦→ R being parameterized (single-objective) functions themselves
• its concrete parameter value 𝜃 ∈ Θ determining the so-called function instance 𝑖,
• the problem dimension 𝑛,
• an underlying quality indicator 𝐼, mapping a set of solutions to its quality, and
• a target value 𝐼target of the underlying quality indicator, see below for details.
We call a problem solved by an optimization algorithm if the algorithm reaches a quality indicator value at least as good as the associated target value. The number of function
evaluations needed to surpass the target value for the first time is COCO‘s central performance measure. [HAN2016co] In case a single quality indicator is used for all problems in
a benchmark suite, we can drop the quality indicator and refer to a problem as a quadruple
𝑓𝜃 , 𝜃, 𝑛, 𝐼target . Note that typically more than one problem for a function instance of (𝑓𝜃 , 𝜃, 𝑛)
is defined by choosing more than one target value.
Pareto set, Pareto front, and Pareto dominance For a function instance, i.e., a function 𝑓𝜃 =
(𝑓𝛼 , 𝑓𝛽 ) with given parameter value 𝜃 and dimension 𝑛, the Pareto set is the set of all (Paretooptimal) solutions for which no solutions in the search space R𝑛 exist that have either an
improved 𝑓𝛼 or an improved 𝑓𝛽 value while the other value is at least as good (or in other
1
Time is considered to be number of function evaluations and, consequently, runtime is measured in number of
function evaluations.
2
words, a Pareto-optimal solution in the Pareto set has no other solution that dominates it).
The image of the Pareto set in the objective space is called the Pareto front. We generalize
the standard Pareto dominance relation to sets by saying solution set 𝐴 = {𝑎1 , . . . , 𝑎|𝐴| }
dominates solution set 𝐵 = {𝑏1 , . . . , 𝑏|𝐵| } if and only if for all 𝑏𝑖 ∈ 𝐵 there is at least one
solution 𝑎𝑗 that dominates it.
ideal point The ideal point (in objective space) is defined as the vector in objective space that
contains the optimal function value for each objective independently, i.e. for the above
concrete function instance, the ideal point is given by 𝑧ideal = (inf 𝑥∈R𝑛 𝑓𝛼 (𝑥), inf 𝑥∈R𝑛 𝑓𝛽 (𝑥)).
nadir point The nadir point (in objective space) consists in each objective of the worst value obtained by any Pareto-optimal solution. More precisely, if 𝒫𝒪 denotes the Pareto set, the
nadir point satisfies 𝑧nadir = (sup𝑥∈𝒫𝒪 𝑓𝛼 (𝑥), sup𝑥∈𝒫𝒪 𝑓𝛽 (𝑥)).
archive An external archive or simply an archive is the set of non-dominated solutions, obtained
over an algorithm run. At each point 𝑡 in time (that is after 𝑡 function evaluations), we
consider the set of all mutually non-dominating solutions that have been evaluated so far.
We denote the archive after 𝑡 function evaluations as 𝐴𝑡 and use it to define the performance
of the algorithm in terms of a (quality) indicator function 𝐴𝑡 → R that might depend on a
problem’s underlying parameterized function and its dimension and instance.
2 Performance Assessment with a Quality Indicator
For measuring the runtime on a given problem, we consider a quality indicator which is to be
optimized (minimized). In the noiseless single-objective case, the quality indicator is the best sofar observed objective function value (recommendations can replace previous observations). In the
case of the bbob-biobj test suite, the quality indicator is based on the hypervolume indicator of
the archive 𝐴𝑡 .
2.1 Definition of the Quality Indicator
COCO
The indicator 𝐼HV
to be mininized is either the negative hypervolume indicator of the archive
with the nadir point as reference point or the distance to the region of interest [𝑧ideal , 𝑧nadir ] after a
normalization of the objective space2 :
{︂
−HV(𝐴𝑡 , [𝑧ideal , 𝑧nadir ]) if 𝐴𝑡 dominates {𝑧nadir }
COCO
𝐼HV =
.
𝑑𝑖𝑠𝑡(𝐴𝑡 , [𝑧ideal , 𝑧nadir ]) otherwise
where
)︃
⋃︁ [︂ 𝑓𝛼 (𝑎) − 𝑧ideal,𝛼 ]︂ [︂ 𝑓𝛽 (𝑎) − 𝑧ideal,𝛽 ]︂
,1 ×
,1
HV(𝐴𝑡 , 𝑧ideal , 𝑧nadir ) = VOL
𝑧nadir,𝛼 − 𝑧ideal,𝛼
𝑧nadir,𝛽 − 𝑧ideal,𝛽
𝑎∈𝐴
(︃
𝑡
2
We conduct an affine transformation of both objective function values such that the ideal point 𝑧ideal =
(𝑧ideal,𝛼 , 𝑧ideal,𝛽 ) is mapped to (0, 0) and the nadir point 𝑧nadir = (𝑧nadir,𝛼 , 𝑧nadir,𝛽 ) is mapped to (1, 1).
3
is the (normalized) hypervolume of archive 𝐴𝑡 with respect to the nadir point (𝑧nadir,𝛼 , 𝑧nadir,𝛽 ) as
reference point and where (with division understood to be element-wise, Hadamard division),
⃦
⃦
⃦ 𝑓 (𝑎) − 𝑧 ⃦
⃦
⃦
𝑑𝑖𝑠𝑡(𝐴𝑡 , [𝑧ideal , 𝑧nadir ]) =
inf
𝑎∈𝐴𝑡 ,𝑧∈[𝑧ideal ,𝑧nadir ] ⃦ 𝑧nadir − 𝑧ideal ⃦
is the smallest (normalized) Euclidean distance between a solution in the archive and the region of
interest, see also the figures below for an illustration.
f2
znadir
f1
Fig. 1: Illustration of Coco’s quality indicator (to be minimized) in the (normalized) bi-objective
case if no solution of the archive (blue filled circles) dominates the nadir point (black filled circle),
i.e., the shortest distance of an archive member to the region of interest (ROI), delimited by the
nadir point. Here, it is the fourth point from the left (indicated by the red arrow) that defines the
smallest distance.
2.2 Rationales Behind the Performance Measure
Why using an archive? We believe using an archive to keep all non-dominated solutions is relevant in practice in bi-objective real-world applications, in particular when function evaluations are expensive. Using an external archive for the performance assessment has the
additional advantage that no populuation size needs to be prescribed and algorithms with
different or even changing population sizes can be easily compared.
4
f2
znadir
f1
Fig. 2: Illustration of Coco’s quality indicator (to be minimized) in the bi-objective case if the
nadir point (black filled circle) is dominated by at least one solution in the archive (blue filled
circles). The indicator is the negative hypervolume of the archive with the nadir point as reference
point.
Why hypervolume? Although, in principle, other quality indicators can be used in replacement
of the hypervolume, the monotonicity of the hypervolume is a strong theoretical argument
for using it in the performance assessment: the hypervolume indicator value of the archive
improves if and only if a new non-dominated solution is generated [ZIT2003].
2.3 Specificities and Properties
In summary, the proposed bbob-biobj performance criterion has the following specificities:
• Algorithm performance is measured via runtime until the quality of the archive of nondominated solutions found so far surpasses a target value.
• To compute the quality indicator, the objective space is normalized. The region of interest
(ROI) [𝑧ideal , 𝑧nadir ], defined by the ideal and nadir point, is mapped to [0, 1]2 .
• If the nadir point is dominated by at least one point in the archive, the quality is computed
as the negative hypervolume of the archive using the nadir point as hypervolume reference
point.
• If the nadir point is not dominated by the archive, the quality equals the distance of the
5
archive to the ROI.
This implies that:
• the quality indicator value of an archive that contains the nadir point as non-dominated point
is 0.
• the quality indicator value is bounded from below by −1, which is the quality of an archive
that contains the ideal point, and
• because the quality of an archive is used as performance criterion, no population size has to
be prescribed to the algorithm. In particular, steady-state and generational algorithms can be
compared directly as well as algorithms with varying population size and algorithms which
carry along their external archive themselves.
3 Definition of Target Values
For each problem instance of the benchmark suite, consisting of a parameterized function, its
dimension and its instance parameter 𝜃𝑖 , a set of quality indicator target values is chosen, eventually
used to measure algorithm runtime to reach each of these targets. The target values are based on a
target precision ∆𝐼 and a reference hypervolume indicator value, 𝐼𝑖ref , which is an approximation
COCO
of the 𝐼HV
indicator value of the Pareto set.
3.1 Target Precision Values
All target indicator values are computed in the form of 𝐼𝑖ref + ∆𝐼 from the instance dependent
reference value 𝐼𝑖ref and a target precision value ∆𝐼. For the bbob-biobj test suite, 58 target
precisions ∆𝐼 are chosen, identical for all problem instances, as
∆𝐼 ∈ {−10−4 , −10−4.2 , . . . , −10−4.8 , −10−5 , 0, 10−5 , 10−4.9 , 10−4.8 , . . . , 10−0.1 , 100 } .
⏞
⏟
six negative target precision values
Negative target precisions are used because the reference indicator value, as defined in the next
section, can be surpassed by an optimization algorithm.3 The runtimes to reach these target values
are presented as empirical cumulative distribution function, ECDF [HAN2016perf]. Runtimes to
reach specific target precisions are presented as well. It is not uncommon however that the quality
indicator value of the algorithm never surpasses some of these target values, which leads to missing
runtime measurements.
3
In comparison, the reference value in the single-objective case is the 𝑓 -value of the known global optimum and,
consequently, the target precision values have been strictly positive [HAN2016perf].
6
3.2 The Reference Hypervolume Indicator Value
Unlike the single-objective bbob test suite [HAN2009fun], the biobjective bbob-biobj test
suite does not provide analytic expressions of its optima. Except for 𝑓1 , the Pareto set and the
Pareto front are unknown.
Instead of the unknown hypervolume of the true Pareto set, we use the hypervolume of an approximation of the Pareto set as reference hypervolume indicator value 𝐼𝑖ref .4 To obtain the approximation, several multi-objective optimization algorithms have been run and all non-dominated
solutions over all runs have been recorded.5 The hypervolume indicator value of the obtained
set of non-dominated solutions, also called non-dominated reference set, separately obtained for
each problem instance in the benchmark suite, is then used as the reference hypervolume indicator
value.
4 Instances and Generalization Experiment
The standard procedure for an experiment on a benchmark suite, like the bbob-biobj suite,
prescribes to run the algorithm of choice once on each problem of the suite [HAN2016ex]. For the
bbob-biobj suite, the postprocessing part of COCO displays currently by default only 5 out of
the 10 instances from each function-dimension pair.
5 Data Storage and Future Recalculations of Indicator
Values
Having a good approximation of the Pareto set/Pareto front is crucial in assessing algorithm performance with the above suggested performance criterion. In order to allow the reference sets to
approximate the Pareto set/Pareto front better and better over time, the COCO platform records
every non-dominated solution over the algorithm run. Algorithm data sets, submitted through the
COCO platform’s web page, can therefore be used to improve the quality of the reference set by
adding all solutions to the reference set which are currently non-dominated to it.
Recording every new non-dominated solution within every algorithm run also allows to recover the
algorithm runs after the experiment and to recalculate the corresponding hypervolume difference
values if the reference set changes in the future. In order to be able to distinguish between different
collections of reference sets that might have been used during the actual benchmarking experiment
and the production of the graphical output, COCO writes the absolute hypervolume reference
4
Using the quality indicator value of the true Pareto set might not be desirable, because the set contains an infinite
number of solutions, which is neither a possible nor a desirable goal to aspire in practice.
5
Amongst others, we run versions of NSGA-II [DEB2002] via Matlab’s gamultiobj function, SMS-EMOA
[BEU2007], MOEA/D [ZHA2007], RM-MEDA [ZHA2008], and MO-CMA-ES [VOS2010], together with simple
uniform RANDOMSEARCH and the single-objective CMA-ES [HAN2001] on scalarized problems (i.e. weighted
sum) to create first approximations of the bi-objective problems’ Pareto sets.
7
values together with the performance data during the benchmarking experiment and displays a
version number in the plots generated that allows to retrieve the used reference values from the
Github repository of COCO.
Acknowledgements
This work was supported by the grant ANR-12-MONU-0009 (NumBBO) of the French National
Research Agency.
The authors would like to thank Thanh-Do Tran for his contributions and assistance with the preliminary code of the bi-objective setting and for providing us with his extensive experimental data.
We also thank Tobias Glasmachers, Oswin Krause, and Ilya Loshchilov for their bug reports, feature requests, code testing, and many valuable discussions. Special thanks go to Olaf Mersmann
for the inital rewriting of the COCO platform without which the bi-objective extension of COCO
would not have happened.
References
[BEU2007]
N. Beume, B. Naujoks, and M. Emmerich (2007). SMS-EMOA: Multiobjective
selection based on dominated hypervolume. European Journal of Operational
Research, 181(3), pp. 1653-1669.
[DEB2002]
K. Deb, A. Pratap, S. Agarwal, and T. A. M. T. Meyarivan (2002). A fast and
elitist multiobjective genetic algorithm: NSGA-II. IEEE Transactions on Evolutionary Computation, 6(2), pp. 182-197.
[HAN2001]
N. Hansen and A. Ostermeier (2001). Completely derandomized self-adaptation
in evolution strategies. Evolutionary computation, 9(2), pp. 159-195.
[HAN2016perf] N. Hansen, A. Auger, D. Brockhoff, D. Tušar, T. Tušar (2016). COCO: Performance Assessment, ArXiv e-prints, arXiv:160x.xxxxx.
[HAN2016co]
N. Hansen, A. Auger, O. Mersmann, T. Tušar, D. Brockhoff (2016). COCO: A
Platform for Comparing Continuous Optimizers in a Black-Box Setting, ArXiv
e-prints, arXiv:1603.08785.
[HAN2009fun] N. Hansen, S. Finck, R. Ros, and A. Auger (2009). Real-parameter black-box optimization benchmarking 2009: Noiseless functions definitions. Technical Report
RR-6829, Inria, updated February 2010.
[HAN2016ex]
N. Hansen, T. Tušar, A. Auger, D. Brockhoff, O. Mersmann (2016). COCO: The
Experimental Procedure, ArXiv e-prints, arXiv:1603.08776.
8
[TUS2016]
T. Tušar, D. Brockhoff, N. Hansen, A. Auger (2016). COCO: The Bi-objective
Black Box Optimization Benchmarking (bbob-biobj) Test Suite, ArXiv e-prints,
arXiv:1604.00359.
[VOS2010]
T. Voß, N. Hansen, and C. Igel (2010). Improved step size adaptation for the
MO-CMA-ES. In Genetic and Evolutionary Computation Conference (GECCO
2010), pp. 487-494. ACM.
[ZHA2007]
Q. Zhang, and H. Li (2007). MOEA/D: A multiobjective evolutionary algorithm based on decomposition. IEEE Transactions on Evolutionary Computation,
11(6), pp. 712-731.
[ZHA2008]
Q. Zhang, A. Zhou, and Y. Jin (2008). RM-MEDA: A regularity model-based
multiobjective estimation of distribution algorithm. IEEE Transactions on Evolutionary Computation, 12(1), pp. 41-63.
[ZIT2003]
E. Zitzler, L. Thiele, M. Laumanns, C. M. Fonseca, and V. Grunert da Fonseca
(2003). Performance Assessment of Multiobjective Optimizers: An Analysis and
Review. IEEE Transactions on Evolutionary Computation, 7(2), pp. 117-132.
9
| 9cs.NE
|
SKELETA OF G-PARKING FUNCTION IDEALS
arXiv:1708.04712v2 [math.CO] 12 Sep 2017
ANTON DOCHTERMANN
Abstract. Given a graph G, the G-parking function ideal MG is an artinian monomial
ideal in the polynomial ring S with the property that a linear basis for S/MG is provided by
the set of G-parking functions. It follows that the dimension of S/MG is given by the number
of spanning trees of G, which by the Matrix Tree Theorem is equal to the determinant of the
reduced Laplacian of G. The ideals MG and related algebras were introduced by Postnikov
and Shapiro [23] where they studied their Hilbert functions and homological properties. In
[14] it was shown that a minimal resolution of MG can be constructed from the graphical
hyperplane arrangement associated to G, providing a combinatorial interpretation for the
Betti numbers.
Motivated by constructions in the theory of chip-firing on graphs, we study certain ‘skele(k)
ton’ ideals MG ⊂ MG generated by subsets of vertices of G of size at most k + 1. We study
(k)
monomial bases of MG and provide formulas and combinatorial interpretations for the di(1)
(n−2)
mensions of S/MG and S/MG
for the case that G = Kn+1 is the complete graph. These
monomial bases have connections to various combinatorial objects including Cayley trees
and determinants of the signless Laplacians, and in some cases lead to new enumerative
(1)
formulas. We furthermore study resolutions of MG and show that for certain graphs G a
minimal resolution is supported on the decomposition of Euclidean space coming from the
theory of tropical hyperplane arrangements. This leads to combinatorial interpretations of
the Betti numbers of these ideals.
1. Introduction
Let G = (V, E) be a simple undirected graph with vertex set V = {0, 1, . . . , n}, with
distinguished sink vertex 0. In this paper we study various algebraic objects associated
with combinatorial properties of G. Our point of departure will be the well-known formula
of Cayley that says that the number of spanning trees of the complete graph is equal to
(n+1)n−1 . This number is related to various other seemingly disparate combinatorial objects,
and in particular counts the number of parking functions of length n. This correspondence
can be generalized to the case of arbitrary graphs G in the context of sequence of integers
known as G-parking functions.
On the algebraic side this data can be encoded in the G-parking function ideal MG , a
monomial ideal living in the polynomial ring S = K[x1 , x2 , . . . , xn ]. By construction the
standard monomials of MG correspond to G-parking functions (this explains the name), and
hence are in bijection with the number of spanning trees of G. The standard monomials form
a monomial basis for the algebra S/MG and, in the case of G = Kn+1 the complete graph,
Cayley’s formula says that S/MG has dimension (n + 1)n−1 . A generating set of monomials
for MG is indexed by all nonempty subsets of [n] = {1, 2, . . . , n}.
MG = hmσ : ∅ =
6 σ ⊆ [n]i
Date: September 14, 2017.
1
2
ANTON DOCHTERMANN
We refer to the next section for the definition of mσ . For the case of the complete graph
Kn+1 this set gives a minimal set of generators, but for all other graphs G this generating
set is redundant.
The ideal MG was introduced by Postnikov and Shapiro in [23], where they study various
properties of AG = S/MG . In particular they seek to compare AG to a related algebra BG ,
defined as the quotient of S by certain powers of linear forms. In [23] it is shown that AG and
BG have the same dimension. The motivation in [23] came from certain algebras generated
by curvature forms on generalized flag manifolds, in an attempt to lift Schubert calculus to
the level of differential forms.
In [23] the authors study homological properties of MG and show that a cellular resolution
of MG is supported on B(∆n−1 ), the barycentric subdivision of the (n − 1)-dimensional
simplex, which is minimal only for the case of the complete graph G = Kn+1 . The ideals
MG are examples of more general monotone monomial ideals which have cellular resolutions
supported on the geometric realization of certain posets. In [14] it is shown that a minimal
resolution of MG is supported on a certain polyhedral complex obtained from the graphical
hyperplane arrangement associated with G.
In this paper we study generalizations of MG and similarly relate their algebraic properties
to well-known combinatorial structures. Our motivation comes from the theory of chip-firing
on G. In this context one studies the dynamical system associated with distributing chips on
the vertices of a G according to edge adjacency information. The dynamics of chip-firing is
governed by the Laplacian matrix LG of the graph, and it can be shown that the G-parking
function ideal MG is in fact a certain initial ideal of the (binomial) lattice ideal determined
by LG .
Recently Backman and Paoletti and others ([1], [6]) introduced the notion of ‘hereditary
set’ chip-firing, where certain subsets of vertices are allowed to fire simultaneously. This
rather general notion of chip-firing interpolates between the ‘abelian sandpile’ model (where
only singletons fire) and the ‘cluster model’ (where any subset is allowed to fire). One
can see that the ‘stable’ configurations in the cluster model are precisely the G-parking
functions. Hereditary set chip-firing shares desirable properties with the more traditional
models, including stabilization that is independent of the chosen firings and a well-defined
notion of a recurrent configuration. In [1] an explicit bijection between these recurrent
configurations and the set of spanning trees is described.
Motivated by these constructions, we study subideals of MG described by k-dimensional
‘skeleta’. Recall that MG has a (possibly redundant) set of generators mσ for every nonempty
(k)
subset σ ⊆ [n] = {1, 2, . . . , n}. For each integer k, define the ideal MG to be the sub ideal
of MG generated by elements corresponding to subsets of cardinality at most k + 1:
(k)
MG = hmσ : ∅ =
6 σ ⊆ [n], |σ| ≤ k + 1i.
For k = 0 the ideal is generated by powers of the variables (corresponding to the degree of
the corresponding vertex), and for k = n − 1 we recover the ideal MG . We will mostly focus
on the case k = 1 and k = n − 2, the one-dimensional and codimension one skeleta of MG .
It turns out for these cases there is a story of free resolutions and monomial bases that runs
somewhat parallel to that of MG described above.
Recall that the standard monomials of MG are by construction given by the G-parking
functions (and hence in bijection with the spanning trees of G). By the matrix-tree theorem
SKELETA OF G-PARKING FUNCTION IDEALS
3
this number is given by the determinant of L̃G , the reduced Laplacian matrix of G. A
(k)
formula for the number of standard monomials of MG is not so clear even in the case that
G = Kn+1 is the complete graph. In the cases that we can find such a formula, we see some
intriguing connections to other combinatorial objects. Our main results along these lines are
summarized by the following, we refer to subsequent sections for any undefined terms.
(k)
(k)
Theorem (see Theorems 3.2 and 3.8 and Corollary 3.4). Let Mn := MKn+1 denote the
k-skeleton ideal of the complete graph Kn+1 . Then we have the following.
(1)
(1)
• The number of standard monomials of Mn (and hence dimK S/Mn ) is given by
(2n − 1)(n − 1)n−1 .
(1)
This number is given by det(Q̃Kn+1 ), the determinant of the reduced signless Laplacian
of the complete graph Kn+1 .
(n−2)
(n−2)
• The number of standard monomials of Mn
(and hence dimK S/Mn
by
(n + 1)n−1 + (n − 1)n−1 .
) is given
(2)
(k)
We have not found a closed formula for the dimension of S/Mn for other skeleta. We
remark that the value det(Q̃G ) appearing in (1) counts (in a weighted fashion) certain combinatorial substructures of the graph G, in the same way that the usual reduced Laplacian
det(L̃G ) counts spanning trees (see Section 3 for details). We are unable to find an explicit
(1)
bijection between the standard monomials of Mn and these substructures. For arbitrary
(1)
graphs G we illustrate that det(Q̃G ) does not count standard monomials of MG (see Example 3.6), but some computer experimentation suggests that an inequality may hold (see
Question 3.7).
The value (n − 1)(n−1) appearing in (2) also has combinatorial interpretations, for example
involving Cayley trees in which the root is smaller than its immediate descendant [7], but
again we do not know of a bijection between these objects and the standard monomials of
(n−2)
(n−2)
Mn
. An explicit formula for the standard monomials of Mn
coming directly from
n−1
the definition provides ‘another way of counting’ (n − 1)
in the spirit of [5]. We refer to
Corollary 3.9 for details.
(k)
We next turn to homological properties of MG , and in particular a combinatorial description of minimal free resolutions. Recall in [23] it was shown that a cellular resolution
for MG is supported on the barycentric subdivision of a simplex (with a minimal cellular
(k)
resolution described in [14]). To build resolutions of the skeleton ideals MG , we will search
for polyhedral decompositions of the simplex that agree with B(∆n−1 ) on the appropriate
subcomplex induced by subsets of cardinality at most k + 1. For the case of k = 1 there is
a natural candidate coming from the the theory of tropical convexity. Our next main result
is the following.
(1)
Theorem (see Theorem 4.6). The ideal MKn+1 has a minimal cocellular resolution supported on the labeled polyhedral complex induced by any generic arrangement of two tropical
hyperplanes in Rn−1 .
As was spelled out in [12], arrangements of tropical hyperplanes are intimately tied to
triangulations of a product of simplices, which in turn are related to mixed subdivisions
4
ANTON DOCHTERMANN
and have many pleasing combinatorial properties [11]. In particular by taking the ‘staircase
(1)
triangulation’ we obtain a combinatorial interpretation for the Betti numbers of Mn =
(1)
MKn+1 in terms of certain subgraphs of the complete bipartite graph Kn,2 . In addition, we
are able to obtain a closed formula for the Betti numbers of these ideals.
(1)
Corollary (See Corollary 4.7). The total Betti numbers βin of the ideal MKn+1 are given by
n
X
j−1
n
βi =
j
.
i−1
j=1
(1)
The complex supporting a resolution of MKn+1 does not support a resolution of MG
for arbitrary G. However, we are able to utilize certain degenerations of the underlying
arrangement to obtain a minimal resolution for a certain class of graphs.
Proposition (See Proposition 4.11). Suppose G is a graph of the form G = H ∗ {v},
(1)
where H is obtained by removing disjoint cliques from the complete graph Kn . Then MG
has a minimal resolution supported on the subdivision of Rn−1 induced by a (non-generic)
arrangement of two tropical hyperplanes.
The rest of the paper is organized as follows. In Section 2 we review the basic definitions of
the G-parking function ideals, their skeleta and related constructions, and also recall some
basic facts from combinatorial commutative algebra. In Section 3 we study the standard
(n−2)
(1)
monomials of the skeleta ideals, and provide explicit formulas for the case of MG and MG
for the complete graph. We relate these counts to well-known combinatorial structures
including signless Laplacians and Cayley’s formula. In Section 4 we study homological
properties of the skeleta ideal, and in particular establish our results regarding cocellular
(1)
resolutions of MG for certain G. We end with some concluding remarks and open questions.
Acknowledgements. We thank Spencer Backman, Sam Hopkins, and Suho Oh for useful
conversations. The results in the this paper grew out of discussions initiated at a workshop
on ‘Generalizations of chip-firing and the critical group’ held July 8-12, 2013 at the American
Institute of Mathematics. We thank the organizers for the opportunity to participate and
AIM for providing the working environment.
2. Definitions and objects of study
2.1. G-parking functions, ideals, and their skeleta. We begin with some basic facts
regarding the combinatorial objects involved in our study. Recall that a parking function
of size n is a sequence (a1 , a2 , . . . , an ) of nonnegative integers such that its rearrangement
c1 ≤ c2 ≤ · · · ≤ cn satisfies ci < i. This seemingly innocent construction turns out to have
connections and applications to many areas of mathematics. For instance, in [20] it is shown
that the parking functions of size n are in bijection with the number of trees on n + 1 labeled
vertices which, by Cayley’s formula, is given by (n + 1)n−1 .
One can extend these constructions to arbitrary graphs (with the classical case recovered
by the complete graph Kn+1 ). For this suppose that G is an (undirected) graph on vertex
set {0, 1, . . . , n} with distinguished sink vertex 0 (much of the theory can be extended to
directed graphs with multiple edges, but here we will focus on the simple undirected case).
SKELETA OF G-PARKING FUNCTION IDEALS
5
Associated to the graph G is the reduced Laplacian matrx LG , an n × n symmetric matrix
with entries given by
(L̃G )i,j
(
deg(vi ) if i = j
=
−1
if i 6= j
The matrix-tree theorem says that the number NG of spanning trees of G is given by
NG = det L̃G . Our definition of L̃ corresponds to deleting from the usual Laplacian LG the
row and column corresponding to the vertex 0, but is well known that NG is independent of
which vertex is chosen. For a subset σ ⊆ [n] = {1, 2, . . . , n} and a vertex i ∈ [n], let
dσ (i) = |{j : j ∼ i, j ∈
/ σ}|
(3)
Definition 2.1. A sequence (b1 , b2 , . . . , bn ) is said to be a G-parking function if for any
∅=
6 σ ⊆ [n] there exists i ∈ σ such that bi < dS (i).
Note that if G = Kn+1 a complete graph then this recovers the classical parking functions
of size n defined above. In addition we have the following enumerative property.
Theorem 2.2. [17] The number of G-parking functions is given by det L̃G , the number of
spanning trees of G.
In [23] the authors formulate G-parking functions in an algebraic context. As above we
assume that G is an (undirected) graph on vertex set {0, 1, . . . , n} with distinguished sink
vertex 0. Fix a field K and let S = K[x1 , x2 , . . . , xn ] denote the polynomial ring on n
variables. For any ∅ =
6 σ ⊆ [n] define the monomial mσ by
Y outdeg (i)
σ
,
(4)
mσ =
xi
i∈σ
where outdegσ (i) = #{j ∈ [n]\σ : j ∼ i} is the number of vertices outside the set σ that are
adjacent to i.
Definition 2.3. For a graph G on vertex set {0, 1, . . . , n} the G-parking function ideal
MG ⊂ S is generated by all mσ for nonempty σ:
MG = hmσ : ∅ =
6 σ ⊆ [n]i.
The ideal MG was introduced by Postnikov and Shapiro in [23], where they studied monomial bases for the quotient algebra S/MG and also described cellular resolutions of these and
related ideals. The ideals MG also have connections to ‘chip-firing’ and a discrete RiemannRoch theory of graphs. In fact it can be shown that MG is a certain initial ideal of the
so-called toppling ideal IG , the binomial lattice ideal parametrizing linear equivalence classes
of effective divisors on G.
In the context of chip-firing one studies the following dynamical system: a number of ‘chips’
are placed on the vertices of a graph G, and a vertex v is allowed to ‘fire’ if the number
of chips is at least deg(v) (the degree of that vertex), in which case one chip is passed to
each of its neighbors and v loses deg(v) chips. In the Abelian sandpile model vertices are
restricted to fire individually, whereas in the ‘cluster firing model’ any collection of non-sink
vertices is allowed to fire simultaneously. In both models the stabilization of a configuration
is independent of the sequence of firings, and the number of ‘recurrent’ configurations is the
6
ANTON DOCHTERMANN
same (given by the number of spanning trees). In the case of the cluster firing method, the
recurrent configurations are the same as the stable configurations.
Backman and Paoletti and others ([1], [6]) generalized this situation by considering ‘hereditary chip-firing’ models, where the prescribed sets of vertices form any simplicial complex
on the vertex set of G. In [1] it is shown that these models enjoy similar properties, and for
example by adapting the Cori-Le Borgne algorithm from [10] an explicit bijection between
the recurrent configurations of a hereditary chip-firing model on a graph and the set of its
spanning trees is given.
In this paper we consider the algebraic aspects of hereditary chip-firing in the context of
the monomial ideals defined above. Although a general theory seems difficult to describe,
we do see some pleasing structure arising in the case of subideals defined by ‘skeleta’. The
following will be our main objects of study.
Definition 2.4. Suppose G is a graph on vertex set {0, 1, . . . , n}. For an integer k with
(k)
1 ≤ k ≤ n − 1, we define the ideal MG according to
(k)
MG = hmσ : ∅ =
6 σ ⊆ [n], |σ| ≤ k + 1i.
(k)
One can think of the MG as certain ‘k-skeleta’ of the ideal MG . For example for G = K5
and k = 2 we have
(2)
MG = hx41 , x42 , x43 , x44 , x31 x32 , x31 x33 , x31 x34 , x32 x34 , x32 x34 , x33 x34 i,
generated by all monomials mσ with 1 ≤ |σ| ≤ 2.
(k)
(k)
Notation. To simplify notation we will often use Mn := MKn+1 to denote the k-skeleton
ideal of the complete graph G = Kn+1 .
(k)
We note that the ideals MG are non-generic: for example the monomials x31 x32 and x31 x33
(2)
are both generators of the ideal MK5 with the same positive degree in x1 , and yet the only
other generator dividing their least common multiple x31 x32 x33 is the monomial x32 x33 . Hence
certain techniques for studying monomial ideals (as were utilized in [23]) break down. In
addition, the Scarf complex associated to these ideals will not support a resolution. For the
(k)
case of a complete graph, the Scarf complex of Mn will consist of the (k − 1)-skeleton of
the barycentric subdivision of the (n − 1)-dimensional simplex, and for k ≤ n − 1 will carry
a nontrivial top homology group.
These ideals are also not monotone in the sense of [23] (despite the apparent similarity in
the names/construction). In that context, for any generators mσ and mτ one requires that
(k)
lcm(mσ , mτ ) is divisible by a generator mρ for some ρ ⊇ σ ∪ τ . Our construction of MG
provides a condition on subsets of σ for each mσ , and in this sense runs somewhat dual to
the theory of monotone ideals.
2.2. Resolutions and monomial bases. We briefly recall some of the commutative algebra needed for our constructions and results. Further detail and any undefined terms can be
found for example in [22].
Recall that if M is any finely graded S-module, then a (Zn -graded) free resolution of M
is a sequence of free S-modules
0 ← M ← F1 ← F2 ← · · · ← Fd ← 0,
SKELETA OF G-PARKING FUNCTION IDEALS
7
L
where each Fi = m∈Zn S(−m)βi,m . The sum is over all monomials m in S, which are often
(as in this case) described in terms of their exponent vector in Zn . The resolution is said to
be minimal if each of the βi,m is minimum over all free resolutions of M , in which case the
βi,m are called the (Zn -, or finely-)graded Betti numbers of M .
A resolution of a monomial ideal M said to be cellular (resp. cocellular) if there exists a
CW -complex X with monomial labelings mF on its faces F ∈ X such that the chain complex
computing cellular homology (resp. cohomology) ‘supports’ the resolution. More specifically,
suppose X is a CW-complex with a monomial associated to each face {mF : F ∈ X }
satisfying
mF = max{mG : for G ⊆ F a face},
so that every face contained in F has a monomial label that divides the label of F . We
define a complex C∗ (X ) of free S-modules according to
M
Ci =
S(−mF )
F ∈X , dim F =i+1
where the sum is over all (i + 1)-dimensional faces F of X . The differentials ∂i : Ci → Ci−1
of C∗ (X ) are monomial-valued matrices with scalar coefficients determined by the entries in
the maps of the chain complex of K-modules computing homology of X , and with monomial
entries that make the differential a Zn -graded map. A cocellular resolution is defined similarly
using the cochain complex C ∗ (X ) of the underlying space. Note that in the case of a cellular
resolution the monomial ideal is generated by the monomials on the vertices (0-cells), whereas
in the case of a cocellular complex is generated by the facets (maximal faces).
Given a monomial labeled complex X there is a useful criteria for checking where it
supports a resolution. For a monomial m, let X≤m denote the subcomplex of X consisting
of those faces F whose label mF divides m:
X≤m = {F ∈ X : mF divides m}.
From [4] we then have the following observation.
Lemma 2.5. The complex C∗ (X ) of S-modules is exact if and only if the induced complex
X≤m is K-acyclic for all monomials m. In this case C∗ (X ) supports a cellular resolution of
the ideal
I = hmv : v ∈ X a vertexi,
generated by the monomials on the vertices of X . If in addition we have that mF 6= mG for
any proper containment F ( G of faces then C∗ (X ) is a minimal free resolution of I.
If the labeling on X satisfies
mF = max{mG : for G ⊃ F a face}
then X is said to be colabeled and gives rise to a cocellular resolution via the cellular cochain
complex of X . Note that for any monomial m, the set X≤m of cells H with mH dividing m
is not a subcomplex of X , but rather a union of relatively open stars of cells. The criteria
for checking acyclicity of the complex is similar.
Lemma 2.6. The colabeled complex C ∗ (X ) is exact if and only if X≤m is K-acyclic for every
monomial m ∈ Zn (identifying m with its exponent vector). In this case C ∗ (X ) provides a
resolution of S/I, where
I = hmH : H ∈ X is a maximal cell i.
8
ANTON DOCHTERMANN
The resolution is minimal if mF 6= mG for any proper containment F ( G of faces.
Suppose I ∈ S is an ideal such that S/I is finite dimensional as a K-vector space. Then
we say that I is an Artinian ideal and a basis for S/I is called the set of standard monomials
of I (determined by a chosen term order). In the case that I is monomial, one can see that
the standard monomials consist of those monomials m such that m is not divisible by any
generators of I. In [23] it is shown that for the G-parking function ideals I = MG , the
dimension of S/MG is equal to the number of spanning trees of G. More recently [8] the
authors provide a family of combinatorial bijections between the G-parking functions and
the spanning trees of G.
3. Monomial bases of skeleta
Recall that the standard monomials of MG (a basis for the vector space S/MG ) are given
by the G-parking functions (and hence counted by spanning trees). For the case of G = Kn+1
the number of standard monomials is thus given by (n−1)n+1 . In this section we consider the
standard monomials of skeleton ideals, seeking an analogous combinatorial interpretation.
(k)
Note that in general since MG ⊂ MG we have that the usual G-parking functions will be
among the standard monomials for the k-skeleton ideals for all k.
(0)
For any graph G the zero-skeleton ideal MG is generated by each variable raised to the
(0)
power of the degree of the corresponding vertex. Hence a monomial basis for S/MG is easy
to describe as the set of monomials lying within an n-orthotope with edge lengths given by
degree:
o
nY
di
xi : 0 ≤ di < deg(i)
i∈[n]
We next turn to the other skeleta. For the remainder of this section we will primarily
(k)
(k)
focus on the case that G = Kn+1 is a complete graph. Recall that we use Mn := MKn+1 to
denote the k-skeleton ideal of the complete graph Kn+1 .
(2)
Example 3.1. For n = 3 we have M3
standard monomials given by
= hx31 , x32 , x33 , x21 x22 , x21 x23 , x22 x23 i, with the set of
{G-parking functions of length 3} ∪ {x1 x2 x3 , x21 x2 x3 , x1 x22 x3 , x1 x2 x23 },
giving a total of 16 + 4 = 20 standard monomials. Here as usual we identify a G-parking
function with the monomial having that sequence as an exponent vector, so that for example
(1, 0, 2) is identified with the monomial x1 x23 .
In general we have the following formula.
(1)
Theorem 3.2. The number of standard monomials of Mn (and hence the dimension of the
(1)
K-vector space S/Mn ) is given by
dimK (S/Mn(1) ) = (2n − 1)(n − 1)n−1 .
(2)
Proof. Suppose (a1 , a2 , . . . , an ) is the exponent vector of a standard monomial of Mn . Then
by definition each entry ai is strictly less than n, and at most one entry equals n − 1.
SKELETA OF G-PARKING FUNCTION IDEALS
9
If no entry equals n − 1 then every entry satisfies 0 ≤ ai ≤ n − 2 so we have (n − 1)n
possibilities. If exactly one entry (say ai ) equals n − 1 then every other entry satisfies
0 ≤ aj ≤ n − 2 for j 6= i. Hence we have n(n − 1)n−1 possibilities. Adding these up we get
(n − 1)n + n(n − 1)n−1 = (2n − 1)(n − 1)n−1
standard monomials, as desired.
It turns out this number has a determinantal interpretation analogous to the case of Gparking functions. Once again recall that the dimension of S/MKG (and the number of
G-parking functions) is equal to det L̃G , the determinant of the reduced Laplacian of G. For
the case of G = Kn+1 this number is provided by the well known formula (n + 1)n−1 . For
the one-skeleton a different but related matrix makes an appearance.
Definition 3.3. For a graph G on vertex set {0, 1, . . . , n} the signless Laplacian Q = QG is
the symmetric (n + 1) × (n + 1) matrix with rows and columns indexed by the vertices of G
and with entries given by
(
deg(i)
if i = j,
Qi,j =
|{edges connecting i and j}| if i 6= j.
Define the reduced signless Laplacian Q̃ to be the matrix obtained from Q by deleting the row
and column corresponding to the vertex 0.
Note that Q has entries given by the absolute values of the entries of the usual Laplacian
L (hence the name). For example, if G = K4 we get the following matrices.
3
1
Q=
1
1
1
3
1
1
1
1
3
1
1
1
1
3
3 1 1
Q̃ = 1 3 1
1 1 3
In this case one has det Q̃ = 20 = (5)(22 ). This should remind the reader of the count from
Example 3.1 above, and in general we have the following observation.
(2)
Corollary 3.4. The number of standard monomials of Mn
(2)
S/Mn ) is given by
dimK (S/Mn(2) ) = det Q̃Kn+1 .
(and hence the dimension of
Proof. According to Theorem 3.2 it is enough to show that det Q̃Kn+1 = (2n − 1)(n − 1)n−1 .
For this we examine the eigenvalues of the matrix Q̃Kn+1 . We have one eigenvalue 2n − 1
with multiplicity 1 corresponding to the all 1’s vector 1. Subtracting the matrix (n − 1)In
from Q̃Kn+1 gives us the matrix J consisting of all ones, which has an (n − 1)-dimensional
kernel. Hence Q̃Kn+1 has one other eigenvalue n − 1 with multiplicity n − 1. The result
follows.
It would be interesting to find a bijective proof of Corollary 3.4, along the lines of wellknown bijections between spanning trees and parking functions that are scattered throughout
the literature (see for instance [8]). For this we need a graph-theoretical interpretation of
the determinant of Q̃, which we have (again) thanks to the Cauchy-Binnet theorem.
10
ANTON DOCHTERMANN
0
0
1
4
3
1
4
2
3
2
Figure 1. The graphs H 0 and H from Example 3.6. The graphs are isomorphic but have different values of det Q̃ (in both cases the sink is given by the
vertex 0).
Proposition 3.5. [8] For any graph G the determinant of Q̃G = Q̃ is given by
X
det Q̃ =
4c(H) ,
H
where the summation runs over all spanning T U -subgraphs H of G with c(H) unicyclic
components, and one tree component which contains the vertex 0.
Here a T U -subgraph is a subgraph of G whose components are trees or unicylic graphs
with odd cycles. Analogous to the usual Laplacian, Q can be obtained as Q = M T M , where
M is the 0 − 1 vertex-edge signless incidence matrix of G. Applying Cauchy-Binet to this
factorization gives the result (see [3] for details).
A bijective proof of Corollary 3.4 would therefore associate to each spanning T U -subgraph
(1)
H ⊂ Kn+1 a collection of 4c(H) standard monomials of Mn . Note that each spanning
tree of G would be assigned 40 = 1 standard monomials, so presumably such a bijection
would extend the correspondence between usual G-parking functions and spanning trees. In
Example 3.1 we have the 16 parking functions coming from the spanning trees of G, and 4
new standard monomials coming from the T U -subgraph consisting of the edges {12, 13, 23}.
For graphs that lack symmetry we note that det(Q̃) depends on the choice of sink vertex.
This is in contrast to the usual Laplacian, where the determinant simply counts the number
of spanning trees containing the sink and in particular is independent of which row/column
is deleted to obtain L̃. In fact Corollary 3.4 does not hold for general G, as the following
example illustrates.
Example 3.6. Let H be the graph obtained from
The reduced signless Laplacian is
4 1
1 4
Q̃H =
1 1
1 1
removing the edge (34) from the graph K5 .
1
1
3
0
1
1
0
3
(1)
with det(Q̃H ) = 99. According to Macaulay2 [18] we have dimK MG = 105, so that there
(1)
are 105 standard monomials of MG in this case. Note however if H 0 is the graph obtained
from K5 by removing the edge 01 then we get det(Q̃H 0 ) = 135, while there are 135 standard
(1)
monomials of MH .
After a number of calculations we have not found an example where the determinant of
(1)
Q̃ is larger than the dimension of S/MG , which begs the following question.
SKELETA OF G-PARKING FUNCTION IDEALS
11
Question 3.7. For any graph G is it true that
(1)
dimK S/MG ≥ det(Q̃G )?
The hope here is that for each T U subgraph H ⊂ G we can find a way to assign 4c(H)
(1)
standard monomials of MG (in an injective fashion).
3.1. Codimension one. Having obtained a formula for the number of standard monomials
(1)
(n−2)
of the one-skeleton S/Mn we next turn to S/Mn
, the ‘codimension one’ skeleton ideal.
(n−2)
Recall that MG
is generated by all monomials mσ where σ ⊆ [n] and 1 ≤ |σ| ≤ n − 1. In
(n−2)
particular MG
has the same generators as MG except for the single monomial m{1,2,...,n}
corresponding to the entire set. Again we restrict ourselves to the complete graph and let
(n−2)
Mn(n−2) = MKn+1
denote the codimension-one skeleton ideal of the complete graph. There is a surprisingly
(n−2)
simple closed formula for the vector space dimension of S/Mn
that again involves familiar
combinatorial objects.
(n−2)
Theorem 3.8. The number of standard monomials of the ideal Mn
(n−2)
mension of the K-vector space S/Mn
) is given by
(and hence the di-
dimK S/Mn(n−2) = (n + 1)n−1 + (n − 1)n−1 .
(n−2)
Proof. Again since Mn
⊂ MKn+1 we have that the set of standard monomials of MKn+1
(n−2)
are contained in those of Mn
. The new contributions come from monomials of the form
xa11 xa22 · · · xann , where ai ≥ 1 for all i.
Let M+
n denote the set of such monomials, so that if Mn denotes the set of standard
(n−2)
monomials of Mn = MKn+1 then the set of standard monomials of Mn
is given by Mn ∪
n−1
+
and hence it suffices to
Mn . From the above discussion we know that |Mn | = (n + 1)
n−1
|
=
(n
−
1)
.
show that |M+
n
Note that an element of M+
n is obtained by taking the monomial x1 x2 · · · xn , choosing a
set S ⊆ [n], and then raising each variable corresponding to a element of S by a nonzero
amount. Let F (n, s) denote the number of standard monomials in M+
n+1 that correspond to
raising the exponent of all variables in a chosen (and fixed) set of n − s variables, fixing the
exponents of all other variables to be one. We claim that F (n, s) satisfies the recursion
n−s
X
n−s
F (n, s) =
F (n − 1, s + j − 1),
(5)
j
j=0
with F (1, 1) = 1 and F (n, 0) = 0 for n ≥ 1.
We have F (1, 1) = 1 since if we have only two variables, we get a single contribution to the
(1)
set M+
that are elements of M2 ) by ‘changing zero variables’,
2 (standard monomials of M2
namely from the monomial x1 x2 . In fact F (n, n) = 1 for all n ≥ 1.
Similarly we have F (n, 0) = 0 for n ≥ 1 since if we have n + 1 variables and raise the
(n−1)
exponent of any n variables we end up with a monomial in Mn+1 . This follows from the
(n−1)
fact that Mn+1 contains the monomial mS for all S ⊆ [n] with |S| ≤ n.
12
ANTON DOCHTERMANN
Next, to establish the recursion, suppose m is a standard monomial in M+
n+1 resulting
from raising the exponents in some set of t = n − s variables. Without loss of generality
suppose that the set is {1, 2, . . . , t} so that
m = xa11 xa22 · · · xat t xt+1 · · · xn+1 ,
with each ai > 1. Suppose j of these variables have exponent equal to 2 (so that t − j
variables have exponent 3 or larger). Take the subset of variables consisting of those with
exponent 3 or larger and subtract one from each of those exponents to get a new monomial
m0 (on that subset of variables, shifting indices if necessary) that is now a standard monomial
(n−2)
of Mn
with the property that t − j = n − s − j variables have exponent (strictly) larger
than 1. The number of such standard monomials is F (n − 1, s + j − 1). If we add up
n−s
F (n − 1, s + j − 1) for all possible choices of j variables with exponent equal to 2 we
j
get F (n, s). The recursion formula (5) follows.
To aid the reader we provide an illustration of the identity for the case
F (4, 2) = F (3, 1) + 2F (3, 2) + F (3, 3).
Here we are considering monomials with 4 + 1 = 5 variables. Assuming as above that the
set of variables with raised coefficients is {x1 , x2 } we have
|{33111, 34111, 43111, 32111, 23111, 42111, 24111, 22111}|
= |{2211, 2311, 3211}| + 2|{2111, 3111}| + |{1111}|,
(6)
where monomials are represented by their exponent vectors.
Now, in [25] it is shown that F (n, s) counts the number of forests with n labeled vertices
consisting of s distinct trees such that s specified vertices belong to distinct trees. There it
is shown that
F (n, s) = snn−s−1 .
In [7] the authors consider labeled rooted trees on vertex set [n + 1] with the property that
the root has no smaller ‘son’ (immediate descendant). They prove that the number of such
trees is nn and show how the tree description leads to certain combinatorial interpretations
of nn . By considering the parameter ‘degree s of the root’, they obtain the identity:
n
X
n+1
n
n =
snn−s−1 .
(7)
s
+
1
s=1
We have seen that snn−s−1 counts the number of monomials in M+
n+1 that correspond to
raising the exponent of all variables in a chosen (and fixed) set
of
n
−
s
variables (while fixing
n+1
n+1
all other variables to have exponent one). There are n−s = s+1 ways to choose which
variables to raise, and hence our result follows.
There is an interesting (and as far as we know, new) enumerative corollary of this result
that follows from a more straightforward way to count the elements of M+
n.
Corollary 3.9. For any integer n ≥ 1 we have
X
n
n − k1
n − (k1 + · · · + kn−3 )
n−1
(n − 1)
=
···
.
k1
k2
kn−2
0≤k ≤1
1
0≤k1 +k2 ≤2
···
0≤k1 +···+kn−2 ≤n−2
(8)
SKELETA OF G-PARKING FUNCTION IDEALS
13
Proof. We prove this by counting the elements of M+
n and applying Theorem 3.8. Note
,
then
no
exponent
can
satisfy
ai ≥ n since xni is a generating
that if xa11 xa22 · · · xann ∈ M+
n
(n−1)
monomial of MG . We can have at most one exponent satisfying ai = n − 1, at most two
exponents satisfying aj = n − 2, and so on. Let k1 denote the number of exponents ai such
that ai = n − 1, k2 the number of exponents aj such that aj = n − 2, etc. Hence the number
of standard monomials of the form xa11 xa22 · · · xann is given by the expression on the right hand
side of (8), which by Theorem 3.8 is equal to (n − 1)n−1 .
The formula is similar to ‘other way to count N N ’ as described in [5]. Curiously, both
formulas involve a summation over the same index set but one can check that the terms in
the summation are not the same (even up to reordering).
Remark 3.10. As discussed in [7] Equation (7) enumerates labeled rooted trees with no
smaller immediate descendants, via the parameter ‘degree s of the root’. In our context this
corresponds to enumerating elements of M+
n+1 in terms of ‘raising the exponent of n − s
variables’. We can alternatively enumerate elements of M+
n+1 according to degree (as a
n
polynomial) and this leads to yet another q-analogue of n that we were unable to find in the
literature.
Example 3.11. As an example of the above discussion, the relevant monomials in M+
4
(described in terms of exponent vectors) are 3211 (all 12 permutations), 2211 (all six permutations), 3111 (all four permutations), 2111 (all four permutations), and 1111. Theorem
7 gives the decomposition
27 = 18 + 8 + 1,
since 18 monomials have two variables raised, 8 have a single variable raised, and one has
no variable raised. If we count using the degree statistic then we get
27 = 1 + 4 + 10 + 12.
We assume that this has an interpretation in terms of trees but have not found any reference.
4. Resolutions of the 1-skeleton
Recall from [23] that a free resolution of the G-parking function ideal MG is supported
on B(∆n ), the barycentric subdivision of the (n − 1)-dimensional simplex. This resolution
is minimal only in the case that G = Kn+1 is the complete graph, in which case the ideal is
generic and the resolution can be seen to coincide with the Scarf complex. For the case of
arbitrary G it was shown in [14] that a minimal free resolution of MG is supported on the
bounded complex of a certain affine slice of the graphical hyperplane arrangement of G. This
provides a convenient description of the resolution itself and also leads to a combinatorial
interpretation of the Betti numbers of MG in terms of Whitney numbers of the intersection
lattice of the graphical arrangement, acyclic orientations of (contractions of) the graph G
with unique sink, etc. We refer to [14] for details.
(k)
In the case of the skeleton ideals MG one can ask whether the Betti numbers have similar
combinatorial meaning, and if a (co)cellular resolution can be easily described. Although in
general this question does not seem to have a concise answer, we do have a way to interpret
(1)
the Betti numbers for a large class of 1-skeleton ideals MG .
14
ANTON DOCHTERMANN
(1)
Recall that the ideal MG is generated by all monomials mσ where 1 ≤ |σ| ≤ 2. Hence
a complex supporting a cocellular resolution must have maximal cells that index nonempty
subsets of [n] of cardinality at most 2 (just as the vertices of B(∆n ) are naturally labeled
by all nonempty subsets of [n]). For our description of such a complex it will be convenient
to use the language of tropical convexity, as developed by Develin and Sturmfels in [12] and
adapted here for our purposes. Most of our techniques will be standard manipulations in
tropical convexity but for the reader unfamiliar with such constructions we will provide a
more or less self-contained treatment in what follows.
Let ∆n denote the (n − 1)-dimensional simplex given by the convex hull of the origin and
the standard basis vectors
∆n = conv{0, e1 , . . . , en }.
Now suppose a = (a1 , a2 , . . . , an ) is a point in Rn , thought of as a linear functional in an
appropriate dual space. For us a tropical hyperplane determined by the functional a, denoted
H(a), is by definition the codimension one skeleton of the outward normal fan of ∆n , with
vertex at a. Note that the complement of the tropical hyperplane consists of n + 1 maximal
cones which we label {1, 2, . . . , n + 1} according to which vertex of ∆n it attains a maximum
value on (with the origin treated as the (n + 1)st vertex).
(1,1)
e2
2
3
0
(3,0)
1
e1
Figure 2. The simplex ∆2 , its normal fan with labeled cones, and the arrangement A = {H(1, 1), H(3, 0)}.
A collection A of tropical hyperplanes then gives rise to a polyhedral decomposition XA
of Rn , whose relatively open cells are determined by which cones of the hyperplanes they lie
in. We take the following notion from [12], amended for our purposes.
Definition 4.1. Suppose (T (a), T (b)) is an (ordered) arrangement of tropical hyperplanes
in Rn , and let x ∈ Rn . Let a∗ = (a1 , . . . , an , an+1 = 0), b∗ = (b1 , . . . , bn , bn+1 = 0), and
x∗ = (x1 , . . . , xn , xn+1 = 0) denote the vectors in Rn+1 obtained by adding a ‘0’ in the last
coordinate. Then the type of x is given by T (x) = (Ta (x), Tb (x)), where
Ta (x) = {i ∈ [n + 1] : xi − ai = max {xj − aj }},
j∈[n+1]
Tb (x) = {i ∈ [n + 1] : xi − bi = max {xj − bj }}.
j∈[n+1]
Alternatively, if we think of each tropical hyperplane in terms of the labeled cones that they
define, then Ta (x) (resp. Tb (x)) is the set of closed cones of H(a) (resp. H(b)) that intersect
the point x. We will often use this description in our arguments.
SKELETA OF G-PARKING FUNCTION IDEALS
15
For example in the arrangement of Figure 2, we have TA (2, 0) = ({1}, {2, 3}). Note that
this notion corresponds to the ‘fine’ types of [15], but we won’t need the distinction here.
In [12] it is shown that the set of x ∈ Rn with a fixed type forms a relatively open convex
(and tropically convex) subset of Rn , the cells of the decomposition XA . If F ∈ XA is a
relatively open cell we will often speak of the type of F , by which we mean the type of any
point x ∈ F .
Given a pair of tropical hyperplanes A = (H(a), H(b)) we label each maximal cell F of
XA with a monomial mF as follows. If x ∈ F is any point in the relative interior of F with
type TA (x) = (Ta , Tb ) we let σ = Ta ∪ Tb and set mF = mσ . This is well-defined since, as we
have seen, every point in the relative interior of F has the same type. For example in Figure
2 the point x = (2, −1) has type ({1}, {3}) and hence the cell containing x is labeled by the
monomial
outdeg(1) outdeg(3)
m{1,3} = x1
x3
.
If G ∈ XA is any other (non maximal) face, we label it with the monomial mG where
mG = lcm{mF : F is a facet with G ⊆ F }.
We refer to Figure 3 for an example.
x23
x12x22
x22x32
x33
x13x22x32
x13x22
x12x32
x13
Figure 3. The monomial labeled complex XA .
We will see that with this monomial labeling the complex XA in fact supports a cocellular resolution of our ideals. To establish this we will need one more notion from tropical
convexity.
Definition 4.2. Given points x, y ∈ Rn the (max-)tropical line segment they determine is
given by the set of points p ∈ Rn of the form
p = max{λ1 + x, µ1 + y},
where λ, µ are arbitary real numbers, and 1 is the vector in Rn consisting of all one’s.
We then have the following results from [15], again adapted to suit our setup.
Proposition 4.3. [15] Let A = {H(a), H(b)} be an arrangement of tropical hyperplanes
in Rn , and let XA denote the induced decomposition of Rn . For every cell G ∈ XA of
codimension ≥ 1 we have
[
TA (G) =
TA (F ),
G⊂F
16
ANTON DOCHTERMANN
where the union of a pair of sets is taken to be the pair of the unions, so that e.g. (Sa , Sb ) ∪
(Ta , Tb ) = (Sa ∪ Ta , Sb ∪ Tb ).
Lemma 4.4. [15] Suppose x, y ∈ Rn , and let z be a point on the tropical line segment between
x and y. Suppose A = {H(a), H(b)} is a arrangement of a pair of tropical hyperplanes in
Rn , and let TA (x) = (Ca , Cb ), TA (y) = (Da , Db ), and TA (z) = (Ea , Eb ) denote the types of
these points with respect to A. Then we have the following inclusions of multisets:
Ca ∩ Da ⊆ Ea ⊆ Ca ∪ Da ,
(9)
Cb ∩ Db ⊆ Eb ⊆ Cb ∪ Db .
Proposition 4.5. [15] Let C and D be distinct cells in XA with types T (C)A = (Ca , Cb ) and
T (C)A = (Da , Db ). If C is contained in the closure of D then as multisets we have
Ca ∪ Cb 6= Da ∪ Db .
Note that cells that are ‘far away’ in the complex XA can violate the inequality in the
previous Proposition. For instance in the arrangement in Figure (2) we have T (1, 1/2) =
({1, 3}, {2}) and T (2, 0) = ({1}, {2, 3}). In this case both of the relevant faces have the
monomial label x21 x22 x23 . With these in statements in place we can now prove our main
results of this section.
Theorem 4.6. Suppose A = {H(a), H(b)} is an arrangement of two tropical hyperplanes
in Rn−1 in generic position, and let XA denote the induced subdivision of Rn−1 with the
(1)
monomial labeling described above. Then the ideal MKn+1 has a minimal cocellular resolution
supported on XA .
(1)
(1)
Proof. Recall that a non-redundant list of generators of the ideal Mn = MKn+1 is indexed
by all subsets σ of [n] = {1, 2, . . . , n} such that 1 ≤ |σ| ≤ 2. Also recall that each maximal
cell F of XA is labeled by the monomial mσ , where σ = Ta ∪ Tb is the union of the sets
occurring as the type of any x ∈ F . Since x does not lie on H(a) ∪ H(b), we have σ = {i, j}
where 1 ≤ i ≤ j ≤ n. Since the arrangement is assumed to be generic we have that every
such (possibly degenerate) pair is realized as the union of type sets for some point x ∈ Rn .
This is equivalent to the fact that in any regular fine mixed subdivision of ∆n−1 + ∆n−1
every lattice point occurs as a 0-cell, and these lattice points are of the form ei + ej for
all 1 ≤ i ≤ j ≤ n. Furthermore, the set of points that have a particular σ label forms a
maximal connected open cell in Rn . We conclude that the ideal generated by the monomials
(1)
corresponding to maximal cells of XA is indeed the ideal Mn .
We next apply Lemma 2.6 to show that the complex of S-modules induced by the cochain
complex of XA with this monomial labeling is indeed K-acyclic. For this suppose m ∈ Zn is
any monomial (identified with its exponent vector) and consider the complex (XA )≤m . We
follow the argument in [15] to show that (XA )≤m is tropically convex, which implies that it
is in fact contractible [12] (and hence K-acyclic).
The key observation here is that the monomial label of any point x in A can be read off
from its type data. Namely, if F is a cell of A (of any dimension), and x ∈ F has type
TA (x) = (Ta , Tb ) then the monomial label of F is given by
n
Y
mF =
xdi i ,
(10)
i=1
SKELETA OF G-PARKING FUNCTION IDEALS
where
17
if i ∈ Ta and i ∈ Tb ,
n
di = n − 1 if i ∈ Ta \Tb or i ∈ Tb \Ta ,
0
if i ∈
/ Ta ∪ Tb .
This follows from the way that we have defined the monomial labeling. Namely, if F ∈ XA
is a maximal cell and x ∈ F then x does not lie on either hyperplane of the arrangement and
hence is in exactly one cone defined by each. In this case the monomial label of the relevant
cell is xn−1
xn−1
if the cones are distinct, or xni if the cones are the same. If x lies in some
i
j
other (non maximal) cell F ∈ XA then Proposition 4.3 implies that the monomial label xF
described in (10) is equal to the lcm of the monomial labels of all faces that contain F .
Now suppose x, y ∈ (XA )≤m and suppose z in a point in the tropical convex hull of x and
y. From Lemma 4.4 we have that TA (z) is contained in TA (x) ∪ TA (y). Hence if z lies in the
cell G ∈ XA we have from (10) that mF divides m. We conclude that z ∈ (XA )≤m so that
(XA )≤m is tropically convex and hence K-acyclic.
We next turn to minimality of the resolution. If F ⊂ G are cells of XA with F 6= G then
by Proposition 4.5 along with the labeling formula (10) we have that mF 6= mG . By Lemma
2.6 this implies that the resolution is minimal.
P
Recall that the ith total Betti number of a module is the sum σ βi,σ of all Betti numbers
(1)
in the ith homological degree. According to Theorem 4.6 the total Betti number of a Mn
is given by the number of codimension i faces in the decomposition of Rn determined by the
tropical hyperplane arrangement. Importing results from [15] we get the following formula.
(1)
Corollary 4.7. The total Betti numbers βin of the ideal Mn are given by
n
X
j−1
n
βi =
j
i−1
j=1
(1)
Example 4.8. For n = 3 the resolution of M3
has a Z-graded resolution given by
(2)
0 ← M3 ← S(−3)3 ⊕ S(−4)3 ← S(−5)6 ⊕ S(−6)2 ← S(−7)3 ← 0.
This can be read off from the monomial labeled complex depicted in Figure 3. The total Betti
numbers are given by β13 = 6, β23 = 8, β33 = 3.
Remark 4.9. Note that since the complexes (XA )≤m considered above are contractible (and
hence K-acylic for any field K), it follows that the resolution and associated Betti numbers
are independent of which field we are working over. Similar properties hold for the resolutions
of MG considered in [23] and [14].
The combinatorics of the decomposition of Rn induced by a generic arrangement of a pair
of tropical hyperplanes is closely related to regular triangulations of a product of simplices
∆1 × ∆n−1 , as spelled out in [12]. By the Cayley Trick [11] this implies that a cellular
(1)
resolution of the ideal MKn+1 is supported on any ‘regular fine mixed subdivision’ of the
Minkowski sum ∆n−1 +∆n−1 (see Figure 4). We will not stress this perspective here and refer
to [15] for more details, where arrangements of tropical hyperplanes are used to construct
18
ANTON DOCHTERMANN
minimal cellular resolutions of similar monomial ideals arising from oriented matroid ‘type’
data.
x34
x13x33
x23x33
x14
x33x43
x13x43
x24
x44
x23x43
(1)
Figure 4. A cellular resolution of M4 is supported on any regular fine mixed
subdivision of ∆3 + ∆3 .
Remark 4.10. Triangulations of a product of simplices (and the related mixed subdivisions
of dilated simplices) are well-studied objects (see for example Chapter 6.2 of [11]). Applying
(1)
this technology we obtain several combinatorial interpretations for the Betti numbers of Mn .
In particular if we consider the ‘staircase’ (or ‘pulling’) triangulation of ∆1 × ∆n−1 we see
that maximal syzygies (corresponding to simplices in the triangulation) can be encoded by
certain lattice paths or, equivalently, certain ‘non-crossing’ spanning trees of the complete
bipartite graph K2,n (see Figure 5). We refer to [11] for more details.
(1)
Figure 5. Maximal syzygies of Mn are indexed by ‘non-crossing’ spanning
trees of K2,n (here n = 3, compare with Example 4.8).
(1)
We next consider resolutions of the ideal MG for an arbitrary graph G. Recall that a
(1)
(possibly redundant) set of generators of MG is again given by the set of monomials mσ for
1 ≤ |σ| ≤ 2, and hence the decomposition of Rn induced by a generic arrangement of a pair
of tropical hyperplanes has a natural monomial labeling with the property that its maximal
cells correspond to these generators. However one can check that with such a labeling a
generic arrangement of hyperplanes does not support a resolution of MG for general G. For
example if G is the graph obtained from K4 by removing the edge 12, then the arrangement
depicted in Figure 3 (with relevant monomial labels) will have the property that the downset
(XA )≤(112) is disconnected. For certain graphs, however, we can perturb our arrangements
to obtain complexes that support (minimal) resolutions.
Proposition 4.11. Suppose G = H ∗ {0} is obtained by coning the vertex 0 over the graph
H, where H is obtained by removing a collection of disjoint cliques from the complete graph
SKELETA OF G-PARKING FUNCTION IDEALS
19
Kn . Then there exists a (degenerate) arrangement A = (T (a), T (b)) ⊂ Rn−1 of tropical hyperplanes with the property that the induced subdivision of Rn−1 supports a minimal cocellular
(1)
resolution MG .
Proof. Without loss of generality we can assume that H is obtained from the complete graph
Kn by removing cliques C1 , C2 , . . . , Cd , where
V (C1 ) = {1, 2, . . . , k1 }, V (C2 ) = {k1 + 1, . . . , k2 }, . . . , V (Cd ) = {kd−1 + 1, . . . , kd }.
We will explicitly write down the coordinates of the hyperplane arrangement that supports
(1)
our resolution of MG . For this let a = (0, . . . , 0) be the origin in Rn−1 and define b ∈ Rn−1
according to:
. . . , 1, 2, . . . , 2, . . . , d, . . . , d, d + 1, d + 2, . . . , n − kd + d),
if kd 6= n
(1,
| {z }
| {z } | {z }
k1 -times
k2 − k1
kd − kd−1
b=
(1, . . . , 1, 2, . . . , 2, . . . , d − 1, . . . , d − 1, 0, 0, . . . , 0 ),
if kd = n.
| {z } | {z }
{z
} | {z }
|
k1 -times
k2 − k1
kd−1 − kd−2
kd − kd−1 − 1
Let A = (T (a), T (b)) be the resulting arrangement of tropical hyperplanes, and let XA
denote the induced decomposition of Rn−1 . We first claim that the maximal regions of XA
(1)
are in one-to-one correspondence with the set of minimal generators of MG . Recall that for
(1)
any G, a redundant set of generators for MG is given by all monomials mσ for all subsets
σ ⊂ [n] with 1 ≤ |σ| ≤ 2. For our given graph G, a minimal set of generators is given by
i
monomials m{i} = xdeg
for all i ∈ [n], along with m{i,j} for all subsets {i, j} ∈ E(G). To see
i
this note that if ij ∈
/ E(G) we have
deg(i) deg(j)
xj
m{i,j} = xi
= m{i} m{j} ,
and hence the monomial m{i,j} is redundant.
On the other hand, if F ∈ XA is a maximal cell with x ∈ F (so that x does not intersect
either of the hyperplanes T (a) nor T (b)) then x sits in exactly one cone of each hyperplane.
Hence the sets Ta and Tb are each singletons (where T (x) = (Ta , Tb ) is the type of x). By
construction the only such types that can occur are of the type ({i}, {i}) and ({i}, {j}),
where ij ∈ E(G). Furthermore, all such pairs of singletons actually do occur exactly once
as types of some region.
To see this, note that since these types occur exactly once in a generic arrangement
it suffices to show that these types are actually achieved in our particular arrangement
A = (T (a), T (b)). We first consider the repeated singletons. To achieve ({i}, {i}) for i 6= n
simply take the x to be a vector of 0’s except for a sufficiently large positive entry in the ith
coordinate. To achieve ({n}, {n}) take x = (−1, −1, . . . , −1), a vector of all −1’s.
To realize a cell with the type ({i}, {j}) with i < j we have to consider a couple of cases
depending on the value of kd (the largest vertex in our removed cliques). In the case that
kd 6= n we achieve the type ({i}, {j}), j 6= n by taking x to be the vector consisting of all 0’s
except the values bi + 1 in the ith coordinate and bj + 1/2 in the jth coordinate. To achieve
the pair ({i}, {n}) take x to be all 0’s except for the value bi + 1/2 in the ith coordinate.
In the case that kd = n we can achieve any type ({i}, {j}) with i < j as follows. If j ∈ Cd
take x to be all 0’s except bi + 1/2 in the ith coordinate and bj + 1 = 1 in the jth coordinate.
20
ANTON DOCHTERMANN
Otherwise if j ∈
/ Cd take x to be all 0’s except bi + 1 in the ith coordinate and bj + 1/2 in
the jth coordinate.
We next describe the mononial labeling on the complex XA , a modification of the monomial
labeling of the case of the complete graph described in the proof of Theorem 4.6. If F is a
cell of A (of any dimension), and x ∈ F has type TA (x) = (Ta , Tb ) then the monomial label
of F is given by
n
Y
mF =
xdi i ,
(11)
i=1
where
if i ∈ Ta and i ∈ Tb ,
deg(i)
di = deg(i) − 1 if i ∈ Ta \Tb or i ∈ Tb \Ta ,
0
if i ∈
/ Ta ∪ Tb .
With this convention, the maximal cells of XA are labeled with monomials corresponding to
(1)
the set of minimal generators of MG . For all other faces F , Proposition 4.3 implies that the
monomial label xF described in (10) is equal to the lcm of the monomial labels of all faces
that contain F . If m ∈ Zn is any monomial then the same argument employed in the proof
of Theorem 4.6 shows that the complex (XA )≤m is tropically convex and hence contractible.
(1)
Applying Lemma 2.6 then implies that XA supports a cocellular resolution of the ideal MG .
Furthermore, if F ⊂ G are cells of XA with F 6= G then Proposition 4.5 along with the
formula (11) implies that mF 6= mG , so that the resolution is minimal. This completes the
proof.
0
0
1
4
2
3
1
4
2
3
x33
x32
x12x32
x13
x22x32
x12
x3x43
x1x43
x12x42
x23
x22x42
x43
x22
x2x43
x44
(1)
Figure 6. Resolutions of MGi for the graphs G1 and G2 from Example 4.12.
Example 4.12. As examples of the previous result, consider the graphs G1 and G2 depicted
in Figure 6. The graph G1 is obtained by removing the complete graphs (edges) 12 and 34, and
in G2 we have removed the triangle 123. The cellular resolutions of each ideal are illustrated
in Figure (6). In both cases the complexes are dual to a non-generic arrangement of a pair
of tropical hyperplanes in R3 . For G1 we have b = (1, 1, 0), for G2 we have b = (1, 1, 1).
SKELETA OF G-PARKING FUNCTION IDEALS
21
5. Discussion and open questions
We end with some discussion regarding other open questions and possible future directions.
5.1. Resolutions of other skeleta. We do not have uniform descriptions of resolutions
(k)
for other skeleton ideals MG , even for the complete graph G = Kn+1 . We have seen that a
cellular resolution of Mn = MKn+1 is supported on B(∆n−1 ), the barycentric subdivision of
the (n − 1)-dimensional simplex. As discussed above, a reasonable candidate for a geometric
(k)
complex supporting a resolution of the ideals Mn would be a subdivision of the simplex
∆n−1 that agrees with B(∆n−1 ) up to the k-dimensional skeleton. For the case of k = 1
we have seen that complex generated by a generic arrangement of two tropical hyperplanes
has this property, but we do not know of any candidates for other skeleta. We remark that
an arrangement of k tropical hyperplanes will not produce the desired complex, since for
instance the number of maximal cells is not equal to the number of minimal generators of
the relevant ideal.
(n−2)
⊂ MG
A particularly tractable case seems to be in codimension one, where the ideal MG
is only missing the generator of MG given by the monomial m[n] . Here we would need a
subdivision of ∆n+1 that agrees with the complex B(∆n+1 ) on the boundary but is missing
the ‘interior’ vertex.
5.2. Enumeration of parking functions. A natural invariant of a parking function (thought
of as a standard monomial in Mn := MKn+1 ) is its degree. With this statistic we can define
a generating function Pn (q) of parking functions of length n as follows:
X
Pn (q) =
q a1 +a2 +···+an
α=(a1 ,a2 ,...,an )
where α ranges over all parking functions of length n (corresponding to standard monomials
of Mn ).
For small values of n we get
P1 (q) = 1
P2 (q) = 2q + 1
P3 (q) = 6q 3 + 6q 2 + 3q + 1
(12)
P4 (q) = 24q 6 + 36q 5 + 30q 4 + 20q 3 + 10q 2 + 4q + 1
Kreweras [20] studied another polynomial In (q) which enumerates (labeled) rooted forests
by number of inversions, and showed that In (q) was closely connected to Pn (q). Here a rooted
forest on [n] is a graph on the vertex set {1, 2, . . . , n} with the property that each connected
component is a rooted tree. An inversion of a rooted forest is a pair (i, j) satisfying i < j
and such that j lies on the unique path connecting k to i, where k is the root of the tree
containing the vertex i. We let inv(F ) denote the number of inversions of F and define the
polynomial In (q) as
X
In (q) =
q inv(F )
F
where F ranges over all labeled forests on [n].
22
ANTON DOCHTERMANN
The polynomials In (q) have the following connection to parking functions of complete
graphs (as shown in [20]):
n
q ( 2 ) In (1/q) = Pn (q).
(k)
Is there an analogous story for the standard monomials of the skeleton ideals Mn ? Using
the degree statistic we can similarly define a generating function
X
Pn(k) (q) =
q b1 +b2 +···+bn
β=(b1 ,b2 ,...,bn )
(k)
where β ranges over all standard monomials of Mn .
For the case of k = 1 and small values of n we have
(1)
P1 (q) = 1
(1)
P2 (q) = 2q + 1
(13)
(1)
P3 (q) = 3q 4 + 7q 3 + 6q 2 + 3q + 1
(1)
P4 (q) = 4q 9 + 13q 8 + 28q 7 + 38q 6 + 40q 5 + 31q 4 + 20q 3 + 10q 2 + 4q + 1
(1)
In Section 3 we have seen that standard monomials of Mn are counted (in a weighted
fashion) by the spanning T U -subgraphs of Kn (with spanning trees appearing as a special
cases). Is there a notion of an ‘inversion’ for a T U -subgraph that recovers this statistic?
5.3. Generalized parking. We briefly recall the meaning of ‘parking’ in the the parking
functions ideals that we have been studying. We imagine a parking lot consisting of n
labeled spots lined up from right to left, and a sequence of positive integers (a1 , a2 , . . . , an )
as encoding the preferences for parking spaces among n cars: the first car prefers spot a1 ,
the second car a2 , and so on. The cars enter the lot from the right and the procedure for
parking is as follows: the first car takes spot a1 , the second car takes spot a2 unless it is
already taken, in which case it takes the next available spot (moving from right to left). If
no spot is available that car is not able to park and gives up. The procedure continues with
a3 , etc. The sequence (a1 , a2 , . . . , an ) is a parking function if all n cars succeed in parking
in the lot. It can be shown that the set of parking functions correspond to the (exponent
vectors of) standard monomials of MKn+1 (to be precise we must add 1 to each coordinate,
so that e.g. if n = 3 then x21 x3 corresponds to the parking function (3, 1, 2)).
Is there a similar ‘parking’ interpretation for the standard monomials of the skeleton ideals
(k)
Mn defined above? We note that these objects do not seem to be related (as far as we
know) to another generalization of parking functions discussed in [26].
5.4. Power ideals. Recall that in [23] the authors were interested in another class of ‘power
ideals’ that one associate to a graph G on vertex set {0, 1, . . . , n}. We briefly recall the
definition. For any nonempty subset σ ⊆ {1, 2, . . . , n} let
X Dσ
pσ =
xi
,
i∈σ
P
where Dσ = i∈σ dσ (i), with dσ defined in Equation (3). Now define JG to be the ideal in
the polynomial ring S = K[x1 , x2 , . . . , xn ] generated by the pσ for all nonempty subsets σ,
and let BG denote the quotient algebra BG = S/JG . Recall that AG = S/MG is the algebra
SKELETA OF G-PARKING FUNCTION IDEALS
23
obtained by factoring out the G-parking function ideal. The algebras AG and BG are both
graded, and in [23] it is shown that the Hilbert series coincide. In fact both algebras have
a K-basis given by the monomials xb where b ranges over all G-parking functions. This is
particularly surprising since (in general) the monomial ideal MG is not an initial ideal of the
ideal JG for any monomial order, and hence standard Gröbner degeneration techniques do
not apply.
In our context we have a natural definition for the skeleton of these power ideals, mimicking
the construction of the monomial case discussed above. Namely, for any graph G and integer
(k)
k with 0 ≤ k ≤ n − 1 we define the k-skeleton power ideal JG to be the subideal of JG given
by
(k)
JG = hpσ : 1 ≤ |σ| ≤ k + 1i.
(k)
(k)
One can check that for 0 ≤ k ≤ 1 we have MG = JG (equality as ideals), but already
for the 2-skeleton the ideals differ. In fact the quotient algebras can have different Hilbert
series, as the next example illustrates.
Example 5.1. Let G be the graph obtained from removing the edge 12 from the complete
graph on vertex set {0, 1, 2, 3, 4}. Then we have
(2)
MG = hx31 , x32 , x43 , x44 , x21 x33 , x21 x34 , x22 x33 , x22 x34 , x33 x34 , x21 x22 x23 , x21 x2 x24 , x1 x23 x24 , x2 x23 x24 i
(14)
(2)
JG = hx31 , x32 , x43 , x44 , (x1 + x2 )6 , (x1 + x3 )5 , (x1 + x4 )5 , (x2 + x3 )5 , (x2 + x4 )5 ,
(x3 + x4 )6 , (x1 + x2 + x3 )6 , (x1 + x2 + x4 )5 , (x1 + x3 + x4 )5 , (x2 + x3 + x4 )5 i
(15)
(2)
According to our calculations with Macaulay2 [18], the graded components of S/MG and
(2)
S/JG have the same dimension except in degree 6, where
(2)
(2)
dimK (S/MG )6 = 7 > 6 = dimK (S/JG )6 .
(k)
(k)
We do not know if termwise inequalities Hilb S/MG ≥ Hilb S/JG hold for the Hilbert
functions of the two algebras, as was the case for deformations of more general monotone
monomial ideals [23] (and as in Example 5.1). We will point that in this example the
(2)
(2)
monomial basis for (S/MG )6 does not span the linear space (S/JG )6 .
References
[1] S. Backman, A Bijection Between the Recurrent Configurations of a Hereditary Chip-Firing Model and
Spanning Trees, preprint available at arxiv.org:1401.3309.
[2] M. Baker and S. Norine, Riemann-Roch and Abel-Jacobi Theory on a Finite Graph, Adv. Math. 215
(2007), pp. 766–788.
[3] R. P. Bapat, Graphs and matrices, Universitext, Springer-Verlag London, 2010.
[4] Dave Bayer and Bernd Sturmfels, Cellular resolutions of monomial modules, J. Reine Angew. Math.
502 (1998), pp. 123–140.
[5] A. Benjamin and F. Juhnke, Another way of counting N N , SIAM J. Disc. Math. 5, No. 3 (1992), pp.
377–379.
[6] S. Caracciolo, G. Paoletti, A. Sportiello, Multiple and inverse topplings in the Abelian Sandpile Model,
Eur. Phys. J. Special Topics 212 (2012).
[7] C. Chauve, S. Dulucq, O. Guibert, Enumeration of some labelled trees, Proceeding of SFCA/FPSAC
2000 (Moscow, June 2000), D. Krob and A. Mikhalev eds., Springer, pp. 146–157, 2000.
24
ANTON DOCHTERMANN
[8] Denis Chebikin, Pavlo Pylyavskyy, A family of bijections between G-parking functions and spanning
trees, J. Combin. Theory Ser. A 110, Issue 1 (2005), pp. 31–41.
[9] R. Cori, D. Rossin, B. Salvy, Polynomial ideals for sandpiles and their Grobner bases, Theor. Comp.
Sci. 276 (2002), pp. 1–15.
[10] R. Cori, Y. Le Borgne, The sand-pile model and Tutte polynomials, Adv. Appl. Math. 30 (2003), pp.
44–52.
[11] J. De Loera, J. Rambau, F. Santos, Triangulations, Structures for Algorithms and Applications, Algorithms and Computation in Mathematics, vol. 25, Springer-Verlag Berlin Heidelberg, 2010.
[12] M. Develin, B. Sturmfels, Tropical convexity, Doc. Math. 9 (2004), pp. 1–27.
[13] Deepak Dhar, Self-organized critical state of sandpile automaton models, Phys. Rev. Lett. 64 (1990),
no. 14, pp. 1613–1616.
[14] A. Dochtermann and R. Sanyal, Laplacian ideals, arrangements, and resolutions, J. Algebr. Comb.
(2014) 40, pp. 805–822.
[15] A. Dochtermann, M. Joswig, R. Sanyal, Tropical types and associated cellular resolutions, J. Algebra.
356 (2012), pp. 304–324.
[16] J. Francon, Acyclic and parking functions, J. Combin. Theory Ser. A 18 (1975), pp. 27–35.
[17] Andrei Gabrielov, Abelian avalanches and Tutte polynomials, Phys. A 195 (1993), no. 1-2, pp. 253–274.
[18] D. Grayson, M. Stillman, Macaulay2, a software system for research in algebraic geometry, Available at
http://www.math.uiuc.edu/Macaulay2/
[19] Alan G. Konheim and Benjamin Weiss, An occupancy discipline and applications, SIAM J. Applied
Math. 14 (1966), 1266–1274.
[20] G. Kreweras, Une famille de polynômes ayant plusieurs propriétés énumeratives. (French) Period. Math.
Hungar. 11 (1980), no. 4, pp. 309–320.
[21] C.L. Mallows, John Riordan, The inversion enumerator for labeled trees, Bull. Amer. Math. Soc. Volume
74, Number 1 (1968), pp. 92–94.
[22] E. Miller, B. Sturmfels, Combinatorial commutative algebra. Graduate Texts in Mathematics 227.
Springer-Verlag, New York, 2005.
[23] Alexander Postnikov and Boris Shapiro, Trees, parking functions, syzygies, and deformations of monomial ideals, Trans. Amer. Math. Soc. 356 (2004), pp. 3109–3142.
[24] Richard Stanley, Hyperplane Arrangements, Parking Functions and Tree Inversions, Mathematical Essays in honor of Gian-Carlo Rota Volume 161 of the series Progress in Mathematics, pp. 359–375.
[25] L. Takács, On Cayley’s Formula for Counting Forests, J. Combin. Theory Ser. A 53, (1990), pp. 321–323.
[26] Catherine Huafei Yan, On the enumeration of generalized parking functions, Proceedings of the Thirtyfirst Southeastern International Conference on Combinatorics, Graph Theory and Computing (Boca
Raton, FL, 2000), 2000, pp. 201–209.
Department of Mathematics, Texas State University, San Marcos
E-mail address: [email protected]
| 0math.AC
|
Optimal Gossip Algorithms for Exact and Approximate Quantile
Computations
arXiv:1711.09258v1 [cs.DS] 25 Nov 2017
Bernhard Haeupler
CMU
Jeet Mohapatra
MIT
Hsin-Hao Su
UNC Charlotte
Abstract
This paper gives drastically faster gossip algorithms to compute exact and approximate
quantiles.
Gossip algorithms, which allow each node to contact a uniformly random other node in
each round, have been intensely studied and been adopted in many applications due to their
fast convergence and their robustness to failures. Kempe et al. [KDG03, FOCS’03] gave gossip
algorithms to compute important aggregate statistics if every node is given a value. In particular,
they gave a beautiful O(log n + log 1 ) round algorithm to -approximate the sum of all values
and an O(log2 n) round algorithm to compute the exact φ-quantile, i.e., the the dφne smallest
value.
We give an quadratically faster and in fact optimal gossip algorithm for the exact φ-quantile
problem which runs in O(log n) rounds. We furthermore show that one can achieve an exponential speedup if one allows for an -approximation. We give an O(log log n + log 1 ) round
gossip algorithm which computes a value of rank between φn and (φ + )n at every node.Our
algorithms are extremely simple and very robust - they can be operated with the same running
times even if every transmission fails with a, potentially different, constant probability. We also
give a matching Ω(log log n + log 1 ) lower bound which shows that our algorithm is optimal for
all values of .
1
Introduction
Today, due to the vast amount of data and advances in connectivity between computers, distributed
data processing has become increasingly important. In distributed systems, data is stored across
different nodes. When one requires some aggregate properties of the data, such as, sums, ranks,
quantiles, or other statistics, nodes must communicate in order to compute these properties. Aggregating data in an efficient and reliable way is a central topic in distributed systems such as P2P
networks and sensor networks [YG02, MFHH02, SBAS04].
We consider uniform gossip protocols, which are often very practical due to their fast convergence,
their simplicity, and their stability under stress and disruptions. In uniform gossiping protocols,
computation proceeds in synchronized rounds. In each round, each node chooses to pull or push.
In a push a node chooses a message, which is delivered to a uniformly random other node. In a
pull each node receives a message from a random node. The message size is typically restricted to
O(log n) bits. The (time) complexity of an algorithm is measured by the number rounds executed.
One typically wants algorithms that succeed with high probability, i.e., with probability at least
1 − 1/ poly(n).
In this paper, we study the quantile computation problem. In the exact φ-quantile problem
each node v is given a O(log n) bit value xv and wants to compute the dφne smallest value overall.
In the -approximate φ-quantile problem each node wants to compute a value whose rank is
between (φ + )n and (φ − )n.
Previously, Kempe et al. [KDG03] gave a beautiful and simple O(log n + log 1 ) round gossip
algorithm to approximate the sum of all values up to a (1 ± ) factor. They also showed how to
use this algorithm to solve the exact φ-quantile problem in O(log2 n) rounds with high probability.
The main result of this paper is a quadratically faster gossip algorithm for the φ-quantile problem.
The O(log n) round complexity of our algorithm means that the exact φ-quantile can be computed
as fast as broadcasting a single message.
Theorem 1.1. For any φ ∈ [0, 1] there is a uniform gossip algorithm which solves the exact φquantile problem in O(log n) rounds with high probability using O(log n) bit messages.
Clearly the running time of this algorithm is optimal as Ω(log n) rounds are known to be neccesary
to even just broadcast to each node the φ-quantile value, after it has been identified.
Equally interestingly we show that one achieve even faster algorithms if one considers the approximate φ-quantile problem. While a O( 12 log n) round algorithm for an -approximation follows from
q
simple sampling and a O(log n) round algorithm computing the median up to a ±O( logn n ) (but
not general quantiles) was given by Doerr et al. [DGM+ 11], no approximation algorithm for the
quantile problem with a sub-logarithmic round complexity was known prior to this work. We give
an O(log log n + log(1/)) round algorithm for -approximating any φ-quantile which, for arbitrarily
good constant approximations, is exponentially faster than our optimal exact algorithm:
Theorem 1.2. For any constant or non-constant (n) > 0 and any φ ∈ [0, 1], there exists a uniform
1
) rounds
gossip algorithm that solves the -approximate φ quantile problem in O(log log n + log (n)
with high probability using O(log n) bit messages.
We also give a Ω(log 1 ) and a Ω(log log n) lower bound for the -approximate φ-quantile problem,
which shows that our algorithm is optimal for essentially any value of .
1
n
< < 1/8 and φ ∈ [0, 1], any gossip algorithm that uses less
Theorem 1.3. For any 10 log
n
1
than 2 log log n or less than log4 8 round fails to solve the -approximate φ-quantile problem with
probability at least 1/3. This remains true even for unlimited message sizes.
We furthermore show that our algorithms can be made robust to random failures, i.e., the
same round complexities apply even if nodes fail with some constant probability. We remark that a
small caveat of excluding an exp(−t) fraction of nodes after a runnig time of t rounds is necessary
and indeed optimal given that this is the expected fraction of nodes which will not have participated
in any successful push or pull after t rounds.
Theorem 1.4. Suppose in every round every node fails with a, potentially different, probability
bounded by some constant µ < 1. For any φ ∈ [0, 1] there is a gossip algorithm that solves the φ
quantile problem in O(log n) rounds. For any t and any (n) > 0, there furthermore exists a gossip
1
algorithm that solves the -approximate φ-quantile problem in O(log log n + log (n)
+ t) rounds for
n
all but 2t nodes, with high probability.
Despite being extremely fast and robust our algorithms remain very simple. In fact they do little
more than repeatedly sampling two or three nodes, requesting their current value and selecting
the largest, smallest, or median value. Given that in many cases an -approximation is more than
sufficient we expect this algorithm to easily find applications in areas like sensor networks and
distributed database systems. For instance, suppose that a sensor network consisting of thousands
of devices is spread across an object to monitor and control the temperature. Say the top and
bottom 10%-quantiles need special attention. By computing the 90%- and 10%-quantile, each node
can determine whether it lies in the range. It is unlikely that such a computation needs to exact.
In fact, for any 0 < < 1, running O(1/) approximate quantile computations suffices for each
node to determine its own quantile/rank up to an additive . The fact that this can be done in
O(log log n) rounds further demonstrates the power of approximations as there is not even a o(n)
algorithm known1 which allows each node to compute its quantile/rank exactly.
Corollary 1.5. For any constant or non-constant > 0 and given that each node has a value
there is a gossip algorithm that allows every node to approximate the quantile of its value up to an
additive in 1 · O(log log n + log 1 ) rounds with high probability using O(log n) bit messages.
Technical Summary. While computing exact quantiles seems to be a very different problem from
the approximate problem at first we achieve our exact algorithm by first designing an extremely
efficient solution to the -approximate quantile problem running in O(log log n + log 1 ) rounds.
This algorithm is in some sense based on sampling and inherently is not able to work for the exact
quantile problem or too small itself, as it might randomly discard the target value in its first
iteration. However, we show that the algorithm does works for larger than some polynomial in n.
This allows us to bootstrap this algorithm and repeatedly remove a polynomial fraction of values
within O(log n) rounds until, after a constant number of such iterations, only the target value
persists. Overall this leads to an O(log log n + log 1 ) algorithm for the -approximate φ-quantile
problem which works for values that can be any function of n. This is actually generelizes the
1
O(log n) algorithm for the exact φ-quantile problem, too, given that the 2n
-approximate φ-quantile
problem is accurate enough to compute exact ranks and quantiles.
1
The trivial algorithm which broadcasts the maximum value n times requires O(n log n) rounds. Using network
coding gossip [Hae16] one can improve this to O(n) rounds. It is likely that computing the exact rank at each node
cannot be done faster.
2
Next we outline an, indeed very simple, intuition of why a time complexity of O(log log n + log 1 )
is conceivable for the quantile computation problem: Suppose we sample Θ(log n/2 ) many values
uniformly and independently at random. With high probability the φ-quantile of the sampled
values is an -approximation to the φ-quantile in the original data. Since each node can sample t
node values (with replacement) in t rounds, one immediately get an O(log n/2 ) round algorithm
that uses O(log n) bit messages.
Using larger message sizes, it is possible to reduce the number of rounds. Consider the following
doubling algorithm: Each node v maintains a set Sv such that initially Sv = {xv }. In each round,
let t(v) be the node contacted by v. Node v updates Sv by setting Sv ← Sv ∪St(v) . Since the set size
essentially doubles each round, after log O(log n/2 ) = O(log log n + log 1 ) rounds, we have sampled
Ω(log n/2 ) values uniformly, albeit not quite independently, at random. A careful analysis shows
that indeed a O(log log n + log 1 ) running time can be achieved using messages of size Θ(log2 n/2 )
bits.
Our first approach for reducing the message sizes tried to utilize the quantile approximation
sketches from the streaming algorithm community (see related work). We managed to reduce the
message complexity to O( 1 · log n · (log log n + log(1/))) by adopting the compactor ideas from
these streaming sketches. A write-up of this might be of independent interest and can be found
in the Appendix A. Unfortunately even if one could losslesly port the state-of-the-art compactors
scheme from [KLL16] into this setting, the Ω( 1 log log(1/δ)) for getting -approximate quantile
sketches with probability 1 − δ suggests that one cannot achieve a o(log n log log n) message size
this way, even if is a constant. In contrast to most other distributed models of computation gossip
algorithms furthermore do not allow to easily shift extra factors in the message size over to the
round complexity. In fact, we did not manage to devise any o(log n) round gossip algorithm based
on sketching which adheres to the standard O(log n) bound on message sizes.
Instead of sticking to the centralized mentality where each node tries to gather information to
compute the answer, consider the following 3-TOURNAMENT mechanism. In each iteration, each
node uniformly sample three values (in three rounds) and assign its value to the middle one2 .
Intuitively nodes close to the median should have a higher probability of surviving and being
replicated. Indeed, we show that the number of nodes that have values that are close to the
median grows exponentially for the first O(log 1 ) iterations. After that, the number of nodes with
values more than n away from the median decreases double exponentially. For sufficiently large
this simple 3-TOURNAMENT algorithm gives an approximation of the the median in O(log 1 +log log n)
iterations. In general, if we want to approximate the φ-quantile, we shift the [φ − , φ + ] quantiles
to the quantiles around the median by the following 2-TOURNAMENT mechanism. If φ < 1/2, each
node samples two values and assign its value to the higher one. The case for φ > 1/2 is symmetric.
Intuitively this process makes it more likely for nodes with higher/smaller values to survive. Indeed,
we show that with a bit of extra care in the last iterations, O(log 1 ) invocations suffice to shift
the values around the φ-quantile to almost exactly the median, at which point one can apply the
median approximation algorithm.
Finally, our lower bound comes from the fact that if one chooses Θ(n) nodes and either gives
them a very large or a very small value, then knowing which of the two cases occurred is crucial
for computing any -approximate quantiles. However, initially only these Θ(n) nodes have such
information. We show that it takes Ω(log log n + log 1 ) rounds to spread the information from these
2
Such gossip dynamics have also been analyzed in [DGM+ 11] for the setting of adversarial node failures and in
[FHK17] for the setting of random message corruptions.
3
nodes to every node, regardless of the message size.
Related Work The randomized gossip-based algorithms dates back to Demers et al. [DGH+ 87].
The initial studies are on the spreading of a single message [FG85, Pit87, KSSV00], where Karp et
al. [KSSV00] showed that O(log n) round and O(n log log n) total messages is sufficient to spread
a single message w.h.p. Kempe et al. [KDG03] studied gossip-based algorithms for the quantile
computation problem as well as other aggregation problems such as computing the sum and the
average. Kempe et al. developed O(log n) rounds algorithm to compute the sum and average
w.h.p. Later, efforts have been made to reduce the total messages to O(n log log n) for computing
the sum and the average [KDN+ 06, CP12]. Using the ability to sample and count, Kempe et
al. implemented the classic randomized selection algorithm [Hoa61, FR75] in O(log2 n) rounds.
Doerr et al. [DGM+ 11] considered gossip algorithms for the problem of achieving a stabilizing
consensus algorithm under adversarial node failures. They analyze the median rule, i.e., sample
three values and keepq
the middle value, in this setting and show that O(log n) rounds suffice
√
to converge to an ±O( logn n )-approximate median even if O( n) adversarial node failures occur.
Similar gossip dynamics were also studied in [FHK17] which considers randomly corrupted (binary)
messages.
The exact quantile computation is also known as the selection problem, where the goal is to
select the k’th smallest element. The problem has been studied extensively in both centralized
and distributed settings. Blum et al. [BFP+ 73] gave a deterministic linear time algorithm for the
problem in the centralized setting. In the distributed setting, Kuhn et al. [KLW07] gave an optimal
algorithm for the selection problem that runs in O(D logD n) rounds in the CONGEST model, where
D is the diameter of the graph. Many works have been focused on the communication complexity
aspect (i.e. the total message size sent by each node) of the problem [SFR83, RSS86, SSS88, SS89,
SSS92, NSU97, GK04, PS07]. Most of them are for complete graphs or stars. Others studied
specific class of graphs such as two nodes connected by an edge [Rod82, CT87], rings, meshes, and
complete binary trees [Fre83].
The quantile computation problem has also been studied extensively in the streaming algorithm
literature [MP80, ARS97, MRL99, GK01, GM09, HT10, ACH+ 13, MMS13, FO15, KLL16], where
the goal is to approximate φ-quantile using a small space complexity when the data comes in a
single stream.
2
The Tournament Algorithms
In this section, we present our algorithm for the -approximate quantile computation problem for
sufficiently large . For convenience, we use a ± b to denote the interval [a − b, a + b].
Theorem 2.1. For any constant or non-constant (n) = Ω(1/n0.096 ) and any φ ∈ [0, 1], there exists
1
a uniform gossip algorithm that solves the -approximate φ quantile problem in O(log log n+log (n)
)
rounds with high probability using O(log n) bit messages.
The algorithm is divided into two phases. In the first phase, each node adjusts its value so that
the quantiles around the φ-quantile will become the median quantiles approximately. In the second
phase, we show how to compute the approximate median.
4
2.1
Phase I: Shifting the Target Quantiles to Approximate Medians
Let H denote the nodes whose quantiles lie in (φ + , 1], L denote the nodes whose quantiles lie in
[0, φ − ), and M denote the nodes whose quantiles lie in [φ − , φ + ]. Let Li , Mi , and Hi denote
the set L, M , and H respectively at the end of iteration i for i ≥ 1 and L0 , M0 , and H0 denote
the sets in the beginning. The goal is to show that by the end of the algorithm at iteration t, the
size of |Lt | and |Ht | are 1/2 − Ω() so that an approximate median lies in Mt , which consists of our
target quantiles.
Let h0 = 1 − (φ + ) and l0 = φ − . We first consider the case where h0 ≥ l0 and the other case is
symmetric. Initially, we have |Hn0 | ∈ h0 ± 1/n. Let hi+1 = h2i for i ≥ 1. We run the 2-TOURNAMENT
def
algorithm (Algorithm 1) until the iteration t such that ht ≤ T , where T = 1/2 − . We will show
that t = O(log(1/)) and |Hni | concentrates around hi for iteration 1 ≤ i ≤ t − 1 and in the end we
have |Hnt | ∈ T ± 2 .
Algorithm 1 2-TOURNAMENT(v)
1: h0 ← (1 − (φ + ))
2: i ← 0, T = 1/2 − .
3: while hi > T do
4:
hi+1 ← h2i
5:
6:
7:
8:
9:
10:
11:
12:
13:
i −T
δ ← min 1, hih−h
i+1
With probability δ do
Select two nodes t1 (v) and t2 (v) randomly
xv ← min(xt1 (v) , xt2 (v) )
Otherwise do
Select a node t1 (v) randomly
xv ← xt1 (v)
i←i+1
end while
The algorithm ends when hi decreases below T . The lemma below bounds the number of iterations needed for this to happen. It can be shown that since h2i squares in each iteration, the
quantity (1 − hi ) roughly grows by a constant factor in each iteration. Since initially (1 − h0 ) ≥ ,
O(log(1/)) iterations suffice for hi to decrease below T .
Lemma 2.2. Let t denote the number of iterations in Algorithm 1, t ≤ log7/4 (4/) + 2.
Proof. The algorithm ends when ht ≤ 1/2 − . The highest possible value for h0 is 1 − . We show
i
i−1
that hi ≤ 1 − 47 · provided that 1 − 47
· ≥ 3/4.
5
Suppose by induction that hi−1 ≤ 1 −
7 i−1
4
· . We have
hi = h2i−1
i−1 !2
7
≤ 1−
·
4
i−1 !
i−1
7
7
·· 2−
·
=1−
4
4
i
i−1
1
7
7
·· 2−
≤1−
·
≤1−
4
4
4
i−1
7
1−
· ≥ 3/4
4
Therefore, after i0 = log7/4 (4/) iterations we have, hi0 ≤ 1 − ( 47 )i · ≤ 1 −
hi0 +2 ≤ ( 43 )4 < 12 − for < 1/8, it must be the case t ≤ i0 + 2 = O(log(1/)).
1
4
=
3
4.
Since
Note that Line 7 and Line 8 are executed with probability 1 during the first t − 1 iterations. We
show the following.
|
| Hi ] =
Lemma 2.3. For iteration 1 ≤ i < t − 1, E[ |Hi+1
n
|Hi |2
.
n2
Proof.
|Hi+1 |
1 X |Hi | 2
|Hi |2
1X
| Hi =
Pr(xt1 (v) ∈ Hi ∧ xt2 (v) ∈ Hi ) =
=
E
n
n
n
n
n2
v∈V
v∈V
|
|
Therefore, for 1 ≤ i ≤ t − 1, if |Hni | ∼ hi (∼ means they are close), then |Hi+1
∼ hi+1 if |Hi+1
n
n
concentrates around its expectation, since hi+1 = h2i .
In the last iteration, we truncate the probability of doing the tournament by doing it with only
probability δ for each node, so that ideally we hope to have E[ |Hnt | |Ht−1 ] = T . However, since δ is
|
calculated with respect to ht−1 instead of the actual |Ht−1
n , we need the following lemma to bound
|
∼ ht−1 , then indeed we have
the deviated expectation. In the next lemma, we show that if |Ht−1
n
|Ht |
E[ n |Ht−1 ] ∼ T .
00
Lemma 2.4.h Suppose that
i (1 − )ht−1 ≤
T − 300 ≤ E |Hnt | | Ht−1 ≤ T + 300 .
|Ht−1 |
n
≤ (1 + 00 )ht−1 for some 0 < 00 < 1. We have
Proof. First,
|Ht−1 |2
|Ht |
1X
|Ht−1 |
E
| Ht−1 =
δ·
+ (1 − δ) ·
n
n
n2
n
v∈V
|Ht−1 |
|Ht−1 |
· (1 − δ) + δ ·
=
n
n
6
|Ht−1 |
n
≤ (1 + 00 )ht−1 . We have
|Ht |
E
| Ht−1 ≤ (1 + 00 )ht−1 · (1 − δ) + δ · (1 + 00 )ht−1
n
T · (1 − ht−1 ) − 00 ht−1 T + 00 h2t−1
00
= (1 + )ht−1 ·
ht−1 − ht
2
T · (ht−1 − ht−1 )
00 h2t−1 − 00 ht−1 T
00
= (1 + ) ·
+ ht−1 ·
ht−1 − ht
ht−1 − ht
(ht−1 − T )
= (1 + 00 ) · T + 00 h2t−1 ·
ht−1 − ht
00
00
≤ (1 + ) · T + ht ≤ T + 300
Suppose that
|
Similarly, if |Ht−1
≥ (1 − 00 )ht−1 , we have
n
|Ht |
E
| Ht−1 ≥ (1 − 00 )ht−1 · (1 − δ) + δ · (1 − 00 )ht−1
n
T · (1 − ht−1 ) + 00 ht−1 T − 00 h2t−1
00
= (1 − )ht−1 ·
ht−1 − ht
2
00 h2t−1 − 00 ht−1 T
T · (ht−1 − ht−1 )
00
− ht−1 ·
= (1 − ) ·
ht−1 − ht
ht−1 − ht
(h
−
T
)
t−1
= (1 − 00 ) · T − 00 h2t−1 ·
ht−1 − ht
≥ (1 − 00 ) · T − 00 ht ≥ T − 300
In the end, we hope that
each iteration 1 ≤ i ≤ t − 1,
|Ht |
n
|Hi |
n
deviates from T by at most /2. To achieve this, we show that in
deviates from its expectation,
|Hi−1 |2
,
n2
by at most a (1 ± 0 ) factor,
where 0 = /2t+4 is an error control parameter that is much less than . Note that
deviated from
h2i−1
Lemma 2.5. Let = Ω(1/n1/4.47 ) and 0 =
∈ (1 ±
is already
|Hi |
n
|Hi−1 |2
2
n
= hi to some degree. The next lemma bounds the cumulative deviation of
from hi . Note that 0 has to be large enough in order guarantee that
range. This also implies has to be large enough.
|Hi |
n
|Hi−1 |2
n2
2i+1 −1
0 )
.
2t+4
|Hi |
n
lies in the (1 ± 0 ) ·
W.h.p. for iteration 0 ≤ i < t, we have
· hi .
Proof. We will show by induction. Initially, we have
(1 − 0 )h0 ≤ (1 − 1/n) · h0 ≤
Suppose that
|Hi |
n
i+1 −1
∈ (1 ± )2
|H0 |
≤ (1 + 1/n)h0 ≤ (1 + 0 )h0
n
· hi is true. Then by Lemma 2.3,
|Hi |2
n2
i+2
≥ (1 − 0 )2 −2 · h2i
E[|Hi+1 | | Hi ] =
≥ (1 − 0 · (2t+2 − 2)) · h2i
≥ (1 − ) · h2i = Ω(1)
7
(1 − a)b ≥ 1 − ab
0 = t+4 , hi ≥ 1/2 −
2
(1)
Also,
p
0 = · 2log7/4 (/4)−2 /16 ≥ 2.24 /256 = Ω( log n/n)
(2)
Therefore, by (1), (2), and Chernoff Bound, we have
log n
|Hi+1 |
|Hi+1 |
02
0
Pr
∈ (1 ± ) · E[
| Hi ] ≤ 1−2·e−Ω( E[|Hi+1 ||Hi ]) ≤ 1−2·e−Ω( n ·n) = 1−1/ poly(n)
n
n
Thus, w.h.p.
|Hi+1 |
n
|
∈ (1 ± 0 ) · E[ |Hi+1
| Hi ].
n
|Hi+1 |
≤ (1 + 0 ) ·
n
|Hi |
n
2
2
i+1
≤ (1 + 0 ) · (1 + 0 )2 −1 · hi
= (1 + 0 )2
Similarly, we can show
i+1
(1 ± 0 )2 −1 · hi .
|Hi |
n
≥ (1 − 0 )2
i+2 −1
i+2 −1
· hi+1
· hi+1 . Therefore, we conclude for 0 ≤ i < t,
|Hi |
n
∈
Combining Lemma 2.4 for the deviation on the last round and Lemma 2.5 for the deviation on
first t − 1 rounds, the following lemma summarizes the final deviation.
Lemma 2.6. Let = Ω(1/n1/4.47 ). At the end of the algorithm, w.h.p. T −
2
≤
|Ht |
n
≤ T + 2 .
|
Proof. Let 0 = 2t+4
. By Lemma 2.5, we have |Ht−1
∈ (1 ± 0 )2 −1 · ht−1 w.h.p.
n
t
Note that (1+0 )2 −1 ≤ 1+2(2t − 1)0 ≤ 1+2t+1 0 , since (1+a)b ≤ 1+2ab provided 0 < ab ≤ 1/2
t
and we have (2t − 1)0 ≤ /16 ≤ 1/2. Similarly, (1 − 0 )2 −1 ≥ 1 − 2t+1 0 . Therefore, we can let
00 = 2t+1 0 and apply Lemma 2.4 to conclude that
|Ht |
t+1 0
00
T − 3 · 2 ≤ T − 3 ≤ E
| Ht−1 ≤ T − 300 ≤ T + 3 · 2t+1 0
n
p
|
By Lemma 2.5, (1), and (2), we have |Ht−1
= Ω(1) and 0 = Ω( log n/n). Therefore, we can
n
apply Chernoff Bound to show w.h.p. |Ht | ∈ (1 ± )E[|Ht | | Ht−1 ]. Thus,
t
|Ht |
≤ (1 + 0 )(T + 3 · 2t+1 0 )
n
|Ht |
T − 4 · 2t+1 0 ≤
≤ T + 4 · 2t+1 0
n
|Ht |
T− ≤
≤T +
2
n
2
(1 − 0 )(T − 3 · 2t+1 0 ) ≤
0 = /(2t+4 )
Initially, |Mn0 | ≥ 2. The following three lemmas are for showing Lemma 2.10, which states that
|Mt |
|Mi |
n does not decrease much from 2. This is done by showing that n does not decrease in each
round, except possibly the last. We first derive bounds on the expectation of |Mni+1 | in terms |Mni |
in each round.
Lemma 2.7. For 0 ≤ i < t − 1, suppose that
|Mi |
n · (1 + 2).
|Hi |
n
8
≥ 1/2 and
|Mi |
n
≥
|M0 |
n ,
E[ |Mni+1 | | Mi , Hi ] ≥
Proof. Therefore, we have
!
|Mi | |Hi |
|Mi | 2
2·
·
+
n
n
n
v∈V
|Hi | |Mi |
1 X |Mi |
2·
+
= ·
n
n
n
n
E[|Mi+1 | | Mi , Hi ]
1 X
= ·
n
n
v∈V
1 X |Mi |
≥ ·
· (1 + 2)
n
n
|Hi |
≥ 1/2, |Mi | ≥ |M0 | ≥ 2n
n
v∈V
≥
|Mi |
· (1 + 2)
n
Lemma 2.8. Suppose that
|Ht−1 |
n
≥
1
2
− 3
2 and
|Mt−1 |
n
≥
|M0 |
n ,
E[ |Mnt | | Mt−1 , Ht−1 ] ≥ (1 − ) · |Mnt−1 | .
Proof.
!
!
1 X
|Mt−1 | |Ht−1 |
|Mt−1 | 2
|Mt−1 |
E[|Mt | | Mt−1 , Ht−1 ]
= ·
δ· 2·
·
+
+ (1 − δ) ·
n
n
n
n
n
n
v∈V
|Mt−1 |
|Ht−1 | |Mt−1 |
|Mt−1 |
=δ·
2·
+
+ (1 − δ) ·
n
n
n
n
|Mt−1 |
1 3 |Mt−1 |
|M0 |
|Mt−1 |
· (1 − ) + (1 − δ) ·
− ,
≥
≥ 2
≥δ·
n
n
2
2
n
n
|Mt−1 |
≥ (1 − ) ·
n
Lemma 2.9. For 0 ≤ i ≤ t − 1, suppose that
02
1 − 2 · e−Ω( E[|Mi+1 ||Mi ,Hi ]) .
|Mi |
n
≥
|M0 |
n ,
Pr( |Mni+1 | ∈ (1 ± 0 ) · E[ |Mni+1 | | Mi , Hi ] ≤
Proof. Since the events that v ∈ Mi+1 are independent for all v, by Chernoff Bound,
|Mi+1 |
E[|Mi+1 | | Mi , Hi ]
02
0
≤ e−Ω( E[|Mi+1 ||Mi ,Hi ])
Pr
> (1 + ) ·
n
n
|Mi+1 |
E[|M
|
02
i+1 | Mi , Hi ]
0
Pr
< (1 − ) ·
≤ e−Ω( E[|Mi+1 ||Mi ,Hi ])
n
n
Lemma 2.10. Let = Ω(1/n1/4.47 ). At the end of the algorithm, w.h.p.
|Mt |
n
≥
7
4.
Proof. First note that for 0 ≤ i ≤ t − 2, w.h.p. we must have |Hni | ≥ 1/2. Otherwise, by Lemma
|
2.3 and Chernoff Bound with error 0 = 1/3, w.h.p. we have |Hi+1
≤ (1 + 0 )(1/2)2 ≤ 1/2 − . This
n
implies t = i + 1, a contradiction occurs.
We will show by induction that |Mni | ≥ |Mn0 | for 0 ≤ i ≤ t − 1 first. Initially, |Mn0 | ≥ |Mn0 | is true.
Suppose that |Mni | ≥ |Mn0 | is true.
|Mi+1 |
< (1 − 0 ) · E[|Mi+1n||Mi ,Hi ]
n
w.h.p. |Mni+1 | ≥ (1−0 )· E[|Mi+1n||Mi ,Hi ]
Let 0 = /4. By Lemma 2.9 and Lemma 2.7, we have Pr
e
−Ω(02 E[|M
i+1 ||Mi ,Hi ])
−Ω(02 n)
=≤ e
= 1/ poly(n). Therefore,
9
≤
≥
(1 − /4) · (1 + 2) · |Mni | ≥ |Mni | . From this statement, we can also obtain that in the end of round
t − 1,
|Mt−1 |
|Mt−2 |
|M0 |
≥ (1 − ) · (1 + 2) ·
≥ (1 − ) · (1 + 2) ·
n
4
n
4
n
Also let 00 = /2t+4 . By Lemma 2.5, w.h.p. we have
|Ht−1 |
1 3
i+1
≥ (1 − 00 )2 −1 · ht−1 ≥ (1 − 2 · (2i+1 − 1)00 ) · ht−1 ≥ ht−1 · (1 − /2) ≥ T − ≥ −
n
2
2
2
This satisfy the conditions required for Lemma 2.8. By Lemma 2.9 and Lemma 2.8, we have
w.h.p.
|Mt |
E[|Mt | | Mt−1 , Ht−1 ]
≥ (1 − 0 ) ·
n
n
|M
t−1 |
≥ (1 − 0 ) · (1 − ) ·
n
|M0 |
≥ (1 − /4)2 · (1 + ) ·
n
|M0 |
≥ (1 − ) · (1 + ) ·
n
|M0 |
2
≥ (1 − ) ·
n
7
≥
4
(1 − /4)2 ≥ 1 − for ≤ 1/2.
< 1/8 and
|M0 |
n
= 2
The following lemma shows that at the end of Algorithm 1, the problem has been reduced to
finding the approximate median.
Lemma 2.11. Let = Ω(1/n1/4.47 ). At the end of iteration t of Algorithm 1, w.h.p. any φ0 -quantile
where φ0 ∈ [ 21 − 4 , 12 + 4 ] must be in Mt .
|Ht |
|Mt |
1
≥
Proof. Since 12 − 3
2 ≤ n ≤ 2 − 2 and n
|Ht |
|Mt |
1
n + n ≥ 2 + 4 . Combined with the fact that
where φ0 ∈ [ 12 − 4 , 21 + 4 ] must be in Mt .
2.2
7
4 by Lemma 2.5 and Lemma 2.10, we have
|Ht |
1
0
n ≤ 2 − 2 , we conclude that any φ -quantile
Phase II: Approximating the Median
Let Li , Mi , and Hi denote the nodes whose quantiles lie in [0, 12 − ), [ 12 − , 12 + ], and ( 12 − , 1]
respectively at the end of iteration i, and L0 , M0 , and H0 be the nodes with those quantiles in the
beginning. Note that Li and Hi are the nodes whose values are not our targets. We will show the
quantities of |Lni | and |Hni | decrease in each iteration as our 3-TOURNAMENT algorithm (Algorithm 2)
makes progress.
Initially, l0 = h0 = 21 − . Let hi+1 = 3h2i − 2h3i and li+1 = 3li2 − 2li3 for i ≥ 0, we will show that
|Li |
|Hi |
n and n concentrate around li and hi . Note that hi and li roughly square in each iteration.
Once they decrease below a constant after the first O(log(1/)) iterations, they decrease double
exponentially in each iteration. The tournaments end when li and hi decrease below T = 1/n1/3 .
10
Algorithm 2 3-TOURNAMENT(v)
1:
2:
3:
4:
5:
6:
7:
8:
h0 , l0 ← 12 −
i ← 0, T = 1/n1/3 .
while li > T do
hi+1 ← 3h2i − 2h3i , li+1 ← 3li2 − 2li3
Select three nodes t1 (v), t2 (v), t3 (v) randomly.
xv ← median(xt1 (v) , xt2 (v) , xt3 (v) ).
end while
Sample K = O(1) nodes uniformly at random and output the median value of these nodes.
1
Lemma 2.12. Let t denote the number of iterations in Algorithm 2. We have t ≤ log11/8 ( 4
)+
log2 log4 n = O(log(1/) + log log n).
i−1 ≤ 1/4, we will show that l ≤ 1 − ( 11 )i . First, l = 1 − . Suppose
Proof. Suppose that ( 11
i
0
8 )
8
i−1 . We have:
)
by induction that li−1 ≤ 1 − ( 11
8
2
3
li = 3li−1
− 2li−1
i−1 !2
i−1 !3
1
11
1
11
=3
−
· −2
−
·
2
8
2
8
i−1 !3
1 3 11 i−1
11
= −
·+2
·
2 2 8
8
i−1 !
1
1 3 11 i−1
11
·+
·
≤ −
2 2 8
8
8
i
1
11
= −
·
2
8
1
Therefore, after i0 = log11/8 ( 4
) iterations, we have li ≤
Suppose that i1 = log2 log4 n, we have
1
2
−
11 i0
8
≤
1
2
−
1
4
= 41 .
li0 +i1 ≤ 3li20 +i1 −1
i1
≤ 3i1 li20
2i1
1
i1
≤3 ·
4
log2 3
≤ (log4 n)
=
log4 n
1
·
4
(log4 n)log2 3
≤ n2/3
n
1
) + log2 log4 n = O(log(1/) + log log n).
Therefore, t ≤ i0 + i1 = log11/8 ( 4
|
|
Lemma 2.13. For each iteration 0 ≤ i < t, E[ |Li+1
| Li ] = 3( |Lni | )2 − 2( |Lni | )3 and E[ |Hi+1
| Hi ] =
n
n
|Hi | 3
|Hi | 2
3( n ) − 2( n ) .
11
|
|Hi+1 |
Proof. We will show the proof for E[ |Li+1
| Hi ] is the same. Note that a node v
n |Li ], since E[ n
is in Li+1 if and only if at least 2 ot t1 (v), t2 (v), t3 (v) are in Li . Therefore,
!
1X
|Li |
|Li | 2
|Li | 3
|Li | 2
|Li | 3
|Li+1 |
1−
=3
| Li =
+3
−2
E
n
n
n
n
n
n
n
v∈V
The following lemma shows the probabilities that |Lni | and |Hni | deviate from their expectation by
a (1 + 0 ) factor are polynomially small in n, provided 0 is sufficiently large. We cap the quantity
at T for the purpose of applying concentration inequalities.
1/2
n)
Lemma 2.14. Let 0 = Ω( (logn1/3
max(T, E[ |Hni | | Hi−1 ]) and
|Li |
n
|Hi |
n
), w.h.p. for each iteration 0 ≤ i <= t,
≤ (1 + 0 ) ·
≤ (1 + 0 ) · max(T, E[ |Lni | | Li−1 ]).
Proof. In each iteration, since each node set its value independently, by Chernoff Bound (see Lemma
B.2), we have
|Li |
E[|Li | | Li−1 ]
0
Pr
≤ (1 + ) · max(T,
) ≤ exp(−Ω(02 max(n · T, E[|Li | | Li−1 ])))
n
n
= exp(−Ω(02 n2/3 )) = 1/ poly(n)
The proof for |Hi | is the same.
Then we bound the cumulative deviation of
from li .
|Hi |
n
from hi and the cumulative deviation of
1/2
n
). W.h.p. for each iteration 0 ≤ i <= t,
Lemma 2.15. Let 0 = Ω( logn1/3
li , (1 + 0 )T ) and
|Hi |
n
≤ max((1 + 0 )
3i −1
2
|Li |
n
≤ max((1 + 0 )
3i −1
2
·
· hi+1 , (1 + 0 )T ).
Proof. We only show the proof for |Li | and we will prove by induction. Initially,
3i −1
|Li |
n
|Li |
n
|L0 |
n
≤ l0 . Suppose
that
≤ max((1 + 0 ) 2 · li+1 , (1 + 0 )T ) is true.
|
||Li ]
||Li ]
By Lemma 2.14, we have w.h.p., |Li+1
≤ (1 + 0 ) · max( E[|Li+1
, T ). If E[|Li+1
≤ T , then
n
n
n
|Li+1 |
E[|Li+1 ||Li ]
|Li |
0
0
we are done, since n ≤ (1 + ) ·
= (1 + ) · T . Also, if n ≥ (1 + 0 )T , then
n
E[|Li+1 ||Li ]
n
≤3·
|Li |2
n2
|Li |
n
≤
|Li |
n
1
· O( n2/3
) ≤ T.
Otherwise, if
> (1 + 0 )T then it must be the case that
hypothesis. Therefore,
|Li |
n
≤ (1 + 0 )
3i −1
2
· li+1 by induction
|Li+1 |
E[|Li | | Li−1 ]
≤ (1 + 0 ) ·
n
n
2
|L
|Li | 3
i|
0
≤ (1 + ) · 3
−2
n
n
2
3 !
i
i
0 3 2−1
0
0 3 2−1
≤ (1 + ) · 3 (1 + )
· li − 2 (1 + )
· li
0
≤ (1 + ) ·
0
(1 + )
3i+1 −3
2
· 3·
12
li2
−2·
li3
=
0
(1 + )
3i+1 −1
2
· li+1
Now, by setting 0 roughly equal to 1/3t ∼ 3.45 /(log4 n)1.59 , we can bound the final deviations
of |Hni | and |Lni | from T by a factor of 2 w.h.p.
0.61
n
Lemma 2.16. Let = Ω( log
) and 0 = 3.45 /(log4 n)1.59 , then
n0.096
3.45
|Lt |
n
≤ 2T and
1/2
Proof. First, 0 = Ω( log 1.59 n ) = Ω( logn1/3 n ). By Lemma 2.15, w.h.p. we have either
|Ht |
n
≤ 2T w.h.p.
|Lt |
n
≤ (1 + 0 ) · T
3i −1
or |Lnt | ≤ (1 + 0 ) 2 · lt . If it is the former, then we are done since |Lnt | ≤ (1 + 0 ) · T ≤ 2T .
Otherwise, we have
3t −1
|Lt |
·T
≤ (1 + 0 ) 2
n
t
0 3 −1
≤ 1+2· ·
·T
since (1 + x)n ≤ 1 + 2nx for nx ≤ 12
2
1
≤ 1 + 0 · 3log11/8 ( 4 )+log2 log4 n · T
!
log11/8 3
1
log2 3
0
≤ 1+ ·
· log4
n ·T
4
0 = 3.45 /(log4 n)1.59
≤2·T
Finally, we show that when |Lnt | and |Hnt | are O(1/n2/3 ), if we sample a constant number of values
randomly and output the median of them, then w.h.p. the median is in Mt .
Lemma 2.17. W.h.p. every node outputs a quantile in [ 21 − , 12 + ].
Proof. By Corollary 2.16, w.h.p. at the end of iteration t, |Lnt | ≤ 2/n2/3 and |Hnt | ≤ 2/n2/3 . The
algorithm outputs a quantile in [ 12 − , 12 + ] if there are less than K/2 nodes in Lt are sampled
and less than K/2 nodes in Ht are sampled.
The probability that at least K/2 nodes in |Lt | are sampled is at most
K/2
K
2
·
≤
K/2
n2/3
eK
K
2
!K/2
K/2
4e K/2
2
≤
·
n2/3
n2/3
K/2
Similarly, the probability that at least K/2 nodes in |Ht | are sampled is also at most n4e
.
2/3
By an union bound, the probability that less than K/2 nodes in Lt are sampled and less than K/2
K/2
nodes in Ht are sample is at least 1 − 2 · n4e
= 1 − 1/ poly(n).
2/3
Theorem 2.1 follows from Lemma 2.11 and Lemma 2.17.
3
Exact Quantile Computation
To fill the gap of approximating the φ-quantile with error for = O(1/n0.096 ), we show that the
exact quantile computation can be done in O(log n) rounds using O(log n) message size. Since we
def
are to compute the exact quantile φ, we can assume that k0 = φ · n is an integer. The problem
is to compute a value whose rank is k0 . Again, w.l.o.g. we assume that every value has a distinct
value initially.
13
Algorithm 3 Exact Quantile Computation
1: k0 ← φ · n
2: for i = 1, 2, . . . , 25 do
3:
Each node v computes an 2 -approximate of the ( ki−1
n − 2 )-quantile and an 2 -approximate of
−0.05 /2.
the ( ki−1
n + 2 )-quantile with = n
4:
Each node learns the max and the min of these approximates of all nodes.
5:
Compute the rank of min among the original xv ’s and denote it by R.
6:
Each node v set xv ← ∞ if xv ∈
/ [min, max]. Call these nodes valueless, otherwise valued.
7:
Let mi be the smallest power of 2 that is larger than (n0.99 /2)/(# valued nodes). Each valued
node makes mi copies of its value and distribute them to valueless nodes so that there are
at least n0.99 /2 valued nodes. (For convenience, we let the duplicated values to have smaller
ranks than the original one.)
8:
Set ki ← mi · (ki−1 − R + 1).
9: end for
k
10: Every node outputs an (/3)-approximate ( n25 − 2 )-quantile.
The following is a detailed implementation of each step.
Step 3: Since = n−0.05 /2, we can 2 -approximate the quantiles in O(log n) rounds by Theorem
2.1.
Step 4: The maximum (the minimum) can be computed by having each node forwarding the
maximum (the minimum) value it has ever received. Since it takes O(log n) rounds to spread a
message by [FG85, Pit87], this step can be done in O(log n) rounds.
Step 5: The rank of the minimum can be computed by performing a counting. The nodes whose
xv values are less than or equal to the minimum will be assigned 1. Otherwise they are assigned 0.
By Kempe et al. [KDG03], the sum can be aggregated in O(log n) rounds w.h.p.
Step 7: Consider the following process for distributing the values. Initially, every valued node
v generates a token with weight equal to mi , the targeted number of copies. We denote the token
by a value-weight pair (xv , mi ). Recall that mi is a power of 2 and it can be computed by counting
the number of valued nodes in O(log n) rounds. The goal is to split and distribute the tokens so
that every node has at most one token of weight 1 in the end.
The process consists of O(log n) phases and each phase uses O(1) rounds. Suppose that node v
holds tokens (x1 , w1 ), (x2 , w2 ), . . . (xm , wm ) at the beginning of a phase. For each token (xi , wi ), if
wi 6= 1, v splits (xi , wi ) into two tokens (xi , wi /2) and push each one to a node randomly. If wi = 1,
then the token remains at v. Note that when two tokens of the same value are on the same node,
they do not merge.
First note lg mi = O(log n) phases are needed to split to tokens into tokens of weight 1. Now,
we show that it takes constant number of rounds to implement one phase.
Let N (v, i) denote the number of tokens at v at the end i’th phase. Let N = maxv,i N (v, i).
It takes N rounds to implement one phase. The value of N (v, i) can be bounded by following
calculation. Since at any phase, there cannot be more than n0.99 tokens and each token appears at
a node chosen uniformly at random, we have
Pr(N (v, i) ≥ 100K) ≤
n0.99
100K
·
1
n100K
≤
n0.99
n
100K
= (1/n)K .
By taking an union bound over each v and i, we conclude that N < 100K w.h.p. Therefore,
14
w.h.p. each phase can be implemented in 200K = O(1) rounds. At the end of O(log n) phase, every
node has at most 100K tokens of weight 1 w.h.p.
Next, in the subsequent phases, if a node holds more than one token, then it will push every
token except one to a random node (one random node per token). We say a token succeeded at
the end of the phase if it was pushed to a node without any other tokens. Consider a token, the
probability a token did not succeed in a phase is at most n0.99 /n = 1/n0.01 , since there are at most
n0.99 nodes with tokens at the end of the phase.
Therefore, after 100K phases, the probability that a token did not succeed is at most
1/n0.01·100K = 1/nK . By a union bound over the tokens, we conclude that w.h.p. all tokens
succeeded. Moreover, again, each phase can be implemented using 100K rounds. Therefore, we
can split and distribute the tokens in O(log n) rounds. Each original token is duplicated with exactly mi copies. Then, those nodes holding a token set its value to the value of the token. Nodes
without a token will remain valueless.
Q
Correctness:
Let ans be the answer (the value whose initial rank is k0 ). Let Mi = ij=1 mi
denote the number of copies of each value from the beginning. We show by induction that the
values whose ranks lying in (ki − Mi , ki ] are ans after iteration i. Initially, M0 = 1, the statement
trivially holds.
Suppose that at the end of iteration i − 1, the values whose ranks lying in (ki−1 − Mi , ki−1 ] are
ans. After Step 6 in iteration i, the rank of ans is exactly ki−1 − R + 1. Since every value is
duplicated to have mi copies after Step 7, the rank of ans becomes ki = mi · (ki−1 − R + 1) (Recall
that we let the original value to have a larger rank than the duplicated values). Since the values
whose ranks lying in (ki−1 − Mi−1 , ki−1 ] were ans at the end of iteration i − 1 and ans ∈ [min, max],
it must be the case that every value in this range is duplicated with mi copies. Therefore, all values
whose ranks lying in (ki − mi · Mi−1 , ki ] = (ki − Mi , ki ] are ans after iteration i.
Next, we show that after 25 iterations, M25 ≥ n. Suppose that Mi−1 ≤ n. The number of
valueless node after Step 6 is at most 2n + 2Mi−1 ≤ 4n. The 2n comes from the fact that at
most 2n values can lie in the range (min, max). The 2Mi−1 term is because at most Mi−1 values
are equal to min and max. Therefore, mi ≥ (n0.99 /2)/(4n) = (n0.04 /4) and Mi ≥ (n0.04 /4) · Mi−1 .
Therefore, M25 ≥ min(n, (n0.04 /4)25 ) = n. Now the the values of the items whose ranks are
in the range of (k25 − n, k25 ] must be ans. We then compute a quantile in ( kn25 − , kn25 ] using our
/3-approximate quantile computation algorithm to approximate the ( kn25 − 2 )-quantile in O(log n)
rounds.
4
A Lower Bound
We show that it takes at least Ω(log(1/) + log log n) rounds to approximate the φ-quantile up to
error w.h.p.
n
< < 1/8 and φ ∈ [0, 1], any gossip algorithm that uses less
Theorem 1.3. For any 10 log
n
1
than 2 log log n or less than log4 8 round fails to solve the -approximate φ-quantile problem with
probability at least 1/3. This remains true even for unlimited message sizes.
Proof. Consider the following two scenarios. The first is when each node is associated with a
distinct value from {1, 2, . . . , n}. The second is when each node is associated with a distinct value
from {1 + b2nc, . . . , n + b2nc}.
15
def
A node is able to distinguish between these cases only if it recieves an value from S =
{1, 2, . . . , 1 + b2nc} ∪ {n + 1, n + 2, . . . , n + b2nc}. Otherwise, it can only output a value whose
quantile is in [1/2 − , 1/2 + ] with probability 1/2, since the difference of the φ-quantile in both
scenarios is at least b2nc ≥ n.
Call a node good if it ever received a value from S, and bad otherwise. Note that a bad node cannot
outputs a correct answer with probability more than 1/2. Initially there are at most 2 · b2nc ≤ 4n
good nodes. We will show that with probability 1 − 1/n there exists at least one bad node at the
end of round t.
Let X0 = 2 · b2nc and let Xi denote the number of good nodes at the end of round i. Given a
bad node v, it can become good if it pulls from a good node or some good node pushes to it. Let
Yv denote the event that v pulls from a good node, we have Pr(Yv | Xi ) = Xi /n. Also, the pushes
from the good nodes can only generate another at most Xi good nodes. Therefore,
X
P
E[Xi+1 | Xi ] ≤ 2Xi + E
Yv ≤ 3Xi ,
since E[ v∈Bi Yv | Xi ] ≤ Xi
v∈Bi
P
By Chernoff Bound, we have Pr( v∈Bi Yv > 2Xi | Xi ) ≤ e−Xi /2 ≤ e−5 log n ≤ 1/n5 , since
Xi ≥ X0 ≥ 10 log n. Therefore, with probability at least 1 − 1/n5 , Xi+1 ≤ 4Xi . By taking a union
bound over such events for the first t0 = log4 (8/) rounds, we conclude with probability at least
0
1 − 1/n4 , Xt0 ≤ (4n)4t ≤ n/2.
Let t0 be the last round such that Xt0 ≤ n/2. Define Zi = |Bi |/n. A node v remains in Bi if it
did not pull from a good node and it was not pushed from any good nodes. Denote the event by
Wv , we have Pr(Wv | Bi ) ≥ Zi · (1 − n1 )n−1 ≥ Zi · e−1 . Therefore,
X
X
Zi · e−1 = |Bi | · Zi · e−1
Wv | Bi ≥
E[|Bi+1 | | Bi ] = E
v∈Bi
v∈Bi
Note that the events {Wv }v∈Bi are negatively dependent [DR98] (the number of empty bins in
the balls into bins problem). Suppose that Zi · |Bi | ≥ 60e log n. By Chernoff Bound,
Pr (|Bi+1 | ≤ |Bi | · Zi /(2e) | Bi ) ≤ e−Zi ·|Bi |/(12e) ≤ 1/n5
Therefore, Pr(Zi+1 ≤ Zi2 /(2e) | Zi ) ≤ 1/n5 . Suppose that t1 = O(1). We can take an union bound
over the subsequent t1 rounds to show that with probability at least 1 − 1/n4 , Zi+1 ≥ Zi2 /(2e)
holds for these rounds, as long as Zt20 +t1 ≥ (60 log n)/n.
Let zt0 = 1/2 and zi+1 = zi2 /(2e). If zt20 +t1 ≥ (60 log n)/n, with probability at least 1 − 1/n4 , we
have Zt0 +t1 ≥ zt0 +t1 . Let t1 = 12 lg log n, by definition of zi , we have
t1
1
t1
zt0 +t1 =
· zt20
2e
t1 +2t1
1
zt0 = 1/2 ≥ 1/(2e)
≥
2e
√
(lg log n)/2+ log n
1
=
2e
2√log n
1
≥
for sufficiently large n
2e
16
√
√
Since (2e)−2 log n = Ω((log n)/n), zt0 +t1 ≥ 2−2 log n ≥ (60 log n)/n for sufficiently large n.
Therefore, with probability at least 1 − 1/n4 , Zt0 +t1 > 0.
Since with probability 1 − 1/n4 , t0 ≥ log4 8/. By taking an union bound over these two events,
we conclude with probability at most 2/n4 , all nodes are good by the end of round t0 + t1 .
Therefore, the probability that every node computes a correct output at the end of round t0 + t1
is at most 1/2 + 2/n4 .
5
Robustness
In this section, we show that our algorithm is robust against failures. We consider the following
model of failures: Let 0 < µ < 1 be a constant. Every node v in every round i is associated with
a probability 0 ≤ pv,i ≤ µ. Note that pv,i is pre-determined before the execution of the algorithm.
During the execution of round i, each node v fails with probability pv,i to perform its operation
(which may be either push or pull).
Theorem 1.4. Suppose in every round every node fails with a, potentially different, probability
bounded by some constant µ < 1. For any φ ∈ [0, 1] there is a gossip algorithm that solves the φ
quantile problem in O(log n) rounds. For any t and any (n) > 0, there furthermore exists a gossip
1
algorithm that solves the -approximate φ-quantile problem in O(log log n + log (n)
+ t) rounds for
n
all but 2t nodes, with high probability.
We show how to modify the algorithm and the analysis for the tournament algorithms and the
exact quantile computation algorithms in the following.
5.1
The Tournament Algorithms
First consider the 2-TOURNAMENT algorithm. Initially, every node is good. Consider node v, instead
1
1
of only pulling from 2 neighbors each iteration, now pull from Θ( 1−µ
· log( 1−µ
)) neighbors. We say
a pull is good if the node performing the pull operation did not fail and it pulls from a node who
is good at the end of iteration i − 1. A node remains good at the end of iteration i if there are at
least two good pulls. Then, v uses the first two good pulls to execute the tournament procedure.
Also, note that in the last iteration, we let a good node to have probability of δ to do the two
tournament using the first two good pulls and probability 1 − δ to set the value equal to the first
good pull. We show that the good nodes always consists of a constant fraction of nodes.
Lemma 5.2. For 0 ≤ i ≤ t − 1, w.h.p. at the end of each iteration i of 2-TOURNAMENT , there are
1
1
at least n/2 good nodes if every node pulls from k = Θ( 1−µ
· log( 1−µ
)) other nodes.
Proof. We will prove by induction. Initially, every node is good. Suppose that there are at least
n/2 good nodes at the end of iteration i − 1. During iteration i, a node with more than k − 2 bad
pulls will become bad at the end of iteration i.
4
4
Let k = 1−µ
log 1−µ
+ 1. The probability that there are at least k − 1 bad pulls is at most:
(1−µ)
k
(1 − µ) k−1
· 1−
≤ k · e− 2 ·(k−1)
k−1
2
4
4
1−µ 2
≤
log
+1 ·
1−µ
1−µ
4
≤ 1/e + 1/16 ≤ 0.44
x log(1/x) maximized at 1/e
17
Therefore, the expected number of bad nodes is at most 0.44n for iteration 1 ≤ i ≤ t − 1. Since
each node becomes bad independently of other nodes, by Chernoff Bound, w.h.p. there are at most
0.5n bad nodes at the end of iteration i.
At the end of iteration t − 1, there are at least
P n/2 good nodes w.h.p. The expected number of
bad nodes at the end of iteration t is at most v∈V δ · 0.44 + (1 − δ)/2 ≤ 0.5n. By Chernoff Bound
again, we can conclude that w.h.p. there are at least n/3 good nodes.
We can also modify the process in the same way for the 3-TOURNAMENT algorithm. That is,
1
1
in each iteration each node pulls from Θ( 1−µ
log 1−µ
) other nodes. If there are less than 3 good
pulls, then the node becomes bad. Otherwise, it uses the first 3 good pulls to do the tournament
procedure. We can show similarly that there are at most a constant fraction bad nodes in each
iteration.
Consider a node v in the modified processes. Suppose that v is good in iteration i, then v must
have at least two (or three) good pulls. Note that the probability that v pulls from a particular
good node, conditioned v is good at the end of iteration, is uniform among all nodes that are good
at the end of iteration i − 1.
Let Vi denote the set of good nodes and ni = |Vi |. Given any subset of good nodes S ⊆ Vi−1 ,
the probability of choosing a node in |S| is therefore |S|/ni . Therefore, we can replace |Lni | , |Hni | ,
|Mi |
i|
and |Mni | in our proofs in the previous section with |Lnii | , |H
ni , and ni and observe that all the
statements hold. Note that ni ≥ n/3 = Ω(n) for 0 ≤ i ≤ t w.h.p. Thus, all the concentration
inequalities also hold.
K
K
In the last step of Algorithm 2, all nodes pulls from Θ( 1−µ
log 1−µ
) nodes. If there are K good
pulls, then each node outputs the median directly. Otherwise, it becomes bad and outputs nothing.
We can similarly show that there are at least constant fraction of good nodes. Therefore, at least
a constant fraction of nodes output a correct answer and all the others output nothing. Note that
we can use additional O(t) rounds of pulling from valueless nodes to have all but 2nt nodes learn a
correct answer.
5.2
Exact Quantile Computation
Consider Algorithm 3. We know Step 3 (Section 5.1), Step 4 [ES09], and Step 5 [KDG03] tolerate
such failures with a constant factor delay on the running time.
The only step that remains to discuss is Step 7. We run the same algorithm described. Initially
each valued node v generates a token (xv , mi ). In the first O(log n) phases, each node tries to split
every token whose weight is larger than 1 into two tokens with halved weight, and push one of them
to a random node. If the push succeed, the token has been pushed is considered a new token and
the one with halved weight remaining at the same node is considered as the old one. If the push
fails, then we will merge them back to the original one.
We will show that at the end of O(log n) phase, each node holds at most O(1) tokens and their
weights are 1. In each subsequent phases, if a node has more than one tokens, then it will push
every token except one to a random neighbor. We argue that after O(log n) phases, every node
contains exactly one token.
First, we show that the number of tokens at each node is bounded by a constant w.h.p. so that
each phase can be implemented in O(1) rounds. Recall that N (v, i) is the number of tokens at v
at the end of i’th phase. Suppose the total number of phases is C log n for some constant C > 0.
For a token (except the one that is initially at v) to be at node i, it must be the case that it was
pushed to i during one of the phase. Note that at any phase, there are at most n0.99 tokens. By
18
taking an union bound over all possible set of tokens of size 200K and all possible combinations of
phases on when each of the token was pushed to i, we have
n0.99
200K
Pr(N (v, i) ≥ 200K + 1) ≤
· (C log n)
200K
·
1
n200K
≤
Cn0.99 log n
n
200K
≤ 1/nK .
Next, we show that for a token (xv , mi ), it takes O(log n) rounds to split into tokens of weight
1 w.h.p. Let
PTv (i) denote the set of all v’s tokens with weight at least 2 at the end of Phase i.
Let Φ(i) = (xv ,w)∈Tv (i) w2 be a potential function. Consider a token (xv , w) ∈ Tv (i), since with
probability at most µ it fails to split, the expected contribution of the token to the Φ(i + 1) is at
most:
1−µ
2
2
µw + (1 − µ)2 · (w/2) = 1 −
· w2
2
Therefore, E[Φ(i + 1) | Φ(i)] ≤ (1 − (1 − µ)/2) · Φ(i). Since Φ(0) ≤ n, after t =
we have
2K
1 − µ 1−µ log n
E[Φ(t)] ≤ 1 −
· n ≤ e−K log n · n ≤ 1/nK−1
2
2K
1−µ
log n rounds,
Since Φ(t) must be an integer, we conclude that with probability at least 1 − 1/nK−1 , Φ(t) = 0
by Markov’s inequality. Therefore, after t rounds w.h.p. every token of value v has weight 1. We
can further take an union over all tokens to show that it holds for all tokens w.h.p.
Therefore, w.h.p. the weight of all tokens are 1 at the end of O(log n) round. In the subsequent
phases, the probability a token fail to be pushed to a node without any other tokens is at most
µ + n0.99 /n = O(1). Therefore, the probability a token fails in the all of the next O(log n) phases
is at most 1/ poly(n). Thus, w.h.p. after O(log n) phases, every node has at most one token. Since
each phase can be implemented using O(1) rounds, we conclude that Step 7 can be done in O(log n)
rounds in our failure model.
Acknowledgement
We thank Frederik Mallmann-Trenn for pointing out [DGM+ 11] to us.
References
[ACH+ 13] Pankaj K. Agarwal, Graham Cormode, Zengfeng Huang, Jeff M. Phillips, Zhewei Wei,
and Ke Yi. Mergeable summaries. ACM Trans. Database Syst., 38(4):26:1–26:28, 2013.
[ARS97]
Khaled Alsabti, Sanjay Ranka, and Vineet Singh. A one-pass algorithm for accurately
estimating quantiles for disk-resident data. In Proc. 23rd Int’l Conference on Very
Large Data Bases (VLDB), pages 346–355, 1997.
[BFP+ 73] Manuel Blum, Robert W. Floyd, Vaughan Pratt, Ronald L. Rivest, and Robert E.
Tarjan. Time bounds for selection. J. Comput. Syst. Sci., 7(4):448–461, 1973.
[CP12]
Jen-Yeu Chen and Gopal Pandurangan. Almost-optimal gossip-based aggregate computation. SIAM Journal on Computing, 41(3):455–483, 2012.
19
[CT87]
Francis Chin and H. F. Ting. An improved algorithm for finding the median distributively. Algorithmica, 2(1):235–249, 1987.
[DGH+ 87] Alan Demers, Dan Greene, Carl Hauser, Wes Irish, John Larson, Scott Shenker, Howard
Sturgis, Dan Swinehart, and Doug Terry. Epidemic algorithms for replicated database
maintenance. In Proc. 6th ACM Symposium on Principles of Distributed Computing
(PODC), pages 1–12, 1987.
[DGM+ 11] Benjamin Doerr, Leslie Ann Goldberg, Lorenz Minder, Thomas Sauerwald, and Christian Scheideler. Stabilizing consensus with the power of two choices. In Proc. 23rd Annual ACM Symposium on Parallelism in Algorithms and Architectures (SPAA), pages
149–158, 2011.
[DR98]
Devdatt Dubhashi and Desh Ranjan. Balls and bins: A study in negative dependence.
Random Structures & Algorithms, 13(2):99–124, 1998.
[ES09]
R. Elsässer and T. Sauerwald. On the runtime and robustness of randomized broadcasting. Theor. Comput. Sci., 410(36):3414–3427, 2009.
[FG85]
A.M. Frieze and G.R. Grimmett. The shortest-path problem for graphs with random
arc-lengths. Discrete Applied Mathematics, 10(1):57 – 77, 1985.
[FHK17]
Ofer Feinerman, Bernhard Haeupler, and Amos Korman. Breathe before speaking: efficient information dissemination despite noisy, limited and anonymous communication.
Distributed Computing, 30(5):339–355, 2017.
[FO15]
David Felber and Rafail Ostrovsky. A randomized online quantile summary in
o((1/) log(1/)) words. In Approximation, Randomization, and Combinatorial Optimization. Algorithms and Techniques, APPROX/RANDOM, pages 775–785, 2015.
[FR75]
Robert W. Floyd and Ronald L. Rivest. Expected time bounds for selection. Commun.
ACM, 18(3):165–172, 1975.
[Fre83]
Greg N. Frederickson. Tradeoffs for selection in distributed networks (preliminary version). In Proc. 2nd ACM Symposium on Principles of Distributed Computing (PODC),
pages 154–160, 1983.
[GK01]
Michael Greenwald and Sanjeev Khanna. Space-efficient online computation of quantile
summaries. SIGMOD Rec., 30(2):58–66, 2001.
[GK04]
Michael B. Greenwald and Sanjeev Khanna. Power-conserving computation of orderstatistics over sensor networks. In Proc. of the 23rd ACM Symposium on Principles of
Database Systems (PODS), pages 275–285, 2004.
[GM09]
Sudipto Guha and Andrew McGregor. Stream order and order statistics: Quantile
estimation in random-order streams. SIAM J. Comput., 38(5):2044–2059, 2009.
[Hae16]
Bernhard Haeupler. Analyzing network coding (gossip) made easy. Journal of the
ACM, pages 26:1–26:22, 2016.
[Hoa61]
C. A. R. Hoare. Algorithm 63 (partition) and algorithm 65 (find). Commun. ACM,
4(7):321–322, 1961.
20
[HT10]
Regant Y. S. Hung and Hingfung F. Ting. An Ω((1/) log(1/)) space lower bound for
finding -approximate quantiles in a data stream. In Proc. 4th International Conference
on Frontiers in Algorithmics (FAW), pages 89–100, 2010.
[KDG03]
David Kempe, Alin Dobra, and Johannes Gehrke. Gossip-based computation of aggregate information. In Proc. 44th IEEE Symposium on Foundations of Computer Science
(FOCS 2003), pages 482–491, 2003.
[KDN+ 06] Srinivas Kashyap, Supratim Deb, K. V. M. Naidu, Rajeev Rastogi, and Anand Srinivasan. Efficient gossip-based aggregate computation. In Proc. of the 25th ACM Symposium on Principles of Database Systems (PODS), pages 308–317, 2006.
[KLL16]
Z. Karnin, K. Lang, and E. Liberty. Optimal quantile approximation in streams. In
Proc. 57th IEEE Annual Symposium on Foundations of Computer Science (FOCS),
pages 71–78, 2016.
[KLW07]
Fabian Kuhn, Thomas Locher, and Rogert Wattenhofer. Tight bounds for distributed
selection. In Proc. of 19th Annual ACM Symposium on Parallel Algorithms and Architectures (SPAA), pages 145–153, 2007.
[KSSV00] R. Karp, C. Schindelhauer, S. Shenker, and B. Vocking. Randomized rumor spreading.
In Proc. 41st IEEE Symposium on Foundations of Computer Science (FOCS), pages
565–574, 2000.
[MFHH02] Samuel Madden, Michael J. Franklin, Joseph M. Hellerstein, and Wei Hong. Tag: A tiny
aggregation service for ad-hoc sensor networks. SIGOPS Oper. Syst. Rev., 36(SI):131–
146, December 2002.
[MMS13]
Qiang Ma, S. Muthukrishnan, and Mark Sandler. Frugal streaming for estimating
quantiles. In Space-Efficient Data Structures, Streams, and Algorithms: Papers in
Honor of J. Ian Munro on the Occasion of His 66th Birthday, pages 77–96, 2013.
[MP80]
J.I. Munro and M.S. Paterson. Selection and sorting with limited storage. Theoretical
Computer Science, 12(3):315 – 323, 1980.
[MRL99]
Gurmeet Singh Manku, Sridhar Rajagopalan, and Bruce G. Lindsay. Random sampling
techniques for space efficient online computation of order statistics of large datasets.
SIGMOD Rec., 28(2):251–262, 1999.
[NSU97]
A. Negro, N. Santoro, and J. Urrutia. Efficient distributed selection with bounded
messages. IEEE Transactions on Parallel and Distributed Systems, 8(4):397–401, 1997.
[Pit87]
Boris Pittel. On spreading a rumor. SIAM Journal on Applied Mathematics, 47(1):213–
223, 1987.
[PS07]
Boaz Patt-Shamir. A note on efficient aggregate queries in sensor networks. Theor.
Comput. Sci., 370(1-3):254–264, February 2007.
[Rod82]
Michael Rodeh. Finding the median distributively. Journal of Computer and System
Sciences, 24(2):162–166, 1982.
21
[RSS86]
Doron Rotem, Nicola Santoro, and Jeffrey B. Sidney. Shout echo selection in distributed
files. Networks, 16(1):77–86, 1986.
[SBAS04]
Nisheeth Shrivastava, Chiranjeeb Buragohain, Divyakant Agrawal, and Subhash Suri.
Medians and beyond: New aggregation techniques for sensor networks. In Proceedings
of the 2nd International Conference on Embedded Networked Sensor Systems, SenSys
’04, pages 239–249, New York, NY, USA, 2004. ACM.
[SFR83]
Liuba Shrira, Nissim Francez, and Michael Rodeh. Distributed k-selection: From a
sequential to a distributed algorithm. In Proceedings of 2nd ACM Symposium on Principles of Distributed Computing (PODC), pages 143–153, 1983.
[SS89]
N. Santoro and E. Suen. Reduction techniques for selection in distributed files. IEEE
Transactions on Computers, 38(6):891–896, 1989.
[SSS88]
Nicola Santoro, Michael Scheutzow, and Jeffrey B. Sidney. On the expected complexity
of distributed selection. Journal of Parallel and Distributed Computing, 5(2):194 – 203,
1988.
[SSS92]
Nicola Santoro, Jeffrey B. Sidney, and Stuart J. Sidney. A distributed selection algorithm and its expected communication complexity. Theoretical Computer Science,
100(1):185 – 204, 1992.
[YG02]
Yong Yao and Johannes Gehrke. The cougar approach to in-network query processing
in sensor networks. SIGMOD Rec., 31(3):9–18, 2002.
22
A
A Sampling Algorithm
We first present a simpler algorithm with a higher message complexity of O(log2 n/2 ) and
O(log(1/) + poly(log log n)) running time. In Manku et al. [MRL99], they observed the following.
Suppose that we sample a multi-set S of nodes (with replacement) from all the node uniformly at
random. Then, the φ’th quantile of S has quantile approximately equal to φ in all the nodes.
Lemma A.1. Suppose that a multi-set S of values are sampled from the n values and |S| =
Ω(log n/2 ). Given z (which is not necessarily in S), define QS (z) to be the number of elements in
S whose values are smaller than z. We have |QS (z) − Q(z)| ≤ w.h.p.
Therefore, we can -approximate the φ-quantile or -approximate the quantile of every node in
O(log n/2 ) rounds. Each node v uniformly samples |Sv | nodes in |Sv | rounds. To approximate the
φ-quantile, output the element z ∈ S such that QS (z) is the closest to φ. Since there are at least 1/
elements in S, we must have |QS (z)−φ| ≤ . Therefore, |φ−Q(z)| ≤ |φ−QS (z)|+|QS (z)−Q(z)| ≤
2. To approximate v’s own quantile, simply output QS (xv ). We can take an union bound over
each node to show w.h.p. every node outputs correctly. The algorithm uses O(log n) message size,
since in each round each node is only pulling the value of the node contacted.
By using larger message sizes, it is possible to reduce the number of rounds. Consider the
following doubling algorithm: Each node v maintains a multi-set buffer Sv . Before the first
round, each node v sample a random node t0 (v) and set Sv (0) = {t0 (v)}. In round i, let ti (v) be
the node contacted by v in that round, node v sets Sv (i) ← Sv (i) ∪ Sti (v) (i). Let Sv denote Sv (T )
at the end of the algorithm.
Since the buffer size doubles each round after the first round, after t = O(log((log n)/2 )) =
O(log log n + log(1/)) rounds, the size of the buffer is Ω(log n/2 ). The message complexity is
O(log2 n) for this algorithm. On the surface, it seems that we can apply Lemma A.1 to show that
QSv (z) is a good approximation of Q(z), since |Sv | = Ω((log n)/2 ). However, the elements in Sv
are not independently uniformly sampled. There are correlations among the elements. In the rest
of the section, we will show that the correlations among the elements are not too strong so that
the algorithm still works.
Lemma A.2. Let = Ω(1/n1/16 ). Suppose that Sv is sampled by using the doubling algorithm
described above for T = O(log log n + log(1/2 )). Given z, w.h.p. we have |QSx (z) − Q(z)| ≤ .
Proof. For convenience, we also define the multiset Sv0 (0) ← {v} and Sv0 (i) ← Sv0 (i) ∪ Sv0 (ti (v)).
Also let Sv0 = Sv0 (T ). Therefore, Sv = {t0 (u) | u ∈ Sv0 }. First we show that the multiplicity of every
element in Sv0 is bounded by some constant w.h.p. Then, we will reveal t0 (u) for all u ∈ Sv0 to show
that QSv (z) is concentrated around Q(z).
Given a node w ∈ V , we will show that w.h.p. for every x ∈ V , the multiplicity of w in Sv0 ,
mSv0 (T ) (w), is bounded above by a constant. Let Mi = maxy mSy0 (i) (w). Observe that Mi+1 ≤ 2 · Mi
S
due to the update rule. Let S −1 (w, i) = y:w∈Sy0 (i) {y}. Note that Mi > Mi−1 only if a node in
S −1 (w, i − 1) chooses a node in S −1 (w, i − 1). We will first bound the size of S −1 (w, i) over the
execution of the algorithm, then we can bound the probability that Mi > Mi−1 increases.
Initially, S −1 (w, 0) = {w}. In round i, if a node z chose a node in S −1 (w, i − 1), then we must
have z ∈ S −1 (w, i). Let Xz denote whether the node z has chosen a node in S −1 (w, i − 1).
X
E[|S −1 (w, i)|] ≤ |S −1 (w, i − 1)| +
E[Xz ]
z∈V
≤ |S
−1
(w, i − 1)| + n · |S −1 (w, i − 1)|/n = 2 · |S −1 (w, i − 1)|.
23
Moreover, we show that Xz ≤ 4·|S −1 (w, i−1)|+O(log n) w.h.p. Let k = 3·|S −1 (w, i−1)|+K·log n.
We have
k
−1
X
n
|S (w, i − 1)|
Pr(
Xz ≥ k) ≤
k
n
z∈V
k
en |S −1 (w, i − 1)|
≤
·
k
n
K·log n
−1
e · |S (w, i − 1)|
≤
3 · |S −1 (w, i − 1)|
≤ 1/ poly(n)
Therefore, w.h.p. |S −1 (w, i)| ≤ 4 · |S −1 (w, i − 1)| + O(log n) for i = 1 . . . T . As a result,
S −1 (w, T ) = O(4K · log n) = O(log3 n/4 ).
Recall that Mi > Mi−1 only if some node in S −1 (w, i − 1) selects a node in S −1 (w, i − 1) during
round i. Thus, the probability that Mi > Mi−1 is at most
6
log n
def
3
4
O(log3 n/4 )
p = 1 − (1 − O(log n/ )/n)
=O
8 n
The probability that M increases in more than K rounds over the T rounds is at most
K !
K !
T K
e · T log6 n
e log6 n · (log log n + log(1/))
p ≤O
· 8
=O
·
= 1/ poly(n)
K
K
n
K
8 n
Therefore, w.h.p. MT ≤ 2K .
Now, given z ∈ V , we will show that QSv (z) − ≤ Q(z) ≤ QSv (z) + . For every element
y ∈ Sv0 , let Zy be the indicator random variable that denotes whether xt0 (y) ≤ z. Note that
Pr(Zy = 1) = Q(z). Therefore, we have
X
E[rankSx (z)] =
E[Zy ] · mSv0 (y) = Q(z) · |Sv |
y∈Sv0
P
Since the events {Zy }y∈Sv are mutually independent. We can show that the quantity y∈Sv Zy ·
mSv0 (y) concentrates around Q(z) · |Sv (z)| by using Hoeffiding’s inequality.
!
X
22 |Sv |2
Pr(|
Zy · mSv0 (y) − Q(z) · |Sv || ≥ |Sv |) ≤ exp − P
2
y∈Sv mSv0 (y)
y∈Sv
22 |Sv |2
mSv0 (y) ≤ MT ≤ 2K
≤ exp − 2K
2 · |Sv |
≤ exp (−Ω(log n))
≤ 1/ poly(n)
Therefore, w.h.p. |QSv (z) − Q(z)| ≤ .
24
|Sv | = Ω((log n)/2 )
A.1
Compaction
In this section, we show that instead of storing the full buffer, it is possible to only keep a summary
of the buffer, where the size of the summary is Θ( 1 · (log log n + log(1/))). Then, when a node v
contacts t(v), v only pulls the summary of t(v). This reduces the message complexity.
Let k = Θ( 1 · (log log n + log(1/))) be a power of 2 that denotes the size of the buffer. Also,
we assume that n0 = |Sv | is also a power of 2 thus the algorithm runs for T = log2 n0 + 1 rounds.
Suppose that v contacts t(v) in a given round, the new buffer updating rule is:
S̃v ← Compact(S̃v ∪ S̃t(v) )
, where Compact(Z) is a compaction operation. The compaction operation leaves the buffer
unchanged if |Z| ≤ k. Otherwise, it sort the elements in non-decreasing order and then extract the
elements in the even positions. Since |Z| ≤ 2k, we must have | Compact(Z)| ≤ k.
Let Sv denote the buffer of x at end of round T . Let S̃v ⊆ Sv denote the buffer of v with the
compaction updating rule at the end of round T . We will show that QS̃v (z) is a good approximation
of QSv (z) for any value z.
Suppose that the content of Sv is known. At the end of round i, Sv can be expressed as the
union of the buffers of 2T −i nodes. Let Uv (i) denote the multi-set of 2T −i nodes. Let Su (i) and
S̃u (i) denote the buffer and the compacted bufferSof u at the end of round i respectively.
Now let Uv (i) be the set of nodes such that u∈Uv (i) Su (i) = Sv , where |Uv (i)| = 2T −i . Now
def S
we consider the compacted version of the set Cv (i) = u∈Uv (i) S̃u (i). Note that Cv (i) = Sv for
i = 1, . . . , log2 k +1 (that is, before any compaction happens). Also, Cv (T ) = S̃v . We will show that
in each transition along Sv = Cv (log2 k + 1), Cv (log2 k + 2), . . . , Cv (T ) = S̃v , the error introduced
by compaction is sufficiently small. Let a · S denote the multiset where each element s ∈ S is
duplicated to have a occurrence. For a given element z, we will compare the rank of z in wi · Cv (i)
and wi+1 · Cv (i + 1) for log2 k + 1 ≤ i < T to derive the loss of compaction.
Let hi denotes the number of compaction operations done on Sv at the end of round i. Therefore,
hi = max(0, i − lg k − 1). Let wi = 2hi denote the weight of the buffer at the end of round i. Note
that the cardinality of the set |wi · Cv (i)| = |Sv | for 1 ≤ i ≤ T .
Given a multi-set S and any element z (not necessarily in S), we let RS (z) to denote the rank
of z in S (i.e. the number of elements in S that are smaller than or equal to z).
Lemma A.3. Let v be a node and let z by any element. For log2 k + 1 ≤ i < T , |R(wi ·Cv (i)) (z) −
R(wi+1 ·Cv (i+1)) (z)| ≤ |Uv (i + 1)| · wi .
Proof. Note that Cv (i + 1) is obtained from Cv (i) by applying |Uv (i + 1)| compactions, where there
is one compaction operation for each u ∈ Uv (i + 1). Each compaction operation compacts two
buffers S̃u (i) and S̃t(u) (i) into one and also doubles the weight.
Consider the effect of one compaction operation on the rank of z. Let R1 denote the rank of z
in wi · (S̃u (i) ∪ S̃t(u) (i)) (i.e. before the compaction) and R2 denote the rank of z in wi+1 · S̃u (i + 1)
(i.e. after the compaction). If R1 is even, then we must have R1 = R2 . If R1 is odd, then we have
R2 = R1 − wi .
Since there are most |Uv (i + 1)| compactions, the total difference created by all of them is at
most wi · |Uv (i + 1)|.
Corollary A.4. Let v be a node and let z by any element. We have |RSv (z)−RwT ·S̃v (z)| ≤
25
n0
2k
0
log nk .
Proof. First, note that Sv = Cv (log2 k + 1) and S̃v = Cv (T ). For any z, we have:
|RSv (z) − RwT ·S̃v (z)| ≤
T
−1
X
Rwi ·Cv (i) (z) − Rwi+1 ·Cv (i+1) (z)
wlog2 k+1 = 1
i=log2 k+1
≤
T
−1
X
Rwi ·Cv (i) (z) − Rwi+1 ·Cv (i+1) (z)
i=log2 k+1
≤
T
−1
X
|Uv (i + 1)| · wi
By Lemma A.3
i=log2 k+1
=
T
−1
X
2T −i−1 · 2i−log2 k−1
i=log2 k+1
=
T
−1
X
2T −log2 k−2 =
i=log2 k+1
n0
n0
log
2k
k
Therefore, if we set k = Θ( 1 log(n0 )), then we have |RSv (z) − RwT ·S̃v (z)| ≤ n0 by Corollary A.4.
This further implies that |QSv (z) − QS̃v (z)| ≤ . We immediately obtain the following Corollary.
Corollary A.5. Given any 0 < < 1. Let Sv denote the buffer of v with the original updating rule
after T = Θ(log n0 ) rounds, where n0 = |Sv |. Let S̃v denote the buffer of v using the compaction
updating rule with buffer size k = Θ( 1 log(n0 )). For any z, we have |QSv (z) − QS̃v (z)| ≤ .
Theorem A.6. Given = Ω(1/n1/16 ). Let S̃v denote the buffer of v using the compaction updating
rule with buffer size k = Θ( 1 · (log log n + log(1/))) after T = Θ(log log n + log(1/)) rounds. For
any z, w.h.p. we have |Q(z) − QS̃v (z)| ≤ .
Proof. Let 1 = 2 = /2. First, by applying Lemma A.2, we have |QSv (z) − Q(z)| ≤ 1 w.h.p.
where n0 = |Sv | = Θ(log n/2 ). Then, we apply Corollary A.5 to show that the corresponding
compaction update rule obtains a compacted buffer S̃v of size O( 1 (log log n + log(1/))) such that
|QSv (z) − QS̃v (z)| ≤ 2 . By triangle inequality, we have:
|Q(z) − QS̃v (z)| ≤ |QSv (z) − Q(z)| + |QSv (z) − QS̃v (z)| ≤ 1 + 2 ≤
B
Tools
Lemma P
B.1. (Chernoff Bound) Let X1 , . . . , Xn be indicator variables such that Pr(Xi = 1) = p.
Let X = ni=1 Xi . Then, for δ > 0:
E[X]
eδ
Pr(X ≥ (1 + δ)E[X]) <
(1 + δ)(1+δ)
E[X]
eδ
Pr(X ≤ (1 − δ)E[X]) <
(1 − δ)(1−δ)
26
The two bounds above imply that for 0 < δ < 1, we have:
Pr(X ≥ (1 + δ)E[X]) < e−δ
2 E[X]/3
Pr(X ≤ (1 − δ)E[X]) < e−δ
2 E[X]/2
.
Lemma B.2. Suppose that for any δ > 0,
eδ
Pr (X > (1 + δ)np) ≤
(1 + δ)(1+δ)
np
then for any M ≥ np and 0 < δ < 1,
eδ
Pr (X > np + δM ) ≤
(1 + δ)(1+δ)
≤ e−δ
M
2 M/3
Proof. Without loss of generality, assume M = tnp for some t ≥ 1, we have
Pr (X > np + δM )
np
etδ
≤
(1 + tδ)(1+tδ)
M
eδ
=
(1 + tδ)(1+tδ)/t
M
eδ
≤
(1 + δ)(1+δ)
≤ e−δ
2 M/3
(∗)
eδ
(1+δ)(1+δ)
≤ e−δ
2 /3
for 0 < δ < 1
Inequality (*) follows if (1 + tδ)(1+tδ)/t ≥ (1 + δ)(1+δ) , or equivalently, ((1 + tδ)/t) ln(1 + tδ) ≥
(1 + δ) ln(1 + δ). Letting f (t) = ((1 + tδ)/t) ln(1 + tδ) − (1 + δ) ln(1 + δ), we have f 0 (t) =
1
(δt − ln(1 + δt)) ≥ 0 for t > 0. Since f (1) = 0 and f 0 (t) ≥ 0 for t > 0, we must have f (t) ≥ 0 for
t2
t ≥ 1.
27
| 8cs.DS
|
Title: A Quantitative Neural Coding Model of Sensory Memory
Authors: Peilei Liu1*, Ting Wang1
Affiliations:
1
College of Computer, National University of Defense Technology, 410073 Changsha, Hunan,
China.
*Correspondence to: [email protected]
Abstract: The coding mechanism of sensory memory on the neuron scale is one of the most
important questions in neuroscience. We have put forward a quantitative neural network model,
which is self-organized, self-similar, and self-adaptive, just like an ecosystem following
Darwin's theory. According to this model, neural coding is a “mult-to-one” mapping from
objects to neurons. And the whole cerebrum is a real-time statistical Turing Machine, with
powerful representing and learning ability. This model can reconcile some important disputations,
such as: temporal coding versus rate-based coding, grandmother cell versus population coding,
and decay theory versus interference theory. And it has also provided explanations for some key
questions such as memory consolidation, episodic memory, consciousness, and sentiment.
Philosophical significance is indicated at last.
Main Text: Great strides have been made in neuroscience and cognitive science (1-8). Until now
however, the coding mechanism of sensory memory on the neuron scale is still unclear. A gap
exists between the molecular and whole brain research (9). We wish to bridge this gap through a
quantitative coding model. Inspired by the “self-organization” idea (10, 11), we only make local
rules about neuron and synapse based on existing data and theories. Then the hierarchical neural
network will emerge automatically. It has features of considerable interest in neuroscience,
cognitive science, psychology, and even philosophy. The modeling target is mainly the cerebral
cortex. All ci are constants, which have different meanings in different paragraphs.
dq
c1 (1 e c2 ) , mi pi ai , where q is the ion quantity,
dt
m is neuromodulator factor, pi is the postsynaptic signal, ai is the activation or sensitivity of
membrane channels. The neuron will be fired when q>c0. And the postsynaptic AP (action
potential) will induce LTP (long-term potentiation) in this case. Otherwise LTD (long-term
depression) will be induced when lacking postsynaptic AP. N1 is inspired by the MP model (12)
Single neuron model (N): N1)
dpi
c3 wi fi c4 pi (c3 c4 ) , s p c5 pi f , where wi is the
dt
synaptic strength, fi is the presynaptic spike frequency, sp is stimulus quantity inducing LTP
when the neuron is fired, f is the neuron’s firing frequency. Namely pi is determined by two
and Hodgkin-Huxley model (13). N2)
dai
c6 (c7 ai ) 0(c6 0) ,
dt
sd c8 pi (c7 ai ) c9ec t , where sd is stimulus quantity inducing LTD when the neuron can’t be
fired. LTD should be caused by stimulus to fatigue channels namely c7 ai . Postsynaptic AP can
result in complete fatigue of channels namely ai 0 , while EPSP (excitatory postsynaptic
processes: linearly increase of fi and exponentially decay. N3)
6
potential) will result in partial fatigue namely 0 ai c7 . Without stimulus however, they will
recover to c7 quickly since c6>>0.
N1 and N2 reflect the widely acknowledged spatial and temporal summarization of EPSP.
N3 is actually an adaptive mechanism preventing spike frequency becoming too high. This
model is supported by experiments and STDP (Spike timing Dependent Plasticity) model (14).
As in N1, experiments support that inducing LTP needs the coincidence of back-propagating AP
and EPSP (15). Correspondingly, inducing LTD needs stimulus without postsynaptic AP.
According to N1, this require stimulus of low frequency (16) or closely after postsynaptic AP
(namely stimulus in the refractory period) (14). Accordingly to N2 and N3, the time interval
between presynaptic and postsynaptic spikes is important. Specifically, the temporal curves of
LTP and LTD in STDP should be caused by the EPSP decay and channels recovery respectively.
On the other hand however, this neuron model also supports rate-based coding or Hebb coding.
For a usual spike sequence, it can be inferred from N2 that pi c1 fi (1 ec t ) , namely pi c1 fi
when t>>0. Therefore s p c3 pi f c4 fi f , and this is actually the Hebb conjecture (17). This can
2
explain why inducing LTP needs stimulus of high frequency or collaboration of multiple
dendrites (16). In conclusion, this model has reconciled the disputations between rate-based
coding (Hebb coding) and temporal coding (STDP). Specifically, it is temporal coding for single
isolated spike, but rate-based coding for natural spike sequence. We mainly discuss natural spike
sequence in this paper, because the isolated spike is usually man-made rather than natural.
For usual spike sequence, σ in N1 is approximately constant. Therefore N1 becomes
q c1 (1 ec2 )t . Namely, the neuron’s firing frequency is f 1/ t c3 (1 ec2 ) . Some early
models such as MP model also used the form f (i w i fi ) . But the specific form of φ here is
important for the statistical significances. Suppose that the probability of an object O occurring
on condition of attribute Ai is P(Oi)= P(O|Ai), where Ai are independent events. Let P(¬Oi)=
e c4 pi , and then P(O)=P(O1+O2+…+ Oi)= 1-P(¬O1¬O2…¬Oi)= (1 ec4 ) , where i pi .
Compared with N1, the firing frequency f is proportional to the probability of an object occurring
P(O) on condition of attributes. Similarly in N2, pi c5 fi (1 ec t ) could be viewed as the
probabilistic estimate of an attribute occurring P(Ai) on condition of presynaptic spikes history.
Without presynaptic spikes however, the reliability of this probabilistic estimate should drop
exponentially with time. In essence, N1 and N2 reflect space-time localization which is also a
frequently-used principle in the memory management of computer. Namely objects tend to exist
in continuous local space and time. In conclusion, neuron is real-time statistical machine. This is
also inspired by the statistical physics (18) and prediction of John Von Neumann (19).
Incidentally, the exponential functions in this model can be easily implemented in physics and
biology. For example, the radioactive materials decay exponentially.
6
Synapse model (S): S1) LTP:
dw
dr
c1s p h(c2 w) 0 ,
c3 s p h(1 r ) 0 , and
dt
dt
dw
c4 w log(r ) 0 when s p 0 , where w is the synaptic strength, r is the synaptic decay rate, h
dt
is hormone factor, s p c5 fi f according to N2. Generally speaking, synapses are alive and self-
adaptive like muscles: continuous exercises make them thick and tough, but they will decay
passively without exercises. S2) LTD:
and
dwd
dr
c6 sd wd 0 wd 1 , d c7 sd rd 0 rd 1 ,
dt
dt
dwd
c8 log(rd )(1 wd ) when sd 0 , where wd is the synaptic devaluation, rd is the
dt
recovery rate of wd. Generally speaking, LTD is the devaluation of LTP, and the actual synaptic
strength should be their product wwd. Interestingly, curves in S2 are almost the reversal of S1.
Since LTP is long-term (16), stimulus must have changed the synaptic persistence as well as
strength. And obviously LTP decays with time. This synapse model is inspired by the BCM
model in that synaptic growth rate is influenced by the synaptic strength itself (20). From the
statistical viewpoint, synaptic strength reflects the confidence of attribute based on stimulus
history. Similar to N2, S1 actually reflect the temporal localization principle, which should be
the physiological foundation of “recency effect” or “recency bias”. Differently however, N2 is a
kind of instantaneous memory which could be the foundation of consciousness according to
Francis Crick (1). We mainly discuss LTP here, because LTD is similar. In addition, LTD is
induced by stimulus below the firing threshold. Therefore it should be a regulator for depressing
noises other than kernel encoding mechanism (21), just like the fatigue of photoreceptors in the
retina. Both synaptic increase and decrease are necessary for many neural network models (6, 7).
According to our model however, the synaptic decrease should be due to passive decay instead
of LTD.
Competition model (C): C1)
dpi
n
pd , pd c1 (1 ec ) , j 1, j i f j , where pd is the
dt
2
quantity of retrograde messengers, n is the number of dendrites connected to the same axonal
branch, pi and fi are the postsynaptic potential and spike frequency respectively. Specifically,
back-propagating AP will release retrograde messengers to the presynaptic and transitorily
depress other synapses connecting to the axonal branch (22). And their postsynaptic APs will be
depressed reversely according to N1. As statistics, this “rich-get-richer” will lead to lateral
inhibition known as “softmax” or “winner-take-all” in existing theories (see Fig. 1) (4, 6).
However, our model doesn’t need extra “MAX” layers or lateral inhibitory neurons, which is
consistent with physiological evidences. The presynaptic axonal branches themselves actually
play similar roles. This convergent competition should be the physiological foundation of
“primacy effect” in psychology. And the unique winner is actually the coding neuron of this
input, while its revivals can be viewed as background noises. After all, two clocks don’t give
more accurate time. Form this viewpoint, lateral inhibition is a spatial adaptive mechanism, just
like LTD being the temporal adaptive. On the other hand, neurons tend to keep sematic distances
apart and prevent replicative coding due to lateral competition. In brief, neurons are
professionalized in the division of work. And this will lead to sparse coding (23) and the
structure of attribute slot defined as following.
y1
a
y2
b
y3
c
d
Sparse Inputs
Fig. 1. Synaptic competition and attribute slot. Black and white dots mean fired and unfired
neurons respectively. Synapse a and d aren’t activated for lack of presynaptic spike. But b isn’t
activated for its failure in the synaptic competition. However, c will be activated somewhat
because y3 will be fired by other inputs. As a result, neuron y2 will inhibit y1 known as lateral
inhibition, and y2 and y3 will inhibit each other. Lateral inhibition will lead to attribute slots,
which are divided by vertical dotted lines here. For example, y1 and y2 are in the same attribute
slot, while y3 is in a different one.
Definition 1: as in Fig. 1, an attribute slot is a set of neurons: s=(a1, a2, …, an): if ai>0, then
aj=0 for all j≠i, where ai is the neuron’s firing frequency. For example, if every color of a point
corresponds to a single detector, all these detectors can compose an attribute slot. Specially, a
binary bit is actually a special attribute slot with two values, and it needs two neurons for
representing 0 and 1 respectively (see Fig. 2). In some models, 0 is represented by the resting
potential. In our opinion however, the resting potential is meaningless because it can’t transit
neurotransmitters. Therefore the spike is actually unitary other than binary, different from the
binary pulse in computer. On the other hand, an attribute with n values can also be represented
by the combination of m independent neurons (logb (n) m n) other than by n mutuallyexclusive neurons. For example, all colors can be mixed of three-primary colors in different
proportions. In essence, these two coding manners are “grandmother cell” and population coding
respectively (24).
Attribute slot has great representative ability. The “XOR problem” once caused the
“artificial intelligence winter” (25), which was solved by multiple-layered network latter. With
attribute slot however, it can be solved through a single-layer network (see Fig. 2). In fact, it can
be proved that a single-layer network can represent any logic expression P. As well known,
P p1 pi pn , pi a1 a j a m , where aj is either an atom expression or its
negative form. Since aj and ¬aj can be represented by attribute slot, pi can be represented by a
single neuron. Therefore P can be represented by a single-layer network. Moreover, inhibitory
neurons aren’t necessary, different from logic gates in computer. Retrograde massagers actually
play similar roles of logic “NOT”. Similar to the all-or-none AP, the attribute slot is a kind of
digital coding in essence. The digital coding is used widely in computer science for its tolerance
of noise. Similarly, the cerebrum can also ignore flaws known as brain completion. The cost is
that more neurons are needed.
A
x1
y1
¬x1
B
y2
¬x2
x2
¬x1
y1
y3
x1
y2
x2
¬x2
Fig. 2. Representation of XOR function. Panel A and B represent “x1 XOR x2” and “x1 OR x2”
respectively. Feature each of x1 and ¬x1 corresponds to an independent neuron respectively. In
some models however, ¬x1 and ¬x2 are represented by resting potentials.
Self-organized neural network and circuits. The initial model is composed of free
neurons. Free dendrites move randomly and connect to axonal branches nearby with an initial
synaptic strength. Whenever meeting an input, neurons compete and only winners can strengthen
their dendritic synapses. Other synapses will decay until broken. Therefore, any input is actually
either encoded by a free neuron or merged into the most similar coding neuron. Since single
neuron’s dendrites are limited, every neuron only encodes a small part of an input. Therefore
fine-grain encoding is supported. And the whole input is actually encoded by a hierarchy tree
structure, whose root is the coding neuron of this input (see Fig. 3A). Overlapping coding trees
compose neural network or cortical column (26). Since input is sparse (23), for simplicity we
suppose that every input is composed of attribute slots (s1,s2,…,sn). Then every layer should be
composed of attribute slots according to C1. And higher layer contains fewer attribute slots.
Therefore every layer is the input of next layer, and feedback fibers can be viewed as common
inputs. As results, circuits can be self-organized similar to the feedforward network (see Fig. 3B).
Moreover, the neural network is self-similar: coding trees are like large neurons, while columns
are large attribute slots. In some degree, a cortical area or the cerebrum itself is a super attribute
slot.
According to Francis Crick (1), the main cortical structure is determined by genes, while the
fine neural coding is determined by postnatal experiences. Consistent with the neural Darwinism
theory (2), the cortex is like an ecosystem according to our model: inputs are the sunshine;
neurons collaborate and compete like plants and animals; and only the fittest can survive.
Animals at the top of food chains are actually the memories of environment and era. For
instances, lion is the symbol of African grassland, while the tyrannosaurus rex is the memory of
the Cretaceous period. Recent experiments support that neural connections change dynamically
with external stimulus (27). As results, universal free neurons become specialized gradually, just
like embryonic stem-cells irreversibly growing into different organs. For example, visual
pathways in the developing ferret brain can be rewired to auditory cortex and work well. A
neuron’s function is actually determined by its dendritic connections. Just as Karl Marx said,
“Person is the sum of all social relations”. From this viewpoint, people are super neurons of the
society or Twitter.
A
y2
y1
B
Outputs
x
Inputs
Feedback
Fig. 3. Neural network and circuits. In A, black and white dots mean fired and unfired neurons
respectively. Any input will converge to a single neuron such as y1 through a coding tree.
Overlapping trees compose a neural network, and common intermediate nodes such as x are
shared. Due to lateral inhibition, every layer is composed of attribute slots separated by vertical
dotted lines here. In B, the triangle symbol represents coding tree in A. The long feedforward
and feedback fibers can be viewed as common inputs. Therefore circuits can be self-organized
similar to A.
Computational significance. According to our model, neural coding is a “multiple-to-one”
mapping from inputs to neurons due to lateral inhibition, which is a well-defined function in
essence. An input is encoded only when it has been convergent to the coding neuron through a
coding tree. Conversely, whenever a coding neuron is fired, the corresponding input will be
retrieved. Obviously, the retrieval process is an encoding process meanwhile. Neurons are
computing units as well as storage units. Information is saved in the dendritic synapses, while the
firing frequency is actually computing result. From this viewpoint, synapses and spikes are
hardware and software of nerve system respectively, and both of them are distributed. Different
from the von Neumann architecture (19), data and program aren’t separated. Representation of
information is digital but not binary. In addition, dynamic changes of program at runtime are
allowed. From the viewpoint of machine learning, the feedforward encoding is unsupervised
learning in general. But supervised learning can also be conducted by facilitating or inhibiting
neuron’s firing through specific fibers casting. This is important when we need to associate
different objects or distinguish similar objects. For instance, we have to associate words and
pronunciations with pictures as well as distinguish similar faces. Incidentally, that’s why people
have “face cells” while monkeys don’t (1). According to N1 and S1, neuromodulators and
hormones are actually meta-mechanisms influencing neuron’s firing frequency and learning rate.
They could conduct reinforcement learning and selectively reinforcing memories important for
survival, similar to the dopamine system in motor learning (28). However, the supervised
learning and reinforcement learning here are “soft” or statistical rather than absolute. Since fired
neurons strengthen their dendritic synapses towards to the input pattern, the lateral competition
become convergent and stable gradually. Therefore the neural network tends to converge to the
global minimum point, as in the simulated annealing algorithm (29). Similarly, synapses become
stable gradually with stimuli according to S1. And synapses in lower layers are more stable. As
results, iterations in back-propagating algorithm could be avoided when training hierarchical
feedforward networks (7).
Generally speaking, our model supports “grandmother cell” rather than population coding
(24). The “grandmother cell” was questioned for following reasons (1, 4): first, neurons are too
few and too simple for coding infinite complex objects in theory; second, experiments
demonstrate that a neuron isn’t a template of all stimuli in its receptive field. According to our
model however, a neuron encodes more than one object, and not all inputs are stored. Instead,
only those occurring frequently or companied with hormone can be saved. And synaptic decay
and neurogenesis (30) can produce additional memory capacity. Moreover, resources can be
saved through sharing common intermediate nodes in the deep hierarchical network (4, 5),
although retrieval will become slower. And neurons could actually use large trees for encoding
complex objects. On the other hand, not all stimuli in a neuron’s receptive field are effective.
Due to space-time adaptive mechanisms, neurons are more sensitive to features in contrast to
backgrounds. Moreover, the coding is also influenced by the vast top-down connections. A
possible worry is about the dying of neurons. In our opinion, only those lack of stimulus die. In
fact, the influence of accidental death is also limited. As statistical model, lack of partial inputs
won’t result in fatal disaster to a neuron. And due to the low firing threshold, this neuron model
has surprising generalizing ability (4). Almost any input could stir up many neurons, and the
retrieval is actually determined by lateral competition. Therefore the place of a dead neuron can
be taken by the most similar one. In brief, the “grandmother cell” is as robust as population
coding. However, population firing confronts the binding problem (1), namely how to reconcile
distributed information and the unitary consciousness. Brain waves or rhythm was proposed as
the binding mechanism. In our opinion however, binding is computing in essence. And neurons
themselves should be the workshops binding features together.
Psychological significance. S1 actually gives the learning curve and forgetting curve.
However, the forgetting curve in experiments isn’t consistent with exponential function
completely (31). In our opinion, recall in the tests and other rehearsals are actually memory
consolidation process, which will slow down the decay rate. On the other hand, the lateral
inhibition in C1 can explain memories interference (32). Since it influences encoding and
retrieval meanwhile, both proactive interference and retroactive interference can be explained.
Therefore our model has actually reconciled the decay theory and interference theory (33). In
essence, forgetting is retrieval failure due to either lateral inhibition or disappearance of
“memory trace”. Some researchers hold that forgetting should result from random changes of
synapses due to noise (34), because relearning is easier than learning new, namely the implicit
memory in psychology. According to our model, implicit memory are stored in lower layers,
where neurons are common nodes shared by different coding trees. Therefore their synapses are
usually tougher and more likely to remain. And forgetting should mainly be due to synaptic
decay near the root of coding trees. Repetitive stimulus (rehearsal) and hormones can lead to
memory consolidation. For example, we can remember a strange phone number through quick
repetition. And dream should be one of such rehearsals (35). Specifically, neurons will become
sensitive and firing spontaneously when without external inputs in sleep. Due to lateral inhibition,
only strong circuits can be fired. Therefore in essence, dream is the self-reinforcing of strong
memories as well as the fantasy divorced from reality since lacking inputs from nature.
Influential events are often companied with hormones changing, which actually represent their
importance for the survival (36). Since an event seldom repeats, it is hard for the neocortex to
encode. However, it can be stored in hippocampus and limbic structure which are sensitive to
hormones. These memories are usually known as episodic memory or flash memory, which is an
important part of working memory in our opinion. They should be able to influence the sematic
memory in cortex (37), just like computer memory writing and reading disk. From this viewpoint,
sematic memory and rationality serve the hormone system, or instinct dominates rationality.
Consciousness is one of the most important and mystical topics (1, 3, 8), which has different
meanings in different disciplines. Here we try to discuss the neural mechanism correlative with
consciousness, namely what happens when you think of an object. According to our model,
conscious of something means the corresponding coding neuron’s firing. Paying attention to an
object means specific fibers and neuromodulators casting to the coding neuron, while sentiment
should be pervasive hormones. And different areas should be sensitive to different hormones, as
the Lövheim cube of emotion (38). Our consciousness is often believed to be autonomous and
free. For example, a girl can make various sentences she has never heard of. In our opinion, the
consciousness seems autonomous because of the continuous complex inputs, huge memory
volume, and some randomness. Specifically, the cerebrum is a “Turing Machine” in essence,
whose output is determined by current inputs and internal states. Current inputs come from
nature as well as our body, while internal states include memory and instinct system. In fact, the
instinct or sentiment can also be viewed as ancient memories encoded in genes. As Crick said,
we believe our minds are free, because we often know the decision itself but unaware of the
decision-making process (1). In other words, the cerebrum is approximately a black box for us.
Open the mainframe however, you will find a not-so-intelligent machine composed of transistors
and plastic connections, although it is indeed more robust. It is surprising and interesting that
intelligence could come from random processes.
The philosophical significance hasn’t escaped our notice: the consciousness is the second
nature, while the so-called “nature” is actually our subjective cognition of the “real” nature. We
can cognize the nature only in probability and statistics, but never precisely or completely. In
brief, the materialism and idealism can be reconciled.
References and Notes:
1. F. Crick, The Astonishing Hypothesis: The Scientific Search for the Soul (Charles Scribner's
Sons, New York, 1994).
2. G. Edelman, The Theory of Neuronal Group Selection (Basic Books, New York 1987).
3. C. Koch, K. Hepp, Quantum mechanics in the brain. Nature 440, 611-612 (2006).
4. T. Poggio, E. Bizzi, Generalization in vision and motor control. Nature 43, 768-774 (2004).
5. G. E. Hinton, R. R. Salakhutdinov, Reducing the dimensionality of data with neural networks.
Science 313, 504-507 (2006).
6. S. Grossberg, Adaptive resonance theory: how a brain learns to consciously attend, learn, and
recognize a changing world. Neural Networks 37, 1-47 (2013).
7. D. E. Rumelhart, J. L. McClelland. Parallel Distributed Processing (The MIT Press, London,
1986).
8. J. Hawkins, S. Blakeslee, On Intelligence (Times Books, New York, 2004).
9. G. Miller, How are memories stored and retrieved. Science 309, 92 (2005).
10. A. L. Barabási, R. Albert, Emergence of scaling in random networks. Science 286, 509-512
(1999).
11. C. W. Reynolds, Flocks, herds, and schools: a distributed behavioral model. Comput.
Graphics 21, 25-34 (1987).
12. W. S. McCulloch, W. Pitts, A logical calculus of the ideas immanent in nervous activity. Bull.
Math. Biol. 5: 115-133 (1943).
13. A. L. Hodgkin, A. F. Huxley, A quantitative description of membrane current and its
application to conduction and excitation in nerve. J. Physiol. 117, 500-544 (1952).
14. W. Gerstner, R. Kempter, J. L. Hemmen, H. Wagner, A neuronal learning rule for submillisecond temporal coding. Nature 383, 76-78 (1996).
15. H. Markram, J. Lübke, M. Frotscher, B. Sakmann, Regulation of synaptic efficacy by
coincidence of postsynaptic APs and EPSPs. Science 275, 213-215 (1997).
16. C. R. Bramham, B. Srebro, Induction of long-term depression and potentiation by low- and
high-frequency stimulation in the dentate area of the anesthetized rat: Magnitude, time course
and EEG. Brain Res. 405:100-107 (1987).
17. D. O. Hebb, The Organization of Behavior: A Neuropsychological Approach (Jhon & Sons,
New York, 1949).
18. E. Schrödinger, What Is Life? (Cambridge University Press, Cambridge, 1944).
19. J. V. Neumann, The Computer and the Brain (Yale University Press, New Haven, CT, 2000).
20. E. L. Bienenstock, L. N. Cooper, P. W. Munro, Theory for the development of neuron
selectivity: orientation specificity and binocular interaction in visual cortex. J. Neurosci. 2,
28-32 (1982).
21. M. Schonewille et al., Reevaluating the role of LTD in cerebellar motor learning. Neuron 70,
43-50 (2011).
22. W. G. Regehr, M. R. Carey, A. R. Best, Activity-dependent regulation of synapses by
retrograde messengers. Neuron 63, 154-170 (2009).
23. B. A. Olshausen, Sparse codes and spikes. R.P.N. Rao, B. A. Olshausen and M. S. Lewicki
Eds. MIT press, 257-272 (2002).
24. C. G. Gross, Genealogy of the "grandmother cell". Neuroscientist 8, 512-518 (2002).M.
25. Minsky, S. Papert, Perceptrons: An Introduction to Computational Geometry (MIT Press,
London, 1972).
26. R. B. Tootell, M. S. Silverman, R. L. De Valois, Spatial frequency columns in primary visual
cortex. Science 214, 813-815 (1981).
27. G. B. Ermentrout, R. F. Galán, N. N. Urban, Relating neural dynamics to neural coding. Phys.
Rev. 99, 248103 (2007).
28. M. J. Frank, L. C. Seeberger, R. C. O’Reilly, By carrot or by stick: cognitive reinforcement
learning in parkinsonism. Science 306, 1940-1943 (2004).
29. D. H. Ackley, G. E. Hinton, T. J. Sejnowski, A learning algorithm for Boltzmann machine.
Cognit. Sci. 9, 147-169 (1985).
30. C. G. Gross, Neurogenesis in the adult brain: death of a dogma. Nature Rev. Neurosci. 1, 6773 (2000).
31. J. T. Wixted, On common ground: Jost’s (1897) law of forgetting and Ribot’s (1881) law of
retrograde amnesia. Psychol. Rev. 111, 864-879 (2004).
32. T. D. Tomlinson, D. E. Huber, C. A. Rieth, E. J. Davelaar, An interference account of cueindependent forgetting in the no-think paradigm. Proc. Natl. Acad. Sci. U.S.A. 106, 1558815593 (2009).
33. K. Oberauer, S. Lewandowsky, Forgetting in immediate serial recall: decay, temporal
distinctiveness, or interference? Psychol. Rev. 115, 544-576 (2008).
34. J. V. Stone, P. E. Jupp, Falling towards forgetfulness: synaptic decay prevents spontaneous
recovery of memory. PLoS Comput. Biol. 4, (2008). doi:10.1371/journal.pcbi.1000143.
35. R. Stickgold, Sleep-dependent memory consolidation. Nature 437, 1272-1278 (2005).
36. E. Tulving, H. J. Markowitsch, Episodic and declarative memory: Role of the hippocampus.
Hippocampus 8, 198-204(1998).
37. P. W. Frankland, B. Bontempi, The organization of recent and remote memories. Nature Rev.
6, 119-130 (2005).
38. H. Lövheim, A new three-dimensional model for emotions and monoamine neurotransmitters.
Med. Hypotheses 78, 341-348 (2012).
Acknowledgments: This work was supported by the National Natural Science Foundation of
China (Grant No. 61170156 and 60933005).
| 9cs.NE
|
TRACY-WIDOM AT EACH EDGE OF REAL COVARIANCE ESTIMATORS
arXiv:1707.02352v1 [math.PR] 7 Jul 2017
ZHOU FAN AND IAIN JOHNSTONE
Abstract. We study the sample covariance matrix for real-valued data with general population
covariance, as well as MANOVA-type covariance estimators in variance components models under
null hypotheses of global sphericity. In the limit of large n and p, the spectra of such estimators may
have multiple disjoint intervals of support, possibly intersecting the negative half line. We show that
the distribution of the extremal eigenvalue at each regular edge of the support has a Tracy-Widom
F1 limit. Our proof extends a comparison argument of Ji Oon Lee and Kevin Schnelli, replacing a
continuous Green function flow by a discrete Lindeberg swapping scheme.
1. Introduction
Let Y1 , . . . Yn be observations on p traits or features. When the traits are distributed with mean
0 and covariance Σ ∈ Rp×p , the sample covariance matrix provides an unbiased estimate of Σ:
n
X
−1
Σ̂ = n
Yi Yi0 .
i=1
It is common to assume
√ that the vectors Yi are independent, and that the dependence between
traits arises via Yi = nΣ1/2 Xi , where the entries of Xi are i.i.d. with mean 0 and variance 1/n.
Setting X = [X1 , . . . , Xn ] this sample covariance estimator then takes the form
Σ̂ = Σ1/2 XX 0 Σ1/2 .
(1)
Equivalently, we may study its “companion” matrix X 0 ΣX, which has the same eigenvalues at Σ̂
up to |p − n| zeros.
In the large-sample limit where p and n increase proportionally, the empirical spectrum of Σ̂
is well-approximated by a deterministic law ρ [MP67, Yin86, SB95]. This asymptotic regime has
become increasingly relevant for data analysis, due to the prevalence of data where p is of comparable magnitude to n. In this regime and under a “sphericity” null hypothesis that Σ is the
identity, i.e. the traits represent isotropic noise, the largest eigenvalue of the estimate Σ̂ falls close
to the largest point of support of ρ [Gem80, YBK88] and has a Tracy-Widom F1 limiting distribution [Joh01, Sos02, Péc09, PY14, LS16]. This yields a statistical test for whether a large observed
eigenvalue of Σ̂ may be attributed to the isotropic noise, or whether it represents a deviation from
this sphericity null [Joh01, PPR06].
The primary motivation for our work arises from a class of statistical models in which the
observations Yi are dependent. For example, suppose the samples are divided into n/2 pairs (e.g.
pairs of siblings), and samples within each pair are more similar than samples between different
pairs. A classical statistical model for this scenario is a multivariate variance components (or
random effects) linear model, where the p-dimensional observation for sample i, belonging to pair
bi/2c, is modeled as
Yi = αbi/2c + εi .
Here, α1 , . . . , αn/2 are independent draws of a random vector with mean 0 and covariance Σ1 ∈
Rp×p , representing shared “pair effects”, and ε1 , . . . , εn are independent vectors with mean 0 and
Department of Statistics, Stanford University
E-mail addresses: [email protected], [email protected].
1
2
TRACY-WIDOM FOR COVARIANCE MATRICES
a different covariance Σ2 ∈ Rp×p , representing the individual variation. This model is expressed
concisely as
Y = U α + ε,
(n/2)×p
where the rows of α ∈ R
and ε ∈ Rn×p are the above vectors, and U ∈ {0, 1}n×(n/2) is
an incidence matrix encoding the pair memberships. Denoting by π, π ⊥ ∈ Rn×n the orthogonal
projections onto the column span of U and its orthogonal complement, it is verified that
Σ̂1 = n−1 Y 0 (π − π ⊥ )Y,
Σ̂2 = 2n−1 Y 0 π ⊥ Y
(2)
are unbiased estimators of Σ1 and Σ2 based on the observed data Y . These are the classical
multivariate analysis of variance (MANOVA) estimators of covariance in this model.
More generally, a multivariate variance components linear model takes the form
Y = U1 α1 + . . . + Uk αk ,
(3)
RIr ×p
where α1 , . . . , αk are independent matrices, each αr ∈
containing Ir independent rows disp×p
tributed with mean 0 and covariance Σr ∈ R . Each Ur ∈ Rn×Ir is a known and deterministic
incidence matrix. (Often there is an additional fixed effects term Xβ, for example modeling a nonzero mean value for each trait, which we omit here for simplicity.) It is usual to assume αr is Gauss√
1/2
ian, but in the spirit of our earlier discussion, we assume only a representation αr = pXr Σr ,
where the entries of Xr ∈ RIr ×p are i.i.d. mean 0 and variance 1/p. The setting of n independent
samples corresponds trivially to k = 1 and U1 = Id, while the preceding example corresponds to
k = 2, U1 = U , and U2 = Id. Such models encompass crossed and nested classification designs, one
application of which is the modeling of components of variation and covariation of phenotypic traits
in quantitative genetics [LW98, MCM+ 14, BAC+ 15]. In many examples, each Σr has a MANOVA
estimator
Σ̂r = Y 0 BY,
(4)
where B ≡ Br is a symmetric matrix that is constructed based on U1 , . . . , Uk . In the regime where
n, p, I1 , . . . , Ik all increase proportionally, the empirical spectrum of Σ̂r was shown in [FJ16] to also
be well-approximated by a deterministic law, in the sense of global weak convergence.
A relevant null hypothesis for the model (3) is that of “global sphericity”, which asserts that
each Σr is proportional to the identity, Σr = c2r Id for some c1 , . . . , ck ≥ 0. The value cr may be 0,
√
in which case the null hypothesis signifies the absence of the rth effect. Now αr = cr pXr and it
is verified that the MANOVA estimator (4) takes the equivalent form
Σ̂r = X 0 T X
(Xr )kr=1
RM ×p
(5)
(pcr cs Ur0 BUs )kr,s=1
RM ×M
∈
for
upon defining block matrices X =
∈
and T =
M = I1 + . . . + Ik . This is the same form as the companion matrix of (1). Correspondingly, under
a global sphericity null hypothesis, the equations in [FJ16] that describe the empirical spectrum of
Σ̂r reduce to the Marcenko-Pastur equation describing the spectrum of (1).
In this paper, we study the edges of the spectrum of the matrix (1) or (5). Recent work by
Lee and Schnelli [LS16] showed that, under certain regularity conditions, the largest eigenvalue of
(1) or (5) continues to exhibit a Tracy-Widom F1 limit, when Σ or T is positive-definite but not
necessarily identity. Building on the proof in [LS16], we extend this result in two directions:
1. We show that this holds also for symmetric matrices T that have negative eigenvalues.
2. The limiting spectrum of (1) or (5) may have multiple disjoint intervals of support. We establish
a Tracy-Widom F1 limit for the extremal eigenvalue at each “regular” edge of the support.
Both extensions are relevant to variance components models: The matrices B and T in (4) and
(5) typically have both positive and negative eigenvalues. [BM15, Boxes 1 and 2] has an example
from the quantitative genetics literature, and our main result resolves an open question stated
there about Tracy-Widom limits and scaling constants. (The MANOVA estimator Σ̂r may not be
TRACY-WIDOM FOR COVARIANCE MATRICES
3
positive definite. This is not of particular concern in our testing application, whereas for estimation,
it is common to postprocess Σ̂r to make it positive-definite [Ame85].) In spiked perturbations to
the global sphericity null, it is possible for spike eigenvalues of Σ̂r to appear not only beyond the
outer edges of the support, but also in the gaps between support intervals—we will describe this in
forthcoming work. The Tracy-Widom limit at each edge facilitates a statistical test of whether any
such observed eigenvalue may be attributed to a true spike eigenvalue in Σ1 , . . . , Σk . We discuss
applications to variance components models further in Section 2.4.
To our knowledge, the second extension above is new even for the sample covariance matrix (1)
with positive-definite Σ. The analogous result is known when X has complex Gaussian entries,
where convergence of the largest eigenvalue to the Tracy-Widom F2 law was established in [Kar07],
and this was extended to each regular edge in [HHN16]. These analyses use the determinantal form
of the HCIZ integral over the complex unitary group, which has no known real analogue. The proof
of [LS16] for the largest eigenvalue in the real setting is different and relies on a resolvent/Green
function comparison strategy, which we now discuss.
1.1. Strategy of proof. We consider the N × N matrix Σ̂ = X 0 T X where the M × M matrix T
is real symmetric. Let E∗ denote an edge of the deterministic spectral support of Σ̂. (We define
this formally in Section 2.) As in [LS16], the strategy of proof will be to consider
Σ̂(L) = X 0 T (L) X
for a different matrix T (L) , and to compare the eigenvalue behavior of Σ̂ near E∗ with that of Σ̂(L)
(L)
near an edge E∗ . By extending the edge universality established in [KY16] from positive-definite
to general symmetric T , it suffices to consider the case where T is diagonal and X has Gaussian
entries.
In [LS16], E∗ is the rightmost edge of support. The comparison between T and T (L) is achieved
by a continuous interpolation over l ∈ [0, L], where T (0) = T and each T (l) has diagonal entries
(l)
{tα : α = 1, . . . , M } given by
1
1
= e−l (0) + (1 − e−l ).
(6)
(l)
tα
tα
(See [LS16, Eq. (6.1)].) For reasons of the proof, each T (l) is then rescaled so that the largest
eigenvalue of Σ̂(l) ≡ X 0 T (l) X fluctuates with identical scale (of order N −2/3 ) for every l. Taking
L = ∞, T (∞) is a multiple of the identity, so Σ̂(∞) is a white Wishart matrix for which the
(l)
Tracy-Widom distributional limit is known. Along this interpolation, the upper edge E∗ traces a
(0)
(∞)
continuous path between E∗ and E∗ . Defining
(l)
(l)
(l)
(l)
#(E∗ + s1 , E∗ + s2 ) = number of eigenvalues of Σ̂(l) in [E∗ + s1 , E∗ + s2 ],
a resolvent approximation idea from [EYY12b] establishes the smooth approximation
h
i
h
i
(l)
(l)
P #(E∗ + s1 , E∗ + s2 ) = 0 ≈ E K(X(l) (s1 , s2 )) ,
(7)
where K(X(l) (s1 , s2 )) is a smoothed indicator of the integrated Stieltjes transform of Σ̂(l) along an
interval in C+ at height η = N −2/3−ε above the corresponding interval on the real axis. The crux
of the proof in [LS16] is then to show
i
d h
(8)
E K(X(l) (s1 , s2 )) ≤ N −1/3+ε
dl
for a small constant ε > 0 and s1 , s2 on the N −2/3 scale. This is applied to compare the probability
in (7) for l = 0 and l = 2 log N . A simple direct argument compares these probabilities for
l = 2 log N and l = ∞, concluding the proof in [LS16].
4
TRACY-WIDOM FOR COVARIANCE MATRICES
The technical contribution of our paper is to show that the continuous interpolation in (6) may be
replaced by a discrete interpolating sequence. The resulting extra flexibility permits the extension
of this result in the two directions mentioned earlier. Indeed, we note that (6) is not well-defined
(0)
for negative tα , as the right side passes through 0 along the interpolation. More importantly, (6)
does not allow us to study interior edges of Σ̂ when there are multiple disjoint intervals of support,
as the support intervals merge and these edges vanish along the interpolation. We instead consider
a discrete interpolating sequence T (0) , T (1) , . . . , T (L) for an integer L ≤ O(N ), where the diagonal
(l)
entries tα satisfy
M
X
|t(l+1)
− t(l)
α
α |≤ O(1)
(9)
α=1
(l)
for all l = 0, . . . , L − 1. Each matrix Σ̂(l) ≡ X 0 T (l) X will have a corresponding edge E∗ such that
(l+1)
|E∗
(l)
− E∗ |≤ O(1/N ).
(10)
Each of these L discrete steps may be thought of as corresponding to a time interval ∆l = O(N −1 )
in the continuous interpolation (6). We show that the above conditions are sufficient to establish a
discrete analogue of (8),
h
i
h
i
E K(X(l+1) (s1 , s2 )) − E K(X(l) (s1 , s2 )) ≤ N −4/3+ε .
(11)
As L ≤ O(N ), summing over l = 0, . . . , L − 1 establishes the desired comparison between T (0) and
T (L) . Importantly, the requirement (9) is sufficiently weak to allow a Lindeberg swapping scheme,
where each T (l+1) makes a single O(1) perturbation to a single entry of T (l) . Hence we may move
the diagonal entries of T from one interval of support to another, without continuously evolving
them between such intervals. This allows us to preserve the edge E∗ as in (10) along the entire
interpolating sequence, even as the other intervals of support disappear.
Section 3 reviews the requisite proof ingredients, which are similar to those in [LS16]. These include properties of the limiting Stieltjes transform near regular edges, Schur-complement identities
for the resolvent, a local Marcenko-Pastur law as in [BPZ13, KY16], and the resolvent approximation from [EYY12b] that formalizes (7). These ingredients either are drawn from the existing
literature or represent extensions to non-positive-definite T ; we defer proofs or proof sketches of
these extensions to Appendices A and B.
Section 4 constructs an interpolating sequence T (0) , . . . , T (L) for any starting matrix T (0) = T .
As in [LS16], we rescale each T (l) so that the eigenvalue of interest fluctuates with identical scale
for each l. Consequently, the interpolating sequence will not be exactly Lindeberg, but rather will
(l+1)
(l)
(l+1)
(l)
satisfy |tα
− tα |≤ O(1) for a single entry α and |tβ
− tβ |≤ O(1/N ) for all remaining entries
(L)
β 6= α. The final edge E∗ may be either a left or right edge of Σ̂(L) , and we conclude the proof
by applying either the result of [Joh01] for a positive right edge or [FS10] for a positive left edge
of a (real) white Wishart matrix. To ensure that a left edge is not a hard edge at 0, we allow T (L)
to have two distinct diagonal entries {0, t}. Thus, Σ̂(L) may have a different dimensionality ratio
from the starting Σ̂.
In Section 5, we conclude the proof by establishing (11). To achieve this, we generalize the
“decoupling lemma” of [LS16, Lemma 6.2] to a setting involving two different resolvents G and
Ǧ, corresponding to T ≡ T (l) and Ť ≡ T (l+1) . Fortunately, we do not need to perform the same
generalization for the “optical theorems” of [LS16, Lemma B.1], as we may apply (9) to reduce the
higher-order terms arising in the decoupling lemma to involve only G and not Ǧ. We will explain
this later in the proof.
TRACY-WIDOM FOR COVARIANCE MATRICES
5
Acknowledgments. We are indebted to quantitative geneticist Mark Blows for asking the question about Tracy-Widom for variance components models that led to this paper, and for many
stimulating discussions. We would like to also thank Kevin Schnelli for helpful conversations about
[LS16]. ZF is supported in part by a Hertz Foundation Fellowship and an NDSEG Fellowship (DoD
AFOSR 32 CFR 168a). IJ is supported in part by NIH R01 EB001988 and NSF DMS 1407813.
2. Model and results
2.1. Deterministic spectral law. Let T ∈ RM ×M be a deterministic symmetric matrix of
bounded operator norm, whose eigenvalues t1 , . . . , tM may be positive, negative, or zero. Let
X ∈ RM ×N be a random matrix with independent entries of mean 0 and variance 1/N . We study
the matrix
Σ̂ = X 0 T X
in the limit as N, M → ∞ proportionally. In this limit, the empirical spectrum of Σ̂ is wellapproximated by a deterministic law ρ ≡ ρN,M,T .∗ We review in this section the definition of ρ and
its relevant properties.
When T is the identity matrix, ρ is the Marcenko-Pastur law [MP67]. Under our scaling for X,
this has density
p
p
(E+ − x)(x − E− )
1
1(E− ,E+ ) (x),
E± = (1 ± M/N )2
(12)
f0 (x) =
2π
x
on the positive real line, and an additional point mass at 0 when M < N .
More generally, for any symmetric matrix T , the law ρ may be defined by a fixed-point equation
in its Stieltjes transform: For each z ∈ C+ , there is a unique value m0 (z) ∈ C+ which satisfies
z=−
M
1 X
tα
1
+
.
m0 (z) N
1 + tα m0 (z)
(13)
α=1
This is oftentimes called the Marcenko-Pastur equation, and it defines implicitly the Stieltjes transform m0 : C+ → C+ of a law ρ on R [MP67, SB95]. This law ρ admits a continuous density f0 at
each x ∈ R∗ , given by
1
=m0 (z),
(14)
f0 (x) = lim
+
z∈C →x π
where
(
R
if rank(T ) > N
R∗ =
R \ {0} if rank(T ) ≤ N.
For x 6= 0, this is shown in [SC95]; we extend this to x = 0 when rank(T ) > N in Appendix A.
The law ρ is called the free multiplicative convolution of the empirical distribution of t1 , . . . , tM
with the Marcenko-Pastur law (12). In contrast to the case T = Id, if t1 , . . . , tM take more than one
distinct value, then ρ may have multiple disjoint intervals of support. Two such cases are depicted
in Figures 1 and 2. For each support interval [E− , E+ ] of ρ, we will call each endpoint E− and E+
an edge. More formally:
Definition 2.1. The support of ρ is
supp(ρ) = {x ∈ R : ρ([x + δ, x − δ]) > 0 for all δ > 0}.
E∗ ∈ R is a right edge of ρ if (E∗ − δ, E∗ ) ⊂ supp(ρ) and (E∗ , E∗ + δ) ⊂ R \ supp(ρ) for some
δ > 0. E∗ is a left edge of ρ if this holds with (E∗ − δ, E∗ ) and (E∗ , E∗ + δ) exchanged.
Alternatively, supp(ρ) is the closure of {x ∈ R∗ : f0 (x) > 0} together with the isolated point 0 if
rank(T ) < N . When 0 is an isolated point, we do not consider it an edge.
∗We define ρ directly by N/M and T , rather than assuming that N/M and the spectrum of T converge to certain
limiting quantities.
6
TRACY-WIDOM FOR COVARIANCE MATRICES
Figure 1. Left: Density f0 (x) of ρ and simulated eigenvalues of Σ̂, for N = 500,
M = 700, and T having 350 eigenvalues at -2, 300 at 0.5, and 50 at 6. The four soft
edges of ρ are indicated by E1 , . . . , E4 . Right: The function z0 (m), with two local
minima and two local maxima corresponding to the four edges of ρ.
Figure 2. Left: Density f0 (x) of ρ and simulated eigenvalues of Σ̂, for N = M =
500, and T having 400 eigenvalues at -1 and 100 at 4. Here, ρ has three soft edges
E1 , E2 , E4 and one hard edge E3 = 0. Right: The function z0 (m), with three
indicated local extrema, and also a local minimum at m = ∞ corresponding to the
hard right edge E3 = 0.
The support intervals and edge locations of ρ are described in a simple way by (13): Define
P = {0} ∪ {−t−1
α : tα 6= 0}, and consider R̄ = R ∪ {∞}. Consider the formal inverse of m0 (z),
z0 (m) = −
M
1
1 X
tα
+
,
m N
1 + tα m
(15)
α=1
as a real-valued function on R̄ \ P with the convention z0 (∞) = 0. Then z0 is a rational function
with poles P —two examples are plotted in Figures 1 and 2. The following proposition relates the
edges of ρ to the local extrema of z0 . We indicate its proof in Appendix A. (Parts (a), (b), and (d)
follow from [SC95], and part (c) was established for positive-definite T in [KY16].)
TRACY-WIDOM FOR COVARIANCE MATRICES
7
Proposition 2.2. Let m1 , . . . , mn ∈ R̄\P denote the local minima and local maxima† of z0 , ordered
such that 0 > m1 > . . . > mk > −∞ and ∞ ≥ mk+1 > . . . > mn > 0. Let Ej = z0 (mj ) for each
j = 1, . . . , n. Then:
(a) ρ has exactly n/2 support intervals and n edges, which are given by E1 , . . . , En .
(b) Ej is a right edge if mj is a local minimum, and a left edge if mj is a local maximum.
(c) The edges are ordered as
E1 > . . . > Ek > Ek+1 > . . . > En .
00
(d) For each Ej where mj 6= ∞, we have Ep
j ∈ R∗ and z0 (mj ) 6= 0. Defining γj =
the density of ρ satisfies f0 (x) ∼ (γj /π) |Ej − x| as x → Ej with x ∈ supp(ρ).
p
2/|z000 (mj )|,
Definition 2.3. For each edge E∗ of ρ, the local minimum or maximum m∗ of z0 such that
z0 (mp
6 ∞ and hard if m∗ = ∞. For a soft edge,
∗ ) = E∗ is its m -value. The edge is soft if m∗ =
γ = 2/|z000 (m∗ )| is its associated scale.
Hence the local extrema of z0 are in 1-to-1 correspondence with the edges of ρ. Excluding the
point mass at 0 when rank(T ) < N , supp(ρ) is exactly [En , En−1 ] ∪ [En−2 , En−3 ] ∪ . . . ∪ [E2 , E1 ],
where these intervals are disjoint and in increasing order. The density f0 exhibits square-root decay
at each soft edge E∗ , with scale inversely related to the curvature of z0 at m∗ .
When T is positive-semidefinite, supp(ρ) is nonnegative. In this setting, an edge at 0 is usually
called hard and all other edges soft. The above definition generalizes this to non-positive-definite
T : A hard edge is always 0 and can occur when rank(T ) = N . One example is depicted in Figure
2. However, if T has negative eigenvalues, then a soft edge may also be 0 when rank(T ) > N .
2.2. Edge regularity and extremal eigenvalues. We state our assumptions and introduce the
notion of a “regular” edge of ρ. For positive-definite T , regularity was introduced for the rightmost
edge in [Kar07] and generalized to all soft edges in [HHN16, KY16].
Throughout, we denote the dimensionality ratio by
d = N/M.
We make the following assumptions regarding d and the scaling of T :
Assumption 2.4. For some constant Cd,T > 0, the ratio d = N/M satisfies
−1
< d < Cd,T .
Cd,T
T ∈ RM ×M is symmetric, and all of its eigenvalues {tα : α = 1, . . . , M } satisfy
|tα |< Cd,T .
Let us state here, for clarity, that the existence of any regular edge will imply T is also nondegenerate, in the sense
(number of eigenvalues tα such that |tα |> c) > cM
(16)
for a Cd,T -dependent constant c > 0. (See Proposition 3.5.) Thus the largest and average values of
|tα | are both of constant order.
We assume the following weak moment conditions for X:
Assumption 2.5. X ∈ RM ×N is random with independent entries. For every α ∈ {1, . . . , M } and
i ∈ {1, . . . , N },
2
E[Xαi ] = 0,
E[Xαi
] = 1/N.
†m ∈ R̄ \ P is a local minimum of z if z (m) ≥ z (m ) for all m in a sufficiently small neighborhood of m ,
∗
0
0
0
∗
∗
with the convention that m∗ = ∞ is a local minimum if z0 is positive over (C, ∞) ∪ (−∞, −C) for some C > 0. Local
maxima are defined similarly.
8
TRACY-WIDOM FOR COVARIANCE MATRICES
For some constants C1 , C2 , . . . > 0 and every l ≥ 1,
√
E[| N Xαi |l ] ≤ Cl .
We will establish a Tracy-Widom limit at each edge of ρ that is regular in the following sense:
Definition 2.6. Let E∗ ∈ R be a soft edge of ρ with m-value m∗ and scale γ. E∗ is regular if all
of the following hold for a constant τ ∈ (0, 1):
• |m∗ |< τ −1 .
• γ < τ −1 .
• For all α ∈ {1, . . . , M } such that tα 6= 0, |m∗ + t−1
α |> τ .
A smaller constant τ indicates a weaker regularity assumption. We will say E∗ is τ -regular if we
wish to emphasize the role of τ . All subsequent constants may depend on Cd,T , Cl , τ above; we will
usually not explicitly state this dependence.
We discuss implications of edge regularity in Section 3.3. One interpretation of this condition is
the following, whose proof we defer to Appendix A.
Proposition 2.7. Suppose Assumption 2.4 holds and the edge E∗ is regular. Then there exist
constants C, c, δ > 0 such that
(a) (Separation) The interval (E∗ − δ, E∗ + δ) belongs to R∗ and contains no edge other than E∗ .
(b) (Square-root decay) For all x ∈ supp(ρ) ∩ (E∗ − δ, E∗ + δ), the density f0 of ρ satisfies
p
p
c |E∗ − x| ≤ f0 (x) ≤ C |E∗ − x|.
Thus if E∗ is a right edge, then (E∗ − δ, E∗ ) ⊂ supp(ρ) while (E∗ , E∗ + δ) ⊂ R \ supp(ρ). If E∗
is a left edge, this holds with (E∗ − δ, E∗ ) and (E∗ , E∗ + δ) exchanged. Whereas Definition 2.1 and
Proposition 2.2(d) imply the above for C, c, δ depending on N , edge regularity ensures that these
depend only on the constants Cd,T and τ of Assumption 2.4 and Definition 2.6, and not on N .
Hence edge separation and square-root decay hold uniformly in N .
One may check, via Proposition 3.5 below, that Definition 2.6 is equivalent to the definition of a
regular edge in [KY16] when T is positive-definite. The condition |m∗ |< τ −1 quantifies softness of
E∗ , so E∗ cannot converge to a hard edge at 0. The condition γ < τ −1 guarantees non-vanishing
curvature of z0 at m∗ , so E∗ cannot approach a neighboring interval of support. The condition
|m∗ + t−1
α |> τ guarantees separation of m∗ from the poles P of z0 ; this implies, in particular, that
E∗ cannot be the edge of a support interval for an outlier eigenvalue of Σ̂. This last condition was
introduced for the rightmost edge in [Kar07]. In the setting of a simple spiked model [Joh01] where
(t1 , . . . , tM ) = (θ, 1, 1, . . . , 1) for fixed θ > 1, if E∗ is the rightmost edge, then it is easily
√ verified
that this condition is equivalent to θ falling below the phase transition threshold 1 + 1/ d studied
in [BBAP05, BS06, Pau07].
We will study the extremal eigenvalue of Σ̂ at each regular edge. This is well-defined by the
following results establishing closeness of eigenvalues of Σ̂ to the support of ρ.
Theorem 2.8 (No eigenvalues outside support). Suppose Assumptions 2.4 and 2.5 hold. Fix any
constants δ, D > 0, and define the δ-neighborhood of supp(ρ) by
supp(ρ)δ = {x ∈ R : there exists y ∈ supp(ρ) such that |x − y|< δ}.
Then there exists a constant N0 ≡ N0 (δ, D) such that for all N ≥ N0 ,
P[all eigenvalues of Σ̂ belong to supp(ρ)δ ] > 1 − N −D .
Theorem 2.9 (N −2/3 concentration at regular edges). Suppose Assumptions 2.4 and 2.5 hold, and
E∗ is a regular right edge. Then there exists a constant δ > 0 such that for any ε, D > 0, some
N0 ≡ N0 (ε, D), and all N ≥ N0 ,
h
i
P no eigenvalue of Σ̂ belongs to [E∗ + N −2/3+ε , E∗ + δ] > 1 − N −D .
(17)
TRACY-WIDOM FOR COVARIANCE MATRICES
9
The analogous statement holds if E∗ is a regular left edge, with no eigenvalue of Σ̂ belonging to
[E∗ − δ, E∗ − N −2/3+ε ].
Theorem 2.8 ensures that all eigenvalues of Σ̂ are within a δ-neighborhood of the deterministic
support, for any fixed δ as N, M → ∞. (This does not require edge regularity.) Theorem 2.9
then ensures that for sufficiently small δ, within this δ-neighborhood of any regular edge E∗ , the
eigenvalue of Σ̂ farthest from the support is typically only at distance N −2/3 . This is the extremal
eigenvalue at E∗ , whose fluctuations on the N −2/3 scale will follow the Tracy-Widom law.
Results similar to Theorems 2.8 and 2.9 were established in [BS98, KY16] for positive-definite
T . We check in Appendix B that the proof of [KY16] generalizes to any symmetric T .
2.3. Tracy-Widom fluctuations at regular edges. The following is our main result:
Theorem 2.10. Let Σ̂ = X 0 T X. Suppose that Assumptions 2.4 and 2.5 hold for T and X, and
that E∗ is a τ -regular edge of the law ρ ≡ ρN,M,T . Let E∗ have scale γ as defined in Definition 2.3.
Denote by µT W the Tracy-Widom F1 law (corresponding to the β = 1 case from [TW96]). Then
there exists a constant δ ≡ δ(τ ) such that as N → ∞,
(a) If E∗ is a right edge and λmax is the largest eigenvalue of Σ̂ in [E∗ − δ, E∗ + δ], then
L
(γN )2/3 (λmax − E∗ ) → µT W .
(b) If E∗ is a left edge and λmin is the smallest eigenvalue of Σ̂ in [E∗ − δ, E∗ + δ], then
L
(γN )2/3 (E∗ − λmin ) → µT W .
L
The notation → indicates convergence in law. This is interpreted as follows: Let F1 denote the
cumulative distribution function of µT W , and fix x ∈ R. Then
P[(γN )2/3 (λmax − E∗ ) ≤ x] − F1 (x) ≤ o(1),
where o(1) denotes an error term going to 0 as N, M → ∞ and that depends only on x, τ , and the
constants in Assumptions 2.4 and 2.5. (Equivalently, we may interpret the above in the context of
a sequence of matrices of growing size, each satisfying Assumptions 2.4 and 2.5 and each with an
identified τ -regular edge E∗ .) We remark that as we are making no sign convention for T and E∗ ,
part (b) follows from part (a) by considering the reflection T 7→ −T .
When T is positive-semidefinite, the sample covariance matrix Σ̃ = T 1/2 XX 0 T 1/2 has the same
eigenvalues as Σ̂ except for a set of |N − M | zeros. Thus the above result also holds for Σ̃ with the
same values for E∗ and γ:
Corollary 2.11. Under the conditions of Theorem 2.10, suppose t1 , . . . , tM ≥ 0, and let Σ̃ =
T 1/2 XX 0 T 1/2 . Then Theorem 2.10 holds also with Σ̃ in place of Σ̂.
When T = Id, the equation 0 = z00 (m∗ ) may be solved explicitly to yield
!
√
√
√
√
√
1/3
N
( N ± M )2
N± M
1
1
−2/3
√
√ ±√
,
E∗ =
,
(γN )
=
m∗ = − √
N
N
N± M
N
M
for the upper and lower edge. These centering and scaling constants are the same as those of
[Sos02, Péc09, FS10] and differ from those of [Joh01, Ma12] in small O(1) adjustments to N and
M . These adjustments do not affect the validity of Theorem 2.10, although the proper adjustments
are shown in [Ma12] to lead to an improved second-order rate of convergence. The determination
of the rate of convergence and of such adjustments remain open in the setting of general T .
10
TRACY-WIDOM FOR COVARIANCE MATRICES
2.4. Application to variance components models. In the variance components model (3),
when p, n, and I1 , . . . , Ik are large and of comparable magnitude, Theorem 2.10 provides an asymptotic test of the global sphericity null hypothesis
H0 : Σr = c2r Id for every r = 1, . . . , k
based on the largest observed eigenvalue λmax of a MANOVA estimator Σ̂ = Y 0 BY . If H0 specifies
Σr = 0 for some r, then an intuitive choice of B to ensure reasonable power is B = Br for
the MANOVA estimator Σ̂r . Note that by the representation (5), in this context N = p and
M = I1 + . . . + Ik , and the spectral norm of B satisfies kBk n−1 for usual MANOVA estimators.
In detail, the test of H0 may be performed as follows:
1. Define
c1 U10
U = ... ∈ R(I1 +...+Ik )×n ,
ck Uk0
and construct the matrix
T = pU BU 0 .
2. Compute the eigenvalues t1 ≥ . . . ≥ tM of T , plot
M
1X
tα
1
z0 (m) = − +
m p
1 + tα m
α=1
over m ∈ R, and locate the unique local minimum m∗ of this function
in the range (−1/t1 , 0).
p
3. Compute E∗ and γ in Theorem 2.10 as E∗ = z0 (m∗ ) and γ = 2/z000 (m∗ ).
4. Reject H0 at level α if (γp)2/3 (λmax − E∗ ) exceeds the 1 − α quantile of the Tracy-Widom F1
law.
Remark 2.12. If a scaling constant cr is unknown and not specified by H0 , it may be replaced
by any 1/p-consistent estimate ĉr in the above. (For example, one may take the linear spectral
statistic ĉr = p−1 Tr Σ̂r .) The asymptotic validity of this substitution is justified by the fact that
the error in the corresponding estimate of T is O(1/p) in spectral norm, which propagates to a
pointwise O(1/p) error in z0 (m) near m∗ and a O(1/p) error in the estimated centering and scaling
constants E∗ and γ; this follows from the argument of Lemma 4.4. In particular, the error in E∗ is
smaller than O(p−2/3 ) for large p, so Theorem 2.10 remains asymptotically valid for the estimated
center and scale.
A similar procedure may be used to test H0 based on the extremal eigenvalue observed near any
edge E∗ , where m∗ is identified by the correspondence established in Proposition 2.2.
For application of Theorem 2.10 in any finite-sample setting, we expect the accuracy of the TracyWidom approximation to depend on the extent to which the regularity condition of Definition 2.6
holds at E∗ . This is quantified by the separation between m∗ and the asymptotes of z0 , as well as
the curvature of z0 at m∗ . We will not pursue a detailed numerical investigation of this dependence
here, but let us note that one may visually inspect the plot of z0 for a qualitative diagnostic check
of edge regularity. For balanced classification designs where a constant fraction of the eigenvalues
of T equal its largest eigenvalue, edge regularity always holds at the rightmost edge.
Example 2.13. As a concrete example, consider the balanced one-way classification model in the
introduction, where n samples are divided into I = n/J groups of equal size J. The data is modeled
as
Y = U α + ε,
I×p
where the rows of α ∈ R
are distributed as Normal(0, Σ1 ), the rows of ε ∈ Rn×p are distributed as
Normal(0, Σ2 ), and U ∈ {0, 1}n×I has Uk,i = 1 if bk/Jc = i and Uk,i = 0 otherwise. Defining π, π ⊥ ∈
TRACY-WIDOM FOR COVARIANCE MATRICES
11
Table 1. Empirical cumulative probabilities for (γp)2/3 (λmax (Σ̂1 ) − E∗ ) at the theoretical 90th, 95th, and 99th percentiles of the Tracy-Widom F1 law, estimated
across 10000 simulations. Here, Σ̂1 is the MANOVA estimator of Σ1 in the balanced
one-way classification model, for various n, p, J. The final column gives approximate
standard errors based on binomial sampling.
F1
J =2
n=p
J =5
J = 10
J =2
n=4×p
J = 5 J = 10
2 × SE
0.90
0.95
0.99
0.938
0.971
0.995
0.944
0.974
0.995
0.953
0.978
0.995
0.932
0.968
0.993
0.937
0.970
0.994
0.940 (0.005)
0.972 (0.003)
0.995 (0.002)
0.90
p = 100 0.95
0.99
0.926
0.963
0.992
0.934
0.969
0.995
0.931
0.968
0.995
0.923
0.962
0.993
0.919
0.961
0.993
0.918 (0.005)
0.961 (0.004)
0.994 (0.002)
0.90
p = 500 0.95
0.99
0.922
0.961
0.992
0.917
0.958
0.992
0.918
0.959
0.991
0.913
0.956
0.993
0.909
0.954
0.992
0.916 (0.006)
0.959 (0.004)
0.994 (0.002)
p = 20
Rn×n as the orthogonal projections onto the column span of U and its orthogonal complement, the
MANOVA estimators for Σ1 and Σ2 are
1
1
1
0
⊥
0
⊥
Σ̂1 = Y
π−
π
Y,
Σ̂2 = Y
π
Y.
n
n(J − 1)
I(J − 1)
The example of sibling pairs discussed in the introduction corresponds to J = 2.
Suppose we wish to test
H0 : Σ1 = 0, Σ2 = c22 Id
using the largest observed eigenvalue of Σ̂1 . Setting B as the estimation matrix defining Σ̂1 , and
noting π = J −1 U U 0 and π ⊥ = Id −π, we obtain
!
2
0
p c1 J Id
c1 c2 U
T =
∈ R(I+n)×(I+n) .
1
π⊥
n c1 c2 U c22 π − J−1
Explicit diagonalization yields that T has I eigenvalues (p/n)(c21 J + c22 ), I eigenvalues 0, and n − I
eigenvalues −(p/n)(c22 /(J − 1)). Then
z0 (m) = −
1
I
1
n−I
1
+ ·
+
·
−1
m p m + t1
p
m + t−1
2
(18)
where
p 2
p
c22
(c1 J + c22 ),
t2 = − ·
.
n
n J −1
(Figure 2 in fact corresponds to this example with p = n = 500, J = 5, c1 = 0, and c2 = 2.) This is
easily generalized if the model includes a fixed effect term for mean-centering, Y = 1µT + U α + ε
where µ ∈ Rp and where 1 is the length n all-1’s vector. In this case, π in the definition of Σ̂1
also projects out 1 (and π ⊥ remains unchanged), and the factor I/p in the second term of (18) is
replaced by (I − 1)/p.
Table 1 displays the accuracy of the Tracy-Widom approximation for the standardized largest
eigenvalue (γp)2/3 (λmax (Σ̂1 ) − E∗ ), under various settings of n, p, and group size J. The center
and scale E∗ , γ are computed from z0 (m) above, where we have assumed for simplicity that c2 = 1
t1 =
12
TRACY-WIDOM FOR COVARIANCE MATRICES
is known. The approximation is reasonably accurate but has a conservative bias, particularly for
small sample sizes.
3. Preliminaries and tools
The remainder of this paper is devoted to the proof of Theorem 2.10. We collect in this section
the requisite ingredients and tools for the proof.
3.1. Notation. We denote IM = {1, . . . , M } and IN = {1, . . . , N }. Considering the elements
of these index sets as distinct, we define the disjoint union I ≡ IN t IM . For a matrix in
C(N +M )×(N +M ) , we identify {1, 2, . . . , N + M } ' I and index its rows and columns by I, where
IN corresponds to the upper left block and IM to the lower right block. We consistently use lowercase Roman letters i, j, p, q, etc. for indices in IN , Greek letters α, β, etc. for indices in IM , and
upper-case Roman letters A, B, etc. for general indices in I.
=z and <z denote the real and imaginary parts of z. We typically write z = E + iη where
E = <z and η = =z. C+ = {z ∈ C : =z > 0} and C+ = {z ∈ C : =z ≥ 0} denote the open and
closed upper-half complex plane.
X 0 denotes the transpose of a matrix X, and kXk denotes its Euclidean operator norm. For
vectors v, w, hv, wi denotes the Euclidean inner product and kvk= hv, vi1/2 the Euclidean norm.
Throughout, C, c denote positive constants that may change from instance to instance. We write
aN bN for deterministic non-negative quantities aN , bN when cbN ≤ aN ≤ CaN . The constants
C, c may depend on τ in the context of a regular edge.
3.2. Stochastic domination. We use the following the definition introduced in [EKY13] to express bounds that hold with high probability:
Definition 3.1 (Stochastic domination). Let
ξ = (ξ(u) : u ∈ U ),
Ψ = (Ψ(u) : u ∈ U )
be two collections of (N -dependent) random variables indexed by a (N -dependent) set U , where
ξ(u) is complex-valued and Ψ(u) is real-valued and nonnegative. We say that
ξ ≺ Ψ or ξ = O≺ (Ψ)
uniformly over u ∈ U if, for any constants ε, D > 0, there exists N0 ≡ N0 (ε, D) such that for all
N ≥ N0 ,
sup P [|ξ(u)|> N ε Ψ(u)] < N −D .
(19)
u∈U
N0 may implicitly depend on other constants, such as τ in the context of a regular edge. To
explicitly emphasize this dependence for some constant a, we write ξ ≺a Ψ.
In this paper, u ∈ U will typically correspond to indices in IN , IM , I and/or z in bounded
domains of C+ . All uses of the above definition will be implicitly uniform over all such u ∈ U .
With the exception of the proof sketch of the local law in Appendix B, Ψ(u) in this paper will
always be deterministic. If ξ(u) is also deterministic, then ξ ≺ Ψ means |ξ(u)|< Ψ(u)N ε for any
ε > 0, all N ≥ N0 (ε), and all u ∈ U .
We review several properties of ≺ from [EKY13]: First, we may take union bounds, sums, and
products over sets of cardinality N C :
Lemma 3.2. Let ξ(u, v) and Ψ(u, v) be such that ξ(u, v) ≺ Ψ(u, v) uniformly over u ∈ U and
v ∈ V . Suppose the cardinality of V satisfies |V |≤ N C for a constant C > 0. Then:
(a) For any ε, D > 0 and all N ≥ N0 (ε, D),
sup P [there exists v ∈ V such that |ξ(u, v)|> N ε Ψ(u, v)] < N −D .
u∈U
(b)
P
v∈V
ξ(u, v) ≺
P
v∈V
Ψ(u, v) uniformly over u ∈ U .
TRACY-WIDOM FOR COVARIANCE MATRICES
(c)
Q
v∈V
ξ(u, v) ≺
Q
v∈V
13
Ψ(u, v) uniformly over u ∈ U .
Proof. All three parts follow from a simple union bound. For example, for (c),
"
#
i
Y
Y
X h
sup P
|ξ(u, v)|> N ε
Ψ(u, v) ≤ sup
P |ξ(u, v)|> N ε/C Ψ(u, v)
u∈U
v∈V
u∈U v∈V
v∈V
≤ NC
sup
h
i
P ξ(u, v) > N ε/C Ψ(u, v) < N −D
u∈U,v∈V
for any ε, D > 0 and all N ≥ N0 (ε, D). Parts (a) and (b) follow similarly.
When V is a bounded domain of C+ , part (a) above does not directly apply, but we may
oftentimes take the union bound by Lipschitz continuity:
Lemma 3.3. Let ξ(u, z), Ψ(u, z) be such that ξ(u, z) ≺ Ψ(u, z) uniformly over u ∈ U and z ∈ V ,
where V ⊂ C is uniformly bounded. Suppose that for any D > 0, there exists C ≡ C(D) > 0 and
an event of probability 1 − N −D on which
• Ψ(u, z) > N −C for every u, z.
• ξ and Ψ are N C -Lipschitz continuous over z ∈ V for every u.
Then for any ε, D > 0 and all N ≥ N0 (ε, D),
sup P[there exists z ∈ V such that |ξ(u, z)|> N ε Ψ(u, z)] < N −D .
u∈U
Proof. Let D > 0, C = C(D), and set ∆N = N −3C . Take a covering net V 0 ⊂ V such that for
every z ∈ V , there exists z 0 ∈ V 0 with |z − z 0 |< ∆N , and also |V 0 |≤ N 6C+1 . Let u ∈ U and ε > 0.
Then applying Lemma 3.2(a) to union bound over V 0 , with probability at least 1 − 2N −D , for
N ≥ N0 (ε/2, D) and every z ∈ V ,
|ξ(u, z)|≤ |ξ(u, z 0 )|+∆N N C ≤ N ε/2 Ψ(u, z 0 ) + ∆N N C ≤ N ε/2 Ψ(u, z) + 2∆N N ε/2+C ≤ N ε Ψ(u, z).
Under mild conditions, we may also take (conditional) expectations over the above definition:
Lemma 3.4. Let ξ(u) and Ψ(u) be such that ξ ≺ Ψ uniformly over u ∈ U . Suppose furthermore
that for some constant C > 0 and all u ∈ U , Ψ(u) is deterministic with Ψ(u) ≥ N −C , and
E[|ξ(u)|2 ] ≤ N C . Then E[ξ|G] ≺ Ψ uniformly over u ∈ U and all sub-σ-fields G.
Proof. If G is trivial so E[ξ|G] = E[ξ], then this follows from Cauchy-Schwarz: For any ε > 0 and
all N ≥ N0 (ε, C),
|Eξ| ≤ E [|ξ|1{|ξ|≤ N ε Ψ}] + E [|ξ|1{|ξ|> N ε Ψ}]
1/2
≤ N ε Ψ + E |ξ|2
P [|ξ|> N ε Ψ]1/2 ≤ N ε Ψ + N C/2 N −2C ≤ 2N ε Ψ.
For general G, note that for any p ∈ N and any ε > 0,
h
i E[|E[ξ|G]|p ]
E[|ξ|p ]
P |E[ξ|G]|> N ε Ψ ≤
≤
≤ N ε−pε ,
N pε Ψp
N pε Ψp
where the last step holds for all N ≥ N0 (p, ε) by E[|ξ|p ] ≺ Ψp established above. (N0 does not
depend on G.) For any D > 0, this error is at most N −D by taking p > (D + ε)/ε.
14
TRACY-WIDOM FOR COVARIANCE MATRICES
3.3. Implications of edge regularity. We state several consequences of edge regularity. The
first proposition is directly verified from Definition 2.6.
Proposition 3.5. Suppose Assumption 2.4 holds and E∗ is a regular edge with m-value m∗ and
scale γ. Then there exist constants C, c > 0 such that
c < |m∗ |< C,
c < γ < C,
|E∗ |< C,
and for all α = 1, . . . , M ,
|1 + tα m∗ |> c.
Furthermore, if any regular edge E∗ exists, then T satisfies (16), and if T is positive-semidefinite,
then also E∗ > c > 0.
Proof. |m∗ |< τ −1 and γ < τ −1 are assumed in Definition 2.6. Writing
1
1 X
1
0 = z00 (m∗ ) = 2 −
,
2
m∗ dM
(m∗ + t−1
α )
α:t 6=0
α
|m∗ |> c follows from d > c and |m∗ + t−1
α |> τ . |E∗ |< C and γ > c then follow from the definitions
E∗ = z0 (m∗ ) and γ −2 = |z000 (m∗ )|/2 together with the bounds d > c, |m∗ |> c, and |m∗ + t−1
α |> τ .
For |1 + tα m∗ |, take C > 0 such that |m∗ |< C. If |tα |> 1/(2C), then |1 + tα m∗ |> τ /(2C) by the
condition |m∗ + t−1
α |> τ , whereas if |tα |≤ 1/(2C), then |1 + tα m∗ |> 1/2.
P
We verify from 0 = z00 (m∗ ) and |m∗ |< C, d > c, |1 + tα m∗ |> c that M −1 α t2α > c. Together
0
with the assumption
P 2 |tα |< C for all2 α, this implies (16). Finally, note that 0 = z0 (m∗ ) implies
−1
−1
m∗ = (dM )
α tα m∗ /(1 + tα m∗ ) , and hence
E∗ = z0 (m∗ ) =
M
1 X
tα
.
dM
(1 + tα m∗ )2
α=1
If T is positive-semidefinite, then E∗ > c follows from |1 + tα m∗ |< C, d < C, and (16).
The remaining implications heuristically follow from the Taylor expansion
z000 (m∗ )
(m − m∗ )2 + O((m − m∗ )3 ),
2
where there is no first-order term because 0 = z00 (m∗ ). Consequently,
s
2
m0 (z) ≈ m∗ +
(z − E∗ )
00
z0 (m∗ )
z0 (m) − E∗ = z0 (m) − z0 (m∗ ) =
for z ∈ C+ near E∗ and an appropriate choice of square-root. Edge regularity implies uniform
control of the above Taylor expansion; we defer detailed proofs to Appendix A. Similar properties
were established for positive-definite T in [BPZ13, KY16].
Proposition 3.6. Suppose Assumption 2.4 holds and E∗ is a regular edge with m-value m∗ . Then
there exist constants c, δ > 0 such that for all m ∈ (m∗ − δ, m∗ + δ), if E∗ is a right edge then
z000 (m) > c,
and if E∗ is a left edge then z000 (m) < −c.
Proposition 3.7. Suppose Assumption 2.4 holds and E∗ is a regular edge. Then there exist constants C, c, δ > 0 such that the following hold: Define
D0 = {z ∈ C+ : <z ∈ (E∗ − δ, E∗ + δ), =z ∈ (0, 1]}.
Then for all z ∈ D0 and α ∈ {1, . . . , M },
c < |m0 (z)|< C,
c < |1 + tα m0 (z)|< C.
TRACY-WIDOM FOR COVARIANCE MATRICES
15
Furthermore, for all z ∈ D0 , denoting z = E + iη and κ = |E − E∗ |,
√
√
c κ + η ≤ |m0 (z) − m∗ |≤ C κ + η,
cf (z) ≤ =m0 (z) ≤ Cf (z)
where
(√
f (z) =
κ+η
√η
κ+η
if E ∈ supp(ρ)
if E ∈
/ supp(ρ).
3.4. Properties of the resolvent. For z ∈ C+ , denote the resolvent and Stieltjes transform of Σ̂
by
GN (z) = (Σ̂ − z Id)−1 ∈ CN ×N ,
mN (z) = N −1 Tr GN (z).
(20)
The entries of GN and mN are bounded and Lipschitz-continuous:
Lemma 3.8. For any η > 0 and z, z 0 ∈ C+ with =z ≥ η and =z 0 ≥ η, and for any i, j ∈ IN ,
|mN (z)|≤
|mN (z) − mN (z 0 )|≤
1
,
η
|z − z 0 |
,
η2
|Gij (z)|≤
1
,
η
|Gij (z) − Gij (z 0 )|≤
|z − z 0 |
.
η2
P
P
Proof. Let Σ̂ = i λi ξi ξi0 be the (real) spectral decomposition of Σ̂. Then GN (z) = i (λi −z)−1 ξi ξi0 ,
so for any v, w ∈ RN with kvk= kwk= 1,
X hv, ξi ihw, ξi i
d
1 X
|hv, wi|
1
hv, GN (z)wi =
≤ 2
|hv, ξi ihw, ξi i|=
≤ 2.
2
2
dz
(λi − z)
η
η
η
i
i
Taking v, w as standard basis vectors yields the Lipschitz bound for Gij , and averaging over i = j
yields the bound for mN . The bounds |Gij |, |mN |≤ η −1 follow similarly.
As in [LS16, KY16], define the linearized resolvent G(z) by
−z Id
X0
H(z) =
∈ C(N +M )×(N +M ) ,
X
−T −1
G(z) = H(z)−1 .
(21)
We index rows and columns of G(z) by I ≡ IN t IM . The Schur-complement formula for block
matrix inversion yields the alternative form
GN (z)
GN (z)X 0 T
G(z) =
,
(22)
T XGN (z) T XGN (z)X 0 T − T
which is understood as the definition of G(z) when T is not invertible. We will omit the argument
z in m0 , mN , GN , G when the meaning is clear.
For any A ∈ I, define H (A) as the submatrix of H with row and column A removed, and define
G(A) = (H (A) )−1 .
When T is not invertible, G(A) is defined by the alternative form analogous to (22). We index G(A)
by I \ {A}.
Note that G and G(A) are symmetric, in the sense G0 = G and (G(A) )0 = G(A) without complex
conjugation. Restricting to the case of diagonal T , the entries of G and G(A) are related by the
following Schur-complement identities:
Lemma 3.9 (Resolvent identities). Suppose T is diagonal, and fix z ∈ C+ .
16
TRACY-WIDOM FOR COVARIANCE MATRICES
(a) For any i ∈ IN ,
Gii = −
1
z+
.
(i)
P
α,β∈IM
Gαβ Xαi Xβi
For any α ∈ IM ,
Gαα = −
tα
1 + tα
P
.
(α)
i,j∈IN
Gij Xαi Xαj
X
(i)
(b) For any i 6= j ∈ IN ,
Gij = −Gii
Gβj Xβi .
β∈IM
For any α 6= β ∈ IM ,
X
Gαβ = −Gαα
(α)
Gjβ Xαj .
j∈IN
For any α ∈ IM and i ∈ IN ,
Giα = −Gii
X
(i)
Gβα Xβi = −Gαα
β∈IM
X
(α)
Gij Xαj .
j∈IN
(c) For any A, B, C ∈ I with A 6= C and B 6= C,
(C)
GAB = GAB −
GAC GCB
.
GCC
Proof. For T invertible, see [KY16, Lemma 4.4]. (These are algebraic relations derived from the
Schur-complement formula for matrix inversion; they do not require positivity of T .) The identities
for non-invertible T follow by continuity in T and the alternative form for G in (22).
3.5. Local law and reduction to diagonal T . As in [LS16], an important proof ingredient is a
high-probability entrywise bound on G(z) for z ∈ C+ close to a regular edge E∗ . This comes from a
local law establishing a deterministic approximation Π0 for G. Such a law was first established for
the Wigner model in [EYY12a]. In our setting, for T positive-definite, this was proven in [KY16]
near all regular edges (and in regular bulk components); see also [BPZ13, LS16] for the rightmost
edge. We check in Appendix B that the proof of [KY16] generalizes to arbitrary symmetric T .
Theorem 3.10 (Local law at regular edges). Suppose Assumptions 2.4 and 2.5 hold and E∗ is a
τ -regular edge. Then there exists a constant δ ≡ δ(τ ) > 0 such that the following holds: Fix any
constant a > 0 and define
D = {z ∈ C+ : <z ∈ (E∗ − δ, E∗ + δ), =z ∈ [N −1+a , 1]}.
For z ∈ D, define the (N + M ) × (N + M ) matrices
m0 (z) Id
0
Π0 (z) =
,
0
−T (Id +m0 (z)T )−1
Id 0
.
T =
0 T
Then uniformly over z ≡ E + iη ∈ D and vectors v, w ∈ CN +M with kvk= kwk= 1,
s
=m0 (z)
1
−1
−1
hv, T (G(z) − Π0 (z))T wi ≺τ,a
+
,
Nη
Nη
and also
mN (z) − m0 (z) ≺τ,a (N η)−1 .
(Here, τ is the constant from Definition 2.6, and a is the constant defining D.)
(23)
(24)
(25)
TRACY-WIDOM FOR COVARIANCE MATRICES
17
It is verified from (22) that the matrix on the left of (25) is alternatively written as
GN − m0 Id
GN X 0
T −1 (G − Π0 )T −1 =
.
XGN
XGN X 0 − m0 (Id +m0 T )−1
(26)
This is understood as the definition of this matrix when T is not invertible.
Lipschitz continuity of G and mN imply the above result in union bound form:
Corollary 3.11. Under the assumptions of Theorem 3.10, for any ε, D > 0 and all N ≥ N0 (ε, D),
"
P there exist z ∈ D and v, w ∈ CN +M with kvk= kwk= 1 :
s
hv, T
−1
(G(z) − Π0 (z))T
−1
wi > N
ε
=m0 (z)
1
+
Nη
Nη
!#
< N −D .
Proof. This follows from Lemma 3.3. For any D > 0, on an event of probability 1 − N −D , we
have kX ∗ Xk≤ C (see e.g. [BEK+ 14, Theorem 2.10]). The required boundedness and Lipschitz
continuity properties for Lemma 3.3 then follow from (26), Lemma 3.8, Proposition 3.7, and known
properties of any Stieltjes transform m0 .
As in [KY16], Theorem 3.10 is sufficient to establish universality (with respect to the distribution
of entries of X) of the fluctuations of the extremal eigenvalue at E∗ . This follows the argument of
[EYY12b], which we discuss briefly in Appendix B. Hence we may reduce the proof of Theorem
2.10 to the case of Gaussian X. By diagonalizing T and applying rotational symmetry of X, we
may also reduce to diagonal T :
Corollary 3.12. If Theorem 2.10 holds under the additional conditions that T is diagonal and X
has (real) Gaussian entries, then it holds also without these conditions.
3.6. Resolvent approximation. We formalize the approximation (7) stated in the introduction.
This approximation was introduced in [EYY12b].
Fix a regular edge E∗ and define, for s1 , s2 ∈ R and η > 0,
Z E∗ +s2
X(s1 , s2 , η) = N
=mN (y + iη)dy.
(27)
E∗ +s1
For η much smaller than N −2/3 and s1 , s2 on the N −2/3 scale, we expect
#(E∗ + s1 , E∗ + s2 ) ≈ π −1 X(s1 , s2 , η)
where the left side denotes the number of eigenvalues of Σ̂ in this interval.
We apply this in the form of the following lemma, and provide a self-contained proof using the
local law following the approach of [EYY12b, Corollary 6.2]. (For simplicity, we state the result
only for a right edge.)
Lemma 3.13. Suppose Assumptions 2.4 and 2.5 hold, and E∗ is a regular right edge. Let K : R →
[0, 1] be such that K(x) = 1 for all x ≤ 1/3 and K(x) = 0 for all x ≥ 2/3. For all sufficiently small
constants δ, ε > 0, the following holds:
Let λmax be the maximum eigenvalue of Σ̂ in (E∗ − δ, E∗ + δ). Set s+ = N −2/3+ε , l = N −2/3−ε ,
and η = N −2/3−9ε . Then for any D > 0, all N ≥ N0 (ε, D), and all s ∈ [−s+ , s+ ],
E K(π −1 X(s − l, s+ , η)) − N −D ≤ P [λmax (W ) ≤ E∗ + s] ≤ E K(π −1 X(s + l, s+ , η)) + N −D .
Proof. Denote
#(a, b) = number of eigenvalues of Σ̂ in [a, b].
18
TRACY-WIDOM FOR COVARIANCE MATRICES
For any E1 < E2 , any m > 0, and any λ ∈ R, we have the casewise bound
E −E
η
2
1
if λ < E1 − m
π
η 2 +(E1 −λ)2
if E1 − m ≤ λ ≤ E1 + m
Z E2
1
1
η
2 η
1[E1 ,E2 ] (λ) −
dx
≤
if E1 + m < λ < E2 − m
πm
2
2
E1 π η + (x − λ)
1
if
E2 − m ≤ λ ≤ E2 + m
η
E2 −E1
if λ > E2 + m,
π
η 2 +(λ−E2 )2
where the middle case E1 + m < λ < E2 − m follows from
Z E2
Z λ+m
2 η
1
η
1
η
2
−1 η
1−
dx
≤
1
−
dx
=
tan
≤
.
2
2
2
2
π
m
πm
E1 π η + (x − λ)
λ−m π η + (x − λ)
For the first case, we apply also the bound
η2
η
η
2η
m
≤
≤
· 2
,
2
2
+ (E1 − λ)
(E1 − λ)
m m + (E1 − λ)2
and similarly for the last case. Hence, summing over λ as the eigenvalues of Σ̂,
Z
N E2
=mN (x + iη)dx ≤ R(E1 , E2 , m) + S(E1 , E2 , m)
#(E1 , E2 ) −
π E1
(28)
where we set
R(E1 , E2 , m) = #(E1 − m, E1 + m) + #(E2 − m, E2 + m),
2 η
(E2 − E1 )N =mN (E1 + im) + (E2 − E1 )N =mN (E2 + im)
S(E1 , E2 , m) =
πm
+ #(E1 + m, E2 − m) .
We apply the above with E1 , E2 ∈ [E∗ − 2s+ , E∗ + 2s+ ], and with m = N −2/3−3ε . To bound
S(E1 , E2 , m), note that Proposition 3.7 and Theorem 3.10 yield, for j = 1, 2,
=mN (Ej + im) ≺ N −1/3+3ε .
For z = E∗ + i(2s+ ), Proposition 3.7 and Theorem 3.10 also yield N s+ =mN (z) ≺ N 3ε/2 . Applying
#(E∗ − v, E∗ + v) ≤ 2N v =mN (E∗ + iv) for any v > 0, this yields
#(E∗ − 2s+ , E∗ + 2s+ ) ≺ N 3ε/2 .
(29)
Then applying #(E1 + m, E2 − m) ≤ #(E∗ − 2s+ , E∗ + 2s+ ) and η/m = N −6ε , we obtain
S(E1 , E2 , m) ≺ N −2ε uniformly over E1 , E2 ∈ [E∗ − 2s+ , E∗ + 2s+ ]. By Lemmas 3.8 and 3.3,
we may in fact take a union bound over all such E1 , E2 : For any ε0 , D > 0,
h
i
0
P there exist E1 , E2 ∈ [E∗ − 2s+ , E∗ + 2s+ ] such that S(E1 , E2 , m) > N −2ε+ε ≤ N −D (30)
for all N ≥ N0 (ε0 , D).
Now let E = E∗ + s and E+ = E∗ + s+ − l. Then
Z E+ +l
Z
1 E
#(E, E+ ) ≤ 2
#(E1 , E2 )dE2 dE1
l E−l
E+
Z
Z
Z E+ +l
N E+ +l
1 E
≤
=mN (x + iη)dx + 2
R(E1 , E2 , m)dE2 dE1 + O≺ (N −2ε ),
π E−l
l E−l E+
TRACY-WIDOM FOR COVARIANCE MATRICES
19
where we have applied (28) and (30). The first term is π −1 X(s − l, s+ , η). For the second term, we
obtain from the definition of R(E1 , E2 , m)
Z
Z E+ +l
1 E
2m
2m
R(E1 , E2 , m)dE2 dE1 ≤
#(E − l − m, E + m) +
#(E+ − m, E+ + l + m).
2
l E−l E+
l
l
Applying (29) to crudely bound #(E − l − m, E + m) and #(E+ − m, E+ + l + m) by #(E∗ −
2s+ , E∗ + 2s+ ), and noting m/l = N −2ε , we obtain
#(E, E+ ) ≤ π −1 X(s − l, s+ , η) + O≺ (N −ε/2 ).
Theorem 2.9 yields #(E+ , E∗ + δ) = 0 with probability 1 − N −D for N ≥ N0 (ε, D), so
#(E, E∗ + δ) ≤ π −1 X(s − l, s+ , η) + O≺ (N −ε/2 ).
(31)
Similarly, setting E+ = E∗ + s+ + l, we have
Z E+
Z
1 E+l
#(E, E∗ + δ) ≥ 2
#(E1 , E2 )dE2 dE1
l E
E+ −l
≥ π −1 X(s + l, s+ , η) − O≺ (N −ε/2 ).
(32)
For any D > 0 and all N ≥ N0 (ε, D), (31) implies that π −1 X(s − l, s+ , η) ≥ 2/3 whenever
#(E∗ + s, E∗ + δ) ≥ 1, except possibly on an event of probability N −D . Similarly (32) implies
π −1 X(s + l, s+ , η) ≤ 1/3 whenever #(E∗ + s, E∗ + δ) = 0, except possibly on an event of probability
N −D . The result then follows from the definition and boundedness of K.
4. The interpolating sequence
In this section, we construct the interpolating sequence T (0) , . . . , T (L) described in the introduction. We consider only the case of a right edge; this is without loss of generality, as the edge
can have arbitrary sign and we may take the reflection T 7→ −T . For each pair T ≡ T (l) and
Ť ≡ T (l+1) , the following definition captures the relevant property that will be needed in the
subsequent computation.
Definition 4.1. Let T, Ť ∈ RM ×M be two diagonal matrices satisfying Assumption 2.4. Let E∗
be a right edge of the law ρ defined by T , and let Ě∗ be a right edge of ρ̌ defined by Ť . (T, E∗ ) and
(Ť , Ě∗ ) are swappable if, for a constant φ > 0, both of the following hold.
• Letting tα and ťα be the diagonal entries of T and Ť ,
M
X
|tα − ťα |< φ.
α=1
• The m-values m, m̌∗ of E∗ , Ě∗ satisfy
|m∗ − m̌∗ |< φ/N.
We will say that (T, E∗ ) and (Ť , Ě∗ ) are φ-swappable if we wish to emphasize the role of φ. All
subsequent constants may implicitly depend on φ.
We first establish some basic deterministic properties of a swappable pair, including closeness of
the edges E∗ , Ě∗ as claimed in (10) in the introduction.
Lemma 4.2. Suppose T, Ť are diagonal matrices satisfying Assumption 2.4, E∗ , Ě∗ are regular
right edges, and (T, E∗ ) and (Ť , Ě∗ ) are swappable. Let m∗ , γ and m̌∗ , γ̌ be the m-values and scales
of E∗ , Ě∗ . Denote sα = (1 + tα m∗ )−1 and šα = (1 + ťα m̌∗ )−1 . For integers i, j ≥ 0, define
Ai,j =
M
1 X i i j j
tα sα ťα šα .
N
α=1
20
TRACY-WIDOM FOR COVARIANCE MATRICES
Then there exists a constant C > 0 such that all of the following hold:
(a) For all i, j satisfying i + j ≤ 4,
|Ai,j − Ai+j,0 |≤ C/N.
(b) (Closeness of edge location)
|E∗ − Ě∗ |≤ C/N,
and
M
1 X
(tα − ťα )sa šα ≤ C/N 2 .
(E∗ − Ě∗ ) −
N
(33)
α=1
(c) (Closeness of scale)
|γ − γ̌|≤ C/N.
Proof. By Proposition 3.5, |tα |, |sα |, |m∗ |, γ ≤ C, and also |ťα |, |šα |, |m̌∗ |, γ̌ ≤ C. From the definitions of sα and šα , we verify
tα sα − ťα šα = (tα − ťα )sα šα + (m̌∗ − m∗ )tα sα ťα šα .
Then swappability implies
|Ai,j
M
1 X i i j−1 j−1
− Ai+1,j−1 |≤
|tα sα ťα šα ||ťα šα − tα sα |≤ C/N.
N
α=1
Iteratively applying this yields (a).
For (b), recall 0 = z00 (m∗ ) = m−2
∗ − A2,0 . Then
M
1
1
1 X
(tα sα − ťα šα )
E∗ − Ě∗ = −
+
+
m∗ m̌∗ N
α=1
M
1
1 X
= (m∗ − m̌∗ )
− A1,1 +
(tα − ťα )sα šα .
m∗ m̌∗
N
α=1
Part (b) follows from the definition of swappability, together with |A1,1 −m−2
∗ |= |A1,1 −A2,0 |≤ C/N
−2
−1
−1
and |m∗ − m∗ m̌∗ |≤ C/N .
−2 − γ̌ −2 |≤ |m−3 −
For (c), we have γ −2 = z000 (m∗ )/2 = −m−3
∗ + A3,0 . Then (c) follows from |γ
∗
−3
m̌∗ |+|A3,0 − A0,3 |≤ C/N .
In the rest of this section, we prove the following proposition:
Proposition 4.3. Suppose T is diagonal and satisfies Assumption 2.4, and E∗ is a τ -regular
0 , τ 0 , φ > 0, a sequence
right edge with scale γ = 1. Then there exist τ -dependent constants Cd,T
of diagonal matrices T (0) , T (1) , . . . , T (L) in RM ×M for L ≤ 2M , and a sequence of right edges
(0)
(1)
(L)
E∗ , E∗ , . . . , E∗ of the corresponding laws ρ(l) defined by T (l) , such that:
(0)
1. T (0) = T and E∗ = E∗ .
2. T (L) has at most two distinct diagonal entries 0 and t, for some t ∈ R.
0 .
3. Each T (l) satisfies Assumption 2.4 with constant Cd,T
(l)
4. Each E∗ is τ 0 -regular.
(l)
(l+1)
5. (T (l) , E∗ ) and (T (l+1) , E∗
) are φ-swappable for each l = 0, . . . , L − 1.
(l)
6. (Scaling) Each E∗ has associated scale γ (l) = 1.
TRACY-WIDOM FOR COVARIANCE MATRICES
21
(0)
(L)
We first ignore the scaling condition, property 6, and construct T (0) , . . . , T (L) and E∗ , . . . , E∗
satisfying properties 1–5. We will use a Lindeberg swapping construction, where each T (l+1) differs
from T (l) in only one diagonal entry. It is useful to write z00 and z000 as
1 X
1
1
z00 (m) = 2 −
2
m
N
(m + t−1
α )
α:tα 6=0
2
2 X
1
z000 (m) = − 3 +
,
3
m
N
(m + t−1
α )
α:t 6=0
α
and to think about swapping entries of T as swapping or removing poles of z00 and z000 . In particular,
for each fixed m ∈ R, we can easily deduce from the above whether a given swap increases or
decreases the values of z00 and z000 at m.
Upon defining a swap T → Ť , the identification of the new right edge Ě∗ for Ť uses the following
continuity lemma.
Lemma 4.4. Suppose T is a diagonal matrix satisfying Assumption 2.4, and E∗ is a τ -regular right
edge with m-value m∗ . Let Ť be a matrix that replaces a single diagonal entry tα of T by a value
ťα , such that |ťα |≤ kT k and either ťα = 0 or |m∗ + ť−1
α |> τ . Let z0 , ž0 denote the function (15)
defined by T, Ť . Then there exist τ -dependent constants N0 , φ > 0 such that whenever N ≥ N0 :
• Ť has a right edge Ě∗ with m-value m̌∗ satisfying |m∗ − m̌∗ |< φ/N .
• The interval between m∗ and m̌∗ does not contain any pole of z0 or ž0 .
• sign(m∗ − m̌∗ ) = sign(ž00 (m∗ )).
(We define sign(x) = 1 if x > 0, −1 if x < 0, and 0 if x = 0.)
Proof. By Proposition 3.5, |m∗ |> ν for a constant ν. Take δ < min(τ /2, ν/2). Then the given
conditions for ťα imply that (m∗ − δ, m∗ + δ) does not contain any pole of z0 or ž0 , and
|z00 (m) − ž00 (m)|< C/N
for some C > 0 and all m ∈ (m∗ − δ, m∗ + δ). For sufficiently small δ, Proposition 3.6 also ensures
z000 (m) > c for all m ∈ (m∗ − δ, m∗ + δ). Suppose ž00 (m∗ ) < 0 = z00 (m∗ ). Taking N0 = C/(δc), if
N > N0 , then the above implies ž00 (m∗ + C/(cN )) > z00 (m∗ + C/(cN )) − C/N > 0. Hence ž0 has
a local minimum in (m∗ , m∗ + C/(cN )). Similarly, if ž00 (m∗ ) > 0, then ž0 has a local minimum in
(m∗ − C/(cN ), m∗ ), and if ž00 (m∗ ) = 0, then ž0 has a local minimum at m∗ . The result follows from
Proposition 2.2 upon taking Ě∗ = ž0 (m̌∗ ) and φ = C/c.
The basic idea for proving Proposition 4.3 is to take a Lindeberg sequence T (0) , . . . , T (L) and
apply the above lemma for each swap. We cannot do this naively for any Lindeberg sequence,
(l)
(l+1)
because in general if E∗ is τl -regular, then the above lemma only guarantees that E∗
is τl+1 regular for τl+1 = τl − C/N and a τl -dependent constant C > 0. Thus edge regularity, as well as
the edge itself, may vanish after O(N ) swaps.
To circumvent this, we consider a specific construction of the Lindeberg sequence, apply Lemma
4.4 inductively along this sequence to identify an edge Ě∗ for each successive Ť , and use a separate
argument to show that Ě∗ must be τ 0 -regular for a fixed constant τ 0 > 0. Hence we may continue
to apply Lemma 4.4 along the whole sequence.
We consider separately the cases m∗ < 0 and m∗ > 0.
Lemma 4.5. Suppose (the right edge) E∗ has m-value m∗ < 0. Then for some τ -dependent
constant N0 , whenever N ≥ N0 , Proposition 4.3 holds without the scaling condition, property 6.
Proof. We construct a Lindeberg sequence that first reflects about m∗ each pole of z0 to the right
of m∗ , and then replaces each pole by the one closest to m∗ .
22
TRACY-WIDOM FOR COVARIANCE MATRICES
Suppose, first, that there are K1 non-zero diagonal entries tα of T (positive or negative) where
(0) , T (1) , . . . , T (K1 ) where T (0) = T , and each
−t−1
α > m∗ . Consider a sequence of matrices T
T (k+1) replaces one such diagonal entry tα of T (k) by the value ťα such that −ť−1
α < m∗ and
−1 |. For each such swap T → Ť , we verify |ť |≤ |t |≤ kT k, ž 0 (m ) = z 0 (m ) = 0,
|m∗ + ť−1
|=
|m
+t
∗
α
α
∗
∗
α
α
0
0
and ž000 (m∗ ) > z000 (m∗ ) > 0. Thus we may take m̌∗ = m∗ in Lemma 4.4, and the new edge
Ě∗ = ž0 (m∗ ) remains τ -regular for the same constant τ .
All diagonal entries of T (K1 ) are now nonnegative. Let t = kT (K1 ) k be the maximal such entry. By
(K )
the above construction, −t−1 < m∗ < 0. Since E∗ 1 is τ -regular, (16) implies t > c for a constant
c > 0. Let K2 be the number of positive diagonal entries of T (K1 ) strictly less than t, and consider
a sequence T (K1 +1) , . . . , T (K1 +K2 ) where each T (k+1) replaces one such diagonal entry in T (k) by
t. Applying Lemma 4.4 inductively to each such swap T → Ť , we verify ž00 (m∗ ) < z0 (m∗ ) = 0, so
−1
−1
m∗ < m̌∗ < 0. Then |m̌∗ |< |m∗ | and minα |m̌∗ + ť−1
α |> minα |m∗ + tα |. Also m̌∗ + ťα > 0 for
all ťα 6= 0, so ž000 (m̌∗ ) > −2/m̌3∗ > 2t3 . This verifies Ě∗ = ž0 (m̌∗ ) is τ 0 -regular for a fixed constant
τ 0 > 0. (We may take any τ 0 < min(τ, t3/2 ).)
The total number of swaps L = K1 + K2 is at most 2M , and all diagonal entries of T (L) belong
to {0, t}. This concludes the proof, with property 5 verified by Lemma 4.4.
Lemma 4.6. Lemma 4.5 holds also when (the right edge) E∗ has m-value m∗ > 0.
Proof. Proposition 2.2 implies m∗ is a local minimum of z0 . The interval (0, m∗ ) must contain a
pole of z0 —otherwise, by the boundary condition of z0 at 0, there would exist a local maximum m
of z0 in (0, m∗ ) satisfying z0 (m) > z0 (m∗ ), which would contradict the edge ordering in Proposition
2.2(c). Let −t−1 be the pole in (0, m∗ ) closest to m∗ . Note that t < 0 and |t|> |m∗ |−1 > τ . We
construct a Lindeberg sequence that first replaces a small but constant fraction of entries of T by
t, then replaces all non-zero tα > t by 0, and finally replaces all tα < t by 0.
First, fix a small constant c0 > 0, let K1 = bc0 M c, and consider a sequence of matrices
(0)
T , T (1) , . . . , T (K1 ) where T (0) = T and each T (k+1) replaces a different (arbitrary) diagonal entry
of T (k) by t. For c0 sufficiently small, we claim that we may apply Lemma 4.4 to identify an edge
(k)
(k)
E∗ for each k = 1, . . . , K1 , such that each E∗ is τ /2-regular. Indeed, let k ∈ {0, . . . , K1 − 1} and
(j)
suppose inductively that we have identified this edge E∗ for j = 0, . . . , k. Let mj , γj be the m-value
(j)
and scale for E∗ . Then Lemmas 4.4 and 4.2 ensure |mj − mj−1 |< C/N and |γj − γj−1 |< C/N
for a τ /2-dependent constant C > 0. This yields |mk − m0 |< c0 d−1 C and |γk − γ0 |< c0 d−1 C. As
(0)
(k)
the original edge E∗ is τ -regular, for sufficiently small c0 this implies E∗ is in fact 3τ /4-regular.
(k+1)
Applying Lemma 4.4, we may identify an edge E∗
for T (k+1) with m-value mk+1 satisfying
(k+1)
|mk+1 − mk |< φ/N . For N ≥ N0 , this implies E∗
is τ /2-regular, completing the induction.
T (K1 ) now has at least K1 + 1 ≥ c0 M diagonal entries equal to t. By the condition in Lemma
4.4 that the swap m∗ → m̌∗ does not cross any pole of z0 or ž0 , we have that −t−1 is still the
(K )
(K )
pole in (0, m∗ 1 ) closest to m∗ 1 . Let K2 be the number of non-zero diagonal entries tα of T (K1 )
(positive or negative) such that tα > t. Consider a sequence T (K1 +1) , . . . , T (K1 +K2 ) where each
T (k+1) replaces one such entry in T (k) by 0. Applying Lemma 4.4 inductively to each such swap
−1
T → Ť , we verify ž00 (m∗ ) > z00 (m∗ ) = 0, so −t−1 < m̌∗ < m∗ . Then minα:−ť−1
−1 |m̌∗ + ťα |>
α >−t
−1
−1 > c and
minα:−t−1
−1 |m∗ + tα |> τ /2. The conditions m̌∗ > |t|
α >−t
0 = ž00 (m̌∗ ) ≤
1
c0 M
1
−
2
m̌∗
N (m̌∗ + t−1 )2
TRACY-WIDOM FOR COVARIANCE MATRICES
23
ensure that m̌∗ + t−1 > ν for a constant ν > 0, and hence minα |m̌∗ + ť−1
α |> min(ν, τ /2). To bound
ž000 (m̌∗ ), let us introduce the function
f (m) = −
M
t2α m3
2 X
N
(1 + tα m)3
α=1
and define analogously fˇ(m) for Ť . We verify f 0 (m) < 0 for all m, so f (m̌∗ ) > f (m∗ ). Furthermore,
if the swap T → Ť replaces tα by 0, then 1 + tα m̌∗ > 0. (This is obvious for positive tα ; for negative
0
ˇ
tα , it follows from m̌∗ < −t−1
α .) Then f (m̌∗ ) > f (m̌∗ ) > f (m∗ ). Applying the condition 0 = z0 (m∗ ),
4
00
we verify f (m∗ ) = m∗ z0 (m∗ ). Then
ž000 (m̌∗ ) >
m4∗ 00
z (m∗ ) > z000 (m∗ ) > 0.
m̌4∗ 0
This shows that Ě∗ = ž0 (m̌∗ ) is τ 0 -regular for a fixed constant τ 0 > 0. (We may take τ 0 =
min(ν, τ /2) as above.)
Finally, T (K1 +K2 ) now has at least c0 M diagonal entries equal to t, and all non-zero diagonal
entries tα satisfy tα < t < 0. Let K3 be the number of such entries and consider a sequence
T (K1 +K2 +1) , . . . , T (K1 +K2 +K3 ) where each T (k+1) replaces one such entry of T (k) by 0. Applying
Lemma 4.4 inductively to each such swap T → Ť , we verify ž00 (m∗ ) > z00 (m∗ ) = 0, so −t−1 < m̌∗ <
m∗ . As in the K2 swaps above, this implies minα |m̌∗ + ť−1
α |> c for a constant c > 0. The condition
ťα < t implies 1 + ťα m̌∗ < 0 for all ťα , so we have
2 c0 M
t2 m̌3∗
fˇ(m̌∗ ) > −
>c
N N (1 + tm̌∗ )3
for a constant c. Applying again fˇ(m̌∗ ) = m̌4∗ ž000 (m̌∗ ), this yields ž000 (m̌∗ ) > c > 0, so Ě∗ is τ 0 -regular
for a constant τ 0 > 0.
The total number of swaps L = K1 + K2 + K3 is at most 2M (provided c0 < 1 in the first step).
All diagonal entries of T (L) belong to {0, t}, so this concludes the proof.
Proof of Proposition 4.3. Suppose first N ≥ N0 so that both Lemmas 4.5 and 4.6 hold for N .
(0)
(L)
Then there exist T (0) , . . . , T (L) and E∗ , . . . , E∗ satisfying conditions 1–5. By Lemma 4.2, the
associated scales γ0 , . . . , γL satisfy |γl+1 − γl |≤ C/N for a φ, τ 0 -dependent constant C > 0 and each
l = 0, . . . , L − 1.
We verify from the definitions of E∗ , m∗ , γ that under the rescaling T 7→ cT for any c > 0, we
have
E∗ 7→ cE∗ ,
m∗ 7→ c−1 m∗ ,
γ 7→ c−3/2 γ.
2/3
(l)
2/3
(l)
Consider then the matrices T̃ (l) = γl T (l) and edges Ẽ∗ = γl E∗ . We check properties 1–6 for
(l)
T̃ (l) and Ẽ∗ : Properties 1, 2, and 6 are obvious. Since T (0) , . . . , T (L) are all τ 0 -regular, Proposition
3.5 implies c < γl < C for constants C, c > 0 and every l, so properties 3 and 4 hold with adjusted
constants. Property 5 also holds with an adjusted constant φ, since
X
X
X
|γ 2/3 tα − γ̌ 2/3 ťα | ≤ γ 2/3
|tα − ťα | + |γ 2/3 − γ̌ 2/3 |
|ťα |< φ0
α
α
α
and
|γ −2/3 m∗ − γ̌ −2/3 m̌∗ |≤ γ −2/3 |m∗ − m̌∗ |+|γ −2/3 − γ̌ −2/3 | |m̌∗ |< φ0 /N
for a φ, τ 0 -dependent constant φ0 > 0. This concludes the proof in the case N ≥ N0 .
The case N < N0 is trivial, as we may take K = 1, T (1) = t Id where t is chosen to satisfy property
6, and sufficiently large constants such that properties 3–5 hold trivially whenever N < N0 .
24
TRACY-WIDOM FOR COVARIANCE MATRICES
5. Resolvent comparison and proof of Theorem 2.10
The crux of the proof of Theorem 2.10 is the following lemma.
Lemma 5.1 (Resolvent comparison). Fix ε > 0 a sufficiently small constant, and let s1 , s2 , η ∈ R be
such that |s1 |, |s2 |< N −2/3+ε and η ∈ [N −2/3−ε , N −2/3 ]. Let T, Ť ∈ RM ×M be two diagonal matrices
and E∗ , Ě∗ two corresponding regular right edges, such that (T, E∗ ) and (Ť , Ě∗ ) are swappable and
their scales satisfy γ = γ̌ = 1.
Let mN , m̌N be the Stieltjes transforms as in (20) corresponding to T, Ť , and define
Z Ě∗ +s2
Z E∗ +s2
=m̌N (y + iη)dy.
=mN (y + iη)dy,
X̌ = N
X=N
Ě∗ +s1
E∗ +s1
Let K : R → R be any function such that K and its first four derivatives are uniformly bounded by
a constant. Then, uniformly over s1 , s2 , η satisfying the above,
E[K(X) − K(X̌)] ≺ N −4/3+16ε .
(34)
Let us first prove Theorem 2.10 assuming this lemma.
Proof of Theorem 2.10. By symmetry with respect to T 7→ −T , it suffices to consider part (a), the
case of a right edge. By rescaling T 7→ γ 2/3 T , it suffices to consider the case where γ = 1. By
Corollary 3.12, we may assume that T is diagonal and X has Gaussian entries.
(0)
(L)
Let T (0) , . . . , T (L) and E∗ , . . . , E∗ satisfy Proposition 4.3. Define X(k) (s1 , s2 , η) as in (27) for
(k)
each (T (k) , E∗ ). For a sufficiently small constant ε > 0, let η, s+ , l and K : [0, ∞) → [0, 1] be as
in Lemma 3.13, where K has bounded derivatives of all orders.
Fix x ∈ R and let s = xN −2/3 . Applying Lemma 3.13, we have (for all large N )
P[λmax (Σ̂) ≤ E∗ + s] ≤ E[K(π −1 X(0) (s + l, s+ , η)] + N −1 .
Setting ε0 = 9ε and applying Lemma 5.1, we have
0
E[K(π −1 X(k) (s + l, s+ , η)] ≤ E[K(π −1 X(k+1) (s + l, s+ , η)] + N −4/3+17ε
for each k = 0, . . . , L − 1. Finally, defining Σ̂(L) = X 0 T (L) X and λmax (Σ̂(L) ) as its largest eigenvalue
(L)
(L)
in (E∗ − δ 0 , E∗ + δ 0 ) for some δ 0 > 0, applying Lemma 3.13 again yields
(L)
E[K(π −1 X(L) (s + l, s+ , η)] ≤ P[λmax (Σ̂(L) ) ≤ E∗
+ s + 2l] + N −1 .
Recalling L ≤ 2M and combining the above bounds,
(L)
P[N 2/3 (λmax (Σ̂) − E∗ ) ≤ x] ≤ P[N 2/3 (λmax (Σ̂(L) ) − E∗ ) ≤ x + 2N −ε ] + o(1).
The matrix T (L) has all diagonal entries 0 or t, so Σ̂(L) = tX̃ 0 X̃ for X̃ ∈ RM̃ ×N having
Normal(0, 1/N ) entries. The corresponding law ρ(L) has a single support interval and a unique
(L)
(L)
right edge, so E∗ must be this edge. Regularity of E∗ and (16) imply |t| 1 and M̃ /N 1. If
(L)
E∗ > 0, then t > 0. Applying [Joh01, Theorem 1.1] for the largest eigenvalue of a real Wishart
matrix, we have
(L)
P[N 2/3 (λmax (Σ̂(L) ) − E∗ ) ≤ x + 2N −ε ] = F1 (x) + o(1)
(35)
(L)
where F1 is the distribution function of µT W . If E∗ < 0, then t < 0, and edge regularity implies
M̃ /N is bounded away from 1. Then we also have (35) by considering −Σ̂(L) and applying [FS10,
Theorem I.1.1] for the smallest eigenvalue of a real Wishart matrix. (If M̃ < N , we apply this
result to the companion matrix X̃ X̃ 0 .) Combining the above, we obtain
P[N 2/3 (λmax (Σ̂) − E∗ ) ≤ x] ≤ F1 (x) + o(1).
The reverse bound is analogous, concluding the proof.
TRACY-WIDOM FOR COVARIANCE MATRICES
25
In the remainder of this section, we prove Lemma 5.1.
5.1. Individual resolvent bounds. For diagonal T and for z = y +iη as appearing in Lemma 5.1,
we record here simple resolvent bounds that follow from the local law. Similar bounds were used in
[EYY12b, LS16]. We also introduce the shorthand notation that will be used in the computation.
Let E∗ be a regular right edge. Fix a small constant ε > 0, and fix s1 , s2 , η such that |s1 |, |s2 |≤
N −2/3+ε and η ∈ [N −2/3−ε , N −2/3 ]. Changing variables, we write
Z s2
=mN (y + E∗ + iη)dy.
X ≡ X(s1 , s2 , η) = N
s1
For y ∈ [s1 , s2 ], we write as shorthand
z ≡ z(y) = y + E∗ + iη,
G(α) ≡ G(α) (z(y)),
(α)
mN ≡
G ≡ G(z(y)),
1 X (α)
Gii ,
N
X(α)
mN ≡ mN (z(y)),
Z s2
(α)
=mN (y + E∗ + iη)dy.
≡N
s1
i∈IN
The above quantities depend implicitly on y.
We use the simplified summation notation
X
X
≡
,
i,j
X
i,j∈IN
X
≡
α,β
α,β∈IM
where summations over lower-case Roman indices are implicitly over IN and summations over
Greek indices are implicitly over IM . We use also the simplified integral notation
Z
Z s2
Z
Z s2
G̃AB ≡
G(z(ỹ))AB dỹ,
m̃N ≡
mN (z(ỹ))dỹ,
s1
s1
etc., so that integrals are implicitly over [s1 , s2 ], and we denote by F̃ the function F evaluated at
F (z(ỹ)) for ỹ the variable of integration. In this notation, X and X(α) are simply
X Z
X Z (α)
(α)
X=
= G̃ii ,
X =
= G̃ii .
i
i
We introduce the fundamental small parameter
Ψ = N −1/3+3ε .
(36)
We will eventually bound all quantities in the computation by powers of Ψ. In fact, as shown in
Lemmas 5.2 and 5.3 below, non-integrated resolvent entries are controlled by powers of the smaller
quantity
N −1/3+ε .
However, integrated quantities will require the additional slack of N 2ε . We will pass to using Ψ for
all bounds after this distinction is no longer needed.
We have the following corollaries of Proposition 3.7 and Theorem 3.10:
Lemma 5.2. Under the assumptions of Lemma 5.1, for all y ∈ [s1 , s2 ], i 6= j ∈ IN , and α 6= β ∈
IM ,
1
Gαα
tα
≺ 1,
≺ 1,
≺ 1,
Gii ≺ 1,
Gii
tα
Gαα
Gij ≺ N −1/3+ε ,
Giα
≺ N −1/3+ε ,
tα
Gαβ
≺ N −1/3+ε ,
tα tβ
|mN − m∗ |≺ N −1/3+ε .
When T is not invertible, these quantities are defined by continuity and the form (22) for G.
26
TRACY-WIDOM FOR COVARIANCE MATRICES
Proof. For diagonal T , the matrix Π0 in (24) is diagonal,
t1
tM
Π0 = diag m0 , . . . , m0 , −
,...,−
.
1 + t1 m0
1 + tM m0
√
Proposition 3.7 implies =m0 (z(y)) ≤ C κ + η ≤ CN −1/3+ε/2 , while η ≥ N −2/3−ε by assumption.
Then, defining tA = 1 for A ∈ IN (and tA = tα for A = α ∈ IM ), Theorem 3.10 applied with
standard basis vectors for v, w yields (tA tB )−1 (G − Π0 )AB ≺ N −1/3+ε for all A, B ∈ I. Proposition
3.7 also implies |m0 (z)| 1 and |1+tα m0 (z)| 1, from which all of the entrywise bounds on G follow.
√
The bound on mN follows from |m0 −m∗ |≤ C κ + η ≤ CN −1/3+ε/2 and |mN −m0 |≺ N −1/3+ε .
Lemma 5.3. Under the assumptions of Lemma 5.1, for all i ∈ IN and α ∈ IM ,
X (α)
X
−1/3+ε
G(α)
.
Gik Xαk ≺ N −1/3+ε ,
pq Xαp Xαq − m∗ ≺ N
p,q
k
Proof. Applying Lemmas 3.9(b) and 5.2,
X (α)
Gik Xαk = −Giα /Gαα ≺ N −1/3+ε .
k
Similarly, applying Lemma 3.9(a) and Theorem 3.10,
X
Gαα − Π0,αα
1
t2α
1
1
1
G(α)
X
X
−
m
=
−
−
−
m
=
−
=
≺ N −1/3+ε .
αp αq
∗
∗
pq
2
G
t
Π
G
t
G
Π
αα
α
0,αα
αα
αα
0,αα
α
p,q
(These types of bounds are in fact used in the proof of Theorem 3.10 and may also be derived
directly from concentration inequalities and the independence of Xαk and G(α) .)
Remark 5.4. All probabilistic bounds used in the proof of Lemma 5.11, such as the above, are
derived from Theorem 3.10. Thus they in fact hold in the union bound form of Corollary 3.11. We
continue to use the notation ≺ for convenience, with the implicit understanding that we may take
a union bound over all y ∈ [s1 , s2 ], and in particular integrate such bounds over y.
We record one trivial bound for an integral that will be repeatedly used, and which explains the
appearance of Ψ.
Lemma 5.5. Suppose the assumptions of Lemma 5.1 hold, F (z(y)) ≺ N a(−1/3+ε) for some a ≥ 2,
and we may take a union bound of this statement over y ∈ [s1 , s2 ] (in the sense of Lemma 3.3).
Then, with Ψ = N −1/3+3ε ,
Z
N
F̃ ≺ Ψa−1 .
Proof. This follows, for a ≥ 2, from
N (s2 − s1 )N a(−1/3+ε) ≤ 2N 1/3+ε N a(−1/3+ε) ≤ 2Ψa−1 .
The next lemma will allow us to “remove the superscript” in the computation.
Lemma 5.6. Under the assumptions of Lemma 5.1, for any y ∈ [s1 , s2 ], i, j ∈ IN (possibly equal),
and α ∈ IM ,
(α)
Gij − Gij ≺ N 2(−1/3+ε) ,
(α)
mN − mN ≺ N 2(−1/3+ε) ,
X − X(α) ≺ Ψ.
TRACY-WIDOM FOR COVARIANCE MATRICES
27
Proof. Applying the last resolvent identity from Lemma 3.9,
Giα Gjα
Gjα tα
(α)
Gij − Gij =
= Giα
,
Gαα
tα Gαα
so the first statement follows from Lemma 5.2. Taking i = j and averaging over IN yields the
second statement. The third statement follows from
Z
(α)
(α)
X − X = = N m̃N − m̃N
and Lemma 5.5.
5.2. Resolvent bounds for a swappable pair. We now record bounds for a swappable pair
(T, E∗ ) and (Ť , Ě∗ ), where E∗ , Ě∗ are both regular. We denote by m̌N , Ǧ, X̌ the analogues of
mN , G, X for Ť . For ε, s1 , s2 , η and y ∈ [s1 , s2 ] as in Section 5.1, we write as shorthand
ž ≡ ž(y) = y + Ě∗ + iη,
Ǧ ≡ Ǧ(ž(y)),
m̌N ≡ m̌N (ž(y)),
where these quantities depend implicitly on y. The results of the preceding section hold equally for
Ǧ, m̌N , and X̌.
The desired bound (34) arises from the following identity: Suppose first that T and Ť are
invertible. Applying A−1 − B −1 = A−1 (B − A)B −1 ,
−ž + z
0
G − Ǧ = G
Ǧ.
0
−Ť −1 + T −1
Hence, as z − ž = E∗ − Ě∗ ,
Gij − Ǧij =
X
Gik Ǧjk (E∗ − Ě∗ ) −
X Giα Ǧjα
α
k
tα
ťα
(tα − ťα ).
(37)
This holds also by continuity when T is not invertible, where Giα /tα and Ǧjα /ťα are well-defined
by (22).
The following lemma will allow us to “remove the check” in the computation.
Lemma 5.7. Suppose the assumptions of Lemma 5.1 hold. Let Ψ = N −1/3+3ε . Then for any
y ∈ [s1 , s2 ], i, j ∈ IN (possibly equal), and α ∈ IM ,
Gij − Ǧij ≺ N 2(−1/3+ε) ,
mN − m̌N ≺ N 2(−1/3+ε) ,
X − X̌ ≺ Ψ.
Proof. Applying Lemma 5.2 for both G and Ǧ, and also the definition of swappability and Lemma
4.2, we have from (37)
X
Gij − Ǧij ≺ |E∗ − Ě∗ |·N · N 2(−1/3+ε) +
|tα − ťα |N 2(−1/3+ε) ≺ N 2(−1/3+ε) .
α
(The contribution from k = i or k = j in the first sum of (37) is of lower order.) Taking i = j
and averaging over IN yields the second statement, and integrating over y ∈ [s1 , s2 ] and applying
Lemma 5.5 yields the third.
In many cases, we may strengthen the above lemma by an additional factor of Ψ if we take an
expectation. (This type of idea is an important part of the argument in [EYY12b, LS16]. For
example, setting a = 0 and Y ≡ Y (α) ≡ 1 in Lemma 5.9 below yields the second-order bound
E[X − X̌] ≺ Ψ2 .) To take expectations of remainder terms, we will invoke Lemma 3.4 combined
with the following basic bound:
28
TRACY-WIDOM FOR COVARIANCE MATRICES
Lemma 5.8. Under the assumptions of Lemma 5.1, let P ≡ P (z(y)) be any polynomial in the
entries of X and G with bounded degree, bounded (possibly random) coefficients, and at most N C
terms for a constant C > 0. Then for a constant C 0 > 0 and all y ∈ [s1 , s2 ],
0
E[|P |] ≤ N C .
Proof. By the triangle inequality and Holder’s inequality, it suffices to consider a bounded power
of a single entry of G or X. Then the result follows from Assumption 2.5, Lemma 3.8, and the
form (22) for G.
Lemma 5.9. Under the assumptions of Lemma 5.1, let Y be any quantity such that Y ≺ Ψa
for some constant a ≥ 0. Suppose that for each α ∈ IM , there exists a quantity Y (α) such that
Y − Y (α) ≺ Ψa+1 , and Y (α) is independent of row α of X. Suppose furthermore that E[|Y |4 ] ≤ N C
for a constant C > 0.
Then, for all i, j ∈ IN (possibly equal) and y ∈ [s1 , s2 ],
E[(Gij − Ǧij )Y ] ≺ N 2(−1/3+ε) Ψa+1 ≺ Ψa+3 ,
E[(mN − m̌N )Y ] ≺ N 2(−1/3+ε) Ψa+1 ≺ Ψa+3 .
E[(X − X̌)Y ] ≺ Ψa+2 .
Proof. Applying (33), the trivial bound N −1 ≺ Ψ3 , and Lemma 5.2 to (37),
X Giα Ǧjα
X
(tα − ťα )Y
(Gij − Ǧij )Y =
Gik Ǧjk (E∗ − Ě∗ )Y −
tα ťα
α
k
!
X
1 X
Giα Ǧjα
=
(tα − ťα ) sα šα
Gik Ǧjk −
Y + O≺ (Ψa+5 ).
N
t
ť
α
α
α
k
By swappability and Lemma 5.2, the explicit term on the right is of size O≺ (N 2(−1/3+ε) Ψa ). (The
contributions from k = i and k = j in the summation are of lower order.) Applying the assumption
(α)
Y − Y (α) ≺ Ψa+1 as well as Lemma 5.6, we may replace Y with Y (α) , Gik with Gik , and Ǧjk with
(α)
Ǧjk above while introducing an O≺ (N 2(−1/3+ε) Ψa+1 ) error. Hence,
!
X
1 X (α) (α) Giα Ǧjα
Gik Ǧjk −
Y (α) + O≺ (N 2(−1/3+ε) Ψa+1 ). (38)
(Gij − Ǧij )Y =
(tα − ťα ) sα šα
N
t
ť
α
α
α
k
Applying the resolvent identities from Lemma 3.9,
Giα
Gαα X (α)
=
Gik Xαk = −
P
tα
tα
1 + tα
k
Recalling sα = (1 + tα m∗
around x = m∗ ,
)−1 ,
1
X
(α)
p,q Gpq Xαp Xαq
(α)
Gik Xαk .
k
and applying Lemma 5.3 and a Taylor expansion of (1 + tα x)−1
X (α)
Giα
= −sα
Gik Xαk + O≺ (N 2(−1/3+ε) ),
tα
k
where the explicit term on the right is of size O≺ (N −1/3+ε ) ≺ Ψ. A similar expansion holds for
Ǧjα /ťα . Substituting into (38),
X
X
X
1
(α) (α)
(α)
(α)
(Gij − Ǧij )Y =
(tα − ťα )sα šα
Gik Ǧjk −
Gik Xαk Ǧjl Xαl Y (α)
N
α
k
+ O≺ (N
2(−1/3+ε)
a+1
Ψ
).
k,l
TRACY-WIDOM FOR COVARIANCE MATRICES
29
Denoting by Eα the partial expectation over only row α of X (i.e. conditional on Xβj for all β 6= α),
we have
X (α) (α) X (α)
1
(α)
Eα
Gik Ǧjk −
Gik Xαk Ǧjl Xαl = 0,
N
k
k,l
(N 2(−1/3+ε) Ψa+1 )
while the remainder term remains O≺
by Lemmas 3.4 and 5.8, the assumption on
E[|Y |4 ], and Cauchy-Schwarz. Then the first statement follows. The second statement follows from
applying this with i = j and averaging over i ∈ IN . The third statement follows from integrating
over y ∈ [s1 , s2 ] and noting N 1/3+ε N 2(−1/3+ε) = Ψ as in Lemma 5.5. (If Y depends on the spectral
parameter z(y), this integration is performed by fixing this parameter for Y , evaluating mN and
m̌N at a different parameter ỹ, and integrating over ỹ. The preceding arguments do not require Y
and mN , m̌N to depend on the same spectral parameter.)
P
Finally, recall the notation sα = (1 + tα m∗ )−1 and Ai = N −1 α tiα siα . (This corresponds
to Ai,0 in Lemma 4.2.) We derive a deterministic consequence of swappability and the scaling
condition γ = γ̌ = 1. In the proof of [LS16] for a continuous interpolation T (l) , denoting ṫα and
ṁ∗ the derivatives with respect to l, the differential analogue of the following lemma is the pair of
identities
X
X
ṫα tα s3α = N ṁ∗ ,
ṫα t2α s4α = N ṁ∗ (A4 − m−4
∗ ).
α
α
These may be derived by implicitly differentiating 0 = z00 (m∗ ) and 1 = z000 (m∗ ) with respect to l.
We show that discrete versions of these identities continue to hold, with O(N −1 ) error:
Lemma 5.10. Suppose T, Ť satisfy Assumption 2.4, E∗ , Ě∗ are associated regular right edges with
scales γ = γ̌ = 1, and P
(T, E∗ ) and (Ť , Ě∗ ) are swappable. Define sα = (1 + tα m∗ )−1 , šα =
4 4
−1
−1
(1 + ťα m̌∗ ) , A4 = N
α tα sα , and
Pα = sα šα (tα sα + ťα šα ),
Qα = sa šα (t2α s2α + tα sα ťα šα + ť2α š2α ).
Then for some constant C > 0, both of the following hold:
2N (m∗ − m̌∗ ) −
M
X
(tα − ťα )Pα ≤ C/N
(39)
α=1
3N (m∗ − m̌∗ )(A4 − m−4
∗ )−
M
X
(tα − ťα )Qα ≤ C/N.
(40)
α=1
Proof. For (39), we have from the identity 0 = z00 (m∗ ) applied to T and Ť
1 X 2 2
−2
t s − ť2α š2α .
m−2
∗ − m̌∗ =
N α α α
(41)
The left side may be written as
−2
−2 −2
−3
−2
m−2
),
∗ − m̌∗ = (m̌∗ − m∗ )(m̌∗ + m∗ )m∗ m̌∗ = 2(m̌∗ − m∗ )m∗ + O(N
(42)
where the second equality applies |m∗ |, |m̌∗ | 1 and |m̌∗ − m∗ |≤ C/N . The right side may be
written as
1 X 2 2
1 X
tα sα − ť2α š2α =
(tα − ťα )tα s2α + (s2α − š2α )tα ťα + (tα − ťα )ťα š2α .
N α
N α
Including the identities (1 + tα m∗ )sα = 1 and (1 + ťα m̌∗ )šα = 1,
1 X 2 2
1 X
tα sα − ť2α š2α =
(tα − ťα )(tα s2α (1 + ťα m̌∗ )šα + ťα š2α (1 + tα m∗ )sα ) + (s2α − š2α )tα ťα
N α
N α
30
TRACY-WIDOM FOR COVARIANCE MATRICES
1 X
(tα − ťα )sα šα (tα sα + ťα šα + tα sα ťα m̌∗ + ťα šα tα m∗ ) + (s2α − š2α )tα ťα
N α
1 X
≡
(tα − ťα )sα šα (tα sα + ťα šα ) + Rα ,
(43)
N α
=
where we define Rα as the remainder term. Noting that
s2α − š2α = (sα − šα )(sα + šα ) = (ťα m̌∗ − tα m∗ )sα šα (sα + šα ),
we have
Rα = tα ťα sα šα (tα sα m̌∗ + tα šα m∗ − ťα sα m̌∗ − ťα šα m∗ + ťα sα m̌∗ + ťα šα m̌∗ − tα sα m∗ − tα šα m∗ )
= tα sα ťα šα (m̌∗ − m∗ )(tα sα + ťα šα ).
Then, applying Lemma 4.2(a),
1 X
Rα = (m̌∗ − m∗ )(A2,1 + A1,2 ) = 2(m̌∗ − m∗ )A3,0 + O(N −2 ).
N α
By the scaling γ = 1, we have A3,0 = 1 + m−3
∗ . Combining this with (41), (42), and (43) and
multiplying by N yields (39).
The identity (40) follows similarly: The condition γ = γ̌ implies
1 X 3 3
−3
m−3
t s − ť3α š3α .
∗ − m̌∗ =
N α α α
The left side is
−3
−4
−2
(m̌∗ − m∗ )(m2∗ + m∗ m̌∗ + m̌2∗ )m−3
),
∗ m̌∗ = 3(m̌∗ − m∗ )m∗ + O(N
while the right side is
1 X
1 X 3 3
tα sα − ť3α š3α =
(tα − ťα )t2α s3α + (s2α − š2α )t2α sα ťα + (tα − ťα )tα sα ťα š2α
N α
N α
+ (sα − šα )tα ť2α š2α + (tα − ťα )ť2α š3α
1 X
=
(tα − ťα ) t2α s3α (1 + ťα m̌∗ )šα + tα sα ťα š2α (1 + tα m∗ )sα
N α
2 3
+ ťα šα (1 + tα m∗ )sα + (sα − šα )((sα + šα )t2α sα ťα + tα ť2α š2α )
1 X
=
(tα − ťα )sα šα (t2α s2α + tα sα ťα šα + ť2α š2α )
N α
=
+ tα sα ťα šα (m̌∗ − m∗ )(t2α s2α + tα sα ťα šα + ť2α š2α )
!
1 X
(tα − ťα )Qα + 3(m̌∗ − m∗ )A4 + O(N −2 ).
N α
Combining the above and multiplying by N yields (40).
5.3. Proof of Lemma 5.1. We use the notation of Sections 5.1 and 5.2.
Define the following quantities, depending implicitly on a fixed index i ∈ IN and y ∈ [s1 , s2 ]:
1 X 2
X3,120 = K 0 (X)(mN − m∗ )
Gik
N
k
1 X
0
X3,3 = K (X) 2
Gik Gkl Gil
N
k,l
TRACY-WIDOM FOR COVARIANCE MATRICES
Z
1 X
Gik Gil = G̃jk G̃jl
N2
j,k,l
Z
X
1
00
2
= K (X) 2
Gik = G̃2jl
N
X3,2e2 = K 00 (X)
X3,20e20
j,k,l
1 X 2
Gik
N
k
1 X
= K 0 (X)(mN − m∗ ) 2
Gik Gkl Gil
N
X4,220 = K 0 (X)(mN − m∗ )2
X4,13
k,l
X4,4
1 X
Gij Gjk Gkl Gil
= K 0 (X) 3
N
X4,40
1 X 2 2
= K 0 (X) 3
Gik Gjl
N
j,k,l
j,k,l
X4,12e2
X4,120e20
X4,3e2
X4,30e2
X4,3e20
X4,2f
12
0
X4,20 f
12
X4,2e3
X4,2e30
X4,20e3
X4,2e2e2
X4,20e2e2
Z
1 X
= K (X)(mN − m∗ ) 2
Gik Gil = G̃jk G̃jl
N
j,k,l
Z
1 X 2
00
= K (X)(mN − m∗ ) 2
Gik = G̃2jl
N
j,k,l
Z
X
1
00
= K (X) 3
Gip Giq Gpr = G̃jq G̃jr
N
j,p,q,r
Z
1 X 2
= K 00 (X) 3
Gir Gpq = G̃jp G̃jq
N
j,p,q,r
Z
X
1
00
= K (X) 3
Giq Gir Gqr = G̃2jp
N
j,p,q,r
Z
1 X
00
Gik Gil = (m̃N − m∗ )G̃jk G̃jl
= K (X) 2
N
j,k,l
Z
1 X 2
= K 00 (X) 2
Gil = (m̃N − m∗ )G̃2jk
N
j,k,l
Z
X
1
00
= K (X) 3
Gip Giq = G̃jp G̃jr G̃qr
N
j,p,q,r
Z
1 X
00
= K (X) 3
Gip Giq = G̃2jr G̃pq
N
j,p,q,r
Z
1 X 2
00
= K (X) 3
Gip = G̃jq G̃jr G̃qr
N
j,p,q,r
Z
Z
1 X
000
= K (X) 3
Gip Giq = G̃jp G̃jr
= G̃kq G̃kr
N
j,k,p,q,r
Z
Z
1 X
2
000
= G̃kq G̃kr
= K (X) 3
Gip = G̃jq G̃jr
N
00
j,k,p,q,r
31
32
TRACY-WIDOM FOR COVARIANCE MATRICES
Z
Z
1 X
2
=
G̃
G̃
=
G̃
G
G
jp jq
ip iq
kr
N3
j,k,p,q,r
Z
Z
1 X
= K 000 (X) 3
G2ip = G̃2jq
= G̃2kr
N
X4,2e2e20 = K 000 (X)
X4,20e20e20
j,k,p,q,r
Define the aggregate quantities
X3 = X3,12 + X3,3 + X3,2e2
X4 = 3X4,220 + 6X4,13 + 12X4,4 + 3X4,40 + 4X4,12e2 + 8X4,3e2 + 4X4,30e2
+ 2X4,2f
3 + 2X4,2e
30 + 4X4,2e
2e
2,
12 + 4X4,2e
X−
4
= X4,2f
30 + 2X4,2e
3 − X4,12e
2 − X4,30 e
2 − 2X4,3e
2.
12 + X4,2e
(Not all of the above terms appear in these aggregate quantities; we define them because they
appear in intermediate steps of the proof.) The notation signifies that each term X3,∗ is of size at
most O≺ (Ψ3 ), and each term X4,∗ is of size at most O≺ (Ψ4 ), as may be verified from Lemmas 5.2
and 5.5. A ∼ in the subscript denotes an integrated quantity, and a 0 in the subscript denotes a
squared resolvent entry.
Lemma 5.1 is a consequence of the following two technical results.
Lemma 5.11 (Decoupling). Under the assumptions of Lemma 5.1, denote Xλ = λX + (1 − λ)X̌
for λ ∈ [0, 1]. For fixed i ∈ IN and y ∈ [s1 , s2 ], define X3 , X4 , and X−
4 as above. For fixed α ∈ IM ,
let sα = (1 + tα m∗ ) and šα = (1 + ťα m̌∗ ), and define
Pα = sα šα (tα sα + ťα šα ), Qα = sα šα (t2α s2α + tα sα ťα šα + ť2α š2α ), Rα = sα šα (tα sα − ťα šα )2 .
Then
Z
0
1
Giα Ǧiα
E K (Xλ )
tα ťα
0
Z
dλ = sα šα
0
1
"
#
1 X
E K (Xλ )
Gik Ǧik dλ
N
0
k
5
− Pα E[X3 ] + 13 Qα E[X4 ] + 31 Rα E[X−
4 ] + O≺ (Ψ ).
Lemma 5.12 (Optical theorems). Under the assumptions
of Lemma 5.1, for fixed i ∈ IN and
P
y ∈ [s1 , s2 ], define X3 and X4 as above. Let A4 = N −1 α t4α s4α . Then
5
2=E[X3 ] = (A4 − m−4
∗ )=E[X4 ] + O≺ (Ψ ).
Lemma 5.11 generalizes [LS16, Lemma 6.2] to a swappable pair. We will present a self-contained
proof in Section 5.4, although many computational ideas are the same. We introduce the interpolation Xλ = λX + (1 − λ)X̌ as a device to bound K(X) − K(X̌). (This is different from a continuous
interpolation between the entries of T and Ť .) We make several additional remarks for the reader
familiar with [LS16]:
1. The proof in [LS16] requires this lemma in “differential form”, where T = Ť . In this case, we
have G = Ǧ, Xλ = X for every λ ∈ [0, 1], sα = šα , and tα = ťα . Then the integral over λ is
irrelevant, and Lemma 5.11 reduces to the full version of [LS16, Lemma 6.2].
2. There is an additional term X−
4 that does not appear in [LS16], and which is not canceled by the
optical theorems of Lemma 5.12. (When T = Ť , we have Rα = 0 so this term is not present.)
The cancellation will instead occur by symmetry of its definition, upon integrating over y.
3. The main additional complexity in our proof comes from needing to separately track the terms
that arise from resolvent expansions of G and Ǧ, and from X and X̌ after Taylor expanding
K 0 (Xλ ). An important simplification is that we may use Lemmas 5.7 and 5.9 to convert O≺ (Ψ3 )
and O≺ (Ψ4 ) terms to involve only G and not Ǧ—hence X3 , X4 , X−
4 are defined only by T and
not Ť . Swappability of (T, E∗ ) and (Ť , Ě∗ ) is used for this simplification.
TRACY-WIDOM FOR COVARIANCE MATRICES
33
The other technical ingredient, Lemma 5.12, is identical to the full version of [LS16, Lemma B.1],
as the terms X3 and X4 depend only on the single matrix T . We briefly discuss the breakdown of
its proof in Section 5.5.
In [LS16], for expositional clarity, these lemmas were stated and proven only in the special case
K 0 ≡ 1. Full proofs were presented for an analogous deformed Wigner model in [LS15]. Although
more cumbersome, we will demonstrate the full proof of Lemma 5.11 for a general function K in
Section 5.4, as much of the additional complexity due to two resolvents G and Ǧ arises from the
interpolation Xλ and the Taylor expansion of K 0 .
We conclude this section by establishing Lemma 5.1 using the above two results:
Proof of Lemma 5.1. We write
1
Z
K(X) − K(X̌) =
0
d
K(Xλ )dλ =
dλ
Z
1
K 0 (Xλ )(X − X̌)dλ.
(44)
0
Recalling
X=
X
Z
=
G̃ii
i
and applying (37),
X − X̌ =
X
Z
X
=
i
˜ (E − Ě ) −
G̃ik Ǧ
∗
∗
ik
˜
X G̃iα Ǧ
iα
α
k
tα ťα
!
(tα − ťα ) .
˜ denote G and Ǧ evaluated at the variable of integration ỹ.) Further applying (33),
(G̃ and Ǧ
Lemma 5.2, and the trivial bound N −2/3+ε ≺ Ψ2 ,
!
˜
X Z X
G̃iα Ǧ
1 X
iα
˜
X − X̌ =
G̃ik Ǧik −
+ O≺ (Ψ4 ).
=
(tα − ťα ) sα šα
N
t
ť
α
α
α
i
k
Applying this to (44), taking the expectation, exchanging orders of summation and integration,
and noting that K 0 (Xλ ) is real,
E[K(X) − K(X̌)]
1
Z Z
XX
=
(tα − ťα )=
i
0
α
˜
1 X
˜ − G̃iα Ǧiα
E K (Xλ ) sα šα
G̃ik Ǧ
ik
N
tα ťα
"
0
!#
dλ dỹ + O≺ (Ψ4 ),
k
where the expectation of the remainder term is still O≺ (Ψ4 ) by Lemmas 3.4 and 5.8. Denoting
−
by X̃3 (i), X̃4 (i), and X̃−
4 (i) the quantities X3 , X4 , and X4 defined by ỹ and the outer index of
summation i, Lemma 5.11 implies
Z
XX
1/3+ε 5
E[K(X)−K(X̌)] =
(tα −ťα )= (Pα E[X̃3 (i)]− 13 Qα E[X̃4 (i)]− 31 Rα E[X̌−
Ψ ),
4 (i)])dỹ+O≺ (N
i
α
where the error is N 1/3+ε Ψ5 because
[−N −2/3+ε , N −2/3+ε ]. We note
P
α |tα
Z
=
− ťα |≤ C and the range of integration is contained in
X̃−
4 (i)dỹ = 0
by symmetry of the terms defining X̃−
4 , so this term vanishes. Then, applying Lemma 5.12,
Z
XX
A4 − m−4
Qα
∗
E[K(X) − K(X̌)] =
(tα − ťα ) Pα
−
= E[X̃4 (i)]dỹ + O≺ (N 1/3+ε Ψ5 ).
2
3
α
i
34
TRACY-WIDOM FOR COVARIANCE MATRICES
Finally, applying Lemma 5.10, we have
X
A4 − m−4
Qα
∗
(tα − ťα ) Pα
−
≤ C/N.
2
3
α
Hence E[K(X) − K(X̌)] ≺ N 1/3+ε Ψ5 = N −4/3+16ε .
5.4. Proof of Lemma 5.11. In this section, we prove the decoupling lemma. We will implicitly
use the resolvent bounds of Lemma 5.2 throughout the proof.
Step 1: Consider first a fixed value λ ∈ [0, 1]. Let Eα denote the partial expectation over row α
of X (i.e. conditional on all Xβj for β 6= α). In anticipation of computing Eα for the quantity on
the left, we expand
Giα Ǧiα
K 0 (Xλ )
tα ťα
as a polynomial of entries of row α of X, with coefficients independent of all entries in this row.
Applying the resolvent identities,
X (α)
Giα
1
Gαα X (α)
Gik Xαk .
Gik Xαk = −
=
P
(α)
tα
tα
1
+
t
G
X
X
pq
α
αp
αq
k
k
p,q
Applying Lemma 5.3 and a Taylor expansion of the function (1 + tα x)−1 around x = m∗ ,
!
X (α)
X
X (α)
Giα
= −sα
Gik Xαk + tα s2α
G(α)
Gik Xαk
pq Xαp Xαq − m∗
tα
p,q
k
k
!2
X
X (α)
− t2α s3α
G(α)
Gik Xαk + O≺ (Ψ4 )
pq Xαp Xαq − m∗
p,q
k
4
≡ U1 + U2 + U3 + O≺ (Ψ ),
(45)
where we defined the three explicit terms of sizes O≺ (Ψ), O≺ (Ψ2 ), O≺ (Ψ3 ) as U1 , U2 , U3 . Similarly
Ǧiα
= Ǔ1 + Ǔ2 + Ǔ3 + O≺ (Ψ4 ),
ťα
(46)
where Ǔi are defined analogously with šα , ťα , m̌∗ , Ǧ in place of sα , tα , m∗ , G.
(α)
(α)
For K 0 (Xλ ), define Xλ = λX(α) + (1 − λ)X̌(α) and note from Lemma 5.6 that Xλ − Xλ ≺ Ψ.
(α)
Taylor expanding K 0 (x) around x = Xλ ,
(α)
0
K (Xλ ) = K
0
(α)
(Xλ )
+K
00
(α)
(Xλ )(Xλ
−
(α)
Xλ )
K 000 (Xλ )
(α)
+
(Xλ − Xλ )2 + O≺ (Ψ3 ).
2
(47)
Applying the definition of X, X(α) and the resolvent identities,
Z X
Z
Z X 2
X (α)
G̃jα
(α)
(α)
(α)
X−X ==
(G̃jj − G̃jj ) = =
= = G̃αα
G̃jp Xαp G̃jq Xαq .
G̃αα
j
j
j,p,q
Further applying the resolvent identity for G̃αα , a Taylor expansion as above, and Lemma 5.5,
Z X
(α)
(α)
(α)
X − X = −tα sα =
G̃jp Xαp G̃jq Xαq
j,p,q
+ t2α s2α =
Z X
G̃(α)
rs Xαr Xαs − m∗
r,s
X
(α)
(α)
G̃jp Xαp G̃jq Xαq + O≺ (Ψ3 )
j,p,q
3
≡ V1 + V2 + O≺ (Ψ ),
(48)
TRACY-WIDOM FOR COVARIANCE MATRICES
35
where V1 ≺ Ψ and V2 ≺ Ψ2 . Analogously we may write
X̌ − X̌(α) = V̌1 + V̌2 + O≺ (Ψ3 ),
(49)
where V̌1 , V̌2 are defined with šα , ťα , m̌∗ , Ǧ in place of sα , tα , m∗ , G. Substituting (48) and (49) into
(47), and combining with (45) and (46), we obtain
K 0 (Xλ )
Giα Ǧiα
= W2 + W3 + W4 + O≺ (Ψ5 )
tα ťα
(50)
where the O≺ (Ψ2 ) term is
(α)
W2 = K 0 (Xλ )U1 Ǔ1 ,
the O≺ (Ψ3 ) term is
(α)
(α)
W3 = K 0 (Xλ )(U2 Ǔ1 + U1 Ǔ2 ) + K 00 (Xλ )(λV1 + (1 − λ)V̌1 )U1 Ǔ1 ,
and the O≺ (Ψ4 ) term is
(α)
(α)
W4 = K 0 (Xλ )(U3 Ǔ1 + U2 Ǔ2 + U1 Ǔ3 ) + K 00 (Xλ )(λV1 + (1 − λ)V̌1 )(U2 Ǔ1 + U1 Ǔ2 )
"
#
000 (X(α) )
K
(α)
λ
+ K 00 (Xλ )(λV2 + (1 − λ)V̌2 ) +
(λV1 + (1 − λ)V̌1 )2 U1 Ǔ1 .
2
Step 2: We compute Eα of W2 , W3 , W4 above. Note that X(α) , X̌(α) , G(α) , Ǧ(α) are independent
of row α of X. Then for W2 , we have
X (α) (α)
(α)
Eα [W2 ] = sα šα K 0 (Xλ )
Gik Ǧil Eα [Xαk Xαl ]
k,l
X (α) (α)
(α) 1
Gik Ǧik ,
= sα šα K 0 (Xλ )
N
k
where we have used E[Xαk Xαl ] = 1/N if k = l and 0 otherwise.
For W3 , let us introduce
1 X (α) (α)
(α)
(α)
(α)
Y3,120 = K 0 (Xλ )(mN − m∗ )
Gik Ǧik ,
N
k
1 X (α) (α)
(α)
(α)
(α)
0
Gik Ǧik ,
Z3,120 = K (Xλ )(m̌N − m̌∗ )
N
k
X (α) (α) (α)
1
(α)
(α)
Y3,3 = K 0 (Xλ ) 2
Gik Gkl Ǧil
N
k,l
(α)
Z3,3
(α)
Y3,20e20
(α)
Z3,20e20
(α)
Y3,2e2
(α)
Z3,2e2
X (α) (α) (α)
(α) 1
Gik Ǧkl Ǧil
= K 0 (Xλ ) 2
N
k,l
Z
X
1
(α)
(α) (α)
(α)
00
= K (Xλ ) 2
Gik Ǧik = (G̃jl )2
N
j,k,l
Z
X
1
(α)
(α) (α)
˜ (α) )2
00
= K (Xλ ) 2
Gik Ǧik = (Ǧ
jl
N
j,k,l
X (α) (α) Z (α) (α)
(α) 1
00
= K (Xλ ) 2
Gik Ǧil = G̃jk G̃jl
N
j,k,l
X (α) (α) Z ˜ (α) ˜ (α)
(α) 1
00
= K (Xλ ) 2
Gik Ǧil = Ǧjk Ǧjl ,
N
j,k,l
(51)
36
TRACY-WIDOM FOR COVARIANCE MATRICES
which are versions of X3,∗ that don’t depend on row α of X and with various instances of mN , m∗ , G, X
replaced by m̌N , m̌∗ , Ǧ, Xλ . Consider the first term of W3 and write
(α)
Eα [K 0 (Xλ )U2 Ǔ1 ]
!
(α)
X
= Eα −tα s2α šα K 0 (Xλ )
G(α)
pq Xαp Xαq − m∗
p,q
(α)
= −tα s2α šα K 0 (Xλ )
X
X
(α)
(α)
Gik Xαk Ǧil Xαl
k,l
G(α)
pq Eα [Xαp Xαq Xαk Xαl ] −
k,l,p,q
1
(α) (α)
m∗ 1{p = q}Eα [Xαk Xαl ] Gik Ǧil .
N
The summand corresponding to (k, l, p, q) is 0 unless each distinct index appears at least twice in
(k, l, p, q). Furthermore, the case where all four indices are equal is negligible:
X (α)
1
(α) (α)
2
4
Gkk Eα [Xαk ] − m∗ Eα [Xαk ] Gik Ǧik ≺ N · N −2 · Ψ2 ≺ Ψ5 .
N
k
(The k = i case of the sum may be bounded separately as O≺ (N −2 ).) Thus up to O≺ (Ψ5 ), we need
only consider summands where each distinct index appears exactly twice. Considering the one case
where k = l and the two cases where k = p and k = q,
(k)
XX
(α) (α)
(α)
(α) 1
(α)
0
2
0
Gpp − m∗ Gik Ǧik
Eα [K (Xλ )U2 Ǔ1 ] = −tα sα šα K (Xλ )
N2
p
k
(k)
2 X X (α) (α) (α)
+ 2
Gik Ǧil Gkl
+ O≺ (Ψ5 ).
N
k
l
Re-including p = k and l = k into the double summations introduces an additional O≺ (Ψ5 ) error;
hence we obtain for the first term of W3
(α)
(α)
(α)
Eα [K 0 (Xλ )U2 Ǔ1 ] = −tα s2α šα (Y3,120 + 2Y3,3 ) + O≺ (Ψ5 ).
(52)
Similar arguments apply for the remaining three terms of W3 . For the terms involving an integral,
we may apply Lemma 5.5 and also move Xαk outside of the integral and imaginary part because
X is real and does not depend on the variable of integration ỹ. We obtain
(α)
(α)
(α)
Eα [K 0 (Xλ )U1 Ǔ2 ] = −ťα š2α sα (Z3,120 + 2Z3,3 ) + O≺ (Ψ5 ),
(α)
(α)
(53)
(α)
Eα [λK 00 (Xλ )V1 U1 Ǔ1 ] = −λtα s2α šα (Y3,20e20 + 2Y3,2e2 ) + O≺ (Ψ5 ),
(α)
(α)
(54)
(α)
Eα [(1 − λ)K 00 (Xλ )V̌1 U1 Ǔ1 ] = −(1 − λ)ťα š2α sα (Z3,20e20 + 2Z3,2e2 ) + O≺ (Ψ5 ),
and Eα [W3 ] is the sum of (52–55).
For W4 , consider the first term and write
(α)
(α)
Eα [K 0 (Xλ )U3 Ǔ1 ] = Eα t2α s3α šα K 0 (Xλ )
(α)
!2
X
G(α)
pq Xαp Xαq − m∗
p,q
= t2α s3α šα K 0 (Xλ )
(55)
X
X
(α)
(α)
Gik Xαk Ǧil Xαl
k,l
(α)
G(α)
pq Grs Eα [Xαp Xαq Xαr Xαs Xαk Xαl ]
p,q,r,s,k,l
1
− m∗ 1{p = q}G(α)
rs Eα [Xαr Xαs Xαk Xαl ]
N
1
− m∗ 1{r = s}G(α)
pq Eα [Xαp Xαq Xαk Xαl ]
N
TRACY-WIDOM FOR COVARIANCE MATRICES
37
!
1 2
(α) (α)
+ 2 m∗ 1{p = q}1{r = s}E[Xαk Xαl ] Gik Ǧil .
N
A summand corresponding to (k, l, p, q, r, s) is 0 unless each distinct index in (k, l, p, q, r, s) appears at least twice. Furthermore, as in the computations for W3 above, all summands for which
(k, l, p, q, r, s) do not form three distinct pairs may be omitted and reincluded after taking Eα ,
introducing an O≺ (Ψ5 ) error. Considering all pairings of these indices,
(α)
(α)
(α)
Eα [K 0 (Xλ )U3 Ǔ1 ] = t2α s3α šα K 0 (Xλ ) (mN − m∗ )2
1 X (α) (α)
Gik Ǧik
N
k
1 X (α) (α) (α) (α)
1 X (α) (α) (α)
(α)
Gik Gkl Ǧil + 8 3
Gik Gjk Gjl Ǧil
+ 4(mN − m∗ ) 2
N
N
k,l
j,k,l
!
1 X (α) (α) (α) 2
Gik Ǧik (Gjl ) + O≺ (Ψ5 ).
+2 3
N
j,k,l
At this point, let us apply Lemmas 5.6 and 5.7 to remove each superscript (α) above and to
convert each Ǧ to G, introducing an O≺ (Ψ5 ) error. (We could not do this naively for W2 and W3 ,
because the errors would be O≺ (Ψ3 ) and O≺ (Ψ4 ) respectively.) We may also remove the superscript
(α)
(α) and convert Xλ to X in K 0 (Xλ ), via the second-derivative bounds
(α)
(α)
K 0 (Xλ ) − K 0 (Xλ ) ≤ kK 00 k∞ |Xλ − Xλ |≺ Ψ.
K 0 (Xλ ) − K 0 (X) ≤ kK 00 k∞ |Xλ − X|≺ Ψ.
We thus obtain
(α)
Eα [K 0 (Xλ )U3 Ǔ1 ] = t2α s3α šα (X4,220 + 4X4,13 + 8X4,4 + 2X4,40 ) + O≺ (Ψ5 ).
Applying a similar computation to each term of W4 , we obtain
(α)
Eα [K 0 (Xλ )(U3 Ǔ1 + U2 Ǔ2 + U1 Ǔ3 )]
(56)
= sα šα (t2α s2α + tα sα ťα šα + ť2α š2α )(X4,220 + 4X4,13 + 8X4,4 + 2X4,40 ) + O≺ (Ψ5 ),
(57)
(α)
Eα [K 00 (Xλ )(λV1
+ (1 − λ)V̌1 )(U2 Ǔ1 + U1 Ǔ2 )]
= sα šα (λtα sα + (1 − λ)ťα šα )(tα sα + ťα šα )×
(X4,120e20 + 2X4,12e2 + 2X4,3e20 + 2X4,30e2 + 8X4,3e2 ) + O≺ (Ψ5 ),
(58)
(α)
Eα [K 00 (Xλ )(λV2 + (1 − λ)V̌2 )U1 Ǔ1 ]
5
0 + 2X
= sα šα (λt2α s2α + (1 − λ)ť2α š2α )(X4,20 f
3 + 2X4,2e
30 + 8X4,2e
3 ) + O≺ (Ψ ),
4,2f
12 + 2X4,20 e
12
"
#
(α)
K 000 (Xλ )
Eα
(λV1 + (1 − λ)V̌1 )2 U1 Ǔ1
2
(59)
sα šα
(λtα sα + (1 − λ)ťα šα )2 (X4,20e20e20 + 2X4,20e2e2 + 4X4,2e2e20 + 8X4,2e2e2 ) + O≺ (Ψ5 ),
(60)
2
and Eα [W4 ] is the sum of (57–60).
The O≺ (Ψ5 ) remainder in (50) is given by the difference of the left side with W2 , W3 , W4 . As
this is an integral over a polynomial of entries of G(α) and X, its partial expectation is still O≺ (Ψ5 )
by Lemmas 3.4 and 5.8.
Summarizing the results of Steps 1 and 2, we collect (50), (51), (52–55), and (57–60):
Giα Ǧiα
0
Eα K (Xλ )
tα ťα
=
38
TRACY-WIDOM FOR COVARIANCE MATRICES
(α)
= sα šα K 0 (Xλ )
−
−
+
1 X (α) (α)
Gik Ǧik
N
k
(α)
(α)
(α)
(α)
tα s2α šα (Y3,120 + 2Y3,3 ) − ťα š2α sα (Z3,120 + 2Z3,3 )
(α)
(α)
(α)
(α)
λtα s2α šα (Y3,20e20 + 2Y3,2e2 ) − (1 − λ)ťα š2α sα (Z3,20e20 + 2Z3,2e2 )
sα šα (t2α s2α + tα sα ťα šα + ť2α š2α )(X4,220 + 4X4,13 + 8X4,4 + 2X4,40 )
+ sα šα (λtα sα + (1 − λ)ťα šα )(tα sα + ťα šα )(X4,120e20 + 2X4,12e2 + 2X4,3e20 + 2X4,30e2 + 8X4,3e2 )
0 + 2X
+ sα šα (λt2α s2α + (1 − λ)ť2α š2α )(X4,20 f
3 + 2X4,2e
30 + 8X4,2e
3)
4,2f
12 + 2X4,20 e
12
+
sα šα
(λtα sα + (1 − λ)ťα šα )2 (X4,20e20e20 + 2X4,20e2e2 + 4X4,2e2e20 + 8X4,2e2e2 ) + O≺ (Ψ5 ).
2
(61)
Step 3: In (61), we consider the first term on the right (of size O≺ (Ψ2 )) and remove the
superscripts (α), keeping track of the O≺ (Ψ3 ) and O≺ (Ψ4 ) terms that arise.
Applying the resolvent identities and a Taylor expansion for Gαα , we write
(α)
Giα Gkα
Gαα
X (α)
(α)
= Gik − Gαα
Gir Xαr Gks Xαs
Gik = Gik −
r,s
!
= Gik + tα sα
X
(α)
(α)
Gir Xαr Gks Xαs
−
t2α s2α
X
r,s
G(α)
pq Xαp Xαq
− m∗
p,q
X
(α)
(α)
Gir Xαr Gks Xαs
r,s
4
+ O≺ (Ψ )
≡ Gik + R2k + R3k + O≺ (Ψ4 ),
(62)
where we defined the two remainder terms of sizes O≺ (Ψ2 ), O≺ (Ψ3 ) as R2k , R3k . Similarly we write
(α)
Ǧik = Ǧik + Ř2k + Ř3k + O≺ (Ψ4 ).
(63)
(α)
For K 0 (Xλ ), we apply the Taylor expansion (47) and recall V1 , V̌1 , V2 , V̌2 from (48,49) to obtain
(α)
(α)
(α)
(α)
K 0 (Xλ ) = K 0 (Xλ ) − K 00 (Xλ )(Xλ − Xλ ) −
K 000 (Xλ )
(α)
(Xλ − Xλ )2 + O≺ (Ψ3 )
2
(α)
(α)
= K 0 (Xλ ) − K 00 (Xλ )(λV1 + (1 − λ)V̌1 ) − K 00 (Xλ )(λV2 + (1 − λ)V̌2 )
(α)
K 000 (Xλ )
−
(λV1 + (1 − λ)V̌1 )2 + O≺ (Ψ3 ).
2
(64)
Taking the product of (62), (63), and (64), applying the identity
xyz = (x − δx )(y − δy )(z − δz ) + xyδz + xδy z + δx yz − xδy δz − δx yδz − δx δy z + δx δy δz
(α)
(with x = Gik , x − δx = Gik , and δx = R2k + R3k , etc.), and averaging over k ∈ IN , we obtain
X (α) (α)
(α) 1
K 0 (Xλ )
Gik Ǧik ≡ S2 + S3,1 + S3,2 + S4,1 + S4,2 + S4,3 + S4,4 + S4,5 + O≺ (Ψ5 ),
(65)
N
k
where the O≺ (Ψ2 ) term is
S2 = K 0 (Xλ )
1 X
Gik Ǧik ,
N
k
TRACY-WIDOM FOR COVARIANCE MATRICES
39
the O≺ (Ψ3 ) terms are
X
1 X (α)
(α) 1
(α)
Gik Ř2k + K 0 (Xλ )
R2k Ǧik ,
N
N
k
k
X
1
(α)
(α) (α)
= −K 00 (Xλ )(λV1 + (1 − λ)V̌1 )
Gik Ǧik ,
N
(α)
S3,1 = K 0 (Xλ )
S3,2
k
and the O≺ (Ψ4 ) terms are
X
1 X (α)
(α) 1
(α)
Gik Ř3k + K 0 (Xλ )
R3k Ǧik ,
N
N
k
k
X
1
(α) (α)
(α)
Gik Ǧik ,
= −K 00 (Xλ )(λV2 + (1 − λ)V̌2 )
N
(α)
S4,1 = K 0 (Xλ )
S4,2
k
S4,3 = −
S4,4
(α)
K 000 (Xλ )
2
(λV1 + (1 − λ)V̌1 )2
1 X (α) (α)
Gik Ǧik ,
N
k
X
(α) 1
R2k Ř2k ,
= −K 0 (Xλ )
N
k
(α)
S4,5 = K 00 (Xλ )(λV1 + (1 − λ)V̌1 )
1 X (α)
1 X
(α)
(α)
Gik Ř2k + K 00 (Xλ )(λV1 + (1 − λ)V̌1 )
R2k Ǧik .
N
N
k
k
Recalling the definition of R2k and applying Eα to the O≺ (Ψ3 ) terms, we obtain
(α)
(α)
Eα [S3,1 ] = tα sα Y3,3 + ťα šα Z3,3 ,
(α)
(α)
Eα [S3,2 ] = λtα sα Y3,20e20 + (1 − λ)ťα šα Z3,20e20 .
Similarly, we apply Eα to each of the O≺ (Ψ4 ) terms, considering all pairings of the four summation
indices as in Step 2. Then applying Lemmas 5.6 and 5.7 to remove superscripts and convert Ǧ to
G, we obtain
Eα [S4,1 ] = −(t2α s2α + ť2α š2α )(X4,13 + 2X4,4 ) + O≺ (Ψ5 ),
5
0 + 2X
Eα [S4,2 ] = −(λt2α s2α + (1 − λ)ť2α š2α )(X4,20 f
4,20 e
3 ) + O≺ (Ψ ),
12
1
Eα [S4,3 ] = − (λtα sα + (1 − λ)ťα šα )2 (X4,20e20e20 + 2X4,20e2e2 ) + O≺ (Ψ5 ),
2
Eα [S4,4 ] = −tα sα ťα šα (X4,40 + 2X4,4 ) + O≺ (Ψ5 ),
Eα [S4,5 ] = −(λtα sα + (1 − λ)ťα šα )(tα sα + ťα šα )(X4,3e20 + 2X4,3e2 ) + O≺ (Ψ5 ).
40
TRACY-WIDOM FOR COVARIANCE MATRICES
Then applying Eα to (65), noting that the remainder is again O≺ (Ψ5 ) by Lemmas 3.4 and 5.8, and
substituting into (61),
"
#
Giα Ǧiα
1 X
(α)
(α)
0
0
Eα K (Xλ )
Gik Ǧik − tα s2α šα (Y3,120 + Y3,3 )
= sα šα Eα K (Xλ )
tα ťα
N
−
(α)
ťα š2α sα (Z3,120
+
k
(α)
Z3,3 )
(α)
(α)
− 2λtα s2α šα Y3,2e2 − 2(1 − λ)ťα š2α sα Z3,2e2
+ sα šα (t2α s2α + ť2α š2α )(X4,220 + 3X4,13 + 6X4,4 + 2X4,40 )
+ sα šα (tα sα ťα šα )(X4,220 + 4X4,13 + 6X4,4 + X4,40 ) + sα šα (λtα sα
+ (1 − λ)ťα šα )(tα sα + ťα šα )(X4,120e20 + 2X4,12e2 + X4,3e20 + 2X4,30e2 + 6X4,3e2 )
+ sα šα (λt2α s2α + (1 − λ)ť2α š2α )(2X4,2f
30 + 8X4,2e
3)
12 + 2X4,2e
sα šα
(λtα sα + (1 − λ)ťα šα )2 (4X4,2e2e20 + 8X4,2e2e2 ) + O≺ (Ψ5 ).
+
2
(66)
Step 4: In (66), we remove the superscript (α) from Y3,∗ and Z3,∗ , keeping track of the O≺ (Ψ4 )
(α)
(α)
errors that arise. For each quantity Y3,∗ or Z3,∗ , let Y3,∗ or Z3,∗ be the analogous quantity with
(α)
(α)
each instance of mN , G(α) , G̃(α) , Xλ
(α)
For Y3,120 , recall from (62)
replaced by mN , G, G̃, Xλ .
(α)
Gik = Gik + R2k + O≺ (Ψ3 ),
and from (64)
(α)
(α)
K 0 (Xλ ) = K 0 (Xλ ) − K 00 (Xλ )(λV1 + (1 − λ)V̌1 ) + O≺ (Ψ2 ).
(α)
For mN − m∗ , we apply the resolvent identities and write
(α)
mN − m∗ = mN − m∗ −
1 X G2jα
N
Gαα
j
= mN − m∗ − Gαα
1 X (α)
(α)
Gjk Xαk Gjl Xαl
N
j,k,l
= mN − m∗ + tα sα
1 X (α)
(α)
Gjk Xαk Gjl Xαl + O≺ (Ψ3 )
N
j,k,l
≡ mN − m∗ + Q + O≺ (Ψ3 ),
where Q is the O≺ (Ψ2 ) term. Multiplying the above and averaging over k yields
1 X (α)
(α)
(α)
(α)
Gik Ř2k
Y3,120 = Y3,120 + K 0 (Xλ )(mN − m∗ )
N
k
1 X (α)
1 X (α) (α)
(α)
(α)
(α)
+ K 0 (Xλ )(mN − m∗ )
Ǧik R2k + K 0 (Xλ )Q
Gik Ǧik
N
N
k
k
1 X (α) (α)
(α)
(α)
− K 00 (Xλ )(λV1 + (1 − λ)V̌1 )(mN − m∗ )
Gik Ǧik + O≺ (Ψ5 ),
N
k
where each term except Y3,120 on the right is of size O≺ (Ψ4 ). Taking Eα and applying Lemmas 5.6
and 5.7 to remove superscripts and checks,
(α)
Y3,120 = Eα [Y3,120 ] + (tα sα + ťα šα )X4,13 + tα sα X4,40 + (λtα sα + (1 − λ)ťα šα )X4,120e20 + O≺ (Ψ5 ). (67)
TRACY-WIDOM FOR COVARIANCE MATRICES
41
Similar arguments yield
(α)
Z3,120 = Eα [Z3,120 ] + (tα sα + ťα šα )X4,13 + ťα šα X4,40 + (λtα sα + (1 − λ)ťα šα )X4,120e20 + O≺ (Ψ5 ),
(α)
Y3,3 = Eα [Y3,3 ] + (2tα sα + ťα šα )X4,4 + (λtα sα + (1 − λ)ťα šα )X4,3e20 + O≺ (Ψ5 ),
(α)
Z3,3 = Eα [Z3,3 ] + (tα sα + 2ťα šα )X4,4 + (λtα sα + (1 − λ)ťα šα )X4,3e20 + O≺ (Ψ5 ),
(α)
Y3,2e2 = Eα [Y3,2e2 ] + (tα sα + ťα šα )X4,3e2 + 2tα sα X4,2e3 + (λtα sα + (1 − λ)ťα šα )X4,2e2e20 + O≺ (Ψ5 ),
(α)
Z3,2e2 = Eα [Z3,2e2 ] + (tα sα + ťα šα )X4,3e2 + 2ťα šα X4,2e3 + (λtα sα + (1 − λ)ťα šα )X4,2e2e20 + O≺ (Ψ5 ).
Substituting into (66),
"
#
X
Ǧ
1
G
iα iα
= sα šα Eα K 0 (Xλ )
Eα K 0 (Xλ )
Gik Ǧik − tα s2α šα Eα [Y3,120 + Y3,3 ]
tα ťα
N
−
k
2
ťα šα sα Eα [Z3,120 + Z3,3 ] − 2λtα s2α šα Eα [Y3,2e2 ] − 2(1 − λ)ťα š2α sα Eα [Z3,2e2 ]
sα šα (t2α s2α + tα sα ťα šα + ť2α š2α )(X4,220 + 2X4,13 + 4X4,4 + X4,40 )
+
+ sα šα (λtα sα + (1 − λ)ťα šα )(tα sα + ťα šα )(2X4,12e2 + 2X4,30e2 + 4X4,3e2 )
+ sα šα (λt2α s2α + (1 − λ)ť2α š2α )(2X4,2f
3)
30 + 4X4,2e
12 + 2X4,2e
+ 4sα šα (λtα sα + (1 − λ)ťα šα )2 X4,2e2e2 + O≺ (Ψ5 ).
(68)
Step 5: We take the full expectation of both sides of (68), applying Lemma 5.9 to convert Y3,∗
and Z3,∗ into X3,∗ . We illustrate the argument for Z3,120 :
For k 6= i, denote
(α)
Y = K 0 (Xλ )(m̌N − m̌∗ )Gik ,
(α)
(α)
Y (α) = K 0 (Xλ )(m̌N − m̌∗ )Gik .
Then Y ≺ Ψ2 , and Y − Y (α) ≺ Ψ3 for all α ∈ IM , the latter from Lemma 5.6 and the secondderivative bound for K. Then applying Lemma 5.9,
E[Y Ǧik ] = E[Y Gik ] + O≺ (Ψ5 ).
Hence
#
1 X
Gik (Ǧik − Gik ) = O≺ (Ψ5 ),
E K (Xλ )(m̌N − m̌∗ )
N
"
0
(69)
k
where the k = i term is controlled directly by Lemma 5.7. Applying this argument again with
Y = K 0 (Xλ )G2ik , together with the bound m̌∗ − m∗ ≤ C/N ≺ Ψ3 , we may convert the term
m̌N − m̌∗ :
"
#
1 X 2
0
E K (Xλ )(m̌N − m̌∗ − mN + m∗ )
Gik = O≺ (Ψ5 ).
(70)
N
k
Finally, a Taylor expansion of
0
K 0 (x)
around X yields
0
K (Xλ ) = K (X) + (1 − λ)K 00 (X)(X̌ − X) + O≺ (Ψ2 ),
(71)
where we have used X̌ − X ≺ Ψ by Lemma 5.7. Applying the third implication of Lemma 5.9 with
Y = K 00 (X)(mN − m∗ )G2ik ≺ Ψ3 for k 6= i, we obtain
"
#
X
1
E K 00 (X)(X̌ − X)(mN − m∗ )
G2ik = O≺ (Ψ5 ).
(72)
N
k
Then combining (69–72), we obtain E[Z3,120 ] = E[X3,120 ] + O≺ (Ψ5 ).
42
TRACY-WIDOM FOR COVARIANCE MATRICES
The same argument holds for the other terms Y3,∗ and Z3,∗ . Then taking the full expectation
of (68),
"
#
X
G
Ǧ
1
iα iα
E K 0 (Xλ )
= sα šα E K 0 (Xλ )
Gik Ǧik − (tα s2α šα + ťα š2α sα )E[X3,120 + X3,3 ]
tα ťα
N
−
k
2
2(λtα sα šα + (1 − λ)ťα š2α sα )E[X3,2e2 ]
sα šα (t2α s2α + tα sα ťα šα + ť2α š2α )E[X4,220
+
+ 2X4,13 + 4X4,4 + X4,40 ]
+ sα šα (λtα sα + (1 − λ)ťα šα )(tα sα + ťα šα )E[2X4,12e2 + 2X4,30e2 + 4X4,3e2 ]
+ sα šα (λt2α s2α + (1 − λ)ť2α š2α )E[2X4,2f
30 + 4X4,2e
3]
12 + 2X4,2e
+ 4sα šα (λtα sα + (1 − λ)ťα šα )2 E[X4,2e2e2 ] + O≺ (Ψ5 ).
(73)
R
R
R
λ2
R
we integrate (73) over λ ∈ [0, 1], applying λ = (1−λ) = 1/2 and
= 2λ(1−λ) =
R Finally,
−
2
(1 − λ) = 1/3. Simplifying the result and identifying the terms X3 , X4 , X4 , Pα , Qα , and Rα
concludes the proof of the lemma.
5.5. Proof of Lemma 5.12. We discuss briefly the proof of the optical theorems, Lemma 5.12.
In the setting K 0 ≡ 1, Lemma 5.12 corresponds to [LS16, Lemma B.1] upon taking the imaginary
part.
The proof for general K is the same as that of [LS16, Lemma B.1], with additional terms arising
from the Taylor expansion of K 0 as in the proof of Lemma 5.11. The computation may be broken
down into the following intermediate identities:
1
0
E[K 0 (X)] + 2m−1
∗ E[K (X)(mN − m∗ )]
N
−1
−4
5
= 2E[X3 ] − 2m−1
(74)
∗ (z − E∗ )E[X2 ] − (A4 − 2m∗ − m∗ )E[X4 ] + O≺ (Ψ ),
1
E[K 0 (X)(mN − m∗ )] − 2E[X4,220 + X4,13 + X4,4 + X4,12e2 ] = O≺ (Ψ5 ),
N
E[2X4,13 + 3X4,4 + X4,40 + 2X4,3e2 ] = O≺ (Ψ5 ),
(77)
5
(78)
E[X4,12e2 + 2X4,3e2 + X4,30e2 + X4,2f
3 + X4,2e
30 + 2X4,2e
2e
2 ] = O≺ (Ψ ),
12 + 2X4,2e
X2 = K 0 (X)
(76)
5
(z − E∗ )E[X2 ] − E[X4,220 + 4X4,4 + X4,40 + 2X4,30e2 ] = O≺ (Ψ ),
where
(75)
1 X 2
Gik .
N
k
For K 0 ≡ 1, (74) reduces to [LS16, (B.29)], (75) to [LS16, (B.33)], (76) to [LS16, (B.38)], and a
linear combination of (76) and (77) to [LS16, (B.51)]. The last identity (78) is trivial for K 0 ≡ 1,
as the left side is 0. It is analogous to [LS15, Eq. (C.42)] in the full computation for the deformed
Wigner model, and may be derived as an “optical theorem” from X3,2e2 in the same manner as (75)
and (76).
The derivations of these identities do not require positivity of T or E∗ : They use only the resolvent
identities of Lemma 3.9, the resolvent bounds from Section 5.1, the conditions |z − E∗ |≤ CN −2/3+ε
and γ = 1, and the additional bound
|m∗ + z + mM |≺ N −1/3+ε
for
mM = N −1
X
α
Gαα .
TRACY-WIDOM FOR COVARIANCE MATRICES
43
P
This additional bound follows from (13) written in the form z +m−1
0 =−
α Π0,αα , and the bounds
P
−1
−1
−1/3+ε/2
−1/3+ε
|m0 − m∗ |≤ CN
and |mM − α Π0,αα |≺ N
as implied by Proposition 3.7 and
Theorem 3.10. We omit further details and refer the reader to [LS16].
−1
Lemma 5.12 follows from substituting (75) and (77) into (74), adding 4m−1
∗ times (76) and 4m∗
0
times (78), and taking the imaginary part (noting K is real-valued). This concludes the proof.
Appendix A. Properties of ρ
A.1. Fixed N, M, T . We extend the results in [SC95] to the case x = 0 and prove Proposition 2.2
following ideas of [KY16, Lemma 2.5].
The following results are from [SC95]:
Proposition A.1. The Stieltjes transform m0 : C+ → C+ of ρ extends continuously to R \ {0}.
At each x ∈ R \ {0}, ρ admits a continuous density given by
1
f0 (x) = =m0 (x).
π
Proof. See [SC95, Theorem 1.1].
Proposition A.2. Let S = {m ∈ R \ P : z00 (m) > 0}, and denote z0 (S) = {z0 (m) : m ∈ S}. Then
R \ supp(ρ) = z0 (S).
More specifically, z0 : S → R \ supp(ρ) defines a bijection with inverse m0 .
Proof. See [SC95, Theorems 4.1 and 4.2].
Let us first extend Proposition A.1 to handle the case x = 0 (cf. Proposition A.5 below).
Lemma A.3. Denote m0 (C+ ) = {m0 (z) : z ∈ C+ }. For any m ∈ R \ P such that z00 (m) < 0, m
cannot belong to the closure of m0 (C+ ).
Proof. z0 defines an analytic function on C \ P . For any such m, the inverse function theorem
implies z0 has an analytic inverse in a neighborhood B of m in C \ P . If m belongs to the closure
of m0 (C+ ), then B ∩ m0 (C+ ) is non-empty. As z0 (m0 (z)) = z for z ∈ C+ by definition of m0 , the
inverse of z0 on B is an analytic extension of m0 to z0 (B). By the open mapping theorem, z0 (B)
is an open set in C containing m. On the other hand, as m0 is the Stieltjes transform of ρ, it
permits an analytic extension only to C \ supp(ρ), and this extension is real-valued and increasing
on R \ supp(ρ). Then z0 (B) ∩ R must belong to R \ supp(ρ) and z0 must be increasing on B ∩ R,
but this contradicts that z00 (m) < 0.
Lemma A.4. Define
g(q) = z0 (1/q) = −q +
M
1 X
t2α
tα −
.
N
q + tα
(79)
α=1
Then for any c ∈ R, there is at most one value q ∈ R for which g(q) = c and g 0 (q) ≤ 0.
Proof. Denote by P 0 = {−tα : tα 6= 0} the distinct poles of g, and let I1 , . . . , I|P 0 |+1 be the intervals
of R \ P 0 in increasing order. For any c ∈ R, boundary conditions of g at P 0 imply that g(q) = c has
at least one root q in each interval I2 , . . . , I|P 0 | , and hence at least |P 0 |−1 total roots. In addition,
every q ∈ R where g(q) = c and g 0 (q) ≤ 0 contributes two additional roots to g(q) = c, counting
multiplicity. As g(q) = c may be written as a polynomial equation in q of degree |P 0 |+1 by clearing
denominators, it can have at most |P 0 |+1 total roots counting multiplicity, and hence there is at
most one such q.
Proposition A.5. If rank(T ) > N , then m0 extends continuously to x = 0, and ρ has continuous
density f0 (x) = (1/π)=m0 (x) at x = 0.
If rank(T ) ≤ N , then for any sequence zn → 0 with zn ∈ C+ \ {0}, we have |m0 (zn )|→ ∞.
44
TRACY-WIDOM FOR COVARIANCE MATRICES
Proof. Suppose rank(T ) > N . Taking imaginary parts of (13) yields
M
1 X |tα m0 (z)|2
1−
N
|1 + tα m0 (z)|2
=m0 (z)
=z =
|m0 (z)|2
!
.
α=1
C+ ,
Both =z > 0 and =m0 (z) > 0 for z ∈
whereas if |m0 (zn )|→ ∞ along any sequence zn ∈ C+ ,
then
!
M
rank(T )
1 X |tα m0 (zn )|2
→1−
1−
.
N
|1 + tα m0 (zn )|2
N
α=1
When rank(T ) > N , this implies m0 (z) is bounded on all of C+ . In particular, it is bounded in
a neighborhood of x = 0, and the result follows from the same proof as [SC95, Theorem 1.1]. (If
m0 (zn ) → m ∈ C+ for any zn → x, then the limit exists and equals m by the open mapping
theorem applied to z0 in a neighborhood of m, and the uniquness of m0 (z) as the root to (13) in
C+ . If m0 (zn ) → m and m0 (zn0 ) → m0 for m, m0 ∈ R and two sequences zn → x and zn0 → x, then
for each m00 ∈ [m, m0 ] there exists a sequence zn00 → x such that m0 (zn00 ) → m00 . Then z0 is constant
on the interval [m, m0 ], implying m = m0 .)
Suppose now rank(T ) ≤ N . Note (13) holds for z ∈ C+ \{0} by continuity of m0 . If m0 (zn ) → m
for some finite m along any sequence zn ∈ C+ \{0} with zn → 0, then z0 (m) = limn z0 (m0 (zn )) = 0,
and m ∈
/ P . Rearranging (13) yields
zm0 (z) = −1 +
M
1 X
1
rank(T )
−
,
N
N
1 + tα m0 (z)
α=1
and taking real and imaginary parts followed by zn → 0 yields
M
1 X tα =m
.
N
|1 + tα m|2
α=1
α=1
P
When rank(T ) ≤ N , the first equation implies <m 6= 0 and α tα /|1 + tα m|2 6= 0, and the second
equation then implies =m = 0. Thus m ∈ R \ P . But recalling g(q) from (79), we have g(0) = 0
and g 0 (0) ≤ 0 when rank(T ) ≤ N , so Lemma A.4 implies g 0 (q) > 0 for every other q where g(q) = 0.
Thus z00 (m) < 0, but this contradicts Lemma A.3. Hence |m0 (zn )|→ ∞.
1−
M
1 X 1 + tα <m
rank(T )
=−
,
N
N
|1 + tα m|2
0=
We record the following consequence:
Proposition A.6. If E∗ is a soft edge of ρ with m-value m∗ , then E∗ ∈ R∗ , m0 extends continuously
to E∗ , and m0 (E∗ ) = m∗ .
Proof. Recalling g(q) from (79), if E∗ = 0 is a soft edge, then g(−1/m∗ ) = 0 and g 0 (−1/m∗ ) = 0.
Hence Lemma A.4 implies g 0 (0) > 0, so rank(T ) > N . Thus any soft edge E∗ belongs to R∗ .
Propositions A.1 and A.5 then imply continuous extension of m0 to E∗ . Considering m ∈ R with
z00 (m) > 0 and m → m∗ , Proposition A.2 implies m0 (z0 (m)) = m, while continuity of z0 and m0
yield z0 (m) → z0 (m∗ ) = E∗ and m0 (z0 (m)) → m0 (E∗ ). Hence m0 (E∗ ) = m∗ .
Using the above, we may establish Proposition 2.2.
Proof of Proposition 2.2. Let g(q) be as in Lemma A.4. If mj is a local minimum (or maximum) of
z0 , then qj = 1/mj is a local minimum (resp. maximum) of g, where qj = 0 if mj = ∞. Furthermore
these are the only local extrema of g, and they are ordered as q1 < . . . < qn . We have Ej = g(qj )
for each j = 1, . . . , n.
Let P 0 = {−tα : tα 6= 0} be the poles of g, and let I1 , . . . , I|P 0 |+1 be the intervals of R \ P 0 in
increasing order. Denoting
S 0 = {q ∈ R \ P 0 : g 0 (q) < 0},
TRACY-WIDOM FOR COVARIANCE MATRICES
45
Proposition A.2 is rephrased in terms of g as
R \ supp(ρ) = g(S 0 \ {0}).
(80)
(We must remove 0 from S 0 , as m = ∞ is not included in S.) As g 000 (q) > 0 for all q ∈ R \ P 0 , we
have that g 0 (q) is convex on each Ij . Together with the boundary conditions g 0 (q) → ∞ as q → P 0
and g 0 (q) → −1 as q → ±∞, this implies I1 contains the single local extremum q1 (a minimum),
I|P 0 |+1 contains the single local extremum qn (a maximum), and each Ij for j = 2, . . . , |P 0 | contains
either 0 or 2 local extrema (a maximum followed by a minimum). Hence S 0 is a union of open
intervals, say J1 , . . . , Jr , with at most one such interval contained in each Ij . Lemma A.4 verifies
g(Jj ) ∩ g(Jk ) = ∅
(81)
for all j 6= k. Together with (80), this verifies that the edges of ρ are precisely the values g(qj ),
with a local maximum qj corresponding to a left edge and a local minimum qj corresponding to a
right edge. If 0 ∈ S 0 , then it belongs to the interior of some open interval Jj , and supp(ρ) contains
an isolated point at 0 which is not considered an edge. This establishes (a) and (b).
The ordering in part (c) follows from a continuity argument as in [KY16, Lemma 2.5]: Define
for λ ∈ (0, 1]
M
λ X
t2α
gλ (q) = −q +
tα −
.
N
q + tα
α=1
is increasing in λ for each fixed q ∈ R \ P 0 . Hence for each local minimum (or
g, we may define a path qj (λ), continuous and increasing (resp. decreasing) in λ,
such that qj (1) = qj and qj (λ) remains a local minimum (resp. maximum) of gλ for each λ ∈ (0, 1].
As λ & 0, each qj (λ) converges to a pole −tα in P 0 , with gλ (qj (λ)) & tα if qj (λ) % −tα and
gλ (qj (λ)) % tα if qj (λ) & −tα . Hence for sufficiently small λ > 0,
Note that gλ0 (q)
maximum) qj of
gλ (q1 (λ)) > . . . > gλ (qn (λ)).
Lemma A.4 applies to gλ for each fixed λ, implying in particular that gλ (qj (λ)) 6= gλ (qk (λ)) for any
j 6= k. Hence by continuity in λ, the above ordering is preserved for all λ ∈ (0, 1]. In particular it
holds at λ = 1, which establishes (c).
Finally, for part (d), Proposition A.6 yields mj ∈ R∗ and m0 (Ej ) = mj for any soft edge Ej . The
previous convexity argument implies g 00 (qj ) 6= 0 for any local extremum qj , and hence z000 (mj ) 6= 0.
Taking x ∈ supp(ρ) with x → Ej , continuity of m0 implies m0 (x) → mj . As z0 is analytic at mj
and z00 (mj ) = 0, a Taylor expansion yields, as x → Ej ,
x − Ej = z0 (m0 (x)) − z0 (mj ) =
Since =m0 (x) > 0 and =mj = 0, this yields
s
m0 (x) − mj =
z000 (mj )
(1 + o(1))(m0 (x) − mj )2 .
2
2
(x − Ej )(1 + o(1)),
z000 (mj )
where we take the square root with branch cut on the positive real axis and having positive imaginary part. Taking imaginary parts and recalling f0 (x) = (1/π)=m0 (x) yields (d).
A.2. Behavior near regular edges. We now turn to the implications of edge regularity and
prove Propositions 2.7, 3.6, and 3.7; the arguments are similar to those of [KY16, Appendix A].
We first quantify continuity of m0 , uniformly in N , near a regular edge E∗ . In particular this
implies that when |z − E∗ | is small, |m0 (z) − m∗ | is also small.
46
TRACY-WIDOM FOR COVARIANCE MATRICES
Lemma A.7. Suppose Assumption 2.4 holds and E∗ is a regular edge with m-value m∗ . Then
there exist constants C, δ > 0 such that
(E∗ − δ, E∗ + δ) ⊂ R∗ ,
and for every z ∈ C+ with |z − E∗ |< δ,
|m0 (z) − m∗ |2 < C|z − E∗ |.
Proof. Applying Proposition 3.5, take a constant ν > 0 such that |m∗ |> ν. Fix a constant c ≡
c(τ ) < min(ν, τ ) to be determined later, and define
δN = min c, inf(δ > 0 : |m0 (z) − m∗ |< c for all z ∈ C+ ∪ R∗ such that |z − E∗ |≤ δ) .
As m0 (E∗ ) = m∗ , continuity of m0 at E∗ implies δN > 0. Furthermore, if rank(T ) ≤ N so that
0∈
/ R∗ , then the divergence of m0 at 0 from Proposition A.5 implies (E∗ − δN , E∗ + δN ) ⊂ R∗ .
A priori, δN may depend on N . We will first establish that |m0 (z) − m∗ |2 < C|z − E∗ | when
|z − E∗ |≤ δN . This will then imply that δN is bounded below by a constant δ ≡ δ(τ ).
Consider z ∈ C+ with |z − E∗ |≤ δN . Let us write as shorthand m = m0 (z). Then
|z − E∗ | = |z0 (m) − z0 (m∗ )|
M
1
1 X
t2α
= |m − m∗ | −
+
mm∗ N
(1 + tα m)(1 + tα m∗ )
= |m − m∗ |2 −
α=1
M
X
1
1
+
mm2∗ N
α=1
t3α
,
(1 + tα m)(1 + tα m∗ )2
(82)
where the last line adds to the quantity inside the modulus
0=
z00 (m∗ )
M
1
1 X
t2α
= 2−
.
m∗ N
(1 + tα m∗ )2
α=1
As |m − m∗ |< c by definition of δN , we have for each non-zero tα
1
1
c
−
,
<
m m∗
ν(ν − c)
1
c
1
−1 −
−1 < τ (τ − c) .
m + tα
m∗ + tα
Applying this to (82) and recalling γ −2 = |z000 (m∗ )|/2 yields
c
c
|z − E∗ |> |m − m∗ |2 γ −2 − 3
− d−1 3
.
ν (ν − c)
τ (τ − c)
As γ −2 > τ 2 , this implies |m0 (z) − m∗ |2 < C|z − E∗ | when c is chosen sufficiently small, as desired.
By continuity of m0 and definition of δN , either δN = c or there must exist z ∈ C+ such that
|z − E∗ |= δN and |m0 (z) − m∗ |= c. In the latter case, for this z we have c2 = |m0 (z) − m∗ |2 <
C|z − E∗ |= CδN , implying δN > c2 /C. Thus in both cases δN is bounded below by a constant,
yielding the lemma.
Next we bound the third derivative of z0 near the m-value of a regular edge.
Lemma A.8. Suppose Assumption 2.4 holds and E∗ is a regular edge with m-value m∗ . Then
there exist constants C, δ > 0 such that z0 is analytic on the disk {m ∈ C : |m − m∗ |< δ}, and for
every m in this disk,
|z0000 (m)|< C.
TRACY-WIDOM FOR COVARIANCE MATRICES
47
Proof. Proposition 3.5 ensures |m∗ |> ν for a constant ν > 0. Taking δ < min(ν, τ ), the disk
D = {m ∈ C : |m − m∗ |< δ} does not contain any pole of z0 , and hence z0 is analytic on D. We
compute
6
1 X
6
z0000 (m) = 4 −
,
−1
m
N
(tα + m)4
α:tα 6=0
|z0000 (m)|<
so
d > c.
C for m ∈ D and sufficiently small δ by the bounds |m∗ |> ν, |m∗ + t−1
α |> τ , and
Propositions 2.7, 3.6, and 3.7 now follow:
Proof of Proposition 3.6. This follows from Taylor expansion of z000 at m∗ , the condition |z000 (m∗ )|=
2γ −2 > 2τ 2 implied by regularity, and Lemma A.8.
Proof of Proposition 2.7(a). Let C, δ > 0 be as in Lemma A.7. Reducing δ as necessary and
applying Lemma A.8, we may assume z0 is analytic with |z0000 (m)|< C 0 over the disk
√
D = {m ∈ C : |m − m∗ |< Cδ},
for a constant C 0 > 0.
Let E ∗ be the closest other edge to E∗ , and suppose E ∗ ∈ (E∗ −δ, E∗ +δ). Let m∗ be the m-value
for E ∗ . Then Lemma A.7 implies m∗ ∈ D. Applying a Taylor expansion of z00 ,
z0000 (m) ∗
(m − m∗ )2
2
for some m between m∗ and m∗ . Applying 0 = z00 (m∗ ) = z00 (m∗ ), |z000 (m∗ )|= 2γ −2 > 2τ 2 , and
|z0000 (m)|< C 0 , we obtain |m∗ − m∗ |> 4τ 2 /C 0 . Then Lemma A.7 yields |E ∗ − E∗ |> c for a constant
c > 0. Reducing δ to c if necessary, we ensure (E∗ − δ, E∗ + δ) contains no other edge E ∗ . The
condition (E∗ − δ, E∗ + δ) ⊂ R∗ was established in Lemma A.7.
z00 (m∗ ) = z00 (m∗ ) + z000 (m∗ )(m∗ − m∗ ) +
Proof of Propositions 3.7 and 2.7(b). Taking δ sufficiently small, Lemma A.7 implies |m0 (z)−m∗ |<
√
Cδ for all z ∈ D0 . Then |m0 (z)| 1 and |1 + tα m0 (z)| 1 by Proposition 3.5. Reducing δ if
necessary, by Lemma A.8 we may also ensure z0 is analytic with |z0000 (m)|< C 0 on
√
D = {m ∈ C : |m − m∗ |< Cδ}.
Note z = z0 (m0 (z)) by (13) while E∗ = z0 (m∗ ). Then taking a Taylor expansion of z0 and applying
the conditions z00 (m∗ ) = 0, z000 (m∗ ) = 2γ −2 , and |z0000 (m̃)|< C 0 for all m̃ ∈ D, we have
where |r(z)|< C
√
0
z − E∗ = z0 (m0 (z)) − z0 (m∗ ) = (γ −2 + r(z))(m0 (z) − m∗ )2
(83)
Cδ/6. Taking δ sufficiently small, we ensure
|γ −2 + r(z)| 1,
arg(γ −2 + r(z)) ∈ (−ε, ε)
(84)
for an arbitrarily small constant ε > 0, where arg(z)pdenotes the complex argument. Taking the
√
modulus of (83) on both sides yields |m0 (z) − m∗ | |z − E∗ | κ + η.
For =m0 (z), suppose E∗ is a right edge. (The case of a left edge is similar.) By Proposition
2.7(a), we may assume (E∗ − δ, E∗ ) ⊂ supp(ρ) and (E∗ , E∗ + δ) ⊂ R \ supp(ρ). First suppose z 6= E∗
and E ≡ <z ≤ E∗ . As =m0 (z) > 0 by definition, (83) yields
p
m0 (z) − m∗ = (z − E∗ )/(γ −2 + r(z))
where the square-root has branch cut on the positive real
positive imaginary
part. Applying
√ axis and √
√
arg(z −E∗ ) ∈ [π/2, π) and (84), we have =m0 (z) = z − E∗ | z − E∗ | κ + η. By continuity
of m0 , this extends to z ∈ (E∗ − δ, E∗ ) on the real axis. Hence Proposition 2.7(b) also follows, as
f0 (x) = π −1 =m0 (x).
48
TRACY-WIDOM FOR COVARIANCE MATRICES
Finally, suppose E > E∗ . Let us write
Z
Z
η
η
=m0 (z) =
ρ(dλ) +
ρ(dλ) ≡ I + II.
2
2
2
2
|λ−E∗ |<δ (λ − E) + η
|λ−E∗ |≥δ (λ − E) + η
2
Reducing δ to δ/2, we may
√ assume the closest edge to E is E∗ . Then we have II ∈ [0, η/δ ]. For I,
as ρ has density f0 (x) E∗ − x for x ∈ (E∗ − δ, E∗ ) while (E∗ , E∗ + δ) ⊂ R \ supp(ρ),
Z E∗
Z δ
p
√
η
η
I
E∗ − λ dλ =
x dx.
2
2
2
2
E∗ −δ (λ − E) + η
0 η + (κ + x)
√
Considering separately the integral over x ∈ √
[0, κ + η] and x ∈ [κ + η, δ], we obtain I η/ η + κ.
Then II ≤ C · I, and this yields =m0 (z) η/ η + κ.
Appendix B. Local law
We show that the proof of the local law in [KY16] extends to non-positive-definite T with minor
and cosmetic modifications. As in [KY16], we may deduce from this Theorems 2.8 and 2.9 and
Corollary 3.12.
We follow the approach of [KY16] and establish a local law over a spectral domain that satisfies
an abstract stability condition. We will use a modification of the original stability condition in
[BEK+ 14, Lemma 4.5], rather than the one in [KY16], as we find its statement and verification
a bit clearer. (The constant C is not appearing in both the assumption at w ∈ L(z) and the
implication at z, so larger C indicates a weaker condition.)
Definition B.1. Fix a bounded set S ⊂ R and a constant a > 0, and let
D = {z ∈ C+ : <z ∈ S, =z ∈ [N −1+a , 1]}.
For z = E + iη ∈ D, denote
L(z) = {z} ∪ {w ∈ D : <w = E, =w ∈ [η, 1] ∩ (N −6 N)}.
For a positive bounded function g : D → (0, C), the Marcenko-Pastur equation (13) is g -stable
on D if the following holds for some constant C > 0: Let u : C+ → C+ be the Stieltjes transform
of any probability measure, and let ∆ : D → (0, ∞) be any function satisfying
• (Boundedness) ∆(z) ∈ [N −2 , (log N )−1 ] for all z ∈ D,
• (Lipschitz) |∆(z) − ∆(w)|≤ N 2 |z − w| for all z, w ∈ D,
• (Monotonicity) η 7→ ∆(E + iη) is non-increasing for each E ∈ S.
If z ∈ D is such that |z0 (u(w)) − w|≤ ∆(w) for all w ∈ L(z), then
|u(z) − m0 (z)|≤
C∆(z)
p
.
g(z) + ∆(z)
(85)
Theorem B.2 (Abstract local law). Suppose Assumptions 2.4 and 2.5 hold. Fix S ⊂ R and a
constant a > 0, and denote
D = {z ∈ C+ : <z ∈ S, =z ∈ [N −1+a , 1]}.
Suppose for constants C, c > 0 and a bounded function g : D → (0, C) that (13) is g-stable on D
and
c < |m0 (z)|< C,
|1 + tα m0 (z)|> c,
=m0 (z) < Cg(z)
for all z = E + iη ∈ D. Then, letting T , Π0 , G, and mN be as in Theorem 3.10, and denoting
s
=m0 (z)
1
Ψ(z) =
+
,
Nη
Nη
TRACY-WIDOM FOR COVARIANCE MATRICES
49
(a) (Anisotropic law) Uniformly over z ∈ D and all v, w ∈ CN +M with kvk= kwk= 1,
v ∗ T −1 (G(z) − Π0 (z))T −1 w ≺ Ψ(z).
(b) (Averaged law) Uniformly over z ∈ D,
mN (z) − m0 (z) ≺ min
1 Ψ(z)2
,
N η g(z)
.
Proof. The proof is the same as [KY16, Theorems 3.21 and 3.22], with only cosmetic differences
which we indicate here. The notational identification with [KY16] is T ↔ Σ and tα ↔ σi . (We
continue to use Greek indices for IM and Roman indices for IN , although this is reversed from the
convention in [KY16].) The proof follows three steps:
1. The averaged law and an entrywise law (i.e. for v, w being standard basis vectors) when T is
diagonal and X has Gaussian entries.
2. The anisotropic law when X has Gaussian entries.
3. The averaged law and anisotropic law for general X.
As in [KY16], we may assume T is invertible. The non-invertible case follows by continuity.
Step 1 follows [KY16, Section 5], which in turn is based on [PY14] and [BEK+ 14]. Denoting
(S)
by GM and GM the lower-right blocks of G and G(S) , all of the resolvent identities of [KY16,
Lemma 4.4] remain valid for non-positive-definite T , and [KY16, Eqs. (4.17)–(4.20)] and [KY16,
Lemmas 4.8 and 4.9] also hold with identical proofs. Denote tA = 1 for A = i ∈ IN and tA = tα
for A = α ∈ IM , and define
X
X (α)
(i)
(i)
(α)
Zi =
Gαβ Xαi Xβi − N −1 Tr GM ,
Zα =
Gij Xαi Xαj − N −1 Tr GN ,
i,j∈IN
α,β∈IM
X
1 X
t2α
[Z] =
Zα ,
Zi +
N
(1 + tα m0 )2
i∈IN
α∈IM
s
Θ = N −1
X
(G − Π0 )ii + M −1
i∈IN
Λo = max
A6=B∈I
X
(G − Π0 )αα ,
ΨΘ =
α∈IM
|GAB |
,
|tA tB |
Λ = max
A,B∈I
|(G − Π0 )AB |
,
|tA tB |
=m0 + Θ
,
Nη
Ξ = {Λ ≤ (log N )−1 }.
These all implicitly depend on an argument z ∈ D. Then the same steps as in [KY16, Section 5]
yield, either for η = 1 or on the event Ξ, for all z ∈ D and A ∈ I,
|ZA |, Λo ≺ ΨΘ ,
z0 (mN (z)) − z − [Z] ≺
Ψ2Θ
≺ (N η)
(86)
−1
.
(87)
(In the argument for η = 1, the use of [KY16, Eq. (4.16)] may be replaced by [KY16, Lemmas
4.8 and 4.9]. In [KY16, Eqs. (5.4), (5.11), etc.], σi is to be replaced by the positive quantity |tα |.)
Applying (86) and the resolvent identities for Gii and Gαα , we may also obtain on the event Ξ
Θ ≺ |mN − m0 |+|[Z]|+(N η)−1 ,
Λ ≺ |mN − m0 |+ΨΘ .
(88)
The bound (86) yields the initial estimate [Z] ≺ ΨΘ ≺ (N η)−1/2 on Ξ. The conditions of
Definition B.1 hold for ∆ = (N η)−1/2 , so (87), the assumed stability of (13), and the stochastic
continuity argument of [BEK+ 14, Section 4.1] yield that Ξ holds with high probability (i.e. 1 ≺
1{Ξ}) and Λ ≺ (N η)−1/4 on all of D. Next, applying the fluctuation averaging result of [KY16,
50
TRACY-WIDOM FOR COVARIANCE MATRICES
Lemma 5.6] (whose proof uses resolvent identities as in Step 2 below and does not require positivedefinite T ), we obtain for any c ∈ (0, 1] the implications
s
=m0 + (N η)−c
=m0 + (N η)−c
⇒ [Z] ≺
≡ ∆(z).
Θ ≺ (N η)−c ⇒ ΨΘ ≺
Nη
Nη
The conditions of Definition B.1 hold for this ∆(z) by known properties of the Stieltjes transform,
so applying (87), stability of (13), and 1 ≺ 1{Ξ(w)} for all w ∈ L(z), we have the implications
∆(z)
∆(z)
p
p
⇒Θ≺
+ ∆(z) + (N η)−1 .
g(z) + ∆(z)
g(z) + ∆(z)
Θ ≺ (N η)−c ⇒ |mN − m0 |≺
(89)
We bound ∆(z) ≤ C(N η)−1 and
∆(z)
=m0 (z)
p
+ (N η)−(1+c)/2 < (N η)−1 + (N η)−(1+c)/2 ,
≤
N
η
g(z)
g(z) + ∆(z)
where this applies =m0 (z) < g(z). Hence
Θ ≺ (N η)−c ⇒ Θ ≺ (N η)−(1+c)/2 .
Initializing to c = 1/4 and iterating, we obtain Θ ≺ (N η)−1+ε for any ε > 0, so |mN − m0 |≤ Θ ≺
(N η)−1 . Applying (89) once more with c = 1, we have for c = 1 that ∆(z) ≤ Ψ(z)2 and hence also
|mN − m0 |≺ Ψ2 /g. This yields both bounds in the averaged law. The entrywise law Λ ≺ Ψ follows
from (88).
Step 2, the anisotropic law for Gaussian X, follows [KY16, Section 6] with details as in [BEK+ 14,
Section 5]. Upon applying rotational invariance of X and diagonalizing T , we must establish, for
diagonal T , the bounds E|Z|p ≺ Ψp and E|Z 0 |p ≺ Ψp where
X vα Gαi
X vα Gαβ wβ
,
Z0 =
,
Z=
|tα tβ |
|tα |
α∈IM
α6=β∈IM
v, w ∈ CIM are fixed unit vectors, i ∈ IN is a fixed index, and p is an arbitrarily large constant
power. Let α, β, δ ∈ IM and S ⊂ IM denote distinct indices. For E|Z|p , we carry out the graph
operations (a,b,c) of [BEK+ 14, Section 5.7], where operation (a) is implemented by the identities
(S)
(Sδ)
(S)
(S)
Gδβ |tδ |
G
=
− |tδ | αδ
,
|tα tβ |
|tα tβ |
|tα tδ | |tδ tβ | G(S)
δδ
!
(S) 2
(S)
(Sδ)
Gαδ
Gαα
Gαα
|tδ |
=
− |tα tδ |
,
(S)
|tα |
|tα |
|tα tδ |
Gδδ
!
(S) 2
Gαδ
|tα |
|tα |
|tα | |tα | |tδ |
,
= (Sδ) − |tα tδ |
(S)
(S) (Sδ) (S)
|tα tδ |
Gαα
Gαα
Gαα Gαα Gδδ
Gαβ
Gαβ
operation (b) is implemented by the identity
(S)
(Sα)
(S)
Gαα Gββ
=
|tα tβ |
|tα | |tβ |
Gαβ
X
i,j∈IN
and operation (c) is implemented by the identities
X
|tα |
(Sα)
=
±1
±
t
Xαi Gij Xαj ,
α
(S)
Gαα
i,j∈IN
(Sαβ)
Xαi Gij
Xβj ,
TRACY-WIDOM FOR COVARIANCE MATRICES
(S)
Gαα
=±
P
|tα |
1 + tα
1
(Sα)
i,j∈IN
=±
L−1
X
−
k=0
51
Xαi Gij
1
1 + tα m0
k+1
Xαj
tkα
k
X
(Sα)
Xαi Gij
Xαj − m0 + O≺ (ΨL ),
i,j∈IN
which all follow from the resolvent identities in Lemma 3.9. Applying these operations according
(S)
to [BEK+ 14, Section 5], and finally applying the bounds |tα |≤ C, |1 + tα m0 |≥ c, and Gαα /tα ≺ 1,
(S)
(S)
tα /Gαα ≺ 1, and Gαβ /tα tβ ≺ Ψ for α 6= β ∈ IM \ S from the entrywise law, we obtain the bound
on E|Z|p . The bound on E|Z 0 |p follows similarly.
Step 3, the full result for general X, follows exactly as in [KY16, Sections 7, 8, and 9]. The only
proof modification required is in [KY16, Lemma 7.12], where the spectral decomposition in [KY16,
Eq. (4.15)] for G may be replaced by
N
X
uk u0k
0 0
+
,
G=
0 Σ
λk − z
k=1
where Σ̂ =
PN
0
k=1 λk ζk ζk
is the spectral decomposition of Σ̂ = X 0 T X and where
ζk
uk =
.
T Xζk
This follows from (22) and the definition of GN . Then the proof of [KY16, Lemma 7.12] holds with
this definition of uk , where we still have
=Gxx (z) =
N
X
k=1
hx, uk i2 η
(λk − E)2 + η 2
for z = E + iη. The remainder of Step 3 is identical to the proof in [KY16]. In particular, as g(z)
is bounded, the averaged law bound
1 Ψ(z)2
,
Φ(z) ≡ min
N η g(z)
from Step 1 satisfies [KY16, Eq. (9.1)], so [KY16, Proposition 9.1] carries through.
We now verify Theorems 3.10, 2.8, 2.9, and Corollary 3.12. Most of the remaining work is to
verify the stability condition in Definition B.1.
Lemma B.3. Suppose Assumption 2.4 holds and E∗ is a regular edge. Then for some constants
C, δ > 0, the following holds:
Define D by (23) for any a > 0. For z = E + iη ∈ D, denoting
√
κ = |E − E∗ |, let g(z) = κ + η. Then (13) is g-stable on D.
Proof. The proof follows [BEK+ 14, Lemma 4.5] and [KY16, Lemma A.5]: Fix z ∈ D and suppose
|z0 (u(w)) − w|≤ ∆(w) for all w ∈ L(z). We wish to show
C∆(z)
|u(z) − m0 (z)|≤ p
.
κ + η + ∆(z)
(90)
Suppose first =z > ν for a constant ν > 0. Then =m0 (z) > cν. By assumption |z0 (u(z)) − z|≤
∆(z) ≤ (log N )−1 , so also =z0 (u(z)) > cν for a constant c > 0. Since m0 is (cν)−2 -Lipschitz over
{z : =z > cν}, this yields
|u(z) − m0 (z)|= |m0 (z0 (u(z))) − m0 (z)|≤ ∆(z)/(cν)2 ,
52
TRACY-WIDOM FOR COVARIANCE MATRICES
and hence (90) holds at z. Thus it remains to check the case =z ≤ ν. Taking δ and ν sufficiently
small, and applying Lemma A.7 and |z0 (u(w)) − w|≤ ∆(w) ≤ (log N )−1 , we may assume
|m0 (w) − m∗ |< ε,
|u(w) − m∗ |< ε
(91)
for all w ∈ L(z) ∩ {w : =w ≤ ν}, where ε > 0 is an arbitrarily small constant.
For any w ∈ L(z) ∩ {w : =w ≤ ν}, writing u = u(w), m = m0 (w), and ∆0 (w) = z0 (u) − w, we
verify
!
M
m−u
1 X
t2α um
∆0 (w) = z0 (u) − z0 (m) =
−1 +
um
N
(1 + tα u)(1 + tα m)
α=1
2
= α(w)(m − u) + β(w)(m − u)
for
M
t2α
1 1 X
,
·
u N
(1 + tα u)(1 + tα m)2
α=1
!
M
1 X
t2α m2
m
−1 +
= − z00 (m).
N
(1 + tα m)2
u
α(w) = −
β(w) =
1
um
α=1
Denoting the two values
R1 (w), R2 (w) = −
β(w) ±
p
β(w)2 + 4α(w)∆0 (w)
,
2α(w)
(92)
we obtain m0 (w) − u(w) ∈ {R1 (w), R2 (w)}.
We bound α(w) and β(w) under (91): Applying 0 = z00 (m∗ ), we have
M
M
m∗ z000 (m∗ )
1
1 X
t2α m∗
1 X
t2α
=− 2 +
=
−
,
2
m∗ N
(1 + tα m∗ )3
N
(1 + tα m∗ )3
α=1
α=1
Z m
Z mZ x
0
00
00
z0000 (y)dy dx.
z0 (m) =
z0 (x)dx = (m − m∗ )z0 (m∗ ) +
m∗
m∗
m∗
Applying Propositions 3.5,
√ 3.7, and Lemma A.8, for sufficiently small ε in (91) we obtain |α(w)| 1
and |β(w)| |m − m∗ | κ + ηw . (Here and below, ηw = =w.)
The remainder of the argument is as in [BEK+ 14, Lemma 4.5]: Applying the bounds on |α(w)|
and |β(w)| to (92), there exists a constant C0 > 0 for which, for all w ∈ L(z) ∩ {=w ≤ ν},
C0 |∆0 (w)|
C0 ∆(w)
min(|R1 (w)|, |R2 (w)|) ≤ p
≤p
,
κ + ηw + |∆0 (w)|
κ + ηw + ∆(w)
p
p
|R1 (w) − R2 (w)| ≤ C0 κ + ηw + |∆0 (w)| ≤ C0 κ + ηw + ∆(w),
p
p
|R1 (w) − R2 (w)| ≥ C0−1 (κ + ηw − |∆0 (w)|)+ ≥ C0−1 (κ + ηw − ∆(w))+ ,
(93)
(94)
(95)
where the second inequality in each case applies |∆0 (w)|≤ ∆(w) and monotonicity. Increasing C0
as necessary, we may assume also
|u(w) − m0 (w)|≤ p
C0 ∆(w)
κ + ηw + ∆(w)
for w ∈ L(z) ∩ {w : =w > ν}, as we have established (90) in this case.
If
C0 ∆(w)
1 p
p
>
(κ + ηw − ∆(w))+
2C0
κ + ηw + ∆(w)
(96)
TRACY-WIDOM FOR COVARIANCE MATRICES
53
for any w ∈ L(z), then by monotonicity of ∆(w) in ηw , this holds also at w = z. Then rearranging
yields κ + η < C∆(z), so (93), (94), and m0 (z) − u(z) ∈ {R1 (z), R2 (z)} yield (90) as desired.
Otherwise,
C0 ∆(w)
1 p
p
≤
(κ + ηw − ∆(w))+
2C0
κ + ηw + ∆(w)
for all w ∈ L(z). If there exists w ∈ L(z) for which (96) does not hold, then letting w∗ be the one
with maximal imaginary part, we have =w∗ ≤ ν, so (93–95) hold at w = w∗ . This implies
1 p
2C0 ∆(w∗ )
|u(w∗ ) − m0 (w∗ )|= max(|R1 (w∗ )|, |R2 (w∗ )|) >
(κ + ηw∗ − ∆(w∗ ))+ ≥ p
,
C0
κ + ηw∗ + ∆(w∗ )
while (96) holds at w = w+ ≡ w∗ + iN −6 . However, from the conditions N −2 ≤ ∆(w) ≤ 1,
ηw ≥ N −1 , N 2 -Lipschitz continuity of u and m0 , and N 2 -Lipschitz continuity of ∆, we verify
|u(w∗ ) − u(w+ )|≤ CN −4 ,
|m0 (w∗ ) − m0 (w+ )|≤ CN −4 ,
∆(w+ )
∆(w∗ )
p
−p
≤ CN −2.5 ,
κ + ηw+ + ∆(w+ )
κ + ηw∗ + ∆(w∗ )
∆(w∗ )
p
≥ cN −2
κ + ηw∗ + ∆(w∗ )
for constants C, c > 0, which contradicts the above. Thus (96) holds at each w ∈ L(z), and in
particular at w = z, concluding the proof.
√
Proof of Theorem 3.10. This follows from Theorem B.2 applied with g(z) = κ + η, and Proposition 3.7 and Lemma B.3.
Proof of Theorem 2.9. The argument follows [PY14, Eq. (3.4)]. Consider the case of a right edge
E∗ . (A left edge is analogous.) For each E ∈ [E∗ + N −2/3+ε , E∗ + δ], denoting κ = E − E∗ , consider
z = E + iη for
η = N −1/2−ε/4 κ1/4 ∈ [N −2/3 , 1],
where the inclusion holds for all large N because κ ∈ [N −2/3+ε , δ]. Proposition 3.7 implies
=m0 (z) ≤ √
Cη
Cη
≤ √ = C(N η)−1 N −ε/2 .
κ+η
κ
√
Also by Proposition 3.7 and Lemma B.3, we may apply Theorem B.2 with g(z) = κ + η. The
above bound on =m0 (z) yields Ψ(z)2 ≤ C/(N η)2 , and hence Theorem B.2(b) implies
|mN (z) − m0 (z)|≺
(N η)2
1
1
1
√ = 3+ε/2 4 ≤ (N η)−1 N −ε/2 ,
√
≤
2
κ+η
(N η) κ
N
η
where the last bound uses η ≥ N −2/3 . Thus we obtain
=mN (z) ≺ C(N η)−1 N −ε/2 .
As the number of eigenvalues of W in [E − η, E + η] is at most 2N η · =mN (z), this implies W has
no eigenvalues in this interval for all large N . The result follows from a union bound over a grid of
such values E.
Proof of Theorem 2.8. By the bound kW k≤ kT kkX ∗ Xk and a spectral norm bound on X ∗ X,
e.g. from [BEK+ 14, Theorem 2.10], we may take C0 sufficiently large such that kW k≤ C0 with
probability at least N −D , for any D > 0 and all N ≥ N0 (D). This implies also supp(ρ) ⊂ [−C0 , C0 ].
For some small a > 0, let
D = {z ∈ C+ : <z ∈ [−C0 , C0 ] \ supp(ρ)δ , =z ∈ [N −1+a , 1]}.
54
TRACY-WIDOM FOR COVARIANCE MATRICES
For each z = E + iη ∈ D, we have δ ≤ |E − λ|≤ 2C0 for all λ ∈ supp(ρ), hence
Z
η
=m0 (z) =
ρ(dλ) η
(λ − E)2 + η 2
and
Z
|m0 (z)|≤
1
ρ(dλ) < C,
|λ − z|
|m00 (z)|≤
Z
1
ρ(dλ) < C
|λ − z|2
for a constant C > 0.
We verify the conditions of Theorem B.2 on D, for g(z) ≡ 1. For any E ∈ [−C0 , C0 ] \ supp(ρ)δ ,
Proposition A.2 implies there is m ∈ R such that z0 (m) = E, m0 (E) = m, and z00 (m) > 0. If m
is close to 0, i.e. |m|< 1/(2kT k), then m is separated from the poles of z0 : |m + t−1
α |> 1/(2kT k)
for each non-zero tα . Note z00 (x) ≤ 1/x2 ≤ 4kT k2 when |x|≥ 1/(2kT k). Thus if m is far from 0,
i.e. |m|≥ 1/(2kT k), then |m − m∗ |> δ/(4kT k2 ) for any local minimum or maximum m∗ of z0 , as
|E − E∗ |> δ for the corresponding edge E∗ . m is separated from all non-zero poles of z0 by such
0
an m∗ , so we have |m + t−1
α |> c0 for some c0 > 0 in both cases. Applying |m0 (z)|< C, this yields
|m0 (z) + t−1
α |> c0 /2
for all z ∈ D with =z < C/(2c0 ). For =z ≥ C/(2c0 ), we have |m0 (z) + t−1
α |≥ =m0 (z) > c. As
|m0 (z)|< C, considering separately the cases |tα |> 1/(2C) and |tα |≤ 1/(2C) yields |1 +tα m0 (z)|> c
for all non-zero tα . Applying |z|≤ C0 + 1 and |m0 (z) + t−1
α |> c to (13), we also obtain |m0 (z)|> c.
It remains to check that (13) is stable in the sense of Definition B.1 for g(z) ≡ 1, but this follows
easily: Suppose z ∈ D satisfies |z0 (u(z)) − z|≤ ∆(z) ≤ (log N )−1 . As |m00 (z)|< C still holds on a
(log N )−1 neighborhood of D, we have
p
|u(z) − m0 (z)|= |m0 (z0 (u(z)) − m0 (z)|≤ C|z0 (u(z)) − z|≤ C∆(z) = C min( ∆(z), ∆(z)).
Thus Theorem B.2(b) applies to yield |mN (z) − m0 (z)|≺ Ψ(z)2 N −1 + (N η)−2 for any z ∈ D.
Taking η = N −2/3 and applying also =m0 (z) η, we obtain =mN (z) ≺ N −2/3 < 1/(2N η). Then
for any D > 0, W has no eigenvalues in [E − η, E + η] with probability 1 − N −D for N ≥ N0 (D).
The result follows from a union bound over a grid of values E ∈ [−C0 , C0 ] \ supp(ρ)δ , together with
bound kW k≤ C0 .
Proof of Corollary 3.12. The proof is similar to [EYY12b, Theorem 2.4]: Consider two matrices
Σ̂ = X 0 T X and Σ̃ = Y 0 T Y where Y has independent Normal(0, 1/N ) Gaussian entries. Applying
Lemma 5.1 as in the proof of Theorem 2.10, it suffices to show
|E[K(X) − K(X̌)]|= o(1),
where X ≡ X(s, N ε , N −2/3−9ε ) for fixed s ∈ R, and X̌ is the same quantity defined by Σ̃. Introducing
a Lindeberg sequence that swaps the entries of X for Gaussian entries of Y one at a time, it suffices
to show
E[K(X) − K(X̌)] ≺ N −13/6+Cε
for each such swap X ↔ X̌ and some constant C > 0. Writing N =mN (z) = η Tr GN (z)GN (z), the
proof then follows [EYY12b, Theorem 2.4], where we expand the linearized resolvent G from (21)
as
Ǧ = G + GV Ǧ = . . . = G + GV G + (GV )2 G + (GV )3 G + (GV )4 Ǧ
with V = G−1 − Ǧ−1 . V has a symmetric pair of non-zero entries, and all other entries 0. The
proof of [EYY12b, Theorem 2.4] then carries through using the resolvent bounds of Lemma 5.2,
where we may ignore the tα ’s (as they are bounded) and use simply Giα , Gαβ ≺ Ψ and Gαα ≺ 1.
We note that there is no separate case of “diagonal entry swaps” in our setting, and in [EYY12b,
Eq. (6.58)] we can only have k = 0 or 1, as one index of the replaced entry belongs to IM and thus
TRACY-WIDOM FOR COVARIANCE MATRICES
55
cannot coincide with i, j ∈ IN . The analogue of [EYY12b, Lemma 6.6] that is required to conclude
the proof is
E[K 0 (X)Gij Gjk Gαi ] ≺ N −4/3+Cε
(97)
for distinct entries i, j, k ∈ IN and α ∈ IM . This follows by applying Lemma 5.6 and the bounded
(α)
(α)
second-derivative of K to replace X, Gij , Gjk by X(α) , Gij , Gjk , and then applying
Gαi = Gαα
X
k
(α)
Gik Xαk = −
X (α)
tα
Gik Xαk + O≺ (N −2/3+2ε )
1 + tα m∗
k
as in the proof of Lemma 5.9. The partial expectation Eα of the leading order term is 0, thus
yielding (97). We omit further details and refer the reader to [EYY12b].
References
[Ame85]
Yasuo Amemiya. What should be done when an estimated between-group covariance matrix is not nonnegative definite? The American Statistician, 39(2):112–117, 1985.
[BAC+ 15] Mark W Blows, Scott L Allen, Julie M Collet, Stephen F Chenoweth, and Katrina McGuigan. The
phenome-wide distribution of genetic variance. The American Naturalist, 186(1):15–30, 2015.
[BBAP05] Jinho Baik, Gérard Ben Arous, and Sandrine Péché. Phase transition of the largest eigenvalue for nonnull
complex sample covariance matrices. The Annals of Probability, 33(5):1643–1697, 2005.
[BEK+ 14] Alex Bloemendal, László Erdos, Antti Knowles, Horng-Tzer Yau, and Jun Yin. Isotropic local laws for
sample covariance and generalized Wigner matrices. Electronic Journal of Probability, 19(33):1–53, 2014.
[BM15]
Mark W Blows and Katrina McGuigan. The distribution of genetic variance across phenotypic space and
the response to selection. Molecular ecology, 24(9):2056–2072, 2015.
[BPZ13]
Zhigang Bao, Guangming Pan, and Wang Zhou. Local density of the spectrum on the edge for sample
covariance matrices with general population. Technical report, 2013.
[BS98]
Zhi-Dong Bai and Jack W Silverstein. No eigenvalues outside the support of the limiting spectral distribution of large-dimensional sample covariance matrices. The Annals of Probability, 26(1):316–345, 1998.
[BS06]
Jinho Baik and Jack W Silverstein. Eigenvalues of large sample covariance matrices of spiked population
models. Journal of Multivariate Analysis, 97(6):1382–1408, 2006.
[EKY13]
László Erdős, Antti Knowles, and Horng-Tzer Yau. Averaging fluctuations in resolvents of random band
matrices. Annales Henri Poincaré, 8(14):1837–1926, 2013.
[EYY12a] László Erdős, Horng-Tzer Yau, and Jun Yin. Bulk universality for generalized Wigner matrices. Probability Theory and Related Fields, 154(1):1–67, 2012.
[EYY12b] László Erdős, Horng-Tzer Yau, and Jun Yin. Rigidity of eigenvalues of generalized Wigner matrices.
Advances in Mathematics, 229(3):1435–1515, 2012.
[FJ16]
Zhou Fan and Iain Johnstone. Eigenvalue distributions of variance components estimators in highdimensional random effects models. arXiv preprint 1607.02201, 2016.
[FS10]
Ohad N Feldheim and Sasha Sodin. A universality result for the smallest eigenvalues of certain sample
covariance matrices. Geometric And Functional Analysis, 20(1):88–123, 2010.
[Gem80]
Stuart Geman. A limit theorem for the norm of random matrices. The Annals of Probability, 8(2):252–261,
1980.
[HHN16]
Walid Hachem, Adrien Hardy, and Jamal Najim. Large complex correlated Wishart matrices: Fluctuations and asymptotic independence at the edges. The Annals of Probability, 44(3):2264–2348, 2016.
[Joh01]
Iain M Johnstone. On the distribution of the largest eigenvalue in principal components analysis. The
Annals of Statistics, 29(2):295–327, 2001.
[Kar07]
Noureddine El Karoui. Tracy-Widom limit for the largest eigenvalue of a large class of complex sample
covariance matrices. The Annals of Probability, 35(2):663–714, 2007.
[KY16]
Antti Knowles and Jun Yin. Anisotropic local laws for random matrices. Probability Theory and Related
Fields, pages 1–96, 2016.
[LS15]
Ji Oon Lee and Kevin Schnelli. Edge universality for deformed Wigner matrices. Reviews in Mathematical
Physics, 27(08):1550018, 2015.
[LS16]
Ji Oon Lee and Kevin Schnelli. Tracy-Widom distribution for the largest eigenvalue of real sample covariance matrices with general population. The Annals of Applied Probability, 26(6):3786–3839, 2016.
[LW98]
Michael Lynch and Bruce Walsh. Genetics and analysis of quantitative traits. Sinauer Sunderland, 1998.
[Ma12]
Zongming Ma. Accuracy of the Tracy-Widom limits for the extreme eigenvalues in white Wishart matrices. Bernoulli, 18(1):322–359, 2012.
56
TRACY-WIDOM FOR COVARIANCE MATRICES
[MCM+ 14] Katrina McGuigan, Julie M Collet, Elizabeth A McGraw, H Ye Yixin, Scott L Allen, Stephen F
Chenoweth, and Mark W Blows. The nature and extent of mutational pleiotropy in gene expression
of male drosophila serrata. Genetics, 196(3):911–921, 2014.
[MP67]
Vladimir A Marčenko and Leonid Andreevich Pastur. Distribution of eigenvalues for some sets of random
matrices. Sbornik: Mathematics, 1(4):457–483, 1967.
[Pau07]
Debashis Paul. Asymptotics of sample eigenstructure for a large dimensional spiked covariance model.
Statistica Sinica, 17(4):1617, 2007.
[Péc09]
Sandrine Péché. Universality results for the largest eigenvalues of some sample covariance matrix ensembles. Probability Theory and Related Fields, 143(3-4):481–516, 2009.
[PPR06]
Nick Patterson, Alkes L Price, and David Reich. Population structure and eigenanalysis. PLoS Genetics,
2(12):e190, 2006.
[PY14]
Natesh S Pillai and Jun Yin. Universality of covariance matrices. The Annals of Applied Probability,
24(3):935–1001, 2014.
[SB95]
Jack W Silverstein and Zhidong Bai. On the empirical distribution of eigenvalues of a class of large
dimensional random matrices. Journal of Multivariate Analysis, 54(2):175–192, 1995.
[SC95]
Jack W Silverstein and Sang-Il Choi. Analysis of the limiting spectral distribution of large dimensional
random matrices. Journal of Multivariate Analysis, 54(2):295–309, 1995.
[Sos02]
Alexander Soshnikov. A note on universality of the distribution of the largest eigenvalues in certain
sample covariance matrices. Journal of Statistical Physics, 108(5):1033–1056, 2002.
[TW96]
Craig A Tracy and Harold Widom. On orthogonal and symplectic matrix ensembles. Communications in
Mathematical Physics, 177(3):727–754, 1996.
[YBK88]
Yong-Quan Yin, Zhi-Dong Bai, and Pathak R Krishnaiah. On the limit of the largest eigenvalue of the
large dimensional sample covariance matrix. Probability Theory and Related Fields, 78(4):509–521, 1988.
[Yin86]
Y. Q. Yin. Limiting spectral distribution for a class of random matrices. J. Multivariate Anal., 20(1):50–
68, 1986.
| 10math.ST
|
arXiv:1705.02768v1 [math.RA] 8 May 2017
Typical ranks of semi-tall real 3-tensors
Toshio SUMI∗, Mitsuhiro MIYAZAKI†and Toshio SAKATA‡
Abstract
Let m, n and p be integers with 3 ≤ m ≤ n and (m − 1)(n −
1) + 1 ≤ p ≤ (m − 1)m. We showed in previous papers that if
p ≥ (m − 1)(n − 1) + 2, then typical ranks of p × n × m-tensors
over the real number field are p and p + 1 if and only if there exists
a nonsingular bilinear map Rm × Rn → Rmn−p . We also showed that
the “if” part also valid in the case where p = (m − 1)(n − 1) + 1. In
this paper, we consider the case where p = (m−1)(n−1)+1 and show
that the typical ranks of p × n × m-tensors over the real number field
are p and p + 1 in several cases including the case where there is no
nonsingular bilinear map Rm × Rn → Rmn−p . In particular, we show
that the “only if” part of the above mentioned fact does not valid for
the case p = (m − 1)(n − 1) + 1.
Keywords: tensor rank, typical rank, tall tensor, semi-tall tensor, Bezout’s theorem, determinantal variety
MSC:15A69, 14P10, 14M12, 13C40
1
Introduction
Tensor rank is a subject which is widely studied in both pure and applied
mathematics. A high dimensional array of datum is called a tensor in the
field of data analysis. Precisely, let N1 , . . . , Nd be positive integers. A ddimensional array datum T = (ti1 ···id )1≤ij ≤Nj ,1≤j≤d is called a d-way tensor or
simply a d-tensor of format N1 ×· · ·×Nd . For a set S, the set of N1 ×· · ·×Nd
tensors with entries in S is denoted by S N1 ×···×Nd .
Let K be a field and Vj an Nj -dimensional vector space over K with fixed
basis vj1 , . . . , vjNj for 1 ≤ j ≤ d. Then there is a one
Pto one correspondence
between K N1 ×···×Nd and V1 ⊗ · · · ⊗ Vd by (mi1 ···id ) ↔ i1 ···id mi1 ···id v1i1 ⊗ · · · ⊗
∗
Faculty of Arts and Science, Kyushu University, Fukuoka, Japan
Department of Mathematics, Kyoto University of Education, Kyoto, Japan
‡
Emeritus professor, Kyushu University, Fukuoka, Japan
†
1
vdid . A non-zero tensor corresponding to an element of V1 ⊗ · · · ⊗ Vd of the
form a1 ⊗ · · · ⊗ ad is called a rank 1 tensor. For a general tensor T of format
N1 ×· · ·×Nd , the rank of T , denoted by rankT , is by definition the minimum
integer r such that T can be expressed as a sum of r rank 1 tensors, where we
set the empty sum to be zero. Thus, the rank is a measure of the complexity
of a tensor. Further, for a 2-tensor, i.e., a matrix, the rank is identical with
the one defined in linear algebra.
However, for the case where d ≥ 3, the behavior of rank is much more
complicated than the matrix case. In the matrix case, the rank is the maximum size of non-zero minors. Thus, if K is an infinite field, the set of m × n
matrices with rank min{m, n} form a Zariski dense open subset of K m×n .
However, there are non-empty Euclidean open subsets of R2×2×2 such that
one consists of rank 2 tensors and the other one consists of rank 3 tensors. In
particular, it is not possible to characterize the rank of a tensor by vanishing
and/or non-vanishing of polynomials.
Let m, n, p be positive integers. If the set of rank r tensors of format
p × n × m over R contains a non-empty Euclidean open subset of Rp×n×m ,
we say that r is a typical rank of p × n × m tensors over R. The set of typical
ranks of p × n × m tensors over R is denoted as trankR (p, n, m) or simply
trank(p, n, m).
If the base field is C, the set of tensors of format p × n × m with rank
at most r contains a non-empty Zariski open set if and only if its Zariski
closure is Cp×n×m (cf., Chevalley’s Theorem, see e.g., [Har92, p. 39]). Therefore, there exists exactly one “typical rank of p × n × m tensors over C”.
This is called the generic rank of p × n × m tensors over C and denoted as
grankC (p, n, m) or simply grank(p, n, m).
It is fairly easy to show that grank(p, n, m) = min trank(p, n, m) (see e.g.,
[SSM16, Chapter 6]). Further r ≥ grank(p, n, m) if and only if the r-th higher
secant variety of the image of Segre embedding PC Cp × PC Cn × PC Cm →
PC Cp×n×m is the whole space PC Cp×n×m , where PK V denotes the projective
space consisting of one dimensional subspaces of the K-vector space V . Thus,
mnp
by counting the dimensions, we see that grank(p, n, m) ≥ ⌈ m+n+p−2
⌉.
mnp
Suppose that 3 ≤ m ≤ n ≤ p. Then p ≥ ⌈ m+n+p−2 ⌉ if and only if
p ≥ (m − 1)(n − 1) + 1. Catalisano, Geramita, and Gimigliano [CGG02]
(see also [CGG08]) proved that if (m − 1)(n − 1) + 1 ≤ p ≤ mn, then
grank(p, n, m) = p. Thus, min trank(p, n, m) = p in these cases. ten Berge
[tB00] called a p × n × m-tensor with (m − 1)n < p < mn a tall array or a
tall tensor and proved that trank(p, n, m) = {p} for these cases (see [SSM16,
Chapter 6] for another proof). Here we define a p × n × m-tensor a semi-tall
tensor if (m − 1)(n − 1) + 1 ≤ p ≤ (m − 1)n. We [SSM13, SMS15, SMS17]
2
studied the plurality of typical ranks of semi-tall tensors and proved that if
(m − 1)(n − 1) + 2 ≤ p ≤ (m − 1)n, then trank(p, n, m) = {p, p + 1} if there
exists a nonsingular bilinear map Rm ×Rn → Rmn−p and trank(p, n, m) = {p}
otherwise, where a bilinear map ϕ : V1 ×V2 → W is nonsingular if ϕ(x, y) = 0
implies x = 0 or y = 0.
We also showed in [SMS17] that the former part of the above mentioned
result also valid in the case where p = (m−1)(n−1)+1. Therefore, the latter
part of the above mentioned result in the case where p = (m − 1)(n − 1) + 1 is
left open. In this paper, we treat the case where p = (m − 1)(n − 1) + 1 and
show that trank(p, n, m) = {p, p + 1} in several cases. In particular, we show
that the latter part does not valid in the case where p = (m − 1)(n − 1) + 1.
2
Preliminaries
Let K be a field and T = (tijk ) ∈ K ℓ×m×n . For 1 ≤ k ≤ n, we set Tk =
(tijk ) ∈ K ℓ×m and denote T = (T1 ; · · · ; Tn ). For P ∈ GL(ℓ, K) and Q ∈
GL(m, K), we set P T Q = (P T1 Q; · · · ; P Tn Q). Note rankP T Q = rankT by
the definition of rank.
We first state the definition of the typical rank over R.
Definition 2.1 If the set of rank r tensors over R of format ℓ × m × n
contains a non-empty Euclidean open subset of Rℓ×m×n , then we say r is a
typical rank of ℓ × m × n tensors over R. We denote the set of typical ranks
of ℓ × m × n tensors over R by trankR (ℓ, m, n) or simply trank(ℓ, m, n).
By the definition of the rank, we see the following fact.
Lemma 2.2 Let n1 , n2 and n3 be positive integers.
Then
trank(ni1 , ni2 , ni3 ) = trank(n1 , n2 , n3 ) for any permutation i1 , i2 , i3 of
1, 2, 3.
Definition 2.3 For T = (T1 ; · · · ; Tn ) ∈ K ℓ×m×n , we set
fl1 (T ) := (T1 , . . . , Tn ) ∈ K ℓ×mn
and
and call flattenings of T .
T1
fl2 (T ) := ... ∈ K ℓn×m
Tn
3
By the correspondence K ℓ×m×n ↔ V1 ⊗ V2 ⊗ V3 , where V1 (resp. V2 , V3 ) is a
vector space over K of dimension ℓ (resp. m, n) with fixed basis, flattenings
correspond to natural isomorphisms V1 ⊗ V2 ⊗ V3 → V1 ⊗ (V2 ⊗ V3 ) and
V1 ⊗ V2 ⊗ V3 → (V1 ⊗ V3 ) ⊗ V2 . In particular, rank(fl1 (T )) ≤ rankT and
rank(fl2 (T )) ≤ rankT .
m(1)
Definition 2.4 For M = (m1 , . . . , mn ) = ... ∈ K ℓ×n , we set
m(ℓ)
m(1)
M≤j := (m1 , . . . , mj ), j< M := (mj+1 , . . . , mn ), M ≤i := ... and
m(i)
m(i+1)
i<
M := ... .
m(ℓ)
Definition 2.5 Let R be a commutative ring and M ∈ Rℓ×n . We denote by
It (M) the ideal of R generated by t-minors of R.
3
A condition of an n × p × m-tensor to be of
rank p
From now on, let m, n and p be integers with 3 ≤ m ≤ n and (m − 1)(n −
1) + 1 ≤ p ≤ mn. We set u = mn − p.
Fact 3.1 (1) grank(p, n, m) = p. In particular, min trank(p, n, m) = p
[CGG02].
(2) If p > (m − 1)n, then trank(p, n, m) = {p} [tB00].
(3) Suppose p ≤ (m − 1)n. If there exists a nonsingular bilinear map Rm ×
Rn → Ru , then trank(p, n, m) = {p, p + 1}. Moreover, if (m − 1)(n −
1) + 2 ≤ p, then the converse also hold true [SSM13, SMS15, SMS17].
Therefore, the case where p = (m−1)(n−1)+1 and there is no nonsingular
bilinear map Rm × Rn → Ru is still left open. In the following, we consider
the case where p = (m − 1)(n − 1) + 1 and study if there are plural typical
ranks of p × n × m tensors over R.
Before concentrating on the case where p = (m − 1)(n − 1) + 1, we state
notations and a criterion of an n × p × m tensor to be of rank p in the case
4
where (n − 1)(m − 1) + 1 ≤ p ≤ (m − 1)n. Note that trank(p, n, m) =
trank(n, p, m).
Definition 3.2 We set V := {T ∈ Rn×p×m | fl2 (T )≤p is nonsingular.},
O := {Y ∈ Ru×n×m | p< fl1 (Y ) is nonsingular.}, σ : V → Ru×p , σ(T ) =
(p< fl2 (T ))(fl2 (T )≤p )−1 , ν : O → Ru×p , ν(Y ) = −(p< fl1 (Y ))−1 (fl1 (Y )≤p ),
Ep
and µ : Ru×p → O, µ(W ) = fl−1
τ : Ru×p → V , τ (W ) = fl−1
1 (W, −Eu ).
2
W
Remark 3.3
Rn×p×m ⊃ V
τ տց σ
Ru×p
µ ւր ν
Ru×n×m ⊃ O
and σ(τ (W )) = ν(µ(W )) = W for W ∈ Ru×p .
Definition 3.4 Let x = (x1 , . . . , xm ) be a row vector of indeterminates,
i.e., x1 , . . . , xm are independent indeterminates. For A = (A1 ; · · · ; Am ) ∈
Ru×n×m , we set M(x, A) := x1 A1 + · · · + xm Am ∈ R[x1 , . . . , xm ]u×n .
Definition 3.5 Let a and b be column vectors with entries in R of dimension
m and n respectively. We set ψ(a, b) := (a ⊗Kr b)≤p , where
the
⊗Kr denotes
a1 b
a1
..
..
.
Kronecker product, i.e., if a = . , then ψ(a, b) =
.
am−2 b
am
≤p−(m−2)n
am−1 b
For Y ∈ Ru×n×m , we define U(Y ) to be the vector subspace of Rp generated by {ψ(a, b) | M(a⊤ , Y )b = 0}.
Lemma 3.6 For T ∈ V , rankT = p if and only if dim U(µ(σ(T ))) = p.
Proof Set µ(σ(T )) = W = (W1 ; · · · ; Wm ) and ℓ = u − n. Then by [SMS17,
Theorem 6.5 (1) ⇐⇒ (3)], we see that rankT = p if and only if there are
B = (b1 , . . . , bp ) ∈ Rn×p and p × p diagonal matrices D1 , . . . , Dm such that
Dk = Diag(d1k , . . . , dpk )
for 1 ≤ k ≤ m,
(d
W
+
·
·
·
+
d
W
)b
=
0
for 1 ≤ j ≤ p and
j1 1
jm m j
BD1
(∗)
..
.
is nonsingular.
BD
m−2
B ≤n−ℓ D
m−1
5
First suppose that there are B = (b1 , . . . , bp ) ∈ Rn×p and D1 , . . . , Dm
which satisfy (∗). If we set aj = (dj1 , . . . , djm)⊤ for 1 ≤ j ≤ p, then
M(a⊤
j , W )bj = 0
for 1 ≤ j ≤ p and
BD1
..
.
(ψ(a1 , b1 ), . . . , ψ(ap , bp )) =
BDm−2
B ≤n−ℓ Dm−1
is nonsingular. Therefore dim U(µ(σ(T ))) = p.
Conversely, assume that dim U(µ(σ(T ))) = p. Then there are a1 , . . . ,
ap ∈ Rm and b1 , . . . , bp ∈ Rn such that ψ(a1 , b1 ), . . . , ψ(ap , bp ) are linearly
independent. Set aj = (dj1, . . . , djm)⊤ for 1 ≤ j ≤ p, Dk = Diag(d1k , . . . , dpk )
for 1 ≤ k ≤ m and B = (b1 , . . . , bp ) ∈ Rn×p . Then it is easily verified that
B and D1 , . . . , Dm satisfy (∗).
4
Determinantal varieties and Bezout’s theorem
From now on, we consider the case where p = (m − 1)(n − 1) + 1. Then
u = m + n − 2.
Definition 4.1 We set
O(k−1)×n
∈ Ru×n
En
Ak :=
O(m−k−1)×n
for 1 ≤ k ≤ m − 1 and
Am :=
O(m−1)×(n−1) −e1
En−1
0
∈ Ru×n ,
1
0
where e1 = .. ∈ Rm−1 and A = (A1 ; · · · ; Am ) ∈ Ru×n×m .
.
0
The next fact is the key lemma of this paper.
6
Lemma 4.2 Let y be an indeterminate and a1 , . . . , am−1 ∈ C. Then the
following conditions are equivalent, where Va (I) denotes the affine variety
defined by an ideal I.
(1) (a1 , . . . , am−1 , −1) ∈ Va (In (M(x, A))).
(2) y m−1 − am−1 y m−2 − · · · − a2 y − a1 is a factor of y u + 1.
In order to prove this lemma, we need some preparation. First we make
the following
Definition 4.3 Let y be an indeterminate and {µt }t≥1 an infinite sequence
P
k
of
P complex numbers. We set I ({µt }t≥1 ) = {f (y) ∈ C[y] | f (y) = k ck y ,
k ck µk+t = 0 for any t ≥ 1}.
It is easily verified that I ({µt }t≥1 ) is an ideal of C[y].
Now let a1 , . . . , am−1 ∈ C. Set λt = 0 for 1 ≤ t ≤ m − 2, λm−1 = 1 and
am−1 am−2 · · · a1
..
..
..
−1
.
.
.
..
.. ..
.
.
.
a
1
λm−1+s = det
..
.. .. ..
.
.
.
.
.
.
.
.
.
. am−2
−1 am−1
for s ≥ 1, where the right hand side is an s × s-determinant (some ai ’s may
not appear for small s).
By the first row expansion, we see the following
Lemma 4.4 For t ≥ m, we have λt =
Pm−1
k=1
am−k λt−k .
Set h(y) = y m−1 − am−1 y m−2 − · · · − a2 y − a1 . By the above lemma, we
see that h(y) ∈ I ({λt }t≥1 ). Further, since λt = 0 for 1 ≤ t ≤ m − 2 and
λm−1 = 1, there is no polynomial in I ({λt }t≥1 ) whose degree is less than
m − 1 except the zero polynomial, i.e., I ({λt }t≥1 ) is generated by h(y).
7
Set
a1
a2
.
..
am−1
N = M((a1 , . . . , am−1 , −1), A) =
−1
1
..
.
..
.
..
..
..
.
..
.
.
.
..
..
..
.
..
.
.
.
..
a1
a2
..
.
.
−1 am−1
and for integers c1 , . . . , cn with 1 ≤ c1 < · · · < cn ≤ u, we denote by
[c1 , . . . , cn ]N the maximal minor of N consisting of the c1 -th, . . . , cn -th rows
of N.
Now we state the following
Lemma 4.5 Under the notation above, the following conditions are equivalent.
(1) rankN < n.
(2) [i, m, m + 1, . . . , u]N = 0 for 1 ≤ i ≤ m − 1.
(3) λu+t = 0 for 1 ≤ t ≤ m − 2 and λu+m−1 = −1.
(4) λu+t = −λt for t ≥ 1.
(5) y u + 1 ∈ I ({λt }t≥1 ).
Proof Let
am−1 am−2 · · ·
..
..
−1
.
.
..
..
.
.
U =
.
..
a1
1
..
.
..
.
a1
..
.. ..
.
.
.
.. ..
.
. am−2
−1 am−1
be a u × u matrix. Then det(t<
t< U) = λu+m−1−t + δ0,t for 0 ≤ t ≤ u − 1, where
δ0,t is the Kronecker’s delta.
(1)=⇒(3): Since m−2< U = N and rankN < n by assumption, we see that
det(t<
t< U) = 0 for 0 ≤ t ≤ m − 2. Thus, we see that λu+m−1−t + δ0,t = 0 for
0 ≤ t ≤ m − 2.
8
(3)=⇒(2): We see by the first row expansions of det(t<
t< U) and [t +
1, m, . . . , u] and the assumption that
0 = λu+m−1−t + δ0,t
m−1
X
am−k λu+m−1−t−k + δ0,t
=
=
k=1
m−1
X
am−k λu+m−1−t−k + δ0,t
k=m−t−1
=
t+1
X
at+2−s λu+1−s + δ0,t
s=1
= [t + 1, m, . . . , u]N
for 0 ≤ t ≤ m − 2.
(2)=⇒(1) follows from the fact that the last n − 1 rows of N are linearly
independent, (3) ⇐⇒ (4) follows from the facts that λ1 = · · · = λm−2 = 0
and λm−1 = 1 and Lemma 4.4 and (4) ⇐⇒ (5) follows from the definition of
I ({λt }t≥1 ).
Since (a1 , . . . , am−1 , −1) ∈ Va (In (M(x, A))) if and only if rankN < n and
y + 1 ∈ I ({λt }t≥1 ) if and only if h(y) divides y u + 1, we see Lemma 4.2 by
Lemma 4.5.
Now we recall the following facts about determinantal varieties (see e.g.
[Har92, p.151 and p.243]).
u
Fact 4.6 Let X be a u × n matrix of indeterminates.
Then the projective
u
variety in PC Cu×n defined by In (X) has degree n−1
and codimension u −
n + 1.
u
u
Note that n−1
= m−1
and u − n + 1 = m − 1 since u = m + n − 2. Note
u
also that there are m−1 monic factors of y u + 1 of degree m − 1 in C[y].
In view of this fact, we make the following
Definition 4.7 For B = (B1 ; · · · ; Bm ) ∈ Cu×n×m , we set ϕB : Cm → Cu×n ,
(α1 , . . . , αm ) 7→ α1 B1 + · · · + αm Bm .
Then by Lemma 4.2, Fact 4.6 and Bezout’s theorem, we see the following
Corollary 4.8 Let PC (ImϕA ) be the linear subspace of PC Cu×n defined
by
u
ImϕA . Then PC (ImϕA ) and Vp (In (X)) intersect transversely at m−1 distinct
points, where Vp (I) denotes the projective variety defined by the homogeneous
ideal I.
9
By the implicit function theorem, we see the following fact.
Corollary 4.9 There is a Euclidean open neighborhood U of A in Ru×n×m
such that if B ∈ U , then ϕB is injective and the number of real points of
Vp (In (M(x, B))) ⊂ PC Cm is the number of real monic polynomials of degree
m − 1 which divide y u + 1, where we say a point of a complex projective space
is real if all possible ratios of its homogeneous coordintes are real numbers.
We denote the number of real monic polynomials of degree m − 1 which
divides y u + 1 by α(m, n). Then we see the following
Lemma 4.10
α(m, n) =
u/2
(m−1)/2
(u−1)/2
(m−2)/2
(u−1)/2
(m−1)/2
0
if
if
if
if
m
m
m
m
and n are odd,
is even and n is odd,
is odd and n is even and
and n are even.
By replacing U to a smaller neighborhood if necessary, we may assume
that rank(M(a, B)≤n−1 ) = n − 1 for any B ∈ U and a ∈ Rm \ {0} (cf.
[SMS17, Corollary 4.20]). Then we have the following fact.
Lemma 4.11 Suppose B ∈ U . Then #{[ψ(a, b)] ∈ PR Rp | M(a, B)b =
0, a ∈ Rm \ {0}, b ∈ Rn \ {0}} = α(m, n), where [x] denotes the point of
PR Rp defined by x ∈ Rp \ {0}.
5
Plural typical ranks of some formats of 3tensors
In this section, we show that in certain formats of 3-tensors, there are plural
typical ranks. We use the notation of the previous section.
First we recall the following fact.
Fact 5.1 ([SMS17, Proposition 2.4, Lemma 3.5, Theorems 7.3 and 8.1])
If m − 1 and n − 1 are not bit-disjoint, then trank(n, p, m) = {p, p + 1},
where two positive integers are bit-disjoint if there are no 1’s in the same
place of their binary notation.
Example 5.2 trank(n, p, m) = {p, p + 1} in the following cases.
(1) Both m and n are even.
10
(2) m = 5 and n ≡ 5, 6, 7, 8 (mod 8).
(3) m = 6 and n ≡ 2, 4, 5, 6, 7, 8 (mod 8).
(4) m = 7 and n ≡ 3, 4, 5, 6, 7, 8 (mod 8).
(5) m = 8 and n ≡ 2, 3, 4, 5, 6, 7, 8 (mod 8).
(6) m = 9 and n ≡ 9, 10, 11, 12, 13, 14, 15, 16 (mod 16).
Set A′′ = (A2 ; · · · ; Am−2 ; Am ; −Am−1 ; −A1 ). Then there is a permutation
matrix P ∈ GL(u, R) such that p< fl1 (P A′′ ) = −Eu . Set A′ = P A′′ and
W0 = fl1 (A′ )≤p . Further set ρ : Ru×n×m → Ru×n×m , B = (B1 ; · · · ; Bm ) 7→
P (B2 ; · · · ; Bm−2 ; Bm ; −Bm−1 ; −B1 ) and U = µ−1 (ρ(U )), where µ is the map
defined in Definition 3.2. Then U is an open neighborhood of W0 . Further,
we see the following fact by Lemma 4.11.
Lemma 5.3 If T ∈ σ −1 (U), then #{[ψ(a, b)] ∈ PR Rp | M(a, µ(σ(T )))b =
0, a ∈ Rm \ {0}, b ∈ Rn \ {0}} = α(m, n).
Since σ −1 (U) contains τ (W0 ), σ −1 (U) is not an empty set. Therefore, we see
by Lemma 3.6 that if α(m, n) < p, then there exists a non-empty Euclidean
open subset σ −1 (U) of Rn×p×m consisting of tensors of rank larger than p.
Further, since we see by [SMS17, Theorem 8.1] that typical ranks of n×p×m
tensors are less than or equal to p + 1, we see the following fact.
Lemma 5.4 If α(m, n) < p, then trank(n, p, m) = {p, p + 1}.
Now we state the following
Theorem 5.5 Suppose 3 ≤ m ≤ n and p = (m − 1)(n − 1) + 1. Then
trank(n, p, m) = {p, p + 1} in the following cases.
(1) m = 3 or m = 4.
(2) m = 5 and n ≤ 26 or n = 28.
(3) m = 6 and n ≤ 34.
(4) m = 7 and n ≤ 16.
(5) m = 8 and n ≤ 16.
11
Proof By Lemma 4.10 and computation, we see that α(m, n) < p in the
following cases: (1) m = 3 or m = 4, (2) m = 5 and n ≤ 26 or n = 28,
(3) m = 6 and n ≤ 34, (4) m = 7 and n ≤ 12, (5) m = 8 and n ≤ 14 and
(6) m = 9 and n = 10. Thus, we see the result by Example 5.2 and Lemma
5.4.
Remark 5.6 Set m = 3 and n = 5. Then m − 1 and n − 1 are bit-disjoint.
However, by Theorem 5.5, we see that trank(n, p, m) = {p, p + 1}. Thus the
converse of Fact 5.1 does not valid.
References
[CGG02] M. V. Catalisano, A. V. Geramita, and A. Gimigliano, Ranks of
tensors, secant varieties of Segre varieties and fat points, Linear
Algebra Appl. 355 (2002), 263–285.
[CGG08] M. V. Catalisano, A. V. Geramita, and A. Gimigliano, On the
ideals of secant varieties to certain rational varieties, J. Algebra
319 (2008), no. 5, 1913–1931.
[Har92]
Joe Harris, Algebraic geometry, A first course., Graduate Texts in
Mathematics, vol. 133, Springer-Verlag, New York, 1992,
[SSM16] Toshio Sakata, Toshio Sumi and Mitsuhiro Miyazaki, Algebraic and
Computational Aspects of Real Tensor Ranks, SpringerBriefs in
Statistics, Springer, 2016.
[SMS15] Toshio Sumi, Mitsuhiro Miyazaki and Toshio Sakata, Typical ranks
of m × n × (m − 1)n tensors with 3 ≤ m ≤ n over the real number
field, Linear Multilinear Algebra 63 (2015), no. 5, 940–955.
[SMS17] Toshio Sumi, Mitsuhiro Miyazaki and Toshio Sakata, Typical ranks
for 3-tensors, nonsingular bilinear maps and determinantal ideals,
Journal of Algebra, 471 409–453 (Feb. 2017)
[SSM13] Toshio Sumi, Toshio Sakata and Mitsuhiro Miyazaki, Typical ranks
for m × n × (m − 1)n tensors with m ≤ n, Linear Algebra and its
Application, 438 953–958 (Jan. 2013).
[tB00]
Jos M. F. ten Berge, The typical rank of tall three-way arrays,
Psychometrika 65 (2000), no. 4, 525–532.
12
| 0math.AC
|
N.B. A USB memory stick, storing my draft PhD thesis - comprising material partly used in this work, ”disappeared” from
my hotel room safe in July 2017. For this reason, I’m unnecessarily forced to post this preprint prior publication acceptance.
Golden Angle Modulation
Abstract—Quadrature amplitude modulation (QAM) exhibits
a shaping-loss of πe/6, (≈ 1.53 dB) compared to the AWGN
Shannon capacity. With inspiration gained from special (leaf,
flower petal, and seed) packing arrangements (spiral phyllotaxis)
found among plants, a novel, shape-versatile, circular symmetric,
modulation scheme, the Golden Angle Modulation (GAM) is
introduced. Disc-shaped, and complex Gaussian approximating
bell-shaped, GAM-signal constellations are considered. For bellGAM, a high-rate approximation, and a mutual information
optimization formulation, are developed. Bell-GAM overcomes
the asymptotic shaping-loss seen in QAM, and offers Shannon
capacity approaching performance. Transmitter resource limited
links, such as space probe-to-earth, and mobile-to-basestation,
are cases where GAM could be particularly valuable.
Index Terms—Modulation, golden angle, golden ratio, shaping,
inverse sampling, Shannon capacity, optimization.
2.5
2
1.5
1
0.5
Im(x)
arXiv:1708.05947v1 [cs.IT] 20 Aug 2017
Peter Larsson Student Member, IEEE
0
-0.5
-1
-1.5
-2
-2.5
-2
-1
0
1
2
Re(x)
I. I NTRODUCTION
M
ODULATION formats, of great number and variety,
have been developed and analyzed in the literature.
Examples are PAM, square/rectangular-QAM, phase shift
Keying (PSK), Star-QAM [1], and amplitude-PSK (APSK)
[2]. Square-QAM, hereon referred to as QAM, is the defacto-standard in existing wireless communication systems.
However, at high signal-to-noise-ratio (SNR), QAM is known
to asymptotically exhibit an 1.53 dB SNR-gap (a shapingloss) to the additive white Gaussian noise (AWGN) Shannon
capacity [3]. This is attributed to the square shape, and the
uniform discrete distribution, of the QAM-signal constellation
points. Geometric and probabilistic shaping techniques have
been proposed to mitigate the shaping-gap [3]. An early work
on geometric shaping is nonuniform-QAM in [4]. More recent
works in this direction are, e.g., [5], [6], [7], [8]. Existing
work on modulation schemes have, in our view, not completely
solved the shaping-loss problem, nor offered a modulation
format practically well-suited for the task. This leads us to
examine new modulation formats.
Inspired by the beautiful, and equally captivating,
cylindrical-symmetric packing of scales on a cycad cone, the
spherical-symmetric packing of seeds on a thistle seed head,
or the circular-symmetric packing of sunflower seeds, we have
recognized that this shape-versatile spiral-phyllotaxis packing
principle, found among plants, is applicable to modulation
signal constellation design. Based on the spiral phyllotaxis
packing, the key contribution of this letter is proposing a
novel, shape-versatile, high-performance modulation framework - the Golden angle modulation (GAM). We consider
discrete modulation, with equiprobable constellation points,
that approximate a r.v. with continuous complex Gaussian
The author is with the ACCESS Linnaeus Center and the School of
Electrical Engineering at KTH Royal Institute of Technology, SE-100 44
Stockholm, Sweden. E-mail: [email protected].
Figure 1: Geometric-bell-GAM signal constellation, N = 210 .
(bell)-shaped distribution, as well as a baseline case, with a
disc-shaped distribution. We find, as expected, that the MIasymptote of geometric-bell GAM (GB-GAM), for increasing
number of constellation points, coincide with the AWGN
Shannon capacity. This also supports the complex Gaussian
communication signal assumption, used in many performance
analysis works, for the low-to-high SNR-range.
II. G OLDEN A NGLE M ODULATION
The core design of GAM is given below.
Definition 2.1: (Golden angle modulation) The nth constellation point of GAM has the probability of excitation pn , and
the complex amplitude
xn = rn ei2πϕn , n ∈ {1, 2, . . . , N },
(1)
where rn is the radius of constellation point
n, 2πϕ denotes
√
the golden angle in rads, and ϕ = 1 − ( 5 − 1)/2.
We will assume that rn+1 > rn for an increasing spiral winding. For the probability, it may be equiprobable, pn = 1/N ,
or dependent on index n. The later, where pn+1 ≤ pn , corresponds to GAM with probabilistic shaping and is explored
extensively in [9]. Hence, a constellation point, is located
ϕ ≈ 0.381 turns (137.5◦) relative√to the previous constellation
point. Replacing ϕ, with (1 + 5)/2 ≈ 1.618, the golden
ratio, or its fractional part, gives an equivalent spiral winding,
but, in the opposite direction. Note that phase rotation value
deviating with approximately 1% from the golden angle (ratio)
destroys the locally uniform packing. The mathematical design
of the phase rotation in Def. 2.1 is inspired from the work
by Vogel, [10], who described √
an idealized growth pattern
for the sunflower seeds, xn = nei2πϕn (in our notation).
Vogel did however not consider modulation. More importantly,
This work has been submitted to the IEEE for possible publication.
Copyright may be transferred without notice, after which this version may no longer be accessible.
2
a key insight here, enabling the approximation of a complex
√
Gaussian pdf, is to not restrict the radial function rn to n
as in [10]. Allowing for arbitrary radial growth of rn , gives
the geometric shaping capability, and allowing for arbitrary
probabilities pn , gives the dimension of probabilistic-shaping.
GAM features the following advantages:
• Natural constellation point indexing: In contrast to QAM,
APSK, and other, without any natural index order, GAM
enables a unique indexing based on signal phase, 2πϕn,
or magnitude, rn , alone.
• Practically near-ideal circular design: A circular design
can offer enhanced MI-, distance-, symbol error rate- and
PAPR- performance over a square-QAM design.
• Shape-flexibility in radially distribution of constellation
points while retaining an evenly distributed packing: We
recognize this as a central feature of GAM which allows
approximation of (practically) any radial-symmetric pdfs.
• Naturally lends itself for circular-symmetric probabilistic
shaping: We recognize this as a central feature of GAM
which allows approximation of (practically) any radialsymmetric pdfs.
• Any number of constellation points, while retaining the
overall circular shape: This gives full flexibility, e.g., in
alphabet size of a channel coder, or a probabilistic shaper.
• Rotation (and gain) invariant: The uniquely identifiable
gain and rotation of signal constellation could, e.g., allow
for blind channel estimation.
Some comments. First, the index range is not necessarily
limited to n ∈ {1, 2, . . . , N }. Second, GAM has a complex
valued DC component. Possible remedies, if a problem, is to
subtract the DC component, or negate every second symbol.
Third, while hexagonal packing is the densest 2D-packing (as
desirable in the high SNR-range), it does not share the above
listed features of GAM.
A. Disc-GAM
We first introduce disc-GAM, i.e. GAM without shaping,
below. Besides its own merits, this also serves as a base-line
to shaping. If the expression for rn is altered (fixed), but pn
is fixed (altered), geometric (probabilistic) shaping result.
Definition 2.2: (Disc-GAM) Let pn = 1/N , N be the
number of constellation points, and P̄ be the average power
constraint. Then, the complex amplitude of the nth constellation point is
√
rn = cdisc n, n ∈ {1, 2, . . . , N },
(2)
s
2P̄
.
(3)
cdisc ,
N +1
In the above, (3)
from P
the power normalization
P is found
N
1
2
2
2 N +1
condition, P̄ = N
n=1 pn rn = cdisc
n=1 N n = cdisc 2 .
We illustrate the disc-GAM constellation in Fig. 2, where
its approximate disc-shape is clearly depicted.
A few remarks about disc-GAM. The entropy is simply
Hdisc = log2 N . The PAPR is P AP Rdisc = 2P̄ N/(N +
1)P̄ ≃ 2 (≃ 3dB) when N → ∞. From PAPR-pointof-view, this makes disc-GAM favorable over QAM, since
1.5
1
0.5
0
-0.5
-1
-1.5
-1.5
-1
-0.5
0
0.5
1
1.5
Figure 2: Disc-GAM signal constellation, N = 210 .
P AP RQAM = 4.8 dB. (PSK has P AP RPSK = 0 dB,
but with very poor MI-performance for N → ∞). Letting
N → ∞, QAM asymptotically requires 10 log2 (π/3) (≈ 0.2
dB) higher average power than disc-GAM for the same average
constellation point distances . Also when N → ∞, QAM
asymptotically requires 10 log2 (π/2) (≈ 1.96 dB) higher peak
power than disc-GAM for the same average distance between
(uniformly packed) constellation points.
B. Geometric bell-GAM
Next, two geometric shaped GAM schemes are introduced.
1) High-rate approach: This first GB-GAM design builds
on the inverse sampling method (a high-rate (HR) approximation), and approximates a complex Gaussian distributed r.v.
Theorem 2.1: (Geometric-bell-GAM (HR)) Let pn = 1/N ,
N be the number of constellation points, and P̄ be the average
power constraint. Then, the complex amplitude of the nth
constellation point is
s
N
, n ∈ {0, 1, . . . , N − 1},
rn = cgb ln
N −n
s
N P̄
cgb ,
.
(4)
N ln N − ln(N !)
Proof: The proof is given in Appendix A.
We illustrate the geometric-bell-GAM signal constellation
in Fig. III, and note that it is densest at its center, i.e. where
the pdf for the complex Gaussian r.v. peaks.
We note the following characteristics.
When N → ∞,
−1
N
since
lim
N
ln
N
/N
!
=
1,
we get rn ≃
N →∞
p
P̄ ln (N/(N − n)). The entropy is Hgb = log2 N . The
PAPR is P AP Rgb = c2gb ln(N ) = P̄ /(1 − ln (N !)/N ln N ) ≃
P̄ ln (N ), which tends to infinity with N . This is expected as
the PAPR of a complex Gaussian r.v. is infinite. Note here that
the index range in (4) can not be n ∈ {1, 2, . . . , N }, but is
chosen as n ∈ {0, 1, . . . , N − 1}, as rN = ∞ otherwise.
3
1.53 dB
C=log 2(1+SNR)
MI and Shannon capacity [b/Hz/s]
10
N=212
MI GB-GAM (IS)
MI asymptote QAM
11
MI and Shannon capacity [b/Hz/s]
12
N=210
10
9
N=28
8
7
N=26
6
5
Formulation-G1
N=24
4
3
9.5
9
8.5
MI GB-GAM (HR), N=210
8
MI Disc-GAM, N=210
MI QAM, N=210
C=log2(1+S)
7.5
2
5
10
15
20
25
30
35
40
22
24
26
SNR [dB]
Figure 3: MI of GB-GAM (HR), asymptote of QAM, and
Shannon capacity. MI of GB-GAM (G1) with N = 24 .
2) MI-optimization of GB-GAM: Formulation-G1: In this
method, we let pn = 1/N , and vary rn in order to maximize
the MI, for a desired SNR S. The optimized signal constellation points are x∗n = rn∗ ei2πϕn . More formally, allowing for a
complex valued output r.v. Y , and a complex valued (discrete
modulation) input r.v. X, the optimization problem is
maximize
I(Y ; X),
subject to
rn+1 ≥ rn , n = {1, 2, . . . , N },
rn
r1 ≥ 0,
N
X
n=1
pn rn2
σ2
28
30
32
34
SNR [dB]
(5)
= S.
Remark 2.1: For some applications,
PAPR is of interest.
Pthe
N −1
2
A PAPR-inequality constraint, rN
/ n=1 pn rn2 ≤ P AP R0 ,
P AP R0 being the target PAPR, can be amended to the optimization problem. Other constraints may also be of interest.
When optimizing GAM in AWGN, the MI is I(Y ; X) =
2
h(Y
R ) − h(W ), where h(W ) = log2 (πeσ ), h(Y ) =
− C fY log2 (fY ) dy, integrating over the complex domain,
|y−xn |2
PN
PN
with fY = n=1 pn f (y|xn ) = πσ1 2 n=1 pn e− σ2 .
Due to the non-linearities in the MI, the optimization problem in formulation-G1 is hard to solve analytically. Hence, a
numerical optimization solver is used in Section III.
III. N UMERICAL R ESULTS
Here, we now examine the MI-performance of GAM (with
its irregular cell-shapes and varying cell-sizes) by using a
Monte-Carlo simulation approach. For the optimization in
formulation-G1, MATLAB’s fmincon function is used together
with numerical integration for the MI.
In Fig. 3, we illustrate the MI performance for GB-GAM
(HR) together with the Shannon capacity. As expected, for
larger constellation size N , a greater overlap with the Shannon
capacity is seen. The MI approximation is good up to about
≈ 2H/3. Naturally, the MI is limited by the entropy of the
signal constellation. We observe an intermediate SNR region,
Figure 4: MI of disc-GAM, GB-GAM (HR), QAM, with N =
210 , and Shannon capacity.
SNR
≈ 4.8 dB
≈ 11.8 dB
15 dB
AWGN capacity
2
4
≈ 5.03
HR
1.921
3.440
3.828
G1
1.961
3.549
3.926
Table I: MI in [b/Hz/s] for HR- and G1-schemes with N = 16.
a region where the MI does neither reach the channel capacity,
nor the entropy of the signal constellation. It is in this SNRregion that further constellation optimization, i.e. formulationG1 , is of interest. The MI for GB-GAM (G1) is also shown,
but due to optimization complexity only, for N = 16.
In Tab. I, the MI of GB-GAM with the HR-, and G1-,
formulations are given. As expected, the optimized scheme,
G1, perform better than HR. While the MI-improvements are
modest, the optimization formulation is substantiated. For G1,
when the MI is as large as the constellation entropy, we
have observed that the signal constellation approaches the
disc-GAM solution, whereas in the low-MI region, we have
noted that the optimized signal constellation approaches the
HR GB-GAM solution. It is further observed that the extreme
magnitudes of the highest constellation indices for GB-GAM
(HR), as seen in Fig. , are attenuated with the MI-optimization
and leads to improved PAPR performance.
In Fig. 4, we compare disc-GAM, GB-GAM (HR), and
QAM, together with the Shannon capacity when the MI is
nearly as large as the constellation entropy. We note the
(expected) ≈ 1.53 dB SNR-gap for QAM to the Shannoncapacity. The (expected) SNR-gap from QAM to disc-GAM
is ≈ 0.2 dB. When the MI is approximately as large as
the constellation entropy, disc-GAM, and QAM, perform
slightly better than GB-GAM (HR) scheme. This observation
prompted us to develop the MI-optimization formulation-G1.
IV. S UMMARY
AND
C ONCLUSIONS
In this letter, we have introduced a new modulation
format, the golden angle modulation. With geometric- (or
probabilistic-) shaping, GAM can approximate virtually any
circular-symmetric pdf. We studied geometrically-shaped
4
GAM to approximate the pdf of a continuous complex
Gaussian r.v. A high-rate solution, was developed. We also
introduced the notion of MI-optimized GAM under an average SNR-constraint, and optionally also a PAPR-inequalityconstraint. The MI-performance was observed to asymptotically approach the Shannon capacity as the number of signal
constellation points tended to infinity. In contrast to QAM’s
1.53 dB shaping-loss, GAM exhibit no asymptotic loss. The
complex Gaussian communication signal model assumption,
as often used for performance analysis, was substantiated from
a practical modulation point-of-view.
We believe that GAM may find many applications in
transmitter-resource-limited links, such as space probe-toearth, satellite-to-earth, or mobile-to-basestation communication. This is so since high data-rate is desirable from the
power-, energy-, and complexity-limited transmitter side, but
higher decoding complexity is acceptable at the receiver side.
Certain cases may also benefit from using the PAPR-inequality
constrained optimization. With bell-GAM, the power reduction
in the high-MI regime could be up to 30% (1 − 1/101.53/10 ≈
0.3), which is of environmental interest. Moreover, cellularsystem operators, could potentially also reduce energy consumption (and cost) with up to 30%. It is hoped that, GAM,
with its attractive characteristics and performance, could be of
interest for most, wireless, optical, and wired, communication
systems.
A PPENDIX
A. Proof in Theorem 2.1
Proof: The distribution in phase, for the constellation
points, is already given by the e2πϕn -factor. However, the
radial distribution need to be determined. A complex Gaussian
r.v. with variance σ 2 is initially considered. The inverse
sampling method, assumes a uniform continuous pdf on (0, 1).
We modify the inverse sampling method and use a discrete
uniform pmf at steps n/N, n ∈ {0, 1, . . . , N − 1}, where N
is assumed large (for the high-rate approximation).
Z rn − r22
Z rn
e σ
2πr dr
fR (r)dr =
F (rn ) =
πσ 2
0
0
2
rn
= 1 − e− σ 2 .
Setting F (rn ) = n/N , and solving for rn , yields
s
N
.
rn = σ ln
N −n
Thus, the general
for the signal constellation has the
r solution
form rn = cgb ln NN−n . The constant cgb is given by the
average power constraint as follows,
⇒ cgb
N
−1
X
−1
c2gb NX
N
,
N n=0
N −n
n=0
s
v
u
N
P̄
N P̄
u
=
= tP
.
N −1
N ln N − ln N !
ln N
P̄ =
pn rn2 =
n=0
N −n
ln
R EFERENCES
[1] L. L. Hanzo, S. X. Ng, T. Keller, and W. Webb, Star QAM Schemes for
Rayleigh fading channels. Wiley-IEEE Press, 2004, pp. 307–335.
[2] C. Thomas, M. Weidner, and S. Durrani, “Digital amplitude-phase
keying with M-ary alphabets,” IEEE Transactions on Communications,
vol. 22, no. 2, pp. 168–180, February 1974.
[3] G. D. Forney and G. Ungerboeck, “Modulation and coding for linear
Gaussian channels,” IEEE Transactions on Information Theory, vol. 44,
no. 6, pp. 2384–2415, Oct 1998.
[4] W. Betts, A. R. Calderbank, and R. Laroia, “Performance of nonuniform constellations on the Gaussian channel,” IEEE Transactions on
Information Theory, vol. 40, no. 5, pp. 1633–1638, Sep 1994.
[5] L. Szczecinski, S. Aissa, C. Gonzalez, and M. Bacic, “Exact evaluation of bit- and symbol-error rates for arbitrary 2-D modulation
and nonuniform signaling in AWGN channel,” IEEE Transactions on
Communications, vol. 54, no. 6, pp. 1049–1056, June 2006.
[6] Z. Mheich, P. Duhamel, L. Szczecinski, and M. L. A. Morel, “Constellation shaping for broadcast channels in practical situations,” in 2011
19th European Signal Processing Conference, Aug 2011, pp. 96–100.
[7] X. Xiang and M. C. Valenti, “Closing the gap to the capacity of APSK:
Constellation shaping and degree distributions,” in 2013 International
Conference on Computing, Networking and Communications (ICNC),
Jan 2013, pp. 691–695.
[8] F. Buchali, W. Idler, L. Schmalen, and Q. Hu, “Flexible optical transmission close to the Shannon limit by probabilistically shaped QAM,”
in Optical Fiber Communications Conference and Exhibition (OFC)’17,
March 2017, pp. 1–3.
[9] P. Larsson, “Golden angle modulation: Geometric- and probabilisticshaping approaches,” To be Submitted, 2017.
[10] H. Vogel, “A better way to construct the sunflower head,” Mathematical
Biosciences, vol. 44, no. 3-4, pp. 179–189, Jun. 1979.
| 7cs.IT
|
Book review
arXiv:cs/0512096v2 [cs.PL] 22 Jun 2006
The Haskell Road to Logic, Maths and Programming, Kees Doets and Jan van
Eijck, Texts in Computing Series, King’s College Publications, London, 2004.
Price: £14.00 pounds, $ 25.00, 444 pages, Paperback, ISBN: 0-9543006-9-6.
Reviewer: Ralf Lämmel
Microsoft Corp., Redmond, U.S.A.
1
Preamble
In university curricula, the subjects of programming and maths and logic tend to
be separated. For instance, in the typical computer-science curriculum, mathematics and logic are taught according to the (dry) mathematical tradition,
perhaps with modern concessions to help the less abstract minds. One sort of
concession is to illustrate the implementability of selected mathematical concepts in some programming language, be it C, Fortran, or others. Another sort
of concession is to hint at the usefulness of systems like Maple or Mathematica — systems that are probably more appreciated by mathematicians than
applied, programming-minded computer scientists. By contrast, programming
courses are almost certainly focused on a specific programming paradigm and
on a specific language. Logic and maths may show up in such courses as a
sort of “application domain”, while competing with many other domains. For
instance, a typical course on the programming language Haskell is likely to emphasize Haskell’s strengths in the areas of abstract-data type specification and
combinator libraries for domain-specific languages.
The authors of the book at hand have succeeded in amalgamating the themes
of maths, logic and Haskell programming in a mutually beneficial manner. The
book really stands out in so far that it leverages Haskell in diving deep into
maths and logic (as deep as covering power series or co-inductive proofs). The
use of Haskell makes the mathematical concepts quite digestible; Haskell lends
itself so well as a strongly typed and executable modeling language. The authors’
style is quite relaxing. The mathematical concepts are explained in a way that
carefully helps with an intuitive understanding.
The “Haskell road to Logic, Maths and Programming” by Kees Doets and
Jan van Eijck does not (mean to) comprise a detailed and comprehensive introduction to functional programming or the programming language Haskell. There
exist such introductions to Haskell; cf. the excellent textbooks by Thompson
and Hudak [14, 6], and these introductions should still be useful for readers who
want to deeply appreciate Haskell as a general-purpose programming language
and Haskell’s strengths other than maths and logic. The book at hand does
not assume prior knowledge of Haskell, though; all Haskell concepts are introduced incrementally. The combination of texts on maths/logic with computer
programming is not unprecedented. For instance, Barwise and Etchemendy’s
“Language, Proof, and Logic” [2] accompanies a text on logic with several computer programs. Even more to the point, Hall and O’Donnell’s “Discrete math1
ematics using a computer” [5] covers maths and logic in similar ways than the
book at hand, and it also facilitates Haskell.
2
Chapter by chapter
Preface
This is an outstanding book, but the preface could be improved. In particular,
the preface seems to be somewhat ambiguous about the purpose and the subject
of the book. Citation: “The purpose of the book is to teach logic and mathematical reasoning in practice and to connect logical reasoning with computer
programming.” The mentioning of “practice” may count as confusing. Elsewhere in the preface: “The subject of the book is the use of logic in practice,
more in particular the use of logic in reasoning about programming tasks.” It
is not immediately obvious to the reader what sorts of “programming tasks”,
what sorts of “reasoning” are meant here. (For instance, one may think that
program validation and verification is included, but this turns out not to be the
case. Likewise, one may expect reasoning about complexity.) It would help if
the following questions were answered more clearly: What are the mathematical
fields covered by the book? What sorts of reasoning about either mathematical
notions or programs or both are carried out?
According to the preface, Haskell is a member of the LISP family with the
lambda calculus as foundation. This characterization does not hint at two of
Haskell’s strengths: its advanced type system and its lazy semantics. (LISP,
in particular, does not share these assets.) Both strengths are exploited in the
book, and this is worth emphasizing. The choice of Haskell is less arbitrary than
it may seem from the preface.
The outline of the chapters could be more helpful. Regarding chapters 1–7,
there is basically one sentence. The explanations for the remaining chapters may
overwhelm newcomers with technical terms (cf. Chapter 9: “closed forms for
polynomial sequences”). It would be wonderful to read a preface that provides
an accessible overview of the chapters, including a story how these subjects fit
together, making brief links to the three components of the book’s title: logic,
maths, programming.
Chapter 1: Getting started
The text starts with some basic explanations of literate programming, lambda
calculus and typed functional programming. Then, the text picks up the problem of prime-number testing for a first illustration of the “Haskell Road to
Maths”. The text is pleasant to read, and it introduces a good deal of mathematical conventions and Haskell idioms including primitive types, products and
lists. The text also states clearly an important goal of the book early on: to
communicate skills for the formulation of properties about programs and their
proofs. The textbook style of the book also settles immediately: a very good
2
mix of definitions, explanations, examples and exercises.
Eventually, the chapter describes prime factorization and prime-number generation. Surprisingly, the authoritative definition of the former problem is given
in an imperative language notation with assignments! The Haskell implementation of the latter puts the higher-order combinators map and filter to work.
This is an indication that the first chapter has come to real speed. Some bits of
the Haskell exposition in this chapter lack context in the sense that idioms and
examples do not contribute to the running theme of algorithms around prime
numbers.
The chapter pleasantly closes by emphasizing a major strength of Haskell:
this pure, declarative, functional language admits equational reasoning, which is
out of reach for imperative languages like Java due to destructive assignments
and other side effects in these languages. Like most chapters in the book,
Chapter 1 carries a short tail “Further Reading”.
Chapter 2: Talking about Mathematical Objects
The chapter describes the basics of logic, in particular the usual connectives
(not, if, and, or, iff, for all, there exists) complete with Haskell encodings of connectives, formulas, and validity checks for formulas. The text is very pleasant
to read; the alternation between mathematical notation and Haskell encoding
really makes sense. The text carefully explains the many aspects of quantification: free variables, open and closed formulas, domains for quantification. The
text makes a laudable effort to help the beginning logician. For instance, the
text identifies “bad habits” regarding the ambiguous use of quantifiers in textual
formulations of mathematical statements.
The text makes clear that Platonism is adopted in this book: every mathematical statement is either true or false. Other forms of logic, such as intuitionistic logic or fuzzy logic are not covered.
A side note on Haskell’s type classes
Throughout the book, there are a few occasions where Haskell’s type classes
could show off nicely. For instance, when checks for validity or logical equivalence are provided (cf. Chapter 2), they are coded separately for several arities,
as indicated by the following examples:
valid1 :: (Bool -> Bool) -> Bool
valid1 bf = (bf True)
&& (bf False)
3
valid2 :: (Bool -> Bool -> Bool)
valid2 bf = (bf True True)
&& (bf True False)
&& (bf False True)
&& (bf False False)
-> Bool
valid3 :: (Bool -> Bool -> Bool -> Bool) -> Bool
-- ... and so on ...
A single, generic valid function would suffice for all arities, as defined below.
The same holds for logical equivalence. (It is left as an exercise for the reader
to eliminate the minor violation of strict Haskell 98 rules.)
class Valid f
where
valid :: f -> Bool
instance Valid Bool
where
valid = id
instance Valid f => Valid (Bool -> f)
where
valid f = valid (f True)
&& valid (f False)
It is understandable that the text does not want to overwhelm the reader
in the early chapters. However, the power of Haskell’s type classes could have
been discussed in a later chapter. Fortunately, the reader will still hear about
this important part of Haskell because a few chapters take advantage of some
type classes from the Haskell library.
Chapter 3: The Use of Logic: Proof
The chapter communicates style recommendations and recipes for simple proofs.
This is done without much reference to Haskell or Haskell examples. The idea
seems to be that the chapter provides the foundation for carrying out mathematical proofs and proofs about program properties in all subsequent chapters.
A typical style recommendation is “Make sure the reader knows exactly what
you are up to.” The various recipes basically cover all logical connectives —
their introduction and elimination. And then again, even the use of the recipes
is complemented by further recommendations such as not to consider proof by
contradiction all too easily.
The authors deserve extra credit for their clear and detailed enumeration and
rationalization of the proof recipes, complete with careful coverage of style issues. This textbook may really enable readers to acquire proving skills. Some of
4
the examples may be potentially too demanding for the beginning mathematician. For instance, the first example for ∃-elimination refers to transcendent
reals and their definition as not being roots of a polynomial with integer coefficients. (Polynomials and number theory are only covered much later in the
book.)
The text readily admits that programming languages like Haskell are of
little help with proving general properties about programs. One may use the
computer though to refute single cases. Perhaps, this is too much to ask for,
but two topics could have been touched upon in the context of this chapter:
• The Curry-Howard isomorphism, i.e., the correspondence between types
and logical formulas, as much as expressions and logical proofs; see [13]
for a functional-programming-minded discussion of this correspondence.
• The use of theorem provers to check or even to find proofs.
Chapter 4: Sets, Types and Lists
The chapter, according to the authors, presents ample opportunity to exercise
(perhaps better: acquire) the reader’s skills in writing implementations for set
operations and for proving things about sets. This is perhaps a bit misleading.
The text does certainly not discuss different (efficient) implementations of sets or
other data types. In reality, the chapter focuses on the axiomatization of sets as
well as pairs, products and lists. The text also explains Russell’s paradox and the
provisions by the Haskell type system that help avoiding such problems in actual
Haskell programs. (The text also touches upon the relation between the Halting
problem and Russell’s paradox.) Eventually, several list- and set-processing
primitives are encoded. These implementations facilitate the type classes Eq
(“equality”) and Ord (“ordering”) from the Haskell prelude. Afterwards, the
chapter spends some time on Haskell’s special idioms for list comprehensions,
which are clearly related to the mathematical notion of set comprehensions. List
comprehensions are illustrated with a detailed example in the realm of relational
database queries.
Regarding “Further Reading”, some more references spring to mind. The
reader could appreciate pointers to existing functional-programming-minded or
Haskell-specific work on the efficient implementation of data types [11, 12]. The
chapter’s major example for list comprehensions, a database query scenario, is
weakly typed: the type String is used as the universal data type for all cells
in the database instance and all column names. Haskell can do much better
using more sophisticated type-system features. Admittedly, substantially more
strongly typed approaches require non-Haskell 98 features [8, 7]. It is an asset
of the book at hand that it stays with the Haskell 98 standard.
Chapter 5: Relations
The chapter describes the theory of relations (“sets of ordered pairs”); it defines the common properties of relations and other related terms (reflexivity,
5
transitivity, equivalence relations, quotients, characteristic functions, etc.). The
properties are also implemented as checks that can be effectively computed for
relations that are given as finite sets of pairs. The chapter focuses on elementary
material about relations. N-ary relations and the corresponding relational algebra are not covered. The chapter offers a huge amount of simple but insightful
exercises with which the reader can develop skills in mathematical proofs. These
exercises are particularly manageable since they assume mostly basic properties
of sets and the chapter’s simple definitions for relations.
Chapter 6: Functions
The chapter is about functions as mathematical objects, not to be confused with
Haskell functions. There is clearly a plethora of important mathematical notions
around functions: function definition by case discrimination, the relational view
on functions, surjections, injections, bijections, inverse, partiality, congruences,
and so on.
As in previous chapters, the text captures properties of mathematical objects
as Haskell expressions that perform “checks”. That is, a certain property of a
function, such as injectivity, is checked under the assumption that the domain
of the function is given as a (finite) list:
injective :: Eq b => (a -> b) -> [a] -> Bool
injective f [] = True
injective f (x:xs) =
notElem (f x) (image f xs) && injective f xs
One could take this idea a tiny step further, by writing down QuickCheck [4]
properties in Haskell. The QuickCheck tool would automatically check all
these properties. For instance, rather than stating the mere expectation that
fromEnum should be the left-inverse of toEnum (cf. page 222), one could instead
formulate QuickCheck properties for different instances of Enum (i.e., for different enumeration types). Here is a sample property for the Char type:
prop_EnumChar = forAll charInts $ \x ->
fromEnum ((toEnum x) :: Char) == x
charInts :: Gen Int
-- restrict Int generator to valid character codes
Chapter 7: Induction and Recursion
The chapter starts with the basic proof method for mathematical induction.
The method is explained in all detail and with great care. In particular, the
text clarifies that induction proofs are particularly straightforward whenever
the function “under study” is defined by (primitive) recursion on the natural
numbers. Induction and recursion is then also demonstrated for trees and lists.
The text clearly hints at the full generality of the concepts: recursion schemes
6
can be advised for a large class of data types, and corresponding induction proof
schemes are thereby enabled.
The “Further Reading” section would be more complete with some references
to seminal work on recursion schemes, e.g., [9]. There is also a wonderful tutorial
by Augusteijn, which applies morphisms (such as primitive recursion) to the
rigorous design of sorting algorithms and their encoding in Haskell [1].
Chapter 8: Working with Numbers
The chapter presents the different number systems that exist: natural numbers,
integers, rational numbers, irrational numbers, and complex numbers. By developing these number systems for Haskell, the text also provides deeper insight
into Haskell’s built-in and library support for numbers.
As usual, natural numbers are shown to suffice as basis for integers and rational numbers. In fact, different representations for these two derived number systems are discussed and implemented in Haskell. Irrational numbers are defined
informally (as infinite, non-periodic decimal expansions). Implementation-wise,
the Haskell primitive type for reals is used. Key notions for reasoning about
reals are defined: continuous functions, limits, Cauchy sequences.
Throughout the chapter, it is shown how the different number systems participate in standard type classes for equality, ordering and arithmetic operations.
Standard algebraic properties of the operations on the numbers (associativity
etc.) are routinely proved. For several of the number representations, geometric
interpretations are visualized. These illustrations should be of great help for the
reader.
Chapter 9: Polynomials
The chapter discusses polynomials with the initial goal to automate the search
for polynomials as closed forms for sequences of integers. To this end, Babbage’s
classic difference method is explained and implemented in Haskell. (Recall that
the method works as follows: suppose we are given a “long enough” sequence
of integers a0 , . . . , an , for which we presume that it may be computed from a
polynomial f (x) such that a0 = f (0), . . . , an = f (n), then we may attempt
to find the closed form f (x) in two steps. (i) Determine the degree of f (x)
by difference analysis on the given sequence of integers. (ii) Perform Gaussian
elimination to determine the coefficients of f (x).)
The authors succeed in capturing the beauty of this method, making clear
that no magic is involved. Again, the combination of mathematical notation and
Haskell encoding helps enormously to get all the fine details across. This chapter
is a master piece: a concise but still comprehensive and insightful explanation
of the method. All of Babbage, complete with the Haskell encoding on just 13
pages!
The chapter also discusses the link between polynomials and combinatorics.
In this context, Newton’s binomial theorem is derived and put to work. These
7
parts of the chapter are slightly less clear in stating up-front what they aim to
achieve.
We recall the simple version of Newton’s binomial theorem:
n
X
n k
(z + 1) =
(
)z
k
n
k=0
The left-hand side is not in the form of a polynomial (it uses a binomial — a
sum of two terms), whereas the right-hand side is in the form of a polynomial.
Hence, one way to read Newton’s theorem is that it enables binomial expansions
such that pure polynomials are derived.
It is straightforward to prove Newton’s theorem (by induction). The text
however conveys the beauty of Newton’s findings by deriving the theorem through
operations on polynomials: substitution and calculus. Again, the text communicates deep insight combined with a Haskell model.
The final part of the chapter discusses the representation of polynomials as
lists of their coefficients. Polynomial arithmetics and calculus is then derived
on top of this representation. The main goal is to reason about combinatorial
problems. In particular, polynomials (as lists of coefficients) are derived as
solutions to combinatorial problems. For instance, the polynomial expansion of
(z + 1)10 solves the problem of picking k elements from a set of 10.
Chapter 10: Corecursion
The chapter discusses infinite data structures (mostly streams). Haskell laziness’s makes this language clearly suited to study such data structures even
programmatically. One important goal of the chapter is to describe proof methods for corecursive functionality. The text also makes excursions to the following
two related subjects. Non-deterministic processes are modeled as functions from
random integer streams (for coding “decisions”) to streams (for coding the actions performed). The earlier discussion of the link between polynomials and
combinatorics is generalized to power series (so to say polynomials with infinite series of coefficients). Power series naturally show up once polynomials are
divided; the results of division may have an infinite degree.
The authors make the point (in the introduction) that they deliver the first,
general textbook treatment of recursion. This already indicates the challenge
of this section. Some readers may expect from the chapter to become wellversed in coinduction proofs. Unfortunately, the text is scarce in this respect.
There are essentially just 5 pages for the recipe of coinduction proofs. There
is essentially just one example of a coinduction proof for streams, and one may
argue whether or not the proof is detailed enough. The “Further reading”
section should perhaps refer to more scenarios for coinductive reasoning, and
more sample proofs. In particular, Barwise and Moss’ book on “Vicious circles”
[3] springs to mind.
8
Chapter 11: Finite and Infinite Sets
The chapter can be seen as a more advanced continuation of Chapter 4: Sets,
Types and Lists. The text begins with a deeper analysis of mathematical induction. Many additional insights are provided, when compared to the intuitive
use of induction proofs in earlier chapters. For instance, the notion of strong
induction is presented, which does not require a base step and uses a stronger induction hypothesis, thereby enabling proofs of more involved problems. Another
subject of the chapter is equipollence, preparing ultimately for the discussion
of cardinalities of sets including infinite cardinalities such as ℵ0 (i.e., the cardi0
nality of the natural numbers) and 2ℵ (i.e., the cardinality of the powerset of
the natural numbers). A few ideas are nicely illustrated through Haskell code,
e.g., the diagonalization for the rational numbers, which shows that this set of
numbers is countably infinite.
3
Non-Haskell roads to logic and maths
When reading this book, it is a useful exercise to question every now and then
whether or not the Haskell road is a convenient one, what the issues would be
when mainstream programming languages were used at times.
It seems to be undisputed that Haskell allows for a more or less immediate
transcription of mathematical notation to program code. This is made possible
by Haskell’s declarativeness, purity and conciseness. Haskell’s laziness further
helps with encoding mathematical problems that involve infinite data structures. Haskell’s higher-orderness is essential to provide mathematical concepts
as library functionality. (For instance, think of the higher-order function that
turns a sequence of a polynomial’s coefficients in the actual function for the
polynomial.) Finally, Haskell’s type system serves an important documentary
purpose; machine-checked type information and executability are in fact the two
major ways in which Haskell notation complements mathematical notation.
Without any claim of completeness, we would like to illustrate some of the
issues that arise when encoding maths in mainstream languages such as C# or
Java. For the code examples that follow, we will use C# (2.0), but we could
have used Java (1.5) just as well, without changing the tenor of the observations.
Consider the following Haskell function, difs, for computing a difference
sequence. This operation is at the heart of Babbage’s difference engine. (Recall:
one uses the function difs to determine the degree of a potential polynomial
f (x) that should be derived as the closed form from a “long enough” sequence
of integers a0 , . . . , an , for which we assume that it corresponds to f (0), . . . ,
f (n). The degree of f (x) is the number of iterations of applying difs until the
resulting sequence is a constant sequence, if this happens at all.)
difs
difs
difs
difs
:: [Integer] -> [Integer]
[] = []
[_] = []
(m:n:ks) = n-m : difs (n:ks)
9
The following C# encoding may serve as an efficient transcription.
static Decimal[] difs(Decimal[] given)
{
Decimal[] result = new Decimal[
given.Length == 0 ? 0 : given.Length-1];
for (int i=1; i<given.Length; i++)
result[i-1] = given[i]-given[i-1];
return result;
}
The C# function operates on arrays of values of type Decimal (which is C#’s
counterpart to Haskell’s arbitrary precision integers). The array for the result
is explicitly allocated, leveraging the observation that the difs function returns
a sequence that is one element shorter than the (non-empty) input sequence.
All sense of equational reasoning is gone. The allocation code requires an extra
proof in order to establish its correctness. There is actually no straightforward
way to retain the recursive formulation of the problem. (For the record, with
C++ pointer arithmetics, we were able to process the input array recursively.
However, the recursive synthesis of the result would not work so easily.)
Hence, the mainstream road to maths is bumpy because of the encoding
efforts; it will be rather painful once we wanted to reason about these programs.
We attempt some variations in the sequel. The following C# variation uses the
generic collection type List in place of arrays:
static List<Decimal> difs(List<Decimal> given)
{
List<Decimal> result = new List<Decimal>();
if (given.Count > 1)
{
result.Add(given[1]-given[0]);
result.AddRange(difs(given.GetRange(1,given.Count-1)));
}
return result;
}
The earlier array location reduces to a simple object construction for the
initially empty list. (Hence, we eliminated the need for the extra insight regarding the length of the result.) The List type is rich enough to re-establish the
recursive coding style. The exercised style is rather inefficient. For instance, the
GetRange method involves cloning of the “tail”. We also note that we still fail
to provide any sort of equational style; so equational reasoning remains out of
reach. Furthermore, we fail to obtain the conciseness of pattern matching.
Yet another deficiency is worth mentioning: The previous C# encodings
are all eager. This is perhaps acceptable for the use of difs in the context of
Babbage’s method, but eventually all sorts of similar list processing functionality
needs to be lazy. (Think of moving from polynomials to power series.) Here is
10
a more clever C# solution, which engages in less cloning, and which is lazy at
the same time:
static IEnumerable<Decimal> difs(IEnumerable<Decimal> given)
{
Decimal? i = null;
foreach (Decimal j in given)
{
if (i.HasValue) yield return j-i.Value;
i = j;
}
}
That is, we use the IEnumerable interface and the yield return idiom to
compute the resulting list lazily. (We also use C#’s nullable types; cf. “?”.) It is
not an accident that this solution uses an imperative loop again. That is, C#’s
streams (or lazy lists) are idiomatically restricted. Re-establishing recursion is
a non-trivial undertaking. Here is a (prohibitively inefficient) attempt:
static IEnumerable<Decimal> difs(IEnumerable<Decimal> given)
{
if (empty(given)) yield break;
if (empty(tail(given))) yield break;
Decimal m = head(given);
Decimal n = head(tail(given));
yield return n-m;
foreach (Decimal i in difs(tail(given)))
yield return i;
}
One major problem is that the streams returned from the recursive calls
needs to be pulled into the local result stream explicitly through the (expensive)
foreach loop. Also, the code is baroque anyhow, even though we anticipated
the possibility of helpers to take apart lists; cf. empty, head, tail (definitions
omitted).
Bottom line This detour has illustrated that relatively simple pieces of mathematical definitions can be difficult to encode in mainstream programming languages. So anyone in search of a road to logic and maths, is well advised to favor
the Haskell road, not to get lost in the programming bits of such an endeavor.
It is a unique capability of Haskell that mathematical notation (in the form of
equations, sums, products, compound expressions, etc.) is type-able and executable in this programming language more or less as is. The book under review
leverages this capability very appropriately.
11
4
Concluding remarks
Doets and van Eijck’s “The Haskell Road to Logic, Maths and Programming” is
an astonishingly extensive and accessible textbook on logic, maths, and Haskell.
The book adopts a systematic but relaxed mathematical style (definition, example, exercise, ...); the text is very pleasant to read due to a small amount
of anecdotal information, and due to the fact that definitions are fluently integrated in the running text. The overall selection of mathematical subjects is
non-surprising (which is good), except that some advanced topics are included
such as corecursion and Cantor’s infinities. One would perhaps expect to see a
more substantial treatment of calculus and some coverage of statistics.
Regarding the targeted audience, the text seems to be valuable and advisable
for undergraduates with mathematics and logic as minor subjects. However,
even mathematically ambitious readers should find this book interesting, be it
then more for the Haskell part of it. The book was indeed developed through
undergraduate courses at the University of Amsterdam.
The Haskell samples in the book appear to be highly idiomatic, and readers
can certainly expect to learn Haskell programming, so to say as a side effect.
Not surprisingly, the book does not cover all of a typical functional programming text. For instance, type classes are used only in a minimalistic way. Also,
monads do not show up anywhere in the book. While this is normally “unimaginable” for a Haskell introduction, it makes total sense for this specific book.
The book’s write-up and program development has been based on literate programming. (However, the sources that are obtainable from the authors’ website
do not currently comprise the text of the chapters; presumably for copyright
reasons.)
Regarding the availability of the book: the “Texts in Computing Series” of
King’s College Publications are published on a ”Print on Demand” basis. This
has as the benefit that the price is much lower than with regular publishers — at
the time of writing: 25$ vs. 14£. Some may argue that it also has the minor
drawback that the book is not for sale in local bookstore, but has to be ordered
through Amazon instead.
I take the liberty to close this review with an admittedly personal comment.
I would have wanted this very book during my computer science studies. The
Haskell way of modeling mathematical concepts in a declarative, executable
and strongly typed manner is very supportive in learning. The small but telling
distance between mathematical notation and Haskell notation makes it also
more rewarding to actually engage in reasoning about mathematical problems
and programs. I still remember wasting time on engaging in Pascal encodings
for some of my maths homework. Should I end up giving lectures on maths and
logic, I will love to use this book in such a course.
12
References
[1] L. Augusteijn. Sorting morphisms. In S. Swierstra, P. Henriques, and
J. Oliveira, editors, 3rd International Summer School on Advanced Functional Programming, volume 1608 of LNCS, pages 1–27. Springer-Verlag,
Sept. 1998.
[2] J. Barwise and J. Etchemendy. Language, Proof, and Logic. Center for the
Study of Language and Information, Stanford, CA, USA, 1999. Accompanied by software.
[3] J. Barwise and L. Moss. Vicious circles: on the mathematics of nonwellfounded phenomena. Center for the Study of Language and Information,
Stanford, CA, USA, 1996.
[4] K. Claessen and J. Hughes. QuickCheck: a lightweight tool for random
testing of Haskell programs. In Proceedings of the ACM Sigplan International Conference on Functional Programming (ICFP-00), volume 35.9 of
ACM Sigplan Notices, pages 268–279, N.Y., Sept. 18–21 2000. ACM Press.
[5] C. Hall and J. O’Donnell. Discrete mathematics using a computer. SpringerVerlag, 2000. Second Edition to appear [10].
[6] P. Hudak. The Haskell School of Expression: Learning Functional Programming through Multimedia. CUP, Jan. 2000.
[7] O. Kiselyov, R. Lämmel, and K. Schupke. Strongly typed heterogeneous
collections. In Haskell ’04: Proceedings of the ACM SIGPLAN workshop
on Haskell, pages 96–107. ACM Press, 2004.
[8] D. Leijen and E. Meijer. Domain specific embedded compilers. In Proceedings of the 2nd conference on Domain-Specific Languages, pages 109–122.
ACM Press, 1999.
[9] E. Meijer, M. Fokkinga, and R. Paterson. Functional Programming with
Bananas, Lenses, Envelopes and Barbed Wire. In FPCA91: Functional
Programming Languages and Computer Architecture, volume 523 of LNCS,
pages 124–144. Springer-Verlag, 1991.
[10] J. O’Donnell, C. Hall, and R. Page. Discrete Mathematics Using a Computer. Springer-Verlag, 2006. Second Edition.
[11] C. Okasaki. Purely functional data structures. PhD thesis, Carnegie Mellon
University, 1996.
[12] C. Okasaki. An Overview of Edison. In G. Hutton, editor, Electronic Notes
in Theoretical Computer Science, volume 41. Elsevier, 2001.
[13] S. Thompson. Type Theory and Functional Programming. International
Computer Science Series. Addison Wesley, 1991.
13
[14] S. Thompson. The Craft of Functional Programming. Addison Wesley,
1996. 2nd edition in 1999.
14
| 6cs.PL
|
EQUATIONS OF REES ALGEBRAS OF IDEALS IN TWO VARIABLES
arXiv:1511.04073v1 [math.AC] 12 Nov 2015
JEFF MADSEN
Abstract. Let I be an ideal of height two in R = k[x0 , x1 ] generated by forms of the same degree,
and let K be the ideal of defining equations of the Rees algebra of I. Suppose that the second
largest column degree in the syzygy matrix of I is e. We give an algorithm for computing the
minimal generators of K whose degree is at least e, as well as a simple formula for the bidegrees of
these generators. In the case where I is an almost complete intersection, we give a generating set
for each graded piece Ki,∗ with i ≥ e − 1.
Let I be an ideal in a Noetherian ring R. The Rees algebra of I, defined as R(I) = R[It] =
L∞ j j
j=0 I t , is a central object in both commutative algebra, where it is used to study the asymptotic
behavior of I, and algebraic geometry, where it corresponds to blowing up the closed subscheme
V (I) in Spec R. A common approach to understanding Rees algebras is by representing them as
quotients of polynomial rings. Namely, if I = (f1 , . . . , fn ), then R(I) is a quotient of the polynomial
ring S = R[T1 , . . . , Tn ] via the R-linear map Ti 7→ fi t. The kernel of this map is an ideal K ⊂ S,
called the ideal of defining equations of R(I). Much work has been done to determine the ideal K
under various conditions (such as [1, 3, 5, 14, 15, 16, 23, 24, 25, 26, 31, 32]). However, even in the
relatively simple case where R = k[x0 , x1 ] is a polynomial ring in two variables over a field and I
is a homogeneous ideal generated in a single degree d, the problem of determining the equations of
R(I) remains open.
This problem is of interest not only in commutative algebra and algebraic geometry, but also in
elimination theory and geometric modeling, where it is connected to implicitization of parametrized
varieties (e.g. [4, 9, 10, 28, 27]). We focus on the case R = k[x0 , x1 ] where k is an algebraically
closed field. To any height two homogeneous ideal I ⊂ R generated by forms f1 , . . . , fn of degree
(f1 :···:fn )
d, we may associate a morphism Φ : P1k −−−−−−→ Pkn−1 , which parametrizes a curve C ⊂ Pkn−1 .
Conversely, to any map Φ : P1k → C parametrizing a curve in Pkn−1 by equations (f1 : · · · : fn ), we
may associate the ideal I = (f1 , . . . , fn ) ⊂ R. Then the Rees algebra R(I) is the bihomogeneous
coordinate ring of the graph of Φ,
ΓΦ = {(p, Φ(p)) | p ∈ P1k } ⊂ P1k × Pkn−1 .
Date: January 9, 2018.
2010 Mathematics Subject Classification. Primary: 13A30, Secondary: 13D02, 14E05, 14H50, 14Q05.
Key words and phrases. Bidegree, elimination theory, Hilbert-Burch matrix, integral closure, parametrization,
plane curve, rational curve, rational normal scroll, Rees algebra, Sylvester form, symmetric algebra.
1
Since S = R[T1 , . . . , Tn ] is the bihomogeneous coordinate ring of P1k × Pkn−1 , this means that the
ideal K ⊂ S of defining equations of R(I) is the ideal of ΓΦ . That is, for any bihomogeneous h ∈ S,
h ∈ K ⇔ h(x0 , x1 , f1 (x0 , x1 ), . . . , fn (x0 , x1 )) = 0.
In particular, the elements of K having x-degree 0 are just the implicit equations of the curve C.
(By x-degree we mean the degree in the variables x0 , x1 , as opposed to the degree in T1 , . . . , Tn , or
T -degree.)
The first step in computing the equations of the Rees algebra is to compute the syzygy matrix
ϕ of I, which determines the T -linear part of K. Since I is generated in one degree, each column of
the matrix ϕ has the same degree. Several results have been proved by imposing restrictions on the
syzygy matrix. For ideals generated by three forms, Cox, Hoffman, and Wang [7] gave a recursive
algorithm, using Sylvester forms, for computing a minimal generating set for the equations of the
Rees algebra when I has a syzygy of degree 1. This was generalized by Cortadellas and D’Andrea
[6] to the case where a syzygy of I of minimal degree µ, when viewed as a column vector, has
linearly dependent entries (a “generalized zero”). However, their method only produces part of a
minimal generating set. In particular, the equations produced by this algorithm all have x-degree
≥ µ − 1. These are in fact all the minimal generators of K having x-degree ≥ µ − 1, as proved by
Kustin, Polini, and Ulrich [19].
Kustin, Polini, and Ulrich [22] took another approach to determine the equations of K when I
has any number of generators and has a presentation matrix ϕ which is almost linear, meaning all
but one of the columns of ϕ have linear entries. They showed that the Rees algebra is a quotient
of a ring A, which is the coordinate ring of a rational normal scroll, by a height one prime ideal
KA. The equations of the scroll being easy to compute, this reduces the problem to computing the
generators of the divisor KA on A, which was carried out in [21]. The equations are even given
explicitly, assuming ϕ is put into a canonical form.
In this paper, we extend the method of Kustin, Polini, and Ulrich to arbitrary height two ideals
I in k[x0 , x1 ] generated by forms of the same degree. Unlike the almost linearly presented case, we
cannot produce all the defining equations of the Rees algebra in this way, but we do obtain all the
minimal generators of x-degree greater than or equal to the second largest degree of a column of
ϕ, and in some cases, in degree one less. Theorem 3.9 gives an abstract form of these generators,
and Corollary 3.17 gives a recursive algorithm for computing them. In particular, their bidegrees
follow a simple formula; see Corollary 3.13.
In Example 3.20, we show how to use Theorem 3.9 to solve the almost linearly presented case,
reproving [22, Theorem 3.6]. In Example 3.21, we show how the algorithm in Corollary 3.17 reduces
to the Sylvester form algorithm in [6, Theorem 2.10], when I has three generators and has a syzygy
of minimal degree with a generalized zero. In Example 3.22, we give another example, applying
Corollary 3.17 to the case where I has three generators and a syzygy of minimal degree 2 without
a generalized zero, a case which was previously solved in [2] and [6] using Morley forms.
2
In Section 1, we introduce the setup of the problem and the necessary definitions. In particular,
we construct a sequence of modules Em which are successive approximations of I.
In Section 2, we determine the structure of the modules R(Em ) by embedding each Em in a free
L
module. This allows us to embed R(Em ) in a ring of the form R(M ) where M = si=1 mσi (σi ),
which we show in Proposition 2.5 is the coordinate ring of a rational normal scroll. This embedding is actually an integral extension, and the two rings are equal in large degree, as we show in
Theorem 2.9.
In Section 3, we use the results of Section 2 to compute the ideal of equations of R(Em+1 ) in
R(Em ) by first computing it in R(M ) and then contracting to R(Em ). In Theorem 3.9, we give
the elements in R(M ) which contract to a minimal generating set of the x-degree ≥ dm part of the
ideal of equations of R(Em+1 ). The contraction may be made explicit using the algorithm given
in Corollary 3.17. Corollary 3.13 applies this in particular to the case m = n − 2 to obtain the
bidegrees of the minimal generators of K with x-degree ≥ dn−2 .
In Section 4, we consider specifically plane curves, that is, the case n = 3. In this case, we can
be more explicit about the R(E1 )-module structure of R(M ), in Proposition 4.7. We use this to
give a generating set, in general not minimal, for each graded component Ki,∗ with i ≥ d1 − 1.
1. Setup
Throughout this paper, we assume k is a field, R is the two-variable polynomial ring R = k[x0 , x1 ],
and m is the homogeneous maximal ideal m = (x0 , x1 ). We use the convention that N is the set of
nonnegative integers.
Although we wish to compute Rees algebras of ideals, it will be helpful to use Rees algebras of
more general modules. If E is a finitely generated R-module, then following [29], we define the
Rees algebra R(E) to be the quotient of the symmetric algebra S(E) by its R-torsion. As R is a
domain, this is consistent with the other standard definitions of Rees algebras of modules, such as
in [13], and it is also consistent with the usual definition of the Rees algebra of an ideal. Recall
∼ R[T1 , . . . , Tn ]/(g1 , . . . , gm ) where
that if ϕ is the m × n presentation matrix of E, then S(E) =
i
i h
h
(1.1)
g1 · · · gm = T1 · · · Tn ϕ.
The Krull dimension of the Rees algebra is dim R(E) = dim R + rank E = 2 + rank E ([29, Propo-
sition 2.2]).
Let S = R[T1 , . . . , Tn ], with deg xi = (1, 0) and deg Ti = (0, 1). We refer to the first degree in this
bigrading as the x-degree or degx , and the second degree as the T -degree or degT . By Si,j we mean
the bidegree (i, j) part of S, viewed as a k-vector space. Write Si,∗ for the k[T1 , . . . , Tn ]-module
L
L
j Si,j , and S∗,j for the R-module
i Si,j .
We shall use the same bigrading on any quotient of S, such as R(I). Note that this is not the
same as the bigrading induced by the inclusion R(I) = R[It] ⊂ R[t]. For example, the element fi t
has bidegree (d, 1) in R[t], however we will think of it as having bidegree (0, 1) since it corresponds
3
to the element Ti ∈ S. Thus to be more precise, we consider R(I(d)) rather than R(I), so that the
grading induced by S is the natural grading on the Rees algebra.
If I is a height two ideal of R minimally generated by n forms of degree d, then I(d) has a
Hilbert-Burch resolution
(1.2)
0→
n−1
M
ϕ
R(−di ) −
→ Rn → I(d) → 0
i=1
where I is generated by the maximal minors of ϕ, and thus
Pn−1
i=1
di = d. We may assume the
column degrees of ϕ are nondecreasing, that is, 1 ≤ d1 ≤ · · · ≤ dn−1 .
For 1 ≤ m ≤ n − 1, let ϕm be the n × m matrix consisting of the first m columns of ϕ and let
Em = coker ϕm , so that ϕn−1 = ϕ and En−1 = I(d). There is a sequence of surjections
R3 ։ E1 ։ E2 ։ · · · ։ En−1 = I(d)
which induce surjections
S ։ R(E1 ) ։ R(E2 ) ։ · · · ։ R(En−1 ) = R(I(d)).
We think of the rings R(Em ) as successive approximations of R(I(d)). Instead of directly computing
the equations of R(I(d)) in S, we may compute the equations in R(En−2 ), where it will be simpler.
We then need to compute the equations of R(En−2 ) in R(En−3 ), and so on. That is, we have a
sequence of inclusions
0 ( K1 ( K2 ( · · · ( Kn−1 = K
where each Km is a prime ideal with ht Km = dim S − dim R(Em ) = (2 + n) − (2 + rank Em ) = m.
We study the modules Km+1 R(Em ) ∼
= Km+1 /Km . A generating set for K may be obtained by
combining the lifts of generating sets of all the modules Km+1 /Km .
We summarize the data that we will be assuming.
Data 1.3. Let I be a height 2 ideal in R generated by n forms f1 , . . . , fn of degree d, with
presentation matrix ϕ, and resolution as given in (1.2). Let S = k[T1 , . . . , Tn ], let g1 , . . . , gn−1 ∈ S
be the equations of the symmetric algebra of I as in (1.1), and let K ⊂ S be the ideal of equations
of the Rees algebra of I. For 1 ≤ m ≤ n − 1, let ϕm be the n × m matrix consisting of the first m
columns of the presentation matrix ϕ and let Em = coker ϕm . Let Km ⊂ S be the ideal of equations
of the Rees algebra of Em .
By (1.1), the equations of S(Em ) are g1 , . . . , gm . Thus for each m, there is a short exact sequence
(1.4)
0 → Km /(g1 , . . . , gm ) → S(Em ) → R(Em ) → 0.
Let us deduce some simple properties of the modules Em .
Lemma 1.5. Assume Data 1.3 and let 1 ≤ m ≤ n − 1. Then Em is torsionfree of rank n − m, and
for any w ∈ m, (Em )w is a free Rw -module.
4
Proof. Because ϕ is injective, and ϕm is the restriction of ϕ to a submodule, ϕm is injective. Thus
Em has a free resolution
(1.6)
0→
m
M
ϕm
R(−di ) −−→ Rn → Em → 0.
i=1
By expanding the determinants along the last n−1−m columns, we see that I = In−1 (ϕ) ⊂ Im (ϕm ).
Because ht I = 2, we get ht Im (ϕm ) = 2. So if we invert w ∈ m, we get Im (ϕm )w = Rw . This means
∼ Rm−n .
that after inverting w, (1.6) splits, yielding (Em )w =
w
It remains to show that Em is torsionfree, or equivalently that if p 6= 0 is a homogeneous prime
ideal of R, then p 6∈ AssR (Em ). If p 6= m, then (Em )p is free by the above, so pRp 6∈ AssRp ((Em )p )
and thus p 6∈ AssR (Em ). On the other hand, pd Em ≤ 1 by (1.6), so depth Em ≥ 1 by the
Auslander-Buchsbaum formula. Hence m 6∈ AssR (Em ), therefore Em is torsionfree.
As a consequence, we obtain a formula for Km . While this description does not provide any
information about the generators of Km , it is the basis upon which we will perform our calculations
involving Km .
Proposition 1.7. Assume Data 1.3 and let 1 ≤ m ≤ n − 1. Then Km = (g1 , . . . , gm ) :S m∞ .
Proof. By Lemma 1.5, for any w ∈ m, (Em )w is a free Rw -module. Therefore R(Em )w = R((Em )w ) =
S((Em )w ) = S(Em )w . It follows from (1.4) that Km Sw = (g1 , . . . , gm )Sw , which means Km :S m∞ =
(g1 , . . . , gm ) :S m∞ . But Km is a prime ideal, so Km = Km :S m∞ = (g1 , . . . , gm ) :S m∞ .
2. The modules R(Em )
Fix m. To determine the Rees algebra of Em , we begin by embedding Em in a free module.
∗∗ is free (where −∗ = Hom (−, R)); we have an
Namely, we show in Lemma 2.1 that F = Em
R
∗∗ since E
injection Em ֒→ Em
m is torsionfree by Lemma 1.5. The free module F is generated in
nonpositive degrees, while Em is generated in degree 0. Thus E is a submodule of M = F≥0 . This
section is concerned with computing the Rees algebras of F and M and comparing them to the
Rees algebra of Em .
The lemma below is the graded version of [18, Proposition 2.1].
Ls
Lemma 2.1. Let E be an R-module of rank s generated in degree 0. Then E ∗∗ ∼
= i=1 R(σi ) with
all σi ≥ 0.
Proof. Set F = E ∗∗ . Then F is reflexive, so it satisfies Serre’s condition S2 . Since F is torsionfree,
dim F = dim R = 2, hence F is maximal Cohen-Macaulay. Because R is regular, the AuslanderL
Buchsbaum formula implies F is free. Thus we may write F = si=1 R(σi ).
To see that all σi ≥ 0, choose a surjection Rn ։ E and consider the map ξ : Rn → F obtained by
composing this surjection with the natural map θ : E → E ∗∗ = F . Assume by way of contradiction
that some σi , say σs , is negative. Then, viewing ξ as an s × n matrix, the last row of ξ must be zero.
5
This means im ξ ⊂
Ls−1
i=1
R(σi ), so rank(im ξ) < s. But im ξ = im θ, so rank(im ξ) = rank E = s.
By contradiction, all σi are nonnegative.
Recall from Lemma 1.5 that Em is a torsionfree R-module of rank n − m. Using Lemma 2.1, we
extend Data 1.3:
Data 2.2. Assume Data 1.3, fix 1 ≤ m ≤ n − 1, and set s = n − m. Let E = Em and let
Ls
F = E ∗∗ ∼
= i=1 R(σi ), with σ1 ≥ · · · ≥ σs ≥ 0. Let r be the unique integer with 1 ≤ r ≤ s such
L
that σ1 ≥ · · · ≥ σr > 0 = σr+1 = · · · = σs . Let M = F≥0 = si=1 mσi (σi ). Let ξ be the composite
map Rn ։ E ֒→ F .
Because E is generated in degree 0, the inclusion E ֒→ F factors as E ֒→ M ֒→ F . Therefore
there are inclusions of Rees algebras R(E) ֒→ R(M ) ֒→ R(F ). This will allow us to do computations in R(E) by extending to R(M ) or R(F ) and then contracting, as in Proposition 3.1. The
Rees algebra of F is easy to understand.
Remark 2.3. Assume Data 2.2.
Then R(F ) = R[w1 , . . . , ws ] = k[x0 , x1 , w1 , . . . , ws ], where
deg xi = (1, 0) and deg wi = (−σi , 1). Viewing R(E) as a subset of R(F ), we have
i
i h
h
T1 · · · Tn = w1 · · · ws ξ.
L
Proof. Since F = si=1 R(σi ) is a free module, the Rees algebra R(F ) is the same as the symmetric
algebra S(F ), which is a polynomial ring R[w1 , . . . , ws ]. Each variable wi corresponds to one
generator of F . Since the ith generator of F has degree −σi , the variable wi has bidegree (−σi , 1).
The R-module map ξ : Rn ։ E ֒→ F induces the R-algebra map R(ξ) : S = R(Rn ) ։ R(E) ֒→
R(F ) which lets us view R(E) as a submodule of R(F ). Therefore
i
i h
h
T1 · · · Tn = w1 · · · ws ξ.
The Rees algebra of M is also not difficult to describe. First, will to define the R-algebra
Scr(σ), which we will then show is isomorphic to R(M ) in Proposition 2.5. The ring Scr(σ) is the
coordinate ring of a rational normal scroll if all the σi are positive; otherwise, it is a cone over a
rational normal scroll. In any case, it is a Cohen-Macaulay normal domain of dimension s + 2 (see
[11] or [12]).
Definition 2.4. Let σ1 , . . . , σs be nonnegative integers, and let V = R[{vi,j | 1 ≤ i ≤ s, 0 ≤ j ≤
σi }]. Let Γ be the matrix with entries in V :
"
x0 v1,0 v1,1 · · · v1,σ1 −1 · · ·
Γ=
x1 v1,1 v1,2 · · ·
v1,σ1
···
vs,0 vs,1 · · ·
vs,σs −1
vs,1 vs,2 · · ·
vs,σs
#
.
Define the ring Scr(σ) = Scr(σ1 , . . . , σs ) = V /I2 (Γ).
Proposition 2.5. Assume Data 2.2. Then
R(M ) = R[{xσ0 i −j xj1 wi | 1 ≤ i ≤ s, 0 ≤ j ≤ σi }] ⊂ R[w1 , . . . , ws ] = R(F )
6
and R(M ) ∼
= Scr(σ).
Proof. From Remark 2.3, R(F ) = R[w1 , . . . , ws ] where deg xi = (1, 0) and deg wi = (−σi , 1). Thus
the monomials in R(F ) of x-degree 0 are xσ0 i −j xj1 wi , so
R(M ) = R(F )≥0,∗ = R[{xσ0 i −j xj1 wi | 1 ≤ i ≤ s, 0 ≤ j ≤ σi }].
Now define V and Γ as in Definition 2.4, and define a surjective R-linear map θ : V ։ R(M )
by vi,j 7→ xσ0 i −j xj1 wi . It is easy to see that I2 (Γ) ⊂ ker θ, so θ descends to a map θ : Scr(σ) =
V /I2 (Γ) ։ R(M ). Since dim Scr(σ) = s + 2 = 2 + rank M = dim R(M ), θ is a surjective map
between domains of the same dimension, hence it is an isomorphism.
Because R(F ) and R(M ) have straightforward descriptions, we want to compare R(E) to these
rings. The first step is to compare the modules E and F by looking at F/E. In the remark below,
we see that F/E is Artinian and is zero in degree ≥ dm − 1. This will be used in Theorem 2.9 to
show that R(F ) and R(E) agree in these degrees. As another consequence, the resolution (2.8)
gives an easy way of computing the matrix ξ, as the transpose of the syzygy matrix of ϕT
m . We
need to know ξ in order to make the homomorphism R(E) ֒→ R(F ) explicit (Remark 2.3).
Lemma 2.6. Assume Data 2.2.
(a) F/E has free resolution
(2.7)
0→
m
M
ξ
ϕm
→F =
R(−di ) −−→ Rn −
s
M
R(σi ) → F/E → 0.
i=1
i=1
(b) F/E is Artinian.
(c) (F/E)≥dm −1 = 0.
(d) Ext2R (F/E, R) has free resolution
(2.8)
0→
s
M
ξT
ϕT
R(−σi ) −→ Rn −−m
→
(e) s = n − m and
i=1 σi
R(di ) → Ext2R (F/E, R) → 0.
i=1
i=1
Ps
m
M
=
Pm
i=1 di .
Proof. (a) This resolution is obtained by combining the short exact sequence (1.6) with the short
exact sequence
0 → E → F → F/E → 0.
(b) If p ⊂ R is a prime ideal with p 6= m, then Ep is free by Lemma 1.5. Therefore Fp = Ep∗∗ = Ep ,
so (F/E)p = 0. Hence Supp(F/E) ⊂ {m}, meaning F/E is Artinian.
(c) Since 1 ≤ d1 ≤ · · · ≤ dm , it follows from (2.7) that reg(F/E) = dm − 2. Since F/E is
Artinian, this means that (F/E)≥dm −1 = 0.
(d) Since F/E is Artinian, grade(F/E) = 2. By (2.7), pd(F/E) ≤ 2, therefore F/E is perfect of
grade 2. Thus the resolution of Ext2R (F/E, R) is dual to the resolution of F/E in (2.7).
7
(e) From (2.7), we see that the Hilbert polynomial of F/E is
m
s
X
X
R(t − di + 1)
(t + σi + 1) − n(t + 1) +
HPF/E (t) =
i=1
s
X
i=1
= (s + m − n)t + (s + m − n +
σi −
i=1
m
X
di ).
i=1
On the other hand, F/E is Artinian, so its Hilbert polynomial must be zero. Therefore s = n − m
P
P
and si=1 σi = m
i=1 di .
The next theorem relates the Rees algebras R(E), R(M ), and R(F ). It allows us, when looking
in large x-degrees, to pass from R(E) to R(M ), where computation is easier due to the structure
of R(M ) given in Proposition 2.5.
Theorem 2.9. Assume Data 2.2.
(a) R(E)≥dm −1,∗ = R(M )≥dm −1,∗ = R(F )≥dm −1,∗ .
(b) R(M ) is the integral closure of R(E) in its field of fractions.
(c) E is a reduction of M in F .
Proof. (a) The second equality is true because R(M ) = R(F≥0 ) = R(F )≥0,∗ . For the first equality,
we show by induction on j ≥ 0 that R(E)i,j = R(M )i,j for all i ≥ dm − 1. If j = 0, this is true
because R(E)∗,0 = R = R(M )∗,0 . Let j = 1 and let i ≥ dm − 1. As E and M are torsionfree
R-modules, R(E)∗,1 = E and R(M )∗,1 = M . But E≥dm −1 = F≥dm −1 = M≥dm −1 by Lemma 2.6(c),
therefore R(E)≥dm −1,1 = R(M )≥dm −1,1 . In particular, R(E)i,1 = R(M )i,1 since i ≥ dm − 1. Now
suppose j > 1 and i ≥ dm − 1. Certainly R(E)i,j ⊂ R(M )i,j . On the other hand,
R(M )i,j = R(M )0,1 R(M )i,j−1
= R(M )0,1 R(E)i,j−1 by induction
= R(M )0,1 R(E)i,0 R(E)0,j−1
⊂ R(M )i,1 R(E)0,j−1
= R(E)i,1 R(E)0,j−1 by the case j = 1
= R(E)i,j .
Therefore R(E)i,j = R(M )i,j .
(b) Since R(M )≥dm −1,∗ = R(E)≥dm −1,∗ is a faithful R(M )-module which is a finitely generated
R(E)-module, the determinant trick ([30, Lemma 2.1.9]) shows that the extension R(E) ⊂ R(M )
is integral. Since R(M ) ∼
= Scr(σ) by Proposition 2.5, R(M ) is normal. It remains to show that
R(E) and R(M ) have the same field of fractions. For each 1 ≤ i ≤ s, define pi = xσ0 i +dm −1 wi .
i −dm +1
Then deg pi = (dm − 1, 1), so pi ∈ R(E) by (a). Also, x0 ∈ R ⊂ R(E), thus wi = x−σ
pi ∈
0
Quot(R(E)). Since x0 , x1 ∈ Quot(R(E)) and wi ∈ Quot(R(E)) for all i, we get Quot(R(E)) ⊃
8
k(x0 , x1 , w1 , . . . , ws ) = Quot(R(F )) ⊃ Quot(R(M )), therefore Quot(R(E)) = Quot(R(M )). Hence
R(M ) is the integral closure of R(E) in its field of fractions.
(c) This is equivalent to R(M ) being integral over R(E) ([30, Theorem 16.2.3]), which we already
proved in (b).
3. Equations of Rees algebras
Now that we have some understanding of R(E), we wish to compute the ideal Km+1 R(E),
as discussed in Section 1.
First, in Proposition 3.1, we use the fact that R(E) ⊂ R(M ) is
an integral extension of rings to show that Km+1 R(E) is the contraction of the R(M )-ideal
(gm+1 R(F ))≥0,∗ = gm+1 R(F )≥−dm+1 ,∗ .
Theorem 3.7 gives a generating set for ideals of this
form. We will put these together to obtain the x-degree ≥ dm part of a minimal generating
set of Km+1 R(E) in Theorem 3.9. While these generators are given in terms of the variables
w1 , . . . , ws ∈ R(M ), we desire an expression for the generators as elements of S = R[T1 , . . . , Tn ].
We cannot give a closed form, but in Corollary 3.17 we give a recursive algorithm for computing
the generators of Km+1 R(E) with x-degree ≥ dm .
Proposition 3.1. Assume Data 2.2. Then
Km+1 R(E) = (Km+1 R(M )) ∩ R(E) = (gm+1 R(F )≥−dm+1 ,∗ ) ∩ R(E).
Proof. According to Theorem 2.9(b), R(E) ⊂ R(M ) is an integral extension of rings. Since
Km+1 R(E) is a prime ideal, the first equality is true by the lying over property of integral extensions.
By Proposition 1.7, Km = (g1 , . . . , gm ) :S m∞ . Since Km is the ideal of defining equations of
R(E), this means g1 , . . . , gm are zero in R(E), and thus also in R(M ). Proposition 1.7 also gives
Km+1 = (g1 , . . . , gm+1 ) :S m∞ . Therefore
Km+1 R(M ) = (g1 , . . . , gm+1 )R(M ) :R(M ) m∞ = gm+1 R(M ) :R(M ) m∞ .
It remains to show that gm+1 R(M ) :R(M ) m∞ = gm+1 R(F )≥−dm+1 ,∗ .
The containment “⊃” comes because
mdm+1 (gm+1 R(F )≥−dm+1 ,∗ ) ⊂ gm+1 R(F )≥0,∗ = gm+1 R(M ),
which means
gm+1 R(F )≥−dm+1 ,∗ ⊂ gm+1 R(M ) :R(M ) mdm+1 ⊂ gm+1 R(M ) :R(M ) m∞ .
For the containment “⊂”, first note that
gm+1 R(M ) :R(M ) m∞ ⊂ (gm+1 R(F ) :R(F ) m∞ ) ∩ R(M ).
But since R(F ) is a polynomial ring, gm+1 R(F ) is an unmixed ideal of height 1, while mR(F ) is a
prime ideal of height 2, so gm+1 R(F ) :R(F ) m∞ = gm+1 R(F ). Thus
gm+1 R(M ) :R(M ) m∞ ⊂ (gm+1 R(F )) ∩ R(M ) = (gm+1 R(F ))≥0,∗ = gm+1 R(F )≥−dm+1 ,∗ .
9
This shows that gm+1 R(M ) :R(M ) m∞ = gm+1 R(F )≥−dm+1 ,∗ .
According to Proposition 3.1, there are two steps to computing Km+1 R(E). First, we must compute gm+1 R(F )≥−dm+1 ,∗ , then we must contract to R(E). We begin with the task of determining
the generators of R(F )≥−c,∗ as an R(M )-module for any c ≥ 0.
3.1. Generators of truncations of R(F ). Assume Data 2.2. For any s-tuple α = (α1 , . . . , αs ) ∈
Ns , write wα = w1α1 · · · wsαs ∈ R(F ), and define α · σ = α1 σ1 + · · · + αs σs . Then since degx xi = 1
and degx wi = −σi , we may compute the x-degree of any monomial in R(F ) with the formula
degx (xj0 xk1 wα ) = j + k − α · σ.
In what follows, it will be convenient to use the following nonstandard notation: For h, h′ ∈ R(F ),
say h divides h′ (or write h | h′ ) to mean that there exists ℓ ∈ R(M ) such that h′ = ℓh. Note
that this is different from the standard notion of divisibility in R(F ) because we require ℓ to have
nonnegative degree (that is, to be in R(M ) and not just R(F )).
One more definition will make it easier to work with exponent vectors α ∈ Ns . Recall that r is
the unique integer with 1 ≤ r ≤ s such that σi > 0 for i ≤ r and σi = 0 for i > r. In the definition
below, we think of α+ as the part of the vector corresponding to those σi which are positive, and
α0 as the part corresponding to those σi which are zero.
Definition 3.2. Assume Data 2.2, and let α ∈ Ns . Define α+ = (α1 , . . . , αr , 0, . . . , 0) ∈ Ns and
α0 = (0, . . . , 0, αr+1 , . . . , αs ) ∈ Ns .
The following properties are clear from the definition.
Remark 3.3. Assume Data 2.2, and let α ∈ Ns .
(a) α = α+ + α0 .
(b) (α+ )0 = 0 and (α0 )+ = 0.
(c) α · σ = α+ · σ.
The importance of distinguishing α+ from α0 can be seen in the following lemma, which shows
that we may ignore wr+1 , . . . , ws for the purpose of computing minimal generators.
Lemma 3.4. Assume Data 2.2. Then after row operations on ξ corresponding to an automorphism
of k[w1 , . . . , ws ] which fixes w1 , . . . , wr , we get w α ∈ R(E) for all α ∈ Ns with α+ = 0.
Proof. By definition of r, for every i > r we have σi = 0. By (2.7), this means the last s−r rows of ξ
have entries in k. Now by Lemma 2.6(a,b), grade Is (ξ) = grade Fitt0 (F/E) = grade ann(F/E) = 2,
so Is (ξ) 6= 0. Thus there is an invertible n × n matrix χ with entries in k such that
#
"
A B
ξχ =
0 Is−r
where Is−r is the (s − r) × (s − r) identity matrix, and A and B are arbitrary matrices of sizes
r × (n − s + r) and r × (s − r), respectively. We may perform row operations on ξ that involve
10
subtracting multiples of the last s − r rows from the first r rows to assume that B = 0. Such row
operations correspond to automorphisms of k[w1 , . . . , ws ] fixing w1 , . . . , wr . Now Remark 2.3 yields
#
"
i A
i
h
i
h
h
0
,
T1 · · · Tn χ = w1 · · · ws ξχ = w1 · · · ws
0 Is−r
from which we get Tn−s+i = wi for all r < i ≤ s. Because all Tj ∈ R(E), this means wr+1 , . . . , ws ∈
R(E). Therefore if α ∈ Ns has α+ = 0, meaning α = (0, . . . , 0, αr+1 , . . . , αs ), we get w α =
α
r+1
wr+1
· · · wsαs ∈ R(E).
Now we are ready to give minimal generating sets for R(M )-modules of the form R(F )≥−c,∗ .
Definition 3.5. Assume Data 2.2, and fix c ≥ 0. Set
Λc = {α ∈ Ns | α0 = 0 and α · σ ≥ c}.
For 1 ≤ i ≤ r, let
Ωc,i = {α ∈ Ns | αi > 0 and αj = 0 for j > i and c ≤ α · σ < c + σi } ⊂ Λc
S
and set Ωc = ri=1 Ωc,i ⊂ Λc . Define
Ac = {w α | α ∈ Ns and α0 = 0 and α · σ < c}
and
Bc = {xj0 xk1 wα | α ∈ Ωc and j + k = α · σ − c}.
We will see in Theorem 3.7 that Ac ∪ Bc is a minimal generating set for R(F )≥−c,∗ . Observe that
for any monomial h ∈ Ac , we have −c < degx h ≤ 0, while any monomial h ∈ Bc has degx h = −c.
In fact, the set Ac consists of all monomials of x-degree > −c which do not involve x0 , x1 , or any wi
with i > r (which are the i with σi = 0). The set Bc is more complicated, but it may be thought of
as the set of all monomials of x-degree −c which do not involve any wi with i > r, and only include
just as many other wi as necessary to make the x-degree small enough. This is made precise in the
next lemma.
Lemma 3.6. Assume Data 2.2, and fix c > 0. Place a partial order on Ns such that α ≤ β if and
only if αi ≤ βi for all i. Then Ωc is the set of minimal elements of Λc under this partial order.
Proof. First, we show that every minimal element of Λc is in Ωc . Let α be a minimal element of
Λc . Choose i so that αi > 0 and αj = 0 for all j > i. From the definition of Λc , we have α0 = 0.
This means αr+1 = · · · = αs = 0, therefore i ≤ r. We claim that α ∈ Ωc,i . Suppose not. Then since
α · σ ≥ c by definition of Λc , we must have α · σ ≥ c + σi . Let β = (α1 , . . . , αi−1 , αi − 1, 0, . . . , 0) =
α − (0, . . . , 0, 1, 0, . . . , 0) (with the 1 in position i), which is in Ns since αi > 0. Moreover, β 0 = 0,
and β · σ = α · σ − σi ≥ c. Therefore β ∈ Λc . But β < α, contradicting the minimality of α. Hence
α ∈ Ωc,i ⊂ Ωc .
11
Now we must show that every element of Ωc is minimal in Λc . Take α ∈ Ωc ; then α ∈ Ωc,i
for some 1 ≤ i ≤ r. Consider β ∈ Λc with β ≤ α. Choose k so that βk > 0 and βj = 0 for
all j > k. Since β ≤ α, the largest nonzero index of α must be greater than or equal to the
largest nonzero index of β, that is, i ≥ k. This means that for j > i, αj = βj = 0. Further, the
inequalities β · σ ≥ c and α · σ < c + σi yield (α − β) · σ = α · σ − β · σ < (c + σi ) − c = σi . Thus
P
P
σi > (α − β) · σ = sj=1 (αj − βj )σj = ij=1 (αj − βj )σj , since αj = βj for j > i. But each coefficient
P
αj − βj is nonnegative, so if some αℓ 6= βℓ , we would have ij=1 (αj − βj )σj ≥ σℓ ≥ σi , the latter
inequality because the σj are in non-increasing order. This contradiction means that we must have
αj = βj for all j, thus α = β. Hence α is minimal in Λc .
Theorem 3.7. Assume Data 2.2, and fix c > 0. Let Ac and Bc be as defined in Definition 3.5.
(a) The image of Ac is a minimal generating set for the R(M )-module R(F )≥−c,∗ /RR(F )−c,∗ .
(b) Bc is a minimal generating set for the R(M )0,∗ -module R(F )−c,∗ .
(c) Ac ∪ Bc is a minimal generating set for the R(M )-module R(F )≥−c,∗ .
Proof. (a) Let Ac be the image of Ac in R(F )≥−c,∗ /RR(F )−c,∗ . Since R(F )≥−c,∗ is generated by
monomials, to show that Ac is a generating set, it suffices to show that for any monomial h ∈ R(F )
with degx h ≥ −c, the image h ∈ R(F )≥−c,∗ /RR(F )−c,∗ is in the R(M )-module generated by Ac .
Write h = xj0 xk1 w β . First suppose β + · σ < c. Then since (β + )0 = 0, if we set h′ = w β
h′ ∈ Ac . Since
0
xj0 xk1 wβ
has nonnegative x-degree, it is in R(M ), therefore h =
+
we have
0
xj0 xk1 w β h′
is in
+
the R(M )-module generated by Ac . On the other hand, suppose c ≤ β · σ = β · σ. Then since
−c ≤ degx h = j + k − β · σ, we have 0 ≤ β · σ − c ≤ j + k. Thus there are 0 ≤ j ′ ≤ j and 0 ≤ k′ ≤ k
′
such that j ′ + k′ = β · σ − c. If we set h′ = xj0 xk1 w β , then we have degx h′ = j ′ + k′ − β · σ = −c,
j−j ′
meaning h′ ∈ R(F )−c,∗ . Therefore h = x0
′
′
x1k−k ∈ RR(F )−c,∗ , so h = 0, which is certainly in the
R(M )-module generated by Ac . Hence Ac generates R(F )≥−c,∗ /RR(F )−c,∗ .
Because Ac consists of monomials, to see that it is a minimal generating set, it is enough to show
that no element of Ac divides another. Suppose there are h, h′ ∈ Ac with h′ | h, meaning there
′
is ℓ ∈ R(M ) with h = h′ ℓ. Write h = w α and h′ = w α . The only way we can have h = h′ ℓ is if
′
α′ ≤ α, in which case ℓ = w α−α . Because ℓ ∈ R(M ), we must have deg ℓ = −(α − α′ ) · σ ≥ 0,
P
so (α − α′ ) · σ ≤ 0. But each term in the sum (α − α′ ) · σ = si=1 (αi − α′i )σi is nonnegative, so
each (αi − α′i )σi must be 0. If i ≤ r, then σi > 0, so αi = α′i . If i > r, then αi = α′i = 0 by
the assumption that α0 = (α′ )0 = 0. Thus α = α′ and therefore h = h′ . Hence Ac is a minimal
generating set for R(F )≥−c,∗ /RR(F )−c,∗ .
(b) Since every monomial in Bc has x-degree −c, we have Bc ⊂ R(F )−c,∗ . Since R(F )−c,∗
is generated by monomials, to show that Bc is a generating set it will suffice to show that any
monomial h ∈ R(F )−c,∗ is divisible by an element of Bc . Write h = xj0 xk1 wβ . Then
(3.8)
− c = degx h = j + k − β · σ = j + k − β + · σ,
12
so β + · σ ≥ c. Also, (β + )0 = 0, therefore β + ∈ Λc . By Lemma 3.6, there is α ∈ Ωc such that
α ≤ β + . Since α ∈ Λc , α · σ ≥ c. Thus 0 ≤ α · σ − c ≤ β + · σ − c = j + k (using (3.8)), so there
′
are 0 ≤ j ′ ≤ j and 0 ≤ k′ ≤ k such that j ′ + k′ = α · σ − c. Set h′ = xj0 xk1 w α . Then h′ ∈ Bc since
′
α ∈ Ωc and j ′ + k′ = α · σ − c. Because j ′ ≤ j, k′ ≤ k, and α ≤ β + ≤ β, there is ℓ ∈ R(F ) such
that h = ℓh′ . Since degx ℓ = degx h − degx h′ = (−c) − (−c) = 0, we have ℓ ∈ R(M )0,∗ . This proves
that h is divisible by an element of Bc , so Bc is a generating set for R(F )−c,∗ .
To show minimality, since Bc consists of monomials, it is enough to show that no element of Bc
′
divides another. Suppose there are h, h′ ∈ Bc with h′ | h. Write h = xj0 xk1 w α and h′ = xj0 xk1 w α .
′
′
Then α, α′ ∈ Ωc and α′ ≤ α. By Lemma 3.6, α′ = α. Therefore j ′ + k′ = α′ ·σ − c = α ·σ − c = j + k,
but also j ′ ≤ j and k′ ≤ k, therefore j ′ = j and k′ = k. This shows that h′ = h, hence Bc is a
minimal generating set for R(F )−c,∗ .
(c) It follows from (b) that Bc minimally generates the R(M )-module L defined in (a), and Ac
minimally generates R(F )≥−c,∗ /L by (a). Thus, taken together, Ac ∪ Bc generates R(F )≥−c,∗ . For
minimality, it is enough to show that no element of Ac ∪ Bc divides another. But no element of
Ac is divisible by another element of Ac ∪ Bc by (a), and no element of Bc is divisible by another
element of Bc by (b). Finally, no element of Bc may be divisible by an element of Ac since every
element of Bc has x-degree −c while every element of Ac has x-degree > −c. Thus Ac ∪ Bc is a
minimal generating set for R(F )≥−c,∗ .
3.2. Generators of the equations of the Rees algebra. Now we will prove the main theorem.
In Theorem 2.9, we showed the equality R(E)≥dm −1,∗ = R(M )≥dm −1,∗ . This allows us to compute
the contraction from Proposition 3.1 in x-degree ≥ dm , and thus determine the generators of
Km+1 R(E) in these degrees using Theorem 3.7.
It may seem like we should be able to compute the minimal generators in x-degree dm − 1 as
well. However, because Theorem 3.7 gives a generating set as an R(M )-module, and we want a
generating set as an R(E)-module, the proof of Theorem 3.9 is not as simple as just applying
Theorem 3.7. Instead, it works by factoring out the x-degree dm − 1 part of Km+1 R(E), which only
allows us to compute the generators in x-degree ≥ dm . In §4, we will determine the generators in
x-degree dm − 1 when n = 3 and m = 1.
Theorem 3.9. Assume Data 2.2. Then the set
(3.10)
{gm+1 w α | α ∈ Ns and α0 = 0 and α · σ ≤ dm+1 − dm }
is equal to the x-degree ≥ dm part of a minimal generating set for Km+1 R(E) as an R(E)-module.
Proof. Observe that the set (3.10) is equal to gm+1 Ac for c = dm+1 − dm + 1, where Ac is defined
in Definition 3.5. Let J = Km+1 R(E). The claim that gm+1 Ac is the part of a minimal generating
set of J having x-degree ≥ dm is equivalent to saying that the image of gm+1 Ac in R(E)/RJdm −1,∗
is a minimal generating set of J≥dm −1,∗ /RJdm −1,∗ .
13
From Proposition 3.1, J = (gm+1 R(F )≥−dm+1 ,∗ ) ∩ R(E). Therefore
J≥dm −1,∗ = (gm+1 R(F ))≥dm −1,∗ ∩ R(E)≥dm −1,∗ = (gm+1 R(F )≥−c,∗ ) ∩ R(E)≥dm −1,∗ .
But R(E)≥dm −1,∗ = R(F )≥dm −1,∗ by Theorem 2.9, therefore
(3.11)
J≥dm −1,∗ = gm+1 R(F )≥−c,∗ .
We claim that it is enough to show that R(F )≥−c,∗ = Ac R(E) + RR(F )−c,∗ . Indeed, if we
multiply this equation by gm+1 and use (3.11), we obtain J≥dm −1,∗ = gm+1 Ac R(E) + RJdm −1,∗ .
This shows that gm+1 Ac is a generating set for J≥dm −1,∗ /RJdm −1,∗ over R(E). Minimality follows
because Ac , and therefore gm+1 Ac , is minimal over the larger ring R(M ) (Theorem 3.7(a)).
We now prove that R(F )≥−c,∗ = Ac R(E) + RR(F )−c,∗ . Theorem 3.7(a) says that Ac generates
R(F )≥−c,∗ /RR(F )c,∗ as an R(M )-module, that is to say, R(F )≥−c,∗ = Ac R(M ) + RR(F )−c,∗ .
It is clear that Ac R(E) + RR(F )−c,∗ ⊂ Ac R(M ) + RR(F )−c,∗ , so we just need to show that
Ac R(M ) ⊂ Ac R(E) + RR(F )−c,∗ . Because R(M ) is generated by monomials, it suffices to show
that for any monomials h ∈ Ac and f ∈ R(M ), we have hf ∈ Ac R(E) + RR(F )−c,∗ .
By Lemma 3.4, we may apply an automorphism of k[w1 , . . . , ws ] which fixes w1 , . . . , wr , and
0
therefore fixes Ac , to assume that w β ∈ R(E) for all β. From the definition of Ac , we may
write h = wα with α0 = 0 and α · σ < c. Also, write f = xj0 xk1 w β , so that hf = xj0 xk1 w α+β . If
+
(α + β) · σ < c, then since (α + β) · σ = (α + β + ) · σ, we have wα+β ∈ Ac . Thus hf is the product
+
0
of wα+β ∈ Ac and xj0 xk1 wβ ∈ R(E), meaning hf ∈ Ac R(E).
On the other hand, if (α + β) · σ ≥ c, then d := degx (w α+β ) ≤ −c. Since degx h > −c (as h ∈ Ac )
and degx f ≥ 0 (as f ∈ R(M )), we obtain degx (hf ) ≥ −c. But degx (hf ) = degx (xj0 xk1 w α+β ) =
j + k + d, so 0 ≤ −c − d ≤ j + k. Thus there are 0 ≤ j ′ ≤ j and 0 ≤ k′ ≤ k such that
′
j ′ + k′ = −c − d. Set ℓ = xj0 xk1 w α+β . Then degx ℓ = j ′ + k′ + d = −c, so ℓ ∈ R(F )−c,∗ . Therefore
j−j ′
hf = (x0
′
′
x1k−k )ℓ ∈ RR(F )−c,∗ .
In either case, hf ∈ Ac R(E) + RR(F )−c,∗ , which proves the claim that R(F )≥−c,∗ = Ac R(E) +
RR(F )−c,∗ . Therefore the image of gm+1 Ac minimally generates J≥dm ,∗ /RJdm −1,∗ as an R(E)module. For every α ∈ Ns with α0 = 0 and α · σ < c, we get a generator gm+1 w α ∈ gm+1 Ac having
P
bidegree (dm+1 − α · σ, ri=1 αi + 1).
The bidegrees of the minimal generators given in Theorem 3.9 may be computed:
Corollary 3.12. Assume Data 2.2. Then Km+1 R(E) has a minimal generating set whose elements
P
of x-degree ≥ dm lie in bidegrees (dm+1 − α · σ, ri=1 αi + 1) for each α ∈ Ns with α0 = 0 and
α · σ ≤ dm+1 − dm . In particular, the only minimal generator of Km+1 R(E) having x-degree dm+1
is gm+1 , and all other minimal generators of Km+1 R(E) have x-degree < dm+1 .
Proof. Since gm+1 has bidegree (dm+1 , 1), for every α ∈ Ns with α0 = 0 and α · σ ≤ dm+1 − dm ,
P
the element gm+1 wα has bidegree (dm+1 − α · σ, ri=1 αi + 1), so the first claim follows from
14
Theorem 3.9. The second claim comes from noticing that if α0 = 0 ad α 6= 0, then degx (w α ) < 0,
therefore degx (gm+1 wα ) < dm+1 .
As a consequence, we get Corollary 3.13, which gives the bidegrees of the minimal generators
of K having x-degree at least dn−2 . For example pictures of these bidegrees, see Tables 1, 2, and
3. The row represents T -degree, the column represents x-degree, and the numbers in the table are
the number of generators in the corresponding bidegree. Only the minimal generators lying on the
right of the vertical line are given in Corollary 3.13. Note the patterns in these tables: in cases (i)
and (iii), the generators lie on a line of slope −1/σ1 ; while in case (ii), the generators lie between
a line of slope −1/σ1 and a line of slope −1/σ2 .
Corollary 3.13. Assume Data 2.2 with m = n − 2. Then s = 2 and σ1 + σ2 =
Pn−2
i=1
di , and the
minimal generators of K having x-degree ≥ dn−2 consist of the generator gn−2 and any other gm
having bidegree (dn−2 , 1), and also:
(i) If σ2 = 0: one generator of bidegree (dn−1 − jσ1 , j + 1) for each integer j with 0 ≤ j ≤
dn−1 −dn−2
.
σ1
(ii) If σ1 > σ2 > 0: one generator of bidegree (dn−1 − iσ1 − (j − i)σ2 , j + 1) for each pair of
integers (j, i) with 0 ≤ j ≤
dn−1 −dn−2
σ2
n−2 −jσ2
and 0 ≤ i ≤ min{j, dn−1 −d
}.
σ1 −σ2
(iii) If σ1 = σ2 : j + 1 generators of bidegree (dn−1 − jσ1 , j + 1) for each integer j with 0 ≤ j ≤
dn−1 −dn−2
.
σ1
Proof. The equalities s = 2 and σ1 + σ2 =
Pn−2
di come from Lemma 2.6. Since K = Kn−1 , a
generating set for K consists of the lifts of generating sets of each quotient Km+1 /Km ∼
= Km+1 R(Em )
i=1
for 1 ≤ m ≤ n − 2. For m < n − 2, Corollary 3.12 says that besides gm+1 , all the minimal
generators of Km+1 R(Em ) have x-degree less than dm+1 . Since dm+1 ≤ dn−2 , the only generator
of Km+1 R(Em ) that could possibly have x-degree at least dn−2 is gm+1 . The remaining generators
with x-degree at least dn−2 must come from generators of Kn−1 R(En−2 ).
The bidegrees of the minimal generators of Kn−1 R(En−2 ) are given in Corollary 3.12. If σ2 = 0,
then r = 1, so we only consider those pairs (α1 , α2 ) with α2 = 0. Setting j = α1 , we see that
Kn−1 R(En−2 ) has a minimal generator in bidegree (dn−1 −jσ1 , j+1) for all j with jσ1 ≤ dn−1 −dn−2 .
If σ2 > 0, then r = 2. Set i = α1 and j = α1 + α2 . Then Kn−1 R(En−2 ) has a minimal generator in
bidegree (dn−1 − iσ1 − (j − i)σ2 , j + 1) for all i, j with iσ1 + (j − i)σ2 ≤ dn−1 − dn−2 . If σ1 > σ2 , then
these bidegrees are all distinct. For if (dn−1 −iσ1 −(j −i)σ2 , j +1) = (dn−1 −i′ σ1 −(j ′ −i′ )σ2 , j ′ +1),
then certainly j = j ′ . But then i(σ1 − σ2 ) = i′ (σ1 − σ2 ), whence i = i′ . If σ1 = σ2 , on the other
hand, these bidegrees reduce to (dn−1 − jσ1 , j + 1), and there is one generator in this bidegree for
each 0 ≤ i ≤ j, for a total of j + 1 generators.
The condition σ2 = 0 in Corollary 3.13 has geometric significance, as we see in the next proposition, a consequence of the results in [8]. The proposition only holds if dn−2 < dn−1 , for if
dn−2 = dn−1 , then the matrix ϕn−2 consisting of all but the last column of ϕ is not an invariant.
15
7
6
5
1
4
1
3
1
2
1
1
1
1
0 1 2 3 4 5 6 7 8 9 10 11 12 13 14 15 16
Table 1. An example of case (i), with n = 3, d1 = 3, d2 = 16, σ1 = 3, and σ2 = 0.
7
6
1 1
5
1 1 1 1
1 1 1
4
3
1
1
1
1
2
1
1
1
1
1
0 1 2 3 4 5 6 7 8 9 10 11 12 13 14 15 16
Table 2. An example of case (ii), with n = 3, d1 = 5, d2 = 16, σ1 = 3, and σ2 = 2.
7
7
6
6
5
5
4
4
3
3
2
1
2
1
1
0 1 2 3 4 5 6 7 8 9 10 11 12 13 14 15 16
Table 3. An example of case (iii), with n = 3, d1 = 4, d2 = 16, σ1 = 2, and σ2 = 2.
However, if dn−2 = dn−1 , then the only minimal generators of K having x-degree ≥ dn−2 are some of
the gm , so there is no difference between the cases in Corollary 3.13. The birationality assumption
is also mild, as any rational curve C may be reparametrized so that the morphism Φ : P1k → C is
birational, as shown in [20]. It is not known whether there is any geometric meaning of the values
of σ1 and σ2 when they are both positive.
Proposition 3.14. Assume Data 2.2 with m = n − 2. Let f1 , . . . , fn be a minimal generating set
of I and let Φ : P1k → C ⊂ Pkn−1 be the morphism defined by (f1 : · · · : fn ). Assume k is algebraically
16
closed and Φ is birational onto its image C, and assume dn−2 < dn−1 . Then σ2 = 0 if and only if
C has a point of multiplicity greater than dn−2 (which must then have multiplicity dn−1 ).
Proof. Let p = (a1 : · · · : an ) ∈ Pkn−1 be a point. By [8, Theorem 1.8], the multiplicity of p on C is
deg gcd I1 (aϕ) where a = [a1 · · · an ]. We may write aϕ = [h1 · · · hn−1 ] where hi is homogeneous
of degree di , so that the multiplicity of p on C is deg gcd(h1 , . . . , hn−1 ). By assumption, di < dn−1
for 1 ≤ i ≤ n − 2. Thus the only way that the gcd of h1 , . . . , hn−1 can have degree greater than
dn−2 is if h1 = · · · = hn−2 = 0, in which case the gcd will have degree dn−1 . Therefore C has a
point of multiplicity greater than dn−2 if and only if some linear combination of the rows of ϕ is
zero except for the last column. This is the same as saying that some linear combination of the
rows of ϕn−2 is zero. From the exact sequence (2.7), this will occur if and only if σ2 = 0.
3.3. A recursive procedure for the generators. While Theorem 3.9 gives the generators as elements of R[w1 , . . . , ws ], we seek the generators as elements of S = R[T1 , . . . , Tn ]. In Corollary 3.17,
we show how the generators in S may be computed. For specific examples, see Example 3.21 and
Example 3.22. The algorithm hinges on the following way to multiply by wi in S.
Proposition 3.15. Assume Data 2.2. For 1 ≤ i ≤ s, let ξi be the ith row of ξ and let ξ i be the
matrix obtained by deleting the ith row from ξ. Let ρi = [ρi1 · · · ρim+1 ] be the n × (m + 1) matrix
whose columns generate the kernel of ξ i . Set pij = ξi ρij ∈ R and set qji = [T1 · · · Tn ]ρij ∈ S∗,1 . Then
(a) For each i, (pi1 , . . . , pim+1 )dm −1+σi = Rdm −1+σi .
(b) For each i and j, we have qji = pij wi in R(E) ⊂ R(F ).
Proof. (a) Let h ∈ R be a form of degree dm − 1 + σi . Then [0 · · · h · · · 0]T ∈ F (with h in the ith
position) has degree dm − 1. By Lemma 2.6, (F/E)dm −1 = 0, and since F/E = coker ξ by (2.7), ξ
is surjective in degree dm − 1. Thus there is u ∈ Rn such that
0
.
.
.
ξu =
h .
.
..
0
(3.16)
This means that u ∈ ker ξ i = im ρi , so u = a1 ρi1 + · · · + am+1 ρim+1 for some a1 , . . . , am+1 ∈ R. Then
the ith row of (3.16) yields
h = ξi u = a1 ξi ρi1 + · · · + am+1 ξi ρim+1 = a1 pi1 + · · · + am+1 pim+1 .
Therefore h ∈ (pi1 , . . . , pim+1 ) for all h ∈ Rdm −1+σi , hence (pi1 , . . . , pim+1 )dm −1+σi = Rdm −1+σi .
17
(b) Since ρij is in the kernel of ξ i (the matrix consisting of all but the ith row of ξ),
0
0
. .
. .
. .
i
i
i
ξρj =
ξi ρj = pj
. .
.. ..
0
0
with pij appearing in the ith row. Thus, using Remark 2.3,
h
qji = T1 · · ·
i
h
Tn ρij = w1 · · ·
i
h
ws ξρij = w1 · · ·
0
.
.
.
i
i
i
ws
pj = pj wi .
.
..
0
Corollary 3.17. Assume Data 2.2. Define the polynomials pij and qji as in Proposition 3.15.
For each α ∈ Ns with α0 = 0 and α · σ ≤ dm+1 − dm , we recursively construct forms hα ∈ S
P
of bidegree (dm+1 − α · σ, ri=1 αi + 1) as follows. Set h0 = gm+1 . For α 6= 0, choose any i
such that αi > 0, and assume we have constructed h(α1 ,...,αi −1,...,αs ) . Write h(α1 ,...,αi −1,...,αs ) =
a1 pi1 + · · · + am+1 pim+1 for some a1 , . . . , am+1 ∈ S (which we may do by Proposition 3.15(a) since
i
. Then the
the x-degree of h(α1 ,...,αi −1,...,αs ) is at least dm + σi ), then set hα = a1 q1i + · · · + am+1 qm+1
images of {hα } in R(E) are uniquely determined and form the x-degree ≥ dm part of a minimal
generating set of Km+1 R(E).
Proof. Let A = {α ∈ Ns | α0 = 0 and α · σ ≤ dm+1 − dm }. By Theorem 3.9, the x-degree ≥ dm
part of a minimal generating set for Km+1 R(E) is given by {gm+1 w α | α ∈ A}. Thus we are done
P
if we can show that hα = gm+1 w α in R(E). We use induction on ℓ :=
αi . If ℓ = 0, then α = 0,
and by definition h0 = gm+1 = gm+1 w0 .
Now suppose ℓ > 0. By construction, there is i for which αi > 0 and a1 , . . . , am+1 ∈ S such that
i
. Then
h(α1 ,...,αi −1,...,αs ) = a1 pi1 + · · · + am+1 pim+1 and hα = a1 q1i + · · · + am+1 qm+1
i
hα = a1 q1i + · · · + am+1 qm+1
= a1 pi1 wi + · · · + am+1 pim+1 wi by Proposition 3.15(b)
= h(α1 ,...,αi −1,...,αs ) wi
= (gm+1 w1α1 · · · wiαi −1 · · · wsαs )wi by induction
= gm+1 wα .
Therefore hα = gm+1 w α for all α ∈ A.
18
When m = 1, the process in Corollary 3.17 may be represented in terms of Sylvester forms (or
Jacobian duals). In fact, when n = 3 (so that there are only two σs) and σ2 = 0, the procedure
in Corollary 3.17 agrees with the one given in [6, Theorem 2.10]. It is also similar to the iterated
Jacobian dual construction in [1, §4], although we use Sylvester forms with respect to multiple
regular sequences pi1 , pi2 . Before stating the result, we give the definition of Sylvester forms.
Let p1 , p2 ∈ R be homogeneous polynomials which form a regular sequence. Consider forms
f, g ∈ S such that f, g ∈ (p1 , p2 )S. Then we may (non-uniquely) write f = f1 p1 + f2 p2 and
g = g1 p1 + g2 p2 . We will, by abuse of notation, refer to the determinant
#
"
f1 g1
sylp1 ,p2 (f, g) = det
f2 g2
as the Sylvester form of f, g with respect to p1 , p2 . The Sylvester form sylp1 ,p2 (f, g) is not uniquely
determined, but its image in S/(f, g) is unique ([17, Proposition 3.8.1.6]).
Proposition 3.18. Assume Data 2.2 with m = 1. Define pi1 , pi2 as in Proposition 3.15. Then
for all i, the polynomials pi1 , pi2 form a regular sequence, and for any form h ∈ S with x-degree
≥ d1 − 1 + σi , the image of the Sylvester form sylpi1 ,pi2 (g1 , h) in R(E) = S/(g1 ) is a nonzero scalar
multiple of hwi . In particular, the polynomials hα of Corollary 3.17 may be defined, up to scalar
multiples, as hα = sylpi ,pi (g1 , h(α1 ,...,αi −1,...,αs ) ) for any i such that αi > 0.
1
2
Proof. That pi1 , pi2 form a regular sequence is immediate from Proposition 3.15(a). Recall the
matrix ρi from Proposition 3.15. Let δ1i , δ2i be the degrees of the columns of ρi . We claim that
δ1i + δ2i = d1 − σi . First, since ξ i is all but the ith row of ξ, there is a surjection coker ξ ։ coker ξ i .
By Lemma 2.6(a,b), coker ξ is Artinian, so coker ξ i is Artinian as well. Since R is a two-dimensional
regular ring, there is a free resolution
ξi
ρi
→
0 → R(−δ1i ) ⊕ R(−δ2i ) −→ Rn −
M
R(σj ) → coker ξ i → 0.
j6=i
By computing the Hilbert polynomial and using that coker ξ i is Artinian, we see that δ1i + δ2i =
P
P
σj = d1 by Lemma 2.6(e), hence δ1i + δ2i = d1 − σi .
j6=i σj . But
Now let h ∈ R(E) be homogeneous of x-degree ≥ d1 − 1 + σi . Then by Proposition 3.15(a), there
are a1 , a2 ∈ R(E) such that h = a1 pi1 + a2 pi2 . Recall that ρi generates the kernel of ξ i , the matrix
obtained by removing the ith row of ξ. Since ξϕ1 = 0 from (2.7), ϕ1 is in the kernel of ξ i , therefore
ϕ1 = ρi λ for some 2 × 1 column vector λ with entries in R. Then
i
h
i
i
p1 p2 λ = ξi ρi λ = ξi ϕ1 = 0.
Since pi1 , pi2 are a regular sequence, this means
λ=
"
−cpi2
cpi1
19
#
for some c ∈ R. Thus
h
(3.19)
g1 = T1 · · ·
i
h
Tn ϕ1 = T1 · · ·
i
i
h
Tn ρi λ = q1i q2i λ = c(pi1 q2i − pi2 q1i ).
Now pij has x-degree σi + δj , while qji has x-degree δj . Because we have shown that δ1i + δ2i = d1 − σi ,
we have degx (pi1 q2i − pi2 q1i ) = σi + δ1i + δ2i = d1 = degx g1 , hence c ∈ k. We cannot have c = 0, or we
would get ϕ1 = ρi λ = ρi · 0 = 0, which is impossible. Therefore c is a unit. Since h = a1 pi1 + a2 pi1
and g1 = (cq2i )pi1 + (−cq1i )pi2 by (3.19), the definition of the Sylvester form yields
#
"
cq2i a1
= c(a1 q1i + a2 q2i ).
sylpi ,pi (g1 , h) = det
i
1 2
−cq1 a2
But by Proposition 3.15(b), a1 q1i + a2 q2i = a1 pi1 wi + a2 pi2 wi = hwi , so sylpi ,pi (g1 , h) = chwi .
1
2
Because we showed in the proof of Corollary 3.17 that hα = h(α1 ,...,αi −1,...,αs ) wi in R(E), the final
claim follows.
3.4. Some examples.
Example 3.20 (compare [22, Theorem 3.6]). Assume I is almost linearly presented, meaning
Pn−2
d1 = · · · = dn−2 = 1 and dn−1 = c ≥ 1. First, Lemma 2.6(e) gives σ1 + σ2 = i=1
di = n − 2. By
Theorem 2.9, the modules E = En−2 and M = mσ1 (σ1 ) ⊕ mσ2 (σ2 ) agree in degree ≥ 0, but they
are both generated in degree 0, so E ∼
= M . Thus by Proposition 2.5, after an automorphism of
S = R[T1 , . . . , Tn ], the ideal of equations of R(E) is generated by the 2 × 2 minors of the matrix
#
"
Tσ1
Tσ1 +2 Tσ1 +3 · · · Tn−1
x0 T1 T2 · · ·
.
x1 T2 T3 · · · Tσ1 +1 Tσ1 +3 Tσ1 +4 · · ·
Tn
In particular, there are n − 2 generators of bidegree (1, 1) and 12 (n − 2)(n − 3) generators of bidegree
(0, 2). The ideal K of equations of R(I) is generated by these equations, plus a lift of a minimal
generating set of KR(E).
By (3.11) with m = n − 2, this ideal is KR(E) = gn−1 R(F )≥−c,∗ where F = R(σ1 ) ⊕ R(σ2 ). We
may compute the generators of this ideal directly from Theorem 3.7: they are just the elements of
gn−1 Ac ∪ gn−1 Bc . The set gn−1 Ac consists of the minimal generators of KR(E) having x-degree
> 0, and equals
{gn−1 w1α1 w2α2 | α1 σ1 + α2 σ2 ≤ c − 1} if σ2 > 0
{gn−1 w1α1 | α1 σ1 ≤ c − 1} if σ2 = 0.
We can translate from w’s to T ’s by the equations Tj+1 = xσ0 1 −j xj1 w1 and Tσ1 +j+2 = xσ0 2 −j xj1 w2
(see Proposition 2.5). The bidegrees of these minimal generators are listed in Corollary 3.13.
We can also compute the minimal generators of x-degree 0, which are the elements of gn−1 Bc .
To do this, we must first determine the set Ωc of Definition 3.5. There are two cases. First, suppose
σ2 = 0. Then r = 1, so
Ωc = Ωc,1 = {(β1 , 0) | c ≤ β1 σ1 < c + σ1 }.
20
This is just Ωc = {(k, 0)} where k = ⌈ σc1 ⌉. Let ℓ = kσ1 − c; note that 0 ≤ ℓ < σ1 . Then gn−1 Bc
consists of ℓ + 1 elements of bidegree (0, k + 1), namely
{gn−1 x0ℓ−j xj1 w1k | 0 ≤ j ≤ ℓ}.
Second, suppose σ2 > 0. Then r = 2, so Ωc = Ωc,1 ∪ Ωc,2 . As above, Ωc,1 = {(k, 0)} where
k = ⌈ σc1 ⌉. Again, setting ℓ = kσ1 − c, gn−1 Bc has ℓ + 1 elements of bidegree (0, k + 1), namely
{gn−1 x0ℓ−j xj1 w1k | 0 ≤ j ≤ ℓ}.
In this case, however, we also have
Ωc,2 = {(α1 , α2 ) | α2 > 0 and c ≤ α1 σ1 + α2 σ2 < c + σ2 }.
2
To have (α1 , α2 ) ∈ Ωc,2 , we must have α1 σ1 < c + σ2 , thus α1 < ⌈ c+σ
σ1 ⌉. Once we have chosen α1 ,
1 σ1
the condition c − α1 σ1 ≤ α2 σ2 < c − α1 σ1 + σ2 forces α2 = ⌈ c−α
σ2 ⌉. Hence (α1 , α2 ) ∈ Ωc,2 come
c−iσ1
2
in pairs of the form (i, v(i)) where 0 ≤ i < ⌈ c+σ
σ1 ⌉ and v(i) = ⌈ σ2 ⌉. Let ℓ(i) = iσ1 + v(i)σ2 − c,
so that 0 ≤ ℓ(i) < σ2 . Therefore the remaining part of gn−1 Bc consists of ℓ(i) + 1 generators of
2
bidegree (0, i + v(i) + 1) for each 0 ≤ i < ⌈ c+σ
σ1 ⌉, namely
ℓ(i)−j j i v(i)
{gn−1 x0
x1 w1 w2
c + σ2
}.
|0≤i<
σ1
Example 3.21 (compare [6, Theorem 2.10] and [19, Corollary 3.10]). Assume n = 3 and assume
that the entries of the first column of ϕ are linearly dependent. Thus after row operations on ϕ,
we have
γ1 δ1
ϕ = γ2 δ2 .
0
δ3
Since ϕ1 = [γ1 γ2 0]T , the resolution (2.8) becomes
ξT
ϕT
1
0 → R(−d1 ) ⊕ R −→ R3 −−→
R(d1 ) → Ext2R (F/E, R) → 0
where the image of ξ T is the kernel of [γ1 γ2 0], meaning
"
#
−γ2 γ1 0
ξ=
.
0
0 1
In particular, σ1 = d1 , σ2 = 0, s = 2, and r = 1. Then by Theorem 3.9, the minimal generators
of K besides g1 having x-degree ≥ d1 are g2 w1α1 for α1 ∈ N such that α1 σ1 ≤ d2 − d1 , or in other
words, 0 ≤ α1 ≤
d2
d1
− 1.
21
To compute these generators, we will use Corollary 3.17. The matrix ρ1 defined in Proposition 3.15
is the syzygy matrix of the matrix obtained by deleting the first row of ξ, which is [0 0 1], thus
1 0
ρ1 = 0 1 .
0 0
Following Proposition 3.15, we have
i
h
i h
i
h
p11 p12 = −γ2 γ1 0 ρ1 = −γ2 γ1
and
h
q11
q21
i
i
i
h
h
= T1 T2 T3 ρ1 = T1 T2 .
Now Corollary 3.17 says that the generators of K can be computed recursively. Let h(0,0) = g2 ,
and for any 1 ≤ α1 ≤
d2
d1
− 1, write h(α1 −1,0) = a1 p11 + a2 p12 = −a1 γ2 + a2 γ1 , then set h(α1 ,0) =
a1 q11 + a2 q21 = a1 T1 + a2 T2 . These h(α1 ,0) , together with g1 , are all the minimal generators of
K having x-degree at least d1 . We will determine the minimal generators of x-degree d1 − 1 in
Example 4.10.
Example 3.22 (compare [2, Proposition 3.4] and [6, Theorem 5.4]). Assume n = 3 and d1 = 2.
We may assume that the entries of the first column of ϕ are linearly independent, for the other
case was considered in Example 3.21. Then after row operations, we may write
2
x0 δ1
ϕ = x0 x1 δ2 .
x21
δ3
Then the resolution (2.8) is
ϕT
ξT
1
0 → R(−1) ⊕ R(−1) −→ R3 −−→
R(2) → Ext2R (F/E, R) → 0
where
ξ=
"
−x1
0
x0
0
−x1 x0
#
.
In particular, σ1 = σ2 = 1, s = 2, and r = 0. By Theorem 3.9, the minimal generators of K besides
g1 having x-degree ≥ 2 are g2 w1α1 w2α2 for α1 , α2 ∈ N such that α1 + α2 ≤ d2 − 2.
Now let us use Corollary 3.17 to compute the precise generating set. Let ρ1 and ρ2 be the syzygy
matrices of the matrices obtained by deleting the first and second row of ξ, respectively. That is,
0 x0
1 0
ρ1 = 0 x0 and ρ2 = 0 x1 .
1
0 x1
0
Therefore
h
i h
i
h
i
p11 p12 = −x1 x0 0 ρ1 = −x1 x20
22
h
h
q11
q12
i
i
h
i h
p21 p22 = 0 −x1 x0 ρ2 = x0 −x21
i
i
h
i h
q21 = T1 T2 T3 ρ1 = T1 x0 T2 + x1 T3
i
i
h
i h
q22 = T1 T2 T3 ρ2 = T3 x0 T1 + x1 T2 .
h
Now we are ready to apply the algorithm from Corollary 3.17. Let h(0,0) = g2 . Suppose we have
computed h(α1 ,α2 ) for some α1 , α2 with α1 + α2 < d2 − 2. Write h(α1 ,α2 ) = −a1 x1 + a2 x20 =
b1 x0 − b2 x21 . (We can do this since degx h(α1 ,α2 ) ≥ 2.) Set h(α1 +1,α2 ) = a1 T1 + a2 (x0 T2 + x1 T3 ) and
h(α1 ,α2 +1) = b1 T3 + b2 (x0 T1 + x1 T2 ). By Corollary 3.17, the polynomials h(α1 ,α2 ) constructed in this
way coincide with the generators g2 w1α1 w2α2 .
4. The case of plane curves
Let n = 3, so that I = (f1 , f2 , f3 ) and C is a plane curve. In this case, we can actually
compute a generating set for each slice Ki,∗ with i ≥ d1 − 1, in Theorem 4.8. In particular, we get
a generating set of K≥d1 −1,∗ by combining the generators of Kd1 −1,∗ from Theorem 4.8 with the
minimal generators of K in x-degree ≥ d1 from Theorem 3.9. Unfortunately, the generating set for
Ki,∗ in Theorem 4.8 is usually not minimal.
Observe that K1 = (g1 ) : m∞ = (g1 ), meaning R(E1 ) = S(E1 ) = S/(g1 ). Thus to determine K,
it suffices to compute K/K1 = KR(E1 ) = K2 R(E1 ). To accomplish this, we will first determine
the k[T1 , T2 , T3 ]-module structure of R(F )i,∗ . This is carried out in the next two lemmas and
Proposition 4.4, where we show that R(F )i,∗ is generated as a k[T1 , T2 , T3 ]-module in T -degrees 0
and 1 for all i ≥ −1.
Lemma 4.1. Let ζ be an n × (n − 1) matrix and let η be an (n − 1) × n matrix with coefficients
in m, such that ηζ = f In−1 for some f ∈ R. Assume that grade In−1 (ζ) = 2. Then f ∈ In−1 (ζ).
Proof. Let χ be the 1 × n row matrix consisting of the signed minors of ζ. By the Hilbert-Burch
theorem, im(χT ) = ker(ζ T ). By the Hilbert syzygy theorem, coker η has minimal free resolution
ϑ
η
→ Rn−1 → coker η → 0
0→F −
→ Rn −
where F is a free R-module. We have f ∈ annR (coker η), so rank(coker η) = 0. Using the additivity
of rank, we see that rank F = 1, hence ϑ is an n × 1 column matrix.
Now η(ζη − f In) = ηζη − f η = (ηζ − f In−1 )η = 0, so im(ζη − f In) ⊂ ker η = im ϑ. Thus there is
a row matrix λ : Rn → R such that ζη − f In = ϑλ. Then ϑ(λζ) = ζηζ − f ζ = ζ(ηζ − f In−1 ) = 0.
Since ϑ is injective, this means λζ = 0. Hence im(λT ) ⊂ ker(ζ T ) = im(χT ), so there is g ∈ R such
that λ = gχ. We conclude that ζη − gϑχ = f In .
Set
h
i
ζ ′ = ζ −gϑ and η ′ =
23
" #
η
χ
.
These are both n × n matrices, and ζ ′ η ′ = ζη − gϑχ = f In . It follows that η ′ ζ ′ = f In . But
"
#
ηζ −gηϑ
′ ′
ηζ =
,
χζ −gχϑ
so we get f = −gχϑ = χ(−gϑ). Therefore f ∈ im χ = In−1 (ζ).
In order to compare R(E) = R(E1 ) and R(F ) in high T -degrees, we first compare them in
bidegree (−1, 2). This will be used as the base case in the proof of Proposition 4.4.
Lemma 4.2. Assume Data 2.2 with n = 3 and m = 1. Then R(F )−1,2 = R(E)0,1 R(F )−1,1 .
Proof. Recall that R(F ) = R[w1 , w2 ], with deg wi = (−σi , 1). Then
R(F )−1,1 = Rσ1 −1 w1 ⊕ Rσ2 −1 w2
R(F )−1,2 = R2σ1 −1 w12 ⊕ Rσ1 +σ2 −1 w1 w2 ⊕ R2σ2 −1 w22 .
Because dimk Rℓ = ℓ + 1 for any ℓ, we get dimk R(F )−1,1 = σ1 + σ2 = d1 , while dimk R(F )−1,2 =
3σ1 + 3σ2 = 3d1 . Set T = [T1 T2 T3 ], considered as a function T : R(F )3−1,1 → R(F )−1,2 . Since
R(E)0,1 = kT1 + kT2 + kT3 , we have R(E)0,1 R(F )−1,1 = im(T). It will suffice to show that T
is injective, for then dimk (R(E)0,1 R(F )−1,1 ) = 3(dimk R(F )−1,1 ) = 3d1 = dimk R(F )−1,2 , which
implies the equality R(E)0,1 R(F )−1,1 = R(F )−1,2 .
Suppose [p1 p2 p3 ]T ∈ ker(T), with pi ∈ R(F )−1,1 . From Remark 2.3, [w1 w2 ]ξ = [T1 T2 T3 ].
Thus
p1
i p1
h
0 = T p2 = w1 w2 ξ p2
p3
p3
which we can rewrite as
h
i
p1 p2 p3 ξ
T
"
w1
w2
#
= 0.
Since w1 , w2 is a regular sequence in R(F ), there is f ∈ R(F )d1 −1,0 = Rd1 −1 such that
i
i
h
h
(4.3)
p1 p2 p3 ξ T = −f w2 f w1 .
For a, b ∈ k, denote by pi (a, b) ∈ R the evaluation of pi at (w1 , w2 ) = (a, b). Let
#
"
p1 (0, −1) p2 (0, −1) p3 (0, −1)
η=
.
p1 (1, 0)
p2 (1, 0)
p3 (1, 0)
Then (4.3) gives
ηξ T =
#
"
−f · (−1) f · 0
−f · 0
f ·1
= f I2 .
By Lemma 4.1, f ∈ I2 (ξ T ).
Recall the exact sequence (2.8), which in this setting is
ξT
ϕT
→ R(d1 ) → Ext2R (F/E, R) → 0.
0 → R(−σ1 ) ⊕ R(−σ2 ) −→ R3 −−m
24
T
By the Hilbert-Burch theorem, I2 (ξ T ) = I1 (ϕT
m ), therefore f ∈ I2 (ξ ) = I1 (ϕm ). But f ∈ Rd1 −1 ,
while I1 (ϕm ) is generated in degree d1 , therefore f = 0. Hence by (4.3), [p1 p2 p3 ]ξ T = 0, so
[p1 p2 p3 ]T ∈ ker ξ. Recall from (2.7) that ker ξ = im ϕm , which is generated in degree d1 . However,
degx pi ≤ σ2 − 1 < d1 , so we must have pi = 0, hence T is injective as claimed.
The next proposition shows that R(M ) is generated as a module over R(E)0,∗ = k[T1 , T2 , T3 ] in
T -degrees 0 and 1.
Proposition 4.4. Assume Data 2.2 with n = 3 and m = 1. Let U = R(E)0,∗ = k[T1 , T2 , T3 ].
Then R(F )i,∗ = U R(F )i,0 + U R(F )i,1 for all i ≥ −1.
Proof. We first show by induction on j that
R(F )−1,j = Uj R(F )−1,0 + Uj−1 R(F )−1,1 .
For j = 0 and j = 1 this is clear. For j ≥ 2,
R(F )−1,j = R(F )−1,2 R(F )0,j−2
= U1 R(F )−1,1 R(F )0,j−2 by Lemma 4.2
= U1 R(F )−1,j−1
= U1 (Uj−1 R(F )−1,0 + Uj−2 R(F )−1,1 ) by induction
= Uj R(F )−1,0 + Uj−1 R(F )−1,1 .
This proves that R(F )−1,∗ = U R(F )−1,0 +U R(F )−1,1 . Multiplying both sides by Ri+1 = R(F )i+1,0
gives R(F )i,∗ = U R(F )i,0 + U R(F )i,1 .
By Proposition 4.4, the U -module R(M ) is generated in T -degrees 0 and 1. In T -degree 0,
R(M )∗,0 is just the ring R = k[x0 , x1 ]. To understand R(M )∗,1 as a U -module, recall that R(E)∗,1 =
E and R(M )∗,1 = M = F≥0 . Therefore R(M )∗,1 /R(E)∗,1 = (F/E)≥0 . The free resolution of F/E
was given in Lemma 2.6(a):
(4.5)
ϕ1
ξ
0 → R(−d1 ) −→ R3 −
→ R(σ1 ) ⊕ R(σ2 ) → F/E → 0.
Remark 4.6. Assume Data 2.2 with n = 3 and m = 1. Let HF/E (i) be the Hilbert function of
F/E. Then for all −1 ≤ i ≤ d1 − 1, we have HF/E (i) = d1 − i − 1.
Proof. Recall from Lemma 2.6(e) that σ1 + σ2 = d1 . Then this is a straightforward computation
from (4.5).
For −1 ≤ i ≤ d1 − 1, we have the following refinement of Proposition 4.4.
Proposition 4.7. Assume Data 2.2 with n = 3 and m = 1. Let U = R(E)0,∗ = k[T1 , T2 , T3 ].
Then for all −1 ≤ i ≤ d1 − 1, there is an isomorphism R(F )i,∗ ∼
= U i+1 ⊕ U d1 −i−1 (−1) under which
R(E)i,∗ ∼
= U i+1 .
25
i
Proof. The vector space R(F )i,0 is generated by xi0 , xi−1
0 x1 , . . . , x1 , so it has dimension i + 1. By
Remark 4.6, dimk (Fi /Ei ) = d1 − i − 1. Thus there are elements pi,1 , . . . , pi,d1 −i−1 ∈ Fi = R(F )i,1
whose images in Fi /Ei form a basis. Since Ei = R(E)i,1 = U1 R(E)i,0 = U1 R(F )i,0 , the vector
space R(F )i,1 /U1 R(F )i,0 is generated by the images of p1 , . . . , pd1 −i−1 . Now consider the map
U i+1 ⊕ U d1 −i−1 (−1) → R(F )i,∗ given by [xi0 · · · xi1 p1 · · · pd1 −i−1 ]. Then the image of this map
contains R(F )i,0 , and it contains R(F )i,1 . Since R(F )i,∗ = U R(F )i,0 +U R(F )i,1 by Proposition 4.4,
it is a surjective map, and it sends U i+1 to R(E)i,∗ = U xi0 + · · · + U xi1 . To show that it is
actually an isomorphism U i+1 ⊕ U d1 −i−1 (−1) ∼
= R(F )i,∗ , it suffices to show that the U -modules
U i+1 ⊕ U d1 −i−1 (−1) and R(F )i,∗ have the same Hilbert function.
The Hilbert function of the former is
j +2
j+1
HU i+1⊕U d1 −i−1 (−1) (j) = (i + 1)
+ (d1 − i − 1)
2
2
j+1
= (i + 1)(j + 1) + d1
.
2
To compute the Hilbert function of R(F )i,∗ , note that a basis for R(F )i,j consists of monomials
xk0 xℓ1 w1j−a w2a where k + ℓ = (j − a)σ1 + aσ2 + i. Thus
j
X
((j − a)σ1 + aσ2 + i + 1)
HR(F )i,∗ (j) =
a=0
= σ1
j
X
(j − a) + σ2
j
X
a + (i + 1)(j + 1)
a=0
a=0
j+1
j+1
= σ1
+ σ2
+ (i + 1)(j + 1)
2
2
j+1
= (i + 1)(j + 1) + d1
2
= HU i+1⊕U d1 −i−1 (−1) (j).
Therefore R(F )i,∗ ∼
= U i+1 ⊕ U d1 −i−1 (−1).
We may now compute the generators of Ki,∗ for any i ≥ d1 − 1.
Theorem 4.8. Assume Data 2.2 with n = 3 and m = 1. Fix i ≥ d1 − 1 and set c = d2 − i. Let
U = k[T1 , T2 , T3 ]. Let Ωc and Bc be as defined in Definition 3.5. For each 0 ≤ ℓ ≤ d1 − 2, choose
elements pℓ,1 , . . . , pℓ,d1 −ℓ−1 ∈ Fℓ = R(F )ℓ,1 whose images in Fℓ /Eℓ form a basis. Then
1
1
{g1 xi−d
, . . . , g1 xi−d
} ∪ g2 Bc ∪ {g2 pα·σ−c,j wα | α ∈ Ωc , 1 ≤ j ≤ d1 − α · σ + c − 1}
0
1
is a generating set (not necessarily minimal) for Ki,∗ as a U -module.
Proof. Since R(E) = R(E1 ) ∼
= S/(g1 ), and g1 has x-degree d1 , we just need to show that
g2 Bc ∪ {g2 pα·σ−c,j w α | α ∈ Ωc , 1 ≤ j ≤ d1 − α · σ + c − 1}
26
is a minimal generating set for Ki,∗ R(E). We showed in (3.11) that Ki,∗ R(E) = g2 R(F )−c,∗ . From
Theorem 3.7(b), we know that g2 Bc is a minimal generating set for Ki,∗ R(E) as an R(M )0,∗ -module.
Recalling the definition of Bc , this means that
X
X
(4.9)
Ki,∗ R(E) = g2
R(M )0,∗ (x0 , x1 )α·σ−c w α = g2
R(M )α·σ−c,∗ wα .
α∈Ωc
α∈Ωc
By Proposition 4.4, R(M )α·σ−c,∗ = U R(M )α·σ−c,0 + U R(M )α·σ−c,1 . The vector space R(M )α·σ−c,0
α·σ−c−j j
x1
is spanned by {x0
| 0 ≤ j ≤ α·σ−c}, while the vector space R(M )α·σ−c,1 /U1 R(M )α·σ−c,0 =
(F/E)α·σ−c is spanned by the images of {pα·σ−c,j | 1 ≤ j ≤ d1 −α·σ +c−1}. Therefore R(M )α·σ−c,∗
is generated as an U -module by the union of these two sets. Then by (4.9), Ki,∗ R(E) is generated
as an U -module by the union of
α·σ−c−j j α
x1 w
g2 Bc = {g2 x0
| α ∈ Ωc , 0 ≤ j ≤ α · σ − c}
and
{g2 pα·σ−c,j w α | α ∈ Ωc , 1 ≤ j ≤ d1 − α · σ + c − 1}.
Example 4.10 (continuation of Example 3.21). Assume n = 3 and assume that the entries of the
first column of ϕ are linearly dependent. In Example 3.21, we computed the minimal generators of
K with x-degree ≥ d1 . As for the generators having x-degree d1 − 1, we claim that the generating
set given in Theorem 4.8 is minimal. Note that Ωc (where c = d2 − d1 + 1) in this case consists
of only one element, namely (α1 , 0) where α1 = ⌈ dd21 ⌉ (as in the σ2 = 0 case of Example 3.20).
Therefore the generating set in Theorem 4.8 is just
g2 w1α1 {xα0 1 σ1 −c , . . . , xα1 1 σ1 −c , pα1 σ1 −c,1 , . . . , pα1 σ1 −c,d1 −α1 σ1 +c−1 }.
By Proposition 4.7, {xα0 1 σ1 −c , . . . , xα1 1 σ1 −c , pα1 σ1 −c,1 , . . . , pα1 σ1 −c,d1 −α1 σ1 +c−1 } is a minimal generating set for R(F )α1 σ1 −c,∗ as a U -module. Thus the generating set from Theorem 4.8 is a minimal
generating set for K in x-degree d1 − 1.
Example 4.11 (continuation of Example 3.22). Assume n = 3 and d1 = 2, and assume that the
entries of the first column of ϕ are linearly independent, as in Example 3.22. To determine the
minimal generators of K having x-degree 1, first note that Ωd2 −1 = {(α1 , α2 ) | α1 + α2 = d2 − 1},
therefore g2 Bd2 −1 = {g2 w1α1 w2α2 | α1 + α2 = d2 − 1}. We claim that this is a minimal generating
set of K1,∗ . Minimality will follow from Theorem 3.7, so we just have to show it is generating. By
Theorem 4.8, the only other generators of K1,∗ have the form g2 pα1 +α2 −d2 +1,j w1α1 w2α2 for (α1 , α2 ) ∈
Ωd2 −1 . But since (α1 , α2 ) ∈ Ωd2 −1 , this means α1 + α2 − d2 + 1 = 0 so the generators have the form
g2 p0,1 w1α1 w2α2 , where p0,1 is a lift of a generator of F1 /E1 . Since α1 + α2 = d2 − 1,
g2 p0,1 w1α1 w2α2 ∈ g2 R(F )0,1 (w1 , w2 )d2 −1 ⊂ g2 R(F )−1,2 (w1 , w2 )d2 −2 .
But R(F )−1,2 = U1 R(F )−1,1 by Proposition 4.4, and R(F )−1,1 is generated by w1 and w2 , thus
g2 p0,1 w1α1 w2α2 ∈ g2 R(F )−1,2 (w1 , w2 )d2 −2 ⊂ U1 (g2 (w1 , w2 )d2 −1 ).
27
Therefore g2 p0,1 w1α1 w2α2 is in the U -submodule generated by g2 Bd2 −1 . Hence g2 p0,1 w1α1 w2α2 is not
part of a minimal generating set, so the only minimal generators are g2 Bd2 −1 .
Combining this with what we already showed in Example 3.22, we conclude that all the minimal
generators of K besides g1 having x-degree ≥ 1 are g2 w1α1 w2α2 for α1 , α2 ∈ N such that α1 + α2 ≤
d2 − 1. Since the only generator of K in degree 0 is the equation of the fiber ring, in bidegree (0, d),
we have all the minimal generators of K.
Example 4.12. Unlike in the previous examples, it is not always possible to determine which, or
even how many, of the generators in Theorem 4.8 are minimal, from only the numerical data of d1 ,
d2 , σ1 , and σ2 . For example, take
x40
ϕ = x20 x21
x41
x71
0 .
x70
Then d1 = 4, d2 = 7, and σ1 = σ2 = 2. A computation in Macaulay2 shows that K3,∗ has seven
minimal generators: three in degree 3, and four in degree 4. On the other hand, consider
4
x0 + x30 x1 x71
ϕ = x20 x21
0 .
x41
x70
Then, as before, we have d1 = 4, d2 = 7, and σ1 = σ2 = 2. But in this case, K3,∗ has only six
minimal generators: three in degree 3, and three in degree 4.
Acknowledgments
This work was carried out as part of the author’s doctoral thesis research. The author is very
grateful to his advisor, Claudia Polini, for suggesting the problem and for helpful discussions, as
well as her comments on an early version of the paper.
References
[1] J. A. Boswell and V. Mukundan. Rees algebras and almost linearly presented ideals. Preprint, arXiv:1505.04851.
[2] L. Busé. On the equations of the moving curve ideal of a rational algebraic plane curve. J. Algebra, 321(8):2317–
2344, 2009.
[3] L. Busé, M. Chardin, and A. Simis. Elimination and nonlinear equations of Rees algebras. J. Algebra, 324(6):1314–
1333, 2010.
[4] L. Busé and J.-P. Jouanolou. On the closed image of a rational map and the implicitization problem. J. Algebra,
265(1):312–357, 2003.
[5] T. Cortadellas Benı́tez and C. D’Andrea. Rational plane curves parameterizable by conics. J. Algebra, 373:453–
480, 2013.
[6] T. Cortadellas Benı́tez and C. D’Andrea. Minimal generators of the defining ideal of the Rees algebra associated
with a rational plane parametrization with µ = 2. Canad. J. Math., 66(6):1225–1249, 2014.
[7] D. Cox, J. W. Hoffman, and H. Wang. Syzygies and the Rees algebra. J. Pure Appl. Algebra, 212(7):1787–1796,
2008.
28
[8] D. Cox, A. R. Kustin, C. Polini, and B. Ulrich. A study of singularities on rational curves via syzygies. Mem.
Amer. Math. Soc., 222(1045):x+116, 2013.
[9] D. A. Cox. The moving curve ideal and the Rees algebra. Theoret. Comput. Sci., 392(1-3):23–36, 2008.
[10] D. A. Cox, T. W. Sederberg, and F. Chen. The moving line ideal basis of planar rational curves. Comput. Aided
Geom. Design, 15(8):803–827, 1998.
[11] D. Eisenbud. The geometry of syzygies, volume 229 of Graduate Texts in Mathematics. Springer-Verlag, New
York, 2005.
[12] D. Eisenbud and J. Harris. On varieties of minimal degree (a centennial account). In Algebraic geometry, Bowdoin, 1985 (Brunswick, Maine, 1985), volume 46 of Proc. Sympos. Pure Math., pages 3–13. Amer. Math. Soc.,
Providence, RI, 1987.
[13] D. Eisenbud, C. Huneke, and B. Ulrich. What is the Rees algebra of a module? Proc. Amer. Math. Soc.,
131(3):701–708, 2003.
[14] J. Hong, A. Simis, and W. V. Vasconcelos. The equations of almost complete intersections. Bull. Braz. Math.
Soc. (N.S.), 43(2):171–199, 2012.
[15] J. Hong, A. Simis, and W. V. Vasconcelos. Extremal Rees algebras. J. Commut. Algebra, 5(2):231–267, 2013.
[16] M. R. Johnson. Second analytic deviation one ideals and their Rees algebras. J. Pure Appl. Algebra, 119(2):171–
183, 1997.
[17] J.-P. Jouanolou. Aspects invariants de l’élimination. Adv. Math., 114(1):1–174, 1995.
[18] V. Kodiyalam. Integrally closed modules over two-dimensional regular local rings. Trans. Amer. Math. Soc.,
347(9):3551–3573, 1995.
[19] A. Kustin, C. Polini, and B. Ulrich. The bi-graded structure of symmetric algebras with applications to rees
rings. J. Algebra. To appear, arXiv:1301.7106.
[20] A. R. Kustin, C. Polini, and B. Ulrich. Blowups and fibers of morphisms. Preprint, arXiv:1410.3172.
[21] A. R. Kustin, C. Polini, and B. Ulrich. Divisors on rational normal scrolls. J. Algebra, 322(5):1748–1773, 2009.
[22] A. R. Kustin, C. Polini, and B. Ulrich. Rational normal scrolls and the defining equations of Rees algebras. J.
Reine Angew. Math., 650:23–65, 2011.
[23] N. P. H. Lan. On Rees algebras of linearly presented ideals. J. Algebra, 420:186–200, 2014.
[24] K.-N. Lin and C. Polini. Rees algebras of truncations of complete intersections. J. Algebra, 410:36–52, 2014.
[25] S. Morey. Equations of blowups of ideals of codimension two and three. J. Pure Appl. Algebra, 109(2):197–211,
1996.
[26] S. Morey and B. Ulrich. Rees algebras of ideals with low codimension. Proc. Amer. Math. Soc., 124(12):3653–
3661, 1996.
[27] T. Sederberg, R. Goldman, and H. Du. Implicitizing rational curves by the method of moving algebraic curves.
J. Symbolic Comput., 23(2-3):153–175, 1997. Parametric algebraic curves and applications (Albuquerque, NM,
1995).
[28] T. W. Sederberg and F. Chen. Implicitization using moving curves and surfaces. In Proceedings of the 22Nd
Annual Conference on Computer Graphics and Interactive Techniques, SIGGRAPH ’95, pages 301–308, New
York, NY, USA, 1995. ACM.
[29] A. Simis, B. Ulrich, and W. V. Vasconcelos. Rees algebras of modules. Proc. London Math. Soc. (3), 87(3):610–
646, 2003.
[30] I. Swanson and C. Huneke. Integral closure of ideals, rings, and modules, volume 336 of London Mathematical
Society Lecture Note Series. Cambridge University Press, Cambridge, 2006.
[31] H. H. Tài. On the Rees algebra of certain codimension two perfect ideals. Manuscripta Math., 107(4):479–501,
2002.
29
[32] B. Ulrich and W. V. Vasconcelos. The equations of Rees algebras of ideals with linear presentation. Math. Z.,
214(1):79–92, 1993.
Department of Mathematics, University of Notre Dame, Notre Dame, IN 46556
E-mail address: [email protected]
30
| 0math.AC
|
COMMUTATIVITY THEOREMS FOR GROUPS AND SEMIGROUPS
arXiv:1706.00381v2 [math.GR] 22 Nov 2017
FRANCISCO ARAÚJO AND MICHAEL KINYON∗
Abstract. In this note we prove a selection of commutativity theorems for various classes
of semigroups. For instance, if in a separative or completely regular semigroup S we have
xp y p = y p xp and xq y q = y q xq for all x, y ∈ S where p and q are relatively prime, then
S is commutative. In a separative or inverse semigroup S, if there exist three consecutive
integers i such that (xy)i = xi y i for all x, y ∈ S, then S is commutative. Finally, if S is a
separative or inverse semigroup satisfying (xy)3 = x3 y 3 for all x, y ∈ S, and if the cubing
map x 7→ x3 is injective, then S is commutative.
1. Introduction
Broadly speaking, a commutativity theorem in group theory is any result concluding that
a group is commutative, i.e. abelian. Perhaps the best known example is the following
standard exercise, usually given to students at the beginning of their study of group theory:
If G is a group satisfying x2 = 1 for all x ∈ G, then G is commutative.
Commutativity theorems can sometimes be extended to various classes of semigroups
properly containing groups. For instance, a semigroup S is cancellative if it satisfies the
conditions xy = xz =⇒ y = z and yx = zx =⇒ y = z for all x, y, z. Every finite
cancellative semigroup is a group; the positive integers under addition provide an example
of a cancellative semigroup which is not a group.
The exercise above extends easily to cancellative semigroups once we reinterpret the condition “x2 = 1”. Since we do not wish to assume the existence of an identity element, we
replace the condition with “x3 = x”, which is clearly equivalent to x2 = 1 in groups.
Proposition 1.1. If S is a cancellative semigroup satisfying x3 = x for all x ∈ S, then S is
commutative (and in fact, is a group satisfying x2 = 1 for all x ∈ S).
The parenthetical part of the assertion suggests the proof: if x3 = x, then x3 y = xy and
so cancelling gives x2 y = y for all x, y ∈ S. Dually, yx2 = y for all x, y ∈ S. Thus for all
x, y ∈ S, x2 = x2 y 2 = y 2. This constant, which we denote by 1, is an identity element.
Hence x3 = x = x1, and cancelling gives x2 = 1 for all x ∈ S. Therefore S is a group and
we have reduced the problem to the original exercise.
Still using this elementary example to illustrate our point, further extensions of the result
are possible. A semigroup is separative if it satisfies the conditions xy = xx & yx = yy =⇒
x = y and xy = yy & yx = xx =⇒ x = y ([7], Def. II.6.2, p. 51). Every cancellative
semigroup is evidently separative.
We also need the notion of a semilattice of semigroups. A semilattice is a partially ordered
set (I, ≤) such that every two elements x, y ∈ S have a greatest lower bound, denoted by
Partially supported by Simons Foundation Collaboration Grant 359872 and by FCT project CEMATCIÊNCIAS UID/Multi/04621/2013.
∗
1
x∧y. A semigroup S is a semilattice of semigroups if there exist a semilattice (I, ≤) and a set
Y = {Sα }α∈I of pairwise disjoint subsemigroups Sα ≤ S indexed by I such that S = ∪α∈I Sα ,
and satisfying this property: for all α, β ∈ I and for all a ∈ Sα , b ∈ Sβ , we have ab ∈ Sα∧β .
(For details, see [7], Def. II.1.4, p. 27).
For our purposes, the following results are key ([7, Thm. II.6.4, p. 51], [6, Thm. 3.12, p.
47]).
Proposition 1.2. Let S be a semigroup.
(1) S is separative if and only if S is semilattice of cancellative semigroups.
(2) S is commmutative and separative if and only if S is a semilattice of commutative
cancellative semigroups.
Now we can extend the original exercise even further.
Corollary 1.3. Let S be a separative semigroup satisfying x3 = x for all x ∈ S. Then S is
commutative, and in fact, is a semilattice of abelian groups satisfying x2 = 1.
Indeed, by Proposition 1.2(1), S is a semilattice of cancellative semigroups Sα , α ∈ I, such
that each Sα satisfies x3 = x. By Proposition 1.1, each Sα is an abelian group satisfying
x2 = 1. By Proposition 1.2(2), S is commutative.
We can also view the generalizations of our exercise from a different perspective. A
semigroup S is regular if for each a ∈ S, there exists b ∈ S such that aba = a. This is
equivalent to asserting that each a ∈ S has an inverse a′ ∈ S satisfying aa′ a = a and a′ aa′ =
a′ . If each a ∈ S has a unique inverse, then S is said to be an inverse semigroup. Equivalently,
an inverse semigroup is precisely a regular semigroup in which all the idempotents commute.
A semigroup S such that each element has a commuting inverse aa′ = a′ a is said to be
completely regular. Equivalently, a completely regular semigroup is a union of groups. A
completely regular, inverse semigroup is called a Clifford semigroup. A Clifford semigroup
is characterized as a semilattice of groups. (For further details on regular semigroups, see,
for instance, [3].) Note in particular that every Clifford semigroup is separative.
This gives us a different way of viewing our exercise. A semigroup satisfying x3 = x for
all x is completely regular, with the commuting inverse of each x being given by x′ = x.
It is easy to see that a regular, cancellative semigroup is a group by essentially the same
argument as above: xx′ xy = xy and yxx′ x = yx, so x′ xy = y and yxx′ = y for all x, y ∈ S.
A semigroup (or more generally, any magma) with both a left identity element and a right
identity element has a (necessarily unique) identity element 1, and we have xx′ = x′ x = 1
for all x ∈ S. Thus we have another proof of Proposition 1.2.
In this paper, we will extend three commutativity theorems from group theory to semigroups. Our first result, which was our original motivation, is based on a recent preprint of
Venkataraman [8]. She proved that in a finite groups, if squares commute with squares and
cubes commute with cubes then the group is commutative; she also proposed the problem
of extending her result to infinite groups. More generally, in the same paper, she asked if
a group satisfying the conditions xp y p = y p xp and xq y q = y q xq for all x where p and q are
relatively prime is necessarily commutative. Although we did not know it when we began our
investigation, this is apparently a folk result in group theory [9], although we have not been
able to find a reference in the literature. (Note that the proofs given in the cited website do
2
not generalize directly to semigroups.) In the spirit of our discussion above, we prove that
Venkataraman’s desired result holds more generally.
Theorem 1.4. Let S be a separative or completely regular semigroup such that, for all
x, y ∈ S, xp y p = y p xp and xq y q = y q xq where p and q are relatively prime positive integers.
Then S is commutative.
In Example 2.2, we note that this theorem cannot be extended from completely regular
semigroups to general regular semigroups.
It is easy to see that a group, or more generally a cancellative semigroup, is commutative
if and only if (xy)2 = x2 y 2 for all x, y. The direct implication is trivial; for the converse,
xyxy = xxyy implies yx = xy after cancellation. In the same vein is the following well-known
exercise ([2], §2.3, Exer. 4):
If G is a group such that (ab)i = ai bi for three consecutive integers i for all
a, b ∈ G, show that G is abelian.
The slightly awkward wording allows two interpretations: that the integers i depend on
the elements a, b, or that the same integers i work for all a, b. The proof for groups is
essentially the same in either case. Our generalizations require both readings.
Theorem 1.5. Let S be a semigroup.
(1) Suppose S is separative and suppose that for each a, b ∈ S, there exist three consecutive nonnegative integers i such that (ab)i = ai bi . Then S is commutative.
(2) Suppose S is an inverse semigroup and suppose that there exist three consecutive
nonnegative integers i such that (ab)i = ai bi for all a, b ∈ S. Then S is commutative.
Part (2) of this theorem cannot be formulated in the same way as part (1) is; see Example
3.2. In addition, part (2) cannot be generalized to other types of regular semigroups; see
Example 3.3.
Another commutativity theorem for groups was motivated for us by another known exercise [2, Exer. 24, p. 48]:
Let G be a finite group whose order is not divisible by 3. Suppose that
(ab)3 = a3 b3 for all a, b ∈ G. Prove that G must be abelian.
The finiteness is not essential and the condition can be replaced with the assumption that
G is a group with no elements of order 3, that is, G satisfies the condition
x3 = 1 =⇒ x = 1
(1)
for all x ∈ G. More generally, a group G satisfying (ab)3 = a3 b3 for all a, b ∈ G can be
described by a more general theorem of Alperin [1] as a quotient of a subgroup of a direct
product of abelian groups and groups of exponent 3. The condition (1) rules out groups of
exponent 3, and so G is abelian. The existence of nonabelian groups of exponent 3, such
as the unique one of order 27, shows that some additional hypothesis like (1) is needed to
conclude commutativity.
There are two reasonable reformulations of (1) for semigroups. First, if a semigroup
S satisfies (ab)3 = a3 b3 for all a, b ∈ S, then this just asserts that the cubing mapping
S → S; x 7→ x3 is an endomorphism. From this point of view, (1) asserts that in groups, the
kernel of this endomorphism is trivial, or equivalently, that the endomorphism is injective.
3
This latter formulation makes sense in any semigroup S:
x3 = y 3 =⇒ x = y
(2)
for all x ∈ S.
Another reformulation of (1) for groups which works for any semigroup S, possibly without
an identity element, is weaker, but more straightforward:
x4 = x =⇒ x2 = x
(3)
for all x ∈ S. To see that this is weaker, suppose (2) holds and x4 = x. Then (x2 )3 = x6 = x3 ,
and so applying (2) yields x2 = x.
Theorem 1.6. Let S be a semigroup satisfying (xy)3 = x3 y 3 for all x ∈ S.
(1) If S is separative and satisfies (2), then S is commutative.
(2) If S is an inverse semigroup and satisfies (3), then S is commutative.
The hypothesis of part (1) of Theorem 1.6 cannot be weakened to (3); see Example 4.2
below. Also, neither part of the theorem extends to other types of regular semigroups; see
Example 4.3.
All of our investigations were aided by the automated deduction tool Prover9 created
by McCune [4]. Automated theorem provers are especially good at equational reasoning,
being able to derive consequences of equational axioms much faster and more efficiently
than humans. Any currently available automated theorem prover would have sufficed for
this project, but Prover9 has the advantage that its input and output are easily readable
by mathematicians with no familiarity with such tools. For example, here are the axioms in
an input file for the special case of Theorem 1.4 where S is a group, p = 2 and q = 3:
% group axioms
(x * y) * z = x * (y * z). % associativity
e * x = x. x * e = x.
% identity element
x’ * x = e. x * x’ = e.
% inverses
% x^2 y^2 = y^2 x^2
(x * x) * (y * y) = (y * y) * (x * x).
% x^3 y^3 = y^3 x^3
(x * (x * x)) * (y * (y * y)) = (y * (y * y)) * (x * (x * x)).
The goal is just
x * y = y * x. % commutativity
Here each equation is interpreted by Prover9 to be universally quantified in the variables.
Everything written after a % symbol is a comment. Notice that the association of terms in
any equation is made explicit; while there are settings in Prover9 which allow one to avoid
parenthesization of the input, they do not generally improve readability of proofs.
Since we are working in semigroups, the associative law is part of any input file, and is
heavily used throughout proofs, mostly as a rewrite rule. While this can lengthen proofs, it
actually causes little to no trouble for a human reader, who can skip many lines where the
rule is being applied.
4
Running Prover9 with its default settings on the above input gives a proof in less than
half a minute on a not particularly fast computer. The proof has 174 steps. Some of the
steps are long and unpleasant; for instance, the longest one has an equation with 67 symbols
in it (variables and occurrences of the operations ∗ and ′ . Here is a more typical one from
the proof, split into two lines:
961 x * (y * (x’ * (y * y))) = y * (y * (x * (y * x’))).
[para(864(a,2),51(a,1,2,2)),rewrite([14(7)])].
The number 961 is a clause identification number which is an internal index that Prover9 uses
to keep track of kept clauses. The part in square brackets is the justification for clause 961.
Here “para” is short for paramodulation, which is the primary inference used in equational
reasoning. Paramodulation refers to the substitution of one side of an equation into a subterm
of another equation. In this case, clause 864 was plugged into a subterm of clause 51. This
was followed by a rewrite of the resulting clause by clause 14.
It is not terribly enlightening to show the details of this particular step in the proof nor
any other step, because it turns out not to be necessary for translation into humanly readable
form. A proof of 167 steps is, by the standards of automated theorem provers, not very long,
and so it is reasonable to try to obtain a human proof. For familiar associative structures such
as groups, lattices, rings, and so on, this is usually easy, albeit sometimes time-consuming.
Given two equations and being told that under the axioms of group theory, the two yield a
third is usually enough for someone familiar with groups to see how the proof goes.
In addition, a human reader can take numerous shortcuts. For example, here is another
step in the proof, omitting the justification:
11605 x * (y * (x * y’)) = x * x.
A human reader can immediately see that the proof is essentially finished: cancel x on each
side of the identity and then multiply on the right by y to get commutativity. Prover9, on
the other hand, took 14 additional steps to reach commutativity. In other words, the “out
of the box” proof that Prover9 found was far from optimal.
Experienced users can tweak Prover9’s many parameters and use various specialized techniques to find proofs faster, to find shorter proofs, and so on. For example, changing the term
ordering from the default lexicographic path ordering (LPO) to the Knuth-Bendix ordering
gets a different proof in just 10 seconds and the new proof is 19 steps shorter than the first
one.
Certainly the most interesting use of Prover9 was in our investigation of Theorem 1.4.
Conditions such as xp y p = y p xp for all x, y where p is an arbitrary but fixed positive integer
cannot be directly encoded in Prover9 (or any other first-order theorem prover) because it
has no built-in description of the integers. Instead, we had to look at several special cases
such as the one above.
It was only after examination of several special cases that we realized that many of the
steps in the proofs were similar. This enabled us to see the pattern of the proof of Theorem
1.4 for groups. (Recall that at this point, we were not aware that the theorem was a
known folk result.) In particular, the special cases led us to our formulation of Lemma 2.1
below as containing the essential idea of the proofs. It was also at this same point in our
investigations that we realized the theorem holds more generally in cancellative, and then
separative semigroups. This motivated us to look at a sampling of other commutativity
theorems in various classes of semigroups.
5
After this paper was submitted, we became aware of the recently published paper of
Moghaddam and Padmanabhan [5]. The results contained therein are different from ours,
but the spirit of the work is exactly the same: by extracting the essential features of syntactic
proofs of commutativity theorems for groups, they were able to extend them to cancellative
semigroups.
2. Proof of Theorem 1.4
The goal of this section is to prove Theorem 1.4. We start with the following key lemma.
Lemma 2.1. Let S be a cancellative semigroup and suppose that there exists a map g : S → S
satisfying the following conditions: for all x, y ∈ S,
(a) xg(x) = g(x)x;
(b) g(x)g(y) = g(y)g(x);
(c) xg(x) · yg(y) = yg(y) · xg(x).
Then the semigroup S is commutative.
Proof. We claim that the following identity holds:
g(g(x)y)y = yg(g(x)y) .
(4)
In fact,
(a)
g(x)y g(g(x)y) = g(g(x)y) g(x)y
| {z } | {z } | {z } | {z }
u
g(u)
u
g(u)
= g(g(x)y) g(x) y
| {z } |{z}
g(u)
g(x)
(b)
= g(x) g(g(x)y) y,
|{z} | {z }
g(x)
g(u)
which, eliminating g(x) by left cancellation, gives (4).
Our next claim is that g(x)y = yg(x). We start by observing that
(c)
yg(y) g(x)y g(g(x)y) = g(x)y g(g(x)y) yg(y).
| {z } | {z } | {z } | {z }
u
u
g(u)
6
g(u)
(5)
Now,
(b)
g(x)y g(y)g(g(x)y) y = g(x)y g(g(x)y)g(y) y
|
|
{z
}
{z
}
g(y)g(u)
g(u)g(y)
= g(x)yg(g(x)y) g(y)y
| {z }
(a)
= g(x)yg(g(x)y) yg(y)
| {z }
(5)
= yg(y)g(x)yg(g(x)y)
(b)
yielding
= y g(x)g(y) yg(g(x)y)
| {z }
(4)
= yg(x)g(y) g(g(x)y)y,
| {z }
g(x)y · g(y)g(g(x)y)y = yg(x) · g(y)g(g(x)y)y ,
which, by right cancellation, implies g(x)y = yg(x), as claimed.
Now the proof that xy = yx is straightforward:
(c)
(b)
x yg(x) g(y) = x g(x)y g(y) = yg(y)xg(x) = y xg(y) g(x) = yxg(x)g(y),
| {z }
| {z }
| {z }
and xy = yx follows by right cancellation. This completes the proof of the lemma.
The previous lemma opens the gate to the proof of our first main theorem.
Proof of Theorem 1.4. We may assume without loss of generality that p, q > 1.
Suppose first that S is cancellative. Since p and q are relatively prime, by Bezout’s identity
there exist integers r, s such that pr + qs = 1. Since one of pr or qs must be negative, we
assume without loss of generality that qs < 0; thus −qs > 0 and x−qs ∈ S for all x ∈ S.
Since q > 0, we have s < 0 so that −s > 0; thus x−s ∈ S for all x ∈ S. As pr > 0 and p > 0,
we have r > 0 and xpr , xr ∈ S for all x ∈ S.
Let g(x) = x−qs . We claim that g(x) satisfies the three properties (a), (b) and (c) of the
previous lemma. By associativity, we have xg(x) = g(x)x so that (a) holds. Regarding (b)
we have
g(x)g(y) = x−qs y −qs = (x−s )q (y −s )q = (y −s )q (x−s )q = y −qs x−qs = g(y)g(x).
The second equality holds because by assumption the qth powers commute. Finally, regarding (c) we have
xg(x)yg(y) = xx−qs yy −qs = x1−qs y 1−qs = xpr y pr
= (xr )p (y r )p = (y r )p (xr )p = y pr xpr
= y 1−qs x1−qs = yy −qsxx−qs = yg(y)xg(x).
The fourth equality holds because pth powers commute.
We have proved that S admits a function g : S → S satisfying the three conditions of the
previous lemma. It follows that S is commutative.
7
Next assume that S is separative. By Proposition 1.2, S is a semilattice of cancellative
semigroups Sα , each of which satisfies the hypotheses of the theorem. It follows that each
Sα is commutative. By Proposition 1.2 again, S is commutative.
Finally, assume that S is completely regular. If e, f ∈ S are idempotents, then ef =
p p
e f = f p ep = f e. It follows that S is an inverse semigroup. Since S is both completely
regular and inverse, it is a Clifford semigroup, hence is a semilattice of groups. In particular,
S is separative and the desired result follows.
Example 2.2. Theorem 1.4 does not generalize from completely regular semigroups to other
types of regular semigroups. For example, let S be the Brandt semigroup of order 5. Then
for every positive integer p and every x ∈ S, xp is an idempotent. Thus the hypotheses of
the theorem are satisfied since idempotents commute in inverse semigroups, but S is not
commutative.
3. Proof of Theorems 1.5
We first need a lemma which will prove useful in both this section and the next.
Lemma 3.1. Let S be an inverse semigroup and suppose there exists an integer k > 1 such
that (xy)k = xk y k for all x ∈ S. Then S is a Clifford semigroup.
Proof. Denote the unique inverse of an element x ∈ S by x′ . We will show that xx′ = x′ x for
all x ∈ S. It will follow that S is completely regular, hence Clifford. First, since (x′ x)k = x′ x,
we have
(x′ )k xk = x′ x
(6)
′ k−1
k−1 ′
for all x ∈ S. Next, recalling that (x )
= (x ) in inverse semigroups, we compute
(6)
(x′ )k−1 xk = (xk−1 )′ xk = (x′ )k−1xk−1 · xx′ x = xx′ · (x′ )k−1xk−1 x = x(x′ )k xk = xx′ x = x ,
|
{z
}
where we used the fact that idempotents commute in the third equality. Thus
(x′ )k−1xk = x .
(7)
Next, we have
(7)
(7)
x′ xx = x′ x(x′ )k−1 xk = (x′ )k−1xk = x ,
where we used k > 1 in the second equality. Thus we have both
x′ xx = x ,
′ ′
(8)
′
xx x = x ,
(9)
where (9) follows from (8) by replacing x with x′ and using x′′ = x. Finally, we compute
(8)
(9)
xx′ = x′ xxx′ = xx′ x′ x = x′ x ,
where we used commuting idempotents in the second equality. This completes the proof of
the lemma.
Proof of Theorem 1.5. We begin with part (1) and suppose first that S is cancellative. One
of the standard proofs of Herstein’s exercise ([2], §2.3, Exer. 4) only uses cancellation and
so applies here. Say that the three consecutive nonnegative integers are i, i + 1, i + 2. Then
ai bi ab = (ab)i ab = (ab)i+1 = ai+1 bi+1 . Cancel ai on the left and b on the right to get
8
bi a = abi . Repeating the same argument with i + 1 in place of i gives bi+1 a = abi+1 . Thus
bi · ab = abi b = abi+1 = bi+1 a = bi · ba. Cancelling gives ab = ba.
Now suppose S is separative. By Proposition 1.2, S is a semilattice of cancellative semigroups Sα . Each Sα satisfies the hypothesis of the theorem, hence is commutative. By
Proposition 1.2, S is commutative.
Now we turn to part (2) and assume that S is an inverse semigroup satisfying the hypotheses of the theorem. Applying Lemma 3.1 with k = i + 1, we have that S is a Clifford
semigroup. In particular, S is separative. Since the hypotheses of part (2) are stronger than
those of part (1), we may now apply part (1) to conclude that S is commutative.
Example 3.2. Let S be any nonClifford inverse semigroup with a zero 0, for instance, the
Brandt semigroup of order 5 will suffice. Let b be any element such that bb′ 6= b′ b. Then
(0b)k = 0k bk for all k ≥ 1. This shows that part (2) of Theorem 1.5 cannot be strengthened
to be like part (1).
Example 3.3. Let S = {e, f } be the 2-element left (say) zero semigroup. Then trivially
(xy)k = xk y k for all x, y ∈ S and all positive integers k, but S is not commutative. Thus
Theorem 1.5 does not extend to arbitrary regular semigroups or even completely regular
semigroups.
4. Proof of Theorem 1.6
We start with a lemma of some independent interest.
Lemma 4.1. Let S be a cancellative semigroup satisfying (xy)3 = x3 y 3 for all x, y ∈ S.
Then for all x, y ∈ S,
x3 y = yx3 .
(10)
Proof. First, cancellation on both sides of (xy)3 = x3 y 3 gives (yx)2 = x2 y 2 for all x, y ∈ S.
Using this, we compute (x · yx)(x · yx) = (yx)2 x2 = x2 y 2 x2 . Cancelling on both sides, we
obtain
xy 2 x = yx2 x
(11)
(11)
for all x, y ∈ S. Next, x · y 2x2 · x = x(xy)2 x = xy · x2 · xy. Cancelling xy on the left gives
yx3 = x3 y for all x ∈ S, as desired.
Finally, we prove our last main result.
Proof of Theorem 1.6. For part (1), assume first that S is cancellative and satisfies (2).
Then (10) shows that the image C = {x3 | x ∈ S} of the cubing map S → S; x 7→ x3 is
commutative. The condition (2) asserts that this map is injective, hence S is isomorphic to
C. In particular, S is commutative.
Now assume S is separative. By Proposition 1.2, S is a semilattice of cancellative semigroups Sα , each of which satisfies both (2) and (xy)3 = x3 y 3 for all x ∈ Sα . By the argument
above, each Sα is commutative. Applying Proposition 1.2 again, we have that S is commutative.
For (2), now let S be an inverse semigroup satisfying both (3) and (xy)3 = x3 y 3 for all
x ∈ S. By Lemma 3.1, S is a Clifford semigroup, hence a semilattice of groups Sα . Each
group Sα satisfies (3) as well, but in groups, (3) is equivalent to (2). In particular, S is a
separative semigroup satisfying the conditions of part (1), and so S is commutative.
9
Example
4.2. The
hypothesisof part (1) of Theorem 1.6 cannot be weakened to (3). Indeed,
1 a
let S =
| a, b ∈ Z+ with matrix multiplication as the operation. Then S is a
0 b
cancellative semigroup without idempotents and thus trivially satisfies (3). However, S is
not commutative.
Example 4.3. Let S be as in Example 3.2 and note once again that (xy)3 = x3 y 3 is trivially
satisfied for all x, y ∈ S. Since S is idempotent, conditions (2) and (3) both hold. However
S is not commutative. Thus neither part of Theorem 1.6 extends to other types of regular
semigroups.
References
[1] J. L. Alperin, The classification of n-abelian groups, Canad. J. Math. 21 (1969), 1238–1244.
[2] I. N. Herstein, Topics in Algebra, Second edition. Xerox College Publishing, Lexington, Mass.-Toronto,
Ont., 1975.
[3] J. M. Howie, Fundamentals of semigroup theory, London Mathematical Society Monographs. New Series,
12. Oxford Science Publications. The Clarendon Press, Oxford University Press, New York, 1995.
[4] W. W. McCune, Prover9 and Mace4, version 2009-11A. http://www.cs.unm.edu/~mccune/prover9/
[5] G. I. Moghaddam and R. Padmanabhan, Commutativity theorems for cancellative semigroups, Semigroup Forum 95 (2017), 448–454.
[6] A. Nagy, Special classes of semigroups, Advances in Mathematics (Dordrecht) 1, Kluwer Academic
Publishers, Dordrecht, 2001.
[7] M. Petrich, Introduction to Semigroups, Merrill Research and Lecture Series. Charles E. Merrill Publishing Co., Columbus, Ohio, 1973.
[8] G. Venkataraman, Groups in which squares and cubes commute, https://arxiv.org/abs/1605.05463.
[9] A group such that am bm = bm am and an bn = bn an (m, n coprime) is abelian?
https://math.stackexchange.com/questions/326702/a-group-such-that-am-bm-bm-am-and-an-bn-bn-an-m-n(Araújo) Colégio Planalto, R. Armindo Rodrigues 28, 1600–414 Lisboa, Portugal
(Kinyon) Department of Mathematics, University of Denver, Denver, CO 80208, USA
(Kinyon) CEMAT-CIÊNCIAS, Departamento de Matemática, Faculdade de Ciências, Universidade de Lisboa, 1749-016, Lisboa, Portugal
E-mail address: [email protected]
10
| 4math.GR
|
CLASSIFICATION OF EXTERNAL ZONOTOPAL
ALGEBRAS
arXiv:1803.09966v1 [math.CO] 27 Mar 2018
GLEB NENASHEV
Abstract. In this paper we work with power algebras associated
to hyperplane arrangements. There are three main types of these
algebras, namely, external, central, and internal zonotopal algebras. We classify all external algebras up to isomorphism in terms
of zonotopes. Also, we prove that unimodular external zonotopal
algebras are in one to one correspondence with regular matroids.
For the case of central algebras we formulate a conjecture.
1. Introduction
In this paper we work with power algebras, which are quotients of
polynomial rings by power ideals. We consider zonotopal ideals, which
are associated to zonotopes. These ideals were independently introduced in two different ways. There are three types of algebras. We will
work with the definition from F. Ardila and A. Postnikov [2]; it comes
from algebras generated by the curvature forms of tautological Hermitian linear bundles [3, 28], see also papers [4, 5, 14, 15, 16, 22, 23, 26,
27, 29], where people work with quotients algebras by these ideals. At
the same time the definition and the name was established by O. Holtz
and A. Ron [12]; it comes from Box-Splines and from Dahmen-Micchelli
space [1, 8, 10], see also the papers [9, 11, 13, 18, 19, 20, 21, 30].
Let A ∈ Rn×m be a matrix of rank n. Denote by y1 , . . . , ym ∈ Rn the
columns and by t1 , . . . , tn ∈ Rm the rows. For a matrix A, we define
the zonotope
M
ZA :=
[0, yi ] ⊂ Rn
i∈[m]
as the Minkovskii sum of intervals [0, yi ], i ∈ [m]. By F (A) we denote
the set of facets of ZA . For any facet H ∈ F (A), we define m(H) as
the number of non-zero coordinates of the vector ηH A ∈ Rm , where
ηH ∈ Rn is a normal for H.
(k)
Let CA be the quotient algebra
(k)
(k)
CA := R[x1 , . . . , xn ]/IA ,
Key words and phrases. Commutative algebra, Power ideals, Zonotopes, Matroids, Lattice points.
1
2
G. NENASHEV
(k)
where IA is the zonotopal ideal generated by the polynomials
(k)
pH = (ηh · (x1 , . . . , xn ))m(H)+k , H ∈ F (A).
There are 3 main cases, where k = ±1 and 0; they were considered
in [2, 12].
(1)
• k = 1 : CAEx = CA is the external zonotopal algebra for A;
(0)
• k = 0 : CAC = CA is the central zonotopal algebra for A;
(−1)
• k = −1 : CAIn = CA is the internal zonotopal algebra for A.
Remark 1. The case k > 1 is not “zonotopal”, because the ideal Î (k)
generated by
ph = (h · (x1 , . . . , xn ))m(h)+k , h ∈ Rn
is different from I (k) . They coincide only for the case where k ≤ 1.
In the case k ≤ −5, Hilbert series is not a specialization of the
corresponding Tutte polynomial, see [2].
Theorem 1 (cf. [2, 4, 12, 20], External [27], Central for graphs [26]).
For a matrix A ∈ Rn×m , the Hilbert series of zonotopal algebras are
given by
• H(CAEx ) = q m−n TA (1 + q, 1q );
• H(CAC ) = q m−n TA (1, 1q );
• H(CAIn ) = q m−n TA (0, 1q ),
where TA is the Tutte polynomial of the vector configuration of the
columns of A (i.e., vectors y1 , . . . , ym ).
There are other definitions of external algebras from [27]. Let Φm
be the square-free commutative algebra generated by φi , i ∈ [m], i.e.,
with relations
φi φj = φj φi , i, j ∈ [m] and φ2i = 0, i ∈ [m].
Theorem 2 (cf. [27]). The external algebra CAEx is isomorphic to the
subalgebra of ΦEx
A := Φm generated by
Xi := ti · (φ1 , . . . , φm ), i ∈ [n].
In papers [15, 16] we constructed the analogue of Theorem 2 in the
case of central and internal zonotopal algebras for totally unimodular
matrices, see the definition below.
The main interesting examples of zonotopal algebras arise for totally
unimodular matrices and for graphs. The matrix A is totally unimodular if any its minors is equal to ±1 or 0. In this case the total dimensions
of the algebras have a nice interpretation.
Theorem 3 (cf. [12]). Let A ∈ Rn×m be a totally unimodular matrix
of rank n. Then the total dimension
CLASSIFICATION OF EXTERNAL ZONOTOPAL ALGEBRAS
3
• dim(CAEx ) is equal to the number of lattice points of ZA ;
• dim(CAC ) is equal to the volume of ZA ;
• dim(CAIn ) is equal to the number of interior lattice points of ZA .
The main examples of totally unimodular matrices are graphs. Namely,
let G be a graph on n vertices; then the incidence matrix of any orientation of G is totally unimodular. To construct the zonotopal algebra,
we should forget exactly one row for each connected component of G.
These algebras are independent (up to isomorphism) of the choice of
orientations and rows.
These graphical algebras were considered in [14, 15, 22, 23, 27]. In the
graphical case Theorem 3 can be written in graph theory terminology.
Theorem 4 (cf. [26]). Let G be a graph. Then the total dimension
• dim(CGEx ) is equal to the number of forests in G;
• dim(CGC ) is equal to the number of trees in G (in the connected
case).
It is well-known that the number of lattice points (volume) of the
corresponding zonotope and the number of forests (trees) of a graph
are the same, see for example [6, 17] (Points of the zonotope correspond
to score vectors).
Example 1. Let G be graph on the vertex set {0, 1, 2} with 4 edges
(0, 1), (0, 2), (1, 2), and (1, 2),
see fig. 1, left. Let us orient all edges in the sense of increasing corresponding number. Consider the incidence matrix after forgetting of
0-th row
−1 0
1
1
A :=
,
0 −1 −1 −1
see zonotope ZA on fig. 1, right.
1
2
b
1
-3
b
-2
b
0
b
b
-1
0b
-1 b
0
1
2
-2 b
b
b
-3
b
b
2
3
4
b
-4
Figure 1. A graph and its corresponding zonotope.
The zonotope ZA has 6 facets. We need the set of its normals (note
that parallel facets have the same normal up to a factor). There are 3
normals
4
G. NENASHEV
•
•
•
It is
η1 = (1, 0);
η2 = (0, 1);
η3 = (1, 1).
easy to check that m(η1 ) = m(η2 ) = 3 and m(η3 ) = 2. Hence,
(k)
3+k
2+k
IA = hx3+k
i.
1 , x2 , (x1 + x2 )
Then
• H(CAEx ) = 1 + 2q + 3q 2 + 3q 3 + q 4 and dim(CAEx ) = 10;
• H(CAC ) = 1 + 2q + 2q 2 and dim(CAC ) = 5;
• H(CAIn ) = 1 + q and dim(CAIn ) = 2.
It is easy to check that 10, 5, and 2 are exactly the number of lattice
points, the area, and the number of interior lattice points of ZA , respectively. Furthermore, 10 and 5 are the number of forests and trees
in G. In this case the Tutte polynomial is given by
TG (x, y) = TA (x, y) = x + y + x2 + xy + y 2 ,
Anyone can check the formulas for the Hilbert series.
The following important property of external graphical algebras was
proved in [23].
Theorem 5 (cf. [23]). Given two graphs G1 and G2 . Then the following
are equivalent:
• CGEx1 and CGEx2 are isomorphic as non-graded algebras;
• CGEx1 and CGEx2 are isomorphic as graded algebras;
• the graphical matroids MG1 and MG2 are isomorphic.
The following conjecture was formulated for the central case:
Conjecture 1 (cf. [23]). Given two connected graphs G1 and G2 , the
following are equivalent:
• CGC1 and CGC2 are isomorphic as non-graded algebras;
• CGC1 and CGC2 are isomorphic as graded algebras;
• the bridge-free matroids MG1 and MG2 are isomorphic.
Here the bridge-free matroid of a graph is its graphical matroid after
deleting all bridges.
In the paper [22], the K-theoretic filtration was considered, see defiEx
nition there. Denote by KG
the K-theoretic filtration of CGEx.
Theorem 6 (cf. [22]). Given two graphs G1 and G2 without isolated
Ex
Ex
vertices, the filtered algebras KG
and KG
are isomorphic if and only
1
2
if G1 and G2 are isomorphic.
The structure of this paper is as follows: in § 2 we present a classification of external zonotopal algebras and a conjecture for the central
case; in § 3 we prove our classification.
CLASSIFICATION OF EXTERNAL ZONOTOPAL ALGEBRAS
5
Acknowledgments. This material is based upon work supported by
the National Science Foundation under Grant DMS-1440140 while the
author was staying at the Mathematical Sciences Research Institute
in Berkeley, California, during the program “Geometric and Topological Combinatorics” in the fall 2017. He also would like to thank the
participiants of the PA-seminar for their comments.
2. Main results
Definition 1. Two linear spaces V1 ⊂ Rm1 and V2 ⊂ Rm2 are called
z-equivalent if m1 = m2 = m and there is an invertible diagonal matrix
D ∈ Rm×m and a permutation π ∈ Sm such that
V1 = V2 (πD).
The matrices A1 ∈ Rn1 ×m1 of rank n1 and A2 ∈ Rn2 ×m2 of rank n2
are called z-equivalent if the span of rows of A1 is z-equivalent to the
span of rows of A2 .
Remark 2. It is easy to see that z-equivalence is an equivalence relation.
In the case when A1 and A2 do not have proportional columns, we can
say that the matrix A1 is equivalent to A2 if and only if their zonotopes
are equivalent (since we can reconstruct the “matrix” from the zonotope
in this case).
This equivalence is weaker than that of matroids.
Proposition 7. If two matrices A1 and A2 are z-equivalent, then the
matroids MA1 and MA2 are isomorphic.
It is easy to check that CAEx1 and CAEx2 are isomorphic if A1 and A2 are
z-equivalent. The converse also holds.
Theorem 8. Let A1 ∈ Rn1 ×m1 and A2 ∈ Rn2 ×m2 be two matrices of
rank n1 and n2 respectively. Then the following are equivalent:
• CAEx1 and CAEx2 are isomorphic as non-graded algebras;
• CAEx1 and CAEx2 are isomorphic as graded algebras;
• A1 and A2 are z-equivalent.
Corollary 1. Let A1 ∈ Rn1 ×m1 and A2 ∈ Rn2 ×m2 be two matrices of
ranks n1 and n2 respectively, with isomorphic external algebras CAEx1 ∼
=
CAEx2 . Then the matroids MA1 and MA2 are isomorphic.
The following theorems shows that unimodular external zonotopal
algebras are in one to one correspondence with regular matroids.
Theorem 9. Let A1 ∈ Rn1 ×m1 and A2 ∈ Rn2 ×m2 be two unimodular
matrices of rank n1 and n2 respectively. Then the following are equivalent:
6
G. NENASHEV
•
•
•
•
CAEx1 and CAEx2 are isomorphic as non-graded algebras;
CAEx1 and CAEx2 are isomorphic as graded algebras;
A1 and A2 are z-equivalent.
the matroids MA1 and MA2 are isomorphic.
Since for graphs we have a totally unimodular matrix, all graphical matroids are regular; the converse is almost true. Every regular matroid may be constructed by combining graphic matroids, cographic matroids, and a certain ten-element matroid R10 , see [25] or
the book [24]. In the graphical case the last theorem says that the
algebra remembers graph up to 2-isomorphism, see [31].
For the central case, we can extend Conjecture 1 for all matrices. For
a matrix A, we say that a column is a bridge-column if after deleting
it the rank decreases.
Conjecture 2. Let A1 ∈ Rn1 ×m1 and A2 ∈ Rn2 ×m2 be two matrices of
ranks n1 and n2 respectively. Then the following are equivalent:
• CAC1 and CAC2 are isomorphic as non-graded algebras;
• CAC1 and CAC2 are isomorphic as graded algebras;
• A′1 and A′2 are z-equivalent, where A′i ∈ R(ni −ki )×(ni −ki ) is the
submatrix of Ai resulting after deleting all ki bridge-columns
and those ki rows such that rk(A′i ) = ni − ki .
3. Proofs
Let
be a finite dimension algebra over R. We say that an element
PB
k
r = i=1 a2i a2i+1 is reducible if ai ∈ B, i ∈ [2k] are nilpotent elements.
For a nilpotent element a ∈ B we define the length ℓ(a) as the maximal ℓ such that aℓ 6= 0.
Proof of Theorem 8. Clearly, we have 1 ⇐= 2 ⇐= 3, so we will prove
1 =⇒ 3. Let CAEx be our algebra. We will work with the square-free
definition, i.e., CAEx is a subalgebra of Φm , where
m = max(ℓ(a) : a ∈ CAEx ).
(note that we can work with Φm only theoretically, i.e., we do not know
this embedding). We know which element is the unit, so we can chose
basis x1 , . . . , xn of nilpotents of CAEx with the following property:
ℓ(x + r) ≥ ℓ(x),
CAEx
for any reducible r ∈
and x ∈ span{x1 , . . . , xn }. Since we can define
the algebra via some matrix A, then there is such basis.
Any element has the representation
m
X
ai,k φi + ri ,
xi =
k=1
CLASSIFICATION OF EXTERNAL ZONOTOPAL ALGEBRAS
7
where ai,k ∈ R and ri is reducible. Let A′ = {ai,k : (i, k) ∈ [n] × [m]}
be the corresponding matrix. Our goal is to reconstruct A′ up to zequivalece.
Consider the projective space Pn−1 over R. To finish the proof we
should find the multiset
A := {(a1,k , a2,k , . . . , an,k ) ∈ Pn−1 , k ∈ [m]}.
Define the set S of all non-zeroes s ∈ R such that there are i 6= j ∈ [n]
and a non-zero t ∈ R for which
ℓ(xi − sxj ) < ℓ(xi − txj ).
It is easy to see that S is exactly the set
ai,k
: i, j ∈ [n], k ∈ [m], and ai,k , aj,k 6= 0 .
aj,k
Then S is a finite set. Define a theoretical set S of rows of X as
S := {(s1 , s2 , . . . , sn ) ∈ Pn−1 , si ∈ S ∪ {0}}.
Any element of A is an element of S, so it is enough to find the
multiplicity of any s ∈ S.
Consider the following partial order on elements of Pn−1 :
(p1 , p2 , . . . , pn ) ≥ (p′1 , p′2 , . . . , p′n )
if there is t ∈ R such that
(
tpi
∀i ∈ [n], p′i =
0.
Note that if, for any s, we know the summary multiplication of all
s′ ≥ s in A, then we can calculate multiplicity of all elements.
Given s ∈ S, then the summary multiplication of all s′ ≥ s is equal
to
!
!
X
X
ℓ
bi xi − ℓ
ci xi ,
i∈I
i∈I
where
• I ⊆ [n] is the support of s;
• bi , i ∈ I are generic;
P
• ci , i ∈ I are generic with the linear condition i∈I ci si = 0.
Let us check it:
!
!
!
m
m
X X
X
X X
ai,k φk ) =
ai,k φk + rk ) = ℓ
bi (
ℓ
bi xi = ℓ
bi (
i∈I
i∈I
i∈I
k=1
m X
X
=ℓ
(
bi ai,k )φk
k=1 i∈I
!
= #{i ∈ k :
k=1
X
i∈I
bi ai,k 6= 0},
8
G. NENASHEV
Similarly we have
ℓ
X
i∈I
ci xi
!
= #{i ∈ k :
X
ci ai,k 6= 0}.
i∈I
P
Since bi , i ∈ I and ci , i ∈ I are generic
P with one condition
P i∈I ci si =
0, we have the following property: if i∈I bi ai,k 6= 0 then i∈I ci ai,k = 0
if and only if (a1,k , . . . , an,k ) ≥ s.
Hence, we can compute the multiplicity of any s.
Proof of Theorem 9. We know 1 ⇐⇒ 2 ⇐⇒ 3 =⇒ 4 by Theorem 8,
where we also reconstructed a matrix, so we know the matroid.
3 ⇐= 4. Let A1 and A2 be two totally unimodular matrices which
give the same regular matroid (we assume that the order of elements
are the same).
Also if M is a regular matroid, then all orientations of M differ only
by reorientations (see Corollary 7.9.4 [7]). Hence, we can multiply some
columns of A2 by −1 and get A′2 such that A1 and A′2 have the same
oriented matroid.
It is well-known that if we have a totally unimodular matrix Ai , then
all minimal linear dependents of its columns have coefficients ±1. We
get that matrices A1 and A′2 have linear dependents with the same
coefficients and, hence, A1 ∼z A′2 ∼z A2 .
References
[1] A. A. Akopyan, A. A. Saakyan, A system of differential equations that is related
to the polynomial class of translates of a box spline, Mat. Zametki 44:6 (1988),
pp. 705–724
[2] F. Ardila, A. Postnikov, Combinatorics and geometry of power ideals, Trans. of
the AMS 362:8 (2010), pp 4357–4384.
[3] V. I. Arnold, Remarks on eigenvalues and eigenvectors of Hermitian matrices,
Berry phase, adiabatic connections and quantum Hall effect, Selecta Mathematica 1:1 (1995), 1–19
[4] A. Berget, Products of linear forms and Tutte polynomials, European J. Combin.
31 (2010), pp. 1924–1935.
[5] A. Berget, Internal Zonotopal Algebras and the Monomial Reflection Groups,
https://arxiv.org/abs/1611.06446
[6] O. Bernardi, Tutte polynomial, subgraphs, orientations and sandpile model: new
connections via embeddings, Electronic J. Combinatorics 15:1 (2008).
[7] A. Björner, M. Las Vergnas, B. Sturmfels, N. White, G. M. Ziegler, Oriented matroids, volume 46 of Encyclopedia of Mathematics and its Applications (1999).
[8] W. Dahmen, A. Micchelli, On the local linear independence of translates of a box
spline, Studia Math. 82:3 (1985), pp.243–263
[9] C. De Boor, A. Ron, On polynomial ideals of finite codimension with applications
to box spline theory, J. Math. Anal. Appl. 158:1 (1991), pp. 168–193
[10] C. De Concini, C. Procesi, Topics in hyperplane arrangements, polytopes and
box-splines, Universitext, Springer, New York, 2011.
CLASSIFICATION OF EXTERNAL ZONOTOPAL ALGEBRAS
9
[11] N. Dyn, A. Ron, Local approximation by certain spaces of exponential polynomials, approximation order of exponential box splines, and related interpolation
problems, Trans. Amer. Math. Soc. 319 (1990), no. 1, 381–403.
[12] O. Holtz, A. Ron, Zonotopal algebra, Advances in Mathematics 227 (2011), pp
847–894.
[13] O. Holtz, A. Ron, Z. Xu, Hierarchical zonotopal, spaces.Transactions of the
American Mathematical Society 364.2 (2012): 745-766.
[14] B. Huang, Monomization of Power Ideals and Generalized Parking Functions,
https://math.mit.edu/research/highschool/primes/materials/2014/Huang.pdf
[15] A. N. Kirillov, G. Nenashev, On Q-deformations of Postnikov-Shapiro algebras,
Séminaire Lotharingien de Combinatoire, 78B.55, FPSAC (2017) 12 pp.
[16] A. N. Kirillov, G. Nenashev, Unimodular zonotopal algebra, in preparation
[17] D. J. Kleitman, K. J. Winston, Forests and score vectors, Combinatorica 1:1
(1981), pp 49–54.
[18] M. Lenz, Zonotopal algebra and forward exchange matroids. Advances in Mathematics 294 (2016), 819–852.
[19] M. Lenz, Lattice Points in Polytopes, Box Splines, and Todd Operators. International Mathematics Research Notices 2015:14 (2016), 5289–5310.
[20] M. Lenz, Hierarchical Zonotopal Power Ideals, European Journal of Combinatorics 33 (2012), pp. 1120-1141.
[21] L. Nan, A. Ron , External zonotopal algebra, Journal of Algebra and its Applications 13.02 (2014):1350097.
[22] G. Nenashev, B. Shapiro, “K-theoretic” analog of Postnikov-Shapiro algebra
distinguishes graphs, Journal of Combinatorial Theory, Series A, 148 (2017),
pp. 316–332.
[23] G. Nenashev, Postnikov-Shapiro algebras, graphical matroids and their generalizations, https://arxiv.org/abs/1509.08736
[24] J. G. Oxley, Matroid theory, Oxford University Press, USA, 2006.
[25] P. D. Seymour, Decomposition of regular matroids, Journal of combinatorial
theory, Series B, 28:3 (1980), pp.305–359.
[26] A. Postnikov, B. Shapiro, Trees, parking functions, syzygies, and deformations
of monomial ideals, Trans. Amer. Math. Soc. 356:8 (2004), pp 3109–3142.
[27] A. Postnikov, B. Shapiro, M. Shapiro, Algebras of curvature forms on homogeneous manifolds, Differential topology, infinite-dimensional Lie algebras, and
applications, Amer. Math. Soc. Transl. Ser. 2, 194 (1999), pp 227–235.
[28] B. Shapiro, M. Shapiro, On the ring generated by Chern 2-forms on SLn /B, C.
R. Acad. Sci. Paris. I Math. 326:1 (1998), pp 75–80.
[29] J. J. Shan,
A
special
case
of
Postnikov-Shapiro
conjecture
https://arxiv.org/abs/1307.5895
[30] B. Sturmfels, Z. Xu, Sagbi bases of Cox-Nagata rings, J. Eur. Math. Soc. 12
(2010), no. 2, 429–459
[31] H. Whitney, 2-isomorphic graphs, American Journal of Mathematics 55 (1933),
pp. 245–254.
Department of Mathematics, Stockholm University, S-10691, Stockholm, Sweden
E-mail address: [email protected]
| 0math.AC
|
Soft Computing Techniques for Change Detection in remotely
sensed images : A Review
Ms Madhu Khurana1 and Dr Vikas Saxena2
1
Dept. Of Computer Science & Engineering, ABES Engineering College
Ghaziabad-201009,India
[email protected]
2
Department of CS & IT, Jaypee Institute of Information Technology,
Noida – 201307, India
[email protected]
Abstract
With the advent of remote sensing satellites, a huge repository
of remotely sensed images is available. Change detection in
remotely sensed images has been an active research area as it
helps us understand the transitions that are taking place on the
Earth’s surface . This paper discusses the methods and their
classifications proposed by various researchers for change
detection. Since use of soft computing based techniques are now
very popular among research community, this paper also
presents a classification based on learning techniques used in
soft-computing methods for change detection.
Keywords : change detection, COP-kmeans, constrained kmeans,k-means, fuzzy k-means, ISODATA, Semi supervised SVM,
remote sensing, soft-computing, HMRF
1. Introduction
In Remote Sensing, change detection means assessing or
measuring the change on the Earth's surface by jointly
processing multi-temporal images of the same
geographical area acquired at different times. This field
has attracted a lot of effort in research due to its
applications in various areas as Land-use and land-cover
(LULC), Forest or Vegetation change, Forest mortality,
defoliation and damage assessment, Wetland change,
Urban expansion, Damage assessment, Crop monitoring,
Changes in glacier mass balance, Environmental change
and Deforestation, regeneration & selective logging[1].
Change detection (CD) on earth's surface is an active
research topic since it can help in monitoring and optimal
planning of Earth's resources and also help to arrest
undesired changes. Any change detection system should
be able to (a) define the change area and change rate; (b)
distribution of change areas; (c) change trajectories; and
(d) the accuracy assessment of the change detection
methods. [2]
The result of any CD is affected by various factors s.a.
spatial, spectral, thematic and temporal constraints;
radiometric resolution, atmospheric conditions and soil
moisture conditions. The methods being developed are
usually application specific. Various techniques have been
developed for CD using both pixel based as well as object
based approaches.
2. Change Detection Process : An overview
The change detection systems designed use the data
generated by different satellites having different
radiometric resolutions, spatial resolution & orbiting
frequency. Most commonly SPOT (Systeme Pour
l'Observation de la Terre(SPOT) satellite system),
NOAA’s AVHRR (Advanced Very High Resolution
Radiometer, http://www.noaa.gov/), MODIS (Moderate
Resolution
Imaging
Spectroradiometer,
http://modis.gsfc.nasa.gov/index.php),
LANDSAT
(http://landsat.gsfc.nasa.gov/?page_id=9) data is used.
Also very high resolution data from optical satellites s.a
QuickBird, IKONOS, EROS and GeoEye can also be
used. For CD systems to give good results, the multitemporal images taken should be of the same spatial and
radiometric resolution, and proper registration of the
images should be done, so as to avoid false alarms. Also
the images should have near anniversary acquisition
dates, so as to overcome the problems of sun angle and
phenological differences.
As discussed by D. Lu et. al in [2], the process of change
detection includes three major steps
Step 1: Image preprocessing –Before applying any
Change Detection algorithm on the multitemporal
images, we need to carry radiometric corrections and
image registration, so as to increase the accuracy of the
CD. It may also include geometric rectification,
atmospheric corrections and also topographic corrections,
if the terrain contains mountains. Radiometric correction
is done to reduce the inconsistency between the values
surveyed by sensors arising due to different imaging
seasons, different solar altitudes, different angles or
different cloud or snow covers[11]. Also image
registration of multi-temporal images is required so as to
avoid false change detection for pixel-based CD methods.
Image registration may not be necessary for feature-based
CD methods where extracted features can be compared for
change detection[11]. Jianya (2008) suggested that image
registration and change detection should be integrated in
a single algorithm.
Step 2 : Change Detection Algorithm or technique : we
choose a technique or algorithm to detect and analyze the
change. Selecting an algorithm for change detection is a
difficult task. Different algorithms can be studied and the
algorithm best suited for the application area, can be
selected, depending on the kind of results expected. If the
application only requires whether a change has occurred
or not, then some set of algorithms may be suited while if
we also want to work on change area and the direction of
change, another set may be suitable. The algorithm will
also depend upon the radiometric and spatial resolutions
of the images, some pixel based method may not be suited
for VHR images[4]. The spatial resolution dictates the
kind of study that can be performed on the RS data, if the
resolution is coarse the change over a large area only can
be done, if we wish to perform a sub-metric change
detection, then the data available should support such
resolutions.
Step 3 : Accuracy Assessment : Accuracy assessment
determines the quality of information retrieved from
remotely sensed data. We can have a qualitative or a
quantitative accuracy assessment. In qualitative
assessment we assess the accuracy by looking at the map
and comparing it with what we see on the ground while in
case of quantitative assessment we try to identify and
measure the remote sensing error by comparing it with
the ground truth.[11]
Designing a change detection system requires a deliberate
effort to understand the kind of changes that need to be
observed. Deciding on the period for which data is
required (or temporal difference between the data
images), spectral resolution and the nature of result
expected will guide in data collection. If the changes to be
observed cover a large geographical area (like land cover
– land change), the images with less radiometric
resolution will suffice, else in case we need to observe the
changes in detail (sub-metric), high resolution images
would be required. Images from different resolutions can
also be combined in an application, using GIS software,
here pre-processing would be required to bring the images
at same resolution, followed by change detection.
3. Existing Classifications of CD Techniques
In the past different classifications of the CD methods
have been proposed. This section discusses some proposed
classifications. A. Singh [3] in his paper on review of the
digital change detection techniques in 1989 categorized
the change detection research on the basis of
(a)
The data transformation used & (b) the analysis technique
used to detect the change. He classified the research work
done till that time into various categories. A brief
overview of all the prevalent techniques s.a. image
differencing, image rationing, image regression,
vegetation index differencing , principal component
analysis, post-classification comparison and Direct
multidate classification, Change Vector Analysis (CVA)
was offered. While evaluating these techniques on the
basis of literature reviewed, it was concluded that
different CD methods produced different change maps. It
was observed that most of the results were not compared
with the ground truth and thus the capability of these
methods were poorly evaluated. The paper provided the
best results for each technique of CD. The accuracy for
these techniques was in the range of 51% to 74%.
Jianya et. al[11] in 2008 proposed two broad groups for
classifying the CD methods, namely, bi-temporal and
temporal trajectory analysis. Bi-temporal methods
measures the change between two date images while
temporal trajectory analysis analyses the change based on
continuous timescale measuring not only the change
between two dates but also the progress of change over the
period. Further he gave seven categories for CD methods,
namely, direct comparison, classification, object-oriented
method, model method, time-series analysis, visual
analysis and hybrid method.
M.Hussain et. al[4] in a recent paper(2013) have reviewed
and categorized the change detection methods based on
the unit of analysis, i.e. pixel based and object based
approaches. In the pixel based methods the analysis is
done on the basis of variation in pixel intensities while
object based approaches, first extract objects from the
image and then try to perform change detection based
upon the objects extracted. The object based methods are
more suitable for Very High Resolution (VHR) images.
The subclasses with each category are defined and the CD
methods are categorized based upon these classes. Figure
1 is the pictorial representation of the classification
provided by the authors.
Figure 1 : Classification of change detection methods(As in M.Hussain et. al (2013))
The advantages and disadvantages as well as examples
where these method have been used, has been discussed
by Hussain et.al(2013), the interested readers can refer it
for more details.
4. Soft Computing Based Change Detection
Methods
Soft computing is a fusion of methodologies designed to
model and enable solutions to real life problems, which
are not modeled or too difficult to model mathematically.
It has the capability to tolerate imprecision, uncertainty,
approximate reasoning and partial truth. Soft computing
is a collection of Evolutionary computing, neural
networks(NN) (including neural computing), fuzzy
systems (including fuzzy logic(FL)) , machine
learning(ML),
and
probabilistic
computing(PC).
Evolutionary computing(EC) encompasses Genetic
Algorithms, Evolutionary strategies & Evolutionary
programming.
Change detection methods using soft computing can be
divide into various categories depending upon learning
paradigm being used by the system, another distinction
can be made on the techniques being used to model to
find the solution. Soft computing techniques, NN, EC,
ML, FL and PC are complementary to each other rather
than competitive. Therefore a combination of these
techniques is usually used to model a real life problem.
Fuzzy logic plays a critical role as it provides the
tolerance for imprecision and allows to take decision in a
human like manner.
Change detection(CD) in remote sensing requires large
amount of data to be analyzed. Automatic and
unsupervised CD techniques can reduce the time required
for manual image analysis. Also it is difficult to gather
the precise ground truth about the entire regions being
covered by the RS images.
The following section provides a classification of CD
methods based on soft-computing techniques being used.
5 Classification based on Learning Paradigm
5.1 Unsupervised Change Detection
Following pre-processing and calculation of the difference
image, we need to classify the data in the difference image
into change and no-change sets. Unsupervised CD
techniques do not require training data. So the task is to
discriminate the data into changed and unchanged sets.
The problem reduces to clustering the data into two
groups, changed and unchanged. We need to partition the
data into disjoint groups where each group contains
similar data.
Unsupervised methods can be divided into contextinsensitive (based on spectral information) (A.Singh.
1989)
and
context
sensitive(using
spatial
information)(Ghosh 20014). Histogram equalization is an
example of context-insensitive method of classification
which does not use neighborhood information(spatial
relations). Markov Random Fields are used to overcome
the problems of context-insensitive methods. [15]
A number of clustering algorithms can be used[12] :
Hard C- Means (HCM) or K-means : The set of
patterns are to be divided into c clusters( or k clusters).
We randomly chose c number of patterns as the initial
cluster centers. Then the distance of each pattern from
each cluster center is calculated and we assign the pattern
to cluster having minimum distance. Once the assignment
has been done for all the patterns in the dataset, we again
calculate the new cluster centers by taking the mean of all
the patterns in the cluster. Again the distance of each
pattern is calculated from the cluster centers and the
pattern reassigned to the nearest cluster centers. This
process continues till the centers become fixed i.e. there is
no change in the cluster centers over an iteration. The
distance measure can be defined for a specific problem but
usually the Euclidean distance is used. The clusters thus
formed can be further analyzed for designating them as
change or no-change clusters.
ISODATA (Iterative Self Organizing Data Analysis
Technique) : ISODATA is a method of unsupervised
classification. Unlike k-means, no a-priori information
about the number of clusters is required. It starts with
some random no. of clusters and the clusters can be split
or merged depending upon the parameters defined by the
user. User gives the parameters to this algorithm as
various threshold values. The program iterates until the
threshold is reached. Some fast variations of the
algorithm are defined, s.a., ISOCLUS by Memarsadeghi
N. et. al (2007).
Fuzzy C-Means (FCM) : The fuzzy C-Means is a
classification technique, where each pattern can have
membership in more than one cluster, the degree of
membership varies in different clusters. The algorithm
uses functional optimization, where it attempts to
minimize the objective function defined as :
where m > 1 is the fuzzifier, which controls the fuzziness
of the algorithm. The higher the value of m, more fuzzy is
the algorithm, mostly the values used are between 1.5 to
2.5. The algorithm outputs clusters with similar shapes,
sizes and densities, which might not give good results in
various situations.[13] [12][15]
Gustafson-Kessel clustering (GKC) : GKC is also a
fuzzy clustering technique. It is a modification of FCM,
where FCM uses Euclidean distance to measure the
distance between the cluster center and the patterns, GKC
uses an adaptive distance norm. Gustafson and Kessel
suggested an adaptive distance measure using fuzzy
covariance matrix. GKC can be used to detect clusters
with ellipsoidal cluster, by adapting the cluster shape
automatically.[13] [12][15]
5.2 Semi-supervised Change Detection
Unsupervised methods are best suited for change detection
applications, since ground truth may not be available in
many scenarios. Also labeling the complete training set
may not be possible. In case, some data is available, where
labels are provided, the advantages of both supervised and
unsupervised learning can be utilized. A few labeled
training patterns can be used along with unlabeled
patterns for learning. Semi supervised learning can
improve the classification results, in case of nonavailability of the sufficient labeled data [13].
S.Basu[16] proposed that semi-supervised clustering can
be improved by defining pairwise constraints i.e. pairs of
instances labeled as belonging to same or different
clusters. Thus supervision is added in the form of these
constraints, which are used to either modify the objective
function or the distance measure.
Various semi-supervised algorithms have been explored
for the classification of remotely sensed images.
(iv)
K-means based algorithms : As discussed earlier kmeans is a very popular algorithm used for clustering the
data. A number of modifications have been proposed to
the basic k-means algorithm to adapt it for semisupervised learning, i.e. incorporating some label data to
help it learn in a better manner. This section discusses
some such variations.
(i)
Seeded k-means : The initial cluster centers are
randomly chosen in standard k-means algorithm,
here the cluster centers are initialized by the mean
of the seeds belonging to that cluster. The labeled
data is only used to find the initial cluster centers.
Once the cluster centers have been initialized, the
algorithm follows the same steps as standard kmeans. The seeds are also reassigned to other
clusters, if they are nearest to the cluster, thus
restricting the impact of defining seeds only to the
initialization step [15].
(ii)
Constrained k-means : This is same as seeded kmeans, with the only difference being that the seeds
are not reassigned to any other clusters, i.e., their
membership is not re-calculated [15].
(iii)
COP k-means : This is an extension of standard kmeans algorithm, defined by Wagstaff et. al (2001).
It defines the labeled information in the form of
‘must-link’ and ‘cannot-link’ constraints. Both the
constraints have to be satisfied for assigning a
pattern to a cluster. The distance of the pattern
from all the clusters is determined and arranged in
the ascending order based on the distance of the
pattern from the cluster. For assigning the pattern
to a cluster, the cluster at the top of the list is
considered first, if assigning the pattern to this
cluster does not violate any constraints, then the
pattern is assigned to this cluster otherwise the next
cluster on the list is considered. Assigning a
pattern to a cluster means that the patterns which
were already assigned to this cluster do not have a
‘cannot-link’ constraint defined for the unlabelled
pattern. Also it indicates that the patterns assigned
to the different clusters do not have a ‘must-link’
constraint with the said pattern. In case any of the
constraints are violated, the next cluster on the list
is checked. The rest of the steps are similar to the
standard k-means algorithm, apart from the
assignment step.
Semi-supervised HMRF k-means : Basu et. al [16]
in 2004 proposed a k-means algorithm based on
hidden Markov Random Field (HMRF). HMRF
has two components :
a) A
hidden field ,L, whose variables are
unobservable, i.e., unobservable cluster labels
for the patterns
b) An observable set, X, of random variable, i.e. the
patterns
The Maximum a-posteriori (MAP) configuration of
HMRF model is same as maximizing the posteriori
probability Pr(L|Ψ). The hidden random variables Ii
are linked with neighborhood Ni. The must-link
constraints M and cannot-link constraints C define
the neighborhood over the hidden labels such that
neighbors of a point xi are all the points which
have must-link or cannot-link relations with the
said data point. The random field defined over the
hidden variables is a Markov Random Field.
The HMRF K-means algorithm, takes as input the
set of data points, must-link constraints M , cannotlink constraints C, the distance measure D, and
constraint violation costs w and w` . It tries to
minimize the objective function. The algorithm
iteratively assigns the cluster-labels on the
datapoints to minimize the objective function,
recalculates the centroids based on assignments,
recalculates the distance measure D. The process
follows till convergence. The other details of the
method can be found in the paper by Basu et. al
[16].
(v)
Semi-supervised kernel k-means : A weighted
kernel-based approach for semi-supervised learning
was proposed by Kulis et. al. The patterns are
mapped to a higher dimensional space with the
help of a non-linear kernel and a positive weight is
associated with each pattern. The algorithm
defined a ‘reward’ for obeying the constraint
instead of a penalty for constraint violation.
Whenever the patterns having an associated mustlink constraint are assigned the same cluster, the
penalty is subtracted from the objective function,
thus reducing the cost
Semi-supervised Support Vector Machine (S3VM) :
Standard Support Vector Machines can handle partially
labeled datasets by assuming unlabeled datasets as
additional optimization variables. Transductive SVM’s
(TSVMs) were introduced by Vapnik and Sterin (1977)
[17]. TSVM uses both labeled and unlabeled data to
discriminate the boundary such that it does not pass
through dense regions. Transductive learner is one which
works on labeled and unlabeled data and cannot work on
unforeseen data. While Vapnik et. al. named it as the
transductive, it actually is an inductive learner as the
resulting classifier works on the entire data space. Given a
set of data, the SVMs classify it such that the hyperplane
distinguishing the two classes has the maximum margin.
The details of the S3VMs can be found in the paper by
Vapnik et. al [17].
5.3 Supervised Change Detection
Supervised change detection methods require the presence
of ground truth, i.e. labeling of training set into various so
as to use this information for proper classification.
Supervised methods are superior in the sense that they
can recognize the kind of change. They are robust to the
different atmospheric and light conditions at time of
acquisition. Many methods discussed by A.Singh and
Hussain et.al. in their classifications fall into this
category. Post classification methods, direct multi-date
classification and kernel based methods fall under this
category[15]. A lot of work has been done in CD using
these techniques. The interested readers can refer
[3],[4],[6] for details of these techniques.
5.3 Active Learning Algorithms
Active learning is to train the classifier on a small set of
well-chosen examples. The Active learning algorithms are
iterative in nature. At each iteration some labeled samples
are provided to the model, the model adapts itself to
samples provided and then it presents the user with a set
of unlabeled samples and the user ranks them, the model
is again adapted with the new set of samples provided.
The model keeps on improving its adaptation to the
classification problem.
Active learning requires
interaction between the user and the model. [18] The
main focus in active learning is that the ranking of the
unlabeled samples be such that it is able to discriminate
between the two classes. The algorithms can be divided
into three families :
(i) Committee-based heuristics : This method quantifies
the uncertainty of a pixel by considering a committee
of learners[18]. Each member of the committee
labels the pixels in the pool to be labeled by
considering a different hypothesis about the
classification problem. The algorithm then selects the
samples showing maximum disagreement between
the different classification models in the committee
[18]. Query-by-boosting and query-by-bagging are
the heuristics used to limit the search. In remote
sensing Normalized Entropy Query-by-Bagging
(nEQB) and Adaptive Maximum Disagreement
(AMD) algorithms have been used for classification
between different classes[18].
(ii) Large margin based heuristics : The algorithms in
this category try to increase the margin between
different classes. SVMs are used as the base methods
for active learning. The distance to the hyperplane in
SVM can be used as the distance measure in the
active learning algorithm to decide the value of the
decision function. Most popular and researched
algorithm in remote sensing is Margin Sampling
(MS), where in a multiclass one-against-all setting,
the distance to each hyperplane is calculated and the
samples are decided on the basis of heuristic
calculated as selecting the minimum distance to any
of the hyperplanes. The samples with minimum
distance are the samples on the margin and are the
ones which become the support vectors.
Other algorithms using Multiclass Level Uncertainty
(MCLU), Significance Space Construction(SSC) have
been discussed in [18], which can be referred for
details. It is important to have diversity among the
samples used for classification, as it helps to better
classify the samples. General diversity based
algorithms have also been designed which try to
constraint each iteration with diversity among the
candidate samples[18].
(iii) Posterior probability-based heuristics : In this class
of methods, posterior probabilities of class
membership are used to rank the candidates. The
posterior probability distribution is considered per
class for each candidate or change in overall posterior
probability. These probabilities are used to select the
samples which might maximize the change in
posterior probability. Heuristics like KL-Max or
Breaking-Ties (BT) can be used [18].
6. Classification based on Soft Computing
Techniques
6.1 Neural Networks based approaches
Neural networks model the way the training sets would be
used to learn. Different types of neural networks are in
use now a days. The multiple layers can be used for
learning from complex data. Using neural networks for
any application requires us to define the architecture of
neural network, the activation functions for neurons and
the learning algorithm. There are various architectures
prominent now a days, for eg, Multi layer perceptron
model, self organizing maps, competitive networks,
Hopfield. Similarly various activation functions can be
defined like sigmoidal, tangent hyperbolic, radial basis
etc. The most important aspect of a neural network is the
training algorithm that we select, back propagation
algorithm, gradient descent has been extensively in NNs.
The NNs are quite suitable for remote sensing images,
since they have the capability of self-learning, which
helps in the NN learning the classification on its own.
Table 1 : Classification based on learning techniques
Technique
Sub-Category
Advantages
Disadvantages
Examples
Unsupervised
Change Detection
HCM
-unsupervised learning
- no labeled data is required, less cost
-May be struck in local minima, if
initial cluster centroids are such
ISODATA
- no apriori information about data is required
- effective at identifying spectral clusters in data
- Little user effort required
FCM
-able to label overlapping pixels into multiple
classes, able to handle uncertainty, imprecision in
classification
- is distribution free
GKC
-uses adaptive distance norm to measure the
distance between clusters
-can extract clusters with different shapes
- more effective results
-uses background knowledge in the form of
instance level constraints
- time consuming if data is very
unstructured
-Algorithm can spiral out of control
leaving only one class
-Can also be struck in local minima
-Extracts clusters of same size, densities
and circular shape, fails to perform if
the clusters have different shapes
-Performance depends on proper
selection of fuzzifiers.
- the parameter ρi for defining the
shape, should be selected properly.
M.Roy(2013,2014)
Wagstaff(2001), L
Gomez(2003),
Mishra et al(2012)
A.Ghosh(2011)
N. Memarsadeghi
(2007)
Bo Li.(2010)
Semi-supervised
Change Detection
k-means
based
COP-k
means
Seeded kmeans
Constraint kmeans
Semisupervised
HRMF kmeans
Semisupervised
kernel kmeans
Semi supervised SVM
(TSVM)
Active Learning
Methods
Supervised Change
Detection
M.Roy(2013,2014)
Mishra et al(2012)
A.Ghosh(2011)
Mishra et al(2012)
A.Ghosh(2011)
M.Roy(2013,2014)
- a poor decision in the early stage of
algorithm, may lead to non assignments
of a data in any cluster.
- execution time increase as the no. of
constraints increase
- the original labels are lost after
reassignment
Wagstaff(2001),
M.Roy[2013)
- it may be unable to assign a data point
to any of the clusters, as that might
violate the constraints
M.Roy(2013,2014)
Wagstaff(2001)
Bennett et. al (2000)
- Uses both constraints & an underlying distortion
measure for clustering
-can cluster input data only in the form
of vectors
Basu(2004),
M.Roy(2013,2014)
-discovers clusters with non-linear boundaries, after
applying kernel function
- can be used for both graph based & vector based
input data
-unlabeled data helps find a boundary between the
classes which is away from dense regions
-selection of kernel function for a
problem domain
Kullis (2005)
-to find exact boundary is a NP hard
problem
- computationally expensive, difficult to
work with large data sets
Vapnik (1977)
M.Chi(2007)
-uses labeled patterns for initialization of cluster
centers
- rate of convergence is higher
-the labeled patterns are not reassigned to any other
cluster, leading to less time requirement
M.Roy(2013,2014)
Committeebased
heuristics
-these methods can be applied to any model or
combination of models
Tuia D (2011)
Large margin
based
heuristics
Posterior
probabilitybased
heuristics
-diversity criteria improves the quality of results.
Tuia D (2011)
- If batch size is large, simple heuristics like BT
should be used
- Should be avoided if the initial
training set is small.
Refer [3], [4], [6] for details
Tuia D (2011)
The learning paradigm can be supervised, semisupervised or un-supervised. In [19] Hopfield type neural
networks and self organizing maps have been successfully
used for change detection. An overall energy function has
been defined for the network and a change detection map
is formed as a result. Eliptical Basis NN, MLP have been
used in [13] and a multiple classifier system has been
designed for classification.
is repeated on all the parent pairs, thus creating a new
population in the process. The above defined steps are
repeated till some termination criteria is met.
GA has been used along with fuzzy clustering in
[12],[13]. Multiobjective cost function optimization is
utilized in [26] for GA and the main advantage is it can
also work without computing the difference image and
can be used in different types of input satellite images. As
per [26] the main drawback was the computational cost, it
increase with the size of image.
6.2 Fuzzy Logic based approaches
6.4 Hybrid approaches
The Change detection problem in remote sensing
demands that the pixels be classified as change or nochange class, but there are pixels that overlap between
these classes. Fuzzy logic helps us define the membership
function for any pixel in different classes, thus it lets us
handle the imprecision we have in assigning the pixel to a
particular class. In [12], [13] fuzzy clustering has been
used for classification using FCM and GKC and fuzzy k
nearest neighbour as explained in the previous section.
6.3 Genetic Algorithms (GA) based methods
Genetic algorithms belong to evolutionary computing
paradigm. This technique is widely used in optimization
problems, where we need to optimize some objective
function, which is highly non-linear. Genetic algorithms
overcome the problem of local maxima that we face using
fuzzy clustering. It helps us to explore a large solution
space, to find the optimal solution.
The probable solutions to a problem are coded as
chromosomes (representation of solution as strings).
Mostly the chromosomes are coded as strings of 0’s and
1’s. The set of chromosomes is defined as population. A
fitness
function
is
associated
with
the
chromosomes(solutions) and a fitness value is evaluated
for each particular chromosome. Initially a random set of
solutions is generated, and the fitness value for each of
these solutions is calculated. Based on this fitness value, a
set of two chromosomes is selected using some selection
mechanism, known as parent chromosomes. Then two
operations crossover and mutation are applied on these
parents(solutions) to generate two offspring’s (Children).
During crossover a crossover point is selected and the
strings are exchanged between the parent strings at this
point to create the new offspring’s with a probability
defined as crossover probability. After crossover
operation, mutation operation with some defined mutation
probability is done on the produced offspring’s. During
mutation, some of the bits are flipped in the produced
offsprings to generate the new offspring’s. This operation
As discussed earlier that the techniques NN, FL and GA
are complimentary to each other, i.e., we can use a
combination of these techniques to create solutions for
any problem domain. Mostly NN is combined with FL,
where FL is used to represent the data while NN is used
for learning and inducting the results.
In change detection a combination of these techniques
have been used by many and they have been found to
produce better results. In [12] A. Ghosh has combined FL
with GA and Simulated Annealing (SA) and analyzed the
results thus obtained. Fuzzy clustering has been found to
be less costly & simple and perform better when
compared to other context sensitive techniques.
7. Conclusion
Change detection in remotely sensed images is an
important aspect of the applications intended for
observing the changes occurring due to natural or manmade processes and restraining the damages being caused
by them to Earth’s Environment, and planning the
utilization of Earths’ resources. This paper discussed the
basic process of change detection along with the
classifications proposed for these methods. The
classification based on learning techniques used in
softcomputing methods in Change detection has been
proposed and discussed.
References
[1] Bruzzone, L., & Bovolo, F. (2013). A Novel Framework for
the Design of Change-Detection Systems for Very-HighResolution Remote Sensing Images. Proceedings of IEEE.
[2] Lu, D., Mausel, P., Brondizio, E., & Moran, E. (2004).
Change detection techniques. International journal of
remote sensing, 25(12), 2365-2401.
[3] Singh, A. (1989). Review Article Digital change detection
techniques using remotely-sensed data. International
journal of remote sensing, 10(6), 989-1003.
[4] Hussain, M., Chen, D., Cheng, A., Wei, H., & Stanley, D.
(2013). Change detection from remotely sensed images:
From pixel-based to object-based approaches. ISPRS
Journal of Photogrammetry and Remote Sensing, 80, 91106.
[5] Dianat, R., & Kasaei, S. (2010). Change detection in
optical remote sensing images using difference-based
methods and spatial information. Geoscience and Remote
Sensing Letters, IEEE, 7(1), 215-219.
[6] A. J. Radke, S. Andra, O. Al-Kofahi, and B. Roysam,
Image change detection algorithms: A systematic survey,
IEEE Trans. on Image Process., vol. 14, no. 3, pp. 294–
307, Mar. 2005.
[7] P. R. Coppin, I. Jonckheere, and K. Nachaerts, Digital
change detection in ecosystem monitoring: A review, Int. J.
Remote Sens., vol. 25, no. 9, pp. 1565–1596, May 2004.
[8] F. Bovolo, A multilevel parcel-based approach to change
detection in very high resolution multitemporal images,
IEEE Geosci. Remote Sens. Lett., vol. 6, no. 1, pp. 33–37,
Jan. 2009.
[9] R. A. Rensink, Change detection, Annu. Rev. Psychol., vol.
53, pp. 245–277, 2002.
[10] S. Spotorno and S. Faure, Change detection in complex
scenes: Hemispheric contribution and the role of perceptual
and semantic factors, Perception, vol. 40, pp. 5–22,2011.
[11] Jianyaa, G., Haiganga, S., Guoruia, M., Qimingb, Z., 2008.
A Review of Multi-Temporal Remote Sensing Data Change
Detection Algorithms, ISPRS Congress, Beijing, China, pp.
757–762.
[12] A.Ghosh,N.S. Mishra, S.Ghosh, Fuzzy clustering
algorithms for unsupervised change detection in remote
sensing images, Information Sciences, pp. 699-715,2011
[13] M.Roy,S.Ghosh, A.Ghosh,A Novel approach for change
detection of remotely sensed images using semi-supervised
multiple classifier system, Information Sciences, pp 35-47,
2014.
[14] Erkan, A.N.; Camps-Valls, G.; Altun, Y., "Semi-supervised
remote sensing image classification via maximum entropy,"
Machine Learning for Signal Processing (MLSP), 2010
IEEE International Workshop on , vol., no., pp.313,318,
Aug. 29 2010-Sept. 1 2010 doi:
10.1109/MLSP.2010.5589199
[15] M.Roy,S.Ghosh, A.Ghosh , Change detection in remotely
sensed images using semi-supervised clustering algorithms,
Int. J. Knowledge Engineering and Soft Data Paradigms,
Vol. 4, No. 2, 2013
[16] Sugato Basu, Mikhail Bilenko, Raymond J. Mooney, A
Probabilistic Framework for SemiSupervised Clustering,
Proceedings of the Tenth ACM SIGKDD International
Conference on Knowledge Discovery and Data
Mining(KDD-2004), pp. 59-68, Seattle, WA, August 2004.
[17] Vladimir N. Vapnik and A. Sterin. On Structural Risk
Minimization or Overall Risk in a Problem of Pattern
Recognition. Automation and Remote Control, 10(3):14951503, 1977.
[18] Tuia, D.; Volpi, M.; Copa, L.; Kanevski, M.; Munoz-Mari,
J., "A Survey of Active Learning Algorithms for Supervised
Remote Sensing Image Classification," Selected Topics in
[19]
[20]
[21]
[22]
[23]
[24]
[25]
[26]
Signal Processing, IEEE Journal of , vol.5, no.3,
pp.606,617, June 2011
S.Ghosh, S. Patra, A.Ghosh ,A Neural Approach to
Unsupervised Change Detection of Remote-Sensing
Images, Speech, Audio, Image and Biomedical Signal
Processing using Neural Networks, Studies in
Computational Intelligence Volume 83, 2008, pp 243-268
Wagstaff, K., Cardie, C., Rogers, S. and Schroedl, S.
‘Constrained K-Means clustering with background
knowledge’, in C.E. Brodley and A.P. Danyluk (Eds.):
Proceedings 18th International Conference on Machine
Learning (ICML-2001), pp.577–584, Morgan Kaufmann,
Williamstown, MA, USA.
Kulis, B., Basu, S., Dhillon, I. and Mooney, R.J. ‘Semisupervised graph clustering: a kernel approach’, in
Proceedings of the 22nd International Conference on
Machine Learning, Germany, 2005.
N. Memarsadeghi, D. M. Mount, N. S. Netanyahu, and J.
L. Moigne, “A fast implementation of the isodata clustering
algorithm,” International Journal of Computational
Geometry & Applications, vol. 17,no. 1, pp. 71–103, 2007.
Mingmin Chi; Bruzzone, L., "Semisupervised
Classification of Hyperspectral Images by SVMs Optimized
in the Primal," Geoscience and Remote Sensing, IEEE
Transactions on , vol.45, no.6, pp.1870,1880, June 2007
Bo Li; Hui Zhao; Zhenhua Lv, "Parallel ISODATA
Clustering of Remote Sensing Images Based on
MapReduce," Cyber-Enabled Distributed Computing and
Knowledge Discovery (CyberC), 2010 International
Conference on , vol., no., pp.380,383, 10-12 Oct. 2010
K.P. Bennett, P.S. Bradley, and A. Demiriz, Constrained
K-Means Clustering, Microsoft Research, MSR-TR-200065, 2000
Çelık, Turgay; Yetgın, Zeki, Change detection without
difference image computation based on multiobjective cost
function optimization, Turkish Journal of Electrical
Engineering & Computer Sciences; Vol. 19 Issue 6, p941,
Nov 2011
Madhu Khurana has done B.Sc(Comp.Sc) from HansRaj College,
Delhi University in 1990, M.Tech(I.T.) from GGSIPU, Delhi in 2009.
Currently pursuing Ph.D from JIIT Noida, India. Worked in various
companies from 1990 to 1993. Followed by working as a freelance
consultant to various National Banks and Trading companies by
providing services
for system development and maintenance.
Academic experience of 11 years, presently working in ABES
Engineering College, Ghaziabad for last 7+ years. . Her research
interest area is Digital Image Processing and Soft Computing.
Vikas Saxena is Associate professor in Dept of CSE&IT at JIIT,
Noida. He has more than 12 years of experience in academics. His
research interest areas are--DIP, Computer vision and
Watermarking. He has more than 40 research paper published in
repute international jounals and conferences.
| 9cs.NE
|
BOUNDED QUANTIFIER INSTANTIATION FOR CHECKING
INDUCTIVE INVARIANTS
arXiv:1710.08668v1 [cs.LO] 24 Oct 2017
YOTAM M. Y. FELDMAN, ODED PADON, NEIL IMMERMAN, MOOLY SAGIV,
AND SHARON SHOHAM
Tel Aviv University, Tel Aviv, Israel
e-mail address: [email protected]
Tel Aviv University, Tel Aviv, Israel
UMass, Amherst, USA
Tel Aviv University, Tel Aviv, Israel
Tel Aviv University, Tel Aviv, Israel
Abstract. We consider the problem of checking whether a proposed invariant ϕ expressed
in first-order logic with quantifier alternation is inductive, i.e. preserved by a piece of code.
While the problem is undecidable, modern SMT solvers can sometimes solve it automatically.
However, they employ powerful quantifier instantiation methods that may diverge, especially
when ϕ is not preserved. A notable difficulty arises due to counterexamples of infinite size.
This paper studies Bounded-Horizon instantiation, a natural method for guaranteeing
the termination of SMT solvers. The method bounds the depth of terms used in the
quantifier instantiation process. We show that this method is surprisingly powerful for
checking quantified invariants in uninterpreted domains. Furthermore, by producing partial
models it can help the user diagnose the case when ϕ is not inductive, especially when the
underlying reason is the existence of infinite counterexamples.
Our main technical result is that Bounded-Horizon is at least as powerful as instrumentation, which is a manual method to guarantee convergence of the solver by modifying the
program so that it admits a purely universal invariant. We show that with a bound of 1
we can simulate a natural class of instrumentations, without the need to modify the code
and in a fully automatic way. We also report on a prototype implementation on top of Z3,
which we used to verify several examples by Bounded-Horizon of bound 1.
Key words and phrases: Decidable Logic, Quantifier Instantiation, EPR, Inductive Invariants, Deductive
Verification, Decision Procedures.
∗
A preliminary version of this paper appeared in [FPI+ 17].
LOGICAL METHODS
IN COMPUTER SCIENCE
© Y.M.Y. Feldman, O. Padon, N. Immerman, M. Sagiv, and S. Shoham
Creative Commons
DOI:10.2168/LMCS-???
1
2
Y.M.Y. FELDMAN, O. PADON, N. IMMERMAN, M. SAGIV, AND S. SHOHAM
1. Introduction
This paper addresses a fundamental problem in automatic program verification: how to
prove that a piece of code preserves a given invariant. In Floyd-Hoare style verification this
means that we want to automatically prove the validity of the Hoare triple {P }C{P } where
P is an assertion and C is a command. Often this is shown by proving the unsatisfiability
of a formula of the form P (V ) ∧ δ(V, V 0 ) ∧ ¬P (V 0 ) (the verification condition) where P (V )
denotes the assertion P before the command, P (V 0 ) denotes the assertion P after the
command, and δ(V, V 0 ) is a two-vocabulary formula expressing the meaning of the command
C as a transition relation between pre- and post-states. When C is a loop body, such a P is
an inductive invariant and can be used to prove safety properties of the loop (if it also holds
initially and implies the desired property).
For infinite-state programs, proving the validity of {P }C{P } is generally undecidable
even when C does not include loops. Indeed, existing Satisfiability Modulo Theory (SMT)
solvers can diverge even for simple assertions and simple commands. Recent attempts to apply
program verification to prove the correctness of critical system’s design and code [HHK+ 15]
identify this as the main hurdle for using program verification.
The difficulty is rooted in powerful constructs used in the SMT-based verification
of interesting programs. Prominent among these constructs are arithmetic and other
program operations modelled using background theories, and logical quantifiers. In this
paper we target the verification of applications in which the problem can be modelled
without interpreted theories. This is in line with recent works that show that although
reasoning about arithmetic is crucial for low-level code, in many cases the verification of
high-level programs and designs can be performed by reasoning about quantification in
uninterpreted theories. Specifically, the decidable Effectively Propositional logic (EPR) has
been successfully applied to application domains such as linked-list manipulation [IBI+ 13],
Software-Defined Networks [BBG+ 14] and some distributed protocols [PMP+ 16, PLSS17].
Without interpreted theories it remains to address the complications induced by the use
of quantifiers, and specifically by the use of alternating universal (∀) and existential (∃)
quantifiers.
In the presence of quantifier alternation, the solver’s ability to check assertions is
hindered by the following issues:
(1) An infinite search space of proofs that must be explored for correct assertions. A
standard form of proofs with quantified formulas is instantiation, in which the solver
attempts to replace universal quantifiers by a set of ground terms. The problem of
exploring the inifnite set of candidates for instantiation is sometimes manifested in
matching loops [DNS05].
(2) A difficulty of finding counterexamples for invalid assertions, notably when counterexamples may be of infinite size. Current SMT techniques often fail to produce models of
satisfiable quantified formulas [GM09, RTG+ 13]. This is somewhat unfortunate since
one of the main values of program verification is the early detection of flaws in designs
and programs. The possibility of infinite counterexamples is a major complication in
this task, as they are especially difficult to find. In uninterpreted domains, infinite
counterexamples usually do not indicate a real violation and are counterintuitive to
programmers, yet render assertions invalid in the context of general first-order logic
(on which SMT proof techniques are based). Hence infinite counter-models pose a real
problem in the verification process.
BOUNDED QUANTIFIER INSTANTIATION FOR CHECKING INDUCTIVE INVARIANTS
3
Previous works on EPR-based verification [IBI+ 13, BBG+ 14, PMP+ 16] used universally
quantified invariants with programs expressed by ∃∗ ∀∗ formulas (EPR programs)1. In that
setting, checking inductive invariants is decidable, hence problems (1) and (2) do not occur.
In particular, EPR enjoys the finite-model property, and so counterexamples are of finite
size. EPR programs are in fact Turing-complete [PMP+ 16], but universal invariants are not
always sufficient to express the program properties required for verification.
For example, [HHK+ 15] describes a client-server scenario with the invariant that “For
every reply message sent by the server, there exists a corresponding request message sent by
a client”. (See Example 3.6 for further details.) This invariant is ∀∗ ∃∗ and thus leads to
verification conditions with quantifier alternation. This kind of quantifier alternation may
lead to divergence of the solver as problems (1) and (2) re-emerge.
This paper aims to expand the applicability of the EPR-based verification approach to
invariants of more complex quantification. We focus on the class of ∀∗ ∃∗ invariants. ∀∗ ∃∗
invariants arise in interesting programs, but, as we show, checking inductiveness of invariants
in this class is undecidable. We thus study problems (1),(2) above for this setting using the
notion of bounded quantifier instantiation, a technique we term Bounded-Horizon.
Main results. This paper explores the utility of limited quantifier instantiations for checking
∀∗ ∃∗ invariants, and for dealing with the problems that arise from quantifier alternation:
divergence of the proof search and infinite counter-models.
We consider instantiations that are bounded in the depth of terms. Bounded instantiations
trivially prevent divergence while maintaining soundness. Although for a given bound the
technique is not complete, i.e. unable to prove every correct invariant, we provide completeness
guarantees by comparing bounded instantiations to the method of instrumentation, a
powerful technique implicitly employed in previous works [IBI+ 13, KBI+ 17, PMP+ 16].
Instrumentation tackles a ∀∗ ∃∗ invariant by transforming the program in a way that allows
the invariant to be expressed using just universal quantifiers, and, accordingly, makes the
verification conditions fall in EPR. We show that for invariants that can be proven using a
typical form of instrumentation, bounded instantiations of a small bound are also complete,
meaning they are sufficiently powerful to prove the original program without modifications
and in a fully automatic way. This is encouraging since instrumentation is labor-intensive
and error-prone while bounded instantiations are completely automatic.
This result suggests that in many cases correct ∀∗ ∃∗ invariants of EPR programs can
be proven using a simple proof technique. Typically in such cases existing tools such as
Z3 will also manage to automatically prove the verification conditions. However, bounded
instantiations guarantee termination a-priori even when the invariant is not correct. In this
case, when the bounded instantiation procedure terminates, it returns a logical structure
which satisfies all the bounded instantiations. This structures is not necessarily a true
counterexample but “approximates” one. Interestingly, this capability suggests a way to
overcome the problem of infinite models. This problem arises when the user provides an
invariant that is correct for finite models but is not correct in general first-order logic. In
such cases, state-of-the-art SMT solvers typically produce “unknown” or timeout as they
fail to find infinite models. The user is thus left with very little aid from the solver when
attempting to make progress and successfully verify the program. In contrast, bounded
quantifier instantiation can be used to find finite models with increasing sizes, potentially
1 ∃∗ ∀∗ transition relations can be extracted from code by existing tools for C code manipulating linked
lists [IBI+ 13, IBR+ 14, KBI+ 17] and for the modeling language RML [PMP+ 16] which is Turing-complete.
4
Y.M.Y. FELDMAN, O. PADON, N. IMMERMAN, M. SAGIV, AND S. SHOHAM
indicating the existence of an infinite model, and provide hints as to the source of the error.
This information allows the user to modify the program or the invariant to exclude the
problematic models. We demonstrate this approach on a real example in which such a
scenario occurred in one of our verification attempts. We show that the provided models
assist in identifying and fixing the error, allowing the user to successfully verify the program.
We also implemented a prototype tool that performs bounded instantiations of bound 1,
and used it to verify several distributed protocols and heap-manipulating programs. The
implementation efficiently reduces the problem of checking inductiveness with bound 1 to a
Z3 satisfiability check on which the solver always terminates, thereby taking advantage of
Z3’s instantiation techniques while guaranteeing termination.
Outline. The rest of the paper is organized as follows: Section 2 provides some technical
background and notations. In Section 3 we define the Bounded-Horizon algorithm and
discuss its basic properties. Section 4 defines the concept of instrumentation as used in this
work, and shows that Bounded-Horizon with a low bound is at least as powerful. Section 5
relates instrumentation to bounded instantiation in the converse direction, showing that
other forms of instrumentation can simulate quantifier instantiation of arbitrarily high depth.
In Section 6 we show how bounded instantiations can be used to tackle the problem of
infinite counterexamples to induction when the verification conditions are not valid. Section 7
describes our implementation of Bounded-Horizon of bound 1, and provides initial evaluation
of its ability to prove some examples correct by bound 1 instantiation. Section 8 discusses
related work, and Section 9 concludes. The discussion of the undecidability of checking
inductiveness of ∀∗ ∃∗ invariants is deferred to Appendix A.
2. Preliminaries
In this section we provide background and explain our notation. Σ will always denote
a relational first-order vocabulary, which may contain constant symbols, ci , and relation
symbols, rj , but no function symbols. For a formula ϕ we denote by const[ϕ] the set of
constants that appear in ϕ. We write that ϕ ∈ ∃∗ (Σ) to mean that ϕ is an existential
formula defined over vocabulary Σ. Similarly, the class of universal formulas is denoted by
∀∗ (Σ). We say that ϕ is quantifier-free, denoted ϕ ∈ QF(Σ) if it contains no quantifiers, and
that it is alternation free, denoted ϕ ∈ AF(Σ), if it can be written as a Boolean combination
of formulas in ∃∗ (Σ). FOL(Σ) stands for arbitrary first-order formulas over Σ. A term or
formula is ground if it does not contain free variables. A sentence is a ground formula.
EPR. The effectively-propositional (EPR) fragment of first-order logic, also known as the
Bernays-Schönfinkel-Ramsey class, consists of ∃∗ ∀∗ (Σ) sentences. Such sentences enjoy the
small model property; in fact, a satisfiable EPR sentence has a model of size no larger than
the number of its constants plus existential quantifiers. Thus satisfiability of EPR sentences
is decidable [Ram30].
EPR Transition Relation. We specify a transition relation via an EPR sentence, δ, over
a vocabulary Σ ] Σ0 where Σ is a relational vocabulary used to describe the source (or pre-)
state of a transition and Σ0 = {a0 | a ∈ Σ} is used to describe the target (or post-) state.
BOUNDED QUANTIFIER INSTANTIATION FOR CHECKING INDUCTIVE INVARIANTS
5
Inductive Invariants. A first-order sentence I over Σ is an inductive invariant for δ if
I ∧ δ → I 0 is valid, or, equivalently, if I ∧ δ ∧ ¬I 0 is unsatisfiable2, where I 0 results from
substituting every constant and relation symbol in I by its primed version (i.e. I 0 ∈ FOL(Σ0 )).
Counterexample to Induction. Given a first-order sentence I over Σ and transition
relation δ (over Σ ] Σ0 ), a counterexample to induction is a structure A (over Σ ] Σ0 ) s.t.
A |= I ∧ δ ∧ ¬I 0 .
Skolemization. Let ϕ(z1 , . . . , zn ) ∈ FOL(Σ). The Skolemization of ϕ, denoted ϕS , is a
universal formula over Σ ] ΣS , where ΣS consists of fresh constant symbols and function
symbols, obtained as follows. We first convert ϕ to negation normal form (NNF) using
the standard rules. For every existential quantifier ∃y that appears under the scope of the
universal quantifiers ∀x1 , . . . , ∀xm , we introduce a fresh function symbol fy ∈ ΣS of arity
n + m. We replace each bound occurrence of y by fy (z1 , . . . , zn , x1 , . . . , xm ), and remove the
existential quantifier. If n + m = 0 (i.e., ϕ has no free variables and ∃y does not appear in
the scope of a universal quantifier) a fresh constant symbol is used to replace y. It is well
known that ϕS → ϕ is valid and ϕS and that ϕ are equi-satisfiable.
3. Bounded-Horizon
In this section, we define a systematic method of quantifier instantiation called BoundedHorizon as a way of checking the inductiveness of first-order logic formulas, and explore
some of its basic properties.
Undecidability. We first justify the use of sound but incomplete algorithms, such as the
Bounded-Horizon algorithm, for the problem of checking inductiveness of ∀∗ ∃∗ formulas. For
a universal formula I ∈ ∀∗ (Σ), the formula I ∧ δ ∧ ¬I 0 is in EPR (recall that δ is specified in
EPR). Hence, checking inductiveness amounts to checking the unsatisfiability of an EPR
formula, and is therefore decidable. The same holds for I ∈ AF (Σ). However, this is no
longer true when quantifier alternation is introduced. In Appendix A we show that checking
inductiveness of ∀∗ ∃∗ formulas is indeed undecidable, even when the transition relation
is restricted to EPR (see Theorem A.3). Thus, an attempt to check inductiveness must
sacrifice either soundness, completeness, or termination. Techniques based on quantifier
instantiation usually prefer completeness over termination. In contrast, the Bounded-Horizon
algorithm guarantees termination a-priori, possibly at the expense of completeness (but is
surprisingly powerful nonetheless). We now move to define the Bounded-Horizon algorithm
for and discuss its basic properties in checking inductiveness.
Bounded-Horizon Instantiations. Let δ ∈ ∃∗ ∀∗ (Σ, Σ0 ) be an EPR transition relation
and I ∈ FOL(Σ) a candidate invariant. We would like to check the satisfiability of I ∧ δ ∧ ¬I 0 ,
and equivalently of Ind = IS ∧ δS ∧ (¬I 0 )S . Recall that ϕS denotes the Skolemization of
ϕ, and note that IS and (¬I 0 )S possibly add Skolem functions to the vocabulary. Roughly
speaking, for a given k ∈ N, Bounded-Horizon instantiates the universal quantifiers in Ind,
while restricting the instantiations to produce ground-terms of function nesting at most k.
We then check if this (finite) set of instantiations is unsatisfiable; if it is already unsatisfiable
then we have a proof that I is inductive. Otherwise we report that I is not known to be
2 In this paper, unless otherwise stated, satisfiability and validity refer to general models and are not
restricted to finite models. Note that for EPR formulas, finite satisfiability and general satisfiability coincide.
6
Y.M.Y. FELDMAN, O. PADON, N. IMMERMAN, M. SAGIV, AND S. SHOHAM
inductive. The idea is to choose a (preferably small) number k and perform instantiations
bounded by k instead of full-blown instantiation. As we will show, this algorithm is sound
but not necessarily complete for a given k.
Below we provide the formal definitions. We start with the notion of instantiations,
and recall Herbrand’s theorem which establishes completeness of proof by unrestricted
instantiations. Suppose that some vocabulary Σ̃ including constants and function symbols is
understood (e.g., Σ̃ = Σ ] ΣS , where ΣS includes Skolem constants and function symbols).
Definition 3.1 (Instantiation). Let ϕ(x) ∈ ∀∗ (Σ̃) be a universal formula with n free variables
and m universal quantifiers. An instantiation of ϕ by a tuple t of n + m ground terms,
denoted by ϕ[ t ], is obtained by substituting t for the free variables and the universally
quantified variables, and then removing the universal quantifiers.
Note that an instantiation is a quantifier-free sentence.
∗
Theorem 3.2
n (Herbrand’s Theorem). Let ϕ ∈ ∀ (Σ̃). Then
o ϕ is satisfiable iff the (potentially
infinite) set ϕ[ t ] | t is a tuple of ground terms over Σ̃ is satisfiable.
We now turn to restrict the depth of terms used in instantiations.
Definition 3.3 (Bounded-Depth Terms). For every k ∈ N, we define BHTk to be the set
of ground terms over Σ̃ with function symbols nested to depth at most k. BHTk is defined
by induction over k, as follows. Let C be the set of constants in Σ̃, F the set of functions,
and for every f ∈ F let Arityf be the arity of f . Then
BHT0 = C
BHTk = BHTk−1 ∪ {f (t1 , . . . tm ) | f ∈ F, m = Arityf , t1 , . . . , tm ∈ BHTk−1 }.
We will also write t ∈ BHTk for a tuple of terms t, to mean that every entry of t is in
BHTk (the number of elements
in t should be clear from the context). Note that the set of
S
ground terms is BHT∞ = k∈N BHTk .
Definition 3.4 (Depth of Instantiation). Let ϕ ∈ ∀∗ (Σ̃) and t ∈ BHT∞ . The depth of
instantiation, denoted depth(ϕ[ t ]), is the smallest k such that all ground terms that appear
in ϕ[ t ] are included in BHTk .
We are now ready to define the algorithm and discuss its basic soundness and completeness properties.
Bounded-Horizon algorithm. Given a candidate invariant I ∈ FOL(Σ), a transition
relation δ over Σ ] Σ0 , and k ∈ N, the Bounded-Horizon algorithm constructs the formula
Ind = IS ∧ δS ∧ (¬I 0 )S , and checks if the set
(3.1)
Ind[ t ] | t ∈ BHTk , depth(Ind[ t ]) ≤ k
is unsatisfiable. If it is unsatisfiable, then I is provably inductive w.r.t. δ with BoundedHorizon of bound k. Otherwise we report that I is not known to be inductive.
Note that the satisfiability check performed by Bounded-Horizon is decidable since the
set of instantiations is finite, and each instantiation is a ground quantifier-free formula.
BOUNDED QUANTIFIER INSTANTIATION FOR CHECKING INDUCTIVE INVARIANTS
7
Bounded-Horizon for ∀∗ ∃∗ Invariants. We illustrate the definition of Bounded-Horizon in
the case that I ∈ ∀∗ ∃∗ (Σ). Let I = ∀x. ∃y. α(x, y) where α ∈ QF. Then IS = ∀x. α(x, f (x))
where f are new Skolem function symbols. δS introduces Skolem constants but no function
symbols, and in this case so does (¬I 0 )S . The Bounded-Horizon check of bound k can be
approximately3 understood as checking the (un)satisfiability of
^
^
^
IS [ t ] ∧
δS [ t ] ∧
¬I 0 S [ t ] .
(3.2)
t∈BHTk−1
t∈BHTk
t∈BHTk
Lemma 3.5 (Soundness). For every k ∈ N, Bounded-Horizon with bound k is sound, i.e., if
Bounded-Horizon of bound k reports that I ∈ FOL(Σ) is inductive w.r.t. δ, then I is indeed
inductive.
Proof. Assume that I is not inductive w.r.t. δ, so there is a structure A such that A |=
IS ∧ δS ∧ (¬I 0 )S . In particular A |= Ind[ t ] for every t ∈ BHT∞ and in particular for every
t ∈ BHTk such that depth(Ind[ t ]) ≤ k. Hence, Bounded-Horizon of bound k will not report
that I is inductive.
As the algorithm is sound for any k, the crucial question that remains is an appropriate
choice of k. A small k is preferable for efficiency, but a larger k could allow for proving more
invariants. In the following example, a bound of even 1 suffices for proving that the invariant
is inductive. We then show that for every correct invariant there is a suitable bound k, but
a single choice of k cannot prove all correct invariants. Later, in Section 4, we show that
bound of 1 or 2 is surprisingly powerful nonetheless.
Example 3.6. Figure 1 presents a simple model of the client-server scenario described
in [HHK+ 15]. The program induces an EPR transition relation, and its invariant is provable
by Bounded-Horizon of bound 1.
We first explain this example while ignoring the annotations denoted by “/@”. The
system state is modeled using three binary relations. The req relation stores pairs of users
and requests, representing requests sent by users. The resp relation similarly stores pairs
of users and replies, representing replies sent back from the server. The match relation
maintains the correspondence between a request and its reply.
The action new request models an event where a user u sends a new request to the
server. The action respond models an event where the server responds to a pending request
by sending a reply to the user. The request and response are related by the match relation.
The action check is used to verify the safety property that every response sent by the server
has a matching request, by aborting the system if this does not hold.
A natural inductive invariant for this system is
I = ∀u, p. resp(u, p) → ∃q. req(u, q) ∧ match(q, p).
The invariant proves that the then branch in action check will never happen and thus the
system will never abort. This invariant is preserved under execution of all actions, and this
fact is provable by Bounded Horizon of bound 1.
3 Equation (3.2) is an under-approximation of the set of instantiations used for bound k; variables that do
not appear in IS under a function symbol can be taken from BHTk in the conjunction without increasing
the total depth of instantiation beyond k, and are therefore allowed in bounded instantiation of bound k.
This approximation is illustrative nonetheless, and will be useful in the proofs in Section 4.
8
Y.M.Y. FELDMAN, O. PADON, N. IMMERMAN, M. SAGIV, AND S. SHOHAM
req := ∅; resp := ∅; match := ∅;
action new request(u) {
q := new request;
req := req ∪ {(u, q)}
/@ r := r ∪ {(u, y) | match(q, y)}
}
action respond(u, q) {
assume req(u, q);
p := new response;
match := match ∪ {(q, p)};
/@ r := r ∪ {(x, p) | req(x, q)}
resp := resp ∪ {(u, p)}
}
action check(u, p) {
if resp(u, p) ∧ ∀q. req(u, q) → ¬match(q, p)
/@ ,→ if resp(u, p) ∧ ¬r(u, p)
then abort
}
Invariant I = ∀u, p. resp(u, p) →
∃q. req(u, q) ∧ match(q, p)
/@ r(x, y) ≡ ∃z. req(x, z) ∧ match(z, y)
/@ Invariant Ib = ∀u, p. resp(u, p) → r(u, p)
Figure 1. Example demonstrating a ∀∗ ∃∗ invariant that is provable with
bound 1. The reader should first ignore the instrumentation code denoted by
/@ (see Example 4.1). This example models a simple client-server scenario,
with the safety property that every response sent by the server was triggered
by a request from a client. Verification of this example requires a ∀∗ ∃∗
invariant. This example is inspired by [HHK+ 15]. The complete program is
provided in [add] (files client server ae.ivy, client server instr.ivy).
Lemma 3.7 (Completeness for some k). For every I ∈ FOL(Σ) and δ such that I is
inductive w.r.t. δ there exists a finite k ∈ N s.t. I is provably inductive w.r.t. δ with
Bounded-Horizon of bound k.
Proof. From Theorem 3.2 and compactness there is a finite unsatisfiable set S of instantiations.
Take k to be the maximal depth of the instantiations in S.
For example, if I ∈ ∀∗ then Bounded-Horizon of bound 0 is complete. However, as
expected due to the undecidability of checking inductiveness (see Appendix A), BoundedHorizon is not complete for a given k for arbitrary invariants.
Example 3.8. An example of a program and an inductive invariant for which a bound of 0
or 1 is insufficient appears in Figure 2. In this example the server operates as a middleman
between clients and the database (DB), and is used to anonymize user requests before they
reach the database. The server performs a translation p between clients’ identity and an
anonymous unique id, sends a translated request to the DB, and forwards the DB’s response
to the clients. The safety property is that every response sent by the server was triggered by
a request from a client. The inductive invariant states, in addition to the safety property,
that every server request to the DB was triggered by a client’s request from the server, and
that every DB response was triggered by a server’s request. Proving that the invariant is
inductive under the action server recv db response requires the prover to understand
that for the response from the DB there is a matching request from the server to the DB,
and that for this request to the DB there is a matching request from the client to the server.
Every such translation requires another level of nesting in the instantiation. In this example,
a bound of 2 manages to prove inductiveness. This example can be lifted to require an even
larger depth of instantiation by adding more translation entities similar to the server, and
describing the invariant in a similar, modular, way.
BOUNDED QUANTIFIER INSTANTIATION FOR CHECKING INDUCTIVE INVARIANTS
9
req := ∅; resp := ∅;
db req := ∅; db resp := ∅; t := ∅;
action new request(u) {
action server recv request(u, q) {
q := new request;
assume req(u, q);
req := req ∪ {(u, q)}
id := new DB request id;
}
t := t ∪ {(id , u)};
action db recv request(id , q) {
db req := db req ∪ {(id , q)}
assume db req(id , q);
}
p := DB response ensuring db(q, p);
db resp := db resp ∪ {(id , p)}
}
action server recv db response(id , p) {
action check(u, p) {
assume db resp(id , p);
assume resp(u, p);
resp := resp ∪ {(x, p) | t(id , x)}
if ∀q. req(u, q) → ¬db(q, p)
}
then abort
}
Invariant I = ∀u, p. resp(u, p) → ∃q. req(u, q) ∧ db(q, p) ∧
∀id , q. db req(id , q) → ∃u. t(id , u) ∧ req(u, q) ∧
∀id , p. db resp(id , p) → ∃q. db req(id , q) ∧ db(q, p) ∧
∀id , u1 , u2 . t(id , u1 ) ∧ t(id , u2 ) → u1 = u2
Figure 2. Example demonstrating a ∀∗ ∃∗ invariant that is provable only
with bound 2. The server anonymizes requests from clients to the database (DB) and forwards the answer to the client. The server performs a
translation t between clients’ identity and an anonymous unique id. The
safety property is that every response sent by the server to a client was
triggered by a request from the client. The inductive invariant further states
that every server request to the DB was triggered by a client’s request from
the server, and that every DB response was triggered by a server’s request.
The complete program corresponding to this Figure appears in [add] (file
client server db ae.ivy).
Small Bounded-Horizon for ∀∗ ∃∗ Invariants. Despite the incompleteness, we conjecture
that a small depth of instantiations typically suffices to prove inductiveness. The intuition
is that an EPR transition relation has a very limited “horizon” of the domain: it interacts
only with a small fraction of the domain, namely elements pointed to by program variables
(that correspond to logical constants in the vocabulary).
When performing the Bounded-Horizon check with bound 1 on a ∀∗ ∃∗ invariant I =
∀x. ∃y. α(x, y), we essentially assume that the existential part of the invariant ψ(x) =
∃y. α(x, y) holds on all program variables — but not necessarily on all elements of the
domain — and try to prove that it holds on all elements of the domain after the transition.
We expect that for most elements of the domain, the correctness of ψ is maintained simply
because they were not modified at all by the transition. For elements that are modified by
the transition, we expect the correctness after modification to result from the fact that ψ
holds for the elements of the domain that are directly involved in the transition. If this is
indeed the reason that ψ is maintained, a bound of 1 sufficiently utilizes ψ in the pre-state
to prove the invariant in the post-state, i.e. to prove that it is inductive.
10
Y.M.Y. FELDMAN, O. PADON, N. IMMERMAN, M. SAGIV, AND S. SHOHAM
This is the case in Example 3.6. Additional examples are listed in Section 7. The
example of Figure 2 itself also admits a different invariant that is provable by bound 1.
Section 4 further studies the power of Bounded-Horizon with a low bound.
4. Power of Bounded-Horizon for Proving Inductiveness
We now turn to investigate the ability of Bounded-Horizon to verify inductiveness. In
this section we provide sufficient conditions for its success by relating it to the notion of
instrumentation (which we explain below). We show that Bounded-Horizon with a low
bound of 1 or 2 is as powerful as a natural class of sound program instrumentations, those
that do not add existential quantifiers. Section 7 demonstrates the method’s power on
several interesting programs that we verified using Bounded-Horizon of bound 1.
4.1. Instrumentation. We present our view of the instrumentation procedure used in
previous works [IBI+ 13, KBI+ 17, PMP+ 16] to eliminate the need for quantifier-alternation,
thus reducing the verification task to a decidable fragment. The procedure begins with a
program that induces a transition relation δ ∈ ∃∗ ∀∗ (Σ ∪ Σ0 ). The purpose of instrumentation
is to modify δ into another transition relation δb that admits an inductive invariant with
simpler quantification (e.g., universal, in which case it is decidable to check). We note
that instrumentation is generally a manual procedure. For simplicity, we describe the
instrumentation process informally, but provide the semantic soundness requirement in
Definition 4.2. The instrumentation process is thoroughly described in a recent work [PLSS17].
The instrumentation procedure consists of the following three steps:
(1) Identify a formula ψ(x) ∈ FOL(Σ) (usually ψ will be existential) that captures information that is needed in the inductive invariant. Extend the vocabulary with an
instrumentation relation r(x) that intentionally should capture the derived relation
b = Σ ∪ {r} denote the extended vocabulary4.
defined by ψ(x). Let Σ
(2) Add update code that updates r when the original (“core”) relations are modified, and
maintains the meaning of r as encoding ψ. The update code must not block executions
of real code, and can possibly be a sound approximation. Sometimes it can be generated
automatically via finite differencing [RSL10].
(3) Modify the program to use r. Often this is performed by rewriting some program
conditions, keeping in mind that r encodes ψ. This means replacing some quantified
expressions by uses of r.
Example 4.1. In the example of Figure 1, to achieve a universal invariant we add an
instrumentation relation r defined by r(x, y) ≡ ∃z. req(x, z) ∧ match(z, y) (step 1). The
simple form of ψ allows us to obtain precise update code, which appears as annotations
marked with /@ in lines that mutate req and match (step 2). We also replace the if condition
in the action check by an equivalent condition that uses r (step 3). The line marked with
/@ ,→ in the check action replaces the line above it. The resulting program has the invariant
Ib = ∀u, p. resp(u, p) → r(u, p), which is universal.
4It is also possible to instrument the program with constants rather than relations. This can be emulated
by adding a unary relation c(x) representing the constant, and adding the assumption that c contains exactly
one element to the invariant. This form is in line with the conditions of Theorem 4.12.
BOUNDED QUANTIFIER INSTANTIATION FOR CHECKING INDUCTIVE INVARIANTS
11
b ∪Σ
b 0 ) denote the transition relation induced by the modified program
Let δb ∈ ∃∗ ∀∗ (Σ
(modifications occur in steps 2,3). The soundness of the instrumentation procedure is
b
formalized in the following connection between ψ, δ, and δ:
b ∪Σ
b 0 ) is a sound instrumentation for
Definition 4.2 (Sound Instrumentation). δb ∈ ∃∗ ∀∗ (Σ
∗
∗
0
δ ∈ ∃ ∀ (Σ ∪ Σ ) and ψ ∈ FOL(Σ) if
b
δ → δ[ψ/r,
ψ 0 /r0 ]
is valid.
Definition 4.2 ensures that the instrumented program includes at least all the behaviors
of the original program, when r is interpreted according to ψ. Thus, if the instrumented
program is safe, then it is sound to infer that the original program is safe.
The instrumentation procedure does not require the user to know an inductive invariant
for the original program. However, if a sound instrumentation which leads to an invariant
exists, then an inductive invariant for the original δ can be produced by substituting back
the “meaning” of r as ψ (thus, safety of the original program is implied):
b be an inductive
Lemma 4.3. Let δb be a sound instrumentation for δ and ψ, and Ib ∈ FOL(Σ)
b
b
invariant for δ. Then I = I[ψ/r] is inductive w.r.t. δ.
Proof. Ib ∧ δb → Ib0 is valid, thus, so is (Ib ∧ δb → Ib0 )[ψ/r, ψ 0 /r0 ]. δb is a sound instrumentation
for δ, so (using Definition 4.2) I ∧ δ → I 0 is valid.
b
Note that typically the quantification structure of I is more complex than that of I.
b
Remark 4.4. In the expression δ[ψ/r,
ψ 0 /r0 ] the update code of r in δb becomes a constraint
over the core relations in Σ. In a sound instrumentation this constraint is required to follow
from the way the core relations are updated by δ, essentially meaning that r is updated in a
way that is consistent with its interpretation as ψ.
Instrumentation without additional existential quantifiers. In order to relate instrumentation to Bounded-Horizon instantiations, we consider the typical case where the
b This happens
instrumentation process of δ does not add new existential quantifiers to δ.
when the update code does not introduce additional existential quantifiers. Formally:
b Σ
b 0 ).
Definition 4.5 (Existential Naming). Let δb = ∃z1 , . . . zm . ϕ(z1 , . . . , zm ) where ϕ ∈ ∀∗ (Σ,
b
b
An existential naming η for (δ, δ) is a mapping η : {z1 , . . . , zm } → const[δS ] ∪ const[δS ]. We
b to be ϕ[η(z1 )/z1 , . . . , η(zm )/zm ].
define η(δ)
An existential naming provides a Skolemization procedure which uses existing constants
rather than fresh ones. If such η exists, it maps the (Skolemized) existential quantifiers in δb
to their counterparts in δ. For example, the instrumentation in Figure 1 results in δb that has
an existential naming w.r.t. the original δ. Note that it is possible that δb has in fact fewer
existential quantifiers than δ, for example due to the rewriting of conditions (as happens in
the example of Figure 1 — see the if statement in action check).
An instrumentation without additional existentials is an instrumentation from δ to
b the
δb whose soundness can be shown while respecting an existential naming from δ to δ;
existential naming matches the existential quantifiers of δ with the vocabulary of δb in a
sound way:
12
Y.M.Y. FELDMAN, O. PADON, N. IMMERMAN, M. SAGIV, AND S. SHOHAM
Definition 4.6 (Instrumentation Without Additional Existenials). δb is a sound instrumentation without additional existentials for δ if there exists an existential naming η such
that
b
δS → η(δ)[ψ/r,
ψ 0 /r0 ]
is valid.
4.2. From Instrumentation to Bounded-Horizon. The results described in this section
show that if there is an instrumentation without additional existentials, then BoundedHorizon with a low bound is able to prove the original invariant, without specific knowledge
of the instrumentation and without manual assistance from the programmer. This is the case
in the example of Figure 1, which admits an instrumentation that transforms the invariant
to a universal invariant (see Example 4.1) in a form that matches Theorem 4.9, and indeed
the original invariant is provable by Bounded-Horizon of bound 1.
Interestingly, in case Bounded-Horizon with a small bound does not prove inductiveness
the results imply that either the invariant is not inductive or no instrumentation that does
not add existential quantifiers can be used to show that it is inductive (even with the
programmer’s manual assistance). This is the case in the example of Figure 2, where a
bound of 1 does not suffice.5
While we show that instrumentation that does not add existentials is at most as powerful
as Bounded-Horizon with a low bound, sound instrumentations that do add existentials
to the program (thereby not satisfying Definition 4.6) can be used to simulate quantifier
instantiation of an arbitrary depth. This topic is explored in Section 5.
In the remainder of this section we will assume that δb is a sound instrumentation without
additional existentials for δ, and η is the corresponding naming of existentials. Further, we
b
assume that Ib is an inductive invariant for δb and denote I = I[ψ/r].
Results. We now state the results whose proofs are presented in the rest of this section.
Theorem 4.9 and Theorem 4.10 consider I ∈ ∀∗ ∃∗ that is transformed to Ib ∈ ∀∗ . In
Theorem 4.9 we show that a bound of 1 suffices to prove that I is inductive for δ when
ψ ∈ ∃∗ (that is, the instrumentation defining formula is existential) and the instrumentation
b or when ψ ∈ ∀∗ and r appears only negatively in I.
b
relation r appears only positively in I,
This is an attempt to explain the success of bound 1 instantiations in proving our examples
(see Section 7). In Theorem 4.10 we show that a bound of 2 suffices in the more general
b
setting of ψ ∈ AF (with no restriction on appearances of r in I).
Theorem 4.12 considers a generalization to I that is 1-alternation and transformed to
Ib ∈ AF. We show that a bound of 2 suffices in this case.
5Strictly speaking this shows that there is no such instrumentation where the instrumentation relation
appears only positively in the invariant, which is the most common case. Examples that require an even
larger bound (sketched above) do not admit any instrumentation without additional existential quantifiers
that transforms the invariant to a universal form.
BOUNDED QUANTIFIER INSTANTIATION FOR CHECKING INDUCTIVE INVARIANTS
13
Proof idea. The rest of the section is devoted to proofs of the these claims. The intuition
behind the proofs is that the instrumented invariant is universal, so the fact that it is inductive
can be shown by instantiating the universal quantifiers with the constants. Relating between
the instrumented and original program, this constitutes a proof that the instrumented
invariant is inductive (for the original program), where the proof is by instantiating the
universal quantifiers with the constants, essentially by the same set of instantiations. With
the existential quantifiers present in the program without instrumentation, this implies
a proof by bounded instantiations. The formal proofs handle the fine details of which
quantifiers are instantiated, and with what constants, to establish the results.
To highlight the main points in the formal proof, the crux of the argument is as follows:
Assume for the sake of contradiction that I cannot be shown to be inductive for δ by BoundedHorizon of a low bound, and take a counterexample to induction of the instantiated I (see
Equation (4.2) in the proof of Lemma 4.8). By the assumption that δb is an instrumentation
without additional existentials for δ, we can utilize properties of substitution to obtain a
counterexample to induction for the instantiated Ib w.r.t. δb (see Equation (4.5)). By the
assumption that Ib ∈ ∀∗ and δb ∈ ∃∗ ∀∗ , we use complete instantiation to argue that we
have obtained a true counterexample to induction of Ib w.r.t. δb (see Equation (4.9)), in
contradiction to the premise.
Remark 4.7. The results of this section also apply when multiple instrumentation relations
ψ1 , . . . , ψt ∈ FOL(Σ) are simultaneously substituted instead of the relation symbols r1 , . . . , rt
b
in δb and I.
4.3. Power for ∀∗ ∃∗ Invariants. We now establish that low bounds are sufficient for
the Bounded-Horizon check, assuming that a sound instrumentation without additional
b and I ∈ ∀∗ ∃∗ . To do so, we first prove the
existentials exists, in the case of Ib ∈ ∀∗ (Σ)
following lemma.
Lemma 4.8. Let δb be a sound instrumentation of δ without new existentials and with
b and let α = α
naming η. Write Ib = ∀x. α
b(x) where α
b ∈ QF(Σ)
b[ψ/r]. Then,
^
0
α(c) ∧ δS ∧ (¬I )S
(4.1)
c∈C n
is unsatisfiable, where C = const[δS ∧ (¬I 0 )S ] and n is the number of universal quantifiers in
b
I.
Proof. Assume not, i.e., there exists a structure A0 such that,
^
A0 |=
α(c) ∧ δS ∧ (¬I 0 )S .
(4.2)
c∈C n
b Let C
b ∧ (¬Ib0 ) ]. Then,
b = const[η(δ)
We will show that Ib is not inductive for δ.
S
^
0
A1 |=
α(c) ∧ δS ∧ (¬I )S
(4.3)
b n
c∈(C∪C)
b \ C as some arbitrary
where A1 is the same as A0 but also interprets any constant in C
constant in C. Thus α(c) holds for the new constants as well.
14
Y.M.Y. FELDMAN, O. PADON, N. IMMERMAN, M. SAGIV, AND S. SHOHAM
Removing some conjuncts from Equation (4.3), we get,
^
A1 |=
α(c) ∧ δS ∧ (¬I 0 )S .
(4.4)
bn
c∈C
By assumption (Definition 4.6), it follows that,
^
b
A1 |=
α(c) ∧ η(δ)[ψ/r,
ψ 0 /r0 ] ∧ (¬I 0 )S .
(4.5)
bn
c∈C
Recall that I 0 = Ib0 [ψ 0 /r0 ]. Since A1 |= (¬Ib0 [ψ 0 /r0 ])S , it follows that A1 |= (¬Ib0 )S [ψ 0 /r0 ].
In the latter formula, some existentially quantified variables from ψ or ¬ψ may remain,
whereas in the former formula they were replaced by Skolem constants. Thus this is just a
corollary of the fact that γS → γ is valid for any γ.
Thus we have shown,
^
b ∧ (¬Ib0 ) [ψ/r, ψ 0 /r0 ].
(4.6)
A1 |=
α
b(c) ∧ η(δ)
S
bn
c∈C
Now, consider the structure Ab that expands A1 by interpreting r and r0 the way that A1
interprets ψ and ψ 0 , respectively. Then,
^
b ∧ (¬Ib0 ) .
α
b(c) ∧ η(δ)
(4.7)
Ab |=
S
bn
c∈C
Since the formula in Equation (4.7) is universal, it is also satisfied by Ab|Cb , the substructure
b Ab, i.e., A’s
b interpretation of the constant symbols C
b (recall that
of Ab with universe C
0
b
b
b
C = const[η(δ) ∧ (¬I )S ]). Thus,
b ∧ (¬Ib0 ) .
Ab|Cb |= ∀x. α
b(x) ∧ η(δ)
(4.8)
S
b → δb (for the same reasons), we know,
Finally, since γS → γ is valid and so is η(δ)
Ab|Cb |= Ib ∧ δb ∧ ¬Ib0 .
(4.9)
b
But this contradicts the fact that Ib is inductive for δ.
The following results are corollaries of Lemma 4.8.
b or ψ ∈ ∀∗
Theorem 4.9. Let Ib ∈ ∀∗ . Assume ψ ∈ ∃∗ and r appears only positively in I,
b Then I = I[ψ/r]
b
and r appears only negatively in I.
is inductive for δ with Bounded-Horizon
∗
∗
of bound 1. (Note that I ∈ ∀ ∃ .)
Proof. Let Ib = ∀x. α
b(x) where α
b ∈ QF. In both cases of the claim α = α
b[ψ/r] ∈ ∃∗ , and so
b This implies that the satisfiability check of
all the universal quantifiers in I are those of I.
Lemma 4.8 is simply the Bounded-Horizon satisfiability check with bound 1, and it shows
that the result must be unsatisfiable.
More formally, assume for the sake of contradiction that I is not inductive w.r.t. δ with
Bounded-Horizon of bound 1. Let α(x) = ∃y1 , . . . , ym . θ(x, y1 , . . . , ym ) where θ ∈ QF, and
let
αS (x) = θ(x, f1 (x), . . . , fm (x))
BOUNDED QUANTIFIER INSTANTIATION FOR CHECKING INDUCTIVE INVARIANTS
15
be its Skolemization with fresh Skolem function symbols f1 , . . . , fm (introduced for y1 , . . . , ym ,
respectively). Then there is a structure A satisfying
^
^
^
αS [ t ] ∧
δS [ t ] ∧
¬I 0 S [ t ] .
(4.10)
t∈BHT0
t∈BHT1
t∈BHT1
Since αS has no universal quantifiers, the instantiation is just a substitution of the free
variables, and A satisfies
^
^
^
αS (t) ∧
δS [ t ] ∧
¬I 0 S [ t ] .
(4.11)
t∈BHT0
t∈BHT1
t∈BHT1
By reducing A to the elements pointed to by BHT1 terms we have that
^
A|BHT1 |=
αS (t) ∧ δS ∧ ¬I 0 S
(4.12)
t∈BHT0
Note that in A|BHT1 the interpretations of the Skolem functions are possibly partial functions.
The functions appear in the formula of Equation (4.12) only grounded, and applied on BHT0 ,
and these cases the interpretations of the functions are defined. (In particular, they can be
extended to total functions in an arbitrary way, and the resulting structure still satisfies
Equation (4.12).)
We now move from the Skolem functions back to existential quantifiers. By the valuation
that to every existentially quantified variable yi in α assigns the interpretation of fi (t) in
A|BHT1 (recall that fi (t) appears in αS instead of the quantifier ∃yi in α), we know that
^
A|BHT1 |=
(4.13)
α(t) ∧ δS ∧ ¬I 0 S .
t∈BHT0
Recall that BHT0 = const[δS ∧
¬I 0
A|BHT1 |=
S
]. Therefore, Equation (4.13) can be rewritten as
^
α(c) ∧ δS ∧ ¬I 0 S
(4.14)
c∈C n
0
b
where C n = const[δS ∧ ¬I S ]n and is the number of universal quantifiers in I (and I).
b and
By Lemma 4.8 this is a contradiction to the assumption that Ib is inductive w.r.t. δ,
the claim follows.
b
Theorem 4.10. Let Ib ∈ ∀∗ . If ψ ∈ AF then I = I[ψ/r]
is inductive for δ with Bounded∗
∗
Horizon of bound 2. (Note that I ∈ ∀ ∃ .)
Proof. As before, Let Ib = ∀x. α
b(x) where α
b ∈ QF. ψ ∈ AF implies that α = α
b[ψ/r] ∈ AF .
Let
α(x) = ∀v1 θ1,1 (x, v1 ) ∨ ∃z1 θ1,2 (x, z1 ) ∧ . . .
(4.15)
∧ ∀vr θr,1 (x, vr ) ∨ ∃zr θr,2 (x, zr )
where θ1,1 , θ1,2 , . . . , θr,1 , θr,2 ∈ QF.
Assume for the sake of contradiction that I is not inductive w.r.t. δ with Bounded-Horizon
of bound 2.
For brevity denote
^
^
ξ(k) =
δS [ t ] ∧
¬I 0 S [ t ]
t∈BHTk
t∈BHTk
16
Y.M.Y. FELDMAN, O. PADON, N. IMMERMAN, M. SAGIV, AND S. SHOHAM
and let gi (x) denote the fresh Skolem function introduced for zi in αS .
By the assumption that inductiveness is not provable using Bounded-Horizon of bound
2,
^
αS [ t ] ∧ ξ(2)
(4.16)
t∈BHT1
is satisfiable by a structure A.
In particular
^
θ1,1 (c, d1 ) ∨ θ1,2 (c, g1 (c)) ∧
. . . ∧ θr,1 (c, dr ) ∨ θr,2 (c, gr (c)) ∧ ξ(1)
(4.17)
is satisfied by A, where the conjunction is over c ∈ BHT0 and d1 , . . . , dr ∈ BHT1 (with
the relevant arity). Note that the difference from Equation (4.16) is that there are fewer
conjuncts here, because the full Bounded-Horizon check with bound 2 has conjuncts for each
c ∈ BHT1 and not just BHT0 , and we have ξ(1) instead of ξ(2).
Reduce A to the elements pointed by BHT1 terms, let A↓ = A|BHT1 .
Now,
^
∀v1 θ1,1 (c, v1 ) ∨ ∃z1 θ1,2 (c, z1 ) ∧ . . .
c∈BHT0
(4.18)
∧ ∀vr θr,1 (c, vr ) ∨ ∃zr θr,2 (c, zr ) ∧ ξ(1)
is satisfied by A↓ . This is because:
• The universal quantifiers are semantically equivalent to a conjunction over all BHT1
elements because the domain was reduced, and
• The existential quantifiers are justified by the following valuation: the valuation assigns
every zi the interpretation of gi (c).
With this valuation the conjunctions of formula 4.18 are all guaranteed by the conjunctions
in formula 4.17.
Now formula 4.18 exactly means that
^
A↓ |=
α(c) ∧ ξ(1).
(4.19)
c∈BHT0
As in the proof of Theorem 4.9, using Lemma 4.8, this is a contradiction to the assumption
b and the claim follows.
that Ib is inductive w.r.t. δ,
4.4. Generalization to 1-Alternation Invariants. We now generalize the results of
Section 4.3 to 1-alternation invariants. A formula is 1-alternation if it can be written as
b and I = I[ψ/r]
b
a Boolean combination of ∀∗ ∃∗ formulas. In the sequel, Ib ∈ AF(Σ)
∈
1-alternation(Σ).
b be an inductive invariant for δb ∈ ∃∗ ∀∗ (Σ).
b
Lemma 4.11. Let ψ ∈ FOL(Σ). Let Ib ∈ AF(Σ)
b = ∀x. α
b Let α1 = α
Write IbS = ∀x. α
c1 (x) and (¬I)
c2 (x), where α
c1 , α
c2 ∈ QF(Σ).
c1 [ψ/r] and
S
BOUNDED QUANTIFIER INSTANTIATION FOR CHECKING INDUCTIVE INVARIANTS
17
b Σ
b 0 ) be a sound instrumentation of δ
α2 = α
c2 [ψ/r].
Let δ ∈ ∃∗ ∀∗ (Σ) and let δb ∈ ∃∗ ∀∗ (Σ,
without new existentials and with naming η. Then,
^
^
α1 (c1 ) ∧ δS ∧
α2 (c2 )
(4.20)
c1 ∈C n
c2 ∈C m
is unsatisfiable, where C = const[IbS ∧ δS ∧ (¬I 0 )S ], n is the number of universal quantifiers
b .
in IbS and m is the number of universal quantifiers in (¬I)
S
Proof. The proof is similar to that of Lemma 4.8, with the transformations applied to the
conjunction are performed both on the invariant in the pre-state and in the post-state.
Assume not, i.e., there exists a structure A0 such that,
^
^
A0 |=
α1 (c1 ) ∧ δS ∧
α2 (c2 ) .
(4.21)
c1 ∈C n
c2 ∈C m
b Let C
b ∧ (¬Ib0 ) ]. Then,
b = const[IbS ∧ η(δ)
We will show that Ib is not inductive for δ.
S
^
^
A1 |=
(4.22)
α1 (c1 ) ∧ δS ∧
α2 (c2 ) .
bn
c1 ∈C
bm
c2 ∈ C
b \ C as some arbitrary
where A1 is the same as A0 but also interprets any constant in C
constant in C.
By the assumption (Definition 4.6), it follows that,
^
^
b
A1 |=
α1 (c1 ) ∧ η(δ)[ψ/r,
ψ 0 /r0 ] ∧
α2 (c2 ) .
(4.23)
bn
c1 ∈ C
Thus we have shown,
^
b ∧
α
c1 (c1 ) ∧ η(δ)
A1 |=
bn
c1 ∈C
bm
c2 ∈C
α
c2 (c2 ) [ψ/r, ψ 0 /r0 ].
^
(4.24)
bm
c2 ∈C
Now, consider the structure Ab that expands A1 by interpreting r and r0 the way that A1
interprets ψ and ψ 0 , respectively. Then,
^
^
b ∧
Ab |=
α
c1 (c1 ) ∧ η(δ)
α
c2 (c2 )
(4.25)
bn
c1 ∈ C
bm
c2 ∈ C
Since the formula in Equation (4.25) is universal, it is also satisfied by Ab|Cb , the subb Ab, i.e., A’s
b interpretation of the constant symbols C.
b Thus,
structure of Ab with universe C
b ∧ ∀x. α
Ab|Cb |= ∀x. α
c1 (x) ∧ η(δ)
c2 (x) .
(4.26)
b ∧ (¬Ib0 ) ].
b was defined as C
b = const[Ib ∧ η(δ)
Recall that C
S
b → δb is valid, we know,
Finally, since γS → γ is valid and for the same reasons η(δ)
Ab|Cb |= Ib ∧ δb ∧ ¬Ib0 .
(4.27)
b
But this contradicts the fact that Ib is inductive for δ.
18
Y.M.Y. FELDMAN, O. PADON, N. IMMERMAN, M. SAGIV, AND S. SHOHAM
The following result is a corollary of Lemma 4.11.
b
Theorem 4.12. Let Ib ∈ AF. If ψ ∈ AF then I = I[ψ/r]
is inductive for δ with BoundedHorizon of bound 2. (Note that I ∈ 1-alternation.)
c1 [ψ/r] ∈ AF and α2 (x) = α
c2 [ψ/r] ∈ AF (recall that
Proof. ψ ∈ AF implies that α1 (x) = α
α
c1 , α
c2 ∈ QF).
By the assumption that inductiveness is not provable using Bounded-Horizon of bound
2,
^
^
^
(α1 )S [ t ] ∧
δS [ t ] ∧
(α2 )S [ t ]
(4.28)
t∈BHT1
t∈BHT2
t∈BHT1
is satisfiable, where (α1 )S , (α2 )S introduce Skolem functions. Let A be such a satisfying
structure, and A↓ = A|BHT1 .
Because α1 ∈ AF, in the same way as in the proof of Theorem 4.10,
^
A↓ |=
(4.29)
α1 (c)
c∈BHT0
and in the same way, since α2 ∈ AF as well, the same structure has
^
A↓ |=
α2 (c).
(4.30)
c∈BHT0
b Overall we have
Note that from Equation (4.28), A↓ |= δ.
^
^
A↓ |=
α1 (c) ∧ δS ∧
c∈BHT0
α2 (c)
(4.31)
c∈BHT0
b
and by Lemma 4.11 this is a contradiction to the assumption that Ib a is inductive w.r.t. δ.
5. Instrumentation for High Depth Instantiations
In this section we discuss the connection between quantifier instantiation and program
instrumentation in the converse direction, i.e. simulating quantifier instantiation with
instrumentation. In Section 4 we showed that instrumentation without adding existential
quantifiers is powerful at most as bounded instantiations with a low bound. In this section
we show that allowing additional existentials does increase the power of instrumentation in
proving ∀∗ ∃∗ invariants. In particular, we show how to systematically construct instrumented
programs in a way that corresponds to quantifier instantiation for ∀∗ ∃∗ -invariants: performing
the required instantiations within the program allows expressing the invariant in a form that
falls in the decidable fragment. Together with Section 4, this makes the point that, in the
context of invariant checking, the form of instrumentation by a derived relation studied in
this paper directly corresponds to quantifier instantiation.
As before, the instrumentation process begins with a program that does not have a
universal inductive invariant. We would like to transform the program in a sound way to a
b As before, we start by identifying some
program that has a universal inductive invariant I.
∗
existential formula ψ(x) ∈ ∃ (Σ) that expresses needed information, and encode it using an
instrumentation relation r, with the meaning that “r(x) ≡ ψ(x)”. As before, we present
the procedure for a single formula and instrumentation relation, but it can be used with
multiple formulas and instrumentation relations (see Remark 4.7).
BOUNDED QUANTIFIER INSTANTIATION FOR CHECKING INDUCTIVE INVARIANTS
19
b express existential
Intuitively, adding the instrumentation relation r lets Ib ∈ ∀∗ (Σ)
information by referring to r. The modifications to the program must encode enough
of “r(x) ≡ ψ(x)” to make Ib inductive. As opposed to the instrumentation described in
Section 4.1, this will be done by instantiating the correspondence between ψ(t) and r(t)
for specific variables in the program, using assume statements. This replaces the rewriting
of program conditions that use ψ to use r. This approach is also extended to obtain deep
instantiations which simulate Bounded-Horizon with an arbitrary bound.
Local Instantiations. Rather than rewriting ψ(t) to r(t) in the program, we would like to
enforce r to be interpreted according to ψ in the pre-state, i.e., to enforce ∀x. ψ(x) ↔ r(x).
The direction ∀x. ψ(x) → r(x) is an EPR formula (since ψ ∈ ∃∗ ), and thus we can simply
conjoin it to the verification conditions without sacrificing decidability.
The converse implication, ∀x. r(x) → ψ(x), is a ∀∗ ∃∗ formula, and adding it to the
verification condition will lead to undecidability. Note that Bounded-Horizon with bound
1 is analogous to enforcing r(t) → ψ(t) for every t that is a program variable. Inspired by
this, we define the following instrumentation that lets the user locally enforce the definition
of r for program variables.
Definition 5.1 (Local Instantiation). Let ψ(x) = ∃y. ϕ(x, y). To generate an instantiation
of ∀x. r(x) → ∃y. ϕ(x, y) on some tuple of program variables t, we instrument the program
by adding new program variables c, and inserting the following code:
c := ∗; assume r(t) → ϕ(t, c)
This code uses the havoc statement, which sets the value of c to arbitrary values, followed
by an assume statement that restricts the execution such that if r(t) holds, then ϕ(t, c)
holds (see [PMP+ 16] for the formal semantics). Thus, this code realizes the restriction
that r(t) → ∃y. ϕ(t, y), and also assigns to the new program variables c, the witnesses for
the existential quantifiers. We call this addition to the program a local instantiation, as it
imposes the connection between r and ψ locally for some program variables t.
Lemma 5.2 (Soundness of Local Instantiations). If δb is obtained from δ by a local instantiation then δb is a sound instrumentation by Definition 4.2.
Proof. The code added by a local instantiation for t that uses new variables c translates a
new constraint in δb of the form γ = ∃c. r(t) → ϕ(t, c). Since ∀x. r(x) ↔ ψ(x) ⇒ γ, we have
b which implies the condition of Definition 4.2.
∀x. r(x) ↔ ψ(x) ∧ δ ⇒ δ,
Remark 5.3. The combination of adding ∀x. ψ(x) → r(x) to the verification condition and
allowing the user to perform local instantiations on the program variables is at least as
powerful as rewriting program conditions, since any rewrite of ψ(t) to r(t) can be simulated
by a local instantiation on t.
Instantiations for the Invariant. The mechanism of local instantiations is designed to
b
support instantiations of ∀x. ψ(x) ↔ r(x) required to prove that the invariant I = I[ψ/r]
is
preserved by the program. This proof is carried out by showing that (∀x. ψ(x) ↔ r(x)) ∧
0 is unsatisfiable. This may require instantiating the definition of r on
b
Ib ∧ δ ∧ ¬(I[ψ/r])
Skolem constants that come from the negation of the invariant in the post-state. Thus, we
extend our instrumentation method by adding new “program variables” that represent the
elements of the domain on which the invariant is potentially violated in the post-state. For
20
Y.M.Y. FELDMAN, O. PADON, N. IMMERMAN, M. SAGIV, AND S. SHOHAM
/@ r1 (x, y) ≡ ∃z. req(x, z) ∧ db(z, y)
/@ r2 (x, y) ≡ ∃z. t(x, z) ∧ req(z, y)
/@ r3 (x, y) ≡ ∃z. db req(x, z) ∧ db(z, y)
/@ Invariant Ib = ∀u, p. resp(u, p) → r1 (u, p) ∧
/@
∀id , q. db req(id , q) → r2 (id , q) ∧
/@
∀id , p. db resp(id , p) → r3 (id , p) ∧
/@
∀id , u1 , u2 . t(id , u1 ) ∧ t(id , u2 ) → u1 = u2
action server process db response(id , p) {
# instantiate r3 on (id , skp ) (depth 1)
/@ c1 := ∗;
/@ assume r3 (id , skp ) → db req(i, c1 ) ∧ db(c1 , skp );
# instantiate r2 on (id , c1 ) (depth 2)
/@ c2 := ∗;
/@ assume r2 (id , c1 ) → t(id , c2 ) ∧ req(c2 , c1 );
...
}
action check(u, p) {
# instantiate r1 on (u, p) (depth 1)
/@ c3 := ∗;
/@ assume r1 (u, p) → req(u, c3 ) ∧ db(c3 , p);
...
}
Figure 3. An illustration of instrumentation by local instantiations for
the example of Figure 2. The instrumentation adds three instrumentation
relations r1 , r2 , r3 , and performs three local instantiations in order to prove
that the invariant is inductive. Note that an instantiation depth of 2 is used,
in accordance with the fact that the original invariant is provable using bound
2 but not bound 1. The complete model corresponding to this Figure appears
in [add] (file client server db instr.ivy).
b we add a special program variable skx which
every universally quantified variable x in I,
can be used in local instantiations, enhancing their power to prove that Ib is inductive.
Obtaining Deep Instantiations. Applying local instrumentation on a tuple t that consists
of original program variables, or variables that represent Skolem constants, corresponds
to instantiations of Bounded-Horizon with bound 1. However, once a local instantiation
is performed, new program variables c are added. Performing a local instantiation on
these new variables now corresponds to instantiation from depth 2. By iteratively applying
local instantiations, where each iteration adds new program variables, we can thus obtain
quantifier instantiations of arbitrary depth.
Illustrating Example. Figure 3 illustrates the local instantiation procedure on the example
of Figure 2. Recall that the ∀∗ ∃∗ invariant I of Figure 2 is not provable using BoundedHorizon of bound 1, but is provable using Bounded-Horizon of bound 2. The instrumentation
BOUNDED QUANTIFIER INSTANTIATION FOR CHECKING INDUCTIVE INVARIANTS
21
presented in Figure 3 introduces three instrumentation relations to encode the existential
b
parts of I, thereby producing the instrumented universal invariant I.
b we use local instantiations in the actions check and
To prove the inductiveness of I,
server process db response. In server process db response, we instantiate the definition of r3 on (id , skp ), and assign the existential witness to c1 . Intentionally, c1 gets the
request that was sent from the server to the DB that led to the response skp being sent
from the DB to the server. skp is the response that supposedly causes a violation of the
invariant when the action server process db response is executed (the instantiations are
used to prove that a violation does not occur). This instantiation is of depth 1. Next, we
make an instantiation of the definition of r2 on (id , c1 ) and obtain a new existential witness
c2 . The use of c1 here makes this instantiation depth 2. The check action includes another
instantiation of depth 1, which is simply used to prove that the abort cannot happen
(similarly to rewriting a program condition).
The reader can observe that the local instantiations introduced during the instrumentation process closely correspond to the instantiations required to prove that the original ∀∗ ∃∗
invariant I (see Section 3).
We note that the process of instrumentation by local instantiations discussed here is
different in spirit from those of Section 4: instrumentation by local instantiation consists
of almost nothing but adding existential quantifiers to the transition relation, as opposed
to the condition in Definition 4.6 where we do not allow the instrumentation to add new
existential quantifiers.
6. Partial Models for Understanding Non-Inductiveness
When conducting SMT-based deductive verification (e.g., using Dafny [Lei10]), the user
constructs both the formal representation of the system and its invariants. In many cases,
the invariant I is initially not inductive w.r.t. the given program, due to a bug in the program
or in the invariant. Therefore, deductive verification is typically an iterative process in which
the user attempts to prove inductiveness, and when this fails the user adapts the program,
the invariant, or both.
In such scenarios, it is extremely desirable to present the user with a counterexample
to induction in the form of a state that satisfies I but makes a transition to a state that
violates it. Such a state can be obtained from a model of the formula Ind = I ∧ δ ∧ ¬I 0 which
is used to check inductiveness. It explains the error, and guides the user towards fixing the
program and/or the invariant [Lei10, FLL+ 02]. However, in many cases where the check
involves quantifier alternation, current SMT solvers are unable to produce counterexamples.
Instead, SMT solvers usually diverge or report “unknown” [GM09, RTGK13]. In such cases,
Bounded-Horizon instantiations can be used to present a concrete logical structure which is
comprehensible to the user, obtained as a model of the (finite) instantiations of the formula
Ind. While this structure is not a true counterexample (as it is only a model of a subset of
the instantiations of the formula), it can still guide the user in the right direction towards
fixing the program and/or the invariant.
We illustrate this using a simple leader-election protocol in a ring [CR79], whose model
is presented in Figure 4. The protocol assumes that nodes are organized in a directional ring
topology with unique IDs, and elects the node with the highest ID as the leader. Each node
sends its own ID to its successor, and forwards messages when they contain an ID higher
than its own ID. A node that receives its own ID is elected as leader. We wish to prove a
22
Y.M.Y. FELDMAN, O. PADON, N. IMMERMAN, M. SAGIV, AND S. SHOHAM
action receive packet(n, m) {
assume pending(m, n)
pending := pending \ {(m, n)}
if m = n then
leader := leader ∪ {n}
else
if n < m then
assume ring next(n, n0 )
pending := pending ∪ {(m, n0 )}
else # do not forward
}
pending := ∅;
. . . # ring topology
action send packet(n) {
assume ring next(n, m)
pending := pending ∪ {(n, m)}
sent := sent ∪ {n}
}
Figure 4. A sketch of the leader election protocol discussed in this section
(the complete program appears in [add], file ring leader termination.ivy).
(a)
(b)
(c)
Figure 5. Leader-election in a ring protocol as an illustration of the use of
partial models for incorrect programs and invariants. (a),(b) show partial
models of bound 1 and 2, respectively, and (c) illustrates an infinite structure
that explains the root cause of the non-inductiveness.
termination property which states that once all nodes have sent their ID, and there are no
pending messages in the network, there must be an elected leader. To verify this we use
a relational model of the protocol, similar to [PMP+ 16], and specify the property via the
following formula:
(∃n. leader(n)) ∨ (∃n1 , n2 . ¬sent(n1 ) ∨ pending(n1 , n2 ))
(6.1)
A natural attempt of proving this using an inductive invariant is by conjoining Equation (6.1)
(which is not inductive by itself) with the following property (this was the authors’ actual
next step in proving this termination property):
∀n1 . sent(n1 ) ∧ ¬leader(n1 ) → ((∃n2 . pending(n1 , n2 )) ∨ (∃n2 . n1 < n2 ))
(6.2)
BOUNDED QUANTIFIER INSTANTIATION FOR CHECKING INDUCTIVE INVARIANTS
23
meaning that if a node has sent its own ID but has not (yet) become leader, then there is
either a message pending in the network with the node’s ID, or a node with a higher ID.
Alas, the conjunction of Equations (6.1) and (6.2) is still not an inductive invariant for
the protocol (as we explain below). Since Equation (6.2) contains ∀∗ ∃∗ quantification, the
associated inductiveness check is outside of the decidable EPR fragment. Indeed, Z3 diverges
when it is used to check Ind. This is not surprising since the formula has no satisfying finite
structures, but has an infinite model (a scenario that is not unusual for ∀∗ ∃∗ formulas).
On the other hand, applying Bounded-Horizon (with any bound) to Ind results in a
formula that has finite models. These concrete models are partial models of Ind. Figs. 5(a)
and (b) show partial models (restricted to the pre-states) obtained with bounds of 1 and 2,
respectively, on this example.
These models are not true counterexamples to induction: the sub-formula of Equation (6.2) residing under the universal quantifier does not hold for all the elements of the
domain. It does, however, hold for all elements with which the quantifier was instantiated,
which are the elements above the dashed line in the figure. These elements have all sent
their own ID, which was blocked by their successor that has a higher ID, so none of them is
the leader. In a finite model, this has to end somewhere, because one of the nodes must
have the highest ID. Hence, no finite counter-model exists. However, extrapolating from
Figure 5(a) and (b), we can obtain the infinite model depicted in Figure 5(c). This model
represents an infinite (“open”) ring in which each node has a lower ID than its successor.
This model is a true model of the formula Ind generated by the invariant in Equations (6.1)
and (6.2), but the fact that it is infinite prevented Z3 from producing it.
Since we use tools that check general (un)satisfiability, which is not limited to finite
structures, the only way to prove that an invariant is inductive is to exclude infinite
counterexamples to induction as well. Using Bounded-Horizon instantiations, we are able to
obtain meaningful partial models that provide hints to the user of what is missing. In this
case, the solution is to add an axiom to the system model which states that there is a node
with maximal ID: ∃n1 . ∀n2 . n2 ≤ n1 . With this additional assumption, the formula Ind
is unsatisfiable so the invariant is inductive, and this is proven both by Z3’s instantiation
heuristics and by Bounded-Horizon with a bound of 1. This illustrates the usefulness of
Bounded-Horizon when the invariant is not inductive.
7. Implementation and Initial Evaluation
In this section, we describe our implementation of Bounded-Horizon of bound 1 and present
an initial evaluation of the technique on several examples. While discussing the examples, we
note that certain types of ghost code that records properties of the history of an execution can
be viewed as a special case of instrumentation, and demonstrate how the Bounded-Horizon
technique circumvents the need for augmenting the program with such ghost code.
7.1. Implementation. We implemented a prototype of Bounded-Horizon of bound 1 on
top of Z3 [DMB08] and used it within Ivy [PMP+ 16] and the framework of [IBI+ 13].
Our implementation works by adding “guards” that restrict the range of universal
quantifiers to the set of constants where necessary. Technically, recall that we are considering
the satisfiability of Ind = IS ∧ δS ∧ (¬I 0 )S .6 Let ∀x. θ be a subformula of Ind. If θ contains
6 Skolemization is performed via Z3, taking advantage of heuristics that reduce the number of different
Skolem functions.
24
Y.M.Y. FELDMAN, O. PADON, N. IMMERMAN, M. SAGIV, AND S. SHOHAM
W
function symbol applications7, we transform the subformula to ∀x.
c x = c → θ where
c ranges over const[Ind]. The resulting formula is then dispatched to the solver. This is a
simple way to encode the termination criterion of bound 1 while leaving room for the solver
to perform the necessary instantiations cleverly. The translation enlarges the formula by
O(#Consts · #∀) although the number of bounded instantiations grows exponentially with
#∀. The exponential explosion is due to combinations of constants in the instantiation, a
problem we defer to the solver.
Z3 terminates on the class of formulas because during the Model-Based Quantifier
Instantiation process every instantiation of a universally quantified formula has the same
truth value in the model as an instantiation using one of the existing ground terms (constants
and then BHT1 terms). Z3’s instantiation engine will produce instantiations using existing
terms rather than create superfluous new terms [Bjø].
An alternative approach of implementation is to integrate the termination criterion of
the bound with the solver’s heuristics more closely (see [BRK+ 15]).
7.2. Examples. We applied the procedure to the incorrect example of Section 6, and also
successfully verified several correct programs and invariants using bound 1. These examples
are (the examples’ code can be found in [add]):
• The client-server example of Figure 1.
• List reverse [IBI+ 13], where the invariant states that the n edges (“next” pointers) are
reversed. The invariant is ∀∗ ∃∗ due to the encoding of n via n∗ as explained in [IBI+ 13].
• Learning switch [BBG+ 14], in which the routing tables of the switches in the network are
automatically constructed through observing the source packets arriving on each switch’s
links. The invariant states that in every routing path, every routing node has a successor.
• Hole-punching firewall [BBG+ 14], in which the firewall allows a packet from an external
host to enter the network only if the host is recorded as trusted by the firewall. The invariant
states that if a packet from an external host then is allowed, then there previously was an
internal host that contacted the external host. We explored two modeling alternatives:
using a ghost history relation, or existentially quantifying over time. We elaborate on this
topic and this example below.
• Leader election in a ring [CR79, PMP+ 16] with the invariant discussed in Section 6. (See
Section 6 for full details.)
Derived Relations Over the History. Sometimes expressing verification conditions
requires modifications to the program in which the state is augmented with additional
relations, but these relations cannot be defined as derived relations over the core program
relations. Often the reason is that the property of interest depends not only on the current
program state, but also on previous states in the execution history. An example where such
a scenario occurs is a hole-punching firewall. In this example, a firewall controls packets
entering and leaving the organization’s network. Packets from outside the organization are
allowed if they originate from a host that is considered trusted. A host hosto is considered
trusted if some host hosti in the organization’s network has previously sent a packet to hosti .
This correctness condition depends on previous states in the execution history, and cannot
be expressed in an inductive invariant without changing the vocabulary.
7 This in fact implements the approximation as of Equation (3.2). The exact bound 1 per Equation (3.1)
can be implemented by a more careful consideration of which universally quantified variables should be
restricted, but this was not necessary for our examples.
BOUNDED QUANTIFIER INSTANTIATION FOR CHECKING INDUCTIVE INVARIANTS
25
A common method to overcome this problem is to introduce a ghost state to record historical information, and ghost code to mutate the ghost state, implementing this book-keeping.
In the hole-punching firewall example, the user can add a relation ever-pending(host1 , host2 ),
and add ghost code that adds the tuple (host1 , host2 ) to ever-pending whenever host1 sends
a packet to host2 .
We observe that it is sometimes possible to think of such ghost relations as standard
derived relations, defined over a vocabulary in which past states are available, and view the
procedure of adding appropriate ghost code as an instrumentation by a derived relation as
defined in this paper (Definition 4.2). The ideas is as follows:
In our first-order setting, we can lift the vocabulary to encode the time explicitly with
a classical encoding (see e.g. [Aba89]), by adding a new parameter t to every relation and
constant symbol. in the transition relation, references to p(·) are replaced by p(tnow , ·) and
p0 (·) by p(t0now , ·), where tnow is a new constant that represents the current time. tnow is
incremented in every transition (according to some total order on time). The transition
relation also need to include the requirement that the previous states are not modified by
the current transition, meaning that ∀t < tnow . ∀x. p(t, x) ↔ p0 (t, x). Call this modified
transition relation δt .
We now proceed to express the ghost relation as a derived relation over the history. In
the example of the hole-punching firewall, the relation ever-pending can now be expressed
by the derived relation ψ(host1 , host2 ) ≡ ∃t. t ≤ tnow ∧ pending(t, host1 , host2 ).
With this construcrtion, we can directly use the derived relation in the inductive invariant
instead of the ghost relation. Viewing the addition of the ghost relation and ghost code as
a sound instrumentation, Lemma 4.3 implies that an inductive invariant for the program
with ghost state enduces an inductive invariant for δt . This invariant is defined over the
vocabulary that records the entire history, but without the ghost relation symbol.
The value of this point of view is that the results of this paper imply that in certain
cases the user can prove an invariant expressed over the history using bounded instantiations
with a low bound, without resorting to ghost code manipulations. The reasoning is as follows:
If adding the ghost code does not add existential quantifiers, then augmenting the program
with the ghost code can be thought of as instrumentation without additional existentials.
If, additionally, the derived relation — now defined over the entire execution history —
can be expressed as a combination of universal and existential properties, Theorem 4.10
applies, showing that Bounded-Horizon with a low-bound is therefore guaranteed to prove
the inductiveness of the invariant expressed for δt , with no need to add the ghost code.
These conditions are satisfied by the hole-punching firewall example. We manually
performed the transformation of the transition relation to a vocabulary over the history in
Ivy [PMP+ 16], and successfully proved the inductive invariant with Bounded-Horizon of
bound 1.
7.3. Initial Evaluation. An initial evaluation of the method’s performance appears in
Table 1.
26
Y.M.Y. FELDMAN, O. PADON, N. IMMERMAN, M. SAGIV, AND S. SHOHAM
Table 1. Experimental results.
Program
Client-server
List reverse
Learning switch
Hole-punching firewall with ghost
Hole-punching firewall ∃ time
Leader-election in a ring (correct)
Leader-election in a ring (incorrect)
#∀
14
47
70
15
32
41
40
#Func
1
3
1
1
2
1
1
#Consts
15
15
7
18
21
21
20
#∀↓
2
4
37
3
3
1
1
B1 Total
58 ms
319 ms
245 ms
75 ms
102 ms
113 ms
1112 ms
B1 Solve
3 ms
211 ms
65 ms
4 ms
4 ms
36 ms
1008 ms
Baseline Z3
3 ms
50 ms
33 ms
4 ms
4 ms
27 ms
—
B1 Total is the time in milliseconds for the bound 1 implementation. It is compared to Baseline Z3 which
is the solving time in milliseconds of Ind as is (with quantifier alternation) by Z3. B1 Solve measures the
solving time of the formula restricted to bound 1, which demonstrates that most of the overhead occurs when
constructing the formula. #∀ is the number of universal quantifiers in Ind, #Func the number of different
Skolem function symbols, and #Consts the number of constants. #∀↓ is the number of universally quantified
variables that were restricted in the bound 1 check. Measurements were performed on a 3.5GHz Intel i5-4690
CPU with 8GB RAM running Linux 3.13 x86 64.
The results are encouraging because they suggest that the termination strategy of
Bounded-Horizon, at least for bound 1, can be combined with existing instantiation techniques
to ensure termination with only a slight performance penalty. Most encouraging is the
satisfiable example of Section 6. On this instance, Z3 was able to return “sat” within seconds,
although to do so, in theory, the solver must exhaust the entire set of bounded instantiations.
This suggests that the Bounded-Horizon termination criterion might indeed be useful for
“sat” instances on which the solver may diverge.
8. Related Work
Quantifier instantiation. The importance of formulas with quantifier-alternation for
program verification has led to many developments in the SMT and theorem-proving
communities that aim to allow automated reasoning with such formulas. The Simplify
system [DNS05] promoted the practical usage of quantifier triggers, which let the user affect
the quantifier instantiation procedure in a clever way. Similar methods are integrated into
modern SMT solvers such as Z3 [DMB08]. Recently, a method for annotating the source
code with triggers has been developed for Dafny [LP16]. The notion of instantiation depth
is related to the notions of matching-depth [DNS05] and instantiation-level [GBT09] which
are used for prioritization within the trigger-based instantiation procedure.
In addition to user-provided triggers, many automated heuristics for quantifier instantiation have been developed, such as Model-Based Quantifier Instantiation [GM09]. Even
when quantifier instantiation is refutation-complete, it is still important and challenging to
handle the SAT cases, which are especially important for program verification. Accordingly,
many works (e.g., [RTGK13]) consider the problem of model finding.
Local Theory Extensions and Psi-Local Theories [Sof05, IJS08, BRK+ 15] identify settings
in which limited quantifier instantiations are complete. They show that completeness is
achieved exactly when every partial model can be extended to a (total) model. In such
settings Bounded-Horizon instantiations are complete for invariant checking. However,
Bounded-Horizon can also be useful when completeness cannot be guaranteed.
BOUNDED QUANTIFIER INSTANTIATION FOR CHECKING INDUCTIVE INVARIANTS
27
Classes of SMT formulas that are decidable by complete instantiations have been studied
by [GM09]. In the uninterpreted fragment, a refined version of Herbrand’s Theorem generates
a finite set of instantiations when the dependencies are stratified. Bounded-Horizon is a way
to bound unstratified dependencies.
Natural Proofs. Natural proofs [QGSM13] provide a sound and incomplete proof technique
for deductive verification. The key idea is to instantiate recursive definitions over the terms
appearing in the program. Bounded-Horizon is motivated by a similar intuition, but focuses
on instantiating quantifiers in a way that is appropriate for the EPR setting.
Decidable logics. Different decidable logics can be used to check inductive invariants. For
example, Monadic second-order logic [HJJ+ 95] obtains decidability by limiting the underlying
domain to consist of trees only, and in particular does not allow arbitrary relations, which
are useful to describe properties of programs. There are also many decidable fragments
of first-order logic [BGG01]. Our work aims to transcend the class of invariants checkable
by a reduction to the decidable logic EPR. We note that the example of Section 6 does
not fall under the Loosely-Guarded Fragment of first-order logic [Hod02] due to a use of a
transitivity axiom, and does not enjoy the finite-model property.
Abstractions for verification of infinite-state systems. Our work is closely related to
abstractions of infinite-state systems. These abstractions aim at automatically inferring
inductive invariants in a sound way. We are interested in checking if a given invariant is
inductive either for automatic and semi-automatic verification.
The View-Abstraction approach [AHH13, AHH14, AHH15] defines a useful abstraction
for the verification of parameterized systems. This abstraction is closely related to universally
quantified invariants. An extension of this approach [AHH14] adds contexts to the abstraction,
which are used to capture ∀∗ ∃∗ invariants in a restricted setting where nodes have finite-state
and are only related by specific topologies. Our work is in line with the need to use ∀∗ ∃∗
invariants for verification, but applies in a more general setting (with unrestricted high-arity
relations) at the cost of losing completeness of invariant checking.
Our work is related to the TVLA system [LS00, SRW02] which allows the programmers to
define instrumentation relations. TVLA also employs finite differencing to infer sound update
code for updating instrumentation relations [RSL10], but generates non-EPR formulas and
does not guarantee completeness. The focus operation in TVLA implements materialization
which resembles quantifier-instantiation. TVLA shows that very few built-in instrumentation
relations can be used to verify many different programs.
Instrumentation and update formulas. The idea of using instrumentation relations and
generating update formulas is not limited to TVLA and was also used for more predictable
SMT verification [LQ06, LQ08].
Instrumentation for Decidable Logics. The technique of instrumentation by derived
relations to allow decidable reasoning is further discussed in a recent work [PLSS17]. Several
variants of Paxos are proved safe using models in the decidable logic of EPR with stratified
functions, whose use is enabled by instrumentation. Efficiently mplementing a bounded
instantiations scheme that bounds instantiations only where they are unstratified is an
interesting challenge, in hope of relieving the need for instrumentation also for such cases.
28
Y.M.Y. FELDMAN, O. PADON, N. IMMERMAN, M. SAGIV, AND S. SHOHAM
9. Conclusion
We have provided an initial study of the power of bounded instantiations for tackling
quantifier alternation. This paper shows that quantifier instantiation with small bounds can
simulate instrumentation. This is a step in order to eliminate the need for instrumenting
the program, which can be error-prone. The other direction, i.e. simulating quantifier
instantiation with instrumentation, was also presented for conceptual purposes, although it
is less appealing from a practical point of view.
We are encouraged by our initial experience that shows that various protocols can be
proven with small instantiation bounds, and that partial models are useful for understanding
the failures of the solver to prove inductiveness. Some of these failures correspond to
non-inductive claims, especially those due to infinite counterexamples. In the future we
hope to leverage this in effective deductive verification tools, and explore meaningful ways
to display infinite counterexamples to the user. Other interesting directions include further
investigation into the automation of program transformations for the purpose of verification
(of which instrumentation is an example), including types of ghost code, and the use of
Bounded-Horizon for automatically inferring invariants with quantifier-alternation.
BOUNDED QUANTIFIER INSTANTIATION FOR CHECKING INDUCTIVE INVARIANTS
29
References
[Aba89]
[add]
[AHH13]
[AHH14]
[AHH15]
[BBG+ 14]
[BGG96]
[BGG01]
[Bjø]
[BRK+ 15]
[CR79]
[DMB08]
[DNS05]
[FLL+ 02]
[FPI+ 17]
[GBT09]
[GM09]
[HHK+ 15]
[HJJ+ 95]
[Hod02]
[IBI+ 13]
[IBR+ 14]
Martı́n Abadi. The power of temporal proofs. Theor. Comput. Sci., 65(1):35–83, June 1989.
Full code materials. http://www.cs.tau.ac.il/research/yotam.feldman/papers/tacas17/
examples_code.zip.
Parosh Aziz Abdulla, Frédéric Haziza, and Lukás Holı́k. All for the price of few. In Verification,
Model Checking, and Abstract Interpretation, 14th International Conference, VMCAI 2013, Rome,
Italy, January 20-22, 2013. Proceedings, pages 476–495, 2013.
Parosh Aziz Abdulla, Frédéric Haziza, and Lukáš Holı́k. Block me if you can! In International
Static Analysis Symposium, pages 1–17. Springer, 2014.
Parosh Abdulla, Frédéric Haziza, and Lukáš Holı́k. Parameterized verification through view
abstraction. International Journal on Software Tools for Technology Transfer, pages 1–22, 2015.
Thomas Ball, Nikolaj Bjørner, Aaron Gember, Shachar Itzhaky, Aleksandr Karbyshev, Mooly
Sagiv, Michael Schapira, and Asaf Valadarsky. Vericon: towards verifying controller programs in
software-defined networks. In ACM SIGPLAN Conference on Programming Language Design and
Implementation, PLDI ’14, Edinburgh, United Kingdom - June 09 - 11, 2014, page 31, 2014.
E. Börger, E. Grädel, and Y. Gurevich. The Classical Decision Problem. Springer-Verlag, 1996.
Egon Börger, Erich Grädel, and Yuri Gurevich. The classical decision problem. Springer Science
& Business Media, 2001.
Nikolaj Bjørner. personal communication.
Kshitij Bansal, Andrew Reynolds, Tim King, Clark W. Barrett, and Thomas Wies. Deciding local
theory extensions via e-matching. In Computer Aided Verification - 27th International Conference,
CAV 2015, San Francisco, CA, USA, July 18-24, 2015, Proceedings, Part II, pages 87–105, 2015.
Ernest Chang and Rosemary Roberts. An improved algorithm for decentralized extrema-finding
in circular configurations of processes. Communications of the ACM, 22(5):281–283, 1979.
L. De Moura and N. Bjørner. Z3: An efficient SMT solver. In TACAS, 2008.
David Detlefs, Greg Nelson, and James B. Saxe. Simplify: a theorem prover for program checking.
J. ACM, 52(3):365–473, 2005.
Cormac Flanagan, K. Rustan M. Leino, Mark Lillibridge, Greg Nelson, James B. Saxe, and
Raymie Stata. Extended static checking for java. In Proceedings of the 2002 ACM SIGPLAN
Conference on Programming Language Design and Implementation (PLDI), Berlin, Germany,
June 17-19, 2002, pages 234–245, 2002.
Yotam M. Y. Feldman, Oded Padon, Neil Immerman, Mooly Sagiv, and Sharon Shoham. Bounded
quantifier instantiation for checking inductive invariants. In TACAS, 2017.
Yeting Ge, Clark W. Barrett, and Cesare Tinelli. Solving quantified verification conditions using
satisfiability modulo theories. Ann. Math. Artif. Intell., 55(1-2):101–122, 2009.
Yeting Ge and Leonardo De Moura. Complete instantiation for quantified formulas in satisfiabiliby
modulo theories. In International Conference on Computer Aided Verification, pages 306–320.
Springer, 2009.
Chris Hawblitzel, Jon Howell, Manos Kapritsos, Jacob R. Lorch, Bryan Parno, Michael L. Roberts,
Srinath T. V. Setty, and Brian Zill. Ironfleet: proving practical distributed systems correct. In
Proceedings of the 25th Symposium on Operating Systems Principles, SOSP, pages 1–17, 2015.
Jesper G. Henriksen, Jakob L. Jensen, Michael E. Jørgensen, Nils Klarlund, Robert Paige, Theis
Rauhe, and Anders Sandholm. Mona: Monadic second-order logic in practice. In Tools and
Algorithms for Construction and Analysis of Systems, First International Workshop, TACAS ’95,
Aarhus, Denmark, May 19-20, 1995, Proceedings, pages 89–110, 1995.
Ian Hodkinson. Loosely guarded fragment of first-order logic has the finite model property. Studia
Logica, 70(2):205–240, 2002.
Shachar Itzhaky, Anindya Banerjee, Neil Immerman, Aleksandar Nanevski, and Mooly Sagiv.
Effectively-propositional reasoning about reachability in linked data structures. In CAV, volume
8044 of LNCS, pages 756–772, 2013.
Shachar Itzhaky, Nikolaj Bjørner, Thomas W. Reps, Mooly Sagiv, and Aditya V. Thakur.
Property-directed shape analysis. In Computer Aided Verification - 26th International Conference,
CAV 2014, Held as Part of the Vienna Summer of Logic, VSL 2014, Vienna, Austria, July 18-22,
2014. Proceedings, pages 35–51, 2014.
30
[IJS08]
Y.M.Y. FELDMAN, O. PADON, N. IMMERMAN, M. SAGIV, AND S. SHOHAM
Carsten Ihlemann, Swen Jacobs, and Viorica Sofronie-Stokkermans. On local reasoning in verification. In Tools and Algorithms for the Construction and Analysis of Systems, 14th International
Conference, TACAS 2008, Held as Part of the Joint European Conferences on Theory and Practice of Software, ETAPS 2008, Budapest, Hungary, March 29-April 6, 2008. Proceedings, pages
265–281, 2008.
[Imm99] Neil Immerman. Descriptive complexity. Graduate texts in computer science. Springer, 1999.
[IRR+ 04] Neil Immerman, Alexander Moshe Rabinovich, Thomas W. Reps, Shmuel Sagiv, and Greta Yorsh.
The boundary between decidability and undecidability for transitive-closure logics. In CSL, 2004.
[KBI+ 17] Aleksandr Karbyshev, Nikolaj Bjørner, Shachar Itzhaky, Noam Rinetzky, and Sharon Shoham.
Property-directed inference of universal invariants or proving their absence. J. ACM, 64(1):7:1–7:33,
2017.
[Lei10]
K Rustan M Leino. Dafny: An automatic program verifier for functional correctness. In Logic for
Programming, Artificial Intelligence, and Reasoning, pages 348–370. Springer, 2010.
[LP16]
K. Rustan M. Leino and Clément Pit-Claudel. Trigger selection strategies to stabilize program
verifiers. In Computer Aided Verification - 28th International Conference, CAV 2016, Toronto,
ON, Canada, July 17-23, 2016, Proceedings, Part I, pages 361–381, 2016.
Shuvendu K. Lahiri and Shaz Qadeer. Verifying properties of well-founded linked lists. In Proceed[LQ06]
ings of the 33rd ACM SIGPLAN-SIGACT Symposium on Principles of Programming Languages,
POPL 2006, Charleston, South Carolina, USA, January 11-13, 2006, pages 115–126, 2006.
Shuvendu K. Lahiri and Shaz Qadeer. Back to the future: revisiting precise program verification
[LQ08]
using SMT solvers. In Proceedings of the 35th ACM SIGPLAN-SIGACT Symposium on Principles
of Programming Languages, POPL 2008, San Francisco, California, USA, January 7-12, 2008,
pages 171–182, 2008.
[LS00]
Tal Lev-Ami and Shmuel Sagiv. TVLA: A system for implementing static analyses. In Static
Analysis, 7th International Symposium, SAS 2000, Santa Barbara, CA, USA, June 29 - July 1,
2000, Proceedings, pages 280–301, 2000.
[PLSS17] Oded Padon, Giuliano Losa, Mooly Sagiv, and Sharon Shoham. Paxos made epr: Decidable
reasoning about distributed protocols. Proc. ACM Program. Lang., 1(OOPSLA):108:1–108:31,
October 2017.
[PMP+ 16] Oded Padon, Kenneth L. McMillan, Aurojit Panda, Mooly Sagiv, and Sharon Shoham. Ivy: safety
verification by interactive generalization. In Proceedings of the 37th ACM SIGPLAN Conference
on Programming Language Design and Implementation, PLDI 2016, Santa Barbara, CA, USA,
June 13-17, 2016, pages 614–630, 2016.
[QGSM13] Xiaokang Qiu, Pranav Garg, Andrei Stefanescu, and Parthasarathy Madhusudan. Natural proofs
for structure, data, and separation. In ACM SIGPLAN Conference on Programming Language
Design and Implementation, PLDI ’13, Seattle, WA, USA, June 16-19, 2013, pages 231–242,
2013.
[Ram30] F. P. Ramsey. On a problem of formal logic. Proceedings of the London Mathematical Society,
s2-30(1):264–286, 1930.
[RSL10]
Thomas W. Reps, Mooly Sagiv, and Alexey Loginov. Finite differencing of logical formulas for
static analysis. ACM Trans. Program. Lang. Syst., 32(6), 2010.
[RTG+ 13] Andrew Reynolds, Cesare Tinelli, Amit Goel, Sava Krstic, Morgan Deters, and Clark Barrett.
Quantifier instantiation techniques for finite model finding in SMT. In Automated Deduction CADE-24 - 24th International Conference on Automated Deduction, Lake Placid, NY, USA, June
9-14, 2013. Proceedings, pages 377–391, 2013.
[RTGK13] Andrew Reynolds, Cesare Tinelli, Amit Goel, and Sava Krstic. Finite model finding in SMT.
In Computer Aided Verification - 25th International Conference, CAV 2013, Saint Petersburg,
Russia, July 13-19, 2013. Proceedings, pages 640–655, 2013.
[Sof05]
Viorica Sofronie-Stokkermans. Hierarchic reasoning in local theory extensions. In Automated
Deduction - CADE-20, 20th International Conference on Automated Deduction, Tallinn, Estonia,
July 22-27, 2005, Proceedings, pages 219–234, 2005.
[SRW02] Shmuel Sagiv, Thomas W. Reps, and Reinhard Wilhelm. Parametric shape analysis via 3-valued
logic. ACM Trans. Program. Lang. Syst., 24(3):217–298, 2002.
BOUNDED QUANTIFIER INSTANTIATION FOR CHECKING INDUCTIVE INVARIANTS
31
Appendix A. Undecidability
For a universal formula I ∈ ∀∗ (Σ), the formula I ∧ δ ∧ ¬I 0 is in EPR (recall that δ is specified
in EPR). Hence, checking inductiveness amounts to checking the unsatisfiability of an EPR
formula, and is therefore decidable. The same holds for I ∈ AF (Σ). However, this is no
longer true when quantifier alternation is introduced. For example, checking inductiveness
of I ∈ ∀∗ ∃∗ (Σ) amounts to checking unsatisfiability of a formula in a fragment for which
satisfiability is undecidable. In this section we show that checking inductiveness of ∀∗ ∃∗
formulas is indeed undecidable, even when the transition relation is restricted to EPR.
The undecidability of the problem justifies sound but incomplete algorithms for checking
inductiveness, one of which is the Bounded-Horizon algorithm (defined in Section 3) which
we study in this paper.
Finite and infinite structures. We begin by showing that the problem is undecidable
when structures, or program states, are assumed to be finite. This is the intention in most
application domains [Imm99] (including the examples in Section 7), especially when the
program does not involve numerical computations. Nevertheless, in this paper we mostly
concern ourselves with the problem of checking inductiveness when structures may also be
infinite. This is because SMT-based deductive verification relies on proof techniques from
standard first-order logic, whose semantics are defined over general structures, i.e. both
finite and infinite. We thus establish an undecidability result for this setting as well. It is
interesting to note that the discrepancy between the intended finiteness of the domain and
the proof techniques, which cannot incorporate this assumption, re-emerges in Section 6.
We refer to inductiveness over finite structures when the validity of I ∧ δ → I 0 is
considered over finite structures, and to inductiveness over general structures when it is
considered over both finite and infinite structures.
Scope of the proofs. The undecidability proofs of this section are by reductions from
tiling problems. Although technically it is also possible to prove the results by a trivial
reduction from the satisfiability of ∀∗ ∃∗ formulas (since invariants for the transition relation
true are necessarily either valid or unsatisfiable), we believe that the proofs presented here
demonstrate the intuition behind the inherent difficulty of checking inductiveness of ∀∗ ∃∗
formulas in a more profound and robust way.
To further provide intuition, we prove the undecidability of a closely related problem,
that of checking inductive invariants for safety of transition systems. Given a transition
relation δ (over Σ ] Σ0 ), a sentence ϕ0 (over Σ) describing the set of initial states and a
sentence ϕP (over Σ) describing the safety property, the problem is to check whether ϕ0 → I
(initiation), I ∧ δ → I 0 (consecution), and I → ϕP (safety) are valid (over finite or general
structures). We will consider this problem when I ∈ ∀∗ ∃∗ , ϕ0 ∈ ∀∗ , ϕP ∈ ∀∗ , in instances
where ϕ0 → I and I → ϕP are valid (so it only remains to check whether I ∧ δ → I 0 is valid).
With these restrictions, the undecidability of the problem of checking inductive invariants
for safety of transition systems over general structures implies the undecidability of the
problem of checking inductiveness as used elsewhere in this paper.
We now proceed to state the undecidability results and their proofs.
A.1. Inductiveness Over Finite Structures.
Theorem A.1. It is undecidable to check given I ∈ ∀∗ ∃∗ and δ ∈ EPR whether I is
inductive for δ over finite structures.
32
Y.M.Y. FELDMAN, O. PADON, N. IMMERMAN, M. SAGIV, AND S. SHOHAM
The proof is based on a reduction from a variant of tiling problems. We start by defining
the specific tiling problem used in the proof of this theorem:
Definition A.2 ([IRR+ 04]). A halting-tiling problem consists of a finite set of tiles T
with designated tiles Tstart , Thalt ∈ T , along with horizontal and vertical adjacency relations
H, V ⊆ T × T . A solution to a halting-tiling problem is an arrangement of instances of the
tiles in a finite rectangular grid (“board”) such that the tile Tstart appears in the top left
position, the tile Thalt appears in the end of a row (the rightmost position in some row), and
the adjacency relationships H, V are respected, meaning: if a tile t2 appears immediately to
the right of t1 it must hold that (t1 , t2 ) ∈ H, and if a tile t2 appears immediately below t1 it
must hold that (t1 , t2 ) ∈ V.
The problem is undecidable [IRR+ 04]. The proof is by a reduction from the halting
problem: given a Turing machine we can compute a halting-tiling problem such that the
problem has a solution iff the machine halts (on the empty input). In the reduction, rows
represent the tape of the Turing machine as it evolves over time (computation steps). The
tiles encode the location of the head and the current (control) state of the machine. The
horizontal and vertical constraints ensure that successive tiled rows correspond to a correct
step of the machine, and the locality of constratins is possible by the locality of computation
in a Turing machine. See [BGG96] for further details.
Proof of Theorem A.1. The proof is by a reduction from non-tilability in the halting-tiling
problem (Definition A.2) to the problem of checking inductive invariants for safety of a
transition system over finite structures where the initiation and safety requirements are
valid. We think of the transition relation δ as incrementally placing tiles in a rectangle.
Vocabulary. To express locations on the board, we use a total order for both the horizontal
and vertical dimensions of the board. We add an immediate predecessor relation j = i − 1
which is true if j < i and there is no element of the order between j, i. We use a constant
0 for the minimal element of the order. These notions can be defined using a universally
quantified formula.8 A location is a pair of elements of the order, a vertical and horizontal
component. We sometimes use the term board order to refer to the lexicographic order of
pairs of elements of the order.
The transition system keeps track of the last tile placed on the board by a relation
M (i, j) which is true only for the last updated location. Since the placing of tiles occurs
in a sequential manner we also call this board location maximal, and a location active if it
comes before the maximal location in the board order. The active area is the set of active
locations.
The state of tiles on the board is represented by a set of relations {Tk }, one for each tile
type, encoding the locations on the board where a tile of type Tk is placed.
In this proof we also use a constant max to be an element of the total order, representing
the width of the rectangle.
8 The assumption that there exists a predecessor is left for the invariant to state explicitly when necessary,
as this is the heart of the ∀∗ ∃∗ quantification in the proof.
BOUNDED QUANTIFIER INSTANTIATION FOR CHECKING INDUCTIVE INVARIANTS
33
Transitions. In every step the transition system places a valid tile in the next board location.
The next board location is considered while respecting the width of the rectangle, moving to
the next row if the horizontal component of the current tile is max.
Placing a tile of type Tnext on the board is done by an EPR update of the (two-vocabulary)
form
∀i, j. M 0 (i, j) ↔ ((j = 0 ∧ M (i − 1, max) ∨ (j 6= 0 ∧ M (i, j − 1)))
0
∀i, j. Tnext
(i, j) ↔ Tnext (i, j) ∨ M 0 (i, j)
(A.1)
∀i, j. Tk0 (i, j) ↔ Tk (i, j)
∀Tk 6= Tnext .
The transition system nondeterministically chooses a tile Tnext type that respects the adjacency relationships. These relationships are with the tiles in the board location preceding the
current location in the horizontal and vertical components, expressible using the immediate
predecessor relation and existential quantification on these predecessors. (Note that the
existential quantifiers do not need to reside in the scope of universal quantifiers — they
depend only on the current location.) Because the set of tile types T is finite, expressing the
allowed tile types given the two adjacent locations can be done by a quantifier-free formula.
Overall the EPR formula describing a step of the system consists of a disjunction between
choices for Tnext . Each of these possible choices is described via a conjunction of the guard
that makes sure that it is legal to place Tnext , and a corresponding update to the relation
that is a conjunction of the formulas in Equation (A.1).
Initial state. Initially we only have Tstart placed in the upper-left corner, so ∀i, j. Tstart (i, j) ↔
(i = 0 ∧ j = 0) and ∀i, j. ¬Tk (i, j) for every other tile type Tk .
Safety property. The safety property states that the special tile Thalt , is not placed on the
board in the end of a row (in a max position) in the active area.
Invariant. The invariant states that in the active area we have a valid partial tiling. We
require this by a ∀∗ ∃∗ formula saying that for every tile placed in an active location (except
for the maximal location) there is a successor tile, placed in the next board location, that
conforms to the (local) ajdacency relations.9 We also conjoin the safety property to the
invariant.
Reduction argument. The invariant holds for the initial state, and trivially implies the
safety property.
If there exists a valid tiling with Thalt in the end of a row, a counterexample to induction
can be obtained by encoding this valid tiling in the post-state and that same tiling without
Thalt , which is the last-placed tile, in the pre-state.
For the converse, assume that the invariant is not inductive over finite structures, i.e.,
there exists a finite counterexample to induction, and show that there exists a solution to
the halting-tiling problem. The reasoning is as follows: A finite state satisfying the invariant
induces a valid finite partial tiling (defined by the active area of the board in the structure).
Since the transition system always places a tile that respects the adjacency relations, it is
easy to see that a counterexample to induction must be such that the transition places Thalt
on the board in the end of a row, and that this also induces a valid partial finite tiling in
9 We specify the requirement in this forward fashion, rather than requiring that every tile has a valid
predecessor, in order to easily reuse this invariant in the proof of Theorem A.3.
34
Y.M.Y. FELDMAN, O. PADON, N. IMMERMAN, M. SAGIV, AND S. SHOHAM
the post-state. Thus a finite counterexample to induction implies the existence of a valid
finite tiling with Thalt in the end of a row, which is a solution to the halting-tiling problem.
Thus the invariant is inductive iff the halting-tiling problem does not have a solution.
A.2. Inductiveness Over General Structures.
Theorem A.3. It is undecidable to check given I ∈ ∀∗ ∃∗ and δ ∈ EPR whether I is
inductive for δ over general (finite and infinite) structures.
The proof is based on a reduction from a variant of tiling problems. We start by defining
the specific tiling problem used in the proof of this theorem:
Definition A.4. A lower-triangular infinite-tiling problem consists of a finite set of tiles
T , along with horizontal and vertical adjacency relations H, V ⊆ T × T . A solution to a
lower-triangular infinite-tiling problem is an arrangement of instances of the tiles in the
lower-triangular plane (i.e. a total function {(i, j) ∈ N×N | i ≤ j} → T ) where the adjacency
relationships H, V are respected, meaning: if a tile t2 appears immediately to the right of t1
it must hold that (t1 , t2 ) ∈ H, and if a tile t2 appears immediately below t1 it must hold that
(t1 , t2 ) ∈ V.
The problem is undecidable The proof is by a reduction from the non-halting problem:
given a Turing machine we can compute a lower-triangular infinite-tiling problem such that
the problem has a solution iff the machine does not halt (on the empty input). The encoding
is similar to [IRR+ 04].
Proof of Theorem A.3. The proof is by a reduction from non-tilability in the lower-triangular
infinite-tiling problem (Definition A.4) to the problem of checking inductive invariants for
safety of a transition system (over general structures) where the initiation and safety
requirements are valid.
We construct a transition relation similar to the one in the proof of Theorem A.1, with
some changes, as described below.
Discussion and motivation. To provide some intuition to the difference between the
reductions, we remark that both of the proofs in this section are in essence a reduction
from the halting (or non-halting) problem. The proof of Theorem A.1 encodes runs of the
machine as finite tilings, and asks whether a tiling that represents a terminating computation,
encoded by Thalt , is possible. This reduction is no longer adequate when structures may be
infinite. The reason is that an infinite valid partial tiling may not correspond to reachable
configuration of the Turing machine, so there may be such an infinite tiling with Thalt even
though the Turing machine never halts.10
In fact, the reduction in this proof must be in the opposite direction: the invariant
should be inductive iff the machine does not terminate, whereas in the proof of Theorem A.1
the invariant is inductive iff the machine does terminate. This is because satisfiability is
recursively-enumerable over finite structures and co-recursively-enumerable over general
structures (due to the existence of proofs), which reflects on checking inductiveness through
the satisfiability check of the formula I ∧ δ ∧ ¬I 0 .
10 One way to construct such a tiling, using a tile in row ω of the board, is utilized in the proof that
follows.
BOUNDED QUANTIFIER INSTANTIATION FOR CHECKING INDUCTIVE INVARIANTS
35
Thus, we would like to have a counterexample to induction when the machine never
halts, i.e. has an infinite run. As before, runs of the machine are encoded via tiling, only
that now an infinite structure can encode an infinite run of the machine. (It is not necessary
that an infinite tiling represents a valid infinite run of the machine, but every infinite run
can be represented by such a structure.) We would like to “detect” this situation. Our
way to do this is by the observation that induction on the number of rows, or execution
steps, must hold when the number of rows is finite (but unbounded), as in Theorem A.1,
but does not necessarily hold when there may be an infinite number of rows. This idea is
implemented by a relation P with the invariant that it is preserved under successive board
locations. In an infinite structure this does not imply that P is true for all locations. A flag
f is used to express a transition that is aware of P not being globally true.
Another technical detail is the lower-triangular formulation of the tiling problem, which
is used to construct the infinite computation of the transition system by placing a single tile
in each step.
Returning to the proof, we describe the reduction and highlight its differences from the
reduction in Theorem A.1. Following the lower-triangular formulation of the tiling proble,
we restrict the board order to the lower-triangular part (locations (i, j) such that i ≤ j) and
ignore other locations when considering successor in the board order.
Vocabulary. We add a relation P over board locations, and a Boolean flag (nullary predicate)
f.
Transitions. In each step the transition system places a valid tile in the next board location,
similar to the proof of Theorem A.1. The difference is that the criterion for moving to place
tiles in the next tile is when the current location (i, j) has i = j (whereas in Theorem A.1
the criterion was j = max).
To maintain the invariant that P is preserved under successor of active tiles in the board,
when we place a new tile, if P holds for the maximal location before the step, set P to true
for the new maximal location.
If P does not hold for the new maximal location, turn f to false.
Initial state. P is true for the first location (0, 0) only, and f is true. In this proof, initially
the board is empty.
Safety property. The safety property now asserts that f is true.
Invariant. As before, the invariant states that the active board represents a valid partial
tiling, i.e. every active tile except for the maximal one has a valid successor.
The invariant also states that P is preserved under successor of board, i.e.: if (i1 , j1 )
and (i2 , j2 ) are successive active board locations w.r.t. the board order, then if P holds for
(i1 , j1 ) is must also hold for (i2 , j2 ). We also conjoin the safety property to the invariant.
Reduction argument. The invariant holds for the initial state, and trivially implies the
safety property.
Assume that there is no solution to the lower-triangular infinite-tiling problem, and
show that the invariant is inductive. The reasoning is as follows: A state satisfying the
invariant induces a partial valid tiling — either finite or infinite — over the active area of
the board. Since there is no valid partial tiling with an infinite number of rows, the number
of active locations must be finite (the number of columns in the active domain is bounded by
36
Y.M.Y. FELDMAN, O. PADON, N. IMMERMAN, M. SAGIV, AND S. SHOHAM
the number of rows, since we are discussing lower-triangular tilings). Because P is preserved
under successor of the board order, by induction on the number of locations, P must hold
for the maixmal location. After a transition is taken, f remains true. Since the transition
system always places a tile that respects the horizontal and vertical ajdacency relations and
sets P to true for the new maximal location, it is easy to see that the rest of the invariant is
preserved by a transition as well.
For the converse direction, if there is a solution to the lower-triangular infinite-tiling
problem, then there is an infinite structure encoding this tiling. The transition begins with
the infinite valid tiling, with a new additional row after this infinite sequence of tiled rows.
(Recall that the board dimensions are axiomatized using a total order; the additional row
index corresponds to ordinal ω of vertical order.) We place some tile in the first column of
this row as in some valid row in the tiling. Note that when placing tiles in this row we need
not worry about vertical constraints, because they were expressed in a forward fashion, and
this row is not a successor of any other row. The first leftmost location in the new row is
set to be the maximal active one, and we set P to be false for this location. Note that this
does not violate the invariant: P is preserved under successor of the board location, but
nonetheless does not hold for the location in the additional row (it is not the successor of
any location). The transition will now place a new tile and turn f to false, P does not hold
for the current maximal location, thereby violating the invariant.
Thus the invariant is inductive iff the infinite tiling problem does not have a solution.
This work is licensed under the Creative Commons Attribution-NoDerivs License. To view
a copy of this license, visit http://creativecommons.org/licenses/by-nd/2.0/ or send a
letter to Creative Commons, 171 Second St, Suite 300, San Francisco, CA 94105, USA, or
Eisenacher Strasse 2, 10777 Berlin, Germany
| 6cs.PL
|
Where is my forearm? Clustering of body parts from simultaneous tactile and
linguistic input using sequential mapping
Karla Štěpánová1,2 , Matěj Hoffmann1,3 , Zdeněk Straka1 , Frederico B. Klein4 , Angelo Cangelosi4 , Michal Vavrečka2
(1)
Center for Machine Perception, Department of Cybernetics,
Faculty of Electrical Engineering, Czech Technical University in Prague
(2)
Czech Institute of Informatics, Robotics, and Cybernetics, CTU in Prague
(3)
iCub Facility, Istituto Italiano di Tecnologia
(4)
School of Computing, Electronics and Mathematics, Plymouth University, Plymouth, UK
arXiv:1706.02490v1 [cs.NE] 8 Jun 2017
Abstract
is an umbrella term uniting higher level representations,
for perception more than for action, and accessible to
consciousness. Schwoebel and Coslett (2005) amassed
evidence for distinguishing between three types of body
representations: body schema, body structural description, and body semantics—constituting a kind of hierarchy. The body structural description is a topological
map of locations derived primarily from visual input
that defines body part boundaries and proximity relationships. Finally, body semantics is a lexical–semantic
representation of the body including body part names,
functions, and relations with artifacts (e.g., shoes are
used on the feet, and feet can be used to kick a football).
While the details of every particular taxonomy
or hierarchy can be discussed, clearly, there is a trend
from continuous, modality-specific representations (like
the tactile homunculus) to multimodal, more aggregated representations. This may be first instantiated by
increasing receptive field size and combining sensory
modalities, as it is apparent in somatosensory processing, e.g. areas relatively specialized on proprioception
or touch and with small receptive fields (like Brodmann
areas 3a and 3b), touch and proprioception are getting
increasingly combined in areas 1 and 2. Then, going
from anterior to posterior parietal cortex, the receptive
fields grow further and somatosensory information is
combined with visual. One can then ask whether this
process of bottom-up integration or aggregation may
give rise to discrete entities, or categories, similar to individual body parts. Vignemont et al. (2009) focused
on how body segmentation between hand and arm could
appear based on a combined tactile and visual perception. They explored category boundary effect which appeared when two tactile stimuli were presented: these
stimuli felt farther away when they were applied across
the wrist than when they were applied within a single
body part (palm or forearm). In conclusion, they suggest that the representation of the body is structured in
categorical body parts delineated by joints, and that this
categorical representation modulates tactile spatial perception.
Next to the essentially bottom-up clustering of
Humans and animals are constantly exposed to a continuous stream of sensory information from different
modalities. At the same time, they form more compressed representations like concepts or symbols. In
species that use language, this process is further structured by this interaction, where a mapping between the
sensorimotor concepts and linguistic elements needs to
be established. There is evidence that children might
be learning language by simply disambiguating potential meanings based on multiple exposures to utterances
in different contexts (cross-situational learning). In existing models, the mapping between modalities is usually found in a single step by directly using frequencies
of referent and meaning co-occurrences. In this paper,
we present an extension of this one-step mapping and
introduce a newly proposed sequential mapping algorithm together with a publicly available Matlab implementation. For demonstration, we have chosen a less
typical scenario: instead of learning to associate objects with their names, we focus on body representations. A humanoid robot is receiving tactile stimulations on its body, while at the same time listening to
utterances of the body part names (e.g., hand, forearm
and torso). With the goal at arriving at the correct “body
categories”, we demonstrate how a sequential mapping
algorithm outperforms one-step mapping. In addition,
the effect of data set size and noise in the linguistic input are studied.
1
Introduction
Body representation has been the topic of psychological, neuroanatomical and neurophysiological studies for
many decades. Spurred by the account of Head and
Holmes (1911) and their proposal of superficial and postural schema, a number of different concepts has been
proposed since: body schema, body image, and corporeal schema being only some of them. Body schema
is usually thought of as more “low-level”, sensorimotor
representation of the body used for action. Body image
1
multimodal body-related information, an additional
“categorization” of body parts is imposed through language, such as when the infant hears her parents naming the body parts. Interestingly, recent research (Majid, 2010) showed that there are some cross-linguistic
variabilities in naming body parts and this may in turn
override or influence the “bottom-up” multimodal (nonlinguistic) body part categorization.
While the field is relatively rich in experimental observations, the mechanisms behind the development and operation of these representations are still not
well understood. Here, computational and in particular
robotic modeling ties in—see (Hoffmann et al., 2010;
Schillaci et al., 2016) for surveys on body schema in
robots. Petit and Demiris (2016) developed an algorithm for the iCub humanoid robot to associate labels
for body parts and later proto-actions with their embodied counterparts. These could then be recombined
in a hierarchical fashion (e.g., “close hand” consists
of folding individual fingers). Mimura et al. (2017)
used Dirichlet process Gaussian mixture model with latent joint to provide a Bayesian body schema estimation based on tactile information. Their results suggest that kinematic structure could be estimated directly from tactile information provided by a moving
fetus without any additional visual information—albeit
with a lower accuracy. Our own work on the iCub
humanoid robot has thus far focused on learning primary representations—tactile (Hoffmann et al., 2017)
and proprioceptive (Hoffmann and Bednarova, 2016).
In this work, we use the former (the “tactile homunculus”) as input for further processing—interaction with
linguistic input.
In this work, we strive to find segmentation of
body parts based on a simultaneous tactile and linguistic information. However, body part categorization and
mapping to body part names is one instance of a more
general problem of segmenting objects from the environment, learning compressed representations (loosely
speaking: concepts, categories, symbols) to stand in
for them and associating them with words to which the
infant is often exposed simultaneously. Borghi et al.
(2004), for example, studied the interaction of object
names with situated action on the same objects.
We made use of a newly proposed sequential mapping algorithm which extends an idea of one-step mapping (Smith et al., 2006) and compared its overall accuracy to one-step mapping as well as to accuracies of segmenting individual body parts. We further explore how
the accuracy of the learned mapping is influenced by a
level of noise in the linguistic domain and data set size.
The sequential mapping strategy was shown to be very
robust as it can find the mapping under circumstances of
very noisy input and clearly outperformed the one-step
mapping.
Complete source code used for generating results
in this article is publicly available at https://github.
com/stepakar/sequential-mapping.
This article is structured as follows. The inputs
and their preprocessing and the mapping algorithms are
described in Section 2. This is followed by Results
(Section 3) and a Discussion and Conclusion.
2
Materials and Methods
In this section, we will first present the inputs and their
preprocessing pipelines: tactile input (Section 2.1) and
linguistic input (Section 2.2). In total, 9 body parts of
the right half of the robot’s upper body were stimulated:
torso/chest, upper arm, forearm, palm and 5 fingertips.
Tactile stimulation coincided with an utterance of the
body part’s name. Then, the one-step and sequential
mapping algorithms (sections 2.3.1 and 2.4) are presented, and a description of the evaluation (Section 2.5).
2.1
Tactile inputs and processing
To generate tactile stimulation pertaining to different
body parts, we built on our previous work on the iCub
humanoid robot. In particular, the “tactile homunculus” (Hoffmann et al., 2017)—a primary representation
of the artificial sensitive skin the robot is covered with
(see Fig. 1 – one half of the robot’s upper body). In
the current work, the skin was not physically stimulated
anymore, but the activations were emulated and then relayed to the “homunculus”, as detailed below.
2.1.1
Emulated tactile input
We created a YARP (Metta et al., 2006) software module to generate virtual skin contacts1 . A skin part was
randomly selected and then stimulated. The number of
pressure-sensitive elements (henceforth taxels) for different skin parts was 440 for the torso, 380 for upper
arm, 230 for forearm, and 104 for the hand (44 for
palm and 5 × 12 for fingertips)—1154 taxels in total.
Once the skin part was randomly selected, a small region was also randomly picked within that part for the
tactile stimulation—10 taxels at a time, corresponding
to the triangular modules the skin is composed of. For
the hand, the situation was slightly different: the entire hand was treated as one skin part. Then, within the
hand, a random choice was made between 5 subregions
on the palm skin (8 to 10 taxels) and 5 fingertips (12
taxels each). Data was collected for 100 minutes, corresponding to approximately 2000 individual 3 second
stimulations. For all skin parts, the stimulation lasted
for 3 seconds and was sampled at 10 Hz. A label–body
part name–was saved along with the tactile data. These
labels are used to generate the linguistic input and for
performance evaluation later, but do not directly take
1 https://github.com/robotology/peripersonal-space/
tree/master/modules/virtualContactGeneration
part in the clustering of tactile information. Please note
that there were separate labels for the palm and individual fingers, while these were all treated as one “skin
part” in the virtual touch generation and hence the number of samples per finger, for example, was lower than
for other non-hand body parts.
2.1.2
First layer – “tactile homunculus”
The input layer of the “tactile homunculus” (Hoffmann
et al., 2017) consists of a vector, a(t), of activations
of 1154 taxels at time t—the output of the previous
section—that have binary values (1 when a taxel is stimulated, 0 otherwise). The output layer then forms a
7 × 24 (168 “neurons” in total) grid – see Figure 1 B.
This layer is a compressed representation of the skin
surface—the receptive fields of neurons (the parts of
skin they respond to) are schematically color-coded.
However, this code (and “clustering”) is not available
as part of the tactile input.
The output layer will be represented as a single
vector x(t) = [x1 (t), ..., x168 (t)]. The activations of the
output neurons, xi (t), are calculated as dot products of
the weight vector ui corresponding to the i-th output
neuron and the tactile activation vector a(t) as follows:
xi (t) = ui · a(t)
2.1.3
(1)
Second layer – GMM
The output of the first layer, vector x(t) (168 elements,
continuous-valued) serves as input to the second tactile
processing layer. This layer aims to cluster individual
body parts and represent them as abstract models. Resulting models T j are subsequently mapped in the multimodal layer to clusters found in the language layer.
To process the outputs from the first layer, we used
a Gaussian mixture model (GMM), which is a convex
mixture of D-dimensional Gaussian densities l(x|θ j ). In
this case, each tactile model T j is described by a set of
parameters θ j . The posterior probabilities p(θ j |x) are
computed as follows:
p(θ j |x) =
J
X
rkj l(x|θ j ),
(2)
j=1
1
1
exp[− (x−m j )T (S j )−1 (x−m j )],
l(x|θ j ) = p
p
D
2
(2π) |S j |
(3)
where x is a set of D-dimensional continuous-valued
data vectors, rkj are the mixture weights, J is the number
of tactile models, parameters θ j are cluster centers m j
and covariance matrices S j .
Mixture of Gaussians is trained by the EM algorithm (Dempster et al., 1977). Number of tactile
models J is in this model preset based on the number of different linguistic labels. In future, we plan to
use an adaptive extension of GMM algorithm such as
gmGMM (Štepánová and Vavrečka, 2016) to detect this
number autonomously.
An output of this layer for each data point x(t)
is the vector y(t) of J output parameters describing the
data point (the likelihood that the data point belongs to
each individual cluster in a mixture). This corresponds
to the fuzzy memberships (distributed representation).
2.2
Linguistic inputs and processing
Tactile stimulation of a body part was accompanied with
the corresponding utterance. In our case, where we have
9 separate body parts, these are ’torso’, ’upper arm’,
’forearm’, ’palm’, ’little finger’, ’ring finger’, ’middle
finger’, ’index finger’ and ’thumb’. Linguistic and tactile inputs are processed simultaneously.
We conducted experiments with spoken language
input—one-word utterances pronounced by a nonnative English speaker. To process this data, we made
use of CMU Sphinx (an open-source flexible Markov
model-based speech recognizer system) (Lamere et al.,
2003) and achieved 100% accuracy of word recognition. The word-forms are extracted from the audio input
and compared to prelearned language models by means
of the log-scale scores p(wnt |Li ) of the audio matching.
Based on these data, posterior probability can be computed.
However, in the current work, we employed a
shortcut and used the labels (ground truth) directly. This
allowed us to fully explore the effect of misclassification
in linguistic subdomain to mapping accuracy. The noise
to the language data was added subsequently and evenly
to all classes (a given proportion of labels was randomly
permuted).
2.3
Cross-situational learning
One possible way how to establish mapping between
sensorimotor concepts and linguistic elements is to use
frequencies of referent and meaning co-occurrences,
that is, the ones with the highest co-occurrence are
mapped together (Smith et al., 2006; Xu and Tenenbaum, 2007). This method is usually called crosssituational learning and supposes the availability of the
ideal associative learner who can keep track and store all
co-occurrences in all trials, internally memorizing and
representing the word–object co-occurrence matrix of
input. This allows the learner to subsequently choose
the most strongly associated referent (Yu and Smith,
2012).
2.3.1
One-step mapping
The simplest one-step word-to-referent learning algorithm only accumulates word-referent pairs. This can be
viewed as Hebbian learning: the connection between a
Fig. 1: iCub skin and tactile homunculus. (A) Photograph of the iCub robot with artificial skin exposed on the right
half of the upper body (1154 taxels in total). (B) Representation of tactile inputs learned using a Self-Organizing Map
– a 24 × 7 neuronal sheet. (C) Schematics with skin patches unfolded and colored to mark the correspondence with B.
Arrows illustrate the relationship in orientation between skin parts and the learned map (Hoffmann et al., 2017).
word and an object is strengthened if the pair co-occurs
in a trial. To extend this basic idea, we can enable also
forgetting by introducing a parameter η, which can capture the memory decay (Yu and Smith, 2012). Supposing that at each trial t we observe an object ont and hear
a corresponding word wnt (Nt possible associations), we
can describe the update of the strength of the association between word model L(i) and object—in our case
tactile model T ( j)—as follows:
A(i, j) =
R
X
t=1
η(t)
Nt
X
δ(wnt , i)δ(ont , j),
(4)
n=1
where R is the number of trials, δ is the Kronecker delta
function (equal to 1 when both arguments are identical
and 0 otherwise), wnt and ont indicate the nth word–object
association that the model attends to and attempts to
learn in the trial t and η(t) is the parameter controlling
the gain of the strength of association.
Now let’s assume that the word w(i) is modeled by
the model Li in the language domain and object (referent) o( j) is modeled by the model T m(i) in the tactile domain. Our goal is to find the corresponding model T m(i)
from tactile subdomain for each model Li from language
domain to assign them together. Indices m(i) are found
as follows:
∀i : m(i) = argmax A(i, j),
(5)
i
where A is the co-occurrence matrix computed in the
Eq. 4 (element A(i, j) captures co-occurrence between
the word w(i) and object o( j)).
2.4
Sequential mapping algorithm
To capture dynamic competition among models, we extend the basic one-step mapping algorithm for crosssituational learning by sequential addition of inhibitory
connections. The inhibitory mechanisms and situationtime dynamics were already partially included into the
model of cross-situational learning proposed by McMurray et al. (2012). Even though our model shares
some similarities with the model proposed by McMurray, it stems from different computational mechanisms.
After a reliable assignment between a language and tactile model is found, inhibitory connections among this
tactile model and all other language models are added.
Thanks to this mechanism, mutual exclusivity principle
(the fact that children prefer mapping where object has
only one label to multiple labels (Markman, 1990)) is
guaranteed.
The assignment between tactile models T j and
language models L j is found using the following iterative procedure:
1. Tactile and language data are clustered separately
and the corresponding posterior probabilities are
found.
2. For each data point the most probable tactile and
language clusters are selected and the data point is
assigned to these clusters.
3. Co-occurrence matrix with elements A(i, j) is computed and the best assignment is selected:
[im, m(im)] = argmax argmax A(i, j).
i
j
(6)
In this step, the tactile model T m(im) is assigned to
the language model Lim .
4. Inhibitory connections are added between the assigned tactile model T m(im) and all language models
Li , where i , im (mutual exclusivity).
5. Assigned data points (data points which belong to
both T m(im) and Lim ) are deleted from the data set.
6. If data set is not empty or not all tactile clusters are
assigned to some language cluster go to (1), else
stop.
2.5
Evaluation
Accuracy of the learned mapping is calculated in the
following manner: We cluster output activations from
the tactile homunculus and assign each data point to the
most probable cluster. Then, we find indices m(i) for all
clusters as defined in equation 5 for one-step mapping
and equation 6 for sequential mapping. Based on this
mapping we can assign each data point to the language
label. These language labels are subsequently compared
to the ground truth (the body part name is equivalent
to the language label prior to the application of noise).
Accuracy is then computed as:
acc = T P/N
(7)
where T P (true positive) is the number of correctly assigned data points and N is the number of all data points.
3
Results
We studied the performance of one-step vs. the sequential mapping algorithms on the ability to cluster individual body parts from simultaneous tactile and linguistic
input. That is, all the skin regions on the same body part
should “learn” that they belong together (to the forearm,
say), thanks to the co-occurrences with the body part labels. In addition, the effect of data set size and levels of
noise in the linguistic domain are investigated (Section
3.1). A detailed analysis of the mapping accuracy for
individual body parts and a backward projection onto
the tactile homunculus are shown in sections 3.2 and
3.3 respectively.
3.1
Comparison of accuracy of one-step mapping
to sequential mapping
The performance of the one-step and sequential mapping algorithms is shown in Fig. 2. The comparison is
provided for different data set sizes (namely for 6 different data sets with number of data points from 64 to
63806) and noise levels. As can be seen, the accuracy
of sequential mapping remains very stable and outperforms one-step mapping for all values of the noise (in
the linguistic domain) and all data set sizes. For smaller
data sets, we can see a steeper drop in accuracy with
increasing noise in the language data.
3.2
Accuracy of mapping for individual body parts
The accuracy calculated in the previous section and Fig.
2 is an overall accuracy and we don’t take into account
the number of data points per individual body part. To
explore the performance in more detail, we focused also
on the accuracy of sequential mapping for individual
body parts. The results for the data set with 3190 and
638 data points can be seen in Fig. 3 top and bottom
panel, respectively. The accuracy for all body parts decreases with increasing noise in the linguistic input. The
accuracy for fingers is significantly lower—this is due
to the lower number of samples per finger (see Section
2.1.1). Comparing the top and bottom panel in Fig. 3
demonstrates poorer performance with higher variance,
especially for the fingers.
3.3
Projecting results of sequential mapping back
onto homunculus
After tactile data from homunculus are clustered and
these clusters are mapped to appropriate language clusters (representing body parts utterances), we can project
these labels back onto the original tactile homunculus.
Considering that xi (t) are activations of neuron i in the
homunculus, D is the whole data set consisting of vector of homunculus activations for each data point, and
LangLabel(d) is the language label assigned to a data
point d based on the sequential mapping procedure described in the Section 2.4, we can project results of sequential mapping onto the homunculus in a following
manner. First, we compute strength of activation nki of
each neuron i for a given language label k as follows:
nki =
X
xi (t), i ∈ {1, . . . , 168},
(8)
x(t)∈Dk
where Dk = {d ∈ D|LangLabel(d) = k} and k = {torso,
upper arm, forearm, palm, little finger, ring finger, middle finger, index finger, thumb}.
Afterwards, we visualize for each neuron how
much it is activated for individual body parts. Results
for data sets of differing size and level of noise in the linguistic domain can be seen in Fig. 4. Clearly, for large
enough data sets and limited noise, the mapping from
language to the tactile modality is successful in delineating the body part categories (the fingers with fewer
data points being more challenging)—as can be seen by
comparing panels A and B.
Fig. 2: Accuracy of one-step vs. sequential mapping for different levels of noise in language. Number denotes the size
of data set, SM - sequential mapping and OM - one-step mapping. The mean and standard deviation from 20 repetitions
is visualized.
Fig. 3: Accuracy of sequential mapping for individual body parts: Visualization of sequential mapping accuracy based
on the noise in linguistic data for 2 data set sizes: 3190 data points (upper) and 638 data points (lower) , noise in language
data 0-100%(random). The mean and standard deviation from 40 repetitions are visualized.
4
Discussion and Conclusion
To study the problem of associating (mapping) between
sensorimotor or multimodal information, concepts or
categories, and language or symbols, we have chosen
a specific but less studied instance of this problem: segmentation and labeling of body parts. Perhaps, from a
developmental perspective, this could be plausible, as
Fig. 4: Projection of mapping results back onto the tactile homunculus – sample runs of the algorithm. Color
code for individual body parts is the same as in Fig. 1.
(A) Original homunculus with true labels. (B) Results
from data set with 6381 data points and 10% noise. (C)
Results from 638 data points and 80% noise.
the body may be the first “object” the infant is discovering. The self-exploration occurs in the sensorimotor domain, but at the same time or slightly later, the infant is
exposed to utterances of body part names. In this work,
we study the mapping between the tactile modality and
body part labels from linguistic input.
We present a new algorithm for mapping language
to sensory modalities (sequential mapping), compare it
to one-step mapping and test it on the body part categorization scenario. Our results suggest that this mapping procedure is robust, resistant against noise, and sequential mapping shows better performance than onestep mapping for all data set sizes and also slower performance degradation with increasing noise in the linguistic input. Furthermore, we explored accuracy of the
sequential mapping for individual body parts, revealing that body parts less represented in the data set—
fingers—were categorized less accurately. This problem might be mitigated with increased overall data set
size; yet, dealing with clusters with uneven data point
number is a common problem of clustering algorithms
(in our case GMM).
Projecting the labels or categories induced by language back onto the tactile homunculus showed that the
body part categories are quite accurate. Given the nature of the tactile input—the skin is a continuous receptor surface—and the random-uniform tactile input generator used, the linguistic input was the only one that
can facilitate cluster formation. However, more realistic, non-uniform touch and, in particular, the addition
of additional modalities (proprioception, vision) should
enable bottom-up non-linguistic body part category formation, as described by (Vignemont et al., 2009), for
example. These constitute possible directions of our
future work: the “modal” cluster formation will interact with the labels imposed by language. Furthermore,
thus far, only one half of the body was considered—
corresponding to the lateralized representations in the
tactile homunculus—, but one can imagine stimulating
both left and right arm, for example, while hearing always the same utterance: ‘upper arm’. Further study of
the brain areas involved in this processing is needed, in
order to develop models more closely inspired by the
functional cortical networks, like in (Caligiore et al.,
2010) that model the experimental findings of (Borghi
et al., 2004).
For our experiments we used artificially generated linguistic input (i.e., body part labels) with added
noise (i.e. wrong labels with a certain probability). In
the future, we are planning to use actual auditory input
(spoken words) with real noise. This will also add the
additional dimension of similarity in the auditory domain: ‘arm’ and ‘forearm’ are phonetically closer to
each other than to, say, ‘torso’. Thus, the linguistic
modality will not constitute crisp, discrete labels anymore, but these will have to be extracted first—opening
up further possibilities for bidirectional interaction with
other modalities.
5
Acknowledgement
K.S. and M.H. were supported by the Czech Science
Foundation under Project GA17-15697Y. M.H. was additionally supported by a Marie Curie Intra European
Fellowship (iCub Body Schema 625727) within the 7th
European Community Framework Programme. Z.S.
was supported by The Grant Agency of the CTU Prague
project SGS16/161/OHK3/2T/13. M.V. was supported
by European research project TRADR funded by the EU
FP7 Programme, ICT: Cognitive systems, interaction,
robotics (Project Nr. 609763).
References
Borghi, A. M., Glenberg, A. M. and Kaschak, M. P.
(2004). Putting words in perspective. Memory &
Cognition, 32(6):863–873.
Caligiore, D., Borghi, A. M., Parisi, D. and Baldassarre,
G. (2010). Tropicals: A computational embodied
neuroscience model of compatibility effects. Psychological Review, 117(4):1188.
Dempster, A. P., Laird, N. M. and Rubin, D. B. (1977).
Maximum likelihood from incomplete data via the
EM algorithm. Journal of the royal statistical society. Series B (methodological), pp. 1–38.
Head, H. and Holmes, H. G. (1911). Sensory disturbances from cerebral lesions. Brain, 34:102–254.
Hoffmann, M. and Bednarova, N. (2016). The encoding of proprioceptive inputs in the brain: knowns and
unknowns from a robotic perspective. Vavrecka, M.,
Becev, O., Hoffmann, M. and Stepanova, K. (eds.), In
Kognice a umělý život XVI [Cognition and Artificial
Life XVI], pp. 55–66.
Hoffmann, M., Marques, H., Hernandez Arieta, A.,
Sumioka, H., Lungarella, M. and Pfeifer, R.
(2010). Body schema in robotics: A review. Autonomous Mental Development, IEEE Transactions
on, 2(4):304–324.
Hoffmann, M., Straka, Z., Farkas, I., Vavrecka, M. and
Metta, G. (2017). Robotic homunculus: Learning
of artificial skin representation in a humanoid robot
motivated by primary somatosensory cortex. IEEE
Transactions on Cognitive and Developmental Systems.
Lamere, P., Kwok, P., Gouvea, E., Raj, B., Singh, R.,
Walker, W., Warmuth, M. and Wolf, P. (2003). The
cmu sphinx-4 speech recognition system. In IEEE
Intl. Conf. on Acoustics, Speech and Signal Processing (ICASSP 2003), Hong Kong, vol. 1, pp. 2–5. Citeseer.
Majid, A. (2010). Words for parts of the body. Words
and the mind: How words capture human experience,
pp. 58–71.
Markman, E. M. (1990). Constraints children place on
word meanings. Cognitive Science, 14(1):57–77.
McMurray, B., Horst, J. S. and Samuelson, L. K. (2012).
Word learning emerges from the interaction of online
referent selection and slow associative learning. Psychological review, 119(4):831.
Metta, G., Fitzpatrick, P. and Natale, L. (2006). Yarp:
yet another robot platform. International Journal on
Advanced Robotics Systems, 3(1):43–38.
Mimura, T., Hagiwara, Y., Taniguchi, T. and Inamura,
T. (2017). Bayesian body schema estimation using
tactile information obtained through coordinated random movements. Advanced Robotics, 31(3):118–
134.
Petit, M. and Demiris, Y. (2016). Hierarchical action
learning by instruction through interactive grounding
of body parts and proto-actions. In Robotics and Automation (ICRA), 2016 IEEE International Conference on, pp. 3375–3382. IEEE.
Schillaci, G., Hafner, V. V. and Lara, B. (2016). Exploration behaviors, body representations, and simulation processes for the development of cognition in
artificial agents. Frontiers in Robotics and AI, 3:39.
Schwoebel, J. and Coslett, H. B. (2005). Evidence for
multiple, distinct representations of the human body.
Journal of cognitive neuroscience, 17(4):543–553.
Smith, K., Smith, A. D., Blythe, R. A. and Vogt, P.
(2006). Cross-situational learning: a mathematical approach. Lecture Notes in Computer Science,
4211:31–44.
Štepánová, K. and Vavrečka, M. (2016). Estimating
number of components in gaussian mixture model using combination of greedy and merging algorithm.
Pattern Analysis and Applications, pp. 1–12.
Vignemont, d. F., Majid, A., Jola, C. and Haggard, P.
(2009). Segmenting the body into parts: evidence
from biases in tactile perception. The Quarterly Journal of Experimental Psychology, 62(3):500–512.
Xu, F. and Tenenbaum, J. B. (2007). Word learning as bayesian inference. Psychological review,
114(2):245.
Yu, C. and Smith, L. B. (2012). Modeling crosssituational word–referent learning: Prior questions.
Psychological review, 119(1):21.
| 2cs.AI
|
arXiv:1708.01396v1 [math.AC] 4 Aug 2017
GRADED COMPONENTS OF LOCAL COHOMOLOGY
MODULES II
TONY J. PUTHENPURAKAL AND SUDESHNA ROY
Abstract. Let A be a commutative Noetherian ring containing a field K of
characteristic zero and let R = A[X1 , . . . , Xm ]. Consider R as standard graded
with deg A = 0 and deg Xi = 1 for all i. We present a few results about the
behavior of the graded components of local cohomology modules HIi (R) where
I is an arbitrary homogeneous ideal in R. We mostly restrict our attention to
the Vanishing, Tameness and Rigidity problems.
1. introduction
1.1. Standard assumption: Throughout this paper A is a commutative Noetherian
ring containing a field K of characteristic zero. Let R = A[X1 , . . . , Xm ] be standard
graded with deg A = 0 and deg Xi = 1 for all i. We also assume m ≥ 1. Let I be
a homogeneous ideal L
in R and M = HIi (R). It is well-known that M is a graded
R-module. Set M = n∈Z Mn .
In paper [4], the first author studied graded components of local cohomology
modules when A is a regular ring containing a field K of characteristic zero and
he proved that they satisfy many good properties. In this paper we show that few
of the results proved in [4] (in particular Theorems 1.2, 1.3 and 1.6) hold true in
considerably general setting.
In this paper we show that under assumptions as in 1.1 the following results hold
true.
We first show
Theorem 1.2. (with hypotheses as in 1.1). If Mn = 0 for all |n| ≫ 0 then M = 0.
L
In view of Theorem 1.2, it follows that if M = HIi (R) = n∈Z Mn is non-zero
then either Mn 6= 0 for infinitely many n ≪ 0, OR, Mn 6= 0 for infinitely many
n ≫ 0. We show that M is tame. More precisely
Theorem 1.3. (with hypotheses as in 1.1). Then we have
(a) The following assertions are equivalent:
(i) Mn 6= 0 for infinitely many n ≪ 0.
(ii) Mn 6= 0 for all n ≤ −m.
(b) The following assertions are equivalent:
(i) Mn 6= 0 for infinitely many n ≫ 0.
(ii) Mn 6= 0 for all n ≥ 0.
Surprisingly non-vanishing of a single graded component of M = HIi (R) is very
strong. We prove the following rigidity result:
Date: August 7, 2017.
Key words and phrases. local comohology, graded local cohomology, Weyl algebra, generalized
Eulerian modules.
1
2
TONY J. PUTHENPURAKAL AND SUDESHNA ROY
Theorem 1.4. (with hypotheses as in 1.1). The we have
(a) The following assertions are equivalent:
(i) Mr 6= 0 for some r ≤ −m.
(ii) Mn 6= 0 for all n ≤ −m.
(b) The following assertions are equivalent:
(i) Ms 6= 0 for some s ≥ 0.
(ii) Mn 6= 0 for all n ≥ 0.
(c) (When m ≥ 2.) The following assertions are equivalent:
(i) Mt 6= 0 for some t with −m < t < 0.
(ii) Mn 6= 0 for all n ∈ Z.
We also give examples to show that Theorems 1.7, 1.8, 1.9, 1.13, 1.14 from [4]
are false if we don’t assume that A is regular.
Techniques used to prove our results: The main idea in [4] is that if A =
K[[Y1 , . . . , Yd ]] then graded local cohomology modules over R = A[X1 , . . . , Xm ]
become holonomic over an appropriate ring of differential operators, see [4, Theorem 4.2]. Furthermore local cohomology modules are “generalized Eulerian” see
[4, Theorem 3.6] (we will discuss this property later). One of the observation we
first made was that in Theorems 1.2, 1.3 and 1.6 in [4] the generalized Eulerian
property of local cohomology modules was used more crucially. Furthermore in [4]
the generalized Eulerian property of local cohomology modules was proved very
generally.
We now discuss the techniques used to prove our results:
Let A be a Noetherian ring containing a field of characteristic zero. Let R =
A[X1 , . . . , Xm ], graded with deg A = 0 and deg Xi = 1 for all i. Let Am (A) be the
mth Weyl-algebra on A. We consider it a graded ring with deg A = 0, deg Xi = 1
and deg ∂i = −1. We note that R is a graded subring of Am (A). If E is a graded
Am (A)-module and e is a homogeneous P
element of E then set |e| = deg e.
m
Consider the Eulerian operator E = i=1 Xi ∂i . If f ∈ R is homogeneous then
it is easy to check that Ef = |f |f . We say a graded Am (A)-module W is Eulerain
if Ew = |w|w for each homogeneous element w of W . Notice R is an Eulerian
Am (A)-module. We say W is generalized Eulerian if for each homogeneous w of W
there exists a depending on w such that (E − |w|)a w = 0.
The notion of Eulerian modules was introduced in the case A is a field K by Ma
and Zhang [9] (they also defined the notion of Eulerian D-modules in characteristic
p > 0, where D is the ring of K-linear differential operators on R = K[X1 , . . . , Xm ]).
Unfortunately however the class of Eulerian D-modules is not closed under extensions (see 3.5(1) in [9]). To rectify this, the first author introduced the notion
generalized Eulerian D-modules (in characteristic zero), see [6].
The next technique that we use is the technique of de Rham cohomology, Koszul
homology of generalized Eulerian modules. We generalize several properties which
were proved when A = K, a field and the relevant module was holonomic over
Am (K).
The final technique we use to prove vanishing and tameness property of local
cohomology modules is that local cohomology modules is countably generated over
R = A[X1 , . . . , Xm ]. We can exploit this fact if A contains an uncountable field.
This was observed first in [7]. By base change one can always assume that A
contains an uncountable field, see 6.1.
GRADED COMPONENTS OF LOCAL COHOMOLOGY MODULES
3
Remark 1.5. Note that practically all the proofs in this paper use techniques which
were first developed by several researches (including the first author) for holonomic
modules over some appropriate ring of differential operators. In the generality we
consider, we do not have the notion of holonomic modules. However we have the
notion of generalized Eulerian modules. After carefully rewriting several proofs we
were able to deduce our results.
The paper is organized as follows. In Section 2, we discuss a few preliminary results that we need. In Section 3, we define generalized Eulerian modules on Am (A),
the mth -Weyl algebra over A and talk about some basic properties of these modules. In Section 4, some properties of Am (A)-modules are discussed. In Section 5,
we show that HIi (M ) has some nice properties due to its countable generation as an
R = A[X1 , . . . , Xm ]-module. In Section 6 and Section 7 we prove our main results
regarding vanishing, tameness and rigidity of graded local cohomology modules. In
the last Section we give examples to show that Theorems 1.7, 1.8, 1.9, 1.13, 1.14
from [4] are false if we don’t assume that A is regular.
Convention: Throughout modules considered are left modules over some ring
(commutative or not). Usually A, B, R, S will denote commutative rings. The only
non-commutative ring we will consider will be Am (A), the mth Weyl algebra over
a commutative ring A.
2. Preliminaries
In this section we define Lyubeznik functor T and discuss a few preliminary
results that we need.
2.1. Lyubeznik functors
Let B be a commutative Noetherian ring and let X = Spec(B). Let Y be a
closed subset of X. If M is a B-module and if Y is a locally closed subscheme of
Spec(B), we denote by HIi (M ) the ith local cohomology module of M supported in
Y . Suppose Y = Y2 − Y1 where Y1 ⊂ Y2 are two closed subsets of X, then we have
an exact sequence of functors
(2.1.1)
(−) → · · · .
· · · → HYi 1 (−) → HYi 2 (−) → HYi (−) → HYi+1
1
A Lyubeznik functor T is any functor of the form T = T1 ◦ T2 ◦ · · · ◦ Tm where each
Ti is either HYi (−) for some locally closed subset Y of X or the kernel, image or
cokernel of some arrow appearing in (2.1.1) for closed subsets Y1 , Y2 of X such that
Y1 ⊂ Y2 .
We need the following result from [10, Lemma 3.1] regarding the behavior of
Lyubeznik functor under flat maps.
Proposition 2.2. Let φ : B → C be a flat homomorphism of Noetherian rings. Let
M be any B-module and T be a Lyubeznik functor on Mod(B). Then there exists
a Lyubeznik functor Tb on Mod(C) and isomorphisms Tb (M ⊗B C) ∼
= T (M ) ⊗B C
which is functorial in M .
2.3. Graded Lyubeznik functors
Let B be a commutative Noetherian ring and let R = B[X1 , . . . , Xm ] be standard
graded. We say Y is homogeneous closed subset of Spec(R) if Y = V (f1 , . . . , fs ),
where fi ’s are homogeneous polynomials in R. We say Y is homogeneous locally
closed subset of Spec(R) if Y = Y ′′ − Y ′ where Y ′ ⊂ Y ′′ are homogeneous closed
4
TONY J. PUTHENPURAKAL AND SUDESHNA ROY
subsets of Spec(R). Let ∗ Mod(R) be the category of graded R-modules. Then we
have an exact sequence of functors on ∗ Mod(R),
(2.3.2)
· · · → HYi ′ (−) → HYi ′′ (−) → HYi (−) → HYi+1
′ (−) → · · · .
Definition 2.4. A graded Lyubeznik functor T is a composite functor of the form
T = T1 ◦ T2 ◦ · · · ◦ Tm where each Tj is either HYi j (−) for some homogeneous locally
closed subset Yj of Spec(R) or the kernel, image or cokernel of any arrow appearing
in (2.3.2) with Y = Yj , Y ′ = Yj′ and Y ′′ = Yj′′ such that Yj = Yj′′ − Yj′ and Yj′ ⊂ Yj′′
are homogeneous closed subsets of Spec(R).
2.5. Graded Lyubexnik functor under flat maps:
Let φ : B → C be a flat homomorphism of Noetherian rings and let R =
B[X1 , . . . , Xm ] be standard graded. Set S = R ⊗B C = C[X1 , . . . , Xm ]. Clearly
φ induces a map φ′ : R → S. Let T be a graded Lyubeznik functor on ∗ Mod(R).
If Y is a homogeneous closed subset of Spec(R), say Y = V (f1 , . . . , fs ). Set Y ′ =
V (f1′ , . . . , fs′ ) where fi′ = φ′ (fi ). We note that Y ′ is a homogeneous closed subset
of Spec(S). Furthermore it is clear that we have a homogeneous isomorphism
HYi (−) ⊗B C = HYi ′ (−). If Y is a homogeneous locally closed subset of Spec(R),
say Y = Y2 − Y1 . Set Y ′ = Y2′ − Y1′ . Then clearly Y ′ is a homogeneous locally
closed subset of Spec(S). Furthermore applying the functor (−) ⊗B C to (2.3.2)
yields an exact sequence
· · · → HYi 1′ (−) → HYi 2′ (−) → HYi ′ (−) → HYi+1
′ (−) → · · · .
1
∗
More generally if T is a graded Lyubeznik functor on Mod(R) then T ⊗B C is a
graded Lyubeznik functor on ∗ Mod(S).
3. Generalized Eulerian modules
Let A be a commutative Noetherian ring containing a field K of characteristic
zero. Let Am (A) be the mth -Weyl algebra over A. Notice Am (A) = A⊗K Am (K) =
AhX1 , . . . , Xm , ∂1 , . . . , ∂m i. We can consider Am (A) graded, by giving deg A =
0, deg Xi = 1, deg ∂i = −1.
The Euler operator on Am (A), denoted by Em , is defined as
Em :=
m
X
Xi ∂i .
i=1
Note that deg Em = 0. Let E be a graded Am (A)-module. For any homogeneous
element e ∈ E, set |e| = deg e.
Definition 3.1. A graded Am (A)-module M is said to be generalized Eulerian if
for each homogeneous element e of E there exists a positive integer a (depending
on e) such that
(Em − |e|)a · e = 0.
The following result from [4, Theorem 3.6] is extremely useful to us.
Theorem 3.2. Let T (R) be a graded Lyubeznik functor on ∗ Mod(R). Then T (R)
is a generalized Eulerian Am (A)-module.
The following result is well-known.
GRADED COMPONENTS OF LOCAL COHOMOLOGY MODULES
5
Lemma 3.3. If Γ is left (resp. right) Noetherian (not necessarily commutative)
ring then for all m ≥ 1, the Weyl Algebra Am (Γ) is also left (resp. right) Noetherian
for all m.
The following property of generalized Eulerian modules is proved in [4].
Lemma 3.4. [4, Proposition 3.11] Let 0 → M1 → M2 → M3 → 0 be a short exact
sequence of Am (A)-modules. Then M2 is generalized Eulerian if and only if M1
and M3 are generalized Eulerian.
First we develop some basic properties.
Proposition 3.5. Let M be a generalized Eulerian Am (A)-module. Then for i =
0, 1, the Am−1 (A)-module Hi (∂m ; M )(−1) is generalized Eulerian.
Proof. From the short exact sequence
∂
m
M → H0 (∂m ; M ) → 0
0 → H1 (∂m ; M ) → M (1) −→
we get H1 (∂m ; M ) = ker ∂m = {a ∈ M | ∂m a = 0} and H0 (∂m ; M ) = M/∂m M .
Note Am−1 (A) is a sub-ring of Am (A). Further as ∂m commutes with all elements
∂
m
M is Am−1 (A)-linear. So H0 (∂m ; M ) and
in Am−1 (A) we get the map M (1) −→
H1 (∂m ; M ) are Am−1 (A)-modules. Note that H1 (∂m ; M )(−1) ⊆ M . Let ξ ∈
H1 (∂m ; M )(−1) be homogeneous. As M is generalized Eulerian so we have
(Em − |ξ|)a ξ = 0 for some a ≥ 1.
Now by definition Em = Em−1 + Xm ∂m . Since Xm ∂m commutes with Em−1 so we
have
0 = (Em−1 − |ξ| + Xm ∂m )a ξ = ((Em−1 − |ξ|)a + (∗)Xm ∂m )ξ.
Here ∗ denotes the irrelevant terms. As ∂m ξ = 0 we get that (Em−1 − |ξ|)a ξ = 0.
Thus H1 (∂m ; M )(−1) is a generalized Eulerian Am−1 (A)-module.
Let ξ ′ ∈ H0 (∂m ; M )(−1) be homogeneous of degree r. Since H0 (∂m ; M )(−1) =
(M/∂m M )(−1) so ξ ′ = α + ∂m M where α ∈ Mr−1 . As M is generalized Eulerian
we get
0 = (Em − r + 1)b α = 0 for some b ≥ 1.
Now Em = Em−1 + Xm ∂m = Em−1 + ∂m Xm − 1. Therefore Em − r + 1 = Em−1 −
r + ∂m Xm . Moreover, ∂m Xm commutes with Em−1 . Thus
0 = (Em−1 − r + ∂m Xm )b α = (Em−1 − r)b α + ∂m · (∗)α.
Going mod ∂m M we get
(Em−1 − r)b ξ ′ = 0.
Hence H0 (∂m ; M )(−1) is a generalized Eulerian Am−1 (A)-module.
Proposition 3.6. Let M be a generalized Eulerian Am (A)-module. Then for i =
0, 1, the Am−1 (A)-module Hi (Xm ; M ) is generalized Eulerian.
Proof. From the short exact sequence
X
m
M → H0 (Xm ; M ) → 0
0 → H1 (Xm ; M ) → M (−1) −→
we get H1 (Xm ; M ) = ker Xm = {a ∈ M | Xm a = 0} and H0 (Xm ; M ) = M/Xm M .
Note Am−1 (A) is a sub-ring of Am (A). Further as Xm commutes with all elements
X
m
M is Am−1 (A)-linear. So H0 (Xm ; M ) and
in Am−1 (A) we get the map M (−1) −→
H1 (Xm ; M ) are Am−1 (A)-modules.
6
TONY J. PUTHENPURAKAL AND SUDESHNA ROY
Let η ∈ H1 (Xm ; M ) be homogeneous. Note that η ∈ H1 (Xm ; M )r ⊆ M (−1)r =
Mr−1 . As M is generalized Eulerian so we have
(Em − (r − 1))b η = 0
for some b ≥ 1.
Now ∂m Xm − Xm ∂m = 1 and by definition Em = Em−1 + Xm ∂m . Therefore
Em = Em−1 + ∂m Xm − 1. Thus
(Em−1 − r + ∂m Xm )b η = 0.
Since ∂m Xm commutes with Em−1 so we have
0 = (Em−1 − r + ∂m Xm )b η = (Em−1 − r)b η + (∗)∂m Xm η.
As Xm η = 0 we get that (Em−1 − r)b η = 0 for some b ≥ 1. Thus H1 (Xm ; M ) is a
generalized Eulerian Am−1 (A)-module.
Let η ′ ∈ H0 (Xm ; M ) be homogeneous of degree r. Since H0 (Xm ; M ) = M/Xm M
so η ′ = β + Xm M where β ∈ Mr . As M is generalized Eulerian we get
0 = (Em − r)a β = 0
for some a ≥ 1.
Now Em = Em−1 + Xm ∂m . Also Xm ∂m commutes with Em−1 . Thus
0 = (Em−1 − r + Xm ∂m )a β = (Em−1 − r)a β + Xm · (∗)β.
Going mod Xm M we get
(Em−1 − r)a η ′ = 0.
Hence H0 (Xm ; M ) is a generalized Eulerian Am−1 (A)-module.
The following two results are crucial.
Proposition 3.7. Let M be a generalized Eulerian A1 (A)-module. Then for l =
0, 1, the module Hl (∂1 ; M ) is concentrated in degree −1 (i.e., Hl (∂1 ; M )j = 0 for
j 6= −1).
∂
1
M is A-linear. So Hl (∂1 ; M ) is an A-module for l = 0, 1.
Proof. The map M (1) −→
Thus we have an exact sequence of A-modules
∂
1
M → H0 (∂1 ; M ) → 0.
0 → H1 (∂1 ; M ) → M (1) −→
Let ξ ∈ H1 (∂1 ; M )(−1) be homogeneous and non-zero. Notice H1 (∂1 ; M )(−1) ⊆
M . Since M is generalized Eulerian so we have
(X1 ∂1 − |ξ|)a ξ = 0 for some a ≥ 1.
Now (X1 ∂1 − |ξ|)a = (∗)∂1 + (−1)a |ξ|a . As ∂1 ξ = 0 we get (−1)a |ξ|a ξ = 0. Since
ξ 6= 0 so we get |ξ| = 0, i.e., ξ ∈ H1 (∂1 ; M )(−1)0 = H1 (∂1 ; M )−1 . Hence H1 (∂1 ; M )
is concentrated at −1.
Let ξ ′ ∈ H0 (∂1 ; M ) be non-zero and homogeneous of degree r. Therefore ξ ′ ∈
(M/∂1 M )r and hence ξ ′ = α + ∂1 M for some α ∈ Mr . Since M is generalized
Eulerian so we have
(X1 ∂1 − r)b α = 0
for some b ≥ 1.
Since ∂1 X1 − X1 ∂1 = 1, we can write
0 = (∂1 X1 − (r + 1))b α = (∂1 · (∗) + (−1)b (r + 1)b )α.
Now going mod ∂1 M we get (−1)b (r + 1)b ξ ′ = 0. As ξ ′ 6= 0 so r = −1, i.e.,
ξ ′ ∈ H0 (∂1 ; M )−1 . Hence H0 (∂1 ; M ) is concentrated at −1.
GRADED COMPONENTS OF LOCAL COHOMOLOGY MODULES
7
Proposition 3.8. Let M be a generalized Eulerian A1 (A)-module. Then for l =
0, 1, the module Hl (X1 ; M ) is concentrated in degree 0 (i.e., Hl (X1 ; M )j = 0 for
j 6= 0).
X
1
M is A-linear. So Hl (X1 ; M ) is an A-module for
Proof. The map M (−1) −→
l = 0, 1. Thus we have an exact sequence of A-modules
X
1
M → H0 (X1 ; M ) → 0.
0 → H1 (X1 ; M ) → M (−1) −→
Let η ∈ H1 (X1 ; M )(1) ⊆ M be homogeneous and non-zero. Since M is generalized
Eulerian so we have
(X1 ∂1 − |η|)b η = 0 for some b ≥ 1.
Since ∂1 X1 − X1 ∂1 = 1, we can write
0 = (∂1 X1 − (|η| + 1))b η = (∗)X1 η + (−1)b (|η| + 1)b η.
As X1 η = 0 and η 6= 0, we get |η| = −1, i.e., η ∈ H1 (X1 ; M )(1)−1 = H1 (X1 ; M )0 .
Hence H1 (X1 ; M ) is concentrated at 0.
Let η ′ ∈ H0 (X1 ; M ) be non-zero and homogeneous of degree r. Therefore η ′ ∈
(M/X1 M )r and hence η ′ = β + X1 M for some β ∈ Mr . Since M is generalized
Eulerian so we have
(X1 ∂1 − r)a β = 0
a
for some a ≥ 1.
a a
Notice (X1 ∂1 − r) = X1 · (∗) + (−1) r . Thus
X1 · (∗)β + (−1)a ra β = 0.
In M/X1 M , we have (−1)a ra η ′ = 0. As η ′ 6= 0, we get r = 0, i.e., η ′ ∈ H0 (X1 ; M )0 .
Hence H0 (X1 ; M ) is concentrated at 0.
4. Some properties of modules over the Weyl algebra of a
Noetherian ring containing a field of characteristic zero
We prove that some properties of Am (K)-modules can be extended to Am (A)modules. The following result is similar to [2, Theorem 6.2]. Note that in [2,
Theorem 6.2], M is assumed to be holonomic.
t
a = 0 for some
Lemma 4.1. Let M be an Am (A)-module. Set N1 = {a ∈ M | ∂m
t
t ≥ 1} and N2 = {a ∈ M | Xm a = 0 for some t ≥ 1}. Then
(1) N1 is a Am (A)-submodule of M .
(2) N1 = ∂m N1 .
(3) N2 is a Am (A)-submodule of M .
(4) N2 = Xm N2 .
s
Proof. (1) Let b ∈ N1 . Then ∂m
b = 0 for some s ≥ 1. Now ∂i ∂j = ∂j ∂i for all i, j
s
s
and Xi ∂j = ∂j Xi for all i 6= j in Am (A). Thus ∂m
∂i b = ∂i ∂m
b = 0 for all i and
s
s
∂m Xi b = Xi ∂m b = 0 for all i 6= m. Therefore ∂i b ∈ N1 for all i and Xi b ∈ N1 for
α
α
α−1
all i 6= m. Moreover, for any α ≥ 1 we have ∂m
Xm = Xm ∂m
+ α∂m
. Therefore
s+1
s+1
s
∂m Xm b = Xm ∂m b + (s + 1)∂m b = 0 and hence Xm b ∈ N1 . It follows that N1 is
a Am (A)-submodule of M .
s
(2) Clearly ∂m N1 ⊆ N1 . Let b ∈ N1 . Then ∂m
b = 0 for some s ≥ 1. We will use
induction on s to prove N1 ⊆ ∂m N1 . If s = 1, then ∂m (Xm b) = Xm (∂m b) + b = b.
Since Xm N1 ⊆ N1 so we get b ∈ ∂m N1 . We now assume that s ≥ 2 and the result
s
s
s
s−1
s−1
is true for s − 1. If ∂m
b = 0 then we get ∂m
Xm b = Xm ∂m
b + s∂m
b = s∂m
b. It
8
TONY J. PUTHENPURAKAL AND SUDESHNA ROY
s
s−1
s−1
follows that ∂m
Xm b − s∂m
b = ∂m
(∂m Xm b − sb) = 0. Thus ∂m Xm b − sb ∈ N1
and hence by induction hypothesis ∂m Xm b − sb ∈ ∂m N1 . Since ∂m Xm b ∈ ∂m N1 so
we get that sb ∈ ∂m N1 . Therefore b ∈ ∂m N1 and the result follows.
s
(3) Let b ∈ N2 . Then Xm
b = 0 for some s ≥ 1. Now Xi Xj = Xj Xi for all i, j
s
s
and Xi ∂j = ∂j Xi for all i 6= j in Am (A). Thus Xm
Xi b = Xi Xm
b = 0 for all i and
s
s
Xm ∂i b = ∂i Xm b = 0 for all i 6= m. Therefore Xi b ∈ N2 for all i and ∂i b ∈ N2 for
α
α
α−1
all i 6= m. Moreover, for any α ≥ 1 we have Xm
∂m = ∂m Xm
− αXm
. Therefore
s+1
s+1
s
Xm ∂m b = ∂m Xm b − (s + 1)Xm b = 0 and hence ∂m b ∈ N2 . It follows that N2 is
a Am (A)-submodule of M .
s
(4)Clearly Xm N2 ⊆ N2 . Let b ∈ N2 . Then Xm
b = 0 for some s ≥ 1. We will use
induction on s to prove N2 ⊆ Xm N2 . If s = 1, then Xm (∂m b) = ∂m (Xm b)−b = −b.
Since ∂m N2 ⊆ N2 so we get b ∈ Xm N2 . We now assume that s ≥ 2 and the result is
s
s
s
s−1
s−1
true for s − 1. If Xm
b = 0 then we get Xm
∂m b = ∂m Xm
b − sXm
b = −sXm
b. It
s
s−1
s−1
follows that Xm ∂m b + sXm b = Xm (Xm ∂m b + sb) = 0. Thus Xm ∂m b + sb ∈ N2
and hence by induction hypothesis Xm ∂m b + sb ∈ Xm N2 . Since Xm ∂m b ∈ Xm N2
so we get that sb ∈ Xm N2 . Therefore b ∈ Xm N2 and the result follows.
4.2. Fourier transform: Let F : Am (A) → Am (A) be an automorphism of Am (A)
defined by F (∂i ) = −Xi and F (Xi ) = ∂i for all i. Clearly F (X∂) = F (X)F (∂).
Notice F (1) = F (∂i Xi − Xi ∂i ) = F (∂i )F (Xi ) − F (Xi )F (∂i ) = −Xi ∂i + ∂i Xi = 1.
Also note that F : R = A[X1 , . . . , Xm ] → A[∂1 , . . . , ∂m ] = S (restriction map) is
an isomorphism. Let M F denote a new module such that M F = M as an abelian
group and the action of R on M F is defined by r · m = F (r)m where r ∈ R and
m ∈ M F . The new module M F is called the Fourier transform of M . Similarly
−1
we define s ∗ n = F −1 (s)n where s ∈ S and n ∈ M F . Now for any s ∈ S
−1
and m ∈ (M F )F we have s ∗ m = F −1 (s) · m = F (F −1 (s))m = sm. Hence
−1
(M F )F = M .
Let R be a commutative
S Noetherian ring and I be an ideal in R. Let M be an
R-module. Set ΓI (M ) = i≥1n∈N (0 :M I n ), the submodule of M consisting of all
elements of M which are annihilated by some power of I.
Lemma 4.3. Let M be an Am (A)-module. Set R = A[X1 , . . . , Xm ] and S =
A[∂1 , . . . , ∂m ]. Let I and J be homogeneous ideals in R and S respectively. Then
ΓI (M ) and ΓJ (M ) are also Am (A)-modules.
Proof. Clearly R is a commutative ring contained in the left Noetherian ring Am (A).
Let I be a homogeneous ideal in R.
Claim: I l+1 ∂i u ⊆ I l u for all 1 ≤ i ≤ m; l ≥ 0 and for any u ∈ M .
Proof. We know that I l+1 = hg1 · · · gl+1 | gi ∈ Ii. Moreover, ∂i g1 · · · gl+1 =
g1 · · · gl+1 ∂i + ∂i (g1 · · · gl+1 ). Thus g1 · · · gl+1 ∂i u = ∂i g1 · · · gl+1 u − ∂i (g1 · · · gl+1 )u.
Pl+1
By chain rule we get ∂i (g1 · · · gl+1 ) =
i=1 g1 · · · ∂i (gi ) · · · gl+1 . It follows that
g1 · · · gl+1 ∂i u ∈ I l u.
Let u ∈ ΓI (M ). Then I s u = 0 for some s ≥ 0. Now I s+1 ∂i u ⊆ ∂i I s u =
0. Therefore ∂i u ∈ ΓI (M ) and hence ∂i ΓI (M ) ⊆ ΓI (M ) for all i. Moreover,
Xi ΓI (M ) ⊆ ΓI (M ) and AΓI (M ) ⊆ ΓI (M ). It follows that ΓI (M ) is a Am (A)submodule of M .
Now ΓF (I) (M F ) = {m ∈ M F | F (I)s m = 0 for some s ≥ 0}. Since F is a ring
homomorphism so we get F (I)s m = F (I s )m = I s · m for any s ≥ 0. Therefore
ΓI (M )F = ΓF (I) (M F ). Let J be a homogeneous ideal in S. Set I = F −1 (J).
Clearly I is a homogeneous ideal in R and F (I) = F (F −1 (J)) = J. Now we have
GRADED COMPONENTS OF LOCAL COHOMOLOGY MODULES
9
ΓI (M F ) ⊆Am (A) M F , i.e., ΓF (J) (M F ) ⊆Am (A) M F . Taking F −1 both side we get
ΓJ (M ) ⊆Am (A) M .
Remark 4.4. Since (X1 , . . . , Xm ) and (∂1 , . . . , ∂m ) are homogeneous ideals in R
and S respectively so by the above lemma it follows that Γ(X) (M ) and Γ(∂) (M ) are
Am (A)-modules.
Proposition 4.5. Let M be a generalized Eulerian Am (A)-module. Then
(1) Γ(X1 ,...,Xm ) (M )j = 0 for all j ≥ −m + 1.
(2) Γ(∂1 ,...,∂m ) (M )j = 0 for all j ≤ −m.
Proof. (1) Use induction on m. Let m = 1. By Lemma 4.1 we have Γ(X1 ) (M ) =
X
1
Γ(X1 ) (M ) is surjective. Thus we get
X1 Γ(X1 ) (M ), i.e., the map Γ(X1 ) (M )(−1) −→
a short exact sequence
X
1
Γ(X1 ) (M ) → 0.
0 → L → Γ(X1 ) (M )(−1) −→
(4.5.3)
Since M is generalized Eulerian we get by Lemma 3.4 and Remark 4.4 that Γ(X1 ) (M )
is a generalized Eulerian A1 (A)-module. Therefore L = H1 (X1 ; Γ(X1 ) (M )) and is
concentrated at degree 0 by Proposition 3.8. Thus by the exact sequence (4.5.3)
X
1
Γ(X1 ) (M )j for all j 6= 0. Therefore
we have 0 → Γ(X1 ) (M )j−1 −→
X1
X1
X1
X1
Γ(X1 ) (M )0 ֒→ Γ(X1 ) (M )1 ֒→ Γ(X1 ) (M )2 ֒→ Γ(X1 ) (M )3 ֒→ · · · .
Let u ∈ Γ(X1 ) (M )i for some i ≥ 0. Then X1l u = 0 for some l ≥ 1. Since
Γ(X1 ) (M )i ֒→ Γ(X1 ) (M )i+l so we get u = 0 and hence Γ(X1 ) (M )i = 0 for all
i ≥ 0.
We now assume the result is true for m − 1. Notice Γ(X) (M ) ⊆ M is generalized
t
a = 0 for some t ≥ 1}.
Eulerian Am (A)-module. Set N = {a ∈ Γ(X) (M ) | Xm
Since (Xm ) ⊆ (X) we have N = Γ(X) (M ). By Lemma 4.1, N = Xm N , i.e.,
X
m
Γ(X) (M ) is a surjective map. Thus we get a short exact sequence
Γ(X) (M )(−1) −→
(4.5.4)
X
m
0 → V → Γ(X) (M )(−1) −→
Γ(X) (M ) → 0.
By Proposition 3.6, V = H1 (Xm ; Γ(X) (M )) is a generalized Eulerian Am−1 (A)module. Notice V = Γ(X1 ,··· ,Xm−1 ) (V ). So by induction hypothesis we have Vj = 0
for all j ≥ −(m − 1) + 1 = −m + 2. Now from exact sequence (4.5.4) we have
Xm
Xm
Xm
Γ(X) (M )−m+1 ֒→ Γ(X) (M )−m+2 ֒→ Γ(X) (M )−m+3 ֒→ · · · .
l
Let u ∈ Γ(Xm ) (M )−m+j for some j ≥ 1. Then Xm
u = 0 for some l ≥ 1. Since
Γ(X) (M )−m+j ֒→ Γ(X) (M )−m+j+l so we get u = 0 and hence Γ(X) (M )i = 0 for all
i ≥ −m + 1.
(2) Similarly like above we will use induction on m. Let m = 1. By Lemma 4.1,
∂
1
Γ(∂1 ) (M ) is surjective. Thus
Γ(∂1 ) (M ) = ∂1 Γ(∂1 ) (M ), i.e., the map Γ(∂1 ) (M )(1) −→
we get a short exact sequence
(4.5.5)
∂
1
Γ(∂1 ) (M ) → 0.
0 → L → Γ(∂1 ) (M )(1) −→
Since M is generalized Eulerian so we get by Lemma 3.4 and Remark 4.4 that
Γ(∂1 ) (M ) ⊆ M is a graded generalized Eulerian A1 (A)-module. Therefore L =
H1 (∂1 ; Γ(∂1 ) (M )) and is concentrated at degree −1 by Proposition 3.7. Thus by
10
TONY J. PUTHENPURAKAL AND SUDESHNA ROY
∂
1
Γ(∂1 ) (M )j for all j 6= −1.
the exact sequence (4.5.5) we have 0 → Γ(∂1 ) (M )j+1 −→
Therefore
∂1
∂1
∂1
∂1
Γ(∂1 ) (M )−1 ֒→ Γ(∂1 ) (M )−2 ֒→ Γ(∂1 ) (M )−3 ֒→ Γ(∂1 ) (M )−4 ֒→ · · · .
Let u ∈ Γ(∂1 ) (M )i for some i ≤ −1. Then ∂1l u = 0 for some l ≥ 1. Since
Γ(∂1 ) (M )i ֒→ Γ(∂1 ) (M )i−l so we get u = 0 and hence Γ(∂1 ) (M )i = 0 for all i ≥ 0.
We now assume the result is true for m − 1. Notice Γ(∂) (M ) ⊆ M is generalized
t
a = 0 for some t ≥ 1}. Since
Eulerian Am (A)-module. Set N = {a ∈ Γ(∂) (M ) | ∂m
∂
m
(∂m ) ⊆ (∂) we have N = Γ(∂) (M ). By Lemma 4.1, N = ∂m N , i.e., Γ(∂) (M )(1) −→
Γ(∂) (M ) is a surjective map. Thus we get a short exact sequence
(4.5.6)
∂
m
Γ(∂) (M ) → 0.
0 → V → Γ(∂) (M )(1) −→
By Proposition 3.6, V = H1 (∂m ; Γ(∂) (M )) is a generalized Eulerian Am−1 (A)module. Notice V = Γ(∂1 ,··· ,∂m−1 ) (V ). So by induction hypothesis we have Vj = 0
for all j ≤ −(m − 1) = −m + 1. From exact sequence (4.5.6) we have
∂m
∂m
∂m
Γ(∂) (M )−m ֒→ Γ(∂) (M )−m−1 ֒→ Γ(∂) (M )−m−2 ֒→ · · · .
l
Let u ∈ Γ(∂m ) (M )−m−j for some j ≥ 0. Then ∂m
u = 0 for some l ≥ 1. Since
Γ(∂) (M )−m−j ֒→ Γ(∂) (M )−m−j−l so we get u = 0 and hence Γ(∂) (M )j = 0 for all
j ≤ −m.
5. Countable generation
Throughout this section B is a commutative Noetherian ring.
We say a B-module M is countably generated if there exists a countable set of
generators of M as a B-module.
The following result is well-known. We give a proof for the convenience of the
reader.
f
g
Lemma 5.1. Let 0 → X1 −→ X2 −→ X3 → 0 be a short exact sequence of Bmodules. Then X2 is countably generated if and only if X1 and X3 are countably
generated.
Proof. We may assume that f is inclusion
S and g is the quotient map. Let X2 be
countably generated. So we have X2 = i∈N Di where Di ’s are finitely generated
B-modules and DS0 ⊆ D1 ⊆ D2 · · · . Now f (X1 ) = X1 is a submodule of X2 . Thus
X1 = X2 ∩ X1 = i∈N (Di ∩ X1 ). Moreover, B is a Noetherian ring and Di ∩ X1 is a
submodule of Di . Therefore Di ∩X1 is finitely generated and hence X1 is countably
generated. Since X3 ∼
= X2 /X1 as B-modules so we get X3 is countably generated
with same generating set as of X2 .
Conversely, let X1 and X3 are generated by {αn }n≥1 and {γn }n≥1 respectively.
Let βi ∈ X2 be such that g(βi ) = βi = γi .
Claim: X2 is generated by {α1 , . P
. . , β1 , . . .}.
Proof.PLet x ∈ X2 . Then g(x) = bi γi where bi = 0 for all but finitely manyPi. Set
∗
x∗ = bi βi . Then g(x) = g(x∗ ) and hence x−x∗ ∈ P
ker g = XP
ai αi
1 . So x−x =
where ai = 0 for all but finitely many i. Thus x = bi βi + ai αi .
The result follows.
The following result is definitely known. We give a proof for the convenience of
the reader.
GRADED COMPONENTS OF LOCAL COHOMOLOGY MODULES
d0
d1
11
d2
Lemma 5.2. Let X : 0 → X 0 −→ X 1 −→ X 2 −→ · · · be a co-chain complex of
countably generated B-modules. Then H i (X) is countably generated for all i.
Proof. Notice ker di is a submodule of X i . Since X i is countably generated so by
Lemma 5.1 we get ker di is countably generated. Thus, it follows from Lemma 5.1
that H i (X) = ker di / image di−1 is countably generated.
Lemma 5.3. Let I be an ideal in B. Let M be a countably generated B-module.
Then HIi (M ) is countably generated.
Proof. Let I = (f1 , . . . , fs ). Consider the Čech complex
C:0→M →
s
M
Mfi → · · · → Mf1 ···fs → 0.
i=1
Claim: If N is a countably generated B-module then Nf is countably generated for
any f ∈ B.
S
Proof. Clearly Nf = i≥1 N1/f i . Choose a countable generating set C of N . Then
i
C
S1/f i = {c/f | c ∈ C} is a generating set of N1/f i as a B-module and hence
C i is a generating set of Nf . Since C is a countable set so is C1/f i . Thus
Si≥1 1/f
C
i≥1 1/f i is a countable set. It follows that Nf is countably generated.
Since M is countably generated so we get C is a complex of countably generated
B-modules. Thus by Lemma 5.2 it follows that HIi (M ) = H i (C) is countably
generated.
Lemma 5.4. Let T be a Lyubeznik functor on Mod(B). Then T (M ) is a countably
generated B-module for any countably generated B-module M .
Proof. Let Y be any locally closed subset of Spec(B) with Y = Y ′′ − Y ′ where
Y ′ , Y ′′ are closed subsets of Spec(B). Then for any countably generated B-module
N we have exact sequence
HYi ′′ (N ) → HYi (N ) → HYi+1
′ (N ).
Now by Lemma 5.3 we get HYi ′′ (N ) and HYi+1
′ (N ) are countably generated. Moreover, B is Noetherian. So by Lemma 5.1 it follows that HYi (N ) is countably generated. Thus kernel and cokernel of any arrow appearing in the long exact sequence
· · · → HYi ′ (N ) → HYi ′′ (N ) → HYi (N ) → HYi+1
′ (N ) → · · · are countably generated.
Set T = T1 ◦ T2 ◦ · · · ◦ Tm . Then by the above argument we get that Tj (N ) is
countably generated for any countably generated B-module N and for all j. Notice
T (M ) = (T1 ◦ T2 ◦ · · · ◦ Tm )(M ) = T1 (· · · (Tm (M ))). Hence T (M ) is countably
generated for any countably generated B-module M .
Now we have the following result from [7, Lemma 2.3].
Lemma 5.5. Let E be a countably generated B-module. Then AssB (E) is a countable set. In particular, AssB T (B) is a countable set for any Lyubeznik functor T
on Mod(B).
Proof. The result follows from Lemma 5.3 and Lemma 5.5.
5.6. Let A be a commutative Noetherian ring containing a field K of characteristic
zero. Let R = A[X1 , . . . , Xm ] and S = A[∂1 , . . . , ∂m ]. We consider R and S are
graded rings with deg A = 0, deg Xi = 1 and deg ∂i = −1.
12
TONY J. PUTHENPURAKAL AND SUDESHNA ROY
Lemma 5.7. Let M
(1) M is a countably
(2) M is a countably
(3) M is a countably
be an Am (A)-module. The following are equivalent:
generated Am (A)-module.
generated R-module.
generated S-module.
Proof. Notice R and S are sub-rings of Am (A). So clearly (2) =⇒ (1) and
(3) =⇒ (1). Set Am (A) = D.
We now prove (1) =⇒ (3).
S
Let M be a countably generated D-module. Therefore M = i∈N Mi where Mi ’s
are finitely generated D-modules and M0 ⊆ M1 ⊆ M2 · · · . Since Mi is a finitely
generated D-module so there exists a surjectiveS map Dli → Mi → 0 for some
li ≥ 0. Moreover, as a S-module we have D = i∈N Di , where Di is a left (resp.
right) S-module generated by all monomials in X1 , . . . , Xm of degree less than or
equal to i. Clearly Di ’s are finitely generated left (resp. right) S-modules and
D0 ⊆ D1 ⊆ D2 · · · . Thus D is a countably generated left (resp. right) S-module
and so is Dli . Furthermore by Lemma 5.1 we get Mi is countably generated Smodule for all i. It follows that M is a countably generated S-module.
To prove (1) =⇒ (2) we just have to replace S by R and Xi by ∂i for all i in
the proof of (1) =⇒ (3).
This completes the proof.
We need the following result from [4, Proposition 12.1] to prove the result next
to it.
Proposition 5.8. Let f : U → V be a homomorphism of commutative Noetherian
rings. Let M be a B-module. Then
AssU M = {P ∩ U | P ∈ AssV M }.
We now prove the following nice result.
L
Proposition 5.9. Let M = i∈Z Mi be a countably generated R-module and is a
generalized Eulerian Am (A)-module. Let K be uncountable of characteristic zero.
Then there exists some homogeneous element η ∈ K[X1 , . . . , Xm ] of degree 1 such
·η
that Mi → Mi+1 is an injective map for all i ≥ −m + 1 and some homogeneous
·ξ
element ξ ∈ K[∂1 , . . . , ∂m ] of degree −1 such that Mi → Mi−1 is an injective map
for all i ≤ −m.
Proof. By Proposition 4.5 we have Γ(X) (M )j = 0 for all j ≥ −m + 1. Now we have
a short exact sequence 0 → Γ(X) (M ) → M → M/Γ(X) (M ) → 0. Therefore Mj =
M/Γ(X) (M ) j for all j ≥ −m + 1. Again R is Noetherian and I = (X1 , . . . , Xm )
is an ideal in R. So by [5, Lemma 6.2],
AssR
M
= {P ∈ AssR M | P + (X)}.
Γ(X) (M )
Set M = M/Γ(X) (M ). Note if P ∈ AssR M then P is homogeneous. Now by
Lemma 5.1 we get M is countably generated R-module and hence by Lemma 5.5
it follows that AssR M is a countable set. Notice C = K[X1 , . . . , Xm ] ⊂ R is a
sub-ring. So by Proposition 5.8 we have AssC M = AssR M ∩ C. Thus for any
Q ∈ AssC M then there exists some P ∈ AssR M such that Q = P ∩ C and
hence Q is homogeneous. Also as AssR M is a countable set so is AssC M . Now
GRADED COMPONENTS OF LOCAL COHOMOLOGY MODULES
13
S
for any P ∈ AssR M we have P ∩ C1 ( C1 . Denote H = P ∈AssR M P ∩ C1 .
Since K is uncountable so H ( C1 . Thus there exists some η ∈ C1 − H. Now
η = b1 X1 + · · · + bm Xm ∈ K[X1 , · · · , Xn ] is a non-zero divisor on M . Thus
·η
Mj → Mj+1 is an injective map for all j ≥ −m + 1.
By Proposition 4.5 we have Γ(∂) (M )j = 0 for all j ≤ −m. Now we have a
short exact sequence
0 → Γ(∂) (M ) → M → M/Γ(∂) (M ) → 0. Therefore Mj =
M/Γ(∂) (M ) j for all j ≤ −m. Again S is Noetherian and I = (∂1 , . . . , ∂m ) is an
ideal in S. So by [5, Lemma 6.2],
AssS
M
= {P ∈ AssS M | P + (∂)}.
Γ(∂) (M )
Set M = M/Γ(∂) (M ). Note if P ∈ AssS M then P is homogeneous. Since M
is a countably generated R-module so by Lemma 5.7 we have M is a countably
generated S-module. Then by Lemma 5.1 we get M is a countably generated
S-module. So by Lemma 5.5 it follows that AssS M is a countable set. Notice
T = K[∂1 , . . . , ∂m ] ⊂ S is a sub-ring. So by Proposition 5.8 we have AssT M =
AssS M ∩ T . Thus for any Q ∈ AssT M then there exists some P ∈ AssS M such
that Q = P ∩ T and hence Q is homogeneous. Also as AssS M is a countable
set soSis AssT M . Now for any P ∈ AssS M we have P ∩ T−1 ( T−1 . Denote
H = P ∈AssS M P ∩ T−1 . As K is uncountable so H ( T−1 . Thus there exists some
ξ ∈ T−1 − H. Now ξ = a1 ∂1 + · · · + am ∂m ∈ K[∂1 , . . . , ∂m ] is a non-zero divisor on
·ξ
M . Thus Mj → Mj−1 is an injective map for all j ≤ −m.
As a consequence of Proposition 5.9 we get the following results.
Corollary 5.10. Let K be uncountable.
Let T be a graded Lyubeznik functor on
L
Mod(R) and set M = T (R) = n∈Z Mn . Then there exists some homogeneous
·η
element η ∈ K[X1 , . . . , Xm ] of degree 1 such that Mi → Mi+1 is an injective map
for all i ≥ −m + 1 and some homogeneous element ξ ∈ K[∂1 , . . . , ∂m ] of degree −1
∗
·ξ
such that Mi → Mi−1 is an injective map for all i ≤ −m.
Proof. By Lemma 5.4 we have M is a countably generated R-module. Moreover,
by Theorem 3.2 we have M is a generalized Eulerian Am (A)-module. So the result
follows from Proposition 5.9.
L
Corollary 5.11. Let K be uncountable. Let M = i∈Z Mi be a countably generated R-module and a generalized Eulerian Am (A)-module. If M 6= 0 then Mj 6= 0
for infinitely many j ≫ 0 OR Mj 6= 0 for infinitely many j ≪ 0.
Proof. Suppose if possible there exists r ≥ −m + 1 such that Mr 6= 0 and Mn = 0
for all n > r. But by Lemma 5.9 there exists some homogeneous element η ∈
·η
K[X1 , . . . , Xm ] of degree 1 such that Mi → Mi+1 is an injective map for all i ≥
·η
−m + 1. In particular, we have an injective map Mr → Mr+1 . Thus Mr+1 6= 0, a
contradiction.
Again let if possible there exists s ≤ −m such that Ms 6= 0 and Mn = 0 for all
n < s. But by Lemma 5.9 there exists some homogeneous element ξ ∈ K[∂1 , . . . , ∂m ]
·ξ
of degree −1 such that Mi → Mi−1 is an injective map for all i ≤ −m. In particular,
·ξ
we have an injective map Ms → Ms−1 . Thus Ms−1 6= 0, a contradiction.
14
TONY J. PUTHENPURAKAL AND SUDESHNA ROY
L
Corollary 5.12. Let K be uncountable. Let M = i∈Z Mi be a countably generated R-module and a generalized Eulerian Am (A)-module. Then
Mn0 6= 0 for some n0 ≥ −m + 1 =⇒ Mn 6= 0 for all n ≥ n0 ,
Mn0 6= 0 for some n0 ≤ −m =⇒ Mn 6= 0 for all n ≤ n0 .
Proof. By Corollary 5.10 there exists some degree 1 homogeneous element η ∈
η
K[X1 , . . . , Xm ] such that Mi → Mi+1 is an injective map for all i ≥ −m + 1. In
·η
particular, for n0 ≥ −m + 1 we have an injective map Mn0 → Mn0 +1 and hence
Mn0 +1 6= 0. Proceeding similarly we get Mn 6= 0 for all n ≥ n0 .
Furthermore by Corollary 5.10 there exists some homogeneous element ξ ∈
·ξ
K[∂1 , . . . , ∂m ] of degree −1 such that Mj → Mj−1 is an injective map for all
·ξ
j ≤ −m. In particular, for n0 ≤ −m we have an injective map Mn0 → Mn0 −1 and
hence Mn0 −1 6= 0. Proceeding similarly we get Mn 6= 0 for all n ≤ n0 .
6. Vanishing and Tameness
6.1. Key fact: Let A be a commutative Noetherian ring containing a field K of
characteristic zero. Let R = A[X1 , . . . , Xm ] be standard graded with deg A = 0
and deg Xi = 1Lfor all i. Let T be a graded Lyubeznik functor on ∗ Mod(R) and set
M = T (R) = n∈Z Mn . For convenience we have to assume K is uncountable. If
this is not the case, then consider the flat extension A → A[[Y ]]Y . Set B = A[[Y ]]Y .
Then by [7, Remark 3.1, Proposition 3.2] we have B contains a uncountable field
K[[Y ]]Y and B is a faithfully flat extension of A. Set S = B[X1 , . . . , Xm ] and
N = M ⊗R S. Now by Lemma 5.4 we have M is a countably generated R-module. So
N is a countably generated S-module. Furthermore the flat extension A → A[[Y ]]Y
induces a flat homomorphism R → S. By Sub-section 2.5 we get T ⊗R S is a graded
Lyubeznik functor
on ∗ Mod(S). Set Tb = T ⊗R S. Notice N = Tb (S) = M ⊗R S =
L
M ⊗A B = n∈Z (Mn ⊗A B). Since B is a faithfully flat extension of A so we have
M = 0 if and only if N = 0 and Mn = 0 if and only if Mn ⊗A B = 0.
Vanishing of the components plays a crucial role in the study of the graded local
cohomology. We first prove the following result.
Theorem 6.2. Let A be a commutative Noetherian ring containing a field K of
characteristic zero. Let R = A[X1 , . . . , Xm ] be a standard graded with deg A = 0
and deg Xi = L
1 for all i. Let T be a graded Lyubeznik functor on ∗ Mod(R) and set
M = T (R) = n∈Z Mn . If Mn = 0 for all |n| ≫ 0 then M = 0.
Proof. By the above discussion it is enough to prove the result considering K is
uncountable. Moreover by Theorem 3.2 we have M is a generalized Eulerian Am (A)module. Therefore if M 6= 0 then by Corollary 5.11 we have Mj 6= 0 for infinitely
many j ≫ 0 OR Mj 6= 0 for infinitely many j ≪ 0, a contradiction.
The following result shows that T (R) is tame.
Theorem 6.3. Let A be a commutative Noetherian ring containing a field K of
characteristic zero. Let R = A[X1 , . . . , Xm ] be a standard graded with deg A = 0
and deg Xi = L
1 for all i. Let T be a graded Lyubeznik functor on ∗ Mod(R) and set
M = T (R) = n∈Z Mn . Then
Mn0 6= 0 for some n0 ≥ −m + 1 =⇒ Mn 6= 0 for all n ≥ n0 ,
Mn0 6= 0 for some n0 ≤ −m =⇒ Mn 6= 0 for all n ≤ n0 .
GRADED COMPONENTS OF LOCAL COHOMOLOGY MODULES
15
Proof. Similarly as Theorem 6.2 it is enough to prove the result considering K is
uncountable. Furthermore, by Theorem 3.2 we have M is a generalized Eulerian
Am (A)-module. So by Corollary 5.12 we get
Mn0 6= 0 for some n0 ≥ −m + 1 =⇒ Mn 6= 0 for all n ≥ n0 ,
Mn0 6= 0 for some n0 ≤ −m =⇒ Mn 6= 0 for all n ≤ n0 .
The result follows.
7. Rigidity
We first show the following result.
L
Theorem 7.1. Let M = n∈Z Mn be a generalized Eulerian Am (A)-module. Then
(I) The following conditions are equivalent:
(a) Mn 6= 0 for infinitely many n < 0.
(b) There exists r such that Mn 6= 0 for all n ≤ r.
(c) Mn 6= 0 for all n ≤ −m.
(d) Ms 6= 0 for some s ≤ −m.
(II) The following conditions are equivalent:
(a) Mn 6= 0 for infinitely many n ≥ 0.
(b) There exists s such that Mn 6= 0 for all n ≥ s.
(c) Mn 6= 0 for all n ≥ 0.
(d) Mt 6= 0 for some t ≥ 0.
Proof. (I) Clearly (c) =⇒ (b) =⇒ (a) =⇒ (d). We only have to prove
(d) =⇒ (c). We will do this by induction on m.
We first assume m = 1. Consider the exact sequence
(7.1.7)
∂
1
Mj → H0 (∂1 ; M )j → 0.
0 → H1 (∂1 ; M )j → Mj+1 −→
Since M is a generalized Eulerian Am (A)-module so by Proposition 3.7, we get that
Hl (∂1 ; M ) is concentrated in degree −1 for l = 0, 1. Thus by exact sequence (7.1.7)
it follows that
Mj ∼
= M−1 for all j ≤ −2.
∼ Ms 6= 0 for all j ≤ −1.
Hence Mj =
We now assume that m ≥ 2 and the result is proved for m − 1. Consider the
exact sequence
(7.1.8)
∂
m
Mj → H0 (∂m ; M )j → 0.
0 → H1 (∂m ; M )j → Mj+1 −→
By Proposition 3.5 we get Hl (∂m ; M )(−1) is a generalized Eulerian Am−1 (A)module for l = 0, 1. We consider the following three cases:
Case 1: H0 (∂m ; M )(−1)j 6= 0 for some j ≤ −m + 1.
By induction hypothesis it follows that H0 (∂m ; M )(−1)j 6= 0 for all j ≤ −m + 1.
Hence H0 (∂m ; M )j 6= 0 for all j ≤ −m. So by exact sequence (7.1.8) we get that
Mj 6= 0 for all j ≤ −m.
Case 2: H1 (∂m ; M )(−1)j 6= 0 for some j ≤ −m + 1.
By induction hypothesis it follows that H1 (∂m ; M )(−1)j 6= 0 for all j ≤ −m + 1.
Hence H1 (∂m ; M )j 6= 0 for all j ≤ −m. So by exact sequence (7.1.8) it follows that
Mj+1 6= 0 for all j ≤ −m which implies Mj 6= 0 for all j ≤ −m + 1. Thus Mj 6= 0
for all j ≤ −m.
Case 3: Hl (∂m ; M )(−1)j = 0 for l = 0, 1 and for ALL j ≤ −m + 1.
16
TONY J. PUTHENPURAKAL AND SUDESHNA ROY
By exact sequence (7.1.8) it follows that Mj ∼
= M−m+1 for all j ≤ −m+1. Hence
Mj ∼
= Ms 6= 0 for all j ≤ −m.
(II) Clearly (c) =⇒ (b) =⇒ (a) =⇒ (d). We only have to prove (d) =⇒ (c).
To do this we use induction on m.
We first assume m = 1. Consider the exact sequence
(7.1.9)
X
1
Mj → H0 (X1 ; M )j → 0.
0 → H1 (X1 ; M )j → Mj+1 −→
Since M is a generalized Eulerian Am (A)-module so by Proposition 3.8, we get that
Hl (X1 , M ) is concentrated in degree 0 for l = 0, 1. Thus by the exact sequence
(7.1.9) it follows that
Mj ∼
= M0 for all j ≥ 0.
∼
Hence Mj = Mt 6= 0 for all j ≥ 0.
We now assume that m ≥ 2 and the result is proved for m − 1. Consider the
exact sequence
(7.1.10)
X
m
Mj → H0 (Xm ; M )j → 0.
0 → H1 (Xm ; M )j → Mj−1 −→
By Proposition 3.6 we have Hl (Xm ; M ) is a generalized Eulerian Am−1 (A)-module.
We consider the following three cases:
Case 1: H0 (Xm ; M )j 6= 0 for some j ≥ 0.
By induction hypothesis we get H0 (Xm ; M )j 6= 0 for all j ≥ 0. So by exact sequence
(7.1.10) it follows that Mj 6= 0 for all j ≥ 0.
Case 2: H1 (Xm ; M )j 6= 0 for some j ≥ 0.
By induction hypothesis We get H1 (Xm ; M )j 6= 0 for all j ≥ 0. So by exact
sequence (7.1.10) it follows that Mj−1 6= 0 for all j ≥ 0, and hence Mj 6= 0 for all
j ≥ −1. Thus the result follows.
Case 3: Hl (Xm ; M )j = 0 for l = 0, 1 and for ALL j ≥ 0.
By exact sequence (7.1.10) we get Mj ∼
= Mt 6= 0
= M−1 for all j ≥ 0. Hence Mj ∼
for all j ≥ 0.
Corollary 7.2. Let A be a commutative Noetherian ring containing a field of characteristic zero. Let R = A[X1 , . . . , Xm ] be standard graded with deg A = 0 and
∗
deg Xi = 1 for
Lall i. Let T be a graded Lyubeznik functor on Mod(R) and set
M = T (R) = n∈Z Mn . Then
(I) The following conditions are equivalent:
(a) Mn 6= 0 for infinitely many n < 0.
(b) There exists r such that Mn 6= 0 for all n ≤ r.
(c) Mn 6= 0 for all n ≤ −m.
(d) Ms 6= 0 for some s ≤ −m.
(II) The following conditions are equivalent:
(a) Mn 6= 0 for infinitely many n ≥ 0.
(b) There exists s such that Mn 6= 0 for all n ≥ s.
(c) Mn 6= 0 for all n ≥ 0.
(d) Mt 6= 0 for some t ≥ 0.
Proof. By [4, 3.2(2)] and [4, Theorem 3.6] we have T (R) is a graded generalized
Eulerian Am (A)-module. Therefore by Theorem 7.1, T (R) satisfies properties (I)
and (II).
We now prove the following crucial result.
GRADED COMPONENTS OF LOCAL COHOMOLOGY MODULES
17
L
Theorem 7.3. Let M = n∈Z Mn be a generalized Eulerian Am (A)-module. If
m ≥ 2, then the following assertions are equivalent:
(i) Mn 6= 0 for all n ∈ Z.
(ii) There exists r with −m < r < 0 such that Mr 6= 0.
Proof. (i) =⇒ (ii) is clear. We only have to prove (ii) =⇒ (i). We use induction
on m.
Let m = 2. Then we have M−1 6= 0.
Claim 1: Mi 6= 0 for infinitely many i ≤ 0.
Proof. Suppose if possible Claim 1 is false. Then by Theorem 7.1 we get Mi = 0
for all i ≤ −2. Now we have exact sequence
(7.3.11)
∂
2
Mi → H0 (∂2 ; M )i → 0.
0 → H1 (∂2 ; M )i → Mi+1 −→
Therefore Hl (∂2 ; M )i = 0 for all i ≤ −3 and l = 0, 1. Moreover, by Proposition
3.5 we have Hl (∂2 ; M )(−1) is a generalized Eulerian A1 (A)-module. Therefore
by Theorem 7.1 it follows that Hl (∂2 ; M )(−1)i = 0 for all i ≤ −1 and hence
Hl (∂2 ; M )i = 0 for i ≤ −2. From the exact sequence (7.3.11) for i = −2, we get
M−1 ∼
= M−2 = 0 which contradicts our hypothesis. So Claim 1 is correct.
Thus by Theorem 7.1, Mi 6= 0 for i ≤ −2.
Claim 2: Mi 6= 0 for infinitely many i ≥ 0.
Proof. Suppose if possible Claim 2 is false. Then by Theorem 7.1 we get Mi = 0
for all i ≥ 0. Consider the exact sequence
(7.3.12)
X
2
Mi → H0 (X2 , M )i → 0.
0 → H1 (X2 , M )i → Mi−1 −→
So H1 (X2 , M )i = 0 = H0 (X2 , M ) for all i ≥ 1. Moreover, by Proposition 3.6 we
have Hl (X2 , M ) is generalized Eulerian for l = 0, 1. Thus by Theorem 7.1 it follows
that Hl (X2 , M ) = 0 for all i ≥ 0. From the exact sequence (7.3.12) for j = 0, we
get
M−1 ∼
= M0 = 0.
This contradicts our hypothesis. So Claim 2 is true.
Thus by Theorem 7.1 it follows that Mi 6= 0 for all i ≥ 0.
Hence the result is true when m = 2.
We now assume m ≥ 3 and the result is known for m − 1.
We have Mr 6= 0 for some r with −m < r < 0. We want to show Mi 6= 0 for all
i ∈ Z.
Claim 3: Mi 6= 0 for infinitely many i ≤ 0.
Proof. Suppose if possible Claim 3 is false. Then by Theorem 7.1 we get Mi = 0
for all i ≤ −m. Consider the exact sequence
(7.3.13)
∂
m
Mi → H0 (∂m ; M )i → 0.
0 → H1 (∂m ; M )i → Mi+1 −→
Therefore Hl (∂m ; M )i = 0 for all i ≤ −m − 1 and for l = 0, 1. Moreover, by
Proposition 3.5 we have Hl (∂m ; M )(−1) is a generalized Eulerian Am−1 (A)-module.
Thus for l = 0, 1; by Theorem 7.1 it follows that Hl (∂m ; M )(−1)i = 0 for all
i ≤ −m + 1 and by induction hypothesis we get that Hl (∂m ; M )(−1)i = 0 for
−m + 1 < i ≤ −1. Hence Hl (∂m ; M )i = 0 for i ≤ −2. Now by exact sequence
(7.3.13) we get
M−1 ∼
= M−m .
= M−m+1 ∼
= ··· ∼
= M−2 ∼
18
TONY J. PUTHENPURAKAL AND SUDESHNA ROY
This implies M−m ∼
= Mr 6= 0, a contradiction to our assumption. So Claim 3 is
correct. Thus by Theorem 7.1 it follows that Mi 6= 0 for i ≤ −m.
Claim 4: Mi 6= 0 for infinitely many i ≥ 0.
Proof. Suppose if possible Claim 4 is false. Then by Theorem 7.1 it follows that
Mi = 0 for all i ≥ 0. Consider the exact sequence
(7.3.14)
X
m
Mi → H0 (Xm ; M )i → 0
0 → H1 (Xm ; M )i → Mi−1 −→
So H1 (Xm ; M )i = 0 = H0 (Xm ; M )i for all i ≥ 1. Moreover, by Proposition 3.6 we
have Hl (Xm ; M ) is generalized Eulerian for l = 0, 1. So for l = 0, 1 by Theorem 7.1
it follows that Hl (Xm ; M )i = 0 for all i ≥ 0 and by induction hypothesis we get
Hl (Xm ; M )i = 0 for −m + 1 < i < 0. Thus Hl (Xm ; M )i = 0 for i ≥ −m + 2. Now
by exact sequence (7.3.14) we get
M−m+1 ∼
= M0 .
= ··· ∼
= M−1 ∼
= M−m+2 ∼
This implies Mr ∼
= M0 = 0, a contradiction to our hypothesis. So Claim 4 is correct.
Thus by Theorem 7.1 it follows that Mi 6= 0 for i ≥ 0.
As m ≥ 3 we also have to prove that if c 6= r and −m < c < r then Mc 6= 0.
Suppose if possible Mc = 0. We will consider the following two cases:
Case 1: c < r.
By Proposition 3.6 we have H1 (Xm ; M ) is a generalized Eulerian Am−1 (A)-module.
Moreover, by exact sequence (7.3.14) we get H1 (Xm ; M )c+1 = 0. Notice −m + 1 <
c + 1 ≤ r < 0 (as −m < c < r). So by induction hypothesis H1 (Xm ; M )i = 0 for
−m + 1 < i < 0.
Again by exact sequence (7.3.14) we also get H0 (Xm ; M )c = 0. Moreover, by
Proposition 3.6 we have H0 (Xm ; M ) is a generalized Eulerian Am−1 (A)-module.
We will consider two sub-cases.
Sub-case 1.1: −m + 1 < c < r < 0.
By induction hypothesis we have H0 (Xm ; M )i = 0 for −m + 1 < i < 0. Thus by
exact sequence (7.3.14) we get,
M−1 ∼
= M−m+1 .
= ··· ∼
= M−m+2 ∼
= M−2 ∼
Hence Mr ∼
= Mc = 0, a contradiction to our hypothesis.
Sub-case 1.2: c = −m + 1.
We have H0 (Xm ; M )−m+1 = 0. Therefore by induction hypothesis we have
H0 (Xm ; M )i = 0 for − m + 1 < i < 0.
∼ Mc = 0, a contradiction
So by an argument similar to Sub-case 1.1, we get Mr =
to our hypothesis.
Thus our assumption is false. Hence Mc =
6 0 for −m + 1 < c < r.
Case 2: c > r.
By Proposition 3.5 we have H1 (∂m ; M )(−1) is a generalized Eulerian Am−1 (A)module. Moreover, by exact sequence (7.3.13) we get that H1 (∂m ; M )(−1)c =
H1 (∂m ; M )c−1 = 0. Notice −m + 1 ≤ r < c < 0 (as −m < r < 0 and −m < c < 0).
So by induction hypothesis H1 (∂m ; M )(−1)i = 0 for −m + 1 < i < 0. Thus
H1 (∂m ; M ) = 0 for −m < i < −1.
GRADED COMPONENTS OF LOCAL COHOMOLOGY MODULES
19
Again by exact sequence (7.3.13) we also get H0 (∂m ; M )c = 0. Moreover, by
Proposition 3.5 we have H0 (∂m ; M )(−1) is a generalized Eulerian Am−1 (A)-module.
We will consider two sub-cases:
Sub-case 2.1: c 6= −1.
We have −m + 2 < c + 1 < 0 and H0 (∂m ; M )(−1)c+1 = 0. So by induction
hypothesis we get H0 (∂m ; M )(−1)i = 0 for −m + 1 < i < 0. Thus H0 (∂m ; M )i = 0
for −m < i < −1. Therefore by exact sequence (7.3.13) we get
M−1 ∼
= M−m+1 .
= ··· ∼
= M−m+2 ∼
= M−2 ∼
Hence Mr ∼
= Mc = 0, a contradiction to our hypothesis.
Sub-case 2.2: c = −1.
Then we have H0 (∂m ; M )(−1)0 = 0. So by induction hypothesis H0 (∂m ; M )(−1)i =
0 for −m + 1 < i < 0 and hence H0 (∂m ; M )i = 0 for −m < i < −1. Therefore by
an argument similar to Sub-case 2.1, we get Mr ∼
= Mc = 0, a contradiction to our
hypothesis.
Thus our assumption is false. Hence Mc 6= 0 for r < c < 0. The result follows.
An important consequence of the Theorem 7.3 is the following.
Corollary 7.4. Let A be a commutative Noetherian ring containing a field of characteristic zero. Let R = A[X1 , . . . , Xm ] be standard graded with deg A = 0 and
∗
deg Xi = 1 for
Lall i. Let T be a graded Lyubeznik functor on Mod(R) and set
M = T (R) = n∈Z Mn . If m ≥ 2, then the following assertions are equivalent:
(i) Mn 6= 0 for all n ∈ Z.
(ii) There exists r with −m < r < 0 such that Mr 6= 0.
Proof. By [4, 3.2(2)] and [4, Theorem 3.6] we have T (R) is a graded generalized
Eulerian Am (A)-module. Therefore by Theorem 7.3, T (R) satisfies the equivalent
conditions.
The above result shows that non-vanishing of a single graded component of
M = T (R) will assure non-vanishing of its infinitely many graded components.
8. Some examples
Let A be a Noetherian ring containing a field K of characteristic zero. Let
R = A[X1 , . . . , Xm ] be standard
graded with m ≥ 1 and I be a homogeneous ideal
L
in R. Set M = HIi (R) = n∈Z Mn .
In this section we give some examples which show that some results of [4] are
false if A is not regular.
8.1. In [4, Theorem 1.7] it is shown that if A is a regular domain and Mn 6= 0 then
Mn is NOT finitely generated as an A-module. This is not true in general.
i
Example. Let (A, m) be a local domain with dimension d such that Hm
(A) is finitely
i
i
generated and non-zero for some i < d. Take I = mR. Then HI (R)0 = Hm
(A) is
non-zero and finitely generated as an A-module.
8.2. In [4, Theorem 1.8] it is shown that if A is a regular ring and P is a prime
ideal in A then
either
or
µj (P, Mn ) = ∞ for all n ∈ Z
µj (P, Mn ) < ∞ for all n ∈ Z.
20
TONY J. PUTHENPURAKAL AND SUDESHNA ROY
This does not hold in general.
Example. There exists example of Noetherian local ring (A, m) with HJi (A) having
infinite zeroth Bass number over m for some ideal J in A, see L
[11] (also see [3,
i
Chapter 20, Section 5]). Now M = HJR
(R) = HJi (A) ⊗A R = n∈N Mn . So in
this case µ0 (m, Mn ) = 0 for n < 0 (as Mn = 0) and µ0 (m, Mn ) is infinite for n ≥ 0.
8.3. Assume m = 1. In [4, Theorem 1.9] it is shown that if A is a regular ring and
P is a prime ideal in A then µj (P, Mn ) < ∞ for all n ∈ Z. This is not true if A is
not regular.
Example. In Example 8.2, take m = 1.
8.4. In [4, Theorem
S 1.13] it is shown that if A is regular local or a smooth affine
K-algebra then n∈Z AssA Mn is a finite set. This is false in general.
Example. There exists affine or local ring A such that Ass(HI2 (A)) is an infinite
set. For example, take K be an arbitrary field and consider the hypersurface
A = K[s, t, u, v, x, y]/(sv 2 x2 − (s + t)vxuy + tu2 y 2 ).
2
Then H(x,y)
(A) has infinitely many associated primes, see [3, Example 22.17]. Localization of A at the homogeneous maximal ideal (s, t, u, v, x, y) gives
L a local ex2
ample. Take I = (x, y)R. Since M = HI2 (R) = H(x,y)
(A) ⊗A R = n∈N Mn so in
both cases AssA M0 is infinite.
8.5. In [4, Theorem 1.14] it is shown that if A is regular then
injdim Mn ≤ dim Mn for all n ∈ Z.
The following example shows that this is not true in general.
Example. Let (A, m) be a Noetherian local ring containing a field with dim A > 0
0
0
0
and depth A = 0. Now M = HmR
(R) = Hm
(A) ⊗A R. Thus M0 = Hm
(A) and
hence M0 is a finitely generated A-module. If injdim M0 < ∞ then by [2, Corollary
9.6.2] we get A is Cohen-Macaulay, a contradiction. So injdim M0 is infinite.
References
[1] M. P. Brodmann and R. Y. Sharp, Local cohomology: an algebraic introduction with geometric
applications, Cambridge Studies in Advanced Mathematics 60, Cambridge University Press,
Cambridge, 2013.
[2] J.-E. Björk, Rings of Differential Operators, North-Holland Mathematical Library, Vol. 21,
North-Holland Publishing Co., Amsterdam, New York, 1979.
[3] S. B. Iyengar, G. J. Leuschke, A. Leykin, C. Miller, E. Miller, A. K. Singh and U. Walther,
Twenty Four Hours of Local Cohomology, Graduate Studies in Mathematics, Vol. 87, American Mathematical Society, 2011.
[4] T. J. Puthenpurakal,
Graded components of local cohomology modules, Preprint:
arXiv:1701.01270.
[5] T. J. Puthenpurakal, de Rham Cohomology of Local Cohomology Modules, Algebra and its
Applications, Springer Proceedings in Mathematics and Statistics, Vol. 174 (2016), 159-181,
Springer, Singapore; Editors: S. Rizvi, A. Ali, V. Filippis.
[6] T. J. Puthenpurakal, de Rham cohomology of local cohomology modules: The graded case,
Nagoya Math. J., Vol. 217 (2015), 1-21.
[7] T. J. Puthenpurakal, Associated primes of Local Cohomology Modules over Regular Rings,
Pacific Journal of Mathematics, Vol. 282 (2016), No. 1, 233-255.
[8] T. J. Puthenpurakal and J. Singh, On derived functors of Graded local cohomology modules,
Preprint: arXiv:1612.02968.
GRADED COMPONENTS OF LOCAL COHOMOLOGY MODULES
21
[9] L. Ma and W. Zhang, Eulerian graded D-modules, Math. Res. Lett., Vol. 21 (2014), No. 1,
149-167.
[10] G. Lyubeznik, Finiteness Properties of Local Cohomology Modules (an Application of Dmodules to Commutative Algebra), Inv. Math., Vol. 113 (1993), 41-55.
[11] R. Hartshorne, Affine duality and cofiniteness, Invent. Math., Vol. 9 (1969/1970), 145-164.
Department of Mathematics, Indian Institute of Technology Bombay, Powai, Mumbai
400 076, India
E-mail address: [email protected]
E-mail address: [email protected]
| 0math.AC
|
Dynamic Actuator Selection and Robust State-Feedback Control of
Networked Soft Actuators
arXiv:1804.01615v1 [cs.SY] 4 Apr 2018
Nafiseh Ebrahimi∗,∗∗ , Sebastian Nugroho† , Ahmad F. Taha†,‡ , Nikolaos Gatsis† , Wei Gao∗ , Amir Jafari∗,∗∗
Abstract—The design of robots that are light, soft, powerful
is a grand challenge. Since they can easily adapt to dynamic environments, soft robotic systems have the potential of changing
the status-quo of bulky robotics. A crucial component of soft
robotics is a soft actuator that is activated by external stimuli
to generate desired motions. Unfortunately, there is a lack of
powerful soft actuators that operate through lightweight power
sources. To that end, we recently designed a highly scalable,
flexible, biocompatible Electromagnetic Soft Actuator (ESA).
With ESAs, artificial muscles can be designed by integrating a
network of ESAs. The main research gap addressed in this work
is in the absence of system-theoretic understanding of the impact
of the realtime control and actuator selection algorithms on the
performance of networked soft-body actuators and ESAs. The
objective of this paper is to establish a framework that guides
the analysis and robust control of networked ESAs. A novel
ESA is described, and a configuration of soft actuator matrix
to resemble artificial muscle fiber is presented. A mathematical
model which depicts the physical network is derived, considering
the disturbances due to external forces and linearization errors
as an integral part of this model. Then, a robust control and
minimal actuator selection problem with logistic constraints and
control input bounds is formulated, and tractable computational
routines are proposed with numerical case studies.
I. I NTRODUCTION , B RIEF L ITERATURE R EVIEW, PAPER
C ONTRIBUTIONS AND O RGANIZATION
The emerging field of soft robotics represents the foundation of future robotic systems with plethora of applications
in human-robot interaction, locomotion, and rehabilitation
technologies [1]. A crucial component of soft robotics is a
soft actuator that is activated to generate desired motions.
Unfortunately, there is a lack of actuators in rehabilitation
applications that are portable, adoptable to different joint
sizes, while still matching the performance of the mammalian
muscles in terms of response time and output power-to-size
ratio. Consequently, the next generation of soft actuators
that achieve assistive functions raises challenges to actuation
design and analysis, materials engineering, modeling, and
realtime robust control and optimization—the main theme
of the proposed research. Motivated by these challenges we
have recently designed a bio-inspired highly scalable, flexible, biocompatible Electromagnetic Soft Actuator (ESA) [2]
which we explain next.
As shown in Fig. 1, a human skeletal muscle with bundles
of sarcomeres (composed by actin and myosin) behaves like
‡ Corresponding
† Department of
author. ∗ Department of Mechanical Engineering.
Electrical and Computer Engineering. ∗∗ Advanced
Robotic Manipulators (ARM) Lab. All authors are with the University of Texas at San Antonio, 1 UTSA Circle, San Antonio, TX
78249. Emails: {nafiseh.ebrahimi, ahmad.taha, amir.jafari, nikolaos.gatsis,
wei.gao}@utsa.edu, [email protected] material is based
upon work supported by the National Science Foundation under Grant
CMMI-DCSD-1728629.
Fig. 1. ExoMuscles: the future of artificial muscles.
a network of soft actuators. A network of ESA can be
integrated inside an artificial muscle (coined as ExoMuscle)
warped around the joints as active braces, mimicking a
human muscle. Notably, it is analytically and experimentally
confirmed that by scaling down the size of ESA, the ratio
between resulting force to cross-section area (F/CSA) of ESA
increases [2]. It is essential in obtaining large forces over a
relatively small space using ESAs.
Latest relevant studies focus either on the design of soft
material or the construction of soft actuators into artificial
muscles. The main research gap lies in the absence of
system-theoretic understanding of the impact of realtime
control algorithms on the performance of networked softbody actuators in general, and ESA networks in specific. The
objective of this paper is to establish a framework that guide
the design, analysis, and robust control of ESA networks
forming artificial muscles under uncertainty.
Series elastic and variable stiffness actuators (SEA/VSAs)
are implemented and designed because of their ability to
minimize large forces due to shocks, to safely interact with
the user, and their ability to store and release energy in
passive elastic elements [3]–[5]. However, they are made
of intrinsically rigid components. As a result, they are
bulky. Actuators based on shape memory alloys have highly
nonlinear behavior, low energy efficiency and low response
speed [6], despite their advantages that include high power
to weight ratio, mechanism simplicity, silent actuation, and
low driving voltage. Pneumatic artificial muscles [7], [8] have
been vastly used in robotics. However, they require stationary
power sources and accessories such as air pump and valves.
Dielectric elastomer actuators [9], [10] are popularly referred
to as artificial muscles because of their actuation speed, low
density, and silent operation. Unfortunately, they demand
high operating voltages—preventing their operation with onboard batteries.
Another dimension for successful operation of soft robotics
in general, and artificial muscles in specific, is the realtime
transient control problem which refers to the optimal current
injections for each actuator in an networked system—as well
as determining the subset of actuators to be activated given a
desired motion and maximum input current. To that end, the
problem of (a) simultaneously obtaining a minimal actuator
subset while (b) designing localized, robust control laws
for activated actuators is needed. Unfortunately, this routine
of actuator selection and robust control is known to be a
combinatorial, NP-hard problem. To address this challenge,
various quantitative notions of network controllability and
observability have been used [11]–[13] and different techniques that are based on heuristics or simple linear timeinvariant dynamics are presented in [14]–[16].
Little is known about the interplay between actuator selection and the robust control problems with disturbances when
maximum input bounds are imposed—for generic dynamic
systems as well as for networked soft-body actuators. In this
paper, we focus on addressing this gap which transcends
ExoMuscles, and can be applied to various soft robotics. The
paper contributions and organization are as follows.
• The novel ESA and its properties are described, and a
configuration of soft actuator matrix to resemble artificial
muscle fiber is presented (Section II). This is followed by
the derivation of a physical network model of soft actuators. Disturbances due to external forces and linearization
errors are considered as an essential part of the model
(Section III).
• A robust control and minimal actuator selection problem
formulation with actuator constraints and maximum input voltage bounds is investigated in Section IV. This
formulation can be written as a nonconvex optimization
problem with mixed-integer nonlinear matrix inequalities.
To address the computational difficulties in solving this
routine, tractable routines are explored in Section V and VI.
• Numerical tests are presented in Section VII showing
the performance of the relaxations. The results suggest
that network of ESAs can be robustly controlled with a
subset of available actuators, while still yielding reasonable
energy and stability performance.
II. S OFT ACTUATOR AND S YSTEM C ONFIGURATION
A. Electromagnetic Soft Actuator (ESA) Components
Fig. 2 shows the electromagnetic soft actuator we recently
designed [2] and its components. It is made of biocompatible
and magnetically permeable soft silicone that surrounds the
copper wire coils and a soft silicone-ferromagnetic core.
This actuator is made mostly of silicone rubber so that
it can have low stiffness. The major components of the soft
actuator include helical coil, soft silicone ferromagnetic core,
inner layer, spring linkage and outer layer. The helical coils
are made of 100 turn of 34 American wire gauge copper. The
outer layer of ESA which is included to shield and boost
the resultant magnetic field consists of a mixture of 40%
iron oxide and 60% silicone rubber. This part is included to
make a layer of electromagnetic suspension to strengthen the
generated field and increase the force.
Fig. 2. Intrinsically soft electromagnetic actuator composed of two wire
coils located on both sides of silicone spring linkage and a soft siliconeferromagnetic core inside the coils having a clearance fit respect to them.
B. Electromagnetic Soft Actuator Test Setup
Two helical coils are embedded in silicone rubber and
powered having a variable input between 0 and 12 Volts to
control the force and position of the soft actuator. The helical
coil provides the electromotive force of the soft actuator
by the reaction of magnetic field to the current passing
through it. The produced magnetic field causes the helical
coil to react to magnetic field from a permanent magnet
core fixed to the soft actuators frame, thereby provides
expansion/contraction to the ESA. The spring linkage is the
system that provides the elongation for the soft actuator as
the electromotive force is axially applied. Moreover, there
are some electrical components and circuitry including: main
power switch, power supply, voltage regulator, the Arduino
Mega Microcontroller, H-bridge and the soft actuator to
facilitate our design and test implementation. In order to
obtain the generated force by ESA, force measurements were
conducted at various voltages by means of a load cell. The
soft actuator was fixed horizontally so that it could push
against the load cell as voltage applied. The load cell is fixed
using some clamps so that it stayed stationary. The force
measurements were taken at various voltages and presented
in Fig. 3. The relationship between the voltage and force is
assumed to be linear for simplicity
F (t) ≈ 0.0146V (t) − 0.0088.
(1)
This experimental part in the paper is needed for any ESA
used to obtain the slop of the above curve. The verification
of the relationship between the input force and voltage is
essential in yielding a tractable computational framework
instead of a nonlinear model that would further complicate
the analysis and design of robust control methods with
disturbances. The next section presents an aggregate statespace model for networked ESAs that extends to networked
soft-body actuators.
III. S TATE -S PACE M ODELING FOR E XO M USCLES
The scalability of the actuator allows us to exert greater
force from a brace with specific dimensions by embedding
of spring linkage. Each actuator connected to the next one
through another spring (k2 ) and damper (c2 ). Fig. 4 depicts
the schematic mass-spring-damper model of a 2-by-4 matrix
of mass-spring-damper networks. In each column of the
matrix, the two masses of each actuator are attached to the
subsequent ones. This structure resembles the actual muscle
mechanism since there are columns of myosin and actin—
akin to our proposed configuration. The motion dynamics of
the ExoMuscle are
Force vs. Voltage
0.2
Force (N)
0.15
0.1
0.05
ẋ(t) = Ax(t) + Bu u(t) + Bd d(t),
0
0
2
4
6
8
10
12
Voltage (V)
Fig. 3. Electromagnetic soft actuator generated force due to applying
voltage to the coils caused interaction between the ferromagnetic core and
magnetized coils.
Fig. 4. Network with limited number of ESAs Forming ExoMuscles.
more smaller soft actuator cells in a constant volume of the
brace. Our proposed structure for the ExoMuscle is a matrix
of several ESA actuators. Every single actuator is generating
a partial force by means of applying voltage. As a result,
the total generated force of the array also is controllable
by the voltage. Our suggested model resembles the structure
and function of human muscle components. We consider the
actuator as building blocks in developing simulated artificial
muscle fiber. Fig. 1 illustrate that a human skeletal muscle is
composed of several longitudinal muscle fibers. Each muscle
fiber can then be further broken down into smaller myofibers,
which themselves are serial chains of sarcomeres. Each
sarcomere is composed of parallel arrangement of muscle
actuation units Actin and Myosin.
Following the same fashion, we first built the sarcomere
part of our model by arranging bunch of soft actuators in
parallel. Then, by connecting the modeled artificial sarcomeres in series and through silicone spring linkage which also
modeled by spring and damper, we constructed an array
representing artificial fibers. It is notable that the ExoMuscles
would be the combination of artificial fibers in different
arrangements and textures and would be used as active braces
for rehabilitation and assistive application. To achieve the
analytical model of the proposed exo-muscle, firstly, we
model each ESA with the typical system of mass, spring
damper as shown in Fig. 4. Each soft actuator can be modeled
by two masses (m) representing two conductive coils, one
spring (k1 ) and one damper (c1 ) simulating the behavior
(2)
where x(t) ∈ R2rc=nx is the state-vector collecting all
the deflection and velocities of each mass-spring subsystem;
u(t) ∈ Rrc=nu collects all the input forces (or current
injections) of all ESAs in the ExoMuscle; d(t) ∈ Rrc=nd
includes the disturbances from external forces; A, Bu and
Bd are a function of the parameters of the ExoMuscles, as
well as the configuration which we discussed in the previous
sections. In short, the ExoMuscle has nx states, nu control
inputs (N = nu /2 ESAs), and nd external disturbances that
are naturally matched by the control input. We give a concrete
example of a network of two series ESAs; the dynamics of
ESAs can be written as
mẍ1 (t) = −(k1 + k2 )x1 (t) − (c1 + c2 )ẋ1 (t)
+ k1 x2 (t) + c1 ẋ2 (t) + F1 (t) + F1ex (t)
(3a)
mẍ2 (t) = k1 x1 (t) − (k1 + k2 )x2 (t) + k2 x3 (t) + c1 ẋ1 (t)
− (c1 + c2 )ẋ2 (t) + c2 ẋ3 (t) + F1 (t) + F1ex (t)
(3b)
mẍ3 (t) = k2 x2 (t) − (k1 + k2 )x3 (t) + k1 x4 (t) + c2 ẋ2 (t)
− (c1 + c2 )ẋ3 (t) + c2 ẋ4 (t) + F2 (t) + F2ex (t)
(3c)
mẍ4 (t) = k1 x3 (t) − (k1 + k2 )x4 (t) + k2 x5 (t) + c1 ẋ3 (t)
− (c1 + c2 )ẋ4 (t) + c2 ẋ5 (t) + F2 (t) + F2ex (t)
(3d)
where k1 and c1 are the stiffness and damping coefficients
between two masses of one actuator and k2 and c2 are the
stiffness and damping coefficients between two ESAs; xi (t)
denote the deflection; F1,2 (t) are the generated forces via
applying voltage to the first and second coils of first actuator
ex
(that we wish to solve for); F1,2
(t) are the unknown, yet
matched external forces and disturbances.
The next section investigates novel methods that optimally
determine the combination of time-varying activated actuators in addition to the optimal input forces (input voltages)
for all actuators, while considering worst-case disturbances
and uncertainties due to the nature of the ESA networks.
It is noteworthy to mention that there are two sources of
disturbance in our model. Firstly, the approximation of actual
force-voltage graph by a straight line. Secondly, the constant
portion of (1) considered as a kind of disturbance inasmuch
as the fact that practically in the absence of applied voltage
there is no exerted force from soft actuator.
IV. ROBUST C ONTROL AND ACTUATOR S ELECTION
The previous sections focus on the modeling and the
design of output-force maximizing methods for networked
soft bodies and ExoMuscles, thereby achieving the high-level
objectives of soft robotics. The objective of the remaining
of the paper is two-fold. First, to investigate combinatorial
optimal control methods that guide the selection of specific
actuators for different time-periods under uncertainty and
actuator logistic constraints. Second, to simultaneously design local control laws for individual actuators while taking
into account bounds on the input energy (maximum input
current/voltage) and worst-case disturbance scenarios. The
objectives/challenges are coupled, and the theme of these
methods is a set of tractable computational methods that
explore solutions to the two challenges.
A. Robust Control Law and Actuator Selection
First, we define Γ ∈ Rnu ×nu to be a diagonal matrix with
binary variables Γi that select the optimal actuators, where
Γi = 1 if the ith actuator is selected and 0 otherwise. The
ExoMuscles dynamics derived in Section III with the timevarying actuator selection in time-period j can be written as
ẋ(t) = Ax(t) + Bu Γj u(t) + Bd d(t),
(4a)
p(t) = Cx(t) + Du(t).
(4b)
The performance index p(t) is a vector that quantifies the
quantities to be minimized where C and D reflect the
weights—akin to the linear quadratic regulator cost function.
The need for designing a robust control law for various time
periods is due to: (a) the need to deactivate actuators to
avoiding overheating, (b) the possibility of fatigue which impacts the performance of actuators, and (c) potential damages
to actuators which means that these have to be deactivated
permanently. Other logistic constraints also justify the need
for a time-varying actuator selection with robust control. We
solve for variable Γj via this framework which is reflected
via the following optimization routine.
minimize
j
j
K ,Γ
subject to
T
X
kp(t)k∞
j=1
kd(t)k∞
+ αΓ trace(Γj )
(5a)
(4), u(t) = K j x(t), ku(t)k ≤ umax (5b)
Γji ∈ {0, 1}N , Γj ∈ Aj .
(5c)
Problem (5) minimizes the impact of the worst-case disturbance kd(t)k∞ on the performance index, in addition
to the weighted number of activated actuators (weight is
αΓ ). The constraints are the uncertain physics-based dynamics, the state-feedback controller through time-varying
gain K j which is fortunately possible through soft sensors
and dynamic state estimation methods, input-voltage budget
constraints umax , binary constraints on the actuator selection,
and logistic constraints Γj ∈ Aj as described above. We
do not elaborate on soft sensors and the observability of
the dynamic system here, but we assume that the states are
available. Since soft sensors cannot measure all the states
of the ExoMuscle, a simple state estimator can be used to
obtain x̂(t), the estimate of x(t). The input constraints can be
considered on the maximum available input voltage, whereas
the logistic constraints represent simple actuator-based rules
for various time-periods. The optimization problem (5) is
a very challenging combinatorial, non-convex optimization
problem.
B. Problem Formulation
Using the notion of L∞ -stability of LTI systems under
disturbances from [17], we can show that Problem (5) can
be written as an optimization problem with mixed-integer
nonlinear matrix inequalities (MI-NMI). Assuming that the
initial conditions are x0 and that kd(t)k∞ ≤ ρ, then, if there
exist real matrices P = P > 0 and Y , binary diagonal
matrix Γ, and scalars {µ0 , µ1 , µ2 } > 0 for all time-periods
that are the solution to optimization problem (6) (given in
the next page) with optimal value f ∗ , then the time-varying,
adaptive feedback controller u(t) = K j x(t) with K j =
−Y j (P j )−1 guarantees that
q
j
j j
µ0 µ1 + µ2 , ∀j
kp(t)k2 ≤ µρ, µ = max
and that the closed loop system with unknown inputs with
the selected actuators is L∞ -stable with performance level∗ µ
with minimal number of actuators. We have recently analytically proved that (6) is in fact an accurate representation
of (5); the proof is not presented in here, and left for
an extended version of this manuscript† . This formulation
implies that regardless of the disturbance that the system
is subject to, then the performance index remains within a
neighborhood of the origin (or the final desired state) with
the the best performance index µ.
Unfortunately, the optimization routine (6) is challenging
to solve due to the MI-NMIs. Section V outlines efficient
computational methods to bound the optimal value f ∗ of (6).
V. M ETHODS TO B OUND THE S OLUTION OF (6)
This section is dedicated to develop numerical computing methods that, in general, solve nonconvex optimization
problems with MI-NMIs and in specific, robust control with
actuator selection for uncertain soft-body dynamics (4).
A. Approach 1: Relaxing The Integer Constraints
The first approach is based on relaxing the integer constraints Γ ∈ {0, 1} to Γ ∈ [0, 1]. This approach is common in
various problems where integer variables appear in nonlinear
and linear optimization [18]–[20]. Later in this section, we
discuss a simple approach to recover the integer variables
from the real, continuous solutions. Given this relaxation,
and after removing the dependence on time-period j simplicity, Problem (6) becomes nonconvex problem with bilinear
matrix inequalities (BMI) that can be written as
L∗ = minimize µ0 µ1 + µ2 + αΓ trace(Γ)
subject to (6b), (6c), Γ ∈ [0, 1], Γ ∈ A,
(7a)
(7b)
where L∗ ≤ f ∗ due to the integer relaxation and hence
the expansion of the feasible space. BMIs appear in various
∗ The reader is referred to [17] for more the definition of L
∞ stability,
but in short this definition entails that the states of the linear system under
disturbances are guaranteed to be bounded in a tube centered at the origin
of radius µ.
† The interested reader can contact the corresponding author for more
information related to the derivation.
f∗ =
T
X
minimize
{P ,Y,µ0 ,µ1 ,µ2
,α}j
µj0 µj1 + µj2 + αΓ trace Γj
P j A> + AP j + αP j
−Y j> Γj> Bu> − Bu Γj Y j
Bd>
j 2
−µ0 ρ
x0
−µj1 P j
Bd
0,
O
−αµj0 I
CP j + DY j
" 2
u
j
− max
x>
0
ρ2 P
0,
j
j
j
−P
µ0 Y
optimal control and state estimation problems in dynamic
systems [21], [22]. Various methods have investigated upper
bounds to optimization problems with BMIs; see [23]. However, these methods have a widely acknowledged limitation:
they offer no intuition or direction on recovering the optimal
solution (unless costly, inefficient global optimization is used)
to nonconvex problems with BMIs, that is L∗ in (6).
B. Upper Bound on L∗
(6a)
j=1
subject to
O
−µj2 I
O
P j C > + Y j> D >
0
O
−I
(6b)
#
µ0 Y j>
0, Γji ∈ {0, 1}N , Γj ∈ Aj , j = 1, . . . , T.
−µj0 I
(6c)
approach can be applied to the other bilinearities and the
objective function, leading to the approximation of (7) to a
convex problem
L∗u = min Co (µ0 , µ1 , µ2 ) + αΓ trace(Γ)
(9a)
s.t. C(P , Γ, Y , µ0 , µ1 ), Γi ∈ [0, 1], Γ ∈ A, (9b)
where C(·) collects all the convex approximations of the
bilinear terms of the constraints in (7), while Co (·) convexifies
the nonconvex terms in the objective function of (7). The
explicit form of (7) with all the linear matrix inequalities
and constraints is not provided in this paper due to the
lack of space. However, following the convex approximation
of the first constraint, other approximations can similarly
be derived. This problem is solved iteratively, given the
linearization points, until a certain stopping criteria is satisfied as described in Algorithm 1—similar to all SCAs.
It is noteworthy to mention that such approximations yield
desirable convergence properties as discussed in our recent
work [24].
Here, we present an approach based on successive convex
approximations (SCA) to obtain an upper bound L∗u on
L∗ . Before discussing the SCA, we note that (6) has the
following BMI terms: −Bu ΓY − Y > Γ> Bu> , µ1 P , µ0 Y ,
in addition to this nonconvex bilinear part of the objective
function µ0 µ1 . We show how the SCA works for the most
important bilinearity −Bu ΓY − Y > Γ> Bu> involving the
actuator variables, and leave the other BMIs for brevity.
First, notice that the aforementioned bilinear term can be
written as
>
1
−Bu ΓY − Y > ΓBu> =
Bu Γ − Y > Bu Γ − Y > − Algorithm 1 Solving SCA for (9).
2|
{z
}
initialize the bilinear terms, k = 0
P(Γ,Y )
input:
MaxIter, tol
>
1
while k < MaxIter do
Bu Γ + Y > Bu Γ + Y > .
2|
{z
}
Solve (9)
H(Γ,Y )
if |L∗uk − L∗uk−1 | < tol then
The term P(Γ, Y ) is convex in Y and Γ, while H(Γ, Y )
break
is concave in Y and Γ, and hence its first-order Taylor
else
approximation is a global over-estimator. Let Γ0 , Y0 be
k ←k+1
the linearization point, and let Hl (Γ, Y ; Γ0 , Y0 ) denote the
end if
linearization of H(Γ, Y ) at the point (Γ0 , Y0 ). It holds that
end while
H(Γ, Y ) Hl (Γ, Y ; Γ0 , Y0 ) for all Γ0 , Y0 and Γ, Y . The
initialization point can be chosen as Γ0 = Inu (that is, all C. Recovering the Integer Variables
ESAs are activated). The bilinear terms can now be upper
The solutions above generate real, continuous values for
bounded and approximated via
the actuator selection decision variable Γi ∈ [0, 1]. To obtain
the binary selection, the real actuator selection from the
AP + P A> + αP
optimal solution of (9) and Algorithm 1 can be ranked in
−Bu ΓY − Y > ΓBu>
Bd
>
decreasing order. Then, actuators are added up until a closed
Bd
−αµ0 I
loop system metric is satisfied. Alternatively, all actuators
AP + P A> + αP
with Γreal
≥ 0.5 can be activated. This approach yields
i
+0.5(Hl + P)
Bd = C1 (P , Γ, Y , µ0 ). (8)
effective
results
in highly nonconvex mixed-integer nonlinear
Bd>
−αµ0 I
programs [25], [26]. Any solution that we obtain from this
Notice that the RHS of (8) is indeed an LMI in terms of approach yields an upper bound U ∗ on f ∗ since the selection
P , Γ, Y , and µ0 . Hence C1 (·) is a convex constraint. This will be feasible for the original problem (6). Therefore,
Approach 1 yields tight lower and upper bounds on the
optimal solution of (6): L∗ ≤ f ∗ ≤ U ∗ .
VI. A PPROACH 2: K EEPING THE I NTEGER VARIABLES
As a departure from relaxing the integer constraints, we
investigate methods that transform Problem (6) to a mixedinteger semidefinite program (MI-SDP). This is useful as
the recent studies [26]–[28] have investigated and developed
efficient, open-source solvers to solve MI-SDPs. To see how
that can be implemented, we examine the major mixedinteger nonlinear term Ξ = ΓY . Due to the binary and
diagonal nature of Γ, we can write Ξi,(a,b) = Γi Zi,(a,b) if
Γi = 1 and Ξi,(a,b) = 0 if Γi = 0 for i = 1, . . . , N, a = 1, 2
(each actuator has two control inputs) and b = 1, . . . , nx ,
This relationship can be equivalently written as
|Ξi,(a,b) − Zi,(a,b) | ≤ M (1 − Γi ), |Ξi,(a,b) | ≤ M Γi
(10)
M is a sufficiently large positive constant; matrix Ξ replaces
ΓY in the first matrix inequality (6b) as follows
Bd
AP + P A> + αP − Bu Ξ − Ξ> Bu>
0.
Bd>
−αµ0 I
(11)
The overall optimization problem using the Big-M method
can be written as
min
T
X
µj0 µj1 + µj2 + αΓ trace(Γj )
(12a)
j=1
−µj1 P j
s.t.
O
CP j + DY j
O
−µj2 I
O
P j C > + Y j> D >
0
O
−I
(12b)
(6c), (10), (11).
(12c)
Considering that α, µ0 , µ1 are all predetermined constants
(not variables), then (12) is in fact a MI-SDP with
Γ, µ2 , Y , Ξ, and P as the optimization variables. MI-SDPs
can be solved via either off-the-shelf and industry-grade
solvers or recently developed solver in [27]. Many of these
implementations use branch-and-bound to solve MI-SDPs.
The user can also choose α, µ0 , µ1 to be variables, but that
requires then performing the SCA with the MI-SDP to obtain
solutions.
VII. N UMERICAL T ESTS
In this numerical tests, we demonstrate the applicability of
the proposed method to the networked of artificial muscles
consisting of 8 actuators, where each actuator has 6 rows
of mass-spring systems. With this particular configuration,
>
>
the state vector can be express as x(t) = x>
p (t) xv (t)
where the vectors xp (t) ∈ R16 and xv (t) ∈ R16 respectively collect the states of position and velocity of the
whole system. The state space of the system is of the form
ẋ(t) = Ax(t) + Bu u(t) + Bd d(t), p(t) = Cx(t) + Du(t)
where A depicts the network of ESAs described in (3)
with parameters m = 2.94 × 10−3 kg, k = 0.343 N/m,
c = 1.75×10−16 N · s/m . The mass of each coil is obtained
by measurement. Stiffness of the silicone rubber linkage is
achievable according to its 100% Modulus, average cross
section and length using Hook’s Law. Damping coefficient
is assumed to be very small [29].
The remaining state-space matrices are specified as follows
0.01 × I8
C = 0.1 × I16 O16×16 , D =
.
O8×8
A sinusoidal signal is chosen to simulate the disturbance,
which is represented as d(t) = cos(0.1t) with kd(t)k∞ =
2.8284. The numerical tests are performed by assuming zero
initial conditions within a timespan of tspan = [0, 60] sec. For
simplicity, we solve the problem in one time period only, that
is j = 1 and leave the multi-time period numerical tests for
extensions of this work. All numerical tests are performed
using MATLAB R2017b running on a 64-bit Windows 10
with 2.5GHz Intel Core i7-6500U CPU and 8 GB of RAM.
The details of the numerical tests are grouped into three
following sections: (Scenario A) We show the results for
the L∞ control problem without actuator selection, to test
the robustness of the system with a full set of 8 actuators;
(Scenario B) We solve the SCA followed by the slicing
algorithm discussed in Section V; (Scenario C) We solve
the formulation based on MI-SDPs shown in Section VI.
Scenarios B and C return a specific feedback gain matrix
K ∗ = Y ∗ (P −1 )∗ and Γ∗ as the computed actuator selection.
YALMIP [30] and MOSEK [31] are used to solve all of the
optimization problems.
A. Results for Different Optimization Methods
We do not list all of the specific details and parameters
in the implementation of the SCAs. However, all the codes
used to run the numerical tests are available upon request by
contacting the corresponding author.
1) Scenario A: Here, we put our interest to find the
best performance index of the system with full set of 8
actuators. Since the problem involves BMIs, we solve the
SCA formulation of (6) by setting Γ = I. The initialization
of SCA is performed by setting umax = 250 (which is much
larger that what controllers require), α = 10−3 , µ0 = 30,
and µ1 = 4. Since the SCA requires a starting point that
lies inside the relative interior of the feasible set, then we
define a constant 1 = 10−4 such that all the LMIs in (6)
are −1 I, which is common in these studies [23], [24].
For the SCA iterations, we follow Algorithm 1 with stopping
criteria specified by maximum number of iterations, that is set
to be 50, and convergence of the objective function defined
by |L∗uk − L∗uk−1 | ≤ 10−2 .
Table I illustrates the values achieved in this scenario
√
for α, µ = µ0 µ1 + µ2 which is the performance level
of the
Pclosed loop system, the computational time ∆t(sec),
and i Γi (the total number of activated actuators). Fig. 5a
depicts the system performance in terms of the norm of the
disturbance signal d(t), the state x(t), and the performance
vector p(t).
TABLE I
N UMERICAL RESULTS FOR THE THREE SCENARIOS : (A) F ULLY
ACTUATED SYSTEM , (B) M INIMALLY ACTUATED SYSTEM VIA CONVEX
PROGRAMMING , (C) M INIMALLY ACTUATED SYSTEM VIA THE MI-SDP.
Scenario
(A)
(B)
(C)
α
0.169
0.104
0.100
µ
0.44
0.62
0.63
P
i
8
3
2
Γi
∆t(s)
58.3
189.9
74.4
Diag(Γ)
{1, 1, 1, 1, 1, 1, 1, 1}
{0, 0, 0, 0, 1, 0, 1, 1}
{0, 0, 0, 1, 0, 0, 0, 1}
2) Scenario B: In this scenario, we use the same parameters as those in Scenario A. First, an optimal solution is
computed from (9) and Algorithm 1. Next, based on the
output of Algorithm 1, an integer solution is computed via
a simple slicing algorithm as discussed in Section V-C. The
numerical results are shown in Scenario B of Table I, whereas
the comparison between performance index and disturbance
is shown in Fig. 5-b.
3) Scenario C: In this particular test, the user can define
α, µ0 , µ1 to be variables, but that requires then performing
the SCA with the MI-SDP to obtain solutions. Alternatively,
we can predefine these variables to be constants. To that end,
we use values similar to the solutions of Scenario B. The
values are α = 0.1 , µ0 = 1, and µ1 = 0.4, with big-M
constant M = 103 ; see (10). The MI-SDP is solved through
YALMIP’s branch-and-bound algorithm. The results of this
test are shown in Table I and Fig. 5-c.
B. Discussion and Insights
The following preliminary conclusions are drawn from the
experiments.
• Scenario A uses all actuators. Consequently, it should
provide the user with the best performance index µ in
comparison with the other scenarios. This is corroborated
by the numerical experiments; see Table I.
• In Scenario B, a suboptimal solution of the actuator selection is given. In this scenario, the system only needs
3 actuators to be activated with performance index µ
comparable to the one in Scenario A. The selection of
actuator can be seen in Scenario B of Table I. This
illustrates an interesting observation: similar performance
index µ can be obtained with fewer number of actuators.
• In Scenario C, the MI-SDP approach gives the least number
of active actuators with computational time similar to that
of Scenario A and performance index slightly worse than
that of Scenario B; see ∆t(s) and µ in Table I. The specific
actuator selection of actuator is shown in Table I.
• The comparison of the input norm ||u(t)|| which measures
the energy of the control input (the voltage) is depicted in
Fig. 6 for the three scenarios. The results illustrate that the
one with fewest number of actuators (Scenario C) requires
much bigger input energy during the first 20 seconds in
comparison with Scenario A or B that utilizes more active
actuators. The control input norm for Scenarios A and B
are comparable, although the latter tends to be bigger due
to the utilization of fewer actuators. This presents another
interesting observation: fewer actuators can be activated
while not requiring much more input energy and also
maintaining a comparable performance index.
•
Fig. 7 illustrates the trajectory of the first state x1 (t) for all
scenarios. We can see that the damping of the oscillation
of x1 (t) for Scenario C is better than those of Scenario
A or B. This happens because Scenario C uses more
input energy to control the states, compared with the other
scenarios.
VIII. S UMMARY AND C ONCLUSIONS
The objective of this paper is to establish a framework
that guides the design, analysis, and robust control of softbody and ESA networks under uncertainty. A novel ESA
is described, and a configuration of soft actuator matrix to
resemble artificial muscle fiber is presented. A mathematical model which depicts the physical network is derived,
considering the disturbances due to external forces and
linearization errors as an integral part of this model. Combinatorial optimal control problems for uncertain dynamic
ExoMuscles with actuator selection are formulated. To address the computational complexity, efficient computational
routines are formulated based on convex programming and
mixed-integer convex programming. Numerical tests show a
promising simulated performance. Future work will focus on
implementing the controller on an actual ExoMuscle testbed.
One of the limitations of the proposed networked ESAs is
their resolution, which is highly dependent to the number
of actuation units embedded in the system. Compared to
the real muscle composed of tremendous number of muscle
actuation units, ExoMuscles are considered discrete and the
resolution depends on the number and consequently the size
of actuators. To achieve higher resolution we need to take
advantage of high precision technologies such as accurate
3D printers.
R EFERENCES
[1] C. Majidi, “Soft robotics: A perspectivecurrent trends and prospects
for the future,” Soft Robotics, vol. 1, no. 1, pp. 5–11, 2013. [Online].
Available: https://doi.org/10.1089/soro.2013.0001
[2] M. Harrison and A. Jafari, “A novel highly scalable soft electromagnetic actuator,” Soft Robotics, Under Review.
[3] V. Grosu, C. Rodriguez-Guerrero, S. Grosu, B. Vanderborght, and
D. Lefeber, “Design of smart modular variable stiffness actuators for
robotic assistive devices,” IEEE/ASME Transactions on Mechatronics,
vol. PP, no. 99, pp. 1–1, 2017.
[4] A. Jafari, N. G. Tsagarakis, I. Sardellitti, and D. G. Caldwell, “A
new actuator with adjustable stiffness based on a variable ratio lever
mechanism,” IEEE/ASME Transactions on Mechatronics, vol. 19,
no. 1, pp. 55–63, 2014.
[5] Y. Liu, Y. Zhang, and Q. Xu, “Design and control of a novel compliant constant-force gripper based on buckled fixed-guided beams,”
IEEE/ASME Transactions on Mechatronics, vol. 22, no. 1, pp. 476–
486, 2017.
[6] A. Borboni, R. Faglia, and M. Palpacelli, “Shape memory actuator with
slider and slot layout and single fan cooling,” in 2014 IEEE/ASME 10th
International Conference on Mechatronic and Embedded Systems and
Applications (MESA), Conference Proceedings, pp. 1–6.
[7] Y. Dang, L. K. Cheng, M. Stommel, and W. Xu, “Technical requirements and conceptualization of a soft pneumatic actuator inspired by
human gastric motility,” in 2016 23rd International Conference on
Mechatronics and Machine Vision in Practice (M2VIP), Conference
Proceedings, pp. 1–6.
[8] Y. L. Park and R. J. Wood, “Smart pneumatic artificial muscle actuator with embedded microfluidic sensing,” in 2013 IEEE SENSORS,
Conference Proceedings, pp. 1–4.
[9] H. Godaba, J. Li, Y. Wang, and J. Zhu, “A soft jellyfish robot driven
by a dielectric elastomer actuator,” IEEE Robotics and Automation
Letters, vol. 1, no. 2, pp. 624–631, 2016.
3
3
3
2.5
2.5
2.5
2
2
2
1.5
1.5
1.5
1
1
1
0.5
0.5
0.5
0
0
0
10
20
30
40
50
60
0
0
10
20
(a)
30
(b)
40
50
60
0
10
20
30
40
50
60
(c)
Fig. 5. Norm of x(t), p(t), and d(t) for the (a) fully-actuated system (Scenario A), (b) the SCA Scenario B, and (c) the MI-SDP Scenario C.
4.5
4
3.5
3
2.5
2
1.5
1
0.5
0
0
10
20
30
40
50
60
Fig. 6. Comparison of the norm of u(t) for Scenarios A, B, and C.
0.015
0.01
0.005
0
-0.005
-0.01
-0.015
0
10
20
30
40
50
60
Fig. 7. Comparison of the trajectory of x1 (t) for Scenarios A, B, and C.
[10] J. Zou, G. Y. Gu, and L. M. Zhu, “Open-loop control of creep
and vibration in dielectric elastomer actuators with phenomenological
models,” IEEE/ASME Transactions on Mechatronics, vol. 22, no. 1,
pp. 51–58, 2017.
[11] J. Ruths and D. Ruths, “Control profiles of complex networks,”
Science, vol. 343, no. 6177, pp. 1373–1376, 2014.
[12] S. Pequito, S. Kar, and A. Aguiar, “A framework for structural
input/output and control configuration selection in large-scale systems,”
IEEE Transactions on Automatic Control, vol. 61, no. 2, pp. 303–318,
2016.
[13] S. Pequito, G. Ramos, S. Kar, A. P. Aguiar, and J. Ramos, “The robust
minimal controllability problem,” Automatica, vol. 82, pp. 261–268,
2017.
[14] T. Summers, F. Cortesi, and J. Lygeros, “On submodularity and
controllability in complex dynamical networks,” IEEE Transactions
on Control of Network Systems, vol. 3, no. 1, pp. 91–101, 2016.
[15] N. K. Dhingra, M. R. Jovanović, and Z.-Q. Luo, “An admm algorithm
for optimal sensor and actuator selection,” in IEEE Conference on
Decision and Control. IEEE, 2014, pp. 4039–4044.
[16] P. V. Chanekar, N. Chopra, and S. Azarm, “Optimal actuator placement
for linear systems with limited number of actuators,” in 2017 American
Control Conference (ACC), May 2017, pp. 334–339.
[17] T. A. Pancake, “Analysis and control of uncertain/nonlinear systems in
the presence of bounded disturbance inputs,” Ph.D. dissertation, Purdue
University, 2000.
[18] M. A. Duran and I. E. Grossmann, “An outer-approximation algorithm
for a class of mixed-integer nonlinear programs,” Mathematical programming, vol. 36, no. 3, pp. 307–339, 1986.
[19] M. R. Bussieck and A. Pruessner, “Mixed-integer nonlinear programming,” SIAG/OPT Newsletter: Views & News, vol. 14, no. 1, pp. 19–22,
2003.
[20] S. Boyd and L. Vandenberghe, Convex optimization.
Cambridge
university press, 2004.
[21] S. Boyd, L. El Ghaoui, E. Feron, and V. Balakrishnan, “Linear matrix
inequalities in system and control theory,” 1994.
[22] J. G. VanAntwerp and R. D. Braatz, “A tutorial on linear and bilinear
matrix inequalities,” Journal of process control, vol. 10, no. 4, pp.
363–385, 2000.
[23] Q. T. Dinh, S. Gumussoy, W. Michiels, and M. Diehl, “Combining
convex–concave decompositions and linearization approaches for solving bmis, with application to static output feedback,” IEEE Transactions on Automatic Control, vol. 57, no. 6, pp. 1377–1390, 2012.
[24] A. F. Taha, N. Gatsis, T. Summers, and S. Nugroho, “Time-varying
sensor and actuator selection for uncertain cyber-physical systems,”
arXiv preprint arXiv:1708.07912, 2017.
[25] P. Belotti, C. Kirches, S. Leyffer, J. Linderoth, J. Luedtke, and
A. Mahajan, “Mixed-integer nonlinear optimization,” Acta Numerica,
vol. 22, pp. 1–131, 2013.
[26] J. Lee and S. Leyffer, Mixed integer nonlinear programming. Springer
Science & Business Media, 2011, vol. 154.
[27] T. Gally, M. E. Pfetsch, and S. Ulbrich, “A framework for solving
mixed-integer semidefinite programs,” Optimization Methods and Software, pp. 1–39, 2017.
[28] M. Tawarmalani and N. V. Sahinidis, “Global optimization of mixedinteger nonlinear programs: A theoretical and computational study,”
Mathematical programming, vol. 99, no. 3, pp. 563–591, 2004.
[29] “Ecoflex series,” https://www.smooth-on.com/tb/files/ECOFLEX
SERIES TB.pdf, accessed: 09-14-2017.
[30] J. Lofberg, “Yalmip: A toolbox for modeling and optimization in
matlab,” in Computer Aided Control Systems Design, 2004 IEEE
International Symposium on. IEEE, 2004, pp. 284–289.
[31] A. Mosek, “The mosek optimization software,” Online at http://www.
mosek. com, vol. 54, pp. 2–1, 2010.
| 3cs.SY
|
Skyline Computation with Noisy Comparisons
Frederik Mallmann-Trenn1,2 , Claire Mathieu1,3 , and Victor Verdugo1,4
arXiv:1710.02058v1 [cs.DS] 5 Oct 2017
1
École normale supérieure, CNRS, PSL Research University, Paris, France,
{mallmann, cmathieu, victor.verdugo}@di.ens.fr
2
Simon Fraser University
3
Collège de France; work partly done while visiting IRIF (Université Paris-Diderot)
4
Universidad de Chile
October 6, 2017
Abstract
Given a set of n points in a d-dimensional space, we seek to compute the skyline, i.e.,
those points that are not strictly dominated by any other point, using few comparisons
between elements. We study the crowdsourcing-inspired setting ([FRPU94]) where
comparisons fail with constant probability. In this model, Groz & Milo [GM15] show
three bounds on the query complexity for the skyline problem.
We provide two output-sensitive algorithms computing the skyline with query
complexity O(nd log(dk)) and O(ndk log(k)), where k is the size of the skyline. These
results improve significantly on the state-of-the-art and are tight for low dimensions.
1
Introduction
Skylines have been studied extensively, since the 1960s in statistics [BS66], then in algorithms
and computational geometry [KLP75] and in databases [BKS01; CCM13; GSG07; KRR02].
Depending on the field of research, the skyline is also known as the set of maximum vectors,
the dominance frontier, admissible points, or Pareto frontier. The skyline of a set of points
consists of those points which are not strictly dominated by any other point. A point p is
dominated by another point q if pi ≤ qi for every coordinate (attribute or dimension) i. It
is strictly dominated if in addition the inequality is strict for at least one coordinate; see
Figure 1 (from [GM15]) for an example.
In many contexts, comparing attributes is not straightforward. Consider the example of
finding optimal cities from [GM15].
To compute the skyline with the help of the crowd we can ask people questions
of the form “is the education system superior in city x or city y?” or “can I
1
salary
skyline point
dominated point
education quality
Figure 1: Given a set of points X, the goal is to find the set of skyline points, i.e.,points are
not dominated by any other points.
expect a better salary in city x or city y”. Of course, people are likely to make
mistakes, and so each question is typically posed to multiple people. Our objective
is to minimize the number of questions that need to be issued to the crowd, while
returning the correct skyline with high probability.
Thus, much attention has recently been given to computing the skyline when information
about the underlying data is uncertain [MWK+11], and comparisons may give erroneous
answers. One may consider that the location of each point is determined by a probability
distribution over a set of locations, or that data is incomplete [KML08; LEB13]. Some
previous work [PJLY07; AAA+11] model uncertainty about the output by computing a
ρ-skyline: points having probability at least ρ to be in the skyline. In this paper, we work in
the noisy comparison model, which was introduced in the seminal paper [FRPU94] and has
been studied in [GM15; BMW16]: We assume queries are of the type is the i-th coordinate
of point p (strictly) smaller than that of point q?, and the outcome of each such query is
independently correct with probability greater than some constant better than 1/2 (for
definiteness we assume probability 2/3). Our goal is to recover the exact skyline, with error
probability at most δ. In the context of crowdsourcing this model has been considered in
order to capture the fact that people might incur in errors when comparing elements. We
refer to [AZH+15] about skyline computation using the crowd and [LWZF16] for a survey
in crowdsourced data management.
Results In many settings the skyline consists of very few points compared to the input
size, motivating the study of output-sensitive algorithms. Our measure of complexity is
the number of queries. This is expressed as a function of three parameters: n = |X|, the
number of data items (points); d, the number of attributes (dimensions); and k = |sky(X)|,
the size of the skyline (output).
2
Theorem 1.1. Given δ ∈ (0, 1/2) and a set X of data items, SkyLowDim-Search(X, δ)
outputs a subset of X which, with probability at least 1 − δ, is the skyline of X. The expected
number of queries is O(nd log(dk/δ)).
Theorem 1.2. Given δ ∈ (0, 1/2) and a set X of data items, SkyHighDim-Search(X, δ)
outputs a subset of X which, with probability at least 1 − δ, is the skyline of X. The expected
number of queries is O(ndk log(k/δ) + dk 2 log(kn/δ)).
Additionally, we prove that the bound of Theorem 1.1 is (up to constant factors)
essentially tight whenever d ≤ k c for any constant c:
Theorem 1.3. Let A be an algorithm that computes the skyline with error probability less
than 1/10. Then the expected number of queries of A is Ω(nd log k).
Techniques In previous work, Groz and Milo [GM15] give three algorithms. Of the three
algorithms, the third (iii) algorithm is simply based on sorting all the input points in each
dimension and thus reduces the problem to the case of computing the skyline in the noiseless
setting. Our algorithm SkylineLowDim uses a natural but quite different idea: it is to
use discretization, by sampling, sorting all the sample points in each dimension to define
buckets, then placing points into buckets, and identifying “skyline buckets”. Eliminating
points in dominated buckets, we reduce the input size significantly allowing us to apply
a cruder algorithm to solve the problem on the smaller input. One interesting aspect of
our discretization is that a fraction of the input will be, due to the low query complexity,
incorrectly discretized yet we are able to recover the correct skyline.
Algorithms (i) and (ii) from [GM15], recover the skyline points one by one. They
iteratively compute the maximum point, in lexicographic order, among those not dominated
by the skyline points already found.1 The idea behind our algorithm SkylineHighDim
is that it is more efficient to separate the two tasks: finding a point p not dominated by
the skyline points already found, on the one hand, and computing a maximum point (in
lexicographic order) among those dominating p, on the other hand; we optimize queries
carefully for the latter of the two tasks (Algorithm MaxLex).
Our lower bound constructs a technical reduction from the problem of identifying null
vectors among a collection of vectors, each having at most one non-zero coordinate. That
problem can be studied using a two-phase process inspired from [FRPU94].
Context Groz and Milo [GM15] start the research line of computing the skyline in the
noisy setting and they give three algorithms showing three upper bounds on the query
complexity: (i) O(ndk log(dk)), (ii) O(ndk 2 log(k)), and (iii) O(nd log(dn)). Here, we
improve the first two bounds by a factor of k and we improve on the third bound.
When do our bounds improve asymptotically on the existing bounds? We are focused
on settings where the output size k is quite small compared to n: assume that k = no(1) (If
k = nΩ(1) then the simple bound (iii) from [GM15] is best.) Moreover,
1
The difference between (i) and (ii) is due to different subroutines to check dominance.
3
• If the dimension is relatively low: d = k O(1) , then Theorem 1.1 is best, beating (i) by
a factor of k, and is in fact optimal by Theorem 1.3. This comprises in particular the
constant dimension setting: Our bound O(n log k) is tight and improves on O(n log n)
([GM15]) for small enough k (e.g. poly-log n).
• If the dimension is quite high: d = 2ω(k) , then Theorem 1.2 is best, beating (ii) by a
factor of k.
Roadmap In Section 2, we provide the preliminaries, which the reader might wish to
skip on first reading. Section 3 introduces our algorithm for low dimensions (Theorem 1.1)
and Section 4.2 introduces the counterpart for high dimensions (Theorem 1.2). Section 5,
contains our lower bound (Theorem 1.3).
2
Preliminaries
Algorithm SkyLowDim-Search(X, δ) guesses an upper bound k for |skyline(X)| by a
super-exponentially increasing sequence of guesses (similarly to [Cha96; GM15]). This
reduces the problem to that of computing skyline(X) given a rough upper bound k on its
cardinality, a problem solved by our Algorithm SkylineLowDim(k, X, δ).
Algorithm SkyLowDim-Search(X, δ)
input: X set of points, δ error probability
output: skyline(X)
error probability: δ
1:
2:
3:
4:
5:
(see Theorem 1.1)
k ← (bd/δc)2
repeat
δ ← δ/2 ; k ← k 2 ; S ← SkylineLowDim(k, X, δ)
until |S| < k
Output S
Similarly, algorithm SkyHighDim-Search(X, δ) guesses an upper bound k for |skyline(X)|
by a exponentially increasing sequence of guesses. This reduces the problem to that of
computing skyline(X) given an upper bound k on its cardinality, a problem which will be
solved by Algorithm SkylineHighDim(k, X, δ).
2.1
Subroutines used by our algorithms
Sorting, searching and skyline. Before we state our Algorithm SkylineLowDim, we
introduce the subroutines it builds on, namely, SkyGM from [GM15], and NoisySearch
and NoisySort from [FRPU94]. In particular, the algorithm SkyGM yields the complexity
in (i) and (ii). The pseudocode of these three routines is provided in the Appendix.
4
Algorithm SkyHighDim-Search(X, δ)
(see Theorem 1.2)
input: X set of points, δ error probability
output: skyline(X)
error probability: δ.
1: Initialize j ← 0, k ← 1
2: repeat
3:
j ← j + 1 ; k ← 2k ; S ← SkylineHighDim(k, X, δ/8j )
4: until |S| < k
5: Output S
Theorem 2.1 ([GM15]). Given δ ∈ (0, 1/2) and a set X of data items, algorithm SkyGM(X, δ)
outputs a subset of X which, with probability at least 1 − δ, is the skyline of X. The expected
number of queries is (i) O(ndk 2 log(k/δ)) or (ii) O(ndk log(dk)). The complexity depends
on the dominance test used.
In the noisy binary search problem the input is the following: an element y, and an
ordered list (y1 , y2 , . . . , ym−1 ), accessible by comparisons that each have error probability at
most p, and a parameter δ; the goal is to output the interval I = (yi−1 , yi ] such that y ∈ I.
Theorem 2.2 ([FRPU94]). There exists an algorithm, NoisySearch, that solves the noisy
binary search problem with success probability 1 − δ and expected number of comparisons
O(log(m/δ)).
In the noisy sort problem the input correspond to an unordered set Y = {y1 , y2 , . . . , ym },
whose elements are accessible by comparisons that each have error probability at most p,
and a parameter δ. The goal is to output an ordering of Y that is the correct non-decreasing
sorted order.
Theorem 2.3 ([FRPU94]). There exists an algorithm, NoisySort, that solves the noisy sorting problem with success probability 1−δ and expected number of comparisons O(m log(m/δ)).
Boosting. A folklore approach to deal with noise is to take an algorithm for the noiseless
setting and repeat each noisy operation enough times to reduce noise and boost the success
probability; this increases the query complexity by logarithmic factors. We use this approach
in a variety of settings, so we formalize it with a (higher order) algorithm which we call
BoostProb. See Section B.2.
Proposition 2.4. Algorithm BoostProb(⊗, δ1 , δ2 ) takes as input two parameters δ1 and
δ2 that are the desired two-sided errors, and a test ⊗ (query or algorithm) that returns either
true or false with error probability at most 1/3. BoostProb incorrectly outputs true (false
positive) w.p. δ1 and incorrectly outputs false (false negative) w.p. δ2 .
Let T⊗ be the expected query time of test ⊗. BoostProb has expected query complexity
O(log(1/δ1 ))T⊗ if ⊗ should return true (YES instances) and O(log(1/δ2 ))T⊗ otherwise (NO
instances).
5
For our algorithm SkylineLowDim to be able to eliminate dominated buckets, we
design an efficient subroutine to test whether a bucket is empty of points.
In the Bucket-emptiness problem the input corresponds to a bucket B, a set of points
Y and two error parameters δ1 and δ2 . The goal is to decide whether Y ∩ B = ∅; the
algorithm incorrectly outputs true w.p. at most δ1 , and incorrectly outputs false w.p. at
most δ2 . (The two-sided error is used to improve the query complexity). We leave the
description of the algorithm IsEmpty and it analysis to Section B.2.
Lemma 2.5. Algorithm IsEmpty(B, Y, δ1 , δ2 ) solves Bucket-emptiness with expected query
complexity O(d|Y | log(1/δ1 )) + O(d log(d|Y |/δ2 )).
3
3.1
Skyline computation in low dimension: Theorem 1.1
Overview
Algorithm SkylineLowDim(k, X, δ)
input: k integer, X set of points, δ error probability
output: min{k, |skyline(X)|} points of skyline(X)
error probability: δ
δ 0 ← δ/(2dk)5 and s ← dk 2 log(d2 k 2 /δ 0 )
2: if k 5 ≥ n or d5 ≥ n then
3:
Output SkyGM(X, δ 0 )
{Phase (i): bucketing}
4: for each dimension i ∈ {1, 2, . . . , d} do
5:
Si ← NoisySort(sample of X of size s, i, δ 0 /d)
6:
Remove duplicates so that, with prob. 1 − δ 0 /d, the values in Si are all distinct
7: for each point p ∈ X do
Q
8:
Place p in set XB associated to B = di=1 Ii , with Ii = NoisySearch(pi , Si , δ 0 /(dk)).
9: for each bucket
B do
true
if XB = ∅,
10:
emptyB ←
unknown otherwise.
{Phase (ii): elimination}
11: while ∃B with emptyB = unknown and ∀B 0 dominating B: emptyB 0 = true do
12:
emptyB ← IsEmpty(B, XB , δ 0 /k, δ 0 /n)
13:
if more than |X|/ log(|X|) buckets B have emptyB = false then Output
SkylineLowDim(k, X, δ)
14: X 0 ← ∪{XB : emptyB = false and ∀B 0 dominating B: emptyB 0 = true }
{Phase (iii): solve reduced problem}
15: Output SkylineHighDim(k, X 0 , δ 0 ).
1:
6
y
bucket a
bucket b
p1
bucket c bucket d bucket e
p2
p3
bucket f bucket g
x
Figure 2: An illustration of the bucket dominance. We say a bucket i dominates another
bucket j if it is non-empty and has strictly larger coordinates. Dominated buckets are
gray-striped and dominating buckets are orange-striped, and boxed. Here bucket b dominates
c and f but not a, d, e or g. Bucket b is undominated because all buckets to the Northeast
are empty. All points in all undominated buckets are handed to the third phase, ensuring
that the skyline points p1 , p2 , and p3 are among these points.
7
Algorithm SkylineLowDim(k, X, δ) uses phases. In the first phase, bucketing: we sort
the i-th coordinate of a random sample to define s + 1 intervals in each dimension
Q i ∈ [d],
d
hence (s + 1) buckets, where each bucket is a product of intervals of the form i Ii ; then
we place each point p of X in those buckets by searching for each dimension for the interval
Ii containing pi .
In the second
uses elimination: Say that
Q phase, Algorithm SkylineLowDim(k, X, δ)Q
a bucket B = i Ii is dominated by a different bucket B 0 = i Ii0 if in every dimension
max Ii ≤ min Ii0 . Then every point in B 0 dominates every point in B, so no skyline point
belongs to a bucket dominated by a non-empty bucket. We test buckets for emptiness and
eliminate points placed in buckets that are dominated by non-empty buckets. See Figure 2
for an illustration.
In the third phase Algorithm SkylineLowDim(k, X, δ) simply calls the Algorithm
SkylineHighDim to find the skyline of the remaining points.2
3.2
Algorithm SkylineLowDim
We can now formally define the Algorithm SkylineLowDim that was outlined in section 3.1.
In the third phase we have eliminated enough points that we can call our other (less efficient)
algorithm SkylineHighDim (see Section 4.2) on the residual instance.
3.3
Analysis: Proof of Theorem 1.1
The error probability analysis follows by carefully considering the operations made by the
algorithm. We leave that analysis to the Appendix, and focus on the query complexity
analysis.
To study the query complexity of algorithm SkylineLowDim we need some structure
about the bucketing. We say the assignment performed in line 8 is decent if the following
holds: For every j ∈ {1, . . . , d} and for every interval I in Sj , at most 4|X|/(dk 2 ) points
with distinct j coordinate are placed in I, that is,
4|X|
.
dk 2
Otherwise, if the condition above does not hold for an interval we say that I is dense.
|{pj : I = NoisySearch(pj , Sj , δ 0 /(dk))}| ≤
Lemma 3.1. With probability at least 1 − 1/k, the assignment performed in line 8 is decent.
We first study the expected number of queries performed during the execution of
i
SkylineLowDim(ki , X, δ), where ki = bd/δc2 . Throughout this section we assume ki5 , d5 ≤
n = |X| since otherwise the algorithm just calls SkyGM and thus the correctness simply
follows from the correctness of SkyGM. We now study the query complexity of each phase
in SkylineLowDim.
2
Alternatively, one could use an algorithm provided by Groz and Milo [GM15], it is only important that
the size of the input set is reduced to n/k to cope with the larger runtime of the mentioned algorithms.
8
• Phase (i) Bucketing. On line 5: by Theorem 2.3 (noisy sorting) the expected
number of queries performed is d · O(dki2 log(16d2 ki2 /δ)) = O(dn log(ki d/δ)), by the
assumption over ki and d. On line 8: by Theorem 2.2 (noisy search) the expected
number of queries is nd · O(log(16d2 ki3 /δ)) = O(nd log(dki /δ)).
• Phase (ii) Elimination. On line 12: by Lemma 2.5, since (conditioned on an
execution where NoisySort and NoisySearch were both correct) at most γ =
ki3 d log(16d2 ki2 /δ) buckets tested are non-empty. Observe that γ < |X|/ log(|X|) due
to line 2. The resulting conditional expected query complexity is O(|X|d log(dki /δ)).
An incorrect execution of algorithms NoisySort and NoisySearch can cause at most
|X|/ log(|X|) buckets to be verified, however, the probability of such an error is at
most 1/k. The probability of failing i times due to line 13 is at most 1/k i . Thus,
resulting conditional expected query complexity is O(|X|d log(dki /δ)). Hence, by law
of total expectation, and since the probability of executing the expected number of
queries is at most O(|X|d log(dki /δ)).
• Phase (iii) Recovering the skyline. On line 15: We claim that on expectation
|X 0 | = O(n/ki ). which gives, by Theorem 4.3, O(|X 0 |dk log(dki /δ)). We proceed by
proving the claim. Assume that the point assignment is decent, which is, by Lemma 3.1,
w.p. 1 − 1/k the case. Let Qp be the set of points which were strictly dominated by p
and in some bucket B which was not dominated resulting in p ∈ X 0 . Let Bp be the
bucket of p. Consider every dimension i. If Ii of Bp is of type (`, r], then because
the assignment was decent there are at most L = 4|X|/(dki2 ) points in Qp w.r.t. to
dimension i. If Ii of Bp is of type I = [x], x ∈ R, then for every q ∈ Qp correctly
assigned to I we see that p dominates q. Thus summing over all dimensions we have
|Qp | ≤ |X|/ki2 and
[
|X 0 | ≤
Qp ≤ |X|/ki .
p∈skyline(X)
Suppose now that an error occurred (which happens w.p. at most 1/k) we derive,
as before, |X 0 | ≤ |X|. Thus, similarly as before, the expected size of |X 0 | is |X 0 | =
O(n/ki ), yielding the claim. Hence, the expected number of queries of that last line is
O(nd log(dki /δ)).
Let Oi be the output of SkylineLowDim(ki , X, δ/2i ). Recall that SkyLowDim-Search(X, δ)
terminates as soon as |Oi | =
6 ki . Whenever ki > |skyline(X)| such an output is returned
w.p. at least 1 − δ/2i . Let Z denote the variable denoting the number of extra iterations
due to failure, Set γ = log(logd/δ |skyline(X)|) and recall that log(·) = log2 (·). Since
i
ki ≤ (d/δ)2 , the number of iterations is bounded by dγe + 1 + Z. Therefore, the runtime
9
of SkylineLowDim(ki , X, δ/2i ) is bounded by O(nd log(2i dki /δ)). This implies that the
total expected query complexity is upper bounded by
dγe+1+Z
X
O(nd log(2i dki /δ)).
i=1
To obtain the claimed bound, we first show that
dγe+1+Z
X
i=1
log(2i dki /δ) ≤ 32 · 2Z log |skyline(X)|.
(1)
By using this inequality and that P ( Z = i ) ≤ δ i and δ < 1/2 is enough to conclude, since
∞
X
|skyline(X)|d
|skyline(X)|d
i
P ( Z = i ) = O nd log
.
O nd2 log
δ
δ
i=0
We now check that (2) holds:
dγe+1+Z
X
i=1
dγe+1+Z
i
log(2 dki /δ) ≤ 2
X
i=1
dγe+1+Z
log(dki ) ≤ 2
dγe+2+Z
≤ 4 log(d/δ) · 2
X
i
log((d/δ)2 +1 )
i=1
≤ 32 · 2Z log(d/δ) logd/δ |skyline(X)|
= 32 · 2Z log |skyline(X)|.
4
Skyline computation in high dimension: Theorem 1.2
Algorithm SkylineHighDim builds the skyline incrementally by discovering skyline points
one by one. We obtain an additional skyline point by first looking for a point p which is not
dominated by the current set of skyline points; p itself is not necessarily a skyline point.
Then we find a skyline point p∗ dominating p: it has to be a new, additional skyline point
and can thus be added to the set.
We note that an algorithm from [GM15] also builds the skyline incrementally, but it
looks for a point that is simultaneously not dominated by the current skyline and is itself a
skyline point: by separating those two tasks and optimizing each, we gain a factor of k in
complexity.
4.1
Subroutines: Domination and Lexicographic Maximum
Algorithm SkylineHighDim uses a subroutine for the Domination Problem: the input
is a set of points S, a point q ∈ S, and two error parameters δ1 and δ2 . The goal is to check
whether the point q is not dominated by S, that is, there is no point in S that dominates q
on every dimension. Our subroutine is a more efficient variant of an approach already taken
in [GM15, Lemma 5]. The algorithm and analysis can be found in Section B.4.
10
Lemma 4.1. There is an algorithm for the domination problem, SetDominates(S, q, δ1 , δ2 ),
that is correct w.r.t. to its specification and has expected query complexity O(|S|d log(1/δ2 ) +
d log(|S|/δ1 )).
Algorithm SkylineHighDim also uses a subroutine for Lex-Maximum: the input
is a point p, a set of points S and an error probability δ. This algorithm computes the
maximum point in the lexicographic order3 in S that is not dominated by p. In [GM15,
Proposition 2], the authors sketch an algorithm doing the same.
Algorithm MaxLex aims to find the maximum point in lexicographic order, among
the ones that dominate p, by keeping a counter for each point, that is increased when
that point is found to dominate p and decreased when it is either less than another point
in lexicographic order or found not to dominate p. The increments are chosen in such a
way that the counter of the unknown desired output point performs a random walk biased
upwards with non-uniform step size, the others perform random walks biased downwards,
and we always compare (and compare to p) the points with the largest current counters.
We leave the full description of the algorithm and its analysis to Section B.4.
Lemma 4.2. There is an algorithm for lex-maximum, MaxLex(p, S, δ), that is correct
w.r.t. to its specification and has expected query complexity O(|S|d log(1/δ))..
4.2
Algorithm SkylineHighDim
We are now ready to give the full description of the algorithm SkylineHighDim.
Theorem 4.3. Suppose k ≥ |skyline(X)|. Then, SkylineHighDim(k, X, δ) outputs
skyline(X) w.p. at least 1 − δ with expected query complexity O(dkn log(k/δ)).
The proof of Theorem 1.2 follows the same lines as the proof of Theorem 1.1, so we leave
it to the Appendix.
5
Lower Bound-Theorem 1.3
In this section, we exhibit an Ω(dn log k) lower bound on the query complexity in the noisy
skyline problem, denoted Skyline. To that end, we define a noisy vector problem, in which
one is given k vectors each of length ` and needs to decide for each vector whether it is the
all-zero vector. We prove a lower bound for this problem and reduce it to Skyline yielding
the desired result.
3
Given two points, p = (p1 , p2 . . . , pd ) and q = (q1 , q2 . . . , qd ) the first one is lexicographically smaller than
the second one, denoted by p ≤lex q , if pi < qi for the first i where pi and qi differ. If there is no such i
meaning that the points are identical we use the id of the points in the input as a tie-breaker, ensuring that
we obtain a total order.
11
Algorithm SkylineHighDim(k, X, δ)
(see Theorem 4.3)
input: X set of points, k upper bound on skyline size, δ error probability
output: min{k, skyline(X)} skyline points w.p. 1 − δ
Initialize S0 ← ∅, C ← X
for i = 1 to k do
{Find a point p not dominated current skyline points}
while C not empty do
Pick an arbitrary p ∈ C
if not SetDominates(Si , p, δ/4k, δ/(4k|X|)) (check if p is already dominated)
then
break
C ← C \ {p}
{Find a skyline point dominating p}
Compute p∗ ← MaxLex(C, p, δ/2k)
Si ← Si−1 ∪ {p∗ }
Output Sk
1:
2:
3:
4:
5:
6:
7:
8:
9:
10:
11:
12:
5.1
(k, `)-Null-Vectors: definition and lower bound
Definition 1. In the (k, `)-Null-Vectors the input
S is a collection {v1 , v2 , . . . , vk } of k
P
vectors such that for each i, vi ∈ {0, 2}` and
vji ≤ 2, and the output is a vector
jP
(w1 , w2 , . . . , wk ) ∈ {0, 2}k such that for each i, wi = j vji .
We define the distribution µ over vectors of {0, 2}` as follows.
(0, 0, . . . , 0) with probability 1/2
(1, 0, . . . , 0) with probability 1/(2`)
v=
.
..
(0, . . . , 0, 1) with probability 1/(2`)
For inputs to (k, `)-Null-Vectors, we will consider the product distribution µk .
Lemma 5.1. For (k, `)-Null-Vectors under the product distribution µk , if A is is a deterministic algorithm with success probability at least 3/4, then the worst case number of queries
of A is Ω(`k log k).
The proof can be found in Section B.5.
5.2
Reduction from (k, `)-Null-Vectors to Skyline
Step 1. Assume, for simplicity, that d − 2 divides k. From an input S = {u1 , u2 , . . . , uk } to
the (k, `)-Null-Vectors, we first show how to construct an input IS for Skyline with n points
12
in d dimensions and a skyline that is likely to be of size k, where n = (` + d − 2)k/(d − 2).
We first randomly permute the entries of each ui , by using k independent permutations,
resulting in Sπ = {v1 , v2 , . . . , vk }. Partition Sπ into k/(d − 2) blocks of d − 2 vectors, where
for j ∈ {0, 1, . . . , k/(d − 2) − 1}, block Sπj = {vj(d−2)+i : i ∈ [d − 2]}.
For each block, define ` + d − 2 points, as displayed (one point per row) on Figure 3,
and the union over all blocks is the input IS to the Skyline. Formally, we define point p(t)
with t = j` + i as follows.
if d0 = d − 1,
j
if d0 = d,
n − j
(i)
pd0 := 1
if d0 = i − ` and ` ≤ i ,
j·(d−2)+i
vi
if d0 = i and i ∈ [d − 2],
0
otherwise.
d−2
pj·(`+d−2)+1
j·(`+d−2)+2
2
p
..
.
pj·(`+d−2)+`
pj·(`+d−2)+`+1
..
.
p(j+1)·(`+d−2)
Sπj
1
0
0
0
0
1
0
0
0
0
1
0
0
0
0
1
j
j
j
j
j
j
j
j
j
j
j
j
n−j
n−j
n−j
n−j `
n−j
n−j
n−j
n−j
n−j
n−j d−2
n−j
n−j
Figure 3: Block (j, n − j) of the reduction. The vectors of Sπj placed in this block are
vj·(d−2)+1 , vj·(d−2)+2 , . . . , vj·(d−2)+(d−2) .
Step 2. Because of the non-domination implied by the last two coordinates of any
point, the skyline of the set of points is the sum over all blocks of the skyline of each block.
Fix an arbitrary block and focus on the first d − 2 dimensions. For each dimension, the
corresponding column (whose first ` coordinates are those of some vector vi ) contains exactly
one 1 (on the row of some point p) and possibly one 2 , the remaining entries being all 0.
Thus it is easy to verify that p is part of the skyline if and only if vi = ~0.
From the output sky(I) it is now easy to construct the output of the (k, `)-Null-Vectors:
For all blocks, for all dimensions ≤ d − 2, if p ∈ sky(I) then wi ← 0 else wi ← 2. This
yields the correct output w = (w1 , w2 , . . . , wk ). Thus we derive the following observation.
Observation 1. Given the set of points sky(IS ), one can recover the solution to the (k, `)Null-Vectors without further queries.
13
In the following we prove that the construction is likely to have k skyline points.
Observation 2. Let E be the event that the input IS has exactly k skyline points. Then,
√
P ( E ) ≥ 1 − 1/k as long as k ≤ 5 n.
Proof. First observe that, by construction, regardless of whether E holds, every block
contains at most d − 2 skyline points: Consider an arbitrary block. The last two dimensions
are identical for each point belonging to that block and we focus thus on the first d − 2
dimensions.
There are exactly d − 2 points with one coordinate being 1 and all of these points are
potential skyline points. In particular, take any such point p and assume that the i’th
coordinate of p is 1. Then p is part of the skyline if and only if the vector vi is the null
vector. Moreover, every block can have at most d − 2 entries with value 2 and each such 2
eliminating one potential skyline point. Thus, there are at most d − 2 skyline points per
block.
Consider the vertices vi1 , vi1 +1 , . . . , vi2 of any block. We say they are collision free if
0
the following holds: if vjj∗ = 2 for j ∈ [i1 , i2 ], then vjj∗ = 0 for all j 0 ∈ [i1 , i2 ] \ {j}. Observe
that if the vertices of any block are collision free, then each of the first d − 2 dimensions
is dominated by a distinct skyline point and thus there d − 2 skyline points in that block.
Thus, if the vectors of every block are collision free, then there d − 2 skyline points per block
and summing up over all k/(d − 2) blocks, we get that there are thus k skyline points in
total.
Thus, in order to bound P ( E ) it suffices to bound the probability that all blocks are
collision free.. Recall that the random permutations π1 , π2 , . . . , πk permute each vector vi
independently. Since in a block at most k 2 pairs may collide, and each collision happens with
probability 1/`, the expected number of collisions per block is at most k 2 /`. The expected
number of collisions over all blocks is thus, by the Union bound, at most (k/(d − 2)) · k 2 /` ≤
1/k, by assumption on k. Thus, the claim follows by applying Markov inequality.
5.3
Proof of Theorem 1.3
Suppose for the sake of contradiction that there exists an algorithm A recovering the skyline
for any input with exactly k skyline points, with error probability at most 1/10, and using
o(nd log k) queries in expectation. By Markov’s inequality, the probability that the number
of queries exceeds 5 times the expectation is at most 1/5, so truncating the execution at
that point adds 1/5 to the error probability, transforming A into an algorithm B that
recovers the skyline for any input with exactly k skyline points, with error probability at
most 1/5 + 1/10 < 1/3, and using o(5nd log k) queries in the worst case. We claim that
this implies that one can solve the (k, `)-Null-Vectors with o(nd log k) w.p. at least 1/3
contradicting Lemma 5.1.
14
Let S be the input of the (k, `)-Null-Vectors. We cast S as an input IS of B as described
in Section 5.2. By Observation 2, the event E holds w.p. at least 1 − 1/k and thus there are
k skyline points.
By assumption, B can thus compute the skyline w.p. at least 1/2 − 1/k ≥ 1/3, where
we used the Union bound. Thus, by Observation 1, one can obtain w.p. at least 1/3 the
solution to (k, `)-Null-Vectors using o(nd log k) queries, a contradiction.
Future Work
We show that that the query complexity is Θ(dn log(dk) whenever d ≤ k c for any constant c.
The arising questions is thus what the query complexity is when d = ω(k c ). In the light of
the upper bound O(dkn log k) of Theorem 1.2 we conjecture that there exists an algorithm
achieving a query complexity of Θ(dn log(k).
Furthermore, we believe that for constant dimensions, the correct bound is O(n log k)
regardless of the instance (assuming that there are exactly k-skyline points) even if the
algorithm knows the instance up to a permutation—in other words, we believe that our
algorithm is instance-optimal for constant dimensions.
Acknowledgements
We gratefully acknowledge Vincent Cohen-Addad for sharing his insights and ideas.
References
[AAA+11]
P. Afshani, P. K. Agarwal, L. Arge, K. G. Larsen, and J. M. Phillips. “(Approximate) Uncertain Skylines”. In: Proceedings of the 14th International
Conference on Database Theory. ICDT ’11. Uppsala, Sweden: ACM, 2011,
pp. 186–196. url: http://doi.acm.org/10.1145/1938551.1938576 (cit. on
p. 2).
[AZH+15]
A. Asudeh, G. Zhang, N. Hassan, C. Li, and G. V. Zaruba. “Crowdsourcing
pareto-optimal object finding by pairwise comparisons”. In: Proceedings of
the 24th ACM International on Conference on Information and Knowledge
Management. ACM. 2015, pp. 753–762 (cit. on p. 2).
[BS66]
O. Barndorff-Nielsen and M. Sobel. “On the Distribution of the Number
of Admissible Points in a Vector Random Sample”. English. In: Theory of
Probability and its Applications 11.2 (1966), pp. 249–21. url: http://search.
proquest.com/docview/915869827?accountid=15867 (cit. on p. 1).
15
[BKS01]
S. Börzsönyi, D. Kossmann, and K. Stocker. “The Skyline Operator”. In:
Proceedings of the 17th International Conference on Data Engineering. IEEE
Computer Society, 2001, pp. 421–430. url: http://dl.acm.org/citation.
cfm?id=645484.656550 (cit. on p. 1).
[BMW16]
M. Braverman, J. Mao, and S. M. Weinberg. “Parallel Algorithms for Select
and Partition with Noisy Comparisons”. In: Proceedings of the Forty-eighth
Annual ACM Symposium on Theory of Computing. STOC ’16. Cambridge,
MA, USA, 2016, pp. 851–862 (cit. on p. 2).
[Cha96]
T. M. Chan. “Optimal output-sensitive convex hull algorithms in two and three
dimensions”. In: Discrete & Computational Geometry 16.4 (1996), pp. 361–368.
url: http://dx.doi.org/10.1007/BF02712873 (cit. on p. 4).
[CCM13]
J. Chomicki, P. Ciaccia, and N. Meneghetti. “Skyline Queries, Front and Back”.
In: SIGMOD Rec. 42.3 (2013), pp. 6–18. url: http://doi.acm.org/10.
1145/2536669.2536671 (cit. on p. 1).
[FRPU94]
U. Feige, P. Raghavan, D. Peleg, and E. Upfal. “Computing with Noisy
Information”. In: SIAM Journal on Computing 23.5 (1994), pp. 1001–1018
(cit. on pp. 1–5, 27–29).
[Fel68]
W. Feller. An Introduction to Probability Theory and Its Applications. Vol. 1.
Wiley, 1968. url: http://www.amazon.ca/exec/obidos/redirect?tag=
citeulike04-20%7B%5C&%7Dpath=ASIN/0471257087 (cit. on p. 17).
[GSG07]
P. Godfrey, R. Shipley, and J. Gryz. “Algorithms and Analyses for Maximal
Vector Computation”. In: The VLDB Journal 16.1 (2007), pp. 5–28. url:
http://dx.doi.org/10.1007/s00778-006-0029-7 (cit. on p. 1).
[GM15]
B. Groz and T. Milo. “Skyline Queries with Noisy Comparisons”. In: Proceedings
of the 34th ACM SIGMOD-SIGACT-SIGAI Symposium on Principles of
Database Systems. PODS ’15. Melbourne, Victoria, Australia: ACM, 2015,
pp. 185–198. url: http://doi.acm.org/10.1145/2745754.2745775 (cit. on
pp. 1–5, 8, 10, 11, 29).
[KML08]
M. E. Khalefa, M. F. Mokbel, and J. J. Levandoski. “Skyline query processing
for incomplete data”. In: 2008 IEEE 24th International Conference on Data
Engineering. IEEE. 2008, pp. 556–565 (cit. on p. 2).
[KRR02]
D. Kossmann, F. Ramsak, and S. Rost. “Shooting Stars in the Sky: An Online
Algorithm for Skyline Queries”. In: Proceedings of the 28th International
Conference on Very Large Data Bases. VLDB ’02. Hong Kong, China: VLDB
Endowment, 2002, pp. 275–286. url: http://dl.acm.org/citation.cfm?
id=1287369.1287394 (cit. on p. 1).
16
[KLP75]
H. T. Kung, F. Luccio, and F. P. Preparata. “On Finding the Maxima of a
Set of Vectors”. In: J. ACM 22.4 (1975), pp. 469–476. url: http://doi.acm.
org/10.1145/321906.321910 (cit. on p. 1).
[LWZF16]
G. Li, J. Wang, Y. Zheng, and M. Franklin. “Crowdsourced data management:
A survey”. In: (2016) (cit. on p. 2).
[LEB13]
C. Lofi, K. El Maarry, and W.-T. Balke. “Skyline queries in crowd-enabled
databases”. In: Proceedings of the 16th International Conference on Extending
Database Technology. ACM. 2013, pp. 465–476 (cit. on p. 2).
[MWK+11] A. Marcus, E. Wu, D. Karger, S. Madden, and R. Miller. “Human-powered
Sorts and Joins”. In: Proc. VLDB Endow. 5.1 (2011), pp. 13–24. url: http:
//dx.doi.org/10.14778/2047485.2047487 (cit. on p. 2).
[PJLY07]
A
J. Pei, B. Jiang, X. Lin, and Y. Yuan. “Probabilistic Skylines on Uncertain
Data”. In: Proceedings of the 33rd International Conference on Very Large
Data Bases. VLDB ’07. Vienna, Austria: VLDB Endowment, 2007, pp. 15–26.
url: http://dl.acm.org/citation.cfm?id=1325851.1325858 (cit. on
p. 2).
Auxiliary Claims
Proposition A.1 ([Fel68, Chapter XIV.2, XIV.3]). Let p ∈ (0, 1) \ {1/2} and b, s ∈ N.
Consider a discrete time Markov chain (Zt )t≥0 with state space Ω = [0, b] where
• Z0 = s ∈ [0, b]
• P ( Zt = i | Zt−1 = i − 1 ) = p for i ∈ [1, b − 1], t ≥ 1
• P ( Zt = i | Zt−1 = i + 1 ) = 1 − p for i ∈ [1, b − 1], t ≥ 1
• P ( Zt = i | Zt−1 = i ) = 1 for i ∈ {0, b}, t ≥ 1
Let T = min{t ≥ 0 | Zt ∈ {0, b}}. Then,
P ( ZT = b ) =
P ( ZT = 0 ) =
1−p
p
1−p
p
s
b
−1
and
−1
s
1−p
1−p
−
p
p
.
b
1−p
−
1
p
17
b
Moreover,
E(T ) =
B
s
b
−
·
1 − 2p 1 − 2p
1−
1−p
p
1−
1−p
p
s
b .
Appendix
B.1
Error probability analysis of SkylineLowDim
The probability that the output of SkyLowDim-Search(X, δ) is incorrect is the probability
that there exists
P an iteration such that SkylineLowDim(k, X, δ) is incorrect, which by the
union bound j≥1 δ/2j ≤ δ, hence the error probability in Theorem 1.1. In the following
we show that
P ( output of SkylineLowDim(k, X, δ) incorrect ) ≤
δ
2(dk)5
(2)
The proof proceeds by examining possible sources of error during the execution of Algorithm
SkylineLowDim(k, X, δ) and bounding the probability that they occur. Let δ 0 = δ/(2dk)5 .
Observe that if k 5 ≥ n or d5 ≥ n and SkyGM returns an incorrect skyline, then by
Theorem E.1 this happens w.p. at most δ 0 .
• Phase (i) Bucketing. Some sample is incorrectly sorted in line 5: by Theorem 2.3
and Union Bound over i ∈ [d], error probability is at most δ 0 . The samples are correctly
sorted in line 5, but some points considered in line 6 are either deleted when they
should have been kept, or kept when they should have been deleted: by Theorem 2.3
and Union Bound over i ∈ [d], j ∈ [s − 1], error probability at most δ 0 . The samples
are sorted correctly, but some skyline point is placed in the wrong buckets: by Theorem 2.2 and Union Bound over i ∈ [d] and p ∈ skyline(X), error probability at most δ 0 .
• Phase (ii) Deleting dominated buckets. The samples are sorted and filtered
correctly and the skyline points are placed in the correct buckets, but on line 12
some bucket that contains a skyline point is (incorrectly) tested as empty: Let p ∈ B
be a skyline point. By assumption, p ∈ XB . By the specification of IsEmpty, the
probability that B is (incorrectly) tested as empty is bounded by δ 0 /k. Applying the
Union Bound on the at most min{k, |skyline(X)|} buckets where this might happen,
the error probability is at most δ 0 . (Note that this never happens more than k times
because as soon as more than k skyline buckets are tested non-empty, the algorithm
stops with an error message.)
The samples are sorted and filtered correctly and the skyline points are placed in
the correct buckets, but on line 12 some bucket that dominates a bucket containing
18
a skyline point is (incorrectly) tested as empty: For each dominating bucket, by
Lemma 2.5 this has probability at most δ 0 /n. There are at most n such buckets (since
each has XB =
6 ∅), so by the Union Bound the probability overall in this case is at
most δ 0 .
• Phase (iii) Recovering the skyline. Finally, if the call to SkylineHighDim on
line 15 results in an error, that has probability at most δ 0 by the specification of
SkylineHighDim.
Overall, the error probability sum up to at most 16δ 0 and thus (2) holds.
Proof of Lemma 3.1. Recall that δ 0 = δ/(2dk)5 . For the ease of presentation we will assume
that the line 6 is executed after the points are assigned to buckets (line 8). Assume the
points of X are ordered w.r.t. to their j’th dimension, breaking ties arbitrarily. Consider
these ordered points to be divided into blocks, each one having ` = |X|/(dk 2 ) consecutive
points. In particular, the number of blocks is at most dk 2 .
Consider now the samples after line 5. Each block contains one sample with probability
at least 1 − (1 − `/|X|)s−1 ≥ 1 − δ 0 /d2 k 2 . Thus, in this event, the distance between any two
samples is at most 2` which implies that the number of distinct values is bounded by 2`.
Furthermore, with probability δ 0 /(dk 2 ) the number of points placed incorrectly between
any two samples in line 5 is at most 2|X|/(dk 2 ) w.p. at least 1 − δ 0 , by Theorem 2.2 and
Chernoff bounds4 . Moreover, recall that by assumption, line 6 was executed correctly. We
thus get that that the number of points in I is bounded by 2` (maximum distance between
two samples) plus 2|X|/(dk 2 ) (the incorrectly sorted points), and the points added through
the removal of duplicates (line 6), but these have all the same j-th coordinate. Therefore,
|{pj : p was sorted into I in line 8}| ≤ 2` +
2|X|
4|X|
=
.
2
dk
dk 2
By taking the union bound over all d dimensions and over all s intervals, the error probability
is at most 3δ 0 .
B.2
Subroutines analysis for SkylineLowDim
Proof of Proposition 2.4. By symmetry, assume that ⊗ should return true. Every query
⊗ returns true w.p. at least 2/3, by assumption, so the difference between the number of
queries returning true and the number of queries returning false is a random walk with
bias 2/3. The algorithm outputs false if and only if a biased random walk starting at
s = log(1/δ2 ) reaches 0 before reaching b = log(1/δ2 ) + log(1/δ1 ). Applying Proposition A.1
part 1, the probability of reaching 0 is bounded by 2− log(1/δ2 ) = δ2 .
4
we assume k5 , d5 ≤ n since otherwise the algorithm just calls SkyGM.
19
Algorithm BoostProb(⊗, δ1 , δ2 )
input: a test ⊗ (query or algorithm) returning either true or false w.p. at least 2/3
output: whether ⊗ should return true
error: incorrectly outputs true w.p. δ1 and incorrectly outputs false w.p. δ2 .
Execute/query ⊗ until one of the following two cases.
The number of times the outcome was true exceeds the number of false by log(1/δ1 ). In
this case output true.
The number of times the outcome was false exceeds the number of true by log(1/δ2 ). In
this case output false.
Furthermore, using Proposition A.1 part 2 we derive the claimed bound on the number
of queries it takes for the algorithm to outputs either true or false since this happens when
the random walk reaches either 0 or b. Each query takes T (⊗) expected time, concluding
the proof.
Algorithm for Bucket-emptiness. To prove Lemma 2.5 we require an algorithm for testing
whether a given point belongs or not to a bucket. The following algorithm, InBucket,
solves this problem with constant error probability.
Q
Algorithm InBucket(p, B = di=1 Ii )
input: p point, B bucket
output: whether p ∈ B
error: incorrect output w.p. 1/16
for each dimension i ∈ [d] do
if BoostProb(pi ∈
/ Ii , 1/(16d), 1/16) then
Output false
Output true
Lemma B.1. InBucket(p, B) is correct w.r.t. to its specification and has expected query
complexity O(d).
Proof. Suppose p ∈ B. In all d dimensions i we have pi ∈ Ii . By the Union bound,
the probability that BoostProb(pi ∈
/ Ii , 1/(16d), 1/16) returns incorrectly true is at most
d × 1/(16d) = 1/16. The query complexity is, by Proposition 2.4 and linearity of expectation,
d · O(1) = O(d).
Now suppose that p 6∈ B. Let i be a dimension such that pi ∈
/ Ii . For that value of i, the
probability that BoostProb(pi ∈
/ Ii , 1/(16d), 1/16) incorrectly returns false is at most 1/16.
The expected query complexity for that i is, by query complexity of BoostProb, O(log(d)).
If the algorithm observes correctly that pi 6∈ Ii , then it terminates without testing any
further dimensions. Thus, the expected total query complexity is given by two terms: the
20
expected query complexity for all dimensions i such that pP
i ∈ Ii , and the expected query
complexity for all dimensions i such that pi ∈
/ Ii : d · O(1) + j≥0 O(log d) · 161j = O(d).
Now we can describe the algorithm IsEmpty, which solves the Bucket-emptiness
problem. The idea is simple: we iterate over every point in Y , calling a boosted version of
the InBucket subroutine, since we look for a two-sided error algorithm.
Q
Algorithm IsEmpty(B = i Ii , Y = {y1 , . . . , ym }, δ1 , δ2 )
input: B bucket, Y set of points, δ1 , δ2 error probabilities
output: whether Y ∩ B = ∅
error: incorrectly outputs true w.p. δ1 and incorrectly outputs false w.p. δ2 .
for each point p ∈ Y do
if BoostProb(InBucket(p, B), δ2 /m, δ1 ) then
Output false
Output true
Proof of Lemma 2.5. Suppose B is empty. For all m points p ∈ Y we have p ∈
/ B. By the
Union bound, the probability that BoostProb(InBucket(p, B), δ2 /m, δ1 ) returns incorrectly true is at most m · δ2 /m = δ2 . The query complexity is, by Proposition 2.4 and
linearity of expectation, mT (InBucket(p, B)) log(1/δ1 ), which is O(dm log(1/δ1 ).
Now suppose that B is non-empty. Let p be a point of Y such that p ∈ B. For that
value of p, the probability that BoostProb(InBucket(p, B), δ2 /m, δ1 ) incorrectly returns
false is at most δ1 . The expected query complexity for this test is O(d log(m/δ2 )). If the
algorithm observes correctly that p ∈ B, then it terminates without testing any further
points. Thus, the expected total query complexity is given by two terms: the expected
query complexity for all points p such that p ∈
/ B, and the expected query complexity for
all points p such that p ∈ B:
X
m · d log(1/δ1 ) +
O(d log(m/δ2 )) · δ1j = O(dm log(1/δ1 )) + O(d log(m/δ2 )),
j≥0
hence the claimed bounds.
B.3
Proof of Theorem 1.2
Proof of Theorem 4.3. We first study the correctness of algorithm SkylineHighDim(k, X, δ),
that is, with error probability at most δ the algorithm returns min{k, |skyline(X)|} skyline points. We will check that at every iteration i ∈ {1, . . . , k} the probability error
is at most δ/k, that is, the probability of not recovering a skyline point is at most δ/k.
Then, the overall error probability by the algorithm is at most δ. Let Si be the skyline
21
points identified after the first i iterations. Here are the possible error sources for Algorithm SkylineHighDim(k, X, δ) at iteration i:
• Phase (i) testing domination. An skyline point is incorrectly certified as dominated
in line 6, or a dominated point is certified to be non dominated by SetDominates. By
Lemma B.2, this happens with probability error at most δ/(4k) for the first source of
error. For the second source of error this happens w.p. at most |X| · δ/(4k|X|) = δ/2k,
by the union bound.
• Phase (ii) computing lex-maximum. The dominance test was performed correctly,
but the MaxLex computation is incorrect in line 10: by Proposition 4.2 this happens
with probability at most δ/2k.
Overall, the probability that all iterations are correct is at least 1 − k(δ/2k + δ/2k) = δ, by
the union bound.
We now consider the query complexity of SkylineHighDim(k, X, δ). Observe that
every point in p ∈ X will consider at most twice In any call of SetDominates(Si , p, ·, ·).
Thus, by Lemma B.2, the expected query complexity is
O(|Si |d log(4k|X|/δ)) = O(dk log(k|X|/δ)
per found skyline point and O(d log(|Si |k/δ)) for each non-skyline point. This gives a
total expected query complexity of O(k 2 d log(k|X|/δ)) + O(|X|d log(k/δ)) as claimed. This
finishes the proof.
Proof of Theorem 1.2. First consider the error probability of SkylineHighDim(ki , X, δ/8i ).
This error probability is bounded by δ/2i for the i’th iteration. Thus, by the union bound,
the probability that for no ki an error occurred is at most δ.
We now consider the query complexity. Let Oi be the set of the output of Algorithm
SkylineHighDim(ki , X, δ/8i ). Recall that SkylineHighDim(X, δ) terminates as soon
as |Oi | 6= ki . Whenever ki > |skyline(X)| such an output is returned w.p. at least 1 − δi .
Let Z denote the variable denoting the number of ’extra’ increase due to failure, i.e.,
ki > |skyline(X)|, but |Oi | =
6 |ki |. Set γ = log(|skyline(X)|) and recall that log(·) = log2 (·).
i
Since ki = 2 , the number of iterations is bounded by dγe + 1 + Z. Therefore, the
query complexity of SkylineHighDim(ki , X, δ/8i ) is bounded by O(ndki · log(2i ki /δ)) +
O(ki2 d log(ki n/δ)), which in turns implies that the total expected query complexity is
bounded by
dγe+1+Z
X
O(ndki · log(8i ki /δ) + ki2 d log(ki n/δ)).
i=1
22
Let k = |skyline(X)|. We first show the following statement
dγe+1+Z
X
ki log ki (ki /δ) ≤ O(2Z Z · k log(k/δ)).
i=1
We have
dγe+1+Z
X
i=1
dγe+1+Z
dγe+1+Z
i
i
2 log(8 ki /δ) ≤
X
2
i
4
/δ)
log(kdγe+1+Z
i=1
≤
X
4
/δ)
log(kdγe+1+Z
2i
i=1
Z
Z
= O(2 · k log(k · 2 /δ)).
Note that P ( Z = i ) ≤ δ i and δ ≤ 1/8. Therefore, the total expected number of queries is
thus
!
∞
X
Z
i
O
2 Z · k log(k/δ)δ = O(k log(k/δ)).
i=0
Similarly, we can show the following statement
dγe+1+Z
X
i=1
ki2 log ki (ki n/δ) ≤ O(4Z Z · k 2 log(kn/δ)).
Using P ( Z = i ) ≤ δ i and δ ≤ 1/8 we get
∞
X
i=0
O(4Z Z · k 2 log(kn/δ)) · P ( Z = i ) = O(k 2 log(kn/δ)).
Overall we obtain the number of queries claimed in the theorem.
B.4
Subroutines analysis for SkylineHighDim
The algorithm SetDominates is based on a two-sided algorithm for the basic problem
Domination: the input is two points p and q and two error parameters δ1 and δ2 . The
goal is to check whether p dominates q. The following algorithm solves this problem.
With that at hand, the algorithm SetDominates runs over every point in S, and breaks
with true as soon as it finds a point dominating q.
23
Algorithm Dominates(p, q)
input: points p, q and two error parameters δ1 and δ2
output: whether p dominates q
error probability: 1/16
for each dimension i do
if BoostProb(pi <i qi , 1/(16d), 1/16) then
Output false
Output true
Algorithm SetDominates(S, q, δ1 , δ2 )
input: Set of points S, a point q, and two error parameters δ1 and δ2
output: whether there exists a p ∈ S that dominates q
error: incorrectly outputs true w.p. δ1 and incorrectly outputs false w.p. δ2 .
for each p ∈ S do
if BoostProb(Dominates(p, q), δ1 /|S|, δ2 ) (i.e., test if p ≥ q) then
Output true
Output false
In the following lemma we study the algorithm Dominates. The proof of Lemma 4.1
follows directly, from the if line and the application of union bound.
Lemma B.2. The algorithm Dominates(p, q) is correct w.r.t. to its specification and has
expected query complexity O(d).
Proof. Suppose p dominates q. Thus in all dimensions i we have pi ≥ qi . Hence, for any
dimensions, the probability that BoostProb(pi < qi , 1/(16d), 1/16) returns incorrectly true
is at most 1/(16d). Taking the union bound over all d dimensions yields the desired bound
on the probability of incorrectly returning false. The query complexity is, by Proposition 2.4
(query complexity of BoostProb) and linearity of expectation,
d · O(1) = O(d).
Now suppose that p does not dominate q. The query complexity for dimension i with
pi ≥ qi is due to the fact that the query complexity of BoostProb is, in expectation,
O(log(1/16)) = O(1). Since p does not dominate q, there exists a dimensions i such that
pi < qi . The expected query complexity is, by Proposition 2.4, O(log(16d)). If the algorithm
observes correctly that pi < qi , then it terminates without testing any further dimensions.
The probability that BoostProb(pi < qi , 1/(16d), 1/16) returns incorrectly false is at most
1/16. Thus, the case that the algorithm does not observes that pi < qi , happens only w.p.
24
1/16 and hence the expected total query complexity is given by
|{i∈{1,2,...,d} : pi <qi }|−1
d · O(1) +
X
O(log d) ·
j=0
1
= O(d).
16j
Algorithm Lex(p, q)
input: points p, q
output: whether p >lex q
error probability: 1/16
for each dimension i do
if BoostProb(pi >i qi , 1/(32d), 1/32) then
Output true
else
if BoostProb(qi , >i pi , 1/(32d), 1/32) then
Output false
Output true
Algorithm MaxLex(p, S, δ)
input: Point p, set S containing p, error probability δ
output: the point p∗ ∈ S which has the maximum lexicographic order among those that
dominate p
error probability: δ
For all q ∈ S, c(q) ← log(1/δ)
repeat
Let q1 ← argmaxq∈S c(q);
Let q2 ← argmaxq∈S\{q1 } c(q)
if Lex(q1 , q2 ) (i.e., q1 >lex q2 ) then
x ← q1 , y ← q2
else
x ← q2 , y ← q1
c(y) ← c(y) − 1
if Dominates(x, p) (i.e., x ≥ p) then
c(x) ← c(x) + 1/2
else
c(x) ← c(x) − 1
until c(q2 ) ≤ −2
Output argmaxq∈S c(q).
25
Observation 3. Consider MaxLex. There exists a iteration t ≤ 10(|S| + 3)(log(1/δ) + 3)
(of the repeat loop) in which all but first counter are smaller than −1, in symbols, if
q 6= arg maxp∈S c(p), then c(q) ≤ −1.
Proof. Let τ = 10(|S| + 3)(log(1/δ) + 3). First observe that after any iteration all but the
largest and second largest counter have a value which differs by at most 1. Furthermore,
the sum of the counters at iteration t0 is at most |S| log(1/δ) − t0 /2 implying that after
2τ /5 = 4(|S| + 3)(log(1/δ) + 3) iterations all but the largest c(q1 ) and second largest counter
c(q2 ) must be smaller than −2. At this iteration 2τ /5, the sum of the first two counters is
at most c(q1 ) + c(q2 ) ≤ 2 log(1/δ) + τ /5. Thus after further 3τ /5 iterations, at least one of
the counters c(q1 ) or c(q2 ) must have been −1. Thus there exists an iteration where all but
one counter are −1 or smaller.
Proof of Lemma 4.2. Let q ∗ be the maximum in lexicographic order in S that is not dominated by p. By Observation 3, there is an iteration where all but the first counter are
smaller than −2. Thus, it suffices to show that, with probability error at most δ, it holds
that c(q ∗ ) > −1, implying that q ∗ must be the point returned by the algorithm.
Let Xt denote the value of the counter c(q ∗ ) at the time where q ∗ took part in t
comparisons and we will say that Xt is the counter of q ∗ after t time steps (the time
steps are thus the number of comparisons q ∗ took part in). Let Yt denote the value of the
counter c(q ∗ ) after point q ∗ took part of t rounds, which are defined as follows. Each round
consists out of one or two time-steps. A round ends when either Xt increased two times or
decreased at least once. Hence, given the outcome of all time-steps one can group them into
rounds. Furthermore, let τ (t) be the time-step at which round t ends. Let Et be the event
that Xτ (t)+1 = Xτ (t) + 1/2. Note that ¬Et implies Xτ (t)+1 = Xτ (t) − 1. Moreover, define
Yt = Xτ (t) and
if ¬Et ,
Xτ (t) − 1
Yt+1 = Xτ (t) − 1/2 if Et and ¬Et+1
Xτ (t) + 1
if Et and Et+1
Let T = min{t ≥ 0 : Xt ∈ {0, 20(|S| + 3)(log(1/δ) + 3)}}. We will show that for all t ∈ Z
we have Yt > 0 which implies that for all t ∈ Z we have Xt > 0 − 1 = −1 since every round
consists of at most two time-steps. Consider the random walk (Zt )t≥0 with state space [0, b]
in Proposition A.1, where b = 2τ = 20(|S| + 3)(log(1/δ) + 3). Let Z0 = Y0 = X0 = log(1/δ).
Conditioning on Ft and Yt = i − 1 (the latter implying that Xτ (t) = i − 1) we have for
i − 1 ∈ [1, b − 1]
P ( Yt+1 = i ) = P ( Et+1 |Et ) · P ( Et ) ≥ (1 − 2/12) · (1 − 2/12)
≥ 2/3 = P ( Zt = i|Zt−1 = i − 1 ) ,
where we used the union bound. Thus
P ( Yt+1 = i|Yt = i − 1 ) ≥ P ( Zt+1 = i|Zt = i − 1 ) .
26
By Proposition A.1 parameters p = 10/12, s = log(1/δ), b we have that
b
s
1−p
1−p
−
p
p
P ( ZT = 0 ) =
≤ 2− log(1/δ) = δ.
b
1−p
−1
p
Thus, w.p. 1 − δ we have (Zt ) > 0. And hence
P ( q ∗ is not returned ) ≤ P ( Xτ ≤ −1 ) ≤ P ( Yτ ≤ 0 )
≤ P ( Zt ≤ 0 ) = P ( ZT = 0 ) ≤ δ,
which proves the correctness w.r.t. to the specification of MaxLex. By Lemma B.2,
executing Lex and Dominates requires O(d log(1/δ2 ) + log(d/δ1 )) = O(d) queries in
expectation. Furthermore, by Observation 3, the total number of iterations is bounded by
10(|S| + 3)(log(1/δ) + 3). This finishes the proof.
B.5
Missing Proofs Section 5
Proof of Lemma 5.1. The proof is by contradiction. Assume that A is an algorithm with
success probability at least 3/4 and worst case number of queries T ≤ (`k log3 k)/1000. We
assume that the adversary is generous, i.e. the adversary tells the truth for every entry (i, j)
such that vji = 0, and that lies with probability 1/3 otherwise.
Generalizing the 2-phase computational model by Feige, Peleg, Raghavan and Upfal [FRPU94], we will give the algorithm more leeway and study a 4-phase computation
model, defined as follows. In the first phase, the algorithm queries every entry vji (log3 k)/100
times. In the second phase, the adversary reveals to the algorithm all remaining hidden
entries (i, j) such that vji = 2, except for a single random one. In the third phase, the
algorithm can strategically and adaptively choose kl/10 entries, and the adversary reveals
their true value at no additional cost. Finally, in phase 4, the algorithm outputs wi = 2 for
every vector where it found an entry equal to 2, and wi = 0 for the rest of the vectors.
To see how the two models are related, observe that since T ≤ (`k log3 k)/20, by Markov’s
inequality at most a set S of `k/10 entries are queried by algorithm A more than (log3 k)/2
times, so at the end of the first phase we have queried every entry at least as many times as
A, except for those `k/10 entries, and in the beginning of the third phase there is all the
necessary information to simulate the execution of A, adaptively finding S (and getting
those values correctly), hence the success probability of the three-phase algorithm is greater
than or equal to the success probability of A. Also observe that, thanks to the definition of
µ and to the generosity of the adversary, any execution where all queries to a vector lead
to 0 answers must lead to an output where wi = 0—else the algorithm would be incorrect
when µ selects the null vector.
We now sketch the analysis of the success probability of the three-phase algorithm.
Due to the definition of µ, with probability at least 9/10 the ground-truth input drawn
27
√
from µk has k/2 ± O( k) vectors that contain an entry equal to 2. At the end of the
first phase, and due the fact that the adversary
√ is generous, we have that at most of them
have been identified. There remain
√ k/2 ± O( k) vectors that appear to be all zeroes, and
about (k/2)(1/3)(log3 k)/2 = (1/2) k of those vectors contain a still-hidden entry whose true
value is 2. During the third phase, all of
√ those hidden 2’s are revealed except for one. At
that point, there still remain k/2 ± O( k) vectors whose entries appear to be all zeroes,
there is a 2 hidden somewhere uniformly at random, but all entries have been queried an
equal number of times, all in vain. To find that remaining hidden entry (and therefore
decide which
√ wi is equal to 2), the algorithm has no information to distinguish between the
`(k/2 ± O( k)) remaining entries. Since, the algorithm may only select `k/10 elements to
query further, the algoirthm’s
√ success probability after the fourth phase cannot be better
than (k`/10)/(`(k/2 ± O( k))) < 1/4, a contradiction.
C
The NoisySearch algorithm
input: a point y, and an ordered list (y1 , y2 , . . . , ym−1 ), accessible by comparisons that each
have error probability at most p, and a parameter δ
output: an interval I = (yi−1 , yi ) or [yi ]
error probability: δ
Here, we recall the algorithm from [FRPU94] whose performance is given in Theorem 2.2.
It is best described by first rephrasing standard binary search to search for y in an ordered
set Y = (y1 , y2 , . . . , yn )
The binary search algorithm can be viewed as a downwards walk from the root in a
search tree whose nodes represent intervals. Letting y0 denote −∞ and yn denote +∞, each
tree node represents an interval [yi , yj ], the root represents (y0 , yn ], each leaf represents
(yi−1 , yi ] for some i, and the left and right child of the node representing (yi , yj ] represent
(yi , yk ] and (yk , yj ] respectively, with k = (i + j)/2. When the walk is at a node representing
(yi , yj ], if the node has two children (yi , yk ] and (yk , yj ] then the algorithm compares y to
yk and proceeds to the left or right child according to the result of the comparison. The
algorithm stops when a leaf is reached, after log2 n steps of the downwards walk, and returns
the interval of the current node.
The noisy search algorithm can be viewed as a biased random walk from the root in a
search tree that extends the noiseless search tree, each leaf (yi−1 , yi ] being the parent of an
infinite chain, each of whose nodes are also labeled (yi−1 , yi ]. When the walk is at a node
representing (yi , yj ], the noisy search algorithm first performs two comparisons to check
whether yi < y ≤ yj . If the answer is negative, then the walk moves to the parent node; if
the answer is positive, then, if the node has only one child then the walk proceeds to the
child node, and if the node has two children (yi , yk ] and (yk , yj ] then the algorithm compares
y to yk and proceeds to the left or right child according to the result of the comparison.
The algorithm stops after c log n/δ steps of the random walk and returns the interval of the
28
current node.
D
The NoisySort algorithm
Here, we recall the algorithm from [FRPU94] whose performance is given in Theorem 2.3.
Algorithm D.1 NoisySort(Y, i, δ) [FRPU94]
input: Set of points Y = (y1 , y2 , . . . , yn ), coordinate i, δ error prob.
output: Sorted list of i’th coordinates of points in Y
error probability: δ
Z←∅
for ` = 1 to n do
3:
(zj−1 , zj ] ← NoisySearch(y`,i , Z, δ/n)
4:
Update list Z by inserting y` between zj−1 and zj in Z
5: Output Z.
1:
2:
E
The SkyGM algorithm
The idea of SkyGM is to sort the points correctly on each dimension and to use any skyline
algorithm in the noiseless setting to deduce the skyline.
Theorem E.1 ( [GM15, Theorem 3] ). The algorithm SkyGM is correct w.r.t. to its
specification and has an expected query complexity of O(dn log(dn/δ)).
Algorithm SkyGM(X, δ)
([GM15, Algorithm 3])
input: X set of points, δ error probability
output: skyline(X)
error probability: δ.
for dimension i ∈ {1, 2, . . . , d} do
2:
Si ← NoisySort(X, i, δ/d)
3: Deduce skyline(X) from S1 , S2 , . . . , Sd and output it.
1:
29
| 8cs.DS
|
Conflict management in information fusion with
belief functions
arXiv:1709.04182v1 [cs.AI] 13 Sep 2017
Arnaud Martin
Abstract In Information fusion, the conflict is an important concept. Indeed, combining several imperfect experts or sources allows conflict. In the theory of belief
functions, this notion has been discussed a lot. The mass appearing on the empty
set during the conjunctive combination rule is generally considered as conflict, but
that is not really a conflict. Some measures of conflict have been proposed and some
approaches have been proposed in order to manage this conflict or to decide with
conflicting mass functions. We recall in this chapter some of them and we propose
a discussion to consider the conflict in information fusion with the theory of belief
functions.
1 Introduction
The theory of belief functions was first introduced by [5] in order to represent some
imprecise probabilities with upper and lower probabilities. Then [30] proposed a
mathematical theory of evidence with is now widely used for information fusion.
Combining imperfect sources of information leads inevitably to conflict. One can
consider that the conflict comes from the non-reliability of the sources or the sources
do not give information on the same observation. In this last case, one must not
combine them.
Let Ω = {ω1 , . . . , ωn } be a frame of discernment of exclusive and exhaustive hypothesis. A mass function m, also called basic belief assignment (bba) is the mapping from elements of the power set 2Ω (composed by all the disjunctions of Ω )
onto [0, 1] such that:
(1)
∑ m(X) = 1.
X∈2Ω
A. Martin
IRISA, University of Rennes 1 e-mail: [email protected]
1
2
Arnaud Martin
A focal element X is an element of 2Ω such that m(X) 6= 0. If the focal elements are
nested, the mass functions is consonant. A simple mass function, noted Aw is given
by:
m(A) = w
(2)
m(Ω ) = 1 − w
This mass function allows to show that we can model an imprecise information (if
A is an union of singletons ωi ) and an uncertain information (if w > 0). All nondogmatic mass functions (with m(Ω ) > 0)) can be decomposed by a set of simple
mass functions [30].
Constraining m(0)
/ = 0 corresponds to a closed-world assumption [30], while
allowing m(0)
/ ≥ 0 corresponds to an open world assumption [32]. Smets interpreted
this mass on the empty set such as an non-expected hypothesis and normalizes it in
the pignistic probability defined for all X ∈ 2Ω , with X 6= 0/ by:
BetP(X) =
|X ∩Y | m(Y )
.
|Y | 1 − m(0)
/
Y ∈2Ω ,Y 6=0/
∑
(3)
The pignistic probability can be used in the decision process such as a compromise
between the credibility and the plausibility. The credibility is given for all X ∈ 2Ω
by:
bel(X) =
∑
m(Y ),
(4)
Y ⊆X,Y 6=0/
The plausibility is given for all X ∈ 2Ω by:
pl(X) =
m(Y ) = bel(Ω ) − bel(X c ) = 1 − m(0)
/ − bel(X c ),
∑
(5)
Y ∈2Ω ,Y ∩X6=0/
where X c is the complementary of X. Hence, if we note the decision function fd
that can be the pignistic probability, the credibility or the plausibility, we choose the
element A ∈ 2Ω for a given observation if:
A = argmax ( fd (X)) .
(6)
X∈Ω
The decision is made on the mass function obtained by the combination of all the
mass function from the sources.
The first combination rule has been proposed by Dempster [5] and is defined for
two mass functions m1 and m2 , for all X ∈ 2Ω , with X 6= 0/ by:
mDS (X) =
1
1−κ
∑
A∩B=X
m1 (A)m2 (B),
(7)
Conflict management in information fusion with belief functions
where κ =
∑
3
m1 (A)m2 (B) is the inconsistence of the combination and generally
A∩B=0/
called conflict. We call it here the global conflict such as the sum of all partial
conflicts.
To stay in an open world, Smets [32] proposes the non-normalized conjunctive
rule given for two mass functions m1 and m2 and for all X ∈ 2Ω by:
mConj (X) =
∑
m1 (A)m2 (B) := (m1
∩
m2 )(X).
(8)
A∩B=X
These both rules allow to reduce the imprecision of the focal elements and to
increase the belief on concordant elements after the fusion. The main assumptions
to apply these rules are the cognitive independence and the reliability of the sources.
Based on the results of these rules, the problem enlightened by the famous
Zadeh’s example [38] is the repartition of the global conflict. Indeed, consider
Ω = {ω1 , ω2 , ω3 } and two experts opinions given by m1 (ω1 ) = 0.9, m1 (ω3 ) = 0.1,
and m2 (ω2 ) = 0.9, m1 (ω3 ) = 0.1, the mass function resulting in the combination
using Dempster’s rule is m(ω3 ) = 1 and using conjunctive rule is m(0)
/ = 0.99,
m(ω3 ) = 0.01. Therefore, several combination rules have been proposed to manage
this global conflict [33, 19].
As observed in [17, 20], the weight of conflict given by κ = mConj (0)
/ is not a
conflict measure between the mass functions. Indeed, the conjunctive-based rules
are not idempotent (as the majority of the rules defined to manage the global conflict): the combination of identical mass functions leads generally to a positive value
of κ. Hence, new kind of conflict measures are defined in [20].
In the following section 2, we recall some measures of conflict in the theory of
belief functions. Then, in section 3 we present the ways to manage the conflict either
before the combination, or in the combination rule. The last section 4 presents some
decision methods in order to consider the conflict during this last step of information
process.
2 Modeling conflict
First of all, we should not mix up conflict measure and contradiction measure. The
measures defined by [16, 34] are not conflict measures, but some discord and specificity measures (to take the terms of [15]) we call contradiction measures. We define
the contradiction and conflict measures by the following definitions:
Definition A contradiction in the theory of belief functions quantifies how a mass
function contradicts itself.
Definition (C1) The conflict in the theory of belief functions can be defined by
the contradiction between two or more mass functions.
Therefore, is the mass of the empty set or the functions of this mass (such as
− ln(1 − mConj (0))
/ proposed by [30]) a conflict measure? It seems obvious that the
property of the non-idempotence is a problem to use this as a conflict measure.
4
Arnaud Martin
However, if we define a conflict measure such as Conf(m1 , m2 ) = mConj (0),
/ we note
that Conf(m1 , mΩ ) = 0 where mΩ (Ω ) = 1 is the ignorance. Indeed, the ignorance is
the neutral element for the conjunctive combination rule. This property seems to be
followed from a conflict measure.
Other conflict measures have been defined. In [14], a conflict measure is given
by:
plT1 .pl2
(9)
Conf(m1 , m2 ) = 1 −
kpl1 kkpl2 k
where pl is the plausibility function and plT1 .pl2 the vector product in 2n space
of both plausibility functions. However, generally Conf(m1 , mΩ ) 6= 0, that seems
counter-intuitive.
Auto-conflict
Introduced by [28], the auto-conflict of order s for one source is given by:
!
s
as =
∩
m (0).
/
(10)
j=1
where ∩ is the conjunctive operator of Equation (8). The following property holds:
as ≤ as+1 , meaning that due to the non-indempotence of ∩ , the more masses m are
combined with itself the nearer to 1 κ is, and so in a general case, the more the
number of sources is high the nearer to 1 κ is. The behavior of the auto-conflict was
studied in [20] and it was shown that we should take into account the auto-conflict
in the global conflict in order to really define a conflict. In [37], the auto-conflict
was defined and called the plausibility of the belief structure with itself. The autoconflict is a kind of measure of the contradiction, but depends on the order s of the
combination. A measure of contradiction independent on the order has been defined
in [31].
Conflict measure based on a distance
With the definition of the conflict (C1), we consider sources to be in conflict
if their opinions are far from each other in the space of corresponding bba’s. That
suggests a notion of distance. That is the reason why in [20], we give a definition of
the measure of conflict between sources assertions through a distance between their
respective bba’s. The conflict measure between 2 experts is defined by:
Conf(1, 2) = d(m1 , m2 ).
(11)
Conflict management in information fusion with belief functions
5
We defined the conflict measure between one source j and the other M − 1 sources
by:
Conf( j, E ) =
M
1
Conf(i, j),
∑
M − 1 i=1,i6= j
(12)
where E = {1, . . . , M} is the set of sources in conflict with j. Another definition is
given by:
Conf( j, M) = d(m j , mM ),
(13)
where mM is the bba of the artificial source representing the combined opinions of
all the sources in E except j.
A comparison of distances in the theory of belief functions is presented in [14].
We consider the distance defined in [13] as the most appropriate. This distance is
defined for two basic belief assignments m1 and m2 on 2Ω by:
r
1
(m1 − m2 )T D(m1 − m2 ),
dJ (m1 , m2 ) =
(14)
2
where D is an 2|Ω | × 2|Ω | matrix based on Jaccard dissimilarity whose elements are:
1, if A = B = 0,
/
(15)
D(A, B) =
|A ∩ B|
, ∀A, B ∈ 2Ω .
|A ∪ B|
An interesting property of this measure is given by Conf(m, m) = 0. That means
that there is no conflict between a source and itself (that is not a contradiction).
However, we generally do not have Conf(m, mΩ ) = 0, where mΩ (Ω ) = 1 is the
ignorance.
Conflict measure based on inclusion degree and distance
We have seen that we cannot use the mass on the empty set as a conflict measure
because of the non-idempotence of the conjunctive rule. We also have seen that the
conflict measure based on the distance is not null in general for the ignorance mass.
The conjunctive rule does not transfer mass on the empty set if the mass functions
are included. We give here some definitions of the inclusion.
Definition 1: strict inclusion
We say that the mass function m1 is included in m2 if all the focal elements of m1
are included in each focal elements of m2 .
6
Arnaud Martin
Definition 2: light inclusion
We say that the mass function m1 is included in m2 if all the focal elements of m1
are included in at least one focal element of m2 .
Definition We note this inclusion by m1 ⊆ m2 . The mass functions are included
if m1 is included in m2 or m2 is included in m1 .
In [25], we propose a conflict measure base on five following axioms. Let note
Conf(m1 , m2 ) a conflict measure between the mass functions m1 and m2 . We present
hereafter essential properties that must verify a conflict measure.
1. Non-negativity:
Conf(m1 , m2 ) ≥ 0
(16)
A negative conflict does not make sense. This axiom is,therefore, necessary.
2. Identity:
Conf(m1 , m1 ) = 0
(17)
Two equal mass functions are not in conflict. This property is not reached by the
global conflict, but seems natural.
3. Symmetry:
Conf(m1 , m2 ) = Conf(m2 , m1 )
(18)
The conflict measure must be symmetric. We do not see any case where the nonsymmetry can make sense.
4. Normalization:
0 ≤ Conf(m1 , m2 ) ≤ 1
(19)
This axiom is may not be necessary to define a conflict measure, but the normalization is very useful in many applications requiring a conflict measure.
5. Inclusion:
Conf(m1 , m2 ) = 0, if and only if m1 ⊆ m2 or m2 ⊆ m1
(20)
This axiom means that if the focal elements of two mass functions are not conflicting (the intersection is never empty), the mass functions are not in conflict
and the mass functions cannot be in conflict if they are included. This axiom is
not satisfied by a distance based conflict measure.
These proposed axioms are very similar to ones defined in [7]. If a conflict measure
satisfied these axioms that is not necessary a distance. Indeed, we only impose the
identity and not the definiteness (Conf(m1 , m2 ) = 0 ⇔ m1 = m2 ).
The axiom of inclusion is less restrictive and makes more senss for a conflict measure.
Moreover,
we
do
not
impose
the
triangle
inequality
(Conf(m1 , m2 ) ≤ Conf(m1 , m3 ) + Conf(m3 , m2 )). It can be interesting to have
Conf(m1 , m2 ) ≥ Conf(m1 , m3 ) + Conf(m3 , m2 ) meaning that an expert given the
mass function m3 can reduce the conflict. Therefore, a distance (with the property
of the triangle inequality) cannot be used directly to define a conflict measure.
We see that the axiom of inclusion seems very important to define a conflict measure. This is the reason why we define in [25] a degree of inclusion to measure how
Conflict management in information fusion with belief functions
7
two mass functions are included. Let the inclusion index: Inc(X1 ,Y2 ) = 1 if X1 ⊆ Y2
and 0 otherwise, where X1 and Y2 are two focal elements of m1 and m2 respectively.
According to the definition 1 and definition 2, we introduce two degrees of inclusion
of m1 in m2 . A strict degree of inclusion of m1 in m2 is given by:
dincS (m1 , m2 ) =
1
|F1 ||F2 | X ∑
∈F
∑
Inc(X1 ,Y2 )
(21)
1 Y2 ∈F2
1
where F1 and F2 are the set of focal elements of m1 and m2 respectively, and |F1 |,
|F2 | are the number of focal elements of m1 and m2 .
This definition is very strict, so we introduce a light degree of inclusion of m1 in
m2 given by:
1
max Inc(X1 ,Y2 ).
(22)
dincL (m1 , m2 ) =
|F1 | X ∑
∈F Y2 ∈F2
1
1
Let δinc (m1 , m2 ) a degree of inclusion of m1 and m2 define by:
δinc (m1 , m2 ) = max(dinc (m1 , m2 ), dinc (m2 , m1 ))
(23)
This degree gives the maximum of the proportion of focal elements from one mass
function included in another one. Therefore, δinc (m1 , m2 ) = 1 if and only if m1 and
m2 are included, and the axiom of inclusion is reached for 1 − δinc (m1 , m2 ).
Hence, we define in [25], a conflict measure between two mass functions m1 and
m2 by:
Conf(m1 , m2 ) = (1 − δinc (m1 , m2 ))dJ (m1 , m2 )
(24)
where dJ is the distance defined by the equation (14). All the previous axioms are
satisfied. Indeed the axiom of inclusion is 1 − δinc (m1 , m2 ) and the distance dJ satisfied the other axioms. Moreover 0 ≤ δinc (m1 , m2 ) ≤ 1, by the product of 1 − δinc
and dJ , all the axioms are satisfied.
For more than two mass functions, the conflict measure between one source j
and the other M − 1 sources can be defined from equations (12) or (13).
3 Managing conflict
The role of conflict is essential in information fusion. Different ways can be used
to manage and reduce the conflict. The conflict can come from the low reliability
of the sources. Therefore, we can use this conflict to estimate the reliability of the
sources if we cannot learn it on databases as proposed in [20]. Hence, we reduce
the conflict before the combination, but we can also directly manage the conflict in
the rule of combination as generally made in the theory of belief functions such as
explained in [33, 19].
According to the application, we do not search always to reduce the conflict.
For example, we can use the conflict measure such as an indicator of the inconsis-
8
Arnaud Martin
tence of the fusion for example in databases [2]. Conflict information can also be an
interesting information in some applications such as presented in [29].
3.1 Managing the conflict before the combination
The conflict appearing while confronting several experts’ opinions can be used as
an indicator of the relative reliability of the experts. We have seen that there exist
many rules in order to take into account the conflict during the combination step.
These rules do not make the difference between the conflict (global or local conflict)
and the auto-conflict due to the non-idempotence of the majority of the rules. We
propose here the use of a conflict measure in order to define a reliability measure,
that we consider before the combination, in a discounting procedure.
When we can quantify the reliability of each source, we can weaken the basic
belief assignment before the combination by the discounting procedure:
α
m j (X) = α j m j (X), ∀X ∈ 2Ω r {Ω }
(25)
mαj (Ω ) = 1 − α j (1 − m j (Ω )).
α j ∈ [0, 1] is the discounting factor of the source j that is, in this case, the reliability
of the source j, eventually as a function of X ∈ 2Ω .
Other discounting procedures are possible such as the contextual discounting
[26], or a discounting procedure based on the credibility or the plausibility functions
[39].
According to the applications, we can learn the discounting factors α j , for example, from the confusion matrix [22]. In many applications, we cannot learn the
reliability of each source. A general approach to the evaluation of the discounting
factor without learning is given in [10]. For a given bba the discounting factor is
obtained by the minimization on α of a distance given by:
Dist α j =
∑
BetP j (A) − δA, j
2
,
(26)
A∈Ω
where BetP j is the pignistic probability (Equation (3)) of the bba given by the source
j and δA, j = 1 if the source j supports A and 0 otherwise.
This approach is interesting with the goal of making decision based on pignistic probabilities. However, if the source j does not support a singleton of Ω , the
minimization on α j does not work well.
In order to combine the bbas of all sources together, we propose here to estimate
the reliability of each source j from the conflict measure Con f between the source
j and the others by:
α j = f (Conf( j, M)),
where f is a decreasing function. We can choose:
(27)
Conflict management in information fusion with belief functions
9
1/λ
α j = 1 − Conf( j, M)λ
,
(28)
where λ > 0. We illustrate this function for λ = 2 and
λ = 1/2 on figure 1. This function allows to give more reliability to the sources
with few conflict with the other.
Fig. 1 Reliability of one source based on conflict of the source with the other sources.
Other definitions are possible. The credibility degree defined in [4] is also based
on the distance given in the Equation (14), and could also be interpreted as the
reliability of the source. However the credibility degree in [4] is integrated directly
in the combination with a weighted average. Our reliability measure allows the use
of all the existing combination rules.
3.2 Managing the conflict in the combination
According to the application, if we cannot reduce the conflict before the combination, we can do it by incorporating it into a combination rule. The choice of the
combination rule is not easy, but it can be done by utilizing the conflict and the
assumption on its origin. Indeed the Dempster rule can be apply if the sources are
independent and reliable. Dempster’s rule is given for S sources for all X ∈ 2Ω ,
X 6= 0/ by:
mDS (X) =
S
mConj (X)
1
m j (Y j ) =
,
∑
∏
1 − mConj (0)
/ Y1 ∩...∩Ys =X j=1
1−κ
(29)
where κ = mConj (0),
/ Y j ∈ 2Ω is the answer of the source S j , and m j (Y j ) the associated mass function. This normalization by 1 − mConj (0)
/ hide the conflict and so this
10
Arnaud Martin
rule is interesting only if we consider the closed-world and if the sources are not
highly conflicting.
If the assumption of independent and reliable sources are not reached, the application of the Dempster’s rule can produce some global conflict.
Conflict coming from a false assumption of closed world
In the closed world, the elements of the frame of discernment are assumed exhaustive. If m(0)
/ > 0, this mass can be interpreted such as another element, and so
the assumption on the exhaustiveness of the frame of discernment is false. Hence,
Smets [33] proposed the use of the conjunctive rule given for S sources for all
X ∈ 2Ω by:
S
mConj (X) =
∑
∏ m j (Y j ).
(30)
Y1 ∩...∩Ys =X j=1
Here, the sources must be cognitively independent and reliable while the open-world
is considered. Hence, the mass on the empty set can be interpreted as another element unknown by the sources. In fact, in the proposed model by Smets, the conflict
is transfered during the decision step by the pignistic probability [32], hiding the
conflict to the end. This rule cannot be used in applications with a high value of κ.
The global conflict come from the sum of the partial conflict. Hence, if the global
conflict can be interpreted as an unknown element, all the partial conflict can be
interpreted as many unknown elements. In that case we can keep the partial conflict
in order to decide on these elements (see section 4 for this consideration). Therefore
the assumption of the exclusivity is considered here as false.
Under this assumption, the mass functions are no more defined on the power set
2Ω but on the so called hyper power set1 DΩ . Therefore the space Ω is closed by
the union and intersection operators. This extension of the power set, lead to a lot
of reflexions around this new expressiveness taking the name of DSmT (DempsterShafer modified Theory).
One can also consider a partial exclusiveness of the frame of discernment. Hence,
we introduce the notation DΩ
r in [23] in order to integrate some constraints on the
exclusiveness of some elements of Ω and so reduce the hyper power set size. Under
these assumptions, we define the PCR6 rule in [18], given by:
1
This notation is introduced by [8] and D come from the Dedeking lattice.
Conflict management in information fusion with belief functions
11
mPCR6 (X) = mConj (X)
s−1
∏ mσ j ( j0 ) (Yσ j ( j0 ) )
j0 =1
, (31)
s−1
m j (X)+ ∑ mσ j ( j0 ) (Yσ j ( j0 ) )
S
+
2
∑ m j (X)
j=1
∑
S−1
∩ Yσ ( j0 ) ∩X=0/
j0 =1 j
s−1
(Yσ (1) ,...,Yσ (S−1) )∈(DΩ
r )
j
j0 =1
j
where σ is given by:
σ j ( j0 ) = j0
if j0 < j,
0
0
σ j ( j ) = j + 1 if j0 ≥ j,
(32)
This rule transfers the partial conflicts on the elements that generate it, proportionally to their masses. This rule has been used in many applications allowing for good
results.
Conflict coming from the assumption of source’s independence
If we consider some dependent sources, the conjunctive rule cannot be used. If we
want to combine mass functions coming from dependent sources, the combination
rule has to be idempotent. Indeed, if we combine two identical dependent mass
functions (coming from different dependent sources), we have to obtain the same
mass function without any global conflict.
The simplest way to obtain a non-idempotent rule is the average of the mass
functions such given in [27] by:
mMean (X) =
1
s
S
∑ m j (Y j ).
(33)
j=1
We showed the interest in a such rule in [28], but in that case the sources have to be
assumed totally reliable. If the sources give high conflicting information, this rule
can provide some errors in the decision process.
The cautious rule [6] could be used to combine mass functions for which independence assumption is not verified. Cautious combination of S non-dogmatic
mass functions m j , j = 1, 2, · · · , S is defined by the bba with the following weight
function:
S
w(A) = ∧ w j (A), A ∈ 2Ω \ Ω .
(34)
j=1
We thus have
S
mCautious (X) =
∩
A(Ω
A
∧ w j (A)
j=1
,
(35)
12
Arnaud Martin
where Aw j (A) is the simple support function focused on A with weight function
w j (A) issued from the canonical decomposition of m j . Note also that ∧ is the min
operator.
When the dependence/independence of the sources is estimated, another rule was
proposed in [3].
Conflict coming from source’s ignorance
Another possible interpretation of the reason for the conflict is the ignorance of
the sources. Indeed, if a source is highly ignorant, it should give a categorical mass
function on Ω .
Therefore, [36] interprets the global conflict coming from the ignorance of the
sources and transfers the mass on the total ignorance (i.e. on Ω ) in order to keep the
closed world assumption. In the case of high conflict, the result of the fusion is the
ignorance. This rule is given by:
mY (X) = mConj (X), ∀X ∈ 2Ω , X 6= 0,
/ X 6= Ω
mY (Ω ) = mConj (Ω ) + mConj (0)
/
mY (0)
/ = 0.
(36)
A source can also be ignorant not on all but only on some focal elements. Hence,
[9] proposed a clever conflict repartition by transferring the partial conflicts on the
partial ignorances. This rule is given for all X ∈ 2Ω , X 6= 0/ by:
S
mDP (X) =
∑
S
∏ m j (Y j ) +
Y1 ∩...∩Ys =X j=1
∑
Y1 ∪ . . . ∪Ys = X
Y1 ∩ . . . ∩Ys = 0/
∏ m j (Y j ),
(37)
j=1
where Y j ∈ 2Ω is a focal element of the source S j , and m j (Y j ) the associated mass
function. This rule has a high memory complexity, such as the PCR6 rule, because
it is necessary to manage the partial conflict.
Conflict coming from source reliability assumption
If we have no knowledge of the reliability of the sources, but we know that at
least on source is reliable, the disjunctive combination can be used. It is given for
all X ∈ 2Ω by:
S
mDis (X) =
∑
∏ m j (Y j ).
Y1 ∪...∪Ys =X j=1
(38)
Conflict management in information fusion with belief functions
13
The main problem of this rule is the lost of specificity after combination.
One can also see the global conflict κ = mConj (0)
/ such as an estimation of the
conflict coming from the unreliability of the sources. Therefore, the global conflict
can play the role of a weight between a conjunctive and disjunctive comportment of
the rule such introduced by [12]. This rule is given for X ∈ 2Ω , X 6= 0/ by:
mFlo (X) = β1 (κ)mDis (X) + β2 (κ)mConj (X),
(39)
where β1 and β2 can be defined by:
κ
,
1 − κ + κ2
1−κ
β2 (κ) =
.
1 − κ + κ2
β1 (κ) =
(40)
In a more general way, we propose in [19] to regulate the conjunctive/disjunctive
comportment taking into consideration the partial combinations. The mixed rule is
given for m1 and m2 for all X ∈ 2Ω by:
mMix (X) =
δ1 m1 (Y1 )m2 (Y2 )
∑
∑
Y1 ∪Y2 =X
+
δ2 m1 (Y1 )m2 (Y2 ).
(41)
Y1 ∩Y2 =X
If δ1 = β1 (κ) and δ2 = β2 (κ) we obtain the rule of [12]. Likewise, if
δ1 = 1 − δ2 = 0 we obtain the conjunctive rule, and if δ1 = 1 − δ2 = 1 the disjunctive rule. With δ1 (Y1 ,Y2 ) = 1 − δ2 (Y1 ,Y2 ) = 1l{0}
/ (Y1 ∩ Y2 ) we get back to the
rule of [9] by taking into account partial conflicts.
The choice of δ1 = 1 − δ2 can also be made from a dissimilarity such as:
δ2 (Y1 ,Y2 ) =
|Y1 ∩Y2 |
,
min(|Y1 |, |Y2 |)
(42)
where |Y1 | is the cardinality of Y1 . Jaccard dissimilarity can also be considered by:
δ2 (Y1 ,Y2 ) =
|Y1 ∩Y2 |
.
|Y1 ∪Y2 |
(43)
Therefore, if we have a partial conflict between Y1 and Y2 , |Y1 ∩Y2 | = 0 and the rule
transfers the mass on Y1 ∪Y2 . In that case Y1 ⊂ Y2 (or the contrary), Y1 ∩Y2 = Y1 and
Y1 ∪Y2 = Y2 , hence with δ2 given by (42) the rule transfers the mass on Y1 and with δ2
given by (43) on Y1 and Y2 according to the ratio |Y1 |/|Y2 | of cardinalities. In the case
Y1 ∩Y2 6= Y1 ,Y2 and 0,
/ the rule transfers the mass on Y1 ∩Y2 and Y1 ∪Y2 according to
equations (42) and (43). With such weights, the Mix rule considers partial conflict
according to the imprecision of the elements at the origin of the conflict.
14
Arnaud Martin
Conflict coming from a number of sources
When we have to combine a many sources, the assumption of the reliability of all
the sources is difficult to consider especially if the sources are human. The disjunctive rule (38) assume that at least one source is reliable but a precise decision will
be difficult to take. Moreover, the complexity of main rules managing the conflict in
a clever way is too high such as the rules given by (41) and (31). That is the reason
why we introduce in [40] a new rule according to the following assumptions:
• The majority of sources are reliable;
• The larger extent one source is consistent with others, the more reliable the source
is;
• The sources are cognitively independent.
w
For each mass function m j we consider the set of mass functions {Ak j , Ak ⊂ Ω }
w
coming from the canonical decomposition. If group the simple mass functions Ak j
in c clusters (the number of distinct Ak ) and denote by sk the number of simple mass
functions in the cluster k, the proposed rule is given by:
sk
1−αk +αk
mLNS =
∩
∏ wj
j=1
(Ak )
(44)
k=1,··· ,c
where
αk =
sk
c
.
(45)
∑ si
i=1
How to choose the combination rule?
To answer the delicate question on which combination rule to choose, many authors propose a new rule. Of course, we could propose a no free lunch theorem
showing that there is no a best combination rule.
To answer this question, we propose in [24] a global approach to transfer the belief. Indeed, the discounting process, the reduction of the number of focal elements,
the combination rules and the decision process can be seen such as a transfer of belief and we can define these transfers in joint operator. However, it seems to difficult
to propose a global approach which will be too general to be applied. In [24], we
define a rule integrating only the reliability given for X ∈ 2Ω by:
S
m(X)=
∑ ∏ m j (Y j )w(X, m(Y), T (Y), α(Y)),
Y∈(2Ω )S j=1
(46)
Conflict management in information fusion with belief functions
15
where Y = (Y1 , . . . ,YS ) is the response vector of the S sources, m j (Y j ) the associated
masses (m(Y) = (m1 (Y1 ), . . . , m j (Ys )), w is a weight function, α is the matrix of
terms αi j of the reliability of the source S j for the element i of 2Ω , and T (Y) is
the set of elements of 2Ω , on which we transfer the elementary masses m j (Y j ) for a
given vector Y. This rule has been illustrated in a special case integrating the local
reliability, but it seems even too complex to be easily applied.
Hence, the best way to choose the combination rule is to identify the assumptions
that we can or we have to make and choose the adapted rule according to these
assumptions.
However, we know that a given rule can provide good results in a context where
the assumptions are satisfy this rule. Hence, another way to evaluate and compare
some rules of combination, is to study the results (after decision) of the combined
masses, e.g. on generated mass functions. In [28], from generated mass functions,
we study the difference of the combination rules in terms of decisions. We showed
that we have to take into account the decision process. We will present some of them
in the next section in the context of conflicting mass functions.
4 Decision with conflicting bbas
The classic functions used for decision making such as the pignistic probability,
the credibility and the plausibility are increasing by sets inclusion. We cannot use
these functions directly to decide on other elements than the singletons. When the
assumption of exclusiveness of the frame of discernment is not made, such as in the
equation (31), we can decide on DΩ
r . That can be interesting from the data mining
point of view such as we show in [21].
The approach proposed by [1] has been extend to the consideration of DΩ
r in [23]
allowing to decide on any element of DΩ
r by taking into account the mass function
and the cardinality. Hence, we choose the element A ∈ DΩ
r for a given observation
if:
A = argmax (md (X) fd (X)) ,
(47)
X∈DΩ
r
where fd is the considered decision function such as the credibility, plausibility or
pignistic probability calculated from the mass function coming from the result of
the combination rule, and md is the mass function defined by:
1
,
(48)
md (X) = Kd λX
CM (X)ρ
CM (X) is the cardinality of X of DΩ
r , defined by the number of disjoint parts in
the Venn diagram, ρ is a parameter with its values in [0, 1] allowing to decide from
the intersection of all the singletons (with ρ = 1) until the total ignorance Ω (with
ρ = 0). The parameter λX enables us to integrate the lost of knowledge on one of
16
Arnaud Martin
the elements X of DΩ
r . The constant Kd is a normalization factor that guaranties
the condition of equation (1). Without any constraint on DΩ , all the focal elements
contain the intersection of the singletons. One cannot choose the plausibility such
as decision function fd .
The choice of the parameter ρ is not easy to make. It depends on the size of Ω .
According to the application, it can be more interesting to define a subset on which
we want to take the decision. Hence, we can envisage the decision on any subset of
DΩ
r , noted D, and equation (47) becomes simply:
A = argmax (md (X) fd (X)) .
(49)
X∈D
Particularly this subset can be defined according to the expected cardinality of the
element on which we want to decide.
With the same spirit, in [11], another decision process is proposed by:
A = argmax (dJ (m, mX )) ,
(50)
X∈D
where mX is the categorical mass function m(X) = 1, and m is the mass function
coming from the combination rule. The subset D is the set of elements on which we
want to decide.
This last decision process allows also a decision on imprecise elements of the
power set 2Ω and to control the precision of expected decision element without any
parameter to fit.
5 Conclusion
In this chapter, we propose some solutions to the problem of the conflict in information fusion in the context of the theory of belief functions. In section 2 we
present some conflict measures. Today, there is no consensus in the community on
the choice of a conflict measure. Measuring the conflict is not an easy task because a
mass function contains some information such as auto-conflict, we can interpret differently. The proposed axioms are a minimum that a conflict measure has to reach.
In section 3, we discuss how to manage the conflict. Based on the assumption that
conflict comes from the unreliability of the sources, with a conflict estimation for
each source, the best to do is to discount the mass function according to the reliability estimation (and so the conflict measure).
Another way to manage the conflict, is the choice of the combination rule. Starting from the famous Zadeh’s example, many combination rules have been proposed
to manage the conflict. In this chapter, we present some combination rules (without exhaustiveness) according to the assumptions that the rules suppose. Hence, we
distinguish the assumptions of: open/closed world, dependent/independent sources,
ignorant/not-ignorant sources, reliable/unreliable sources, few/many sources.
Conflict management in information fusion with belief functions
17
To end this chapter, when the assumption of exclusiveness of the frame of discernment is not make, and so when we postpone the matter of conflict to the decision, we present some adapted decision processes. These decision methods are also
adapted to decide on some imprecise elements of the power set.
Of course, all the exposed methods here must be selected according to the application, to the possible assumptions, and to the final expected result.
References
1. Appriou, A. Uncertainty Theories and Multisensor Data Fusion, Wiley-IEEE Press (2014)
2. Chebbah, M., Ben Yaghlane, B., Martin, A. Reliability estimation based on conflict for evidential database enrichment, Belief, Brest, France (2010)
3. Chebbah, M., Martin, A., Ben Yaghlane, B., Combining partially independent belief functions
Decision Support Systems, vol. 73, pp. 37-46 (2015)
4. L.Z. Chen, W.K. Shi, Y. Deng and Z.F. Zhu, A new fusion approach based on distance of
evidences, Journal of Zhejian University Science, vol. 6A, no. 5, pp. 476-482, (2005)
5. Dempster, A.P. Upper and Lower probabilities induced by a multivalued mapping, Annals of
Mathematical Statistics, vol. 83, pp. 325-339, (1967)
6. T. Denœux, Conjunctive and disjunctive combination of belief functions induced by nondistinct
bodies of evidence, Artificial Intelligence, vol 172, pp. 234-264, (2008)
7. Destercke, S. and Burger, T., Revisiting the notion of conflicting belief functions, Belief,
Compiègne, France, 2012)
8. Dezert, J. Foundations for a new theory of plausible and paradoxical reasoning Information &;
Security. An International Journal, vol. 9, pp. 13-57, (2002)
9. Dubois, D., Prade, H., Representation and combination of uncertainty with belief functions and
possibility measures Computational Intelligence, vol. 4, pp. 244-264, (1988)
10. Z. Elouedi, K. Mellouli and Ph. Smets, Assessing Sensor Reliability for Multisensor Data Fusion within the Transferable Belief Model, IEEE Transactions on Systems; man, and Cybernetics
- Part B: Cybernetics, vol. 34, no. 1, pp. 782-787, (2004)
11. Essaid, A. and Martin, A. and Smits, G., and Ben Yaghlane, B., A Distance-Based Decision in
the Credal Level International Conference on Artificial Intelligence and Symbolic Computation
(AISC 2014), pp. 147-156 (2014)
12. Florea, M. C., Jousselme, A.-L., Bossé, E., Grenier, D., Robust combination rules for evidence
theory Information Fusion, vol. 10, pp. 183-197, (2009)
13. Jousselme, A.-L., Grenier D. and Bossé, E. A new distance between two bodies of evidence,
Information Fusion, vol. 2, pp. 91-101, (2001)
14. Jousselme, A.-L. and Maupin, P. Distances in evidence theory: Comprehensive survey and
generalizations, International Journal of Approximate Reasoning, (2011)
15. Klir, G.J. Measures of uncertainty in the Dempster-Shafer theory of evidence, Advances in
the Dempster-Shafer Theory of Evidence, John Wiley and Sons, New York, R.R. Yager and M.
Fedrizzi and J. Kacprzyk edition, pp. 35-49, (1994)
16. George, T. and Pal, N.R. Quantification of conflict in Dempster-Shafer framework: a new
approach, International Journal of General Systems, vol. 24(4), pp. 407-423, (1996)
17. Liu, W. Analyzing the degree of conflict among belief functions, Artificial Intelligence,
vol. 170, pp. 909-924, (2006)
18. Martin, A. Osswald, C. Human Experts Fusion for Image Classification Information and Security: An International Journal, Special issue on Fusing Uncertain, Imprecise and Paradoxist
Information (DSmT), vol. 20, pp. 122-143, (2006)
19. Martin, A. and Osswald, C. Toward a combination rule to deal with partial conflict and
specificity in belief functions theory, International Conference on Information Fusion, Québec,
Canada, (2007)
18
Arnaud Martin
20. Martin, A., Jousselme, A.-L. and Osswald, C. Conflict measure for the discounting operation
on belief functions, International Conference on Information Fusion, Cologne, Germany, (2008)
21. Martin, A. and Quidu, I., Decision Support with Belief Functions Theory for Seabed Characterization International Conference on Information Fusion, Cologne, Germany, (2008)
22. A. Martin, Comparative study of information fusion methods for sonar images classification,
International Conference on Information Fusion, Philadelphia, USA, (2005)
23. Martin, A., Implementing general belief function framework with a practical codification for
low complexity, chapter 7, Advances and Applications of DSmT for Information Fusion, American Research Press Rehoboth, vol. 3, pp. 217-274, (2009)
24. Martin, A. Reliability and combination rule in the theory of belief functions International
Conference on Information Fusion, pp. 529-536 (2009)
25. A. Martin, About conflict in the theory of belief functions, Belief, Compiègne, France, (2012)
26. D. Mercier, B. Quost and T. Denœux, Refined modeling of sensor reliability in the belief function framework using contextual discounting, vol. 9, pp. 246-258, Information Fusion, (2006)
27. Murphy, C., Combining belief functions when evidence conflicts Decision Support Systems,
vol. 29, pp. 1-9 (2000)
28. Osswald, C. and Martin, A. Understanding the large family of Dempster-Shafer theory’s fusion operators - a decision-based measure, International Conference on Information Fusion, Florence, Italy, (2006)
29. Rominger, C. and Martin, A. Using the conflict: An application to sonar image registration,
Belief, Brest, France, (2010)
30. Shafer, G., A mathematical theory of evidence. Location: Princeton University Press, (1976)
31. Smarandache F., Martin A., and Osswald, C., Contradiction measures and specificity degrees
of basic belief assignments, International Conference on Information Fusion, Boston, USA,
(2011)
32. Smets, Ph., Constructing the pignistic probability function in a context of uncertainty, Uncertainty in Artificial Intelligence, vol. 5, pp. 29-39, (1990)
33. Smets, Ph., Analyzing the combination of conflicting belief functions, Information Fusion,
vol. 8, no. 4, pp. 387-412, (2007)
34. Wierman, M.J., Measuring Conflict in Evidence Theory, IFSA World Congress and 20th
NAFIPS International Conference, vol. 3(21), pp. 1741-1745, (2001)
35. Yager, R.R. Entropy and Specificity in a Mathematical Theory of Evidence, International Journal of General Systems, vol. 9, pp. 249-260, (1983)
36. Yager, R. R. On the Dempster-Shafer Framework and New Combination Rules Informations
Sciences, vol. 41, pp. 93-137 (1991)
37. Yager, R.R. On Considerations of Credibility of Evidence International, Journal of Approximate Reasoning, vol. 7, 45-72, (1992)
38. Zadeh, L.A. A mathematical theory of evidence (book review), AI magazine, vol. 5, 81-83,
(1984)
39. C. Zeng and P. Wu, A reliability discounting strategy based on plausibility function of evidence, International Conference on Information Fusion, Québec, Canada, (2007)
40. Zhou, K. and Martin, A. and Pan, Q., Evidence combination for a large number of sources,
submitted to International Conference on Information Fusion, Xian, China, (2017)
| 2cs.AI
|
SIM-TO-REAL OPTIMIZATION OF COMPLEX REAL WORLD
MOBILE NETWORK WITH IMPERFECT INFORMATION VIA
DEEP REINFORCEMENT LEARNING FROM SELF-PLAY
Yongxi Tan1,Jin Yang1,Xin Chen2,Qitao Song2,Yunjun Chen2,Zhangxiang Ye2,Zhenqiang Su1
1 Futurewei Tech. Inc., NJ Research Center, USA
{yongxi.tan,jin.yang,zhenqiang.su}@huawei.com
2 Huawei Technologies Co. Ltd, P.R.China
{chenxin,songqitao,chenyunjun,yezhangxiang}@huawei.com
ABSTRACT
Mobile network that millions of people use every day is one of the most
complex systems in real world. Optimization of mobile network to meet
exploding customer demand and reduce CAPEX/OPEX poses greater challenges
than in prior works. Actually, learning to solve complex problems in real world
to benefit everyone and make the world better has long been ultimate goal of AI.
However, it still remains an unsolved problem for deep reinforcement learning
(DRL), given incomplete/imperfect information in real world, huge state/action
space, lots of data needed for training, associated time/cost, interactions among
multi-agents, potential negative impact to real world, etc. To bridge this reality
gap, we proposed a DRL framework to direct transfer optimal policy learned
from multi-tasks in source domain to unseen similar tasks in target domain
without any further training in both domains. First, we distilled temporal-spatial
relationships between cells and mobile users to scalable 3D image-like tensor to
best characterize partially observed mobile network. Second, inspired by
AlphaGo, we used a novel self-play mechanism to empower DRL agent to
gradually improve its intelligence by competing for best record on multiple
tasks. Third, a decentralized DRL method is proposed to coordinate multi-agents
to compete and cooperate as a team to maximize global reward and minimize
potential negative impact. Using 7693 unseen test tasks over 160 unseen
simulated mobile networks and 6 field trials over 4 commercial mobile networks
in real world, we demonstrated the capability of our approach to direct transfer
the learning from one simulator to another simulator, and from simulation to real
world. This is the first time that a DRL agent successfully transfers its learning
directly from simulation to very complex real world problems with incomplete
and imperfect information, huge state/action space and multi-agent interactions.
1. INTRODUCTION
Using deep neural network (LeCun et al., 2015) for a rich representation of high-dimensional
visual input and as an universal function approximator, deep reinforcement learning (DRL) have
achieved unprecedented success in some challenging domains, such as Atari game (Mnih et al.,
2015), Go (Silver et al., 2016; Silver et al., 2017). The ultimate goal of AI is creating agent that
can not only learn like human, but also make the world better by solving complex problems in real
world. However, application of DRL in complex real world problems still remains an unsolved
problem due to imperfect information (Moravčík et al., 2017; Brown & Sandholm 2018), huge
state/action space, big gap between simulation and real world (Rusu et al., 2016; Tobin et al.,
2017; Bousmalis & Levine 2017), multi-agent interactions (Vinyals et al., 2017), time/cost, etc.
In this work, we use DRL for one-shot optimization of real world mobile network that millions
of people use every day. Coverage & capacity optimization (CCO) of mobile network is crucial
for mobile carrier to meet exploding customer demand and reduce CAPEX/OPEX (Fan et al.,
2014), e.g., $11B CAPEX for Verizon in 2016 (Celentano, 2016), Cisco acquired Intucell for
$475M (Marketwired, 2013). However, it poses much more difficult challenges than in prior
works. First, mobile network is one of the most complex systems in real world since it is a multiusers, multi-cells (Macro, Small), multi-technologies (3G, 4G, 5G) heterogeneous network, in
!1
!
!
Figure 1: DRL framework
!
Figure 2: Distill tensor
!
Figure 3: Self play
Figure 4: Cooperate
which mobile services (app, video, IOT) are consumed by billions of devices and many resource
management decisions need to be made to provide seamless services. Second, it is critical to take
actions only once (one-shot) since CCO involves time-consuming (1-3 days) and costly site
visits to adjust vertical (Tilt) or horizontal (Azimuth) angle of cell antennas. Third, important
state information is typically missing (e.g., user location, map and material of building) and
erroneous (e.g., wrong Tilt or Azimuth). Fourth, action space is huge: (11 Tilt*13 Azim.)^50≈
5.8e107 possible actions for 50 cells. At last, coordinating actions of multiple cells is crucial
since cell action has significant impact to coverage of itself and interference to neighbor cells.
2.
2.1.
METHODS
DRL FRAMEWORK TO TRANSFER LEARNING FROM SOURCE DOMAIN TO TARGET DOMAIN
As in Figure 1, given the discrepancy between source and target domain, we use the same perception
PNN as input for DRL agent in both domains, by projecting observations OS and OT from source and
target domain to PNN via ψS(OS) and ψT(OT). Second, if tasks in source and target domain are similar,
source and target task distributions can be thought of drawn from the same task population Ω, and direct
transfer of policy can be treated as a generalization problem. Therefore, we design and generate
sufficient amount of diversified tasks in source domain to minimize the difference between source task
distribution and target task distribution from view of agent. Ideally, we want to learn optimal policy πT*
in target domain to transit from initial state S0T to optimal state S*T =TT(S0T, πT*) in one-shot. In practice,
we instead learn optimal policy πS* in source domain to approximate πT*, and further approximate πS*
by a neural network πSθ(PNN) with weight θ: πSθ ≈ πS* ≈ πT*, since S0T is partially observable.
2.2.
DISTILL TEMPORAL-SPATIAL RELATIONSHIPS TO SCALABLE 3D IMAGE-LIKE TENSOR
Given the complex temporal-spatial relationships between cells and mobile users, discrepancy between
simulator and real world, and capability of convolutional neural network (CNN) to exploit spatially
local pattern (LeCun et al., 1998), we distill local observations of each agent/cell into scalable 3D tensor
as field of view for DRL agent. As in Figure 2, for each cell C1, we rank all neighbor cells Ni based on
relationships between C1 and Ni, e.g., inter-site distance (ISD), overall interference. We then select most
important neighbors (e.g., 24) and put C1 in center and arrange Ni around C1 in X-Y axis of tensor based
on its rank. At last, for each channel along Z-axis, we extract relevant information from temporal-spatial
relationships between each pair of cells in X-Y axis, such as, topology (ISD), key performance indicator
(cell load, throughput), measured signal (averaged signal strength, averaged interference), etc.
2.3.
SELF-PLAY TO GRADUALLY IMPROVE INTELLIGENCE VIA COMPETITION
Inspired by AlphaGo, we use a novel self-play mechanism to encourage competition for best record on
multi-tasks, just like athletes compete for world record in decathlon. As in Figure 3, for initial state Si0
of task Ti drawn from a distribution, if new actions achieve better immediate global reward over all cells
Rnew than the best record Rbest in history by a threshold: ∆Rg=Rnew-Rbest>=Thge, we encourage them by
backpropagating a gradient, ge=Te(R–B(s))*dθ, here Te is a function (e.g., 2*Abs(x), Abs is absolute
value function), R is expected total reward, B(s) is baseline in REINFORCE (Williams 1992), dθ is
gradient w.r.t. weights θ. If ∆Rg<=Thgp, we penalize them by gradient gp=Tp(R–B(s))*dθ, here Tp is a
function, e.g, -1*Abs. If Thgp<∆Rg <Thge, we use simulated annealing (SA) to decide if accepting them
by comparing an uniform random number ∈ [0,1] with acceptance probability Pg=1/(1+exp(∆Rg/Tg)),
here Tg is global SA temperature annealed according to certain cooling schedule (e.g., exponential).
2.4.
DECENTRALIZED SELF-PLAY, COMPETITIVE AND COOPERATIVE DRL (S2C)
!2
!
!
Figure 5: 5 Mobile Network
!
!
Figure 6: Validation results of DRL agents
Figure 7: Test results of S2C agent
We proposed a decentralized self-play competitive/cooperative DRL method (S2C), to coordinate multiagents to compete as a team for best global reward via self-play and cooperate with each other to
minimize negative impact. As in Figure 4, each cell/agent takes action by its local view PNN. When
actions are accepted at global level, if local reward Rc for cell ci is larger than a threshold Rc>=Thce, then
accept action for ci with gradient ge; if Rc<=Thcp, then reject it with gradient gp; if Thcp<Rc<Thce, we use
SA with acceptance probability Pc=1/(1+exp(Rc/Tc)), here Tc is cell level SA temperature.
3.
3.1.
EXPERIMENTS AND RESULTS
DECENTRALIZED MULTI-AGENT MULTI-TASK DEEP REINFORCEMENT LEARNING IN SIMULATOR
First, we generated 2,380,000 CCO tasks Ti in Netlab simulator, with 10,000 random Tilt settings
as initial states for each of 238 simulated mobile networks (<=60 cells, 400-620 users; 5 shown in
Figure 5). We designed a SA agent to optimize each training task in 10 steps to generate labels (Si0
tensor for each cell in Ti, best Tilt action in 10 shots) for supervised learning (SL-DNN) by a
depth-14 residual network (He et al., 2015) with 32×32×12 input and 11 output Tilt ∈
! [-5,5].
Using 146k training data, we achieved 78.4% accuracy <=1 degree and 91.5% <=2 degree for 16k
validation data. We then use weights of SL-DNN to initialize CNN for 4 DRL agents, DQN (Mnih
et al., 2015), Double Q (Hasselt et al., 2016), A3C (Mnih et al., 2016), and S2C. We use
160k-320k training tasks over 80-160 mobile networks (distributed over 640 simulators on 80
VMs) to train agents for one-shot CCO in 8 epochs with 4-16 threads, and 300-500 validation
tasks (15-25 mobile networks) per epoch. As in Figure 6, S2C achieved better result, in terms of
immediate global reward averaged over all validation tasks (Left, 6.46% for S2C), and ratio of
validation tasks with positive global reward (Right, 94% for S2C). We also tested the same S2C
policy for 5481 unseen tasks over 238 mobile networks in Netlab without retraining. As in Figure
7 (Upper), it achieved 5.60% average global reward and 92% ratio of test tasks with positive gain.
We further verified cross-domain generalization power by testing the same S2C policy for 7693
unseen tasks over 160 unseen mobile networks (100-140 cells, 2480-19840 users) in another
simulator Unet, without any further training in both simulators. As in Figure 7 (Lower), it also
achieved good results with 4.93% average global reward and 95.7% ratio of test tasks with
positive gain.
3.2.
DIRECT TRANSFER LEARNING FROM SIMULATION TO REAL WORLD MOBILE NETWORK
To verify the generalization capability of our approach to direct transfer learning from simulation
to unseen CCO tasks in unseen very complex real world mobile network without any further
training in both domains, we performed 6 field trials over 4 commercial mobile networks that have
never been simulated in both simulators, and are very different from all simulated mobile
networks, e.g., multi-frequency (MF) or carrier aggregations (CA) has never simulated before,
user distribution/number in real world mobile network is temporal-spatial dynamic and very
different from static distribution/number in simulators, very different cell/building layouts and
radio propagation. We separated commercial mobile network A into 2 neighboring clusters C1/C2
(66/47 cells, MF), and performed a trial for each one, with 2.03% RSRP (coverage indicator) and
5.62% RSRQ (interference/capacity indicator) improvement in C1, and 3.17% RSRP, 4.86%
RSRQ improvement for C2. The 3rd trial was done for whole mobile network A (113 cells, MF),
and no significant improvement was observed since most gain has been achieved in first 2 trials.
In 4th trial, we achieved 10.79% RSRP and 6.74% RSRQ improvement for mobile network B
(151 cells, MF). In 5th trial for mobile network C (131 cells, MF/CA), no significant improvement
was observed due to either little room for optimization or significant difference between mobile
network C and task distributions in simulation. In 6th trial, we achieved 9.55% RSRP and 12.42%
RSRQ improvement for commercial mobile network D (159 cells, MF/CA).
!3
REFERENCES
Noam Brown, Tuomas Sandholm. Superhuman AI for heads-up no-limit poker: Libratus beats top
professionals. Science, 359:418-424, 2018.
Konstantinos Bousmalis, Sergey Levine. Closing the Simulation-to-Reality Gap for Deep Robotic Learning.
Google Research Blog, https://research.googleblog.com/2017/10/closing-simulation-to-reality-gap-for.html,
2017
John Celentano. Verizon Wireless: The Big Spender in 2016. AGL Media Group, http://
www.aglmediagroup.com/verizon-wireless-the-big-spender-in-2016/, 2016
Shaoshuai Fan, Hui Tian, Cigdem Sengul. Self-optimization of coverage and capacity based on a fuzzy neural
network with cooperative reinforcement learning. EURASIP Journal on Wireless Communications and
Networking, 2014:57, 2014.
Hado van Hasselt, Arthur Guez, David Silver. Deep Reinforcement Learning with Double Q-learning.
AAAI'16 Proceedings of the Thirtieth AAAI Conference on Artificial Intelligence, 2094-2100, 2016.
Kaiming He, Xiangyu Zhang, Shaoqing Ren, Jian Sun. Deep Residual Learning for Image Recognition.
Preprint at https://arxiv.org/abs/1512.03385, 2015
Yann LeCun, Léon Bottou, Yoshua Bengio, and Patrick Haffner. Gradient-based learning applied to document
recognition. Proceedings of the IEEE, 86:2278-2324, 1998.
Yann LeCun, Yoshua Bengio and Geoffrey Hinton. Deep Learning. Nature, 521:436-444, 2015.
Marketwired. Cisco Announces Intent to Acquire Intucell. http://www.marketwired.com/press-release/ciscoannounces-intent-to-acquire-intucell-nasdaq-csco-1748745.htm, 2013
Volodymyr Mnih, Koray Kavukcuoglu, David Silver, Andrei A. Rusu, Joel Veness, Marc G. Bellemare, Alex
Graves, Martin Riedmiller, Andreas K. Fidjeland, Georg Ostrovski, Stig Petersen, Charles Beattie, Amir
Sadik, Ioannis Antonoglou, Helen King, Dharshan Kumaran, Daan Wierstra, Shane Legg, Demis Hassabis.
Human-level control through deep reinforcement learning. Nature, 518:529–533, 2015.
Volodymyr Mnih, Adrià Puigdomènech Badia, Mehdi Mirza, Alex Graves, Timothy P. Lillicrap, Tim Harley,
David Silver, Koray Kavukcuoglu. Asynchronous Methods for Deep Reinforcement Learning. Preprint at
https://arxiv.org/abs/1602.01783, 2016
Matej Moravčík, Martin Schmid, Neil Burch, Viliam Lisý, Dustin Morrill, Nolan Bard, Trevor Davis, Kevin
Waugh, Michael Johanson, Michael Bowling. DeepStack: Expert-Level Artificial Intelligence in No-Limit
Poker. Science, 356:508-513, 2017
Josh Tobin, Rachel Fong, Alex Ray, Jonas Schneider, Wojciech Zaremba, Pieter Abbeel. Domain
Randomization for Transferring Deep Neural Networks from Simulation to the Real World. IROS, 2017
Andrei A. Rusu, Matej Vecerik, Thomas Rothörl, Nicolas Heess, Razvan Pascanu, Raia Hadsell. Sim-to-Real
Robot Learning from Pixels with Progressive Nets. Preprint at https://arxiv.org/abs/1610.04286, 2016.
David Silver, Aja Huang, Chris J. Maddison, Arthur Guez, Laurent Sifre, George van den Driessche, Julian
Schrittwieser, Ioannis Antonoglou, Veda Panneershelvam, Marc Lanctot, Sander Dieleman, Dominik Grewe,
John Nham, Nal Kalchbrenner, Ilya Sutskever, Timothy Lillicrap, Madeleine Leach, Koray Kavukcuoglu,
Thore Graepel, Demis Hassabis. Mastering the game of Go with deep neural networks and tree search.
Nature, 529:484–489, 2016.
David Silver, Julian Schrittwieser, Karen Simonyan, Ioannis Antonoglou, Aja Huang, Arthur Guez, Thomas
Hubert, Lucas Baker, Matthew Lai, Adrian Bolton, Yutian Chen, Timothy Lillicrap, Fan Hui, Laurent Sifre,
George van den Driessche, Thore Graepel, Demis Hassabis. Mastering the game of Go without human
knowledge. Nature, 550:354–359, 2017.
Oriol Vinyals, Timo Ewalds, Sergey Bartunov, Petko Georgiev, Alexander Sasha Vezhnevets, Michelle Yeo,
Alireza Makhzani, Heinrich Küttler, John Agapiou, Julian Schrittwieser, John Quan, Stephen Gaffney, Stig
Petersen, Karen Simonyan, Tom Schaul, Hado van Hasselt, David Silver, Timothy Lillicrap, Kevin
Calderone, Paul Keet, Anthony Brunasso, David Lawrence, Anders Ekermo, Jacob Repp, Rodney Tsing.
StarCraft II: A New Challenge for Reinforcement Learning. Preprint at https://arxiv.org/abs/1708.04782,
2017.
!4
Ronald J. Williams. Simple Statistical Gradient-Following Algorithms for. Connectionist Reinforcement
Learning. Machine Learning, 8:229–256, 1992
!5
| 2cs.AI
|
Goal-Driven Dynamics Learning via Bayesian Optimization
arXiv:1703.09260v2 [cs.SY] 22 Sep 2017
Somil Bansal, Roberto Calandra, Ted Xiao, Sergey Levine, and Claire J. Tomlin
Abstract— Real-world robots are becoming increasingly complex and commonly act in poorly understood environments
where it is extremely challenging to model or learn their true
dynamics. Therefore, it might be desirable to take a taskspecific approach, wherein the focus is on explicitly learning the
dynamics model which achieves the best control performance
for the task at hand, rather than learning the true dynamics. In
this work, we use Bayesian optimization in an active learning
framework where a locally linear dynamics model is learned
with the intent of maximizing the control performance, and
used in conjunction with optimal control schemes to efficiently
design a controller for a given task. This model is updated
directly based on the performance observed in experiments
on the physical system in an iterative manner until a desired
performance is achieved. We demonstrate the efficacy of the
proposed approach through simulations and real experiments
on a quadrotor testbed.
I. I NTRODUCTION
Given the system dynamics, optimal control schemes such
as LQR, MPC, and feedback linearization can efficiently
design a controller that maximizes a performance criterion.
However, depending on the system complexity, it can be
quite challenging to model its true dynamics. Moreover, for a
given task, a globally accurate dynamics model is not always
necessary to design a controller. Often, partial knowledge
of the dynamics is sufficient, e.g., for trajectory tracking
purposes a local linearization of a non-linear system is often
sufficient. In this paper we argue that, for complex systems,
it might be preferable to adapt the controller design process
for the specific task, using a learned system dynamics model
sufficient to achieve the desired performance.
We propose Dynamics Optimization via Bayesian Optimization (aDOBO), a Bayesian Optimization (BO) based
active learning framework to learn the dynamics model that
achieves the best performance for a given task based on
the performance observed in experiments on the physical
system. This active learning framework takes into account
all past experiments and suggests the next experiment in
order to learn the most about the relationship between the
performance criterion and the model parameters. Particularly
important for robotic systems is the use of data-efficient
approaches, where only few experiments are needed to obtain
improved performance. Hence, we employ BO, an optimization method often used to optimize a performance criterion
All authors are with the Department of Electrical Engineering and Computer Sciences, University of California, Berkeley. {somil, roberto.calandra,
t.xiao, svlevine, tomlin}@eecs.berkeley.edu
∗ This research is supported by NSF under the CPS Frontiers VehiCal
project (1545126), by the UC-Philippine-California Advanced Research
Institute under project IIID-2016-005, and by the ONR MURI Embedded
Humans (N00014-16-1-2206).
Cost Function
Current Linear
Dynamics
New Linear
Dynamics
Optimal Control
Scheme
Controller
Bayesian
Optimization
Actual System
Cost
Output
Cost
Evaluator
Fig. 1: aDOBO: A Bayesian optimization-based active learning framework for optimizing the dynamics model for a given
cost function, directly based on the observed cost values.
while keeping the number of evaluations of the physical
system small [1]. Specifically, we use BO to optimize the
dynamics model with respect to the desired task, where the
dynamics model is updated after every experiment so as to
maximize the performance on the physical system. A flow
diagram of our framework is shown in Figure 1. The current
linear dynamics model, together with the cost function (also
referred to as task/performance criterion), are used to design
a controller with an appropriate optimal control scheme. The
cost (or performance) of the controller is evaluated in closedloop operation with the actual (unknown) physical plant. BO
uses this performance information to iteratively update the
dynamics model to improve the performance. This procedure
corresponds to optimizing the linear system dynamics with
the purpose of maximizing the performance of the final
controller. Hence, unlike traditional system identification
approaches, it does not necessarily correspond to finding the
most accurate dynamics model, but rather the model yielding
the best controller performance when provided to the optimal
control method used.
Traditional system identification approaches are divided
into two stages: 1) creating a dynamics model by minimizing
some prediction error (e.g., using least squares) 2) using
this dynamics model to generate an appropriate controller.
In this approach, modeling the dynamics can be considered
an offline process as there is no information flow between
the two design stages. In online methods, the dynamics
model is instead iteratively updated using the new data
collected by evaluating the controller [2]. Our approach is
an online method. Both for the online and the offline cases,
creating a dynamics model based only on minimizing the
prediction error can introduce sufficient inaccuracies to lead
to suboptimal control performance [3]–[8]. Using machine
learning techniques, such as Gaussian processes, does not
alleviate this issue [9]. Instead, authors in [3] proposed
to optimize the dynamics model directly with respect to
the controller performance, but since the dynamics model
is optimized offline, the resultant model is not necessarily
optimal for the actual system. We instead explicitly find the
dynamics model that produces the best control performance
for the actual system.
Previous studies addressed the problem of optimizing a
controller using BO. In [10]–[12], the authors tuned the
penalty matrices in an LQR problem for performance optimization. Although interesting results emerge from these
studies, tuning penalty matrices may not achieve the desired
performance when an accurate system model is not available.
Our approach overcomes these challenges as it does not rely
on an accurate system dynamics model. In [13], the authors
directly learn the parameters of a linear feedback controller
using BO. However, a typical controller might be non-linear
and can contain hundreds of parameters; it is not feasible
to optimize such high-dimensional controllers using BO [1].
Since aDOBO does not aim at directly learning the controller,
it is agnostic to the dimensionality of the controller. It
can leverage the low-dimensional structure of the dynamics
to optimize the high-dimensional controllers. Moreover, it
does not need to impose any structure on the controller
and can easily design general non-linear controllers as well
(see Sec. IV). The problem of updating a system model
to improve control performance is also related to adaptive
control, where the model parameters are identified from
sensor data, and subsequently the updated model is used
to design a controller (see [14]–[18]). However, in adaptive
control, the model parameters are generally updated to get a
good prediction model and not necessarily to maximize the
controller performance. In contrast, we explicitly take into
account the observed performance and search for the model
that achieves the highest performance.
To the best of our knowledge, this is the first method
that optimizes a dynamics model to maximize the control
performance on the actual system. Our approach does not
require the prior knowledge of an accurate dynamics model,
nor of its parameterized form. Instead, the dynamics model is
optimized, in an active learning setting, directly with respect
to the desired cost function using data-efficient BO.
II. P ROBLEM F ORMULATION
Consider an unknown, stable, discrete-time, potentially
non-linear, dynamical system
zk+1 = f (zk , uk ),
k ∈ {0, 1, . . . , N − 1} ,
(1)
where zk ∈ Rnx and uk ∈ Rnu denote the system state and
the control input at time k respectively. Given an initial state
z0 , the objective is to design a controller that minimizes the
cost function J subject to the dynamics in (1)
J0∗
= min
−1
uN
0
N −1
J0 (zN
)
0 , u0
= min
N
−1
X
−1
uN
0
i=0
with system matrices (Aθ , Bθ ). The system matrices are
parameterized by θ ∈ M ⊆ Rd , which is to be varied during
the learning procedure. For a given θ and the current system
state zk , let πk (zk , θ) denote the optimal control sequence for
the linear system (Aθ , Bθ ) for the horizon {k, k +1, . . . , N }
N −1
πk (zk , θ) := ūkN −1 = arg min Jk (zN
),
k , uk
−1
uN
k
(3)
subject to zj+1 =Aθ zj + Bθ uj .
The key difference between (2) and (3) is that the controller is designed for the parameterized linear system as
opposed to the true system. As θ is varied, different matrix
pairs (Aθ , Bθ ) are obtained, which result in different controllers π(·, θ). Our aim is to find, among all linear models,
the linear model (Aθ∗ , Bθ∗ ) whose controller π(·, θ∗ ) minimizes J0 (ideally achieves J0∗ ) for the actual system, i.e.,
N −1
θ∗ = arg min J0 (zN
),
0 , u0
θ∈M
subject to zk+1 = f (zk , uk ) ,
uk = πk1 (zk , θ),
(4)
where πk1 (zk , θ) denotes the 1st control in the sequence πk (zk , θ). To make the dependence on θ explicit, we
refer to J0 in (4) as J(θ) here on. Note that (Aθ∗ , Bθ∗ ) in (4)
may not correspond to an actual linearization of the system,
but simply to the linear model that gives the best performance
on the actual system when its optimal controller is applied
in a closed-loop fashion on the actual physical plant.
We choose LTI modeling to reduce the number of parameters used to represent the system, and make the dynamics
learning process data efficient. Linear modeling also allows
to efficiently design the controller in (3) for general cost
functions (e.g., using MPC for any convex cost J). In
general, the effectiveness of linear modeling depends on both
the system and the control objective. If f is linear, a linear
model is trivially sufficient for any control objective. If f
is non-linear, a linear model may not be sufficient for all
control tasks; however, for regulation and trajectory tracking
tasks, a linear model is often adequate (see Sec. V). A
linear parameterization is also used in adaptive control for
similar reasons [18]. Nevertheless, the proposed framework
can handle more general model classes as long as the optimal
control problem in (3) can be solved for those classes.
Since f is unknown, the shape of the cost function, J(θ),
in (4) is unknown. The cost is thus evaluated empirically
in each experiment, which is often expensive as it involves
conducting an experiment. Thus, the goal is to solve the
optimization problem in (4) with as few evaluations as
possible. In this paper, we do so via BO.
III. BACKGROUND
l(zi , ui ) + g(zN , uN ) ,
subject to zk+1 =f (zk , uk ) ,
(2)
N −1
where zN
:=
(z
,
z
,
.
.
.
,
z
).
u
is
similarly
defined.
i i+1
N
i
i
One of the key challenges in designing such a controller is
the modeling of the unknown system dynamics in (1). In this
work, we model (1) as a linear time-invariant (LTI) system
In order to optimize (Aθ , Bθ ), we use BO. In this section,
we briefly introduce Gaussian processes and BO.
A. Gaussian Process (GP)
Since the function J(θ) in (4) is unknown a priori, we
use nonparametric GP models to approximate it over its
domain M. GPs are a popular choice for probabilistic nonparametric regression, where the goal is to find a nonlinear
map, J(θ) : M → R, from an input vector θ ∈ M to the
function value J(θ). Hence, we assume that function values
J(θ), associated with different values of θ, are random variables and that any finite number of these random variables
have a joint Gaussian distribution dependent on the values
of θ [19]. For GPs, we define a prior mean function and
a covariance function (or kernel), k(θi , θj ), which defines
the covariance between any two function values, J(θi ) and
J(θj ). In this work, the mean is assumed to be zero without
loss of generality. The choice of kernel is problem-dependent
and encodes general assumptions such as smoothness of the
unknown function. In the experimental section, we employ
the 5/2 Matèrn kernel where the hyperparameters are optimized by maximizing the marginal likelihood [19]. This
kernel function implies that the underlying function J is
differentiable and takes values within the 2σf confidence
interval with high probability.
The GP framework can be used to predict the distribution
of the performance function J(θ∗ ) at an arbitrary input θ∗
based on the past observations, D = {θi , J(θi )}ni=1 . Conditioned on D, the mean and variance of the prediction are
µ(θ∗ ) = kK −1 J ; σ 2 (θ∗ ) = k(θ∗ , θ∗ ) − kK −1 kT , (5)
where K is the kernel matrix with Kij = k(θi , θj ), k =
[k(θ1 , θ∗ ), . . . , k(θn , θ∗ )] and J = [J(θ1 ), . . . , J(θn )]. Thus,
the GP provides both the expected value of the performance
function at any arbitrary point θ∗ as well as a notion of the
uncertainty of this estimate.
B. Bayesian Optimization (BO)
Bayesian optimization aims to find the global minimum
of an unknown function [1], [20], [21]. BO is particularly
suitable for the scenarios where evaluating the unknown
function is expensive, which fits our problem in Sec. II. At
each iteration, BO uses the past observations D to model
the objective function, and uses this model to determine
informative sample locations. A common model used in BO
for the underlying objective, and the one that we consider,
are Gaussian processes (see Sec. III-A). Using the mean and
variance predictions of the GP from (5), BO computes the
next sample location by optimizing the so-called acquisition
function, α (·). Different acquisition functions are used in
literature to trade off between exploration and exploitation
during the optimization process [1]. For example, the next
evaluation for expected improvement (EI) acquisition function [22] is given by θ∗ = arg minθ α (θ) where
α (θ) = σ(θ)[uΦ(u) + φ(u)];
u = (µ(θ) − T )/σ(θ). (6)
Φ(·) and φ(·) in (6), respectively, are the standard normal
cumulative distribution and probability density functions.
The target value T is the minimum of all explored data.
Intuitively, EI selects the next parameter point where the
expected improvement over T is maximal. Repeatedly evaluating the system at points given by (6) thus improves the
observed performance. Note that optimizing α (θ) in (6) does
not require physical interactions with the system, but only
evaluation of the GP model. When a new set of optimal
Algorithm 1: aDOBO algorithm
1
2
3
4
5
6
7
8
9
10
11
D
←− if available: {θ, J (θ)}
Prior ←− if available: Prior of the GP hyperparameters
Initialize GP with D
while optimize do
0
Find θ∗ = arg minθ α (θ); θ ←− θ∗
N −1
N
z0 = {}, u0
= {}
for i = 0 : N − 1 do
0
Given zi and (Aθ0 , Bθ0 ), compute πi (zi , θ )
0
Apply πi1 (zi , θ ) on the real system and
measure zi+1
N
zN
0 ←− (z0 , zi+1 )
0
N −1
u0
←− (u0N −1 , πi1 (zi , θ ))
0
12
13
N −1
Evaluate J(θ ) := J0 (zN
) using (2)
0 , u0
0
0
Update GP and D with {θ , J(θ )}
parameters θ∗ is determined, they are finally evaluated on
the real objective function J (i.e., the system).
IV. DYNAMICS O PTIMIZATION VIA BO (aDOBO)
This section presents the technical details of aDOBO, a
novel framework for optimizing dynamics model for maximizing the resultant controller performance. In this work, θ ∈
Rnx (nx +nu ) , i.e., each dimension in θ corresponds to an entry
of the Aθ or Bθ matrices. This parameterization is chosen
for simplicity, but other parameterizations can easily be used.
Given an initial state of the system z0 and the current
system dynamics model (Aθ0 , Bθ0 ), we design an optimal
0
control sequence π0 (z0 , θ ) that minimizes the cost function
N −1
J0 (zN
), i.e., we solve the optimal control problem
0 , u0
in (3). The first control of this control sequence is applied
on the actual system and the next state z1 is measured.
0
We then similarly compute π1 (z1 , θ ) starting at z1 , apply
the first control in the obtained control sequence, measure
N −1
z2 , and so on until we get zN . Once zN
are
0 and u0
−1
obtained, we compute the true performance of uN
on
0
−1
the actual system by analytically computing J0 (zN
, uN
)
0
0
0
using (2). We denote this cost by J(θ ) for simplicity. We
next update the GP based on the collected data sample
0
0
{θ , J(θ )}. Finally, we compute θ∗ that minimizes the corresponding acquisition function α (θ) and repeat the process
for (Aθ∗ , Bθ∗ ). Our approach is illustrated in Figure 1 and
summarized in Algorithm 1. Intuitively, aDOBO directly
learns the shape of the cost function J(θ) as a function
of linearizations (Aθ , Bθ ). Instead of learning the global
shape of this function through random queries, it analyzes the
performance of all the past evaluations and by optimizing the
acquisition function, generates the next query that provides
the maximum information about the minima of the cost
function. This direct minima-seeking behavior based on the
actual observed performance ensures that our approach is
data-efficient. Thus, in the space of all linearizations, we
efficiently and directly search for the linearization whose
corresponding controller minimizes J0 on the actual system.
States
Control inputs
0
1
v
x
1.5
-1
0.5
0
10
20
-2
30
0
10
20
30
0.5
!
y
1
-1
-2
Fig. 2: Dubins car: mean and standard deviation of η during
the learning process (over 10 trials). aDOBO reaches within
the 10% of the optimal cost in just 100 iterations, starting
from a random dynamics model. Using a log warping on the
cost function further accelerates the learning.
Since the problem in (3) is an optimal control problem
for the linear system (Aθ0 , Bθ0 ), depending on the form
of the cost function J, different optimal control schemes
can be used. For example, if J is quadratic, the optimal
controller is a linear feedback controller given by the solution
of a Riccati equation. If J is a general convex function, the
optimal control problem is solved through a general convex
MPC solver, and the resultant controller could be non-linear.
Thus, depending on the form of J, the controller designed
by aDOBO can be linear or non-linear. This property causes
aDOBO to perform well in the scenarios where a linear
controller is not sufficient, as shown in Sec. VI-B. More
generally, the proposed framework is modular and other
control schemes can be used that are more suitable for a
given cost function, which allows us to capture a richer
controller space.
Note that the GP in our algorithm can be initialized with
dynamics models whose controllers are known to perform
well on the actual system. This generally leads to a faster
convergence. For example, when a good linearization of the
system is known, it can be used to initialize D. When no
information is known about the system a priori, the initial
models are queried randomly. Finally, note that aDOBO can
also be used when the real system is stochastic. In this case,
aDOBO will minimize the expected cost.
V. N UMERICAL S IMULATIONS
In this section, we present some simulation results on the
performance of the proposed method for controller design.
A. Dubins Car System
For the first simulation, we consider a three dimensional
non-linear Dubins car whose dynamics are given as
ẋ = v cos φ,
ẏ = v sin φ,
φ̇ = ω ,
(7)
where z := (x, y, φ) is the state of system, p = (x, y) is
the position, φ is the heading, v is the speed, and ω is the
turn rate. The input (control) to the system is u := (v, ω).
For simulation purposes, we discretize the dynamics at a
frequency of 10Hz. Our goal is to design a controller that
steers the system to the equilibrium point z ∗ = 0, u∗ = 0
starting from the state z0 := (1.5, 1, π/2). In particular, we
0
10
20
30
Horizon (N)
0
10
20
Horizon (N)
Learned
Optimal
Fig. 3: Dubins car: state and control trajectories for the
learned and the true system. The two trajectories are very
similar, indicating that the learned dynamics model represents the system behavior accurately around the desired state.
want to minimize the cost function
N −1
J0 (zN
)=
0 , u0
N
−1
X
T
zkT Qzk + uTk Ruk + zN
Qf zN . (8)
k=0
We choose N = 30. Q, Qf and R are all chosen as identity
matrices of appropriate sizes. We also assume that the
dynamics are not known; hence, we cannot directly design
a controller to steer the system to the desired equilibrium.
Instead, we use aDOBO to find a linearization of dynamics in
(7) that minimizes the cost function in (8), directly from the
experimental data. In particular, we represent the system in
(7) by a parameterized linear system zk+1 = Aθ zk + Bθ uk ,
design a controller for this system and apply it on the actual
system. Based on the observed performance, BO suggests
a new linearization and the process is repeated. Since the
cost function is quadratic in this case, the optimal control
problem for a particular θ is an LQR problem, and can
be solved efficiently. For BO, we use the MATLAB library
BayesOpt [23]. Since there are 3 states and 2 inputs, we learn
15 parameters in total, one corresponding to each entry of
the Aθ and Bθ matrices. The bounds on the parameters are
chosen randomly as M = [−2, 2]15 . As acquisition function,
we use EI (see eq. (6)). Since no information is assumed
to be known about the system, the GP was initialized with
a random θ. We also warp the cost function J using the
log function before passing it to BO. Warping makes the
cost function smoother while maintaining its monotonic
properties, which makes the sampling process in BO more
efficient and leads to a faster convergence.
For comparison, we compute the true optimal controller
that minimizes (8) subject to the dynamics in (7) using the
non-linear solver fmincon in MATLAB to get the minimum
achievable cost J0∗ across all controllers. We use the percentage error between the true optimal cost J0∗ and the cost
achieved by aDOBO as our comparison metric in this work
ηn = 100 × (J0∗ − J(θn ))/J0∗ ,
(9)
where J(θn ) is the best cost achieved by aDOBO by iteration n. In Fig. 2, we plot ηn for Dubins car. As learning
progresses, aDOBO gathers more and more information
about the minimum of J0 and reaches within 10% of J0∗
in just 100 iterations, demonstrating its effectiveness in
designing a controller for an unknown system just from
the experimental data. Fig. 2 also highlights the effect of
warping in BO. A well warped function converges faster
to the optimal performance. We also compared the control
and state trajectories obtained from the learned controller
with the optimal control and state trajectories. As shown
in Fig. 3, the learned system matrices not only achieve
the optimal cost, but also follow the optimal state and
control trajectories very closely. Even though the trajectories
are very close to each other for the true system and its
learned linearization, this linearization may not correspond
to any actual linearization of the system. The next simulation
illustrates this key property of aDOBO more clearly.
B. A Simple 1D Linear System
not even be required as far as the control performance is
concerned.
C. Cart-pole System
We next apply aDOBO to a cart-pole system
(M + m)ẍ − mlψ̈ cos ψ + mlψ̇ 2 sin ψ = F ,
lψ̈ − g sin ψ = ẍ cos ψ ,
where x denotes the position of the cart with mass M , ψ
denotes the pendulum angle, and F is a force that serves
as the control input. The massless pendulum is of length l
with a mass m attached at its end. Define the system state as
z := (x, ẋ, ψ, ψ̇) and the input as u := F . Starting from the
state (0, 0, π6 , 0), the goal is to keep the pendulum straight
up, while keeping the state within given lower and upper
bounds. In particular, we want to minimize the cost
For this simulation, we consider a simple 1D linear system
zk+1 = zk + uk ,
(10)
where zk and uk are the state and the input of the system
at time k. Although the dynamics model is very simple, it
illustrates some key insights about the proposed method. Our
goal is to design a controller that minimizes (8) starting from
the state z0 = 1. We choose N = 30 and R = Q = Qf = 1.
We assume that the dynamics are unknown and use aDOBO
to learn the dynamics, where θ := (θ1 , θ2 ) ∈ R2 are the
dynamics parameters to be learned.
The learning process
converges in 45 iterations
to the true optimal
performance (J0∗ = 1.61),
which is computed using
LQR on the real system.
The converged parameters
are θ1 = 1.69 and
θ2 = 2.45, which are Fig. 4: Cost of the actual
vastly different from the system in (10) as a function
true parameters θ1 = 1 and of the linearization parameters
θ2 = 1, even though the (θ1 , θ2 ). The parameters obactual system is a linear tained by aDOBO (the pink
system. To understand X) yield to performance very
this, we plot the cost close to the true system paobtained on the true rameters (the green ∗). Note
system J0 as a function that aDOBO does not necesof linearization parameters sarily converge to the true pa(θ1 , θ2 ) in Fig. 4. Since rameters.
the performances of the
two sets of parameters are very close to each other, a direct
performance based learning process (e.g., aDOBO) cannot
distinguish between them and both sets are equally optimal
for it. More generally, a wide range of parameters lead
to similar performance on the actual system. Hence, we
can expect the proposed approach to recover the optimal
controller and the actual state trajectories, but not necessarily
the true dynamics or its true linearization. This simulation
also suggests that the true dynamics of the system may
(11)
N −1
J0 (zN
)=
0 , u0
N
−1
X
T
zkT Qzk + uTk Ruk + zN
Qf zN
k=0
+λ
N
X
max(0, z − zi , zi − z),
i=0
(12)
where λ penalizes the deviation of state zi below z and
above z. We assume that the dynamics are unknown and
use aDOBO to optimize the dynamics. For simulation, we
discretize the dynamics at a frequency of 10Hz. We choose
N = 30, M = 1.5Kg, m = 0.175Kg, λ = 100 and
l = 0.28m. The Q = Qf = diag([0.1, 1, 100, 1]) and
R = 0.1 matrices are chosen to penalize the angular
deviation significantly. We use z = [−2, −∞, −0.1, −∞]
and z = [2, ∞, ∞, ∞], i.e., we are interested in controlling
the pendulum while keeping the cart position within [−2, 2],
and limiting the pendulum overshoot to 0.1. The optimal
control problem for a particular linearization is a convex
MPC problem and solved using YALMIP [24]. The true J0∗
is computed using fmincon.
As shown in Fig. 5, aDOBO reaches within 20% of
the optimal performance in 250 iterations and continue to
make progress towards finding the optimal controller. This
simulation demonstrates that the proposed method (a) is also
applicable to highly non-linear systems, (b) can handle general convex cost functions that are not necessarily quadratic,
and (c) different optimal control schemes can be used within
the proposed framework. Since an MPC controller can in
general be non-linear, this implies that the proposed method
can also design complex non-linear controllers with an LTI
parametrization.
VI. C OMPARISON WITH OTHER METHODS
In this section, we compare our approach with some other
online learning schemes for controller design.
A. Tuning (Q, R) vs aDOBO
In this section, we consider the case in which the cost
function J0 is quadratic (see Eq. (8)). Suppose that the actual
linearization of the system around z ∗ = 0 and u∗ = 0 is
Percentage error in J 0
40
30
,= 0
,= 0.1
,= 0.2
20
10
0
0
100
200
300
400
500
Iteration
Fig. 5: Cart-pole system: mean and standard deviation of η
during the learning process. The learned controller reaches
within 20% of the optimal cost in 250 iterations, demonstrating the applicability of aDOBO to highly non-linear systems.
known and given by (A∗ , B ∗ ). To design a controller for the
actual system in such a case, it is a common practice to use
an LQR controller for the linearized dynamics. However, the
resultant controller may be sub-optimal for the actual nonlinear system. To overcome this problem, authors in [10],
[11] propose to optimize the controller by tuning penalty
matrices Q and R in (8). In particular, we solve
θ∗ = arg min
θ∈M
N −1
J0 (zN
),
0 , u0
sub. to zk+1 = f (zk , uk ),
∗
uk = K(θ)zk ,
(13)
∗
K(θ) = LQR(A , B , WQ (θ), WR (θ), Qf ) ,
where K(θ) denotes the LQR feedback matrix obtained for
the system matrices (A∗ , B ∗ ) with WQ and WR as state and
input penalty matrices, and can be computed analytically. For
further details of LQR method, we refer interested readers to
[25]. The difference between optimization problems (4) and
(13) is that now we parameterize penalty matrices WQ and
WR instead of system dynamics. The optimization problem
in (13) is solved using BO in a similar fashion as we solve
(4) [10]. The parameter θ, in this case, can be initialized by
the actual penalty matrices Q and R, instead of a random
query, which generally leads to a much faster convergence.
An alternative approach is to use aDOBO, except that now
we can use (A∗ , B ∗ ) as initializations for the system matrices
A and B. Actual penalty matrices Q and R are used for
aDOBO.
When (A∗ , B ∗ ) are known to a good accuracy, (Q, R)
tuning method is expected to converge quickly to the optimal
performance compared to aDOBO as it needs to learn
fewer parameters, i.e., (nx + nu ) (assuming diagonal penalty
matrices) compared to nx (nx + nu ) parameters for aDOBO.
However, when there is error in (A∗ , B ∗ ) (or more generally
if dynamics are unknown), the performance of the (Q, R)
tuning method can degrade significantly as it relies on an
accurate linearization of the system dynamics, rendering the
method impractical for control design purposes. To compare
the two methods we use the Dubins car model in Eq. (7).
The rest of the simulation parameters are same as Section
V-A. We compute the linearization of Dubins car around
z ∗ = 0 and u∗ = 0 using (7) and add random matrices
(Ar , Br ) to them to generate A0 = (1 − α)A∗ + αAr and
B 0 = (1−α)B ∗ +αBr . We then initialize both methods with
(A0 , B 0 ) for different αs. As shown in Fig. 6, the (Q, R)
tuning method outperforms aDOBO, when there is no noise
Fig. 6: Dubins car: Comparison between tuning the penalty
matrices (Q, R) [10] (dashed curves), and aDOBO (solid
curves) for different noise levels in (A∗ , B ∗ ), the true linearized dynamics around the desired goal state. When the
true linearized dynamics are known perfectly, the (Q, R) tuning method outperforms aDOBO because fewer parameters
are to be learned. Its performance, however, drops significantly as noise increases, rendering the method impractical
for the scenarios where system dynamics are not known to
a good accuracy.
in (A∗ , B ∗ ). But as α increases, its performance deteriorates
significantly. In contrast, aDOBO is fairly indifferent to the
noise level, as it does not assume any prior knowledge of system dynamics. The only information assumed to be known
is penalty matrices (Q, R), which are generally designed by
the user and hence are known a priori.
B. Learning K vs aDOBO
When the cost function is quadratic, another potential approach is to directly parameterize and optimize the feedback
matrix K ∈ Rnx nu in (13) [13] as
N −1
θ∗ = arg min J0 (zN
),
0 , u0
θ∈M
sub. to zk+1 = f (zk , uk ),
(14)
uk = Kθ zk .
The advantage of this approach is that only nx nu parameters are learned compared to nx (nx + nu ) parameters in
aDOBO, which is also evident from Fig. 7a, wherein the
learning process for K converges much faster than that for
aDOBO. However, a linear controller might not be sufficient
for general cost functions, and non-linear controllers are
required to achieve a desired performance. As shown in
Sec. V-C, aDOBO is not limited to linear controllers; hence,
it outperforms the K learning method in such scenarios.
Consider, for example, the linear system
xk+1 = xk + yk ,
yk+1 = yk + uk ,
(15)
and the cost function in Eq. (12) with state zk = (xk , yk ),
N = 30, z = [0.5, −0.4] and z = [∞, ∞]. Q, Qf and R are
all identity matrices of appropriate sizes, and λ = 100.
As evident from Fig. 7b, directly learning a feedback
matrix performs poorly with an error as high as 80% from
the optimal cost. Since the cost is not quadratic, the optimal controller is not necessarily linear; however, since the
controller in (14) is restricted to a linear space, it performs
rather poorly in this case. In contrast, aDOBO continues
to improve performance and reaches within 20% of the
optimal cost within few iterations, because we implicitly
parameterize a much richer controller space via learning A
(a) Dubins car
(b) System of Eq. (15)
Fig. 7: Mean and standard deviation of η obtained via directly learning the feedback controller K [13] and aDOBO for
different cost functions. (a) Comparison for the quadratic cost function of Eq. (8). Directly learning K converges to the
optimal performance faster because fewer parameters are to be learned. (b) Comparison for the non-quadratic cost function
of Eq. (12). Since the optimal controller for the actual system is not necessarily linear in this case, directly learning K leads
to a poor performance
Iteration
200
400
600
aDOBO
53 ± 50%
27 ± 12%
17 ± 7%
Learning Control Sequence
605 ± 420%
357 ± 159%
263 ± 150%
TABLE I: System in (15): mean and standard deviation of η
for aDOBO, and for directly learning the control sequence.
Since the space of control sequence is huge, the error is
substantial even after 600 iterations.
and B. In this example, we capture non-linear controllers
by using a linear dynamics model with a convex MPC
solver. Since the underlying system is linear, the true optimal
controller is also in our search space. Our algorithm makes
sure that we make a steady progress towards finding that
controller. However, we are not restricted to learning a linear
controller K. One can also directly learn the actual control
sequence to be applied to the system (which also captures the
optimal controller). This approach may not be data-efficient
compared to aDOBO as the control sequence space can
be very large depending on the problem horizon, and will
require a large number of experiments. As shown in Table I,
the performance error is more than 250% even after 600
iterations, rendering the method impractical for real systems.
C. Adaptive Control vs aDOBO
In this work, we aim to directly find the best linearization
based on the observed performance. Another approach is
to learn a true linearization of the system based on the
observed state and input trajectory during the experiments.
The underlying hypothesis is that as more and more data
is collected, a better linearization is obtained, eventually
leading to an improved control performance. This approach
is in-line with the traditional model identification and the
N −1
adaptive control frameworks. Let (j zN
) denotes the
0 , j u0
state and input trajectories for experiment j. We also let Di =
N −1
∪ij=1 (j zN
). After experiment i, we fit an LTI model
0 , j u0
of the form zk+1 = Ai zk + Bi uk using least squares on data
in Di and then use this model to obtain a new controller for
experiment i+1. We apply the approach on the linear system
in (15) and the non-linear system in (7) with the cost function
in (8). For the linear system, the approach converges to the
true system dynamics in 5 iterations. However, this approach
Iteration
200
400
600
aDOBO
6 ± 3.7%
2.2 ± 1.1%
1.8 ± 0.7%
Learning via LS
166.7 ± 411%
75.9 ± 189%
70.7 ± 166%
TABLE II: Dubins car: mean and standard deviation of η
obtained via learning (A, B) through least squares (LS), and
through aDOBO.
performs rather poorly on the non-linear system, as shown in
Table II. When the underlying system is non-linear, all state
and input trajectories may not contribute to the performance
improvement. A good linearization should be obtained from
the state and input trajectories in the region of interest, which
depends on the task. For example, if we want to regulate the
system to the equilibrium (0, 0), a linearization of the system
around (0, 0) should be obtained. Thus, it is desirable to
use the system trajectories that are close to this equilibrium
point. However, a naive prediction error based approach has
no means to select these “good” trajectories from the pool
of trajectories and hence can lead to a poor performance.
In contrast, aDOBO does not suffer from these limitations,
as it explicitly utilizes a performance based optimization.
A summary of the advantages and limitations of the four
methods is provided in Table III.
VII. Q UADROTOR P OSITION T RACKING E XPERIMENTS
We now present the results of our experiments on
Crazyflie 2.0, which is an
open source nano quadrotor platform developed by
Bitcraze. Its small size, low
cost, and robustness make it
Fig. 8: The Crazyflie 2.0
an ideal platform for testing
new control paradigms. Recently, it has been extensively
used to demonstrate aggressive flights [26], [27]. For small
yaw, the quadrotor system is modeled asa rigid body
with
a ten dimensional state vector s := p, v, ζ, ω , which
includes the position p = (x, y, z) in an inertial frame I,
linear velocities v = (vx , vy , vz ) in I, attitude (orientation)
represented by Euler angles ζ, and angular velocities
ω.
The system is controlled via three inputs u := u1 , u2 , u3 ,
where u1 is the thrust along the z-axis, and u2 and u3 are
Method
(Q, R) learning [10]
F learning [13]
(A, B) learning via least
squares
aDOBO
Advantages
Only (nx + nu ) parameters are to be learned so
learning will be faster.
Only nx nu parameters are to be learned so learning
will be faster.
Can lead to a faster convergence when the underlying
system is linear
Does not require any prior knowledge of system
dynamics. Applicable to general cost functions.
Limitations
Performance can degrade significantly if the dynamics are not
known to a good accuracy.
Approach may not perform well for non-quadratic cost functions.
Approach is not suitable for non-linear system.
Number of parameters to be learned is higher, i.e., (n2x + nx nu ).
rolling, pitching moments respectively. The full non-linear
dynamics of a quadrotor are derived in [28], and its physical
parameters are computed in [26]. Our goal in this experiment
is to track a desired position p∗ starting from the initial
position p0 = [0, 0, 1]. Formally, we minimize
% improvement in J
TABLE III: Relative advantages and limitations of different methods for automatic controller design.
Percentage error in cost wrt the nominal controller
10
0
-10
0
10
20
30
40
50
60
Iteration
N −1
J0 (s̄N
)=
0 , u0
N
−1
X
s̄Tk Qs̄ + uTk Ruk + s̄TN Qf s̄ , (16)
k=0
where s̄ := p − p∗ , v, ζ, ω . Given the dynamics in [28], the
desired optimal control problem can be solved using LQR;
however, the resultant controller may still not be optimal for
the actual system because (a) true underlying system is nonlinear and (b) the actual system may not follow the dynamics
in [28] exactly due to several unmodeled effects, as illustrated
in our results. Hence, we assume that the dynamics of vx and
vy are unknown, and model them as
f vx
φ
= Aθ
+ Bθ u1 ,
(17)
f vy
ψ
where A and B are parameterized through θ. Our goal is to
learn the parameter θ∗ that minimizes the cost in (16) for
the actual Crazyflie using aDOBO. We use N = 400; the
penalty matrix Q is chosen to penalize the position deviation.
In our experiments, Crazyflie was flown in presence of a
VICON motion capture system, which along with on-board
sensors provides the full state information at 100Hz. The
optimal control problem for a particular linearization in
(17) is solved using LQR. For comparison, we compute
the nominal optimal controller using the full dynamics in
[28]. Figure 9 shows the performance of the controller
from aDOBO compared with the nominal controller during
the learning process. The nominal controller outperforms
the learned controller initially, but within a few iterations,
aDOBO performs better than the controller derived from the
known dynamics model of Crazyflie. This is because aDOBO
optimizes controller based on the performance of the actual
system and hence can account for unmodeled effects. In
45 iterations, the learned controller outperforms the nominal
controller by 12%, demonstrating the performance potential
of aDOBO on real systems.
VIII. C ONCLUSIONS AND F UTURE W ORK
In this work, we introduce aDOBO, an active learning
framework to optimize the system dynamics with the intent
of maximizing the controller performance. Through simulations and real-world experiments, we demonstrate that
Fig. 9: Crazyflie: percentage error between the learned and
the nominal controller. The nominal controller is obtained
by using the full 12D non-linear dynamics model of the
quadrotor. As learning progresses, aDOBO outperforms the
nominal controller by 12% on the actual system, indicating
the capability of aDOBO to overcome modeling errors.
aDOBO achieves the optimal control performance even when
no prior information is known about the system dynamics.
Several interesting future directions emerge from this
work. The dynamics model learned through aDOBO maximizes the performance on a single task. The obtained
dynamics model thus may not necessarily perform well on a
similar but different task. It will be interesting to generalize
aDOBO to optimize the dynamics for a class of tasks, e.g.,
regulating to different states. Leveraging the state and input
trajectory data, along with the observed performance, to
further increase the data-efficiency of the learning process
is another promising direction. During the learning process,
aDOBO can query parameters which might lead to an
unstable behavior on the actual system and can cause safety
concerns. In such cases, it might be desirable to combine
aDOBO with the exploration methods that explicitly take
safety into account, e.g., such as SafeOpt [29], [30]. Finally,
since BO is not scalable to higher-dimensional systems
(roughly beyond 30-40 parameters) [1], exploring alternative
ways to scale aDOBO to more complex non-linear dynamics
models is another interesting direction.
R EFERENCES
[1] B. Shahriari, K. Swersky, Z. Wang, R. P. Adams, and N. de Freitas,
“Taking the human out of the loop: A review of Bayesian optimization,” Proceedings of the IEEE, vol. 104, no. 1, pp. 148–175, 2016.
[2] M. P. Deisenroth, D. Fox, and C. E. Rasmussen, “Gaussian processes
for data-efficient learning in robotics and control,” Transactions on
Pattern Analysis and Machine Intelligence (PAMI), 2015.
[3] J. Joseph, A. Geramifard, J. W. Roberts, J. P. How, and N. Roy, “Reinforcement learning with misspecified model classes,” in International
Conference on Robotics and Automation, 2013, pp. 939–946.
[4] P. L. Donti, B. Amos, and J. Z. Kolter, “Task-based end-to-end model
learning,” arXiv preprint arXiv:1703.04529, 2017.
[5] C. G. Atkeson, “Nonparametric model-based reinforcement learning,”
in Advances in neural information processing systems, 1998, pp. 1008–
1014.
[6] P. Abbeel, M. Quigley, and A. Y. Ng, “Using inaccurate models
in reinforcement learning,” in International conference on Machine
learning, 2006, pp. 1–8.
[7] M. Gevers, “Identification for control: From the early achievements to
the revival of experiment design,” European journal of control, vol. 11,
no. 4-5, 2005.
[8] H. Hjalmarsson, M. Gevers, and F. De Bruyne, “For model-based
control design, closed-loop identification gives better performance,”
Automatica, vol. 32, no. 12, 1996.
[9] D. Nguyen-Tuong and J. Peters, “Model learning for robot control: a
survey,” Cognitive Processing, vol. 12, no. 4, pp. 319–340, 2011.
[10] A. Marco, P. Hennig, J. Bohg, S. Schaal, and S. Trimpe, “Automatic
LQR tuning based on Gaussian process global optimization,” in
International Conference on Robotics and Automation, 2016.
[11] S. Trimpe, A. Millane, S. Doessegger, and R. D’Andrea, “A selftuning LQR approach demonstrated on an inverted pendulum,” IFAC
Proceedings Volumes, vol. 47, no. 3, pp. 11 281–11 287, 2014.
[12] J. W. Roberts, I. R. Manchester, and R. Tedrake, “Feedback controller
parameterizations for reinforcement learning,” in Symposium on Adaptive Dynamic Programming And Reinforcement Learning (ADPRL),
2011, pp. 310–317.
[13] R. Calandra, A. Seyfarth, J. Peters, and M. P. Deisenroth, “Bayesian
optimization for learning gaits under uncertainty,” Annals of Mathematics and Artificial Intelligence, vol. 76, no. 1, pp. 5–23, 2015.
[14] K. J. Åström and B. Wittenmark, Adaptive control. Courier Corporation, 2013.
[15] M. Grimble, “Implicit and explicit LQG self-tuning controllers,”
Automatica, vol. 20, no. 5, pp. 661–669, 1984.
[16] D. Clarke, P. Kanjilal, and C. Mohtadi, “A generalized LQG approach
to self-tuning control part i. aspects of design,” International Journal
of Control, vol. 41, no. 6, pp. 1509–1523, 1985.
[17] R. Murray-Smith and D. Sbarbaro, “Nonlinear adaptive control using
nonparametric Gaussian process prior models,” IFAC Proceedings
Volumes, vol. 35, no. 1, pp. 325–330, 2002.
[18] S. Sastry and M. Bodson, Adaptive control: stability, convergence and
robustness. Courier Corporation, 2011.
[19] C. E. Rasmussen and C. K. I. Williams, Gaussian Processes for
Machine Learning. The MIT Press, 2006.
[20] H. J. Kushner, “A new method of locating the maximum point of an
arbitrary multipeak curve in the presence of noise,” Journal of Basic
Engineering, vol. 86, p. 97, 1964.
[21] M. A. Osborne, R. Garnett, and S. J. Roberts, “Gaussian processes
for global optimization,” in Learning and Intelligent Optimization
(LION3), 2009, pp. 1–15.
[22] J. Močkus, “On bayesian methods for seeking the extremum,” in
Optimization Techniques IFIP Technical Conference, 1975.
[23] R. Martinez-Cantin, “BayesOpt: a Bayesian optimization library for
nonlinear optimization, experimental design and bandits.” Journal of
Machine Learning Research, vol. 15, no. 1, pp. 3735–3739, 2014.
[24] J. Lofberg, “YALMIP: A toolbox for modeling and optimization in
MATLAB,” in International Symposium on Computer Aided Control
Systems Design, 2005, pp. 284–289.
[25] D. J. Bender and A. J. Laub, “The linear-quadratic optimal regulator
for descriptor systems: discrete-time case,” Automatica, 1987.
[26] B. Landry, “Planning and control for quadrotor flight through cluttered
environments,” Master’s thesis, MIT, 2015.
[27] S. Bansal, A. K. Akametalu, F. J. Jiang, F. Laine, and C. J. Tomlin,
“Learning quadrotor dynamics using neural network for flight control,”
in Conference on Decision and Control, 2016, pp. 4653–4660.
[28] N. Abas, A. Legowo, and R. Akmeliawati, “Parameter identification
of an autonomous quadrotor,” in International Conference On Mechatronics, 2011, pp. 1–8.
[29] Y. Sui, A. Gotovos, J. Burdick, and A. Krause, “Safe exploration for
optimization with Gaussian processes,” in International Conference on
Machine Learning, 2015, pp. 997–1005.
[30] F. Berkenkamp, A. P. Schoellig, and A. Krause, “Safe controller
optimization for quadrotors with Gaussian processes,” in International
Conference on Robotics and Automation, 2016, pp. 491–496.
| 3cs.SY
|
Exploiting generalisation symmetries in accuracy-based
learning classifier systems: An initial study
Larry Bull
Abstract. Modern learning classifier systems typically exploit a
niched genetic algorithm to facilitate rule discovery. When used
for reinforcement learning, such rules represent generalisations
over the state-action-reward space. Whilst encouraging maximal
generality, the niching can potentially hinder the formation of
generalisations in the state space which are symmetrical, or very
similar, over different actions. This paper introduces the use of
rules which contain multiple actions, maintaining accuracy and
reward metrics for each action. It is shown that problem
symmetries can be exploited, improving performance, whilst not
degrading performance when symmetries are reduced. 12
1 INTRODUCTION
Learning Classifier Systems (LCS) [Holland, 1976] are rulebased systems, where the rules are usually in the traditional
production system form of “IF condition THEN assertion”. An
evolutionary algorithm and/or other heuristics are used to search
the space of possible rules, whilst another learning process is
used to assign utility to existing rules, thereby guiding the search
for better rules. LCS are typically used as a form of
reinforcement learner, although variants also exist for supervised
[Bernadó Mansilla & Garrell, 2003], unsupervised [Tammee et
al., 2007] and function [Wilson, 2002] learning. Almost twenty
years ago, Stewart Wilson introduced a form of LCS in which
rule utility is calculated solely by the accuracy of the predicted
consequences of rule assertions/actions – the “eXtended
Classifier System” (XCS) [Wilson, 1995]. Importantly, XCS
makes a clear connection between LCS and modern
reinforcement learning (see [Sutton & Barto, 1998]): XCS uses a
genetic algorithm (GA) [Holland, 1975] to discover regularities
in the problem thereby enabling generalisations over the
complete state-action-reward space. It has been found able to
solve a number of well-known problems optimally (e.g., see
[Butz, 2006]). Modern LCS, primarily XCS and its derivatives,
have been applied to a number of real-world problems (e.g., see
[Bull, 2004]), particularly data mining (e.g., see [Bull et al.,
2008]), to great effect. Formal understanding of modern LCS has
also increased in recent years (e.g., see [Bull & Kovacs, 2005]).
XCS uses a niched GA, that is, it runs the GA over rules
which are concurrently active. Initially, following [Booker,
1985] (see also [Fogarty, 1994]), the GA was run in the match
set [M], i.e., the subset of rules whose condition matches the
current state. The primary motivation for restricting the GA in
this way is to avoid the recombination of rule conditions which
generalise over very different areas of the problem space. Wilson
Dept. of Computer Science & Creative Tech., UWE BS16 1QY, UK.
Email: [email protected].
[1998] later increased the niching to action sets [A], i.e., the
subset of [M] whose action matches the chosen output of the
system. Wilson correctly highlighted that for tasks with
asymmetrical generalisations per action, the GA would still have
the potential to unhelpfully recombine rules working over
different sub-regions of the input space unless it is moved to [A].
Using two simple benchmark tasks, he didn’t show significant
changes in performance but did show a decrease in the number
of unique rules maintained when some asymmetry existed from
the use in [A]. Modern XCS uses the [A] form of GA, which has
been studied formally in various ways (e.g., see [Bull, 2002;
2005][Butz et al., 2004][Butz et al., 2007]). It can be noted that
the first LCS maintained separate GA populations per action
[Holland & Reitman, 1978] (see [Wilson, 1985] for a similar
scheme).
The degree of symmetry within the state-action-reward space
across all problems is a continuum. As noted, running the GA in
niches of concurrently active rules identifies those whose
conditions overlap in the problem space. However, using the GA
in [A] means that any common structure in the problem space
discovered by a rule with one action must wait to be shared
through the appropriate mutation of its action. Otherwise it must
be rediscovered by the GA for rules with another action(s). As
the degree of symmetry in the problem increases, so the
potentially negative effect of using the GA in [A] on the search
process increases.
This paper proposes a change in the standard rule structure to
address the issue and demonstrates it using a slightly simplified
version of XCS, termed YCS [Bull, 2005].
2 YCS: A SIMPLE ACCURACY-BASED LCS
YCS is without internal memory, the rule-base consists of a
number (P) of condition-action rules in which the condition is a
string of characters from the traditional ternary alphabet {0,1,#}
and the action is represented by a binary string. Associated with
each rule is a predicted reward value (r), a scalar which indicates
the error () in the rule’s predicted reward and an estimate of the
average size of the niches in which that rule participates (). The
initial random population has these parameters initialized,
somewhat arbitrarily, to 10.
On receipt of an input message, the rule-base is scanned, and
any rule whose condition matches the message at each position
is tagged as a member of the current match set [M]. An action is
then chosen from those proposed by the members of the match
set and all rules proposing the selected action form an action set
[A]. A version of XCS’s explore/exploit action selection scheme
will be used here. That is, on one cycle an action is chosen at
random and on the following the action with the highest average
fitness-weighted reward is chosen deterministically.
The simplest case of immediate reward R is considered here.
Reinforcement in YCS consists of updating the error, the niche
size estimate and then the reward estimate of each member of the
current [A] using the Widrow-Hoff delta rule with learning rate
:
j j + ( |R - rj| - j )
(1)
rj rj + ( R - rj )
(2)
j j + ( |[A]| - j )
(3)
The original YCS employs two discovery mechanisms, a
panmictic (standard global) GA and a covering operator. On
each time-step there is a probability g of GA invocation. The GA
uses roulette wheel selection to determine two parent rules based
on the inverse of their error:
fj = ( 1 / (jv + 1) )
(4)
Here the exponent v enables control of the fitness pressure
within the system by facilitating tuneable fitness separation
under fitness proportionate selection (see [Bull, 2005] for
discussions). Offspring are produced via mutation (probability )
and crossover (single point with probability ), inheriting the
parents’ parameter values or their average if crossover is
invoked. Replacement of existing members of the rulebase uses
roulette wheel selection based on estimated niche size. If no
rules match on a given time step, then a covering operator is
used which creates a rule with the message as its condition
(augmented with wildcards at the rate p#) and a random action,
which then replaces an existing member of the rulebase in the
usual way. Parameter updating and the GA are not used on
exploit trials.
last system cycle upon which it was part of a GA (a development
of [Booker, 1989]). The GA is applied within the current action
set [A] when the average number of system cycles since the last
GA in the set is over a threshold GA. If this condition is met, the
GA time-stamp of each rule is set to the current system time, two
parents are chosen according to their fitness using standard
roulette-wheel selection, and their offspring are potentially
crossed and mutated, before being inserted into the rule-base as
described above.
YCS is therefore a simple accuracy-based LCS which
captures the fundamental characteristics of XCS: “[E]ach
classifier maintains a prediction of expected payoff, but the
classifier’s fitness is not given by the prediction. Instead the
fitness is a separate number based on an inverse function of the
classifier’s average prediction error” [Wilson, 1995] and a
“classifier’s deletion probability is set proportional to the [niche]
size estimate, which tends to make all [niches] have about the
same size, so that classifier resources are allocated more or less
equally to all niches” [ibid]. However, YCS does not include a
number of other mechanisms within XCS, such as niche-based
fitness sharing, which are known to have beneficial effects in
some domains (see [Butz et al., 2004]).
The pressure within XCS and its derivatives to evolve
maximally general rules over the problem space comes from the
triggered niche GA. Selection for reproduction is based upon the
accuracy of prediction, as described. Thus within a niche,
accurate rules are more likely to be selected. However, more
general rules participate in more niches as they match more
inputs. Rules which are both general and accurate therefore
typically reproduce the most: the more general and accurate, the
more a rule is likely to be selected. Any rule which is less
general but equally accurate will have fewer chances to
reproduce. Any rule which is over general will have more
chances to reproduce but a lower accuracy (see [Butz et al.,
2004] for detailed analysis).
Under the new rule representation scheme introduced here
each rule consists of a single condition and each possible action.
Associated with each action are the two parameters updated
according to equations 1 and 2:
Traditional rule – condition: action: reward: error: niche
New rule –
Figure 1: Schematic of YCS as used here.
The niche GA mechanism used here is XCS’s time-based
approach under which each rule maintains a time-stamp of the
condition: action1: reward: error: niche
action2: reward: error
action3: reward: error
…
actionN: reward: error
All other processing remains the same as described. In this
way, any symmetry is directly exploitable by a single rule whilst
still limiting the possibility for recombining rules covering
different parts of the problem space since the GA is run in [A],
as Wilson [1998] described. Any action which is not correctly
associated with the generalisation over the problem space
represented by the condition will have a low accuracy and can be
ignored in any post processing of rules for knowledge discovery.
The generalisation process of modern LCS is implicitly extended
to evolve rules which are accurate over as many actions as
possible since they will participate in more niches. Note that the
niche size estimate can become noisier than in standard
YCS/XCS. Similarly, any effects from the potential maintenance
of inaccurate generalisations in some niches due to their being
accurate in other niches are not explored here. Initial results do
not indicate any significant disruption however.
3 EXPERIMENTATION
3.1 Symmetry
Following [Wilson, 1995], the well-known multiplexer task is
used in this paper. These Boolean functions are defined for
binary strings of length l = k + 2k under which the k bits index
into the remaining 2k bits, returning the value of the indexed bit.
A correct classification results in a payoff of 1000, otherwise 0.
For example, in the k=4 multiplexer the following traditional
rules form one optimal [M] (error and niche size not shown):
1111###############1: 1: 1000
1111###############1: 0: 0
Figure 2 shows the performance of YCS using the new
multi-action rule representation on the 20-bit multiplexer (k=4)
problem with P=1000, p#=0.6, =0.04, v=10, =0.5, GA=25 and
=0.2. After [Wilson, 1995], performance, taken here to mean
the fraction of correct responses, is shown from exploit trials
only, using a 50-point running average, averaged over twenty
runs. It can be seen that optimal performance is reached around
60,000 trails. Figure 2 also shows the average specificity, taken
here to mean the fraction of non-# bits in a condition, for the
LCS. That is, the amount of generalization produced. The
maximally general solution to the 20-bit multiplexer has
specificity 5/20 = 0.25 and YCS can be seen to produce rulebases with an average specificity very close to the optimum. The
average error of rules can also be seen to decrease over time.
Figure 3 shows the performance of YCS using the traditional
rule representation with the same parameters. As can be seen,
optimal performance is not reliably reached in the allowed time.
Figure 4 shows the performance of the same system with
P=2000, with optimality reached around 60,000 trials (matching
that of XCS with the same equivalent parameters, e.g., [Butz et
al., 2004]). That is, with double the rule-base resource, the GA is
able to reliably (re)discover the problem structure in all [A] over
the same time period using the traditional rule representation.
Hence, in a problem with complete symmetry between [A], the
new rule representation presented here significantly improves the
efficiency of the GA.
Figure 2: Performance of new rule representation.
Figure 3: Performance of traditional rule representation.
3.2 Less Symmetry
To reduce the symmetry in the multiplexer in a simple way, an
extra bit can be added. Here an incorrect response becomes
sensitive to the value of the extra input bit: if it is set, the reward
is 500, otherwise it is 0. That is, using the new rule
representation, it is no longer possible for just one rule to use the
same generalisation over the input space to accurately predict the
reward for each action in a given [M]. The following traditional
rules represent one optimal [M]:
1111###############1#: 1: 1000
1111###############11: 0: 500
1111###############10: 0: 0
Figure 4: As Figure 3 but with larger population size.
Figure 5 shows how YCS is unable to solve the less
symmetrical 20-bit multiplexer using the new rule representation
with P=1000. Figures 6 and 7 show how the performance of
YCS with and without the new representation (respectively) is
optimal and roughly equal with P=2000. Note that the new
representation still only requires two rules per [M], as opposed to
three in the traditional scheme. However, although there is a
slight increase in learning speed with the new scheme, it is not
statistically significant (T-test, time taken to reach and maintain
optimality over 50 subsequent exploit cycles, p>0.05). Figures 8
and 9 show there is significant benefit (p≤0.05) from the new
representation when k=5, i.e., the harder 37-bit multiplexer
(P=5000).
3.3 Multiple Actions
Multiplexers are binary classification problems. To create a
multi-class/multi-action variant in a simple way the case where
the data bit is a ‘1’ is altered to require an action equal to the
value of the address bits for a correct response. In this way there
are 2k possible actions/classes. Under the new format with k=3,
one optimal [M] could be represented as the single rule:
Figure 6: As Figure 5 but with larger population size.
111#######1: 8: 1000
7: 0
6: 0
5: 0
4: 0
3: 0
2: 0
1: 0
0: 0
Figures 10 and 11 show the performance of YCS with and
without the new representation (respectively) with k=3 and
P=2000. As can be seen, both representations are capable of
optimal performance with the parameters used but the new
representation learns significantly faster ((p≤0.05).
Figure 5: Performance of new scheme on less symmetrical
task.
Figure 7: Performance of traditional rules on less
symmetrical task (vs. Figure 6).
Figure 8: Performance of new scheme on less symmetrical
multiplexer when k=5.
Figure 9: Performance of traditional rules on less
symmetrical multiplexer when k=5.
3.4 Imbalance
The frequency of state visitation is rarely close to uniform in
most reinforcement learning tasks. For example, in a spatial
maze navigation task, those states at or near a goal will typically
be visited more often than those states far from a goal. In data
mining, real-world data does not typically contain equal
examples of all cases of the underlying concept space - known as
the class imbalance problem, and often tackled through
under/over sampling. This bias of sampling the problem space
can cause difficulties in the production of accurate
generalisations since over general rules can come to dominate
niches due to their frequency of use (updating and reproduction)
in more frequently visited states. Orriols-Puig and Bernado
Mansilla [2008] introduced a heuristic specifically for (limited
to) binary classification tasks which dynamically alters the
learning rate ( and frequency of GA activity (GA) to address
the issue in accuracy-based LCS. They show improved learning
in both imbalanced multiplexers and well-known data sets.
The new rule representation would appear to have some
potential to address the issue of imbalance generally when there
is symmetry in the underlying problem space, i.e., both for
reinforcement learning and data mining. Since all actions are
maintained by all rules, information about all actions is
maintained in the population. Whilst over general conditions will
quickly emerge for the same reasons as for the traditional
representation, later in the search, the use and updating of the
correct actions for less frequently visited states will indicate their
true value and the GA will (potentially) adjust generalisations
appropriately. An imbalanced multiplexer (akin to [Orriols-Puig
& Bernado Mansilla, 2008]) can be created by simply
introducing a probabilistic bias in sampling action ‘1’ compared
to ‘0’. Figures 12 and 13 show the performance of YCS with and
without the new representation (respectively) with k=4, P=2000
and a bias of 80% (4:1). Exploit cycle testing remains unbiased,
as before. As can be seen, the new representation is able to cope
with the bias, whereas the equivalent traditional rule
representation is not. The same was generally found to be true
for various levels of bias, k, etc. (not shown).
Figure 10: Performance of new scheme on multi-action task.
Figure 11: Performance of traditional rules on multi-action
task.
Figure 12: Performance of new scheme on the imbalanced
task.
REFERENCES
Figure 13: Performance of the traditional scheme on the
imbalanced task.
4 CONCLUSIONS & FUTURE WORK
This paper has proposed the use of rules which contain multiple
actions, maintaining accuracy and reward metrics for each
action. This somewhat minor alteration appears to provide
benefits over the traditional approach in a variety of scenarios.
Future work should also consider the new, general rule structure
proposed here with more complex representations such as realvalued intervals (e.g., see [Stone & Bull, 2003]) or genetic
programming (e.g., see [Preen & Bull, 2013]), together with
delayed reward tasks.
Kovacs and Tindale [2013] have recently highlighted issues
regarding the niche GA, particularly with respect to overlapping
problems. They compare the performance of an accuracy-based
LCS with a global GA (see also [Bull, 2005]), a niche GA, and a
global GA which uses the calculated selective probabilities of
rules under a niche GA. The aim being to avoid the reduced
actual selection of accurate, general rules due to overlap within a
given niche. Using the 11-bit multiplexer (k=3) problem they
show a possible slight increase in performance from their new
scheme over the niche GA, with the global GA performing
worst. Their new scheme shows an increase in the number of
unique rules maintained compared to the niche GA and they
postulate this increase in rule diversity may explain the
suggested difference in performance. This seems likely given the
multiplexer does not contain any overlap. Note that Wilson
[1994] proposed using both a global and niche GA together “to
offset any inbreeding tendency” within niches. Since they used a
supervised form of XCS which only maintains the highest
reward entries of the state-action-reward map (UCS) [Bernado
Mansilla & Garrell, 2003], the exploitation of symmetry does not
help to explain their findings. The effect of the new
representation in overlapping problems remains to be explored.
The related use of multiple conditions per action may be a more
appropriate approach.
Bernado Mansilla, E. & Garrell, J. (2003) Accuracy-Based Learning
Classifier Systems: Models, Analysis and Applications to
Classification Tasks. Evolutionary Computation 11(3): 209-238.
Booker, L.B. (1985) Improving the Performance of Genetic Algorithms
in Classifier Systems. In J.J. Grefenstette (ed) Proceedings of the
First International Conference on Genetic Algorithms and their
Applications. Lawrence Erlbaum Associates, pp80-92.
Booker, L.B. (1989) Triggered Rule Discovery in Classifier Systems. In
J. Schaffer (ed) Proceedings of the Third International Conference
on Genetic Algorithms and their Applications. Morgan Kaufmann,
pp265-274.
Bull, L. (2002) On Accuracy-based Fitness. Soft Computing 6(3-4): 154161.
Bull, L. (2004)(ed) Applications of Learning Classifier Systems.
Springer.
Bull, L. (2005) Two Simple Learning Classifier Systems. In L. Bull & T.
Kovacs (eds) Foundations of Learning Classifier Systems. Springer,
pp63-90.
Bull, L. & Kovacs, T. (2005)(eds) Foundations of Learning Classifier
Systems. Springer.
Bull, L., Bernado Mansilla, E & Holmes, J. (2008)(eds) Learning
Classifier Systems in Data Mining. Springer.
Butz, M. (2006) Rule-based Evolutionary Online Learning Systems.
Springer.
Butz, M., Kovacs, T., Lanzi, P-L & Wilson, S.W. (2004) Toward a
Theory of Generalization and Learning in XCS. IEEE Transactions
on Evolutionary Computation 8(1): 28-46
Butz, M., Goldberg, D., Lanzi, P-L. & Sastry, K. (2007) Problem
solution sustenance in XCS: Markov chain analysis of niche support
distributions and the impact on computational complexity. Genetic
Programming and Evolvable Machines 8(1): 5-37
Fogarty, T.C. (1994) Co-evolving Co-operative Populations of Rules in
Learning Control Systems. In T.C. Fogarty (ed) Evolutionary
Computing. Springer, pp195-209.
Holland, J.H. (1975) Adaptation in Natural and Artificial Systems.
University of Michigan Press.
Holland, J.H. (1976) Adaptation. In R. Rosen & F.M. Snell (eds)
Progress in Theoretical Biology, 4. Academic Press, pp313-329.
Holland, J.H. & Reitman, J.H. (1978) Cognitive Systems Based in
Adaptive Algorithms. In Waterman & Hayes-Roth (eds) Patterndirected Inference Systems. Academic Press.
Kovacs, T. & Tindale, R. (2013) Analysis of the niche genetic algorithm
in learning classifier systems. In Proceedings of the Genetic and
Evolutionary Computation Conference. ACM Press, pp1069-1076.
Orriols-Puig, A. & Bernado Mansilla, E. (2008) Evolutionary Rule-based
Systems for Imbalanced Data Sets. Soft Computing 13(3): 213-225.
Preen, R. & Bull, L. (2013) Dynamical Genetic Programming in XCSF.
Evolutionary Computation 21(3): 361-388.
Stone, C. & Bull, L. (2003) For Real! XCS with Continuous-Valued
Inputs. Evolutionary Computation 11(3): 299-336
Sutton, R.S. & Barto, A.G. (1998) Reinforcement Learning. MIT Press.
Tammee, K., Bull, L. & Ouen, P. (2007) Towards Clustering with XCS.
In D. Thierens et al. (eds) Proceedings of the Genetic and
Evolutionary Computation Conference. ACM Press, pp1854-1860
Wilson, S.W. (1985) Knowledge Growth in an Artificial Animal. J.J.
Grefenstette (ed) Proceedings of the First International Conference
on Genetic Algorithms and their Applications. Lawrence Erlbaum
Associates, pp16-23.
Wilson, S.W. (1994) ZCS: A Zeroth-level Classifier System.
Evolutionary Computation 2(1):1-18.
Wilson, S.W. (1995) Classifier Fitness Based on Accuracy. Evolutionary
Computation 3(2):149-177.
Wilson, S.W. (1998) Generalization in the XCS Classifier System. In
Koza et al. (eds.) Genetic Programming 1998: Proceedings of the
Third Annual Conference. Morgan Kaufmann, pp322-334.
Wilson, S.W. (2002) Classifiers that Approximate Functions. Natural
Computing 1(2-3): 211-234.
| 9cs.NE
|
arXiv:1802.08774v1 [cs.CV] 24 Feb 2018
Tool Detection and Operative Skill Assessment in Surgical Videos Using
Region-Based Convolutional Neural Networks
Amy Jin, Serena Yeung, Jeffrey Jopling, Jonathan Krause, Dan Azagury,
Arnold Milstein, and Li Fei-Fei
Stanford University
Stanford, CA 94305
Abstract
Five billion people in the world lack access to quality
surgical care. Surgeon skill varies dramatically, and many
surgical patients suffer complications and avoidable harm.
Improving surgical training and feedback would help to reduce the rate of complications—half of which have been
shown to be preventable. To do this, it is essential to assess
operative skill, a process that currently requires experts and
is manual, time consuming, and subjective. In this work, we
introduce an approach to automatically assess surgeon performance by tracking and analyzing tool movements in surgical videos, leveraging region-based convolutional neural
networks. In order to study this problem, we also introduce
a new dataset, m2cai16-tool-locations, which extends the
m2cai16-tool dataset with spatial bounds of tools. While
previous methods have addressed tool presence detection,
ours is the first to not only detect presence but also spatially
localize surgical tools in real-world laparoscopic surgical
videos. We show that our method both effectively detects
the spatial bounds of tools as well as significantly outperforms existing methods on tool presence detection. We further demonstrate the ability of our method to assess surgical
quality through analysis of tool usage patterns, movement
range, and economy of motion.
1. Introduction
Five billion people in the world lack access to safe surgical care [1]. According to the World Health Organization,
there is a global mortality rate of 0.5-5% for major procedures, and up to 25% of patients undergoing operations that
require a stay in the hospital suffer complications [2]. Many
of these complications are caused by poor individual and
team performance, for which inadequate training and feedback play an important role [3, 4, 5]. These outcomes can be
improved, as studies show that half of all adverse surgical
Figure 1: Real-time automated surgical video analysis could facilitate objective and efficient assessment of surgical skill and provide feedback on surgical performance. In this work, we introduce
an approach to: (1) automatically detect, classify, and localize surgical instruments in real-world laparoscopic surgical videos; and
(2) efficiently analyze surgical performance based on the extracted
tool information.
events are preventable.
One major challenge to improving these outcomes is that
surgeons currently lack individualized, objective feedback
on their surgical technique and how to improve it. Manual
assessment of surgeon performance requires expert supervision which is both subjective and time consuming, with
many operations lasting hours. Real time automated surgical video analysis could provide a way to objectively and
efficiently assess surgical skill.
A number of works address frame-level tool detection
in laparoscopic surgical videos, including [6, 7, 8, 9], as
a part of the 2016 M2CAI Tool Presence Detection Challenge which includes a m2cai16-tool dataset [10]. However, rich analysis of surgeon performance involves analyz-
ing tool movement such as movement range and economy
of motion, and requires detecting the spatial bounds of tools
in addition to their presence.
In this work, we address this task of spatial tool detection
in laparoscopic surgical videos, which to our knowledge has
not been previously studied. In order to study this, we introduce a new dataset, m2cai16-tool-locations, which extends
the m2cai16-tool dataset with spatial bounds of tools. We
develop an approach leveraging region-based convolutional
neural networks (R-CNNs) to perform spatial detection of
tools, and show that our method is able to both effectively
detect the spatial bounds of tools as well as significantly
outperform previous work on detecting tool presence.
Finally, because our deep learning approach allows for
tool localization in addition to frame-level detection, it enables richer analysis of tool movements. We demonstrate
the ability of our method to assess surgical quality through
analysis of tool usage patterns, movement range, and economy of motion (Figure 1). We extract key quantitative and
qualitative metrics that are proven to reflect surgical skill
[11]. We collaborated with surgeons who manually reviewed each surgical video. Their assessments substantiated our findings, validating the ability of our method to
efficiently and accurately assess surgical quality. This enables not only avoiding the subjectivity inherent to human
assessment, but also significantly reduces the time it takes
to analyze a procedure.
In summary, we introduce m2cai16-tool-locations, a
new dataset extending the m2cai16-tool dataset with spatial
bounds of the tools. We use an approach leveraging regionbased convolutional neural networks to effectively perform
spatial detection of the tools, and show that conversion of
these detections to frame-level presence detections also significantly outperforms state-of-the-art on that task. Importantly, we demonstrate that our spatial detections in these
real-world laparoscopic surgical videos enables automatic
assessment of surgical quality through analysis of tool usage patterns, movement range, and economy of motion.
2. Related Work
Early work on surgical tool detection, categorization,
and tracking include those based on radio frequency identification (RFID) tags [12]; segmentation, contour processing and 3D modelling [13]; and the Viola-Jones detection
framework [14].
Furthermore, deep learning approaches based on convolutional neural networks have shown impressive performance on computer vision tasks [15], and works including
[6, 7, 8, 9] leverage deep learning architectures to achieve
state-of-the-art performance on surgical tool presence detection and phase recognition. As a part of the M2CAI 2016
Tool Presence Detection Challenge [10], they introduced a
benchmark for surgical tool presence detection.
While several existing studies address frame-level presence detection, Sarikaya et al. perform surgical tool localization in videos of robot-assisted surgical training tasks,
using multimodal convolutional neural networks [16]. Automated surgical scene understanding and skill assessment
are further areas of study. A few studies have addressed specific components of surgical video analysis, including surgical phase recognition and activity recognition. As a part
of the M2CAI 2016 Surgical Workflow Challenge, works
including [17, 18, 9, 19] address surgical phase recognition
in cholecystectomy videos. Additionally, Zia et al. specifically analyze task-specific suturing and knot-tying videos,
using symbolic, texture, and frequency features [20]. Similarly, Lalys et al. propose a framework using Hidden
Markov Model and visual features, such as shape, color, and
texture, to identify surgical tasks [14]. Lea et al. use skipchain conditional random field as well as handcrafted features to classify actions in short segments of robotic surgery
training tasks, including suturing and knot tying [21]. Also,
Reiter et al. perform surgical tool localization and pose
estimation. However, their approach is limited to robotic
arms that return kinematic data [22]. A limitation to these
task-specific studies is that surgical training tasks differ substantially from and do not accurately reflect surgical performance in real-world surgeries.
Our work builds on these prior contributions and uses
region-based convolutional neural networks [23] to detect
the spatial bounds of tools, enabling richer, more comprehensive assessment of surgical quality in real-world laparoscopic cholecystectomies, or minimally invasive surgical removal of the gallbladder. Moreover, while studies often only use short segments of procedures or of surgeons performing simulated training tasks to analyze surgical skill, our work leverages unedited, full-length surgical
operations. This differentiation is crucial for performance
assessment, since real-time operations include smoke, lens
fogging, variable anatomy, and different usage patterns not
found in simulation scenarios. Limited segments of real operations may not give a comprehensive assessment of surgical performance. Thus, both simulation and video segments of actual operations are limited in their ability to facilitate assessment of surgical performance. In contrast, our
extracted metrics are generated from comprehensive postoperative assessment of full surgical procedures. They have
the added benefit of being able to be correlated with postoperative clinical results, thus providing the link between
surgical skill and outcomes.
3. Dataset
There are few existing datasets on automated tool identification. Most center around frame-level tool presence
detection, including m2cai16-tool, which was released for
the M2CAI 2016 Tool Presence Detection Challenge, and
Figure 2: Top: the seven tools in m2cai16-tool-locations. Bottom: example frames with their spatial tool annotations. Color of the
bounding box corresponds to tool identity.
Tool
Number of annotated instances
Grasper
Bipolar
Hook
Scissors
Clipper
Irrigator
Specimen Bag
Total
Number of Frames
923
350
308
400
400
485
275
3141
2532
Table 1: Number of annotated frames for each tool.
Cholec80 [9]. Both m2cai16-tool and Cholec80 contain
cholecystectomy surgical videos performed at the University Hospital of Strasbourg in France, where each surgical
video frame is labeled with binary annotations indicating
tool presence. Sarikaya et al. also address tool localization
in videos of robot-assisted surgical training tasks [16]. We
expand upon these earlier works and introduce a new dataset
to study the task of surgical tool localization in real-world
laparoscopic surgeries and to enable higher-level analysis of
surgical videos.
In order to study this task, a dataset containing annotations of spatial bounds of tools is required. However, to the
best of our knowledge, no such dataset currently exists for
real-world laparoscopic surgical videos. We therefore collect and introduce a new dataset, m2cai16-tool-locations,
which extends the m2cai16-tool dataset [10] with spatial annotations of tools. We will publicly release this dataset.
m2cai16-tool consists of 15 videos recorded at 25 fps
of cholecystectomy procedures. Videos 1 to 10 are used
for training the R-CNN and videos 11 to 15 are used for
testing the model. The videos, whose durations range from
20 to 75 minutes, are downsampled to 1 fps for processing.
As a result, the dataset contains 23,000 frames labeled with
binary annotations indicating presence or absence of seven
surgical tools: grasper, bipolar, hook, scissors, clip applier,
irrigator, and specimen bag.
In m2cai16-tool-locations, we label 2532 of the frames,
under supervision and spot-checking from a surgeon, with
the coordinates of spatial bounding boxes around the tools.
We use 50%, 30%, and 20% for training, validation, and
test splits. The 2532 frames were selected from among the
23,000 total frames. We first annotate all frames containing just one tool, and increase the number of annotated instances per tool class by additionally labeling frames with
two and three tools. The breakdown of number of annotations per tool class is detailed in Table 1. Figure 2 shows
each tool in the dataset, its number of spatial annotations,
and examples of annotations in a number of frames.
4. Approach
Our approach for spatial detection of surgical tools is
based on Faster R-CNN [23], a region-based convolutional
neural network described in more detail below. The input
is a video frame, and the output is the spatial coordinates
of bounding boxes around any of the seven surgical instruments in Figure 2. This output allows us to perform qualitative and quantitative analyses of tool movements, from
tracking tool usage patterns to evaluating motion economy,
and to correlate these measures with surgical skill, ultimately setting the stage for higher-level analysis of surgical
performance.
The Faster R-CNN architecture we use is shown in Figure 3. The base network is a VGG-16 convolutional neural
network [24] with 16 convolutional layers, which extracts
powerful visual features. On top of this network is a region proposal network (RPN) that shares convolutional features with object detection networks. For each input image,
the RPN generates region proposals likely to contain an object, and features are pooled over these regions before being
passed to a final classification and bounding box refinement
network. The use of the RPN enables significant computational gains over related previous work, R-CNN [25] and
Fast R-CNN [26].
The RPN is trained by optimizing the following loss
function for each image:
Figure 3: Faster R-CNN architecture. The input to the network is a frame from a surgical video. The base network of Faster R-CNN is a
VGG-16 convolutional neural network. This is connected to a region proposal network (RPN) that shares convolutional features with
object detection networks. For each input image, the RPN generates region proposals likely to contain an object, and features are pooled
over these regions before being passed to a final classification and bounding box refinement network. The output is the spatial bounding
box positions of detected surgical tools in the video frame.
1 X
L({pi }, {ti }) =
Lcls (pi , p∗i )
Ncls i
1 X ∗
+λ
p Lreg (ti , t∗i )
Nreg i i
(1)
Here i indexes “anchors” corresponding to each sliding
window position of the input feature map, pi is the anchor’s
objectness probability, and ti is the coordinates of the predicted bounding box. p∗i is the ground-truth label of whether
an anchor is a true object location based on Intersection
over Union (IoU) with ground-truth annotations, and t∗i is
the coordinates of the ground-truth box corresponding to a
positive anchor. The loss function is therefore a weighted
combination of a classification loss Lcls for the binary objectness label and a regression loss Lreg for bounding box
coordinates. Ncls and Nreg are normalization constants and
λ weights the contributions of classification and regression.
The classification and bounding box refinement networks
on top of the pooled regions of interest are trained using
standard cross-entropy and regression loss functions.
While Faster R-CNN has shown impressive performance
on detection of everyday objects, the domain of surgical
videos and surgical tools has quite different visual characteristics. We pre-train the network on the ImageNet
dataset [15], where a large amount of data is available to
learn general visual features, and then fine-tune the network
on our m2cai16-tool-locations dataset, where a smaller
amount of data is labeled with the surgical tools of interest.
To train the RPN, we assign a binary objectness label
to each anchor at each sliding window position of the feature map. We also assign a positive label to anchors with
an overlap greater than 0.8 with the ground-truth box or if
those do not exist, to an anchor or anchors with the highest Intersection over Union (IoU), and a negative label to
anchors with an IoU of less than 0.3.
We fine-tune the VGG-16 network to optimize model
performance using stochastic gradient descent. We modify the classification layer of the network to output softmax
probabilities over the seven tools. All layers are fine-tuned
for 40K iterations with a mini-batch size of 50, and a 3×3
kernel size is used. We perform data augmentation by randomly flipping frames horizontally. The learning rate is initialized at 10−3 for all layers, and decreased by a factor of
10 every 10K iterations. Total training time was approximately two days on an NVIDIA Tesla K40 GPU, and the
network’s processing speed at deployment is 5 fps, achieving real-time surgical tool detection.
5. Results
In this section, we quantitatively evaluate our approach
on the tasks of spatial detection and frame-level presence
detection of surgical tools, using m2cai16-tool-locations
and m2cai16-tool, respectively. We demonstrate strong performance on the new task of spatial detection, and by leveraging spatial annotations, we significantly outperform existing works on presence detection. Finally, we qualitatively
illustrate the ability of our approach to analyze tool usage
patterns, movement range, and economy of motion for assessment of surgical performance.
5.1. Spatial detection
To the best of our knowledge, this study was the first
to perform surgical tool localization in real-world laparoscopic surgical videos, which would set the stage for richer
analysis of surgical performance. Table 2 presents performance using average precision (AP) on spatial detection of
Figure 4: Example frames of spatial detection results. Bounding box color corresponds to predicted tool identity. Correct predictions are
boxed in green (top), and mistakes are boxed in red (bottom). The model is able to successfully detect, classify, and localize surgical
instruments despite varying tool positions and angles, and despite some parts of the tools being occluded, as shown in column (c).
the surgical tools in m2cai16-tool-locations.
Tool
AP
Grasper
Bipolar
Hook
Scissors
Clipper
Irrigator
Specimen Bag
mAP
48.3
67.0
78.4
67.7
86.3
17.5
76.3
63.1
Table 2: Spatial detection average precision (AP) per-class and
mean average precision (mAP) in m2cai16-tool-locations. Clipper
achieves highest performance, likely due to its well-visualized usage pattern. Irrigator has notably low performance, likely due to
its generic shape (it looks like the pole of every other instrument)
and its lack of use in every procedure.
Intersection over Union (IoU) of 0.5 with a ground-truth
bounding box for a given class is considered to be a correct
detection. The overall mAP is 63.1, indicating strong performance overall. Clipper is the highest performing tool,
likely due to its usage pattern: while only present for a specific step in cholecystectomies, surgeons make sure it is well
visualized the entire time it is being used. Thus, its consistent usage pattern may have contributed to its high performance. The irrigator, on the other hand, is difficult likely
due to its generic shape (it looks like the pole of every other
instrument) and its lack of use in every procedure.
Figure 4 provides example frames of detection results.
Columns (a) through (d) display frames with increasing
numbers of tools present in them; from left to right, either 1, 2, or 3 tools are present per frame. We find that
our model is able to successfully detect, classify, and localize surgical instruments despite varying tool positions and
angles, and despite some parts of the tools being occluded,
as shown in column (c). (1) through (4) present incorrect or
partially incorrect detections. In (1), the clipper is mistaken
as a scissor; perhaps the angle of the tool renders its identity ambiguous. In (2), the center grasper bounding box is
a false positive; rather, the structure and shape of the gallbladder and liver form an outline that may appear to be a
grasper to the model. Interestingly, in both (3) and (4), the
poles of the hook and grasper, respectively, are identified as
an irrigator. This mistake is not surprising, as the irrigator
tip takes on the generic shape of any pole of a tool. Additionally, (4) presents a false negative, as the grasper in the
center of the frame is left undetected by the model. Overall,
our model has strong performance on the task of spatial tool
detection.
5.2. Frame-level presence detection
The spatial detections output from our model can also
be converted to frame-level presence predictions, and evaluated on the m2cai16-tool presence detection benchmark.
Table 3 presents AP performance on frame-level presence
detection of tools in m2cai16-tool.
Tool
AP
Grasper
Bipolar
Hook
Scissors
Clipper
Irrigator
Specimen Bag
mAP
87.2
75.1
95.3
70.8
88.4
73.5
82.1
81.8
Table 3: Frame-level presence detection average precision (AP)
per-class and mean average precision (mAP) in m2cai16-tool.
Hook achieves the highest performance, perhaps due to its distinct tip shape. Scissors has lowest performance, most likely due
to its brief usage during the operation and common two-pronged
tip shape.
Hook achieves the highest performance; one possible explanation is that it has a distinct tip shape, making it easily
distinguishable from other tools. Clipper, grasper, and specimen bag also perform well, while bipolar and irrigator are
more frequently misidentified, perhaps due to their generic
tip shapes as well as sparse and irregular appearances, as
usage of these tools is not absolutely essential during the
cholecystectomy procedure. For the bipolar and irrigator,
usage is conditional on the real-time circumstances of the
operation. The bipolar, an instrument for dissection and
hemostasis, is used to prevent likely bleeding, or to cauterize bleeding when it does occur. The irrigator, an instrument
used for flushing and suctioning an area, is typically used
during a cholecystectomy procedure when blood accumulates in the surgical field. Scissors has lowest performance,
most likely due to its brief usage during the operation and
common two-pronged tip shape.
Figure 5 compares mAP performance with winners of
the 2016 M2CAI Tool Presence Detection Challenge [10].
By leveraging the new spatial annotations during training,
our approach is able to significantly outperform all previous work. In comparison to the M2CAI 2016 Tool Presence
Figure 5: Comparison with winners of the M2CAI Tool
Presence Detection Challenge [10] on frame-level presence
detection in m2cai16-tool. By leveraging the new spatial
annotations during training, our approach is able to significantly
outperform all previous work.
Detection Challenge studies, we achieve a 28% improvement in accuracy, increasing mAP from 63.8 to 81.8, using just 2,500 of the 23,000 frames—roughly a tenth of the
data—in m2cai16-tool, which the previous studies used in
its entirety to train their tool presence detection models. We
also surpass EndoNet [9], a novel CNN architecture based
on AlexNet [27], in the tool presence detection task.
5.3. Assessment of Surgical Performance
The current standard for surgical performance assessment is to have an expert surgeon observe the operation in
its entirety and provide feedback to the surgeon performing
the surgery. The Global Operative Assessment of Laparoscopic Skills (GOALS) rating system, is a validated rubric
for grading surgeons on laparoscopic surgery performance.
For each of the five assessment domains of GOALS—depth
perception, bimanual dexterity, efficiency, and tissue handling, and autonomy, surgeons are rated on a scale of 1 to 5,
from least to most technically proficient, totaling to a final
score out of 25.
Operative skill significantly impacts patient outcomes.
Birkmeyer et al. [28] has demonstrated that performance on
standardized ratings of surgical skill are correlated with patient complication rates and hospital readmissions. Despite
this, most surgeons get no formal feedback on their operative performance. The process of manually rating surgeries
is time-consuming and subject to bias. Motivated by these
factors, we use the output of our deep learning model, which
consists of frame-by-frame tool presence, identity, and location information, to extract key metrics that are proven
to reflect surgical skill, such as instrument usage times and
path length [11]. We also create visual depictions of the
progression of the operation, like instrument usage timelines and tool trajectory maps, to parallel the GOALS rating
system.
By working with a group of surgeons who each independently reviewed and rated each testing video manually, we
are able to validate our approach to assess surgical performance on the five testing videos in our dataset. In the fol-
Figure 6: Timelines (top), heat maps (middle), and trajectories (bottom) of tool usage for the testing videos 1 through 4 in m2cai16-tool.
In the timelines, (a)-(g) correspond to Grasper, Bipolar, Hook, Scissors, Clipper, Irrigator, and Specimen Bag, respectively. These metrics
effectively measure bimanual dexterity, efficiency, and overall operative skill and enable us to efficiently examine back and forth switching
of instruments, movement range, and motion patterns of tools. We find that testing video 2 correlates with the most well-executed surgery,
reflecting focused and skillful execution of each step of the surgical procedure. In contrast, the surgeons in the other testing videos have
much less economy of motion, handle the instruments with less dexterity, and struggle with certain parts of the procedure.
lowing paragraphs, we interweave the surgeons’ independent ratings and comments with our objective analyses of
the surgical performance in each video.
To assess instrument usage patterns, we generate timelines displaying tool usage over the course of each of the
five testing videos (Figure 6, top). In examining the timelines, we find that testing video 2 correlates with the most
well-executed surgery. The surgery itself is efficient, and
the number of times that different instruments are switched
out for one another is minimal, reflecting focused and skillful execution of each step of the surgical procedure. On the
contrary, the timeline for testing video 3 reflects inefficient
and poorer technique, with more frequent switching back
and forth of instruments.
Moreover, the instrument usage timelines and bar graphs
quantifying the total time each instrument is used reflect the
level of technical proficiency in handling tissue (Figure 7).
The longer presence and periodic appearance of the bipolar
in testing video 3 indicates rough tissue handling resulting
in tissue damage, as the bipolar is used to stop bleeding.
To study the movement range of the surgical instruments,
Figure 7: Total instrument usage times, by video. Reflecting level
of skill in handling tissue, the longer presence of the bipolar in
testing video 3 indicates tissue damage, as the bipolar is used to
stop bleeding.
we generate heat maps of bounding box occurrences and locations. Figure 6, middle shows the heat maps of the testing
videos. As an indicator of technical proficiency, especially
of bimanual dexterity and efficiency, the heat maps reveal
the surgical skill demonstrated in testing video 2 to be the
Depth Perception
Bimanual Dexterity
Efficiency
Tissue Handling
Total
Video 1
Video 2
Video 3
Video 4
2.67
3.00
2.00
2.33
10.00
4.67
4.67
4.67
4.67
18.67
2.33
2.00
2.33
2.67
9.33
3.67
3.33
3.00
3.33
13.33
Table 4: Average composite GOALS ratings for the four test-
Figure 8: Total distance traveled by the tools during the clipping
phase, by video. Testing video 2 reflects the most technically excellent surgery, with the greatest economy of motion and most deft
tool handling.
highest among the test set once again. Better surgeons generally handle instruments in a more focused region of the
operative field, exhibiting greater precision and economy of
motion.
Finally, our deep learning model for automated tool detection and localization enabled tool tracking to study motion patterns. To further gain insight into bimanual dexterity and efficiency, we generated tool trajectory maps for the
clipping phase of each of the five testing videos. We chose
to examine this phase in particular because placing clips on
the cystic artery and cystic duct is one of the most important steps of the cholecystectomy procedure. If these clips
are placed in the wrong place, or if they come loose, the
patient can suffer devastating complications.
The clipper (blue) and grasper (red) trajectories in testing
video 2 reflect the deftness with which the surgeon placed
the clips. The trajectory maps for testing videos 2 and 3
also provide valuable information for evaluating bimanual
dexterity. While the clipper is used to clip the cystic artery
and cystic duct, the grasper is used to hold the gallbladder
in place and facilitate clip application. Based on the trajectory maps, the surgeons in testing videos 2 and 3 adeptly
maneuver the grasper in a way such that they do not need to
constantly adjust position, tension, or grip, in contrast to the
trajectory maps for testing videos 1, 4, and 5. Indeed, when
surgeons manually reviewed each of the testing videos to
rate them with GOALS, they found that the surgeon in testing video 1 struggled to place the clips properly. He or she
placed an additional clip, had difficulty prying it loose, and
ultimately removed it.
We also quantitatively measured efficiency by computing total distance traveled by the tools during the clipping
phase (Figure 8). Once again, testing video 2 reflects the
most technically excellent surgery, with the greatest economy of motion and most deft tool handling.
ing videos, including average subcomponent scores, as rated independently by the three surgeons. According to these ratings, the
surgeon in testing video 2 is the most skilled, with a high score
of 18.67 out of a maximum of 20.00, and the surgeons in testing
videos 1 and 3 need to improve their operative technique, with
scores of 10.00 and 9.33, respectively. These findings are consistent with our analyses based on the extracted assessment metrics,
proving our approach to be a much more efficient way to review
surgeon performance.
5.4. Validation of approach with GOALS
In addition to the blinded, subjective assessments of the
test videos embedded in the preceding paragraphs, the three
surgeons independently rated four of the test videos using
a modified version of the GOALS assessment rubric. For
this study the autonomy domain was omitted, which can
only be assessed in person and is not applicable to every
surgery. Thus, the total possible score for each video was
20. Test videos 1-4 received average composite scores of
10.00, 18.67, 9.33, and 13.33. The scores and subcomponent scores of GOALS, such as the efficiency score, directly
correlated with the metrics extracted by our algorithm, such
as the heat maps and tool trajectories. Table 4 presents the
GOALS surgeon ratings for each of the testing videos.
6. Conclusion
In this work, we present a new dataset m2cai16-toollocations and an approach based on region-based convolutional neural networks, to address the task of spatial tool
detection in real-world laparoscopic surgical videos. We
show that our method achieves strong performance on the
new spatial detection task, outperforms previous work on
frame-level presence detection, and runs in real-time at a
frame rate of 5fps. Furthermore, it can be used to extract
rich surgical assessment metrics such as tool usage patterns,
movement range, and economy of motion, which directly
correlate with independent assessments of the same videos
made by experienced surgeons. Future work includes continuing to build on the types of meaningful information that
can be automatically extracted from surgical videos, including smoothness of motion, tissue damage, repeated movements, and phase of procedure, as well as developing an
automated GOALS rating system.
References
[1] B. Alkire, N. Raykar, M. Shrime, T. Weiser, et al. Global access
to surgical care: a modelling study. The Lancet Global Health,
3(6):e316–e323, 2015.
[2] T. Weiser, S. Regenbogen, K. Thompson, et al. An estimation of the
global volume of surgery: a modelling strategy based on available
data. The Lancet, 372(9633):139–144, 2008.
[3] A. Gawande, E. Thomas, M. Zinner, and T. Brennan. The incidence
and nature of surgical adverse events in Colorado and Utah in 1992.
Surgery, 126(1):66–75, 1999.
[4] M. Healey, S. Shackford, T. Osler, F. Rogers, and E. Burns. Complications in surgical patients. Archives of Surgery, 137(5):611–618,
2002.
[5] A. Kable, R. Gibberd, and A. Spigelman. Adverse events in surgical
patients in Australia. International Journal for Quality in Health
Care, 14(4):269–276, 2002.
[6] M. Sahu, A. Mukhopadhyay, A. Szengel, and S. Zachow. Tool and
Phase recognition using contextual CNN features. 2016.
[7] A. Twinanda, D. Mutter, J. Marescaux, M. de Mathelin, and
N. Padoy.
Single- and Multi- Task Architectures for Tool
Presence Detection Challenge at M2CAI 2016. arXiv preprint
arXiv:1610.08851, 2016.
[8] A. Raju, S. Wang, and J. Huang. M2CAI Surgical Tool Detection
Challenge Report. 2016.
[9] A. Twinanda, S. Shehata, D. Mutter, J. Marescaux, M. de Mathelin,
and N. Padoy. EndoNet: A Deep Architecture for Recognition Tasks
on Laparoscopic Videos. IEEE Transactions on Medical Imaging,
36(1):86–97, 2016.
[10] Tool Presence Detection Challenge Results.
http:
//camma.u-strasbg.fr/m2cai2016/index.php/
tool-presence-detection-challenge-results.
[11] N. Stylopoulosu, S. Cotin, S. Dawson, et al. CELTS: a clinicallybased Computer Enhanced Laparoscopic Training System. Medicine
Meets Virtual Reality 11: NextMed: Health Horizon, 94:336, 2003.
[12] M. Kranzfelder, A. Schneider, A. Fiolka, et al. Real-time instrument
detection in minimally invasive surgery using radiofrequency identification technology. Journal of Surgical Research, 185(2):704–710,
2013.
[13] S. Speidel, J. Benzko, S. Krappe, et al. Automatic classification
of minimally invasive instruments based on endoscopic image sequences. In SPIE medical imaging, pages 72610A–72610A. International Society for Optics and Photonics, 2009.
[14] F. Lalys, L. Riffaud, D. Bouget, and P. Jannin. A framework for
the recognition of high-level surgical tasks from video images for
cataract surgeries. IEEE Transactions on Biomedical Engineering,
59(4):966–976, 2012.
[15] O. Russakovsky, J. Deng, H. Su, J. Krause, S. Satheesh, S. Ma,
Z. Huang, A. Karpathy, A. Khosla, M. Bernstein, et al. ImageNet
Large Scale Visual Recognition Challenge. International Journal of
Computer Vision, 115(3):211–252, 2015.
[16] D. Sarikaya, J. Corso, and G. Khurshid. Detection and localization
of robotic tools in robot-assisted surgery videos using deep neural
networks for region proposal and detection. IEEE Transactions on
Medical Imaging, 36(7):1542–1549, 2017.
[17] Y. Jin, Q. Dou, H. Chen, et al. EndoRCN: Recurrent Convolutional
Networks for Recognition of Surgical Workflow in Cholecystectomy
Procedure Video. 2016.
[18] R. Cadene, T. Robert, N. Thome, and M. Cord. M2CAI Workflow
Challenge: Convolutional Neural Networks with Time Smoothing
and Hidden Markov Model for Video Frames Classification. 2016.
[19] R. Stauder, D. Ostler, M. Kranzfelder, et al. The TUM LapChole
dataset for the M2CAI 2016 workflow challenge. 2017.
[20] A. Zia, Y. Sharma, V. Bettadapura, E. Sarin, et al. Automated videobased assessment of surgical skills for training and evaluation in
medical schools. International Journal of Computer Assisted Radiology and Surgery, 11(9):1623–1636, 2016.
[21] C. Lea, G. Hager, and R. Vidal. An improved model for segmentation
and recognition of fine-grained activities with application to surgical
training tasks. IEEE, 2015.
[22] A. Reiter, P. Allen, and T Zhao. Articulated surgical tool detection
using virtually-rendered templates. Computer Assisted Radiology
and Surgery, 2012.
[23] S. Ren, K. He, R. Girshick, and J. Sun. Faster R-CNN: Towards RealTime Object Detection with Region Proposal Networks. In Advances
in neural information processing systems, pages 91–99, 2015.
[24] K. Simonyan and A. Zisserman. Very deep convolutional networks
for large-scale image recognition. arXiv preprint arXiv:1409.1556,
2014.
[25] R. Girshick, J. Donahue, T. Darrell, and J. Malik. Rich feature hierarchies for accurate object detection and semantic segmentation. In
Proceedings of the IEEE conference on computer vision and pattern
recognition, pages 580–587, 2014.
[26] R. Girshick. Fast R-CNN. In Proceedings of the IEEE International
Conference on Computer Vision, pages 1440–1448, 2015.
[27] A. Krizhevsky, I. Sutskever, and G. Hinton. ImageNet Classification
with Deep Convolutional Neural Networks. 2012.
[28] J. Birkmeyer, J. Finks, A. O’reilly, et al. Surgical Skill and Complication Rates after Bariatric Surgery. New England Journal of Medicine,
369(15):1434–1442, 2013.
| 1cs.CV
|
SOME ALGEBRAS WITH THE WEAK LEFSCHETZ PROPERTY
arXiv:1606.01812v1 [math.AC] 6 Jun 2016
DAVID COOK II AND UWE NAGEL
Dedicated to Winfried Bruns at the occasion of his 70th birthday.
Abstract. Using a connection to lozenge tilings of triangular regions, we establish an
easily checkable criterion that guarantees the weak Lefschetz property of a quotient by a
monomial ideal. It is also shown that each such ideal also has a semistable syzygy bundle.
1. Introduction
Recently, there have been many investigations of the presence of the weak Lefschetz property (see, e.g., [1, 2, 4, 9, 13, 14, 15, 16, 17, 18]). A standard graded Artinian algebra A
over a field K is said to have the weak Lefschetz property if there is a linear form ℓ ∈ A such
that the multiplication map ×ℓ : [A]i → [A]i+1 has maximal rank for all i (i.e., it is injective
or surjective). The name is a reminder that the Hard Lefschetz Theorem guarantees this
property for the cohomology ring of a compact Kähler manifold. It is a desirable property,
e.g., its presence constrains the Hilbert function (see [11]). Many algebras are expected to
have the weak Lefschetz property. However, establishing this fact is often very challenging.
The recent lecture notes of Harima, Maeno, Morita, Numata, Wachi, and Watanabe [10]
and the survey [17] provide an excellent review of the state of knowledge on the Lefschetz
properties.
The authors have developed a combinatorial approach towards deciding the presence of
the weak Lefschetz property for monomial algebras in three variables in [5, 7]. It relies on
a study of lozenge tilings of so-called triangular regions. The needed results are recalled in
Subsection 2.1. This approach has been used, for example, to investigate the weak Lefschetz
property of monomial algebras of type two [7] and quotients by ideals with fours generators
[8]. Furthermore, we showed in [8] that there is a connection between lozenge tilings and
semistability of syzygy bundles (see Theorem 2.2). Here we use these methods to establish
sufficient conditions for the presence of the weak Lefschetz property as well as the semistability of the syzygy bundle (see Theorem 3.8). These conditions are easily expressed and
checked using triangular regions (see Remark 3.9). As a consequence, we describe infinite
families of ideals which satisfy these conditions. The number of generators of these ideals
can be arbitrarily large. As a simple example, consider the following ideal
I = (x12 , x6 y 2 z 3 , x3 y 2z 7 , xy 7 z 3 , xy 5 z 5 , xyz 9 , y 12 , z 12 ).
Our criterion immediately gives that its quotient has the weak Lefschetz property.
2010 Mathematics Subject Classification. 05B45, 05E40, 13E10.
Key words and phrases. Monomial ideals, weak Lefschetz property, lozenge tilings.
The second author was partially supported by the National Security Agency under Grant Number H9823012-1-0247 and by the Simons Foundation under grants #208869 and #317096.
1
2
D. COOK II AND U. NAGEL
2. Algebraic triangular regions
We recall facts needed to establish our main results in the following section.
Let R = K[x, y, z] be the standard graded polynomial ring over the field K. Unless
specified otherwise, K is always a field of characteristic zero. All R-modules are assumed to
be finitely generated and graded.
Let I be a monomial ideal of R. Then A = R/I = ⊕j≥0 [A]j is the sum of finite vector
spaces, called the homogeneous components (of A) of degree j. The Hilbert function of
A is the function hA : Z → Z given by hA (j) = dimK [A]j . The socle of A, denoted
soc A, is the annihilator of m = (x, y, z), the homogeneous maximal ideal of R, that is,
soc A = {a ∈ A | a · m = 0}.
2.1. Triangular regions labeled by monomials.
Now we briefly review a connection between monomial ideals and triangular regions; for
a more thorough discussion see [5].
For an integer d ≥ 1, the triangular region is the triangular region (of R) in degree d,
denoted Td , is an equilateral triangle of side length d composed of d2 downward-pointing
(▽) and d+1
upward-pointing (△) equilateral unit triangles. These triangles are labeled by
2
the monomials in [R]d−2 and [R]d−1 , respectively, as follows: place xd−1 at the top y d−1 at the
bottom-left, and z d−1 at the bottom-right; the remaining labels are found via interpolation.
See Figure 2.1(i) for an illustration.
(i) T4
(ii) T4 (xy, y 2, z 3 )
Figure 2.1. A triangular region with respect to R and with respect to R/I.
The triangular region (of R/I) in degree d, denoted by Td (I), is the part of Td that is
obtained after removing the triangles labeled by monomials in I. Note that the labels of the
downward- and upward-pointing triangles in Td (I) form K-bases of [R/I]d−2 and [R/I]d−1 ,
respectively. See Figure 2.1(ii) for an example.
Notice that the regions missing from Td in Td (I) can be viewed as a union of (possibly
overlapping) upward-pointing triangles of various side lengths that include the upward- and
downward-pointing triangles inside them. Each of these upward-pointing triangles corresponds to a minimal generator xa y b z c of I that has, necessarily, degree at most d − 1. The
value d − (a + b + c) is the side length of the puncture associated to xa y b z c , regardless of
possible overlaps with other punctures. See Figure 2.2 for an example.
We say that two punctures overlap if they share at least an edge. Two punctures are said
to be touching if they share precisely a vertex.
SOME ALGEBRAS WITH THE WEAK LEFSCHETZ PROPERTY
(i) Td (xa y b z c )
3
(ii) T10 (xy 3 z 2 )
Figure 2.2. Td (I) as constructed by removing punctures.
2.2. Tilings & the weak Lefschetz property.
A lozenge is a union of two unit equilateral triangles glued together along a shared edge,
i.e., a rhombus with unit side lengths and angles of 60◦ and 120◦ . Lozenges are also called
calissons and diamonds in the literature. See Figure 2.3.
Figure 2.3. A triangular region T ⊂ T8 together with on of its 13 tilings.
Fix a positive integer d and consider the triangular region Td as a union of unit triangles.
Thus a subregion T ⊂ Td is a subset of such triangles. We retain their labels. As above, we say
that a subregion T is ▽-heavy, △-heavy, or balanced if there are more downward pointing
than upward pointing triangles or less, or if their numbers are the same, respectively. A
subregion is tileable if either it is empty or there exists a tiling of the region by lozenges
such that every triangle is part of exactly one lozenge. Since a lozenge in Td is the union of
a downward-pointing and an upward-pointing triangle, and every triangle is part of exactly
one lozenge, a tileable subregion is necessarily balanced.
Let T ⊂ Td be any subregion. Given a monomial xa y b z c with degree less than d, the
monomial subregion of T associated to xa y b z c is the part of T contained in the triangle a
units from the bottom edge, b units from the upper-right edge, and c units from the upperleft edge. In other words, this monomial subregion consists of the triangles that are in T
and the puncture associated to the monomial xa y b z c .
We previously established a characterization of tileable triangular regions associated to
monomial ideals.
Theorem 2.1. [5, Theorem 2.2] Let T = Td (I) be a balanced triangular region, where
I ⊂ R is any monomial ideal. Then T is tileable if and only if T has no ▽-heavy monomial
subregions.
4
D. COOK II AND U. NAGEL
A subregion T ⊂ Td can be associated to a bipartite planar graph G that is an induced
subgraph of the honeycomb graph. Lozenge tilings of T can be then related to perfect
matchings on G. The connection was used by Kuperberg in [12], the earliest citation known
to the authors, to study symmetries on plane partitions.
Using this connection, the bi-adjacency matrix of T is the bi-adjacency matrix Z(T ) :=
Z(G) of the graph G associated to T . If T = Td (I) for some Artinian ideal I, then Z(T ) is the
transpose of the matrix defined by ×(x + y + z) : [R/I]d−2 → [R/I]d−1 using monomial bases
in the reverse-lexicographic order (see [7, Proposition 4.5]). Based on results by Migliore,
Miró-Roig, and the second author [15], we established the following criterion for the presence
of the weak Lefschetz property.
Theorem 2.2. [7, Corollary 4.7] Let I be an Artinian monomial ideal in R = K[x, y, z].
Then R/I has the weak Lefschetz property if and only if, for each positive integer d, the
matrix Z(Td (I)) has maximal rank.
It is well-known that the permanent of Z(G) enumerates the perfect matchings of G. A
perfect matching of G can be signed via the permutation it generates; thus a lozenge tiling
τ of T can be similarly signed (see [5]). Hence the signed tilings of the region are related to
the determinant of the bi-adjacency matrix.
Theorem 2.3. [5, Theorem 3.5] If T ⊂ Td is a non-empty balanced subregion, then the
signed lozenge tilings of T are enumerated by det Z(T ).
We recursively define a puncture of T ⊂ Td to be a non-floating puncture if it touches the
boundary of Td or if it overlaps or touches a non-floating puncture of T . Otherwise we call a
puncture a floating puncture. We further have the regions with only even floating punctures
have an easier enumeration.
Corollary 2.4. [5, Corollary 4.7] Let T be a tileable triangular region, and suppose all floating
punctures of T have an even side length. Then every lozenge tiling of T has the same perfect
matching sign as well as the same lattice path sign, and so perm Z(T ) = | det Z(T )| > 0.
2.3. Tileability & semistability of the syzygy bundle.
Let I be a monomial ideal of R whose punctures in Td (corresponding to the minimal
generators of I having degree less than d) have side lengths that sum to m. Then we define
the over-puncturing coefficient of I in degree d to be od (I) = m − d. If od (I) < 0, od (I) = 0,
or od (I) > 0, then we call I under-punctured, perfectly-punctured, or over-punctured in degree
d, respectively.
Let now T = Td (I) be a triangular region with punctures whose side lengths sum to m.
Then we define similarly the over-puncturing coefficient of T to be oT = m − d. If oT < 0,
oT = 0, or oT > 0, then we call T under-punctured, perfectly-punctured, or over-punctured,
respectively. Note that oT = od (J(T )) ≤ od (I), and equality is true if and only if the ideals
I and J(T ) are the same in all degrees less than d.
Perfectly-punctured regions admit a numerical tileability criterion.
Proposition 2.5. [8, Corollary 2.4] Let T = Td (I) be a triangular region. Then any two of
the following conditions imply the third:
(i) T is perfectly-punctured;
(ii) T has no over-punctured monomial subregions; and
(iii) T is tileable.
SOME ALGEBRAS WITH THE WEAK LEFSCHETZ PROPERTY
5
Let I be an Artinian ideal of S = K[x1 , . . . , xn ] that is minimally generated by forms
f1 , . . . , fm . The syzygy module of I is the graded module syz I that fits into the exact
sequence
m
M
0 → syz I →
S(− deg fi ) → I → 0.
i=1
Its sheafification syzI
f is a vector bundle on Pn−1 , called the syzygy bundle of I. It has rank
m − 1.
c1 (E)
Let E be a vector bundle on projective space. The slope of E is defined as µ(E) := rk(E)
.
Furthermore, E is said to be semistable if the inequality µ(F ) ≤ µ(E) holds for every
coherent subsheaf F ⊂ E.
Using the characterization of semistability for monomial ideals given by Brenner [3], we
previously established a connection to tileability.
Theorem 2.6. [8, Theorem 3.3] Let I be an Artinian ideal in R = K[x, y, z] generated by
monomials whose degrees are bounded above by d, and let T = Td (I). If T is non-empty,
then any two of the following conditions imply the third:
(i) I is perfectly-punctured;
(ii) T is tileable; and
(iii) syzI
f is semistable.
We note that Brenner and Kaid showed in [4, Corollary 3.3] that, for almost complete
intersections, nonsemistability implies the weak Lefschetz property in characteristic zero.
3. The criterion
In this section we will establish sufficient conditions that guarantee the presence of the
weak Lefschetz property. We use this result to exhibit explicit infinite families of ideals such
that their quotients have the weak Lefschetz property.
We begin with a necessary condition for the presence of the weak Lefschetz property.
Proposition 3.1. Let I be a monomial ideal such that Td (I) is a balanced region that is not
tileable. Put J = I +(xd , y d , z d ). Then R/J never has the weak Lefschetz property, regardless
of the characteristic of K.
Proof. Since Td (I) = Td (J) is not tileable, Theorem 2.3 gives det Z(Td (J)) = 0. Thus,
Z(Td (J)) does not have maximal rank. Now we conclude by Theorem 2.2.
We illustrate the preceding proposition with an example.
Example 3.2. Consider the regions depicted in Figure 3.1. These regions are both balanced, but non-tileable as they contain ▽-heavy monomial subregions (see Theorem 2.1).
In particular, the monomial subregion associated to xy 2 z in T and the monomial subregion
associated to xy 2 z 2 in T ′ are both ▽-heavy. Thus, R/(x6 , y 7, z 8 , xy 5 z, xy 2 z 3 , x3 y 2z) and
R/(x6 , y 7, z 7 , xy 4 z 2 , xy 2z 4 , x2 y 2 z 2 ) both fail to have the weak Lefschetz property, regardless
of the characteristic of the base field.
Now we use Proposition 3.1 in order to relate the weak Lefschetz property and semistability
of syzygy bundles (see Section 2.3). In preparation, we record the following observation.
Recall that the monomial ideal of a triangular region T ⊂ Td is the largest ideal J whose
minimal generators have degrees less than d such that T = Td (J).
6
D. COOK II AND U. NAGEL
(i) T = T8 (x6 , y 7 , z 8 , xy 5z, xy 2 z 3 , x3 y 2z)
(ii) T ′ = T8 (x6 , y 7, z 7 , xy 4 z 2 , xy 2 z 4 , x2 y 2z 2 )
Figure 3.1. Two balanced non-tileable triangular regions.
Lemma 3.3. Let J ⊂ R be the monomial ideal of a triangular region T ⊂ Td . Then:
(i) The region T has no overlapping punctures if and only if each degree of a least
common multiple of two distinct minimal generators of J is at least d.
(ii) The punctures of T are not overlapping nor touching if and only if each degree of a
least common multiple of two distinct minimal generators of J is at least d + 1.
Proof. Let m1 and m2 be two distinct minimal generators of J. Then their associated
punctures overlap if and only if there is a monomial of degree d − 1 that is a multiple of
m1 and m2 . The existence of such a monomial means precisely that the degree of the least
common multiple of m1 and m2 is at most d − 1. Now claim (i) follows.
Assertion (ii) is shown similarly by observing that the punctures to m1 and m2 touch if
and only if there is a monomial of degree d that is a multiple of m1 and m2 .
The following consequence is useful later on.
Corollary 3.4. Assume T ⊂ Td is a triangular region whose punctures are not overlapping
nor touching, and let J be the monomial ideal of T . Then R/J does not have non-zero socle
elements of degree less than d − 1.
Proof. Since J is a monomial ideal, every minimal first syzygy of J corresponds to a relation
mi ni − mj nj = 0 for suitable monomials ni and nj , where mi and mj are distinct monomial
minimal generators of J. Applying Lemma 3.3 to the equality mi ni = mj nj , we conclude
that the degree of each first syzygy of J is at least d + 1. It follows that the degree of every
second syzygy of J is at least d + 2. Each minimal second syzygy of J corresponds to a socle
generator of R/J. Hence, the degrees of the socle generators of R/J are at least d − 1 .
Remark 3.5. The converse of Corollary 3.4 is not true in general. For example, the socle
generators of R/(x6 , y 7, z 8 , xy 5z, xy 2 z 3 , x3 y 2z) have degrees greater than 7, but two punctures
of T8 (x6 , y 7, z 8 , xy 5z, xy 2 z 3 , x3 y 2z) touch each other (see Figure 3.1).
Recall that perfectly-punctured regions were defined above Proposition 2.5. This concept
is used in the proof of the following result.
Theorem 3.6. Let I ⊂ R be an Artinian ideal whose minimal monomial generators have
degrees d1 , . . . , dt . Set
d1 + · · · + dt
d :=
.
t−1
Assume the following conditions are satisfied:
(i) The number d is an integer.
SOME ALGEBRAS WITH THE WEAK LEFSCHETZ PROPERTY
7
(ii) For all i = 1, . . . , t, one has d > di .
(iii) Each degree of a least common multiple of two distinct minimal generators of I is
at least d.
Then the syzygy bundle of I is semistable if R/I has the weak Lefschetz property.
Proof. Consider the triangular region T = Td (I). By assumption (iii) and Lemma 3.3, we
obtain that the punctures of T do not overlap. Recall that the side length of the puncture
to a minimal generator of degree di is d − di . The definition of d is equivalent to
t
X
d=
(d − di ).
i=1
We conclude that the region T is balanced and perfectly-punctured. Combined with the weak
Lefschetz property of R/I, the first property implies that T is tileable by Proposition 3.1.
Now Theorem 2.6 gives the semistability of the syzygy bundle of I.
The converse of the above result is not true, in general.
Example 3.7. Consider the ideal
J = (x7 , x4 y 2z 2 , xy 3 z 3 , y 7 , z 7 ).
Its syzygy bundle is semistable, and its triangular region T9 (J) is perfectly-puntured. However, R/J does not have the weak Lefschetz property. Notice though that R/I has the weak
Lefschetz property, where I is the very similar ideal
I = (x7 , x5 yz, xy 3 z 3 , y 7, z 8 ).
Both regions, T9 (I) and T9 (J) are tileable and symmetric. In fact, they are examples of
mirror symmetric regions that are studied in [6]
Under stronger assumptions the converse to Theorem 3.6 is indeed true.
Theorem 3.8. Let I ⊂ R be an Artinian ideal with minimal monomial generators m1 , . . . , mt .
Set
d1 + · · · + dt
d :=
,
t−1
where di = deg mi . Assume the following conditions are satisfied:
(i) The number d is an integer.
(ii) For all i = 1, . . . , t, one has d > di .
(iii) If i 6= j, then the degree of the least common multiple of mi and mj is at least d + 1.
(iv) If mi is not a power of x, y, or z, then d − di is even.
Then the syzygy bundle of I is semistable if and only if R/I has the weak Lefschetz property.
Proof. By Theorem 3.6, it is enough to show that R/I has the weak Lefschetz property if
the syzygy bundle of I is semistable.
Consider the region T = Td (I). In the proof of Theorem 3.6 we showed that T is balanced
and perfectly-punctured. Hence T is tileable by Theorem 2.6. Since all floating punctures
of T have an even side length by assumption (iv), Theorem 2.3 and Corollary 2.4 give that
Z(T ) has maximal rank.
Assumption (iii) means that the punctures of T are not overlapping nor touching (see
Lemma 3.3). Hence, Corollary 3.4 yields that the degrees of the socle generators of R/I are
at least d − 1. Therefore, Theorem 2.2 proves that R/I has the weak Lefschetz property.
8
D. COOK II AND U. NAGEL
Remark 3.9. The assumptions of Theorem 3.8 have the following interpretation using the
triangular region Td (I). Assumptions (ii) means that each minimal generator of I gives
a puncture of positive side length of Td (I). Condition (i) expresses the fact that the side
lengths of all punctures add up to d, that is, Td (I) is perfectly punctured. Assumption
(iv) says that all non-corner punctures have an even side length, and Condition (iii) is the
requirement that no punctures of Td (I) touch or overlap.
We now show that, for all positive integers d1 , . . . , dt with t ≥ 3 that satisfy the numerical
assumptions (i), (ii), and (iv) of Theorem 3.8, there is a monomial ideal I whose minimal
generators have degrees d1 , . . . , dt to which Theorem 3.8 applies and guarantees the weak
Lefschetz property of R/I.
Example 3.10. Let d1 , . . . , dt be t ≥ 3 positive integers satisfying the following numerical
conditions:
t
(i) The number d := d1 +···+d
is an integer.
t−1
(ii) For all i = 1, . . . , t, one has d > di .
(iii) At most three of the integers d − di are not even.
Re-indexing if needed, we may assume that d3 ≤ min{d1 , d2} and that d−di is even whenever
4 ≤ i ≤ t. Consider the following ideal
I = (xd1 , y d2 , z d3 , m4 , . . . , mt ),
where m4 = xd−d3 yz −d−1+d3 +d4 if t ≥ 4, m5 = x2d−d3 −d4 y 2z 2d−2+d3 +d4 +d5 if t ≥ 5, and
(
Pi−1
Pi
xd−d3 y 1+ k=4 (d−dk ) z −d(i−3)−1+ k=3 dk if 6 ≤ i ≤ t and i is even
Pi−1
Pi
mi =
x−1+ k=3 (d−dk ) y 2 z −d(i−3)−1+ k=3 dk
if 7 ≤ i ≤ t and i is odd.
Note that deg mi = di for all i. One easily checks that the degree of the least common
multiple of any two distinct minimal generators of I is at least d + 1, that is, the punctures
of Td (I) do not overlap nor touch each other.
Figure 3.2. The region corresponding to d1 = d2 = d3 = 12 and d4 = · · · =
d8 = 11 in Example 3.10.
Corollary 3.11. If I is an ideal as defined in Example 3.10, then R/I has the weak Lefschetz
property and the syzygy bundle of I is semistable.
SOME ALGEBRAS WITH THE WEAK LEFSCHETZ PROPERTY
9
Proof. By construction, the considered ideals satisfy assumptions (i)–(iv) of Theorem 3.8.
Furthermore, the region Td (I) has no over-punctured monomial subregions. Hence, it is
tileable by Proposition 2.5. (Alternatively, one can exhibit a family of non-intersecting
lattice paths to check tileability.) By Theorem 2.6, it follows that the syzygy bundle of I is
semistable, and hence R/I has the weak Lefschetz property by Theorem 3.8.
Remark 3.12. Given an integer t ≥ 3, there are many choices for the integers d1 , . . . , dt , and
thus for the ideals exhibited in Example 3.10. A convenient choice, for which the description
of the ideal becomes simpler, is d1 = 2t − 4, d2 = d3 = d − 1, and d4 = · · · = dt = d − 2,
where d is any integer satisfying d ≥ 2t − 3. Then the corresponding ideal is
I = (x2t−4 , y d−1, z d−1 , xyz d−4 , x3 y 2z d−7 , m6 , . . . , mt ),
where
mi =
(
xy 2i−7 z d+4−2i
x2i−8 y 2z d+4−2i
if 6 ≤ i ≤ t and i is even
if 7 ≤ i ≤ t and i is odd.
If d = 13 and t = 8, then this gives the ideal I mentioned in the introduction. Its triangular
region T13 (I) is depicted in Figure 3.2.
References
[1] M. Boij, J. Migliore, R. Miró-Roig, U. Nagel, F. Zanello, On the shape of a pure O-sequence, Mem.
Amer. Math. Soc. 218 (2012), no. 1024, vii+78 pp.
[2] M. Boij, J. Migliore, R. Miró-Roig, U. Nagel, F. Zanello, On the Weak Lefschetz Property for Artinian
Gorenstein algebras of codimension three, J. Algebra 403 (2014), 48–68.
[3] H. Brenner, Looking out for stable syzygy bundles, Adv. Math. 219 (2008), no. 2, 401–427.
[4] H. Brenner, A. Kaid, Syzygy bundles on P2 and the Weak Lefschetz Property, Illinois J. Math. 51 (2007),
no. 4, 1299–1308.
[5] D. Cook II, U. Nagel, Signed lozenge tilings, Preprint, 2015; arXiv:1507.02507.
[6] D. Cook II, U. Nagel, Signed lozenge tilings of mirror symmetric regions, in preparation.
[7] D. Cook II, U. Nagel, The weak Lefschetz property for monomial ideals of small type, Preprint, 2015;
arXiv:1507.3853.
[8] D. Cook II, U. Nagel, Syzygy bundles and the weak Lefschetz property of almost complete intersections,
Preprint, 2016.
[9] B. Harbourne, H. Schenck, A. Seceleanu, Inverse systems, Gelfand-Tsetlin patterns and the weak Lefschetz property, J. Lond. Math. Soc. (2) 84 (2011), no. 3, 712–730.
[10] T. Harima, T. Maeno, H. Morita, Y. Numata, A. Wachi, J. Watanabe, The Lefschetz properties, Lecture
Notes in Mathematics, 2080, Springer, 2013.
[11] T. Harima, J. Migliore, U. Nagel, J. Watanabe, The weak and strong Lefschetz properties for Artinian
K-algebras, J. Algebra 262 (2003), no. 1, 99–126.
[12] G. Kuperberg, Symmetries of plane partitions and the permanent-determinant method, J. Combin.
Theory Ser. A 68 (1994), no. 1, 115–151.
[13] A. Kustin, H. Rahmati, A. Vraciu, The resolution of the bracket powers of the maximal ideal in a
diagonal hypersurface ring, J. Algebra 369 (2012), 256–321.
[14] J. Li, F. Zanello, Monomial complete intersections, the weak Lefschetz property and plane partitions,
Discrete Math. 310 (2010), no. 24, 3558–3570.
[15] J. Migliore, R. Miró-Roig, U. Nagel, Monomial ideals, almost complete intersections and the weak
Lefschetz property, Trans. Amer. Math. Soc. 363 (2011), no. 1, 229–257.
[16] J. Migliore, R. Miró-Roig, U. Nagel, On the weak Lefschetz property for powers of linear forms, Algebra
Number Theory 6 (2012), no. 3, 487–526.
[17] J. Migliore, U. Nagel, A tour of the weak and strong Lefschetz properties, J. Commut. Algebra 5 (2013),
329–358.
10
D. COOK II AND U. NAGEL
[18] R. P. Stanley, Weyl groups, the hard Lefschetz theorem, and the Sperner property, SIAM J. Algebraic
Discrete Methods 1 (1980), no. 2, 168–184.
Department of Mathematics & Computer Science, Eastern Illinois University, Charleston,
IL 46616
E-mail address: [email protected]
Department of Mathematics, University of Kentucky, 715 Patterson Office Tower, Lexington, KY 40506-0027
E-mail address: [email protected]
| 0math.AC
|
Fast Simulation of Probabilistic Boolean Networks
(Technical Report)
Andrzej Mizera, Jun Pang, and Qixia Yuan ∗
arXiv:1605.00854v1 [cs.CE] 28 Apr 2016
Computer Science and Communications, University of Luxembourg, Luxembourg
[email protected]
Abstract. Probabilistic Boolean networks (PBNs) is an important mathematical
framework widely used for modelling and analysing biological systems. PBNs
are suited for modelling large biological systems, which more and more often
arise in systems biology. However, the large system size poses a significant challenge to the analysis of PBNs, in particular, to the crucial analysis of their steadystate behaviour. Numerical methods for performing steady-state analyses suffer
from the state-space explosion problem, which makes the utilisation of statistical
methods the only viable approach. However, such methods require long simulations of PBNs, rendering the simulation speed a crucial efficiency factor. For
large PBNs and high estimation precision requirements, a slow simulation speed
becomes an obstacle. In this paper, we propose a structure-based method for fast
simulation of PBNs. This method first performs a network reduction operation
and then divides nodes into groups for parallel simulation. Experimental results
show that our method can lead to an approximately 10 times speedup for computing steady-state probabilities of a real-life biological network.
1
Introduction
Systems biology aims to model and analyse biological systems from a holistic perspective in order to provide a comprehensive, system-level understanding of cellular
behaviour. Computational modelling of a biological system plays a key role in systems
biology. It connects the field of traditional biology with mathematics and computational
science, providing a way to organize and formalize available biological knowledge
in a mathematical model and to identify missing biological information using formal
means. Together with biochemical techniques, computational modelling promotes the
holistic understanding of real-life biological systems, leading to the study of large biological systems. This brings a significant challenge to computational modelling in terms
of the state-space size of the system under study. Among the existing modelling frameworks, probabilistic Boolean networks (PBNs) is well-suited for modelling large-size
biological systems. It is first introduced by Shmulevich et al. [1,2] as a probabilistic
generalisation of the standard Boolean networks (BNs) to model gene regulatory networks (GRNs). The framework of PBNs incorporates rule-based dependencies between
genes and allows the systematic study of global network dynamics; meanwhile, it is capable of dealing with uncertainty, which naturally occurs at different levels in the study
of biological systems.
∗ Supported
by the National Research Fund, Luxembourg (grant 7814267).
2
Mizera, Pang and Yuan
Focusing on the wiring of a network, PBNs is essentially designed for revealing the
long-run (steady-state) behaviour of a biological system. Comprehensive understanding
of the long-run behaviour is vital in many contexts. For example, attractors of a gene
regulatory network (GRN) are considered to characterise cellular phenotypes [3]. There
have been a lot of studies in analysing the long-run behaviour of biological systems for
better understanding the influences of genes or molecules in the systems [4]. Moreover,
steady-state analyses have been used in gene intervention and external control [5,6],
which is of special interest to cancer therapists to predict the potential reaction of a patient to treatment. In the context of PBNs, many efforts have been devoted to computing their steady-state probabilities. In [7,8], efficient numerical methods are provided
for computing the steady-state probabilities of small-size PBNs. Those methods utilise
an important characteristics of PBNs, i.e., a PBN can be viewed as a discrete-time
Markov chain (DTMC) and its dynamics can be studied with the use of the rich theory
of DTMCs. The key idea of those methods relies on the computation of the transition
matrix of the underlying DTMC of the studied PBN. They perform well for small-size
PBNs. However, in the case of large-size PBNs, the state-space size becomes so huge
that the numerical methods are not scalable any more.
Many efforts are then spent on addressing the challenge of the huge state-space
in large-size PBNs. In fact, the use of statistical methods and Monte Carlo methods
remain the only feasible approach to address the problem. In those methods, the simulation speed is an important factor in the performance of these approaches. For large
PBNs and long trajectories, a slow simulation speed could render these methods infeasible as well. In our previous work [9], we have considered the two-state Markov
chain approach and the Skart method for approximate analysis of large PBNs. Taking
special care of efficient simulation, we have implemented these two methods in the tool
ASSA-PBN [10] and successfully used it for the analysis of large PBNs with a few
thousands of nodes. However, the required time cost is still expected to be reduced.
This requirement is of great importance for the construction of a model, e.g., parameter estimation, and for a more precise and deep analysis of the system. In this work,
we propose a structure-based method to speed up the simulation process. The method
is based on analysing the structure of a PBN and consists of two key ideas: first, it
removes the unnecessary nodes in the network to reduce its size; secondly, it divides
the nodes into groups and performs simulation for nodes in a group simultaneously.
We show with experiments that our structure-based method can significantly reduce the
computation time for approximate steady-state analyses of large PBNs. To the best of
our knowledge, our proposed method is the first one to apply structure-based analyses
for speeding up the simulation of a PBN.
Structure of the paper. After presenting preliminaries in Section 2, we describe our
structure-based simulation method in Section 3. We perform an extensive evaluation and
comparison of our method with the previous state-of-art methods in Section 4 on a large
number of randomly generated PBNs and a large real-life PBN model of apoptosis in
hepatocytes. We conclude our paper with some discussions in Section 5.
Fast Simulation of Probabilistic Boolean Networks
2
2.1
3
Preliminaries
Probabilistic Boolean networks (PBNs)
A PBN G(X, F) models elements of a biological system with a set of binary-valued
nodes X = {x1 , x2 , . . . , xn }. For each node xi ∈ X, the update of its value is guided by
(i) (i)
(i)
a set of predictor functions Fi = { f1 , f2 , . . . , f`(i) }, where `(i) is the number of pre(i)
dictor functions for node xi . Each f j is a Boolean function whose inputs are a subset
of nodes, referred to as parent nodes of xi . For each node xi , one of its predictor functions will be selected to update the value of xi at each time point t. This selection is in
(i) (i)
(i)
accordance with a probability distribution Ci = (c1 , c2 , . . . , c`(i) ), where the individual probabilities are the selection probabilities for the respective elements of Fi and they
sum to 1. The value of node xi at time point t is denoted as xi (t) and the state of the PBN
at time point t is denoted as s(t) = (x1 (t), x2 (t), . . . , xn (t)). The state space of the PBN
is S = {0, 1}n and it is of size 2n . There are several variants of PBNs with respect to the
selection of predictor functions and the synchronisation of nodes update. In this paper,
we consider the independent synchronous PBNs, i.e., the choice of predictor functions
for each node is made independently and the values of all the nodes are updated synchronously. The transition from state s(t) to state s(t + 1) is performed by randomly
selecting a predictor function for each node xi from Fi and by applying those selected
predictor functions to update the values of all the nodes synchronously. We denote f (t)
the combination of all the selected predictor functions at time point t. The transition of
state s(t) to s(t + 1) can then be denoted as s(t + 1) = f (t)(s(t)).
Perturbations of a biological system are introduced by a perturbation rate p ∈ (0, 1)
in a PBN. The dynamics of a PBN is guided with both perturbations and predictor
functions: at each time point t, the value of each node xi is flipped with probability
p; and if no flip happens, the value of each node xi is updated with selected predictor
functions synchronously. Let γ(t) = (γ1 (t), γ2 (t), . . . , γn (t)), where γi (t) ∈ {0, 1} and
P(γi (t) = 1) = p for all t and i ∈ {1, 2, . . . , n}. The transition of s(t) to s(t + 1) in PBNs
with perturbations is given by
(
s(t) ⊕ γ(t)
if γ(t) 6= 0
(1)
s(t + 1) =
f (t)(s(t))
otherwise,
where ⊕ is the element-wise exclusive or operator for vectors. According to Equation (1), perturbations allow the system to move from a state to any other state in one
transition, hence render the underlying Markov chain irreducible and aperiodic. Thus,
the dynamics of a PBN with perturbations can be viewed as an ergodic DTMC [1].
Based on the ergodic theory, the long-run dynamics of a PBN with perturbations is
governed by a unique limiting distribution, convergence to which is independent of the
choice of the initial state.
The density of a PBN is measured with its predictor function number and parent
F
φ (i), where n is
nodes number. For a PBN G, its density is defined as D(G) = n1 ∑Ni=1
the number of nodes in G, NF is the total number of predictor functions in G, and φ (i)
is the number of parent nodes for the ith predictor function.
4
2.2
Mizera, Pang and Yuan
Simulating a PBN
A PBN can be simulated via two steps based on its definition. First, perturbation is
verified for each individual node and a node value is flipped if there is a perturbation.
Second, if no perturbation happens for any of the nodes, the network state is updated
by selecting predictor functions for all the nodes and applying them. For efficiency
reason, the selection of predictor functions for each node xi is performed with the alias
method [11], which allows to make a selection among choices in constant time irrespective of the number of choices. The alias method requires the construction of an alias
table of size proportional to the number of choices, based on the selection probabilities
of Ci .
3
Structure-based Parallelisation
The simulation method described in the above section requires to check perturbations,
make a selection and perform updating a node for n times in each step. In the case of
large PBNs and huge trajectory (sample) size, the simulation time cost can become prohibitive. Intuitively, the simulation time can be reduced if the n-time operations can be
speeded up, for which we propose two solutions. One is to perform network reduction
such that the total number of nodes is reduced. The other is to preform node-grouping
in order to parallelise the process for checking perturbations, making selections, and
updating nodes. For the first solution, we analyse the PBN structure to identify those
nodes that can be removed and remove them to reduce the network size; while for the
second solution, we analyse the PBN structure to divide nodes into groups and perform the operations for nodes in a group simultaneously. We combine the two solutions
together and refer to this simulation technique as structure-based parallelisation. We
formalise the two solutions in the following three steps: the first solution is described in
step 1 and the second solution is described in steps 2 and 3.
Step 1. Remove unnecessary nodes from the PBN.
Step 2. Parallelise the perturbation process.
Step 3. Parallelise updating a PBN state with predictor functions.
We describe these three steps in the following subsections.
3.1
Removing unnecessary nodes
We first identify those nodes that can be removed and preform network reudction. When
simulating a PBN without perturbations, if a node does not affect any other node in the
PBN, the states of all other nodes will not be affected after removing this node. If this
node is not of interest of the analysis, e.g., we are not interested in analysing its steadystate, then this node is dispensable in a PBN without perturbations. We refer to such a
dispensable node as a leaf node in a PBN and define it as follow:
Definition 1 (Leaf node). A node in a PBN is a leaf node (or leaf for short) if and only
if either (1) it is not of interest and has no child nodes or (2) it is not of interest and has
no other children after iteratively removing all its child nodes which are leaf nodes.
Fast Simulation of Probabilistic Boolean Networks
5
Algorithm 1 Checking perturbations of leaf nodes in a PBN
1: procedure C HECK L EAF N ODES(t)
2:
if rand() > t then return true;
3:
else return false;
4:
end if
5: end procedure
According to the above definition, leaf nodes can be simply removed without affecting
the simulation of the remaining nodes in a PBN without perturbations. In the case of a
PBN with perturbations, perturbations in the leaf nodes need to be considered. Updating states with Boolean functions will only be performed when there is no perturbation
in both the leaf nodes and the non-leaf nodes. Perturbations of the leaf nodes can be
checked in constant time irrespective of the number of leaf nodes as describe in Algorithm 1. The input t in this algorithm is the probability that no perturbation happens in
all the leaf nodes. It can be computed easily as t = (1 − p)` , where p is the perturbation
rate for each node and ` is the number of leaf nodes in the PBN. With the consideration
of their perturbations, the leaf nodes can be removed without affecting the simulation of
the non-leaf nodes in a PBN with perturbations as well. Since the leaves are not of interest, results of analyses performed on the simulated trajectories of the reduced network,
i.e.,. containing only non-leaf nodes, will be the same as performed on trajectories of
the original network, i.e., containing all the nodes.
3.2
Performing perturbations in parallel
The second step of our method speeds up the process for determining perturbations.
Normally, perturbations are checked for nodes one by one. In order to speed up the
simulation of a PBN, we perform perturbations for k nodes simultaneously instead of
one by one. For those k nodes, there are 2k different perturbation situations. We compute
the probability for each situation and construct an alias table based on the distribution.
With the alias table, we make a choice c among 2k choices and perturb the corresponding nodes based on the choice. The choice c is an integer in [0, 2k ) and for the whole
network the perturbation can then be performed k nodes by k nodes using the logical bitwise exclusive or operation which outputs true only when inputs differ. To save
memory, the alias table can be reused for all the groups since the perturbation rate for
each node is the same. It might happen that the number of nodes in the last perturbation
round will be less than k nodes. Assume there is k0 nodes in the last round and k0 < k.
For those k0 nodes, we can reuse the same alias table to make the selection in order to
save memory. After getting the choice c, we perform c = c&m, where & is a bitwise
and operation and m is a mask constructed by setting the first k0 bits of m’s binary
representation to 1 and the remaining bits to 0.
Theorem 1. The above process for determining perturbations for the last k0 nodes
guarantees that the probability for each of the k0 nodes to be perturbed is still p.
Proof. Without loss of generality, we assume that in the last k0 nodes, t nodes should
be perturbed and the positions of the t nodes are fixed. The probability for those t
6
Mizera, Pang and Yuan
Algorithm 2 The group perturbation algorithm
1:
2:
3:
4:
5:
6:
7:
8:
9:
10:
11:
12:
13:
14:
15:
16:
17:
18:
19:
20:
21:
procedure P REPARE P ERTURBATION(n, k)
g = dn/ke; k = dn/ge; k0 = n − k ∗ (g − 1);
construct the alias table A p and mask;
return [A p , mask].
end procedure
procedure P ERTURBATION(A p , mask, s)
i = 0; perturbed = false;
repeat
c = Next(A p );
//Next(A p ) returns a random integer based on A p
if c! = 0 then
s = s ⊕ (c << (i ∗ k));
//shift c to flip only the bits (nodes) of current group
perturbed = true;
end if
i + +;
until i = g − 1
c = Next(A p ) & mask;
if c! = 0 then
s = s ⊕ (c << (i ∗ k)); perturbed = true;
end if
return [s, perturbed].
end procedure
0
fixed nodes to be perturbed is pt (1 − p)k −t . When we make a selection from the alias
0
table for k nodes, there are 2k−k different choices corresponding to the case that t fixed
0
position nodes in the last k nodes are perturbed. The sum of the probabilities of the
0
0
0
i
k−k0 −i = pt (1 − p)k0 −t .
2k−k different choices is [pt (1 − p)k −t ] · ∑k−k
t
u
i=0 p (1 − p)
We describe the process for constructing groups and performing perturbations based
on the groups in Algorithm 2. Algorithm 2 requires three inputs: n is the number of
nodes,1 k is the maximum number of nodes that can be perturbed simultaneously and
s is the PBN’s current state which is represented by an integer. As perturbing one node
equals to flipping one bit of s, perturbing nodes in a group is performed via a logical
bitwise exclusive or operation (see line 11 of Algorithm 2). Perturbing k nodes simultaneously requires 2k double numbers to store the probabilities of 2k different choices.
The size of k is therefore restricted by the available memory.2
3.3
Updating nodes in parallel
The last step to speed up PBN simulation is to update a number of nodes simultaneously
in accordance with their predictor functions. For this step, we need an initialisation pro1 In our methods, it is clear that steps 2 and 3 are independent of step 1. Thus, we consistently
use n to denote the number of nodes in a PBN.
2 For the experiments, we set k to 16 and k could be bigger as long as the memory allows.
However, a larger k requires larger table to store the 2k probabilities and the performance of
a CPU drops when accessing an element of a much larger table due to the large cache miss rate.
Fast Simulation of Probabilistic Boolean Networks
7
cess to divide the n nodes into m groups and compute the combined predictor functions
for each group. After this initialisation, we can select a combined predictor function for
each group based on a sampled random number and apply this combined function to
update the nodes in the group simultaneously.
We first describe how predictor functions of two nodes are combined. The combination of functions for more than two nodes can be performed iteratively. Let xα
and xβ be the two nodes to be considered. Their predictor functions are denoted as
(α)
(α)
(α)
(β )
(β )
(β )
Fα = { f1 , f2 , . . . , f`(α) } and Fβ = { f1 , f2 , . . . , f`(β ) }. The corresponding selec(α)
(α)
(β )
(β )
β
tion probabilities are denoted as Cα = {c1 , c2 , . . . , cα`(α) } and Cβ = {c1 , c2 , . . . , c`(β ) }.
After the grouping, the number of combined predictor functions is `(α) ∗ `(β ). We de(α)
(β ) (α)
(β )
(α)
note the set of combined predictor functions as F̄αβ = { f1 · f1 , f1 · f2 , . . . , f`(α) ·
(β )
(α)
f`(β ) }, where for i ∈ [1, `(α)] and j ∈ [1, `(β )], fi
(β )
· fj
is a combined predictor func-
(α)
(β )
tion that takes the input nodes of functions fi and f j as its input and
(α)
(β )
the Boolean output of functions fi and f j into integers as output. The
combines
combined
integers range in [0, 3] and their 2-bit binary representations (from right to left) repre(α)
(β )
sent the values of nodes xα and xβ . The selection probability for function fi · f j is
(α)
(β )
`(α)
`(β )
(α
(β )
ci ∗ c j . It holds that ∑i=1 Σ j=1 (ci ∗ c j ) = 1. With the selection probabilities, we
can compute the alias table for each group so that the selection of combined predictor
function in each group can be performed in constant time.
We now describe how to divide the nodes into groups. Our aim is to have as few
groups as possible so that the updating of all the nodes can be finished in as few rounds
as possible. However, fewer groups lead to many more nodes in a group, which will
result in a huge number of combined predictor functions in the group. Therefore, the
number of groups has to been chosen properly so that the number of groups is as small
as possible, while the combined predictor functions can be stored within the memory
limit of the computer performing the simulation. Besides, nodes with only one predictor function should be considered separately since selections of predictor functions for
those nodes are not needed. In the rest of this section, we first formulate the problem for
dividing nodes with more than one predictor function and give our solution afterwards;
then we discuss how to treat nodes with only one predictor function.
Problem description. Let S be a list of n items {µ1 , µ2 , . . . , µn }. For i ∈ [1, n], item µi
represents a node in a PBN with n nodes and its weight is assigned by a function ω(µi ),
which returns the number of predictor functions of node µi . We aim to find a minimum
integer m to distribute the nodes into m groups such that the sum of the combined
predictor functions numbers of the m groups will not exceed a memory limit θ . This is
equivalent to finding a minimum m and an m-partition S1 , S2 , . . . Sm of S, i.e., S =
S1 ∪
S2 ∪ · · · ∪ Sm and Sk ∩ S` = 0/ for k, ` ∈ {1, 2, . . . , m}, such that ∑m
ω(µ
)
≤ θ.
Π
j
µ
∈S
i=1
j
i
Solution. The problem in fact has two outputs: an integer m and an m-partition. We first
try to estimate a potential value of m, i.e., the lower bound of m that could lead to an
m-partition of S which satisfies ∑m
i=1 Πµ j ∈Si ω(µ j ) ≤ θ . With this estimate, we then
try to find an m-partition satisfying the above requirements.
8
Mizera, Pang and Yuan
Algorithm 3 The greedy algorithm
1: procedure F IND PARTITIONS(S, m)
2:
sort S with descending orders based on the weight of items in S;
3:
initialise A, an array of m lists; j = 0;
//initially, each A[i] is an empty list
4:
repeat
//S[ j] ( j starts from 0) is the jth item in S
5:
among the m elements of A,
6:
find the one with the smallest total weight and add S[ j] to it;
7:
j + +;
8:
until j = S.size()
//S.size() returns the number of items in S
9:
return A.
10: end procedure
Denote the weight of a sub-list Si as wi and wi = Πµ j ∈Si ω(µ j ). The inequality in the
problem description can be rewritten as ∑m
i=1 wi ≤ θ . We first compute the minimum
value of m̂ satisfying the following inequality and denote this minimum value as m̂min .
q
n ω(µ ) ≤ θ .
(2)
m̂ · m̂ Πi=1
i
Theorem 2. m̂min is the lower bound of m that allows a partition to satisfy ∑m
i=1 wi ≤ θ .
Proof. This is equivalent to proving that
m̂min −1
∑
0
wi > θ ,
(3)
i=1
0
where wi is the weight of the ith sub-list in an arbitrary partition of S into m̂min − 1
sub-lists. Since m̂p
min is the minimum value of m̂ that satisfies Inequality (2), we have
n ω(µ ) > θ . Hence,
(m̂min − 1) · (m̂min −1) Πi=1
i
(m̂min − 1) ·
q
(m̂min −1)
m̂min −1
Πi=1
wi > θ .
0
(4)
Based on the inequality of arithmetic and geometric means, we have
m̂min −1
∑
0
wi ≥ (m̂min − 1) ·
q
(m̂min −1)
m̂min −1
Πi=1
wi .
0
(5)
i=1
Inequality (3) follows from Inequality (4) and Inequality (5).
t
u
Starting from the lower bound, we try to find a partition of S into m sub-lists that
satisfies ∑m
i=1 wi ≤ θ . Since the arithmetic and geometric means of non-negative real
numbers are equal if and only if every number is the same, we get the heuristic that the
weight of the m sub-lists should be as equal as possible so that the sum of the weights
is as small as possible. Our problem then becomes similar to the NP-hard multi-way
number partition problem: to divide a given set of integers into a collection of subsets,
Fast Simulation of Probabilistic Boolean Networks
9
Algorithm 4 Partition n nodes into groups.
1: procedure PARTITION(G, θ )
2:
compute two lists S and S0 based on G; //S0 contains nodes with one predictor function
3:
compute the lower bound m̂; m = m̂;
4:
repeat
5:
A1 = F IND PARTITIONS(S, m);
6:
compute the sum of the weight of the m partitions and assign it to sum;
7:
m = m + 1;
8:
until sum < θ
9:
divide S0 into A2 ;
//partition nodes with only one predictor function
10:
merge A1 and A2 into A;
11:
return A.
12: end procedure
so that the sum of the numbers in each subset are as nearly equal as possible. We adapt
the greedy algorithm (see Algorithm 3 for details) for solving the multi-way number
partition problem, by modifying the sum to multiplication, in order to solve our partition
problem.3 If the m-partition we find satisfies the requirement ∑m
i=1 wi ≤ θ , then we get
a solution to our problem. Otherwise, we need to increase m by one and try to find a
new m-partition. We repeat this process until the condition ∑m
i=1 wi ≤ θ is satisfied. The
whole partition process for all the nodes is described in Algorithm 4.
Nodes with only one predictor function are treated in line 9. We divide such nodes
into groups based on their parent nodes, i.e., we put nodes sharing the most common
parents into the same group. In this way, the combined predictor function size can be
as small as possible such that the limited memory can handle more nodes in a group.
The number of nodes in a group is also restricted by the combined predictor function
size, i.e., the number of parent nodes in this group.4 The partition is performed with an
algorithm similar to Algorithm 3. The difference is that in each iteration we always add
a node into a group which shares most common parent nodes with this node.
3.4
The new simulation method
We describe our new method for simulating PBNs in Algorithm 5. The procedure
P REPARATION describes the whole preparation process of the three steps (network reduction for Step 1, and node-grouping for Step 2 and Step 3). The four inputs of the
procedure P REPARATION are respectively the PBN network G, the memory limit θ , the
maximum number k of nodes that can be put in a group for perturbation and the maximum number of parent nodes in a group. The P REPARATION procedure takes these
inputs, performs network reduction and node grouping. The reduced network and the
3 There exist other algorithms to solve the multi-way number partition problem, and we
choose the greedy algorithm for its efficiency.
4 In our experiments, the maximum number of parent nodes in one group is set to 18. Similar
to the value of k in step 2, the number can be larger as long as the memory can handle. However,
the penalty from large cache miss rate will diminish the benefits by having fewer groups when
the number of parent nodes is too large.
10
Mizera, Pang and Yuan
Algorithm 5 Structure-based PBN simulation.
1:
2:
3:
4:
5:
6:
7:
8:
9:
10:
11:
12:
13:
14:
15:
16:
17:
18:
19:
20:
21:
22:
23:
24:
procedure P REPARATION(G, θ , k)
perform network reduction for G and store the reduced network in G0
get the number of nodes n and perturbation rate p from G;
get the number of nodes n0 from G0 ; t = pow(1 − p, n − n0 );
[A p , mask] =P REPARE P ERTURBATION(n0 , k);
PA =PARTITION(G0 , θ );
for each group in PA, compute its combined functions F, and its alias table A,
and the cumulative number of nodes in the group cum;
return [A p , mask,t, A, F, cum].
end procedure
procedure PARALLEL S IMULATION(A p , mask, A, F, cum,t, s)
[s, perturbed] =P ERTURBATION(A p , mask, s);
//perform perturbations by group
if perturbed || C HECK L EAF N ODES(t) then return s; //check perturbations of leaves
else i = 0; s0 = 0; count = size(A); //size(A) returns the number of elements in array A
repeat
index = Next(A[i]); //select a random integer based on the alias table of group i
f = F.get(index);
//obtain the predictor function at the given index
v = f [s];
// f [s] returns the integer output of f based on state s
s0 = s0 | (v << cum[i]);
//shift v to update only nodes in the current group
i + +;
until i = count − 1;
end if
return s0 .
end procedure
grouped nodes information are then provided for the PARALLEL S IMULATION procedure via six parameters: A p and mask are the alias table and mask used for performing
perturbations of non-leaf nodes as explained in Algorithm 2; t is used for checking perturbations in leaf nodes as explained in Algorithm 1; A is an array containing the alias
tables for predictor functions in all groups; F is an array containing predictor functions
of all groups; and cum is an array storing the cumulative number of nodes in each group.
Perturbations for leaf nodes and non-leaf nodes have been explained in Algorithms 1
and 2. We now explain how nodes in a group are simultaneously updated with combined
predictor function. It is performed via the following three steps: 1) a random combined
predictor function is selected from F based on the alias table A; 2) the output of the
combined predictor function is obtained according to the current state s; 3) the nodes in
this group are updated based on the output of the combined predictor function. To save
memory, states are stored as integers and updating a group of nodes is implemented via
a logical bitwise or operation. To guarantee that the update is performed on the required
nodes, a shift operation is needed on the output of the selected function (line 19). The
number of bits to be shifted for the current group is in fact the cumulative number of
nodes of all its previous groups, which is stored in the array cum.
Fast Simulation of Probabilistic Boolean Networks
4
11
Evaluation
The evaluation of our new simulation method is performed on both randomly generated
networks and a real-life biological network. All the experiments are performed on a high
performance computing (HPC) machines, each of which contains a CPU of Intel Xeon
X5675 @ 3.07 GHz. The program is written in Java and the initial and maximum Java
virtual machine heap size is set to 4GB and 5.89GB, respectively. The evaluation data
is available at http://satoss.uni.lu/software/ASSA-PBN/benchmark.
4.1
Randomly generated networks
With the evaluation on randomly generated networks, we aim not only to show the
efficiency of our methods, but also to answer how much speedup our method is likely
to provide for a given PBN.
The first step of our new simulation method performs a network reduction technique, which is different from the node-grouping technique in the later two steps. Therefore, we evaluate the contribution of the first step and the other two steps to the performance of our new simulation method separately. We name the original simulation
method as Methodold ; the simulation method applying the network reduction technique
as Methodreduction ; and the simulation method applying both the network reduction and
node-grouping techniques as Methodnew . Methodreduction and Methodnew require preprocessing of the PBN under study, which leads to a certain computational overhead.
However, the proportion of the pre-processing time in the whole computation decreases
with the increase of the sample size. In our evaluation, we first focus on comparisons
without taking pre-processing into account to evaluate the maximum potential performance of our new simulation method; we then show how different sample sizes will
affect the performance when pre-processing is considered.
How does our method perform? Intuitively, the speedup due to the network reduction
technique is influenced by how much a network can be reduced and the performance of
node-grouping is influenced by both the density and size of a given network. Hence, the
evaluation is performed on a large number of randomly generated PBNs covering different types of networks. In total, we use 2307 randomly generated PBNs with different
percentages of leaves ranging between 0% and 90%; different densities ranging between
1 and 8.1; and different network sizes from the set {20, 50, 100, 150, 200, 250, 300, 350,
400, 450, 500, 550, 600, 650, 700, 750, 800, 850, 900, 950, 1000}. We simulate 400 million steps for each of the 2307 PBNs with the three different simulation methods and
compare their time costs. For the network reduction technique the speedups are computed as the ratio between the time of Methodreduction and the time of Methodold . The
obtained speedups are between 1.00 and 10.90. For node-grouping, the speedups are
computed as the ratio between the time of Methodnew and the time of Methodreduction .
We have obtained speedups between 1.56 and 4.99. We plot in Figure 1 the speedups
of the network reduction and node-grouping techniques with respect to their related parameters. For the speedups achieved with network reduction, the related parameters are
the percentage of leaves and the density. In fact, there is little influence from density
to the speedup resulting from network reduction as the speedups do not change much
12
Mizera, Pang and Yuan
(a) Methodreduction over Methodold .
(b) Methodnew over Methodreduction .
Fig. 1: Speedups contributed from network reduction and node-grouping.
with the different densities (see Figure 1a). The determinant factor is the percentage of
leaves. The more leaves a PBN has, the more speedup we can obtain for the network. For
the speedups obtained from node-grouping, the related parameters are the density and
the network size after network reduction, i.e., the number of non-leave nodes. Based
on Figure 1b, the speedup with node-grouping is mainly determined by the network
density: a smaller network density could result in a larger speedup contributed from the
node-grouping technique. This is mainly due to the fact that sparse network has a relatively small number of predictor functions in each node and therefore, the nodes will be
partitioned into fewer groups. Moreover, while the performance of network reduction
is largely influenced by the percentage of leaves, the node-grouping technique tends
to provide a rather stable speedup. Even for large dense networks, the technique can
reduce the time cost almost by half.
The combination of these two techniques results in speedups (time of Methodnew
over time of Methodold ) between 1.74 and 41.92. We plot in Figure 2 the speedups in
terms of the percentage of leaves and density. The figure shows a very good performance
of our new method on sparse networks with large percentage of leaves.
What is the influence of sample size? We continue to evaluate the influence of sample
size on our proposed new PBN simulation method. The pre-processing time for the
network reduction step is relatively very small; therefore our evaluation focuses on the
influence to the total time cost of all the 3 steps, i.e., the speedup of Methodnew with
respect to Methodold . We selected 9 representative PBNs from the above 2307 PBNs,
with respect to their densities, percentages of leaves and the speedups we have obtained.
We simulate the 9 PBNs for different sample size using both Methodold and Methodnew .
We show the average pre-processing time of Methodnew and the obtained speedups of
Methodnew (taking into account pre-processing time costs) with different sample sizes
in Table 1. As expected, with the increase of the sample size, the influence of preprocessing time becomes smaller and the speedup increases. In fact, in some cases, the
pre-processing time is relatively so small that its influence becomes negligible, e.g., for
networks 7 and 8 when the sample size is equal or greater than 100 million. Moreover,
Fast Simulation of Probabilistic Boolean Networks
13
Fig. 2: Speedups of Methodnew over Methodold .
network #
1
2
3
4
5
6
7
8
9
size
900
950
1000
600
800
900
450
550
1000
percentage
pre-processing
density
of leaves
time (second)
1.11
0.84
0.30
67.83
68.38
68.00
89.78
88.55
89.10
6.72
6.96
7.00
4.25
3.94
3.89
1.60
1.72
1.75
28.12
32.35
33.72
162.21
43.17
36.58
0.23
0.24
1.08
speedup with different
sample sizes (million)
1
10
100
400
0.65
0.59
0.58
0.13
0.66
0.69
21.44
20.26
10.04
1.49
1.47
1.45
1.08
3.05
3.56
37.59
35.94
31.83
1.71
1.73
1.71
4.51
6.75
6.90
41.62
36.47
35.09
1.73
1.75
1.73
6.89
7.69
7.70
41.84
36.62
37.19
Table 1: Influence of sample sizes on the speedups of Methodnew over Methodold .
often with a sample size larger than 10 million, the effort spent in pre-processing can
be compensated by the saved sampling time (simulation speedup).
Performance prediction. To predict the speedup of our method for a given network, we
apply regression techniques on the results of the 2307 PBNs to fit a prediction model.
We use the normalised percentage of leaves and the network density as the predictor
variables and the speedup of Methodnew over Methodold as the response variables in
the regression model since network size does not directly affect the speedup based on
the plotted pictures. In the end, we obtained a polynomial regression model shown in
Equation 6, which can fit 90.9% of the data:
y = b1 + b2 ∗ x1 + b3 ∗ x12 + b4 ∗ x2 + b5 ∗ x22 ,
(6)
where [b1 , b2 , b3 , b4 , b5 ] = [2.89, 2.71, 2.40, −1.65, 0.71], y represents the speedup, x1
represents the percentage of leaves and x2 represent the network density. The result of
a 10-fold cross-validation of this model supports this prediction rate; hence we believe
14
Mizera, Pang and Yuan
this model does not overfit the given data. Based on this model, we can predict how
much speedup is likely to be obtained with our proposed method for a given PBN.
4.2
An apoptosis network
In this section, we evaluate our method on a real-life biological network, i.e., an apoptosis network containing 96 nodes [12]. This network contains 37.5% of leaves and
has a density of 1.78, which is suitable for applying our method to gain speedups.
The apoptosis network has been analysed in [9]. In one of the analyses, i.e., the longterm influences [13] on complex2 from each of its parent nodes: RIP-deubi, complex1, and FADD, 7 steady-state probabilities of the apoptosis network need to be computed. In this evaluation, we compute the 7 steady-state probabilities using our proposed
structure-based simulation method (Methodnew ) and compare it with the original simulation method (Methodold ). The precision and confidence level of all the computations,
as required by the two-state Markov chain approach [14], are set to 10−5 and 0.95,
respectively. The results of this computation are shown in Table 2. The sample sizes
required by both methods are very close for computing same steady-state probabilities.
Note that the speedups are computed based on the accurate data, which are slightly
different from the truncated and rounded data shown in the table. We have obtained
speedups (Methodnew over Methodold ) between 7.67 and 10.28 for computing those 7
probabilities. In total, the time cost is reduced from 1.5 hours to about 10 mins.
5
Discussion and Conclusion
In this work, we propose a structure-based method for speeding up simulations of PBNs.
Using network reduction and node-grouping techniques, our method can significantly
improve the simulation speed of PBNs. We show with experiments that our new simulation method is especially efficient in the case of analysing sparse networks with a large
number of leaf nodes.
The node-grouping technique gains speedups by using more memory. Theoretically,
as long as the memory can handle, the group number can be made as small as possible.
However, this causes two issues in practice. First, the pre-processing time increases
dramatically with the group number decreasing. Second, the performance of the method
drops a lot when operating on large memories due to the increase of cache miss rate.
Therefore, in our experiments we do not explore all the available memory to maximise
the groups. Reducing the pre-processing time cost and the cache miss rate would be two
future works to further improve the performance of our method. We also plan to apply
our method for the analysis of real-life large biological networks.
References
1. Shmulevich, I., Dougherty, E.R.: Probabilistic Boolean Networks: The Modeling and Control of Gene Regulatory Networks. SIAM Press (2010)
2. Trairatphisan, P., Mizera, A., Pang, J., Tantar, A.A., Schneider, J., Sauter, T.: Recent development and biomedical applications of probabilistic Boolean networks. Cell Communication
and Signaling 11 (2013) 46
Fast Simulation of Probabilistic Boolean Networks
15
Methodold
Methodnew
probability # sample size
pre-processing sample size
total
speedup
time (minute)
(million)
time (second) (million) time (minute)
1
2
3
4
5
6
7
147.50
452.35
253.85
49.52
315.06
62.22
255.88
9.51
28.65
14.88
2.96
17.73
3.69
16.74
4.57
3.10
3.42
3.38
4.40
3.13
4.01
147.82
452.25
253.99
50.39
305.43
50.28
256.61
1.05
2.79
1.74
0.36
2.05
0.39
1.70
9.09
10.28
8.54
8.31
8.39
7.67
9.88
Table 2: Performance of Methodold and Methodnew on an apoptosis network.
3. Kauffman, S.A.: Homeostasis and differentiation in random genetic control networks. Nature
224 (1969) 177–178
4. Shmulevich, I., Dougherty, E.R., Zhang, W.: From boolean to probabilistic boolean networks
as models of genetic regulatory networks. Proceedings of the IEEE 90(11) (2002) 1778–1792
5. Shmulevich, I., Dougherty, E.R., Zhang, W.: Gene perturbation and intervention in probabilistic Boolean networks. Bioinformatics 18(10) (2002) 1319–1331
6. Shmulevich, I., Dougherty, E.R., Zhang, W.: Control of stationary behavior in probabilistic
Boolean networks by means of structural intervention. Journal of Biological Systems 10(04)
(2002) 431–445
7. Shmulevich, I., Gluhovsky, I., Hashimoto, R., Dougherty, E., Zhang, W.: Steady-state analysis of genetic regulatory networks modelled by probabilistic Boolean networks. Comparative
and Functional Genomics 4(6) (2003) 601–608
8. Trairatphisan, P., Mizera, A., Pang, J., Tantar, A.A., Sauter, T.: optPBN: An optimisation
toolbox for probabilistic boolean networks. PLOS ONE 9(7) (2014)
9. Mizera, A., Pang, J., Yuan, Q.: Reviving the two-state markov chain approach (technical
report). Available online at http://arxiv.org/abs/1501.01779 (2015)
10. Mizera, A., Pang, J., Yuan, Q.: ASSA-PBN: a tool for approximate steady-state analysis of
large probabilistic Boolean networks. In: Proc. 13th International Symposium on Automated
Technology for Verification and Analysis. LNCS, Springer (2015) Available at http://
satoss.uni.lu/software/ASSA-PBN/.
11. Walker, A.: An efficient method for generating discrete random variables with general distributions. ACM Transactions on Mathematical Software 3(3) (1977) 253–256
12. Schlatter, R., Schmich, K., Vizcarra, I.A., Scheurich, P., Sauter, T., Borner, C., Ederer, M.,
Merfort, I., Sawodny, O.: ON/OFF and beyond - a Boolean model of apoptosis. PLOS
Computational Biology 5(12) (2009) e1000595
13. Shmulevich, I., Dougherty, E.R., Kim, S., Zhang, W.: Probabilistic Boolean networks: a
rule-based uncertainty model for gene regulatory networks. Bioinformatics 18(2) (2002)
261–274
14. Raftery, A.E., Lewis, S.: How many iterations in the Gibbs sampler? Bayesian Statistics 4
(1992) 763–773
| 5cs.CE
|
Computer Algebra and Material Design
Akihito Kikuchi
arXiv:1612.02275v4 [cond-mat.mtrl-sci] 17 Feb 2018
Canon Inc, Tokyo, Japan∗
1
Abstract
This article is intended to an introductory lecture in material physics, in which the
modern computational group theory and the electronic structure calculation are in collaboration. The effort of mathematicians in the field of group theory have ripened as a new
trend, called “computer algebra”, outcomes of which now can be available as handy computational packages, and would also be useful to physicists with practical purposes. This
article, in the former part, explains how to use the computer algebra for the applications in
the solid-state simulation, by means of one of the computer algebra package, the GAP system. The computer algebra enables us to obtain various group theoretical properties with
ease, such as the representations, the character tables, the subgroups, etc. Furthermore,
it would grant us a new perspective of material design, which could be executed in mathematically rigorous and systematic way. Some technical details and some computations
which require the knowledge of a little higher mathematics (but computable easily by the
computer algebra) are also given. The selected topics will provide the reader with some
insights toward the dominating role of the symmetry in crystal, or, the “mathematical
first principles” in it. In the latter part of the article, we analyze the relation between the
structural symmetry and the electronic structure in C60 (as an example to the system without periodicity). The principal object of the study is to illustrate the hierarchical change
of the quantum-physical properties of the molecule, in accordance with the reduction of
the symmetry (as it descends down in the ladder of subgroups). As an application, this
article also presents the computation of the vibrational modes of the C60 by means of the
computer algebra. In order to serve the common interest of the researchers, the details
of the computations (the required initial data and the small programs developed for the
purpose) are explained as minutely as possible.
∗
akihito [email protected] (The corresponding author)
2
I.
INTRODUCTION
In the history of physics, the cooperation of physicists and mathematicians has
yielded a great harvest in the first half of the twentieth century, as the installation
of group theory in quantum mechanics[1]. One of the main applications is realized
in the field of solid-state physics[2–4]. In subsequent years, however, such productive relationship between physics and mathematics became enfeebled, as physicists
and mathematicians were pursuing their own interests separately. In material science, the typical tool of study has turned into “first principle electronic structure
computation”, in which rapid computers are intensively used so that the quantitative simulation could be achieved. In contrast, the group theoretical view in the
quantum physics is rather a qualitative one, which could explain the likeness in similar material structures but could not illuminate the origin of subtle but distinct
differences. The standpoints of group theoretical analysis, and of first principles
simulation, are located at cross-purposes. This is one of the reasons which brought
about the breaking-off between the group theory and the first principles electronic
structure computation. There is another hardship which hinders the collaboration
of mathematics and physics. The computation of the “representation” in group
theory[5–9], useful in physical applications, requires special arts, which is recondite
to non-experts. Thus, for the purpose of general use, the character tables are listed
up in literature and textbooks (recently, moreover in databases). Traditionally we
are obliged to consult with the non-electric data, which is prone to troublesome errors. Such a circumstance has little affinity to the modern style of computational
physics, where necessary data should easily be accessible on the computer or computed anew by the researcher. Nevertheless, it must be stressed here that such a
cumbersome situation has already become surmountable. The development of the
effective computer architectures and the program packages for computational group
3
theory in recent decades have benefited mathematicians so well that they could actually obtain definite symbolic or numeric solutions to the problems of their own, not
only in proving the existence of the solution[10, 11]. These computational tools are
not restricted to mathematicians, but available to every scientific researcher. Powerful outcomes by mathematicians, also, will confer a modernized viewpoint to the
material science, so that the conventional computational tools could be assisted and
improved, although, in the present state, most of the material scientists seem to be
content in having old-fashioned knowledge of mathematics. This article takes several
examples in the group theoretical analysis in the electronic structure calculation and
presents the details of the computation so that the readers would know how to use
modern mathematical packages of computational group theory. The necessary group
theoretical data are computed by the desktop computer, without references to other
resources, and are applicable to the analysis of the quantum physics in materials
which is principally governed by the symmetry of the system. The article furthermore expounds the possible style of the systematic material design, by which the
electronic structure might be controlled artificially from the viewpoint of the structural symmetry. From these topics, the potentiality of the cooperation of computer
algebra and first-principles electronic structure (which are legitimate heirs of group
theory and quantum mechanics) may be observed.
To be fair, there are a lot of works of physicists who make use of computer algebra,
although the works of pedagogical purpose are not so many. My article is only one
of them. And here I list up some of such works for the benefit of readers.
The vibrational spectrum of C60 is analyzed in the thesis of Mooij[13]. The representation of the group is computed by the computer algebra system Maple; the
vibrational modes are computed by the reduction of the size of the dynamical matrix (of the empirical potential model); the parameters of these models are fit against
the experimental data for the optically active modes. My article also treats the vibra4
tional modes of this molecule, but it does not arrive at the quantitative computation
of vibrational frequencies. My article, by means of GAP, classifies geometrically possible deformations with respect to irreducible representations and generates the basis
set of deformations in the symbolic formulas, although the obtained results could be
applied to more quantitative studies.
The book by El-Batanouny and Wooten treats the wide range of topics related
to the symmetry in the condensed matter[14]. The interest of the research is almost
the same as my article. One of the principal difference to my article is this: the
authors of this book developed the programs for the group theoretical study by
means of Mathematica, while I adopt the GAP system. The authors of this book
developed Mathematica programs which compute the characters of groups by means
of Dixon’s method and applied them toward various topics. (I should say this:
the GAP package computes the characters of groups by the same method, so we can
directly step into applications in physics, as end-users of mathematics.) I recommend
the readers to read the book along with my article, if possible, in order to solidify
their understanding.
The book by Stauffer et al treats the general topics on the computer simulation
and the computer algebra[15]. The adopted computer algebra system is ”reduce”, for
the purpose of symbolic computations, and the application to the physics is rather
introductory.
Hergert and coworkers made the group-theoretical analysis of electronic and photonic band structure[16, 17]. The symmetry properties of Schrödinger equation and
Maxwell equation are investigated. The computer algebra system Mathematica enables them to simplify group theoretical study.
As for the introduction to the computational physics, the book by Thijssen will
be useful[18].
In the study of quantum chemistry, Sakiyama and Waki utilized the computer
5
algebra system GAP[19]. The possible conformations (for a hexakis-methylamine
nickel(ii) complex cation) are generated by the programs built on GAP, and the
stability of possible formations is computed and compared through the quantitative
electronic structure computation.
In the thesis of Rykhlinskaya, the author developed a computer algebra package
”Bethe” in the framework of Maple, and applied it to the symmetric molecular
geometry and symmetries, the vibrational analysis of the molecule, and the analysis
of the atomic behavior and the atoms in the crystal field[20].The applications of
Bethe are demonstrated in the article of Fritzshche[21].
Barnett gave us the review article in the application of computer algebra, not only
in the life science but in the material science[22]. The interest of the author is centered
on the symbolic computation (especially in chemistry), not of the computer group
theory. However, we can find useful references there, if we extend our researches
beyond group theory.
“Bilbao Crystallographic Server” is an online database of crystallography[23–25].
From this server, we can find data of crystallographic point groups and space groups
and useful programs.
Also, there are studies of the group theoretical property of C60 by mathematicians[26,
27]. The authors of these articles (Chung, Kostant, and Sternberg) discuss the embedding of the icosahedral group (of C60 symmetry) into a larger group, which is
a realization of the statement of Galois. And they analyze the energy spectrum of
the molecule by the representation theory, with the hope that their results will find
applications in the physical properties of the molecule.
Eick and Souvignier give us a survey of the algorithms for space groups and
crystallographic groups available in the computer algebra system GAP and in the
software packages Carat and Cryst[28].
It seems to me, that the application of the GAP system to the group theoretical
6
analysis in the material science is still undeveloped, although a great deal of the
products in the pure mathematics can be attainable easily from GAP. Some of these
related works explain to us how to use the computational packages developed on
some computer algebra system, and provide us with those packages, but they are
not kind enough to explain their implementation. So in my article, I intend to
present ”the raw programs” and ”the raw computed results” so that the readers may
have experience on the actual computations by GAP. Again, I must say this: the
examples taken in the following chapters are not novel, and a lot of group theoretical
analysis is already done, without the aid of computer algebra. I only show an easier
way of these analyses by means of a desktop PC. If readers thought that some of the
methods are unfamiliar or novel, they are not of my invention, in the strict sense:
they would be well-known things in the field of pure mathematics, and they are only
interpretations in the viewpoint of physics.
II.
COMPUTATION OF GROUP THEORETICAL PROPERTIES USING
“GAP”
The computational discrete algebra package GAP is one of the powerful tools for
the computation in the group theory[12]. It can be applied to the determination of the
group theoretical properties which are necessary for solid-state electronic structure
computation. For this purpose, the point group of the crystal must be determined.
The group operations are to be listed up; the multiplication table is to be prepared;
through which the character table is computed.
At first, the concept of the character of the group is explained here. Let us consider
an equilateral triangle as in Fig. 1. There are operations in the plane, such as, the
rotation by 60 degrees at the center of mass, and the reflection at an axis which con7
nects the center of mass and one of the vertexes. These operations exchange vertexes
among themselves but fix the triangle unmoved. In the mathematical terminology,
these operations consist a group C3v . These operations are the permutation of the
vertexes, numbered as 1,2, and 3 in the clockwise direction.
FIG. 1. A triangle, as the object of the symmetric operation. The symmetric operation
can be through the permutations of three symbols, 1,2,3, or the rotations and reflections
in the x-y plane.
They are expressed as
E=
B=
D=
1 2 3
1 2 3
1 2 3
1 3 2
1 2 3
3 1 2
,A =
,C =
,F =
1 2 3
2 1 3
1 2 3
3 2 1
1 2 3
2 3 1
,
,
.
In the above the top and the bottom low denote the initial and the final arrangements
8
of the vertexes respectively.
These operations can also be expressed by six rotation matrices in the x-y plane:
E=
B=
D=
1 0
0 1
,A =
√
3
2
− 12
√
3
2
1 0
,
0 −1
√
3
1
− − 23
2
,
, C = √2
3
1
1
−
2
2
2
√
√
3
1
− − 23
2
.
, F = √2
3
1
− 21
−
2
2
√
− 12
−
The multiplication table of these elements is written in table 1.
E
A
B
C
D
F
E
E
A
B
C
D
F
A
A
E
F
D
C
B
B
B
D
E
F
A
C
C
C
F
D
E
B
A
D
D
B
C
A
F
E
F
F
C
A
B
E
D
TABLE 1. The multiplication table. The multiplications xy between the elements in the
left column x and in the top row y are shown.
Let us examine these maps:
Γ1 : E, A, B, C, D, F → 1
Γ10 : E, D, F → 1; A, B, C → −1 (Determinant of matrices)
Γ2 : E, A, B, C, D, F → E, A, B, C, D, F (Identity map as matrices)
(1)
One can see that the multiplication table is kept unaltered by these maps with the
replacement of the six symbols to corresponding targets. The geometrical operations
9
in the triangle, forming a group, are represented by these maps, even if not always
faithfully; the viewpoint can be switched from geometrical one to numerical one.
This is an example of the representation of the group; all elements in the group are
represented by proxies of scalars or matrices; the multiplications among them are
subject to the same rule as the original group elements. The trace of the matrix
representation, such as Γ2 , is called “character”, having certain properties favorable
to the application of quantum physics, as will be seen later. The character table of
C3v is given in table 2.
E
Γ1 1
Γ10 1
Γ2 2
A
1
-1
0
B
1
-1
0
C
1
-1
0
D
1
1
-1
F
1
1
-1
TABLE 2. The character table of C3v .
As can be seen in the multiplication table, the group elements are divided into
subsets such as
Cl(a) = {g ∈ G| there exists x ∈ G such as g = xax−1 }.
(2)
These subsets are called conjugacy classes. The elements in the same conjugacy
class have a common value of the character; it is customary for character tables to
be labeled by conjugacy classes, not by each group elements. The conjugacy classes
for C3v are given by three subsets: {E}, {A, B, C} and {D, F}.
The gap computation for this example proceeds as follows. The group is defined
by the minimal set of generators including a rotation and a reflection:
gap> G:=Group((1,2,3),(1,2));
Group([ (1,2,3), (1,2) ])
10
The program returns the results in the shorthand notation preferred by mathematicians.
gap> Elements(G);
[ (), (2,3), (1,2), (1,2,3), (1,3,2), (1,3) ]
gap> M:=MultiplicationTable(G);
[ [ 1, 2, 3, 4, 5, 6 ], [ 2, 1, 4, 3, 6, 5 ], [ 3, 5, 1, 6, 2, 4 ],
[ 4, 6, 2, 5, 1, 3 ], [ 5, 3, 6, 1, 4, 2 ], [ 6, 4, 5, 2, 3, 1 ] ]
The multiplication table is given by a matrix, which is represented by a doubly
nested list; each of the inner lists should be read as one of the lows in the matrix.
The entry Mij shows the result of the multiplication gi · gj between elements gi and
gj .
Subgroups and generators are computed by:
gap> AllSubgroups(G);
[ Group(()), Group([ (2,3) ]), Group([ (1,2) ]), Group([ (1,3) ]),
Group([ (1,2,3) ]), Group([ (1,2,3), (2,3) ]) ]
gap> GeneratorsOfGroup(G);
[ (1,2,3), (1,2) ]
gap> G.1;
(1,2,3)
gap> G.2;
(1,2)
gap> S:=Subgroup(G,[(1,2)]);
Subgroup(G,[(1,2)]);
gap> Elements(S);
[ (), (1,2) ]
gap> A:=Group((1,2,3));
11
Group([ (1,2,3) ])
gap> IsSubgroup(G,A);
true
gap> IsSubgroup(A,G);
false
gap> A:=Group((1,2,3,4));
Group([ (1,2,3,4) ])
gap> IsSubgroup(G,A);
false
The group can also be defined from the multiplication table as
gap> G2:=GroupByMultiplicationTable(M);
<group of size 6 with 6 generators>
gap> Elements(G2);
[ m1, m2, m3, m4, m5, m6 ]
The elements in the group constructed from the multiplication tables are represented
by abstract symbols m1,m2,...,m6.
In order to construct the group from the rotation in the Euclidean space, the
generating set of the rotation matrices should be given.
gap> M1:=[[-1/2,ER(3)/2],[-ER(3)/2,-1/2]];
[ [ -1/2, -1/2*E(12)^7+1/2*E(12)^11 ],
[ 1/2*E(12)^7-1/2*E(12)^11, -1/2 ] ]
gap> M2:=[[1,0],[0,-1]];
[ [ 1, 0 ], [ 0, -1 ] ]
gap> G3:=Group(M1,M2);
Group(
12
[
[ [ -1/2, -1/2*E(12)^7+1/2*E(12)^11 ],
[ 1/2*E(12)^7-1/2*E(12)^11, -1/2 ] ], [ [ 1, 0 ], [ 0, -1 ] ]
])
Here ER(n) is
√
n and E(n) is the primitive n-th root of unity exp(2πi/n).
Let us inquire of the GAP package whether the three different definitions should
generate equivalent groups.
gap> G=G;
true
gap> G2=G;
false
gap> G3=G;
false
The GAP program decides that the three groups are not identical in the strict sense,
because they are composed from different resources, i.e. permutations, abstract
symbols, and matrices. However, we can construct isomorphisms among them. The
following GAP command returns the isomorphism on the generators of the groups.
gap> IsomorphismGroups(G,G);
[ (1,2,3), (1,2) ] -> [ (1,2,3), (1,2) ]
gap> IsomorphismGroups(G2,G);
[ m1, m2, m3, m4, m5, m6 ] -> [ (), (2,3), (1,2), (1,2,3), (1,3,2),
(1,3) ]
gap> IsomorphismGroups(G,G2);
[ (1,2,3), (1,2) ] -> [ m4, m3 ]
gap> G4:=Group((1,2,3,4))
13
gap> IsomorphismGroups(G,G4);
fail
The existence of the isomorphism between G and G3 can also be verified in this way.
The conjugacy classes are computed by this command:
gap> ConjugacyClasses(G);
[ ()^G, (2,3)^G, (1,2,3)^G ]
To access entries in the conjugacy classes, a special method is needed. The following cannot work well:
gap> Elements(ConjugacyClasses(G));
[ ()^G, (2,3)^G, (1,2,3)^G ]
Instead, using “List” command and anonymous functions in λ calculus, we can
access each entry in the list:
gap> List(Elements(G),x->x);
[ (), (2,3), (1,2), (1,2,3), (1,3,2), (1,3) ]
gap> List(Elements(G),x->x^-1);
[ (), (2,3), (1,2), (1,3,2), (1,2,3), (1,3) ]
gap> List(ConjugacyClasses(G),Elements);
[ [ () ], [ (2,3), (1,2), (1,3) ], [ (1,2,3), (1,3,2) ] ]
gap> List(ConjugacyClasses(G),Representative);
[ (), (2,3), (1,2,3) ]
The character table is computed by this command:
gap> irrg:=Irr(G);
[ Character( CharacterTable( Sym( [ 1 .. 3 ] ) ), [ 1, 1, 1 ] ),
Character( CharacterTable( Sym( [ 1 .. 3 ] ) ), [ 1, -1, 1 ] ),
Character( CharacterTable( Sym( [ 1 .. 3 ] ) ), [ 2, 0, -1 ] ) ]
14
The character table can be accessible as a numeral list, or as a map on the group
elements:
gap> List([1,2,3],y->List([1,2,3],x->irrg[y][x]));
[ [ 1, 1, 1 ], [ 1, -1, 1 ], [ 2, 0, -1 ] ]
gap> List([1,2,3],y->List(Elements(G),x->x^irrg[y]));
[ [ 1, 1, 1, 1, 1, 1 ], [ 1, -1, -1, 1, 1, -1 ],
[ 2, 0, 0, -1, -1, 0 ] ]
The irreducible representation is computed by
gap> rep:=IrreducibleRepresentations(G);
[ Pcgs([ (2,3), (1,2,3) ]) -> [ [ [ 1 ] ], [ [ 1 ] ] ],
Pcgs([ (2,3), (1,2,3) ]) -> [ [ [ -1 ] ], [ [ 1 ] ] ],
Pcgs([ (2,3), (1,2,3) ]) ->
[ [ [ 0, 1 ], [ 1, 0 ] ], [ [ E(3), 0 ], [ 0, E(3)^2 ] ] ] ]
gap> List(Elements(G),x->x^rep[3]);
[ [ [ 1, 0 ], [ 0, 1 ] ], [ [ 0, 1 ], [ 1, 0 ] ],
[ [ 0, E(3)^2 ], [ E(3), 0 ] ], [ [ E(3), 0 ], [ 0, E(3)^2 ] ],
[ [ E(3)^2, 0 ], [ 0, E(3) ] ], [ [ 0, E(3) ], [ E(3)^2, 0 ] ] ]
This command returns the irreducible representations, giving the relations between
the generators of the group and the matrix representation. The characters are computed by taking traces of the matrix representation.
The character table is also computed by
gap> tbl:=CharacterTable(G);
CharacterTable( Sym( [ 1 .. 3 ] ) )
gap> Display(tbl);
CT1
15
2
1
1
.
3
1
.
1
1a 2a 3a
2P 1a 1a 3a
3P 1a 2a 1a
X.1
1
1
1
X.2
1 -1
1
X.3
2
. -1
To make use of the character table computed in this way, various subsidiary commands are prepared in GAP system. (N.B. The orderings of representations or
characters by these commands do not always coincide with each other.)
gap> ConjugacyClasses(tbl);
[ ()^G, (2,3)^G, (1,2,3)^G ]
gap> Irr(tbl);
[ Character( CharacterTable( Sym( [ 1 .. 3 ] ) ), [ 1, -1, 1 ] ),
Character( CharacterTable( Sym( [ 1 .. 3 ] ) ), [ 2, 0, -1 ] ),
Character( CharacterTable( Sym( [ 1 .. 3 ] ) ), [ 1, 1, 1 ] ) ]
Remark. There are various algorithms for the computation of characters. The simplest one, the Burnside algorithm, starts from the counting in the elements in the
conjugacy classes, sets up a matrix, and the characters are computed through the
eigenvalue problem. The Dixon (or Dixon-Schneider) algorithm goes in a more advanced way, employing the computation in the prime fields and the other techniques
16
of the symbolic computation, but it is effective in the large groups[29–31]. The latter
algorithm is adopted in GAP.
III.
SOME PRELIMINARIES
A.
Projection operator
Once the characters of the group are obtained, “projection operators” are constructed and assigned to each of the irreducible representation. The definition is as
follows:
P (p) =
lp X (p)∗
χ
· OT ,
|G| T ∈G T
(3)
where lp is the dimension of the irreducible representation,|G| is the order of the
(p)
group, and χT is the character of the group element T , which is allowed to be
complex-valued. OT is the operator allotted to each T in the space in which the
group action is defined. This operator acts on functions in the Euclidean space as
follows:
OT f (r) = f (T −1 (r)).
(4)
By this construction, we can distinguish whether a function belongs to the corresponding irreducible representation or not. If the projector is applied to the basis
functions of the corresponding irreducible representations, these functions are invariant as a set: on the other hand, the projector is applied to the basis functions of
different irreducible representations, the result of the projection goes to zero. (The
epithet “invariant” should be understood in the following way: in many cases, plural
basis functions belong to one irreducible representation. Through the group operations, these basis functions are interchanged with each other or re-expressed as the
linear combination of them; moreover, they are never transformed into functions be17
longing to other representations. In a word, they are invariant as a subspace; if the
subspace corresponding to a representation is minimal, no more being divisible, it is
called irreducible. The whole of representations and characters of the arbitrary finite
group could be computed exactly, without omission, by the principle of the group
theory.)
(p)
If we know the explicit forms of the irreducible matrix representation Dij (R),
the projection operator of the following form is useful:
(p)
Pkl =
lp X (p)∗
D (OT ) · OT .
|G| T ∈G kl
(5)
Owing to the definition of the partner functions
OT |p; li =
X
(p)
|p; jiDjl (T )
(6)
j
and the orthogonality theorem of the representation
X
(p)∗
(q)
Dik (T )Dlm (T ) =
T
|G|
δpq δil δkm
lp
(7)
this projector transforms one basis vector into another basis in one irreducible representation:
(p)
Pkl |p; li = |p; ki.
(8)
Moreover it can project out k-th partner function of the irreducible representation
P P (q)
from an arbitrary function F = q j fj |q; ji as
(p)
(p)
Pkk F = fk |p; ki.
(9)
At this junction, the connection between the group theory and the energy spec18
trum of the Schrodinger equations arises. Let the potential term be invariant under
some symmetry operations. With this potential term, certain eigenfunctions should
exist. When the symmetry operations are applied to the Schrodinger equation, the
Laplacian, the potential, and the energy spectrum are invariant: the changeable is
only the wavefunction, which is expressed by the basis functions in the irreducible
representation. If the irreducible representation is one-dimensional, the wavefunction will be invariant up to a certain phase factor by the symmetry operation: on
the other-hand, if the representation is multi-dimensional, the symmetry operation
may generate different wave-functions which are other solutions to the same equation with the same eigenvalue. Such circumstances lead to the degeneracy of the
energy spectrum. The identification of each wavefunction to its proper irreducible
representation enables us to clarify the relation between the energy spectrum and the
symmetry in the system, which is helpful in the analysis of the electronic properties
of materials.
B.
Space group
An n-dimensional space group S is a discrete subgroup in a group of Euclidean
motions in Rn , such that the subgroup T in S, composed of pure translations(without
rotations or reflections) is a free Abelian subgroup of rank n, having finite index in
S. Thus an exact sequence of groups exists as
0 → T → S → P → 1,
(10)
in which S acts on T by means of the conjugation action t → s · t · s−1 , which
represents the action of the point group P , defined as a factor group S/T , on the
lattice translations.
19
In a more intuitive expression, the space group S is the set of symmetry operations
in a periodic system; it is composed of rotations and reflections which fix components
in the unit cell, combined with parallel movements along crystal axes; therefore S
is of infinite order; the parallel movements, by themselves, compose the discrete
subgroup T of infinite order, in which the direction and the stride of the movements
are determined by the primitive lattice vectors. Meanwhile, the point group P is the
set of the reminders of operations in S, taken modulo of the crystalline periodicity, i.e.
with the parallel movements in T being nullified; it is a finite group and may include
fractional translations, which are represented by linear combinations of fractions of
primitive lattice vectors.
C.
Crystallographic group
In the case of crystallographic group, the symmetry operation is the affine mapping, composed of the linear parts RT , including the reflection, rotation, and inversion, and the translation parts τT [28]. The definition is
OT = {RT |τT } : r → RT · r + τT ,
(11)
OT · OS = {RT · RS |τT + RT · τs },
(12)
OT−1 = {RT−1 | − RT−1 · τT }.
(13)
The operation on the plane wave is
OT exp(ikr) = exp(ik(RT−1 r − RT−1 τ )) = exp(i(RT · k)(r − τ )).
(14)
The affine mapping gives rise to the phase factor exp(−i(RT · k)τ )), which is not
20
unity at a general k-point. On the other hand, the matrix representation of the affine
mapping is subject to the following relation:
D(p,k) ({Rk |fRk + τn }) = e−ik·τn D(p,k) ({Rk |fR })
(15)
The trace of the matrix D(p,k) is the character: therefore the phase-factors in the character and the wave-number part in the Bloch-type wavefunction (eikr in eikr uk (r) )
are canceled with each other as complex conjugates; in the general k-point, the phase
shift in the character can actually be negligible in the projection operator.(Indeed
there are textbooks or lectures omitting these terms.) However, if the Bloch-type
P
wavefunction is expressed by plane-wave expansion
CG ei(k+G)r , one should be
cautious. The phase shift in the periodic part is eiRk (G)τn , which remains without
cancellation in the projection operation.
D.
Theoretical set-up for Wyckoff positions
The mathematical definition of the Wyckoff positions is stated as follows.
Let G be a space group, T be its translation lattice, K be its point group, tg be
fractional translations. Each element in G has a form of {k|tk + t} for k ∈ K and
t ∈ T , and acts on a vector v in V = Rn as
{k|tk + y}(v) = k · v + tk + t.
(16)
The stabilizer of v ∈ V under this action is denoted as StabG (v)={g ∈ G|g(v) =
v}. Let v ∈ V . An equivalence relations ∼ is set up if StabG (v) is conjugate
to StabG (w),i.e. StabG (w)=StabG (g(v))=g StabG (w)g −1 for some g ∈ G. This
equivalence classes of ∼ is called the Wyckoff positions of G.
21
This definition, given in the mathematicians’ terminology, is a little recondite for
physicists. For a more intuitive understanding, it can be restated as:
The Wyckoff positions are a set of coordinate points, composed from two
subsets:
[Type 1] The coordinate points which may be fixed by a certain symmetry
operation of the point group. They are regarded as generators of the
Wyckoff positions.
[Type 2] The coordinate points generated from those in Type 1 subset
by all of the symmetry operations.
The command in GAP “WyckoffPositions(S)” returns the type 1 set for a space
group S; the equivalent coordinate points to a generator “W” in type 1 set by the
symmetry operations are computed from an another command “WyckoffOrbit(W)”.
E.
Units in the computation
Throughout the computations of quantum physics in this article, we mainly use
atomic units, abbreviated as a.u. if necessary; for the length, the Bohr radius(a0 u
5.292 × 10−11 m); for the energy, electron volt unit (eV) or the Hartree unit(Eh u
27.211eV). However, some computations adopt arbitrary unit in the energy.
IV.
APPLICATION 1: IDENTIFICATION OF WAVEFUNCTIONS TO IR-
REDUCIBLE REPRESENTATIONS
The application in this section is the classical example of group theory in quantum
physics.
22
A.
The simplest case: at Γ point
In this section, the group-theoretical analysis of the wavefunction is exemplified.
The wavefunctions at the Γ point in the diamond crystal are classified to corresponding irreducible representations. The treatment for the general k-point (k 6= 0) shall
be discussed later.(The knowledge of the distinction between symmorphic or nonsymmorphic crystal is necessary; the existence of these two types of crystal makes
the discussion not a little complicated.)
At first, the character table should be computed. The symmetric operations in the
diamond structure, whose unit cell is the minimal one, including two carbon atoms,
are given in the appendix, as well as the multiplication table of these operations.
When one uses the GAP packages, there are several options for the preparation
of the point group; of which three types can be used.
First: the symmetry operations are given as a set of the three-dimensional matrix.
gap> MT[
1]:=[[
1,
0,
0],[
0,
1,
0],[
0,
0,
1]];;
gap> MT[
2]:=[[
1,
0,
0],[
0, -1,
0],[
0,
0, -1]];;
0,
0],[ -1,
..........................
gap> MT[ 48]:=[[
0,
0,
1],[
1,
0,
0]];;
gap> G:=Group(MT);
Second: the multiplication table in the group is supplied.
gap> M:=[[1,2,3,...,48],[...],...,[...]];;
gap> G:=GroupByMultiplicationTable(M);
gap> Elements(G);
[ m1, m2, m3, m4,............,m45, m46, m47, m48 ]
In this case, the group elements are denoted by the symbolical way, m1,m2,...,m48.
23
Actually, all of the group elements are not necessarily provided to define a group:
it is enough to give the smallest generating set, from which other elements are constructed. The group should be remade by means of the smallest generating set. (In
the present implementation of GAP, especially when the multiplication table is supplied, the following tendency in the computation is observed: the computations for
groups, constructed from the minimal generating set, are much quicker than those for
groups where the all elements are stored as generators.) For the case of the diamond,
the computation goes as:
GS:=SmallGeneratingSet(G);
[ m36, m48 ]
G:=Group(GS);
<group with 2 generators>
The command “SmallGeneratingSet()” yields a reasonably small generating set.
In the cases of finite solvable groups (a typical example of this is the crystal point
group) and of finitely generated nilpotent groups, “MinimalGeneratingSet()” command is also available to get minimal generators, but the computation is timeconsuming. In addition, each element “elm” in the group “G“ can be expressed
by generators, by means of Factorization(G,elm):
gap> List(Elements(G),x->Factorization(G,x));
[ <identity ...>, x2^-1*x1*x2*x1, x2^2, (x2*x1)^2, x2^2*x1^2,
x1*x2^2*x1, x1^2, x2^-1*x1^-1*x2^-1*x1, x1^-1*x2^-1*x1*x2,
x2^-1*x1^2*x2, x2^-1*x1*x2*x1^-1, x1^-2, x2^-1*x1, x1^-1*x2^-1,
x2*x1, x2*x1*x2^2, x2^2*x1*x2, x2*x1^-1, x1*x2, x2^-1*x1^-1,
x1^-1*x2^-1*x1^2, x2^-1*x1^3, x2*x1*x2^2*x1^2, x2*x1^3, x1^3,
x2^-1*x1^-1*x2^-1*x1^2, x2^2*x1^3, x1*x2^2*x1^2, x2^2*x1^-1,
24
x2*x1*x2, x1^-1, x2^-1*x1*x2, x1*x2^2, x2^2*x1, x2^-1*x1^-1*x2^-1,
x1, x1^2*x2, x1*x2*x1^-1, x2*x1^-2, x1*x2^2*x1*x2, x2*x1*x2^2*x1,
x2*x1^2, x1^-1*x2^-1*x1, x2^-1*x1^2, x1*x2*x1, x2^-1,
x2*(x2*x1)^2, x2 ]
Third: the crystallographic groups are defined by four-dimensional augmented
matrices, in which both of the point group operations and the translations are inscribed. The notations for the augmented can be given in the following form
A v
0 1
(17)
acting on column vectors (x, 1) from the right or, alternatively,
T
A
0
v 1
(18)
acting on low vectors (x, 1) from the left, which represent the affine mapping x →
A · x + v.(Crystallographers prefer the later notation.)
Once the crystallographic group is defined, the point group can easily be deduced.
As for the diamond case, in the GAP computation, the crystallographic group is
defined as follows. (The minimal generating set is used for simplicity.)
gap> M1:=[[0,0,1,0],[1,0,0,0],[0,-1,0,0],[1/4,1/4,1/4,1]];;
gap> M2:=[[0,0,-1,0],[0,-1,0,0],[1,0,0,0],[0,0,0,1]];;
gap> S:=AffineCrystGroup([M1,M2]);
<matrix group with 2 generators>
gap> P:=PointGroup(S);
Group([ [ [ 0, 0, 1 ], [ 1, 0, 0 ], [ 0, -1, 0 ] ],
25
[ [ 0, 0, -1 ], [ 0, -1, 0 ], [ 1, 0, 0 ] ] ])
By these preparations the conjugacy classes and the character table of the diamond crystal are computed as tables 3 and 4.(The group elements for the symmetry
operation are given in a table in the appendix. The entries in the conjugacy classes
are given by the numbering of the table of the group elements.)
Class
C.1
C.2
C.3
C.4
C.5
C.6
C.7
C.8
C.9
C.10
Group Elements
1 (The identity)
2,3,4
5 , 6 , 7 , 8 , 9 , 10 , 11 , 12
13 , 16 , 17 , 18 , 21 , 23
14 , 15 , 19 , 20 , 22 , 24
25
26 , 27 , 28
29 , 30 , 31 , 32 , 33 , 34 , 35 , 36
37 , 40 , 41 , 42 , 45 , 47
38 , 39 , 43 , 44 , 46 , 48
TABLE 3. Conjugacy classes of the point group in the diamond structure. The elements
are indexed in the same way as the table of the symmetry operations. )
x.1
x.2
x.3
x.4
x.5
x.6
x.7
x.8
x.9
x.10
C.1
1
1
1
1
2
2
3
3
3
3
C.2 C.3 C.4 C.5 C.6 C.7 C.8 C.9 C.10
1 1 1 1 1 1 1 1
1
1 1 -1 -1 -1 -1 -1 1
1
1 1 -1 -1 1 1 1 -1 -1
1 1 1 1 -1 -1 -1 -1 -1
2 -1 0 0 -2 -2 1 0
0
2 -1 0 0 2 2 -1 0
0
-1 0 -1 1 -3 1 0 1 -1
-1 0 -1 1 3 -1 0 -1
1
-1 0 1 -1 -3 1 0 -1
1
-1 0 1 -1 3 -1 0 1 -1
TABLE 4. The character table of the basic diamond structure at Γ point.
We can apportion the wavefunction at the Γ point to each irreducible representations by applying the projection operator. (It is enough to check whether the
26
numerical result of the projection remains almost as it is, or it nearly vanishes to
zero.) The identification in the example of the diamond is given in table 5. In the
numerical expansion coefficients of the wavefunctions, we can imagine the existence
of a hidden mechanism, as is suggested by the succession of zeros, or the interchanges
of specific values with the alteration of signs. The hierarchy and the symmetry in
them can be unraveled and classified by the group theoretical analysis.
1
2
3
4
5
6
7
8
9
10
11
12
13
14
15
Ene.
Rep.
k1
0
-1
1
0
0
0
0
1
-1
-1
1
0
0
1
-1
k2
0
-1
1
-1
1
0
0
0
0
-1
1
-1
1
0
0
k3
0
-1
1
0
0
1
-1
0
0
0
0
-1
1
1
-1
kGk2
0
1.6243
1.6243
1.6243
1.6243
1.6243
1.6243
1.6243
1.6243
1.8756
1.8756
1.8756
1.8756
1.8756
1.8756
1
-0.9765
0.0742
0.0742
-0.0742
-0.0742
-0.0742
-0.0742
-0.0742
-0.0742
0
0
0
0
0
0
-2.1091
x.1
2
0
0.3519
0.3519
-0.1488
-0.1488
0.3159
0.3159
0.1848
0.1848
0.0285
0.0285
0.1325
0.1325
0.3968
0.3968
19.5799
x.10
3
0
-0.0007
-0.0007
0.1173
0.1173
0.301
0.301
-0.419
-0.419
-0.2391
-0.2391
0.3315
0.3315
-0.0935
-0.0935
19.5799
x.10
4
0
0.2934
0.2934
0.4172
0.4172
-0.1397
-0.1397
0.0159
0.0159
0.3432
0.3432
0.2199
0.2199
-0.0981
-0.0981
19.5799
x.10
5
0
-0.3562
0.3562
-0.0785
0.0785
0.084
-0.084
-0.5186
0.5186
0
0
0
0
0
0
25.1394
x.7
6
0
-0.1637
0.1637
0.5376
-0.5376
0.295
-0.295
0.0788
-0.0788
0
0
0
0
0
0
25.1394
x.7
7
0
-0.3913
0.3913
0.1078
-0.1078
-0.4612
0.4612
0.1777
-0.1777
0
0
0
0
0
0
25.1394
x.7
8
0
-0.3212
0.3212
-0.3212
0.3212
0.3212
-0.3212
0.3212
-0.3212
-0.1671
0.1671
-0.1671
0.1671
0.1671
-0.1671
33.1443
x.2
TABLE 5. The identification between the irreducible representation and the wavefunctions
in diamond crystal at Γ point. Wavefunctions are expressed by coefficients (Ci , k1 , k2 , k3 )
in the plane wave expansion Ci exp(iGr); G = k1 G1 + k2 G2 + k3 G3 , where G1 , G2 , G3 are
reciprocal vectors. The left columns (from the second to the fifth ) indicate (k1 , k2 , k3 , |G|2 )
and the remaining parts give the coefficients Ci . The energies and the irreducible representations are given in last two columns indicated by “Ene.” and “Rep.”, respectively. The
carbon atoms are placed at ±(1/8 · a1 + 1/8 · a2 + 1/8 · a3 ) so that the coefficients of the
wavefunctions in the plane wave expansion are real-valued.
27
B.
Character table computation in super-cell
It is a standard way to use the minimal unit cell. In the diamond structure, the
minimal primitive cell is chosen to be the same as the face-centered one, in which
two carbon atoms are located with the point-group G(I × O, C2 × S4 ). Meanwhile,
one can choose the cubic unit cell; the volume of which is four times as large as the
minimal cell, because the cubic cell is constructed from four of the simplest unit cell
stuck together obliquely. The unit cell enlarged in a similar way, (if possible, in which
extra atoms are embedded) is called super-cell. The point groups of the supercells
are inevitably altered by the presence of extra fractional translations. The group
theoretical computations for such super-cells will be difficult for bare human powers
since the point group becomes larger and more complicated. Nevertheless, if one
can use the computer algebra system, the character tables are easily computed. We
should note that it is more advisable to compute characters of groups by ourselves if
we can. Although the possible super-cells are infinite, the character tables available
in textbooks and databases are limited, possibly only for the minimal unit cell.
For the super-cell, actually, the point group of the minimal crystal can be extended
by the semidirect product with a finite Abelian group, by which extra fractional
translations in the enlarged cell are represented. (The construction of the semidirect
product will be stated elsewhere in the following section of this article.) The easiest
and straightest way for this is to remake the multiplication table. This can be done
in the following way: first, apply possible fractional translations to group elements
in symmetry operations of the minimal cell, so that all symmetry operations in the
supercell are prepared in the following form:
{E|τi }{Rj |σj }.
28
(19)
By means of these new set of operations, we must construct the multiplication
table in the super-cell:
{Rj |σj + τi }{Rk |σk + τl } = {Rp |σp + τq }.
(20)
In the evaluation, we can make use of the multiplication table in the minimal cell:
{Rj |σj }{Rk |σk } = {Rj · Rk |σj + Rj σk } ≡ {Rp |σp }
(21)
There is a relation:
{Rj |σj + τi }{Rk |σk + τl } = {Rj · Rk |σj + τi + Rj σk + Rj τl } = {Rp |σp + τq }. (22)
From the equations (21) and (22), it leads that
τq ≡ σj + τi + Rj σk + Rj τl − σp (modulo translational vectors)
(23)
Using this τq , the extended multiplication table can be constructed with ease.
The exemplary calculation for the diamond structure shall be executed in this section. The basic unit cell in the diamond structure is the face-centered one, including
two atoms, The lattice vectors are defined as
a1 = (0, 1/2, 1/2), a2 = (1/2, 0, 1/2), a3 = (1/2, 1/2, 0)
(24)
This basic unit cell could be extended four-fold, in which atomic coordinates are
given by translations such as x → x + τi (i = 0, 1, 2, 3). The vectors τi are defined
as:
τ0 = (0, 0, 0), τ1 = (0, 1/2, 1/2), τ2 = (1/2, 0, 1/2), τ3 = (1/2, 1/2, 0)
29
(25)
The extended unit cell is the cubic one, including eight atoms, to which the lattice
vectors are defined as:
a01 = (1, 0, 0), a02 = (0, 1, 0), a03 = (0, 0, 1)
(26)
Let F be the group composed of the four fractional translation vectors τi (i =
0, 1, 2, 3). And let T be the group of primitive lattice translations, composed of
a01 , a02 , a03 and the zero vector. The factor group N(:=F/T) is a finite group, whose
generators τ̂ are given by the projection of τi onto N. This factor group is the Klein
four group, or Z2 ⊕ Z2 . The generators satisfy the following relations.
2 · τ̂i = τ̂0
(27)
τ̂1 + τ̂2 + τ̂3 = τ̂0
The point-group G0 in the cubic unit cell is the semidirect product N o G, between
the group of the fractional translations N and the point-group G in the basic unit
cell. The new group includes 192 (= 48 × 4) elements. The textbooks or databases,
however, only show the point group of the basic, minimal unit cells, but not those of
the extended super-cells. In the latter cases, the point group is made of the semidirect
product; one can compute its character tables by straightforwardly constructing the
multiplication table, or, in a more complicated and elegant way, i.e. by making
use of the induced representation of the group, when the group N is Abelian. The
computation of the induced representation will be explained later.
The group elements in the extended group are renumbered as follows:
{E|τi }{Rj |sj } → (i × 48 + j)th element of the extended group
30
(28)
The conjugacy classes and the character tables for the extended group are given
in table 6 and 7.
Class Group Elements
Ĉ.1 [0,1]
Ĉ.2 [0,2],[0,3],[0,4],[1,2],[2,3],[3,4]
[0,5],[0,6],[0,7],[0,8],[0,9],[0,10],[0,11],[0,12],
[1,5],[1,6],[1,7],[1,8],[1,9],[1,10],[1,11],[1,12],
Ĉ.3
[2,5],[2,6],[2,7],[2,8],[2,9],[2,10],[2,11],[2,12],
[3,5],[3,6],[3,7],[3,8],[3,9],[3,10],[3,11],[3,12]
[0,13],[0,17],[0,21],[1,16],[1,17],[1,23],
Ĉ.4
[2,16],[2,18],[2,21],[3,13],[3,18],[3,23]
[0,14],[0,15],[0,19],[0,20],[0,22],[0,24],
[1,14],[1,15],[1,19],[1,20],[1,22],[1,24],
Ĉ.5
[2,14],[2,15],[2,19],[2,20],[2,22],[2,24],
[3,14],[3,15],[3,19],[3,20],[3,22],[3,24]
[0,16],[0,18],[0,23],[1,13],[1,18],[1,21],
Ĉ.6
[2,13],[2,17],[2,23],[3,16],[3,17],[3,21]
Ĉ.7 [0,25],[1,25],[2,25],[3,25]
[0,26],[0,27],[0,28],[1,26],[1,27],[1,28],
Ĉ.8
[2,26],[2,27],[2,28],[3,26],[3,27],[3,28]
[0,29],[0,30],[0,31],[0,32],[0,33],[0,34],[0,35],[0,36],
[1,29],[1,30],[1,31],[1,32],[1,33],[1,34],[1,35],[1,36],
Ĉ.9
[2,29],[2,30],[2,31],[2,32],[2,33],[2,34],[2,35],[2,36],
[3,29],[3,30],[3,31],[3,32],[3,33],[3,34],[3,35],[3,36]
[0,37],[0,40],[0,41],[0,42],[0,45],[0,47],
Ĉ.10
31
[1,41],[1,42],[2,45],[2,47],[3,37],[3,40]
[0,38],[0,39],[0,43],[0,44],[0,46],[0,48],
[1,38],[1,39],[1,43],[1,44],[1,46],[1,48],
Ĉ.11
[2,38],[2,39],[2,43],[2,44],[2,46],[2,48],
[3,38],[3,39],[3,43],[3,44],[3,46],[3,48]
Ĉ.12 [1,1],[2,1],[3,1]
Ĉ.13 [1,3],[1,4],[2,2],[2,4],[3,2],[3,3]
[1,37],[1,40],[1,45],[1,47],[2,37],[2,40],[2,41],[2,42],
Ĉ.14
[3,41],[3,42],[3,45],[3,47]
TABLE 6: Conjugacy classes for the cubic diamond structure. The element {1|τj }{Ri |σi } is denoted as [j,i].
y.1
y.2
y.3
y.4
y.5
y.6
y.7
y.8
y.9
y.10
y.11
y.12
y.13
y.14
Ĉ.1
1
1
1
1
2
2
3
3
3
3
6
6
6
6
Ĉ.2 Ĉ.3 Ĉ.4 Ĉ.5 Ĉ.6 Ĉ.7 Ĉ.8
1 1 1 1 1 1 1
1 1 -1 -1 -1 -1 -1
1 1 -1 -1 -1 1 1
1 1 1 1 1 -1 -1
2 -1 0 0 0 -2 -2
2 -1 0 0 0 2 2
-1 0 -1 1 -1 -3 1
-1 0 -1 1 -1 3 -1
-1 0 1 -1 1 -3 1
-1 0 1 -1 1 3 -1
-2 0 -2 0 2 0 0
-2 0 2 0 -2 0 0
2 0 0 0 0 0 0
2 0 0 0 0 0 0
Ĉ.9 Ĉ.10 Ĉ.11 Ĉ.12 Ĉ.13 Ĉ.14
1
1
1
1
1
1
-1
1
1
1
1
1
1 -1 -1
1
1 -1
-1 -1 -1
1
1 -1
1
0
0
2
2
0
-1
0
0
2
2
0
0
1 -1
3 -1
1
0 -1
1
3 -1 -1
0 -1
1
3 -1 -1
0
1 -1
3 -1
1
0
0
0 -2
2
0
0
0
0 -2
2
0
0 -2
0 -2 -2
2
0
2
0 -2 -2 -2
TABLE 7. The character table of the cubic diamond unit cell at Γ point.
The eight lowest eigenvalues in energy spectra in the minimal and cubic diamond
cell are shown in table 8. There are two sets of threefold degeneracy in the former
32
case. On the other hand, in the latter, there is one sixfold degeneracy, which is
attributable to the overlap of the spectra at the three X points in the Brillouin zone
of the minimal unit cell. However, the character tables of the Γ point in the minimal
diamond cell could not assign them in suitable representation; the character tables
rebuilt for the cubic cell can do this.
Minimal
-2.11 19.58 19.58 19.58 25.14 25.14 25.14 33.14
Representation x.1 x.10 x.10 x.10 x.7 x.7 x.7 x.2
Cubic
-2.11 6.74 6.74 6.74 6.74 6.74 6.74 13.03
Representation y.1 y.14 y.14 y.14 y.14 y.14 y.14 y.12
TABLE 8. The energy spectra at the Γ point in the minimal and cubic diamond unit cell
(unit in eV). The assigned irreducible representations are also shown.
V.
APPLICATION 2: A SYSTEMATIC WAY OF THE MATERIAL DE-
SIGNING
A.
How to manage quantum-dynamics in crystal?
The group theoretical view will be able to bring about a systematic way of material design. In this section, although in the level of table-top exercise, a practicable
form of material design is presented. In a nutshell, through the systematic reduction of the crystal symmetry, the band-structure could tactically be modulated; the
degeneracy of the energy spectrum can be artificially split into different levels and
the new energy-gap can be opened. Toward the systematic reduction of the crystal
symmetry, the crystallographic concept, called Wyckoff positions, is utilized[28, 33].
This mathematical idea is not an abstract one, but actually computable by means
of the computer algebra.
33
One should note this: the degeneracy of the energy spectrum is governed by the
symmetry of the crystal. From the mathematical viewpoint, the point group of the
crystal system is the manifestation of the symmetry; the subgroups of the point
group describe the possible style with the reduced symmetry.(It is well known that
a material has general tendency to reduce its own symmetry in order to release itself
from the energetically unstable situation owing to the existence of the degeneracy
of occupied energy levels.) By tracing the tree structure of the inclusion relations
of the subgroup of the finite groups, one can enumerate every possibility of reduced
symmetry. For each of reduced symmetries, one can compute group representations.
The comparison of the character tables in the subgroups and the parent groups (by
means of analyzing compatibility relations) one can rummage the possibility of the
breaking-down of the energy spectrum degeneracy. The opening of the energy-gaps
in the reduced symmetry will be speculated, sorted and classified with their group
theoretical origins. This plan, therefore, will lead to a kind of systematic way of
material design. It must be admitted that such a tactic still lacks in the sufficiently
quantitative argument, since the only possible existence of energy-gap, together with
their relationship with crystal symmetry reduction, is discussed. In order to evaluate
the stability of the reduced structure or the actual possibility of its existence, one
must employ massive first-principles computation. However, the group theoretical
inspection has its own merit; it can be achieved by a light and quick computation,
frugal in hardware usage; maybe simple computations in the level of elementary
perturbation theory will do. Hence it can be applied to screening purposes for picking
up candidates before serious massive first-principles computation.
As an example, consider the crystallographic group of a cube, including following
operations.
• Three rotations along x, y, or z axes by 90 degrees.
34
FIG. 2. The paradigm of the crystal symmetry, or the subgroup lattice.
• Exchange of x, y, and z axes ( Rotations along the diagonal in the cube by 120
degrees).
• The inversion and reflections are omitted for simplicity.
These rotations generate 24 operations in the cubic lattice to construct the point
group. Using GAP, The relationship of the inclusion in subgroups (the lattice subgroup) can be inquired and visualized(Fig.2). The point group of the cubic lattice
is located at the topmost node of the tree. We can trace the subgroups of lower
symmetries in descending along the branches and finally arrive at the trivial group
(the identity group), which is located at the root of the tree. This means that we
can explore and predict the alternation of the band structure (viz, the splitting of
the degenerated energy levels, or the opening of new band gaps) systematically, by
consulting with the “paradigm” of the crystal symmetry.
Now there arise following questions.
Question 1. How to make up actual crystal structures in correspondence with the
35
subgroups of lower symmetries?
Answer 1. The reduction of the symmetry may be occasioned, for example, by the
presence of alien atoms, or the occurrence of point defects in the perfect crystal. The
change in the electronic structure will be investigated by simple perturbation theory
with sufficient accuracy.
Question 2. In what positions alien atoms or point defects can be located? It is easy
to cause the great change the symmetry, or, to abolish it completely. However, for the
case with the small change in the symmetry, for example, that between neighboring
connected entries in the subgroup lattice, can we set up favorable atomic dispositions?
Answer 2. Choose an arbitrary point. Apply all of the operations of the subgroup
in consideration to this coordinate. The number of the generated points is equal to
the order of the subgroup. The set of these points is invariant with respect to the
operation of the subgroup. By putting alien atoms or defects in this set of points,
we can reduce the symmetry of the crystal, and, by continuing this way, descend
through the subgroup lattice.
Question 3. It is impossible for us to put alien atoms at arbitrary points in the
crystal. With the view of the material design, a certain restriction should be imposed
on the number of alien atoms. We would like to know the set of atomic sites, with a
limitation in number, and invariant by the subgroup operation.
Answer 3. Make use of the concept of the Wyckoff position.
The mathematical definition is given in the appendix. In a familiar word of
material sciences, the Wyckoff positions are the set of equivalent points in the unit
cell, which are transformed among themselves by symmetry operations. In general,
the total number of such points is the same as that of the symmetry operations (the
36
order of the point group). However, if the generating point of the Wyckoff position
remains fixed by certain operations of the point group, the total number of Wyckoff
positions will be fewer, which will be a factor of the order of the point group. In a
mathematical word, it is equal to the order of some factor group. In GAP program,
the Wyckoff position is easily computed from a “Cryst” package.
In this example of the cubic lattice, the Wyckoff positions are classified as follows:
• A discrete point (0,0,0).
• A discrete point (1/2,1/2,1/2).
• The centers of the edges of the cube:(1/2,0,0),(0,1/2,0),(0,0,1/2).
• The centers of the faces:(1/2,1/2,0) (1/2,0,1/2),(0,1/2,1/2)
• A segment (x,0,0) (0<x<1),invariant by 90-degrees rotation along the x axis.
• A segment (0,y,0) (0<y<1), invariant by 90-degrees rotation along the y axis.
• A segment (0,0,z) (0<z<1), invariant by 90-degrees rotation along the z axis.
• A segment (x,1/2,1/2) (0<x<1),invariant by 90-degrees rotation along the x
axis.
• A segment (1/2,y,1/2) (0<y<1), invariant by 90-degrees rotation along the y
axis.
• A segment (1/2,1/2,z) (0<z<1), invariant 90-degrees rotation along the z axis.
• A segment (t,t,t), invariant 120-degrees rotations which exchange three axes.(In
a similar way, all of the diagonal lines of the cube belong to this type.)
• Several segments (not discreet points), connected to the vertexes, the centers
of the edges.
37
• Arbitrary points in the unit cell.
In table 9, the Wyckoff positions of the group of full symmetry of the cubic cell
and those of one of the subgroups are compared. These two groups have common
generators of the Wyckoff positions. However, in some of them, the number of
equivalent positions differs. By making use of such positions and stationing atoms in
them, we can switch the crystal symmetry. For example, take the Wyckoff positions
given in the 7-th low. They are in the orbit of the generating point (t, 0, 1/2).
The group of No.11 and No.10 are isomorphic to the symmetric group S4 and the
alternative group A4 respectively. In the full symmetry, given by the group of No.11,
the orbit is composed of 12 points. On the other hand, in the reduced symmetry of
the group No.10, these 12 points are split into two orbits, each of which includes 6
points. If atoms (say, of type W) are positioned in all of 12 points in the cubic lattice,
the crystal symmetry remains unaltered as that of the group of No.11. However, if,
in the half part of them, at the Wyckoff positions of the Group No.10, the atoms are
replaced by those of type B, the crystal symmetry is reduced to the group of No.10.
Those Wyckoff positions are illustrated in Fig. 3, where two split orbits are shown
in B(black) and W(white) atoms.
B.
Control of band gaps by means of the Wyckoff positions
In this section, the example of the material design discussed in the previous section
is presented; in which ab-initio electronic structure calculations are executed, where
the symmetry of the crystal is artificially altered and the band gaps are controlled.
The exemplary calculation goes in the following way: in order to alter the crystal symmetry, certain coordinate points are chosen; and at these localities, we put
impurities, which interact to electrons with short-ranged potentials; these impurities
38
+--------------+
|
|
|
W
|
|
|
| B
TOP
B |
|
|
|
W
|
|
|
+--------------+--------------+--------------+--------------+
|
|
|
|
|
|
W
|
B
|
W
|
B
|
|
|
|
|
|
| B LEFT
B | W FRONT W | B RIGHT B | W REAR
W |
|
|
|
|
|
|
W
|
B
|
W
|
B
|
|
|
|
|
|
+--------------+--------------+--------------+--------------+
|
|
|
W
|
|
|
| B BOTTOM B |
|
|
|
W
|
|
|
+--------------+
FIG. 3. An illustration of split Wyckoff Orbit in the six faces of the cube.
are assumed to be neutral; therefore they do not alter the net charge. The change in
the symmetry of the crystal will be caused by the presence of impurities, the effects
of which are expressed by model potentials with adjustable parameters. We use the
following form of the scattering potential. It is given in the real space as
1
V (r) = C exp −
2
39
r
rloc
2 !
.
(29)
Group NO.11
(#G=24)
[0,0,0]
[1/2,1/2,1/2]
[0,1/2,1/2]
[1/2,0,0]
[t,0,0]
[t,1/2,1/2]
Size
1
1
3
3
6
6
[t,0,1/2]
12
[t,t,0]
[t,t,1/2]
[t,t,t]
[t1,t2,t3]
12
12
8
24
Group NO.10
(#G=12)
[0,0,0]
[1/2,1/2,1/2]
[0,1/2,1/2]
[1/2,0,0]
[t,0,0]
[t,1/2,1/2]
[0,t,1/2]
[1/2,t,0]
(Split in two)
not existent
not existent
[t,t,t]
[t1,t2,t3]
Size
1
1
3
3
6
6
6
6
0
0
4
12
TABLE 9. The Wyckoff positions in the cubic lattice. The left three columns gives the
generators and the right two show the number of equivalent points.
The reciprocal space representation is
1
1
2
V (K) = Z exp − (rloc K) ,
Ω
2
(30)
where Ω is the volume, required in the normalization of the plane wave √1Ω exp (iKr).
p
3
and Z = C (2π)3 (rloc
). There are two parameters: Z is the strength of the potential
and rloc is the effective range.
A simple example is taken from diamond band structure. The coordinate of the
carbon atoms are
(0 · V1 + 0 · V3 + 0 · V3 )
(1/4 · V1 + 1/4 · V3 + 1/4 · V3 ).
40
(31)
The lattice vectors are
a a
V1 =
, ,0 ,
2 2
a
a
V2 =
, 0,
,
2
2 a a
V3 = 0, ,
.
2 2
(32)
In this structure, the lowest eight eigenvalues at Γ point are shown in table 10.
There are two set of threefold degeneracy above and below the band gap, the energy
Γ
-2.11 19.58 19.58 19.58 25.14 25.14 25.14 33.14
Representation x.1 x.10 x.10 x.10 x.7 x.7 x.7 x.2
TABLE 10. The lowest eight eigenvalues at Γ point in the diamond structure (unit in eV).
levels of which are at 19.58 eV and 25.14 eV. (The irreducible representations to
them are x.10 and x.7 respectively, as given in the previous section.)
The band gap (in exact terms, the degeneracy of the eigenvalues) is modified by
the insertion of a single impurity in the crystal. Its position is chosen at (1/8 · V1 +
1/8 · V3 + 1/8 · V3 ). The point group of the genuine diamond structure ( of the order
of 48) is reduced to its subgroup (of the order of 12.) The symmetry operations of
the subgroup are shown in table 11.
The character table of this subgroup is given in table 12. The rows S.1,..., S.6
show the character values of the subgroup, and the rows x.7 and x.10 shows those of
the group of the full diamond symmetry (without impurity).
In the rows of the table, there are two relations: x.7=S.2+S.5 and x.10=S.1+S.6;
which are called matching relations.(In some mathematical context, they are referred
as branching rules.) These relations indicate the threefold degeneracy in the energy
levels in the full diamond symmetry split into two irreducible representations S.2 and
S.5, or S.1 and S.6 with the presence of an impurity. The character values of the
41
TABLE 11. The symmetry operations with the presence of an impurity. The rows (from the
second to the bottom ) show the rotation matrix R (R11, R12,..., R33) and the translation
T (T1, T2, T3).
1
2
3
4
5
6
7
8
9
10
11
12
R11 R12 R13 R21 R22 R23 R31 R32 R33 T1 T2 T3
1
0
0
0
1
0
0
0
1 0 0 0
0
1
0
0
0
1
1
0
0 0 0 0
0
0
1
1
0
0
0
1
0 0 0 0
0 -1
0 -1
0
0
0
0 -1 1/4 1/4 1/4
-1
0
0
0
0 -1
0 -1
0 1/4 1/4 1/4
0
0 -1
0 -1
0 -1
0
0 1/4 1/4 1/4
-1
0
0
0 -1
0
0
0 -1 1/4 1/4 1/4
0 -1
0
0
0 -1 -1
0
0 1/4 1/4 1/4
0
0 -1 -1
0
0
0 -1
0 1/4 1/4 1/4
0
1
0
1
0
0
0
0
1 0 0 0
1
0
0
0
0
1
0
1
0 0 0 0
0
0
1
0
1
0
1
0
0 0 0 0
TABLE 12. The character table of the diamond structure with an impunity. The characters
for each operation are shown here.
S.1
S.2
S.3
S.4
S.5
S.6
x.7
x.10
1
1
1
1
1
2
2
3
3
2
1
1
1
1
-1
-1
0
0
3
1
1
1
1
-1
-1
0
0
4
1
-1
-1
1
0
0
-1
1
5
1
-1
-1
1
0
0
-1
1
6
1
-1
-1
1
0
0
-1
1
7
1
-1
1
-1
-2
2
-3
3
8
1
-1
1
-1
1
-1
0
0
9
1
-1
1
-1
1
-1
0
0
10
1
1
-1
-1
0
0
1
1
11
1
1
-1
-1
0
0
1
1
12
1
1
-1
-1
0
0
1
1
identity element, shown in the second columns from the left, are the degeneracy of
the energy levels; so we can predict the threefold-degeneracy split into two different
levels, viz. twofold and single levels.
To inspect this, the actual computation is executed; where the impurity potential
is set to be a weak Gaussian form. For example, set the parameters in Eq.30 as
Z = −0.1a.u., rloc = 0.1a.u. The lowest eight eigenvalues are given in table 13. The
42
prediction by group theory is realized there.
Γ -2.18 18.86 19.67 19.67 25.04 25.18 25.18 33.13
TABLE 13. The lowest eight eigenvalues at Γ point in the diamond structure with an
impurity (unit in eV).
In the practice of the artificial material design, for example, the following demand
will arise:
Demand: How to substantiate the situation, in which the band gap is
altered, but the degeneracy of the energy levels is unaffected?
Let us proceed with the example of the diamond. In the diamond structure, there
are plural possible Wyckoff positions.(The computation can be done by GAP.) For
example, chose the position of (1/8V1 +1/8V2 +1/8V3 ). By the symmetry operations,
it is transformed into three other points such as: (5/8V1 + 1/8V2 + 1/8V3 ), (1/8V1 +
5/8V2 + 1/8V3 ), (1/8V1 + 1/8V2 + 5/8V3 ). If impurities are stationed at these four
points, there is no change in the point group; thus the degeneracy of the energy
spectrum remains unchanged. The computed energy spectrum is given in table
14, which guarantees the validity of the group theory. The band dispersions along
Γ(0, 0, 0) − X(2π/a, 0, 0) are shown in Fig. 4, in which that of the basic diamond
structure, and that with single impurity, and that with four impurities are compared.
Γ -2.41 18.83 18.83 18.83 25.11 25.11 25.11 33.07
TABLE 14. The lowest eight eigenvalues at Γ point in the diamond structure with four
impurities (unit in eV).
There is a certain noteworthy point in these two cases (of the single impurity,
and of the four impurities). In the former, the single impurity is located at(1/8V1 +
1/8V2 + 1/8V3 ); this point is one of the Wyckoff positions in the point group of the
43
crystal with the deteriorated symmetry; it remains fixed by any symmetry operation
and not transformed into other positions; however, this point also belongs to the
Wyckoff position of the perfect diamond crystal and there transformed to four points.
We can interpret this circumstance as follows: the former example corresponds to a
situation, where the only one of the four Wyckoff position in the latter case is chosen
so that the symmetry of the crystal could be reduced. In a similar way, (but in a
more general and liberal way, by inserting extra atoms or excavating vacancy site) we
can reduce the crystal symmetry for the purpose of controlling the band structure;
by doing this we can cleave into the degeneracy of the energy spectra and open extra
energy gaps.
C.
A possible plan of material design in the super-cell
Now we have seen that the band structure is controllable by the presence of impurities or defects, although the presented example is that of instructive purpose,
and not realistic, as an only model case for the sake of illustrating a possible form
of theoretical material design. The diamond structure is compressed in such a compact cell that the supernumerary atoms could not be actually stationed at arbitrary
places, such as at Wyckoff positions. Meanwhile, in recent decades, crystalline materials which have a large unit cell such as In3 O2 have been put into practical use.
At a glance, In3 O2 seems to have a huge complex structure. However, actually, it
is a super-cell composed from the stacking of body-centered lattice of In-O, with
oxygen vacancies. The band dispersion is almost free-electron-like; its behavior under the presence of doped Sn, viz. impurities in the cell, can be understood by the
group theoretical view[34]. Furthermore, the modern material composing technologies, such as epitaxial growth, or micro atom-scale probes, may enable us to realize
such microscopic disposition of atoms in the supercell structure, which will lead to
44
FIG. 4. The band dispersions in diamond structures altered by the presence of impurities,
along Γ-L direction. The left figure is that without impurity: the center, with a single
impurity: the right, with four impurities.
an actual phase of material design. In the super-cell, there is a more broad freedom
in the atomic arrangement, in which certain atoms could be replaced or removed so
that the periodical arrangement of impurities or defects should be permitted.
Now a principle of material design based on group theoretical view can be proposed.
45
1 Chose a certain crystal as a starting point: we may start from simple atomicand electronic- structures, such that we can execute the electronic structure
computation easily, as the primitive module of the design.
2 Consider an enlarged cell, composed of the stacking of the minimal unit cell.
(The crystal axes may be cut newly and a supercell should be created.)
3 We can determine the band structure and the band gap of the enlarged cell,
simply by folding the band structure of the minimal cell.(The folding of the
band structure is to re-distribute k-points of the old Brillouin zone to the new
one in accordance with the new periodicity and to redraw the band dispersion
by means of new k-points.)
4 Suppose a situation where the symmetry is reduced by the formation of the
supercell with atomic displacements. Check the possibility of the opening of
the extra band gaps (or the splitting of the degeneracy) at k-points with high
symmetries, such as at Γ points and the boundary of the Brillouin zone. The
group theoretical view, such as the character theory, the compatible relation,
the subgroup lattices, or the Wyckoff positions, should be fully made use of.
The change in the band structure may be quantitatively evaluated by simple
perturbation theory. These sorts of trials-and-errors, by tracing the sequence of
crystal symmetry, will help us to search crystal structures with more desirable
band structure.
VI.
TECHNICAL DETAILS
In the preceding sections of the article, the basics of the usage of computer algebra, concerning the computation of group theoretical properties, are explained and
applied to the analysis of the wave-functions. Moreover, the possible scheme for
46
controlling band gaps, by means of tracing the sequence of subgroups, which corresponds to the reduction of crystal symmetry, is suggested through simple electronic
structure computations. Meanwhile, in this part, the topics are rather promiscuous
and technical, such as the computation of characters of the group of the supercell
by means of “semi-direct product” which is an effective way to extend groups or the
special treatment for non-symmorphic crystal, and so on. Moreover, “mathematical
first principle” of crystal formation is demonstrated based on the cohomology theory,
which shall explain to us how specific types of crystals, finite in number, can exist
in accordance with the actions of certain point groups. The topics might arrive at
somewhat higher mathematics, that may be not well-known to physicists. However,
with the aid of computer algebra, one can conquer such hardships so that he might
actually execute necessary computations. As for the fundamental knowledge of the
group theory and its application, the reader should consult with standard references,
such as Refs.[1–6] for the group theory, such as Refs.[7–9] for the representation theory, such as Refs.[10, 11] for the algorithm in the computer algebra. We make use of
the computer algebra package GAP[12], as well as in the previous sections.
A.
Determination of crystal symmetry
It is essential for us to determine the crystal symmetry in order to execute the
group theoretical analysis. The symmetry operations which fix the crystal axes and
atomic coordinates must be listed to construct the point group. The determination
algorithm of the point group is composed of the following two stages.
Stage 1. Determination of the point group of the lattice.
Let A = (A1 , A2 , A3 ) be the lattice vectors. There are two large categories of the crystal system; the one of which is hexagonal, and the other
47
is cubic. In the former, rotations along the crystal axis by 60 degrees are
allowed; in the latter, the rotations of this kind are composed of those
by 90 degrees. The point group of arbitrary crystal system is given as
a subgroup which is reduced from the full hexagonal or cubic symmetries. The full set of the symmetry operations (i.e. rotation matrices) for
the hexagonal or cubic symmetry should be prepared, and from which
admissible operations are extracted.
For each operation R of the full symmetry group, compute the rotation
of the lattice vectors:
R · Ai (i = 1, 2, 3).
(33)
If the vector R · Ai protrudes out of the unit cell, it must be pulled
back according to the crystal periodicity. Hereafter the pulling back of
a vector x into the unit cell is denoted as P (x). For example, R · Ai =
l · A1 + m · A2 + n · A3 for integers l, m, n, then exactly P (R · Ai ) = 0.
Then store the operation R such as
kP (R · Ai )k ≤
for all Ai (=1,2,3), with a threshold value ( This is an attention needed
only in the floating-point computation.) From the stored set of admissible
operations {R}C , the point group of the crystal system is determined; in
this stage, however, the symmetry preservation in the atomic positions
is not yet checked.
Stage 2. Determination of the point group of the crystal.
The symmetry operations which fix atomic coordinates should be ex48
tracted now. The atom transformation table, the fractional translations
associated to each rotation, and the multiplication table should be provided at the same time. Let X and Y be entries in the set of atomic
positions {X}, and let T(X) and T(Y) be the corresponding type of
atoms.
For each R ∈ {R}L , and for each X ∈ {X}, compute R · X; and chose
every Y ∈ {X} such as T(X) = T(Y). Now the vector τc = P (R · X − Y )
is a candidate for the fractional translation associated to R. If τc is the
admissible fractional translation to R, the following conditions should be
satisfied.
Condition: for all X1 in the atomic coordinates, there exist a Y1
such that
kτc − P (R · X1 − Y1 )k ≤ ,
and
T (X1 ) = T (Y1 ).
If this condition is satisfied, store R and τc as a symmetry operation
{R|τc } in the point group of the crystal.
This algorithm works well if the unit cell is minimal, in the sense that the fractional
translation is uniquely determined; i.e. with respect to one symmetry operation of
{R|t}, other operations, such as {E|s}{R|t} = {R|t + s}, do not coexist in the point
group operations. If the unit cell is not minimal, in other words, if it is a supercell
(the accumulation of the minimal unit cell), the existence of plural τ for one R cannot
be negligible. The readers will be able to find and utilize subroutines or functions
purposed for the identification of the crystal symmetry (possibly with more efficacy
than the algorithm in this section)in the band calculation packages of one’s own.
49
B.
Symmetry operations in the reciprocal space
As for the plane wave, the symmetry operation in the real-space is given as
{RT |τT } exp(ikr) = exp ik RT−1 r − RT−1 τT
= exp (i(RT · k) · (r − τT )) .
(34)
Consequently, it is more convenient to give operations with respect to reciprocal
lattice vectors in the wave number space. If the wave-number-vector k is expressed
as k1 · G1 + k2 · G2 + k3 · G3 , the rotation matrix for coefficients in the rotated vector
RT · k is given as
0
R11
0
R12
0
R13
k
1
0
0
0
R21 R22
k2 .
R23
0
0
0
k3
R31 R31 R33
(35)
Thus the rotation matrix for the coefficients is computed from that in the Euclidean
space as:
R0 := (a1 , a2 , a3 )−1 RT (a1 , a2 , a3 )
T
(36)
where a1 , a2 , a3 are primitive lattice vectors.
For the case of the face-centered unit cell, the gap computation is done as:
gap> A:=[ [ 0, 1/2, 1/2 ], [ 1/2, 0, 1/2 ], [ 1/2, 1/2, 0 ] ];;
gap> RM:=[ [ 1, 0, 0 ], [ 0, -1, 0 ], [ 0, 0, -1 ] ];;
gap> TransposedMat(A^-1*TransposedMat(RM)*A);
[ [ -1, 0, 0 ], [ -1, 0, 1 ], [ -1, 1, 0 ] ]
The rotation matrices in the reciprocal space, as well as those in the real space,
are given in the appendix for cubic and hexagonal symmetries.
50
C.
The computation of the compatibility relation
The irreducible representation of the diamond structure shows branching if the
k-point moves from Γ point to non-zero k-point. The situation of the branching (the
compatibility relation) can be checked by GAP. For this purpose, the computation
goes as follows:
• Compute the character table for the small group. Then, for each representation,
evaluate the characters of elements in the small group, and store them in the
lists S1 , S2 , ..., SN . (Let N be the number of the representations of the Small
group.)
• Compute character values of elements in the small group, using the irreducible
representation at the Γ point, and store them in the lists G1 , G2 , ..., GM .(Let
M be the number of the representations of the point group.)
• If the j-th irreducible representation (G.j) branches into the composition of
those in the small (S.i) as follows
Gj =
N
X
Cji × Si
(37)
i=1
the coefficients are computed by the inner product of the lists as
Cji = Gj · Si / |#K|
where |#K| is the order of the small group.
For this purpose, this function can be used:
GetCompati:=function(A,B)
51
(38)
local repA,repB,AV,BV,CV;
repA:=Irr(A);
repB:=Irr(B);
if ( false=IsSubgroup(A,B) ) then
Print("A < B ; Inverse A and B."); return;
fi;
AV:=List(repA,y->List(Elements(B),x->x^y));
BV:=List(repB,y->List(Elements(B),x->x^y));
CV:=AV*TransposedMat(BV)/Size(B);
return CV;
end;
This function computes the characters of the group elements in the group “B”,
in two ways, those by the irreducible representation of group “A” and group “B”.
The results are stored in two lists “AV” and “BV”. Each row of these lists contains
the characters of each irreducible representation. By the computation of the inner
product between these lists, the ramification of the representations is computed.
Symmetry operations defined in wave-number space are used here. These matrices
are three-dimensional ones given in Table 45.(They are stored in the list “MT[ ]” in
the following computation.) The coordinate of the K point along Γ − X point is
given, as the coefficients of reciprocal lattice vectors.
gap> kx:=[0,1,1];
[ 0, 1, 1 ]
The symmetry operations which fix the k-point are extracted, and from them, the
small group is generated.
gap> FL:=Filtered(MT,i->i*kx=kx);
52
[ [ [ 1, 0, 0 ], [ 0, 1, 0 ], [ 0, 0, 1 ] ],
[ [ -1, 0, 0 ], [ -1, 0, 1 ], [ -1, 1, 0 ] ],
[ [ 0, -1, 1 ], [ 0, 0, 1 ], [ -1, 0, 1 ] ],
[ [ 0, 1, -1 ], [ -1, 1, 0 ], [ 0, 1, 0 ] ],
[ [ 0, 1, -1 ], [ 0, 1, 0 ], [ -1, 1, 0 ] ],
[ [ 0, -1, 1 ], [ -1, 0, 1 ], [ 0, 0, 1 ] ],
[ [ 1, 0, 0 ], [ 0, 0, 1 ], [ 0, 1, 0 ] ],
[ [ -1, 0, 0 ], [ -1, 1, 0 ], [ -1, 0, 1 ] ] ]
We can directly construct the small group from the above list of the output. For
the purpose of the comparison of the representations, however, the point group and
the small group must be expressed by common abstract symbols such as “m1,m2,...”,
which are employed to make up the multiplication table.(There is an alternative: the
point group could be constructed from the 48 matrices of symmetry operations in
wave-number space.)
gap> FL:=List(FL,i->Position(MT,i));
[ 1, 2, 19, 20, 27, 28, 41, 42 ]
gap> FL:=List(FL,i->Elements(G)[i]);
[ m1, m2, m19, m20, m27, m28, m41, m42 ]
gap> GS:=Group(FL);
<group with 8 generators>
The character table of the full point group GN is given as:
gap> ConjugacyClasses(GN);
[ m1^G, m2^G, m5^G, m13^G, m14^G, m25^G, m26^G, m29^G,
m37^G, m38^G
]
gap> Display(Irr(GN));
53
[ [
1,
1,
1,
1,
1,
1,
1,
1,
1,
1 ],
[
1,
1,
1,
-1,
-1,
-1,
-1,
-1,
1,
1 ],
[
1,
1,
1,
-1,
-1,
1,
1,
1,
-1,
-1 ],
[
1,
1,
1,
1,
1,
-1,
-1,
-1,
-1,
-1 ],
[
2,
2,
-1,
0,
0,
-2,
-2,
1,
0,
0 ],
[
2,
2,
-1,
0,
0,
2,
2,
-1,
0,
0 ],
[
3,
-1,
0,
-1,
1,
-3,
1,
0,
1,
-1 ],
[
3,
-1,
0,
-1,
1,
3,
-1,
0,
-1,
1 ],
[
3,
-1,
0,
1,
-1,
-3,
1,
0,
-1,
1 ],
[
3,
-1,
0,
1,
-1,
3,
-1,
0,
1,
-1 ] ]
The character table of the small group is as follows:
gap> List(ConjugacyClasses(GS),Elements);
[ [ m1 ], [ m2 ], [ m19, m20 ], [ m27, m28 ], [ m41, m42 ] ]
gap> Display(Irr(GS));
[ [
1,
1,
1,
1,
1 ],
[
1,
1,
-1,
-1,
1 ],
[
1,
1,
-1,
1,
-1 ],
[
1,
1,
1,
-1,
-1 ],
[
2,
-2,
0,
0,
0 ] ]
The branching relations are computed now:
gap> Display(GetCompati(GN,GS));
[ [
1,
0,
0,
0,
0 ],
[
0,
1,
0,
0,
0 ],
[
0,
0,
1,
0,
0 ],
[
0,
0,
0,
1,
0 ],
54
[
0,
1,
0,
1,
0 ],
[
1,
0,
1,
0,
0 ],
[
1,
0,
0,
0,
1 ],
[
0,
0,
0,
1,
1 ],
[
0,
0,
1,
0,
1 ],
[
0,
1,
0,
0,
1 ] ]
This output list should be read as the indication of the following branching relation:
POINT GROUP
SMALL GROUP
G.1
->
S.1
G.2
->
S.2
G.3
->
S.3
G.4
->
S.4
G.5
->
S.2 + S.4
G.6
->
S.1 + S.3
G.7
->
S.1 + S.5
G.8
->
S.4 + S.5
G.9
->
S.3 + S.5
G.10
->
S.2 + S.5
D.
The character table at the boundary of the Brillouin zone
1.
The necessity of the special treatment for the non-symmorphic crystal
In general non-zero k-points, the wave-functions are apportioned to representations different to that of the Γ point. The new representations are deduced from
subgroups of the point group; the subgroup admissible to each k-point must be
55
composed of operations which fix that k-point. The invariance on the k-point by
symmetry operations should be kept up to the periodicity of the wave-space. Thus
at the boundary of the Brillouin zone, different but equivalent k-points are mingled
together in the representation. In general, it is sufficient to choose the subgroup
which connects equivalent k-points and compute the representation in the same way
as in Γ point. However, in special cases, this naive approach inevitably fails, owing
to the existence of two different kinds of crystal structures, i.e. the distinction between the symmorphic and non-symmorphic ones. Let us review this troublesome
situation.
The matrix representation could be written as
Dk ({α|τ }) = exp(−ik · τ )Γ(α).
(39)
Then the product is
Dk ({α1 |τ1 })Dk ({α2 |τ2 }) = exp(−ik·(τ1 +τ2 ))Γ(α1 )Γ(α2 ) = exp(−ik·(τ1 +τ2 ))Γ(α1 α2 ).
(40)
The product should be equal to this:
Dk ({α1 · α2 |α1 · τ2 + τ1 }) = exp(−ik · (α1 · τ2 + τ1 ))Γ(α1 α2 ).
(41)
These equations lead to the relation:
exp(i(α1−1 · k − k) · τ2 ) = 1.
(42)
Certain conditions are required so that the above relation should be valid:
1 The k point is inside the Brillouin zone; α1−1 does not move k (i.e. α1 · k = k).
56
At the Γ point, this condition is always satisfied.
2 The k point is located at the boundary of the Brillouin zone: the vector α1 ·k−k
may coincide with a certain reciprocal vector due to the periodicity on the wave
space. The subgroups of the operations of these kinds are called small groups.
In addition, the τ2 should be zero or a lattice translation vector in the real
space.
The crystal structure, in which the second condition always holds, is said to be
symmorphic. In certain types of crystals, however, there exist affine mappings with
non-zero τ2 , which are fractions of some lattice translational vectors. This type
of the crystal is called non-symmorphic. Indeed, the diamond structure is nonsymmorphic one, as can be seen in the symmetry operations for this crystal. For
the non-symmorphic case, therefore, the relation to be satisfied with respect to the
product of the matrix representation does not hold at the boundary of the Brillouin
zone. Thus it is necessary to employ special methods to treat this case. One of these
treatments was proposed by Herring[32]. The intuitive interpretation of this method
could be given in the following way. One should bear in mind these two points: First,
at the Γ point, there is no problem in the representation even in the non-symmorphic
case. Secondly, the k-point at the boundary of the Brillouin zone can be pulled back
at the Γ point, if the unit cell is extended. For example, in a one-dimensional crystal
with the lattice constant of a, the k-point k = π/a is at the boundary of the Brillouin
zone. However, when the doubled unit cell (with the lattice constant 2a) is assumed,
the k-point k = π/a agrees with the new primitive reciprocal lattice vector; hence
it is equivalent to the Γ point. The irreducible representation at π/a in the small
original cell can be deduced from that at the Γ point in the enlarged cell. Thus the
difficulty in non-symmorphic representation can be eluded. We shall examine the
actual computation of this method in the next section.
57
2.
Herring Method: An example
FIG. 5. Symmorphic and non-symmorphic one-dimensional crystals.
Assume a double array of atoms, as in the Fig. 5. This figure shows two types of
one-dimensional periodic lattice, consisted by double lines of atoms. In case I, two
rows forms one dimensional periodic system of squares. In case II, the two lows are
shifted by half of the lattice constants. The latter is of non-symmorphic type. The
58
symmetry operations which fix the k-point at k = π/a for the latter case is given as
e = {E|0}
(43)
c = {C2 |0} A rotation by 180 degrees along the origin
m = {σx |a/2} A reflection over x axis
g = {σy |a/2} A reflection over y axis
From the aforementioned reasons, the irreducible representation of this group is not
the proper one to represent the small group at k = π/a.
According to the Herring method, instead of this, a factor group G/I should be
made use of, where G is a group of infinite order, composed of
e0 = {E|2na},
(44)
c0 = {C2 |2na},
m0 = {σx |a/2 + 2na},
g 0 = {σy |a/2 + 2na},
ē0 = {E|a + 2na},
c̄0 = {C2 |a + 2na},
m̄0 = {σx |3a/2 + 2na},
ḡ 0 = {σy |3a/2 + 2na},
and I is the normal subgroup composed of
e0 = {E|2na}.
(45)
This factor group is the point group in the unit cell extended twofold in the x
direction. By extending the unit cell, the k-points at the boundary of the original
59
Brillouin zone are pulled back at the Γ point in the new Brillouin zone. As the newly
allotted Γ-point is the zero vector, the mathematical difficulty in the representation
can be avoided. This situation should rather be interpreted as follows: the wavefunction at k = π/a has the periodicity of 2a, that is two times as large as the original
lattice width. On the other hand, the wavefunction at k = 0 in the original Brillouin
zone also matches into the periodicity of 2a.(Because originally it has periodicity
of a) In the representation at the new Γ point, those in the old Γ point and the
old boundary point are mingled with each other. These two different contributions,
however, can be distinguishable. Let us check the character of the elements of following types {E|τ }, viz. of the simple fractional translations in the extended cell, in
which the linear operation is the identity. With respect to the representation of the
boundary of the Brillouin zone of the original smaller cell, the representation matrix
should be represented as Dk=π/a ({E|a}) = exp(−i πa a)I d = −I d ; character of which
is equal to −d ( d is the dimensionality of the representation). From Table 16 we
can see the proper irreducible representation at k = π/a is given by Γ5 alone.
e0
ē0
c0
c̄0
m0
m̄0
g0
ḡ 0
e0
e0
ē0
c0
c̄0
m0
m̄0
g0
ḡ 0
ē0
ē0
e0
c̄0
c0
m̄0
m0
ḡ 0
g0
c0
c0
c̄0
e0
ē0
g
ḡ 0
m0
m̄0
c̄0
c̄0
c0
ē0
e0
ḡ 0
g0
m̄0
m0
m0
m0
m̄0
ḡ 0
g0
ē0
e0
c0
c̄0
m̄0
m̄0
m0
g0
ḡ 0
e0
ē0
c̄0
c0
g0
g0
ḡ 0
m̄0
m0
c̄0
c0
e0
ē0
ḡ 0
ḡ 0
g0
m0
m̄0
c0
c̄0
ē0
e0
TABLE 15. The Multiplication table of G/I. The first column and the top row show the
representatives in G of the elements in G/I. The elements in the table are the products
between the element a in the first column and the element b in the top row, a · b.
The prediction by the character table can be verified. The electronic structure
calculation is executed to the mismatched double allay of Na atoms: The two atoms
60
Γ1
Γ2
Γ3
Γ4
Γ5
e0
1
1
1
1
2
ē0 [c0 ,c̄0 ] [m0 ,m̄0 ] [g’, ḡ 0 ]
1 1
1
1
1 1
-1
-1
1 -1
1
-1
1 -1
-1
-1
-2 0
0
0
TABLE 16. The character table. The top row shows the elements in the conjugacy classes.
The rows below shows the characters in the five irreducible representations.
are placed at ±(1/4ax +1/8ay ) in the unit cell with ax = (a, 0, 0), ay = (0, 4a, 0), az =
(0, 0, 5a) and a = 8.1au. This computation affirms that every pair of eigenvalues at
the boundary of the Brillouin zone shows the double-fold degeneracy.
k=0 -2.775 -2.124 -0.506 -0.479 0.139 0.753 0.780 1.054
k= π/a -0.717 -0.717 -0.072 -0.072 1.501 1.501 1.528 1.528
TABLE 17. Energy spectrum at k=0 and π/a in the mismatched double allay of Na atoms
(unit in eV).
The above example is a simple one-dimensional model. For a non-symmorphic
three-dimensional crystal, the irreducible representation for the k-point at the boundary of the Brillouin zone can be computed in a similar way. For this purpose we should
set up a supercell structure; its periodicity should be equal to that of the wavefunction at the boundary k-point in consideration; the symmetry operations to be taken
into account of must be confined to ones which fix the boundary k-point up to the periodicity of the old Brillouin zone of the small cell. (To consider the full symmetry of
the new super-cell may be superfluous); the irreducible representation at the Γ point
in the supercell is the mixture of the wavefunctions at the Γ point and the k-point
in consideration, but we can distinguish them by the comparison of the character
of the pure fractional translation. The example of the three-dimensional case is the
irreducible representation of diamond structure in the face-centered cell at X point
61
kX = [0, 2π/a, 0](= 1/2G2 + 1/2G3 ). The necessary super-cell is the cubic unit cell.
We can check this: the primitive lattice vectors in the cubic cell is expressed by those
in the face-centered cell as â1 = a1 + a2 − a3 , â2 = a1 − a2 + a3 and â3 = a1 + a2 − a3 ;
so there are relations as â1 · kX = 0,â2 · kX = 0 and â1 · kX = 2π. These relations
indicate that the wavefunction at kX in the face-centered cell fits itself into the cubic
supercell, in the sense that the translations by the primitive lattice vector in the
cubic cell causes no phase shift in the wavefunction. The symmetry operations for
the small group at kX are constructed from those which fix kX (kX → kX ) and those
which invert kX (kX → −kX ), due to the periodicity in the wave-number space. In
contrast, the full symmetry operations in the cubic cell couples six wavefunctions at
[±2π/a, 0, 0],[0, ±2π/a, 0],and [0, 0, ±2π/a]. As the wave-vector kX cannot be fixed,
the corresponding irreducible representation for the full symmetry is not adequate
for the small group at kX . (Elsewhere in this article, the point group of the full
symmetry operations in the cubic unit cell is computed; from which the small group
at kX can be derived as a subgroup.)
E.
The semidirect product and the induced representation
In the previous section of this article, concerning the computation of the character
tables for the supercell structures, we have to remake the multiplication table of the
extended group by the presence of the fractional translation. However, if we make use
of the mathematical idea of the semidirect product, we can treat such a problem more
effectively. In this section, this concept is explained and the possible computation
GAP is presented.
The semidirect product A n G is obtained by extending a group G by an another
group A in the following way. The elements in the new group are given as the
62
Cartesian product of each group:
{(a, g); a ∈ A, g ∈ G}.
(46)
The multiplication is defined as
(a1 , g1 ) · (a2 , g2 ) = (a1 g1 (a2 ), g1 g2 ),
(47)
in which the notation g1 (a2 ) means the automorphism in A acting on a2 , induced by
g1 in G. The typical example of this is the affine mapping in crystallographic group.
The definition is given as follows:
OT = {RT |τT } : r → RT · r + τT (Affine mapping),
(48)
OT · OS = {RT · RS |τT + RT · τS }.
(49)
If the group A is Abelian, the representation of the semidirect product is given by
χV (a, g) =
1
|Gx |
X
x(h(a))χU (hgh−1 ).
(50)
h∈G;hgh−1 ∈Gx
Here Gx is the set of elements in G, such that the character x in A is fixed by the
automorphism induced by elements in Gx . All of the irreducible representations of the
semidirect product are exhausted by this formula through possible combinations of
the character x in A and the irreducible representation χU of Gx .(For the theoretical
background, consult references of pure mathematics[8, 35].)
One of the simplest example is C3v = Group((1, 2, 3), (1, 2)) ≡ S3 . One can check
the validity of this formula, taking C3 = Group((1, 2, 3)) as A and I = Group((1, 2))
63
as G. Let us follow this computation.
The character tables of those three groups are given as
[()] [(1,2),(2,3),(3,1)] [(1,2,3),(1,3,2)]
χ1 1
1
1
χ2 1
-1
1
χ3 2
0
1
TABLE 18. The character table of C3v .
[eA =()] [aA =(1,2,3)] [a−1
A =(1,3,2)]
x1
1
1
1
1
ω
ω2
x2
2
x3
1
ω
ω
TABLE 19. The character table of C3 , where ω = exp(2π/3).
y1
y2
[eG =()] [gG =(1,2)]
1
1
1
-1
TABLE 20. The character table of I.
The automorphism on A by G is defined as:
eG (eA ) = eA ,
eG (aA ) = aA ,
−1
eG (a−1
A ) = aA ,
gG (eA ) = eA ,
gG (aA ) = a−1
A ,
gG (a−1
A ) = aA .
64
(51)
As a trivial case, take x1 . Since Gx1 = G, y1 can be chosen as χU . Since G is
Abelian, i.e. hgh−1 = g, the formula is
χV (a, g) =
=
1
|Gx1 |
1
2
X
x1 (h(a))y1 (hgh−1 )
(52)
h∈G;hgh−1 ∈G
X
h∈G;hgh−1 ∈Gy1
1
x1 (h(a))y1 (g) = (1 + 1) = 1,
2
for all elements in the semidirect product.
Again take x1 . Since Gx1 = G,y2 can also be chosen as χU . The formula is
χV (a, g) =
1
2
X
x1 (h(a))y2 (hgh−1 ).
(53)
h∈G;hgh(−1) ∈Gx1
It gives
1
χV (eA , eG ) = (1 + 1) = 1,
2
(54)
1
χV (aA , eG ) = χV (a−1
A , eG ) = (1 + 1) = 1,
2
1
χV (eA , gG ) = (−1 − 1) = −1,
2
1
χV (aA , gG ) = χV (a−1
A , gG ) = (−1 − 1) = −1.
2
Take x2 . Since Gx2 = hIdi (the trivial group of the identity), it must be χU = 1.
The formula is
χV (a, g) =
X
x2 (h(a))χU (hgh−1 )
h∈G;hgh−1 ∈G
=
x2
X
x2 (h(a)) · 1.
h∈G;hgh−1 ∈G
65
x2
(55)
It gives
χV (eA , eG ) = (1 + 1) = 2,
(56)
2
χV (aA , eG ) = χV (a−1
A , eG ) = (ω + ω ) = −1,
χV (eA , gG ) = 0,
χV (aA , gG ) = χV (a−1
A , gG ) = 0.
The last two characters are zero, since for any h, hgh−1 = g ∈
/ hIdi, no terms are
summed up.
(N.B. When we take x1 , it seems plausible that we could also choose Gx1 =
hIdi.Then the formula is
X
χV (a, g) =
x1 (h(a))χU (hgh−1 )
h∈G;hgh−1 ∈G
x1
X
=
(57)
x1 (h(a)) · 1.
h∈G;hgh−1 ∈<Id>
It gives
χV (eA , eG ) = 2,
(58)
χV (aA , eG ) = χV (a−1
A , eG ) = 2,
χV (eA , gG ) = 0,
χV (aA , gG ) = χV (a−1
A , gG ) = 0.
However, this result is not irreducible. It is essential to use the maximal set of Gx .)
There are other combinations of characters including x3 of C3v . However, they
only produce the redundant results.
The aforementioned computations are tedious and complicated; a naive imple66
mentation by means of GAP programming language could be given as in appendix
G.
In GAP, the semidirect product can be constructed by a built-in function.
gap> A=Group((1,2,3));
Alt( [ 1 .. 3 ] )
gap> G=Group((1,2));
Sym( [ 1 .. 2 ] )
The automorphism in A is evaluated as follows:
gap> au:=AutomorphismGroup(A);
<group of size 2 with 1 generators>
gap> auE:=Elements(au);
[ [ (1,2,3) ] -> [ (1,2,3) ], [ (1,2,3) ] -> [ (1,3,2) ] ]
gap> gens:=GeneratorsOfGroup(G);
[ (1,2) ]
The automorphism induced by G is defined as follows: (The generator (1,2) in G
causes the automorphism in A as (1,2,3)->(1,3,2). Indeed (1,2,3)*(1,3,2)=1: this is
the inversion.)
gap> hom:=GroupHomomorphismByImages(G,au,gens,[auE[2]]);
[ (1,2) ] -> [ [ (1,2,3) ] -> [ (1,3,2) ] ]
gap> p:=SemidirectProduct(G,hom,A);
Group([ (3,4,5), (1,2)(4,5) ])
gap> Irr(p);
[ Character( CharacterTable( Group([ (3,4,5), (1,2)(4,5) ]) ),
[ 1, 1, 1 ] ), Character( CharacterTable( Group([ (3,4,5), (1,2)
(4,5) ]) ), [ 1, 1, -1 ] ), Character( CharacterTable( Group([ (3,
4,5), (1,2)(4,5) ]) ), [ 2, -1, 0 ] ) ]
67
F.
The computation of the character table in the cubic diamond by GAP
In this section, the character table in the cubic diamond structure is computed by
GAP command for the semidirect product. The point group G of the basic diamond
structure is extended by the Abelian group of fractional translations N.
The point group G is given in the table in the appendix of this article, as the set of
48 three-dimensional matrices.(Although in the table accompanying fractional transformations are listed, the point group G is constructed solely from three-dimensional
rotation matrices; the fractional translations are omitted in the point group construction; they are included in the Abelian group N.) The group N is the factor group
F/T, where generators of F are
τ1 := (1/4, 1/4, 1/4), τ2 := (1/4, −1/4, −1/4), τ3 := (−1/4, −1/4, 1/4)
(59)
and those of T are
(1, 0, 0), (0, 1, 0), (0, 0, 1).
(60)
Using these generators, the group A and T are given as the vector space (or module)
over integers.
For the convenience, the group F/T can be expressed as another factor group
A/B, where
A := Group((1, 2, 3, 4), (5, 6, 7, 8), (9, 10, 11, 12));
(61)
and
B : = Group((1, 2, 3, 4)∧ 2 ∗ (5, 6, 7, 8)∧ 2, (5, 6, 7, 8)∧ 2 ∗ (9, 10, 11, 12)∧ 2,
(62)
(9, 10, 11, 12)∧ 2 ∗ (1, 2, 3, 4)∧ 2)
( : = Group((1, 3)(2, 4)(5, 7)(6, 8), (5, 7)(6, 8)(9, 11)(10, 12), (1, 3)(2, 4)(9, 11)(10, 12));
68
This definition corresponds to the relations of the fractional translations in F/I;
4 ∗ τ̄i := (0, 0, 0) [Identity]
(63)
2 ∗ (τ̄1 + τ̄2 ) = 2 ∗ (τ̄2 + τ̄3 ) = 2 ∗ (τ̄3 + τ̄1 ) ≡ (0, 0, 0) [Identity]
(64)
and
The gap computation goes as follows:
gap> A:=Group((1,2,3,4),(5,6,7,8),(9,10,11,12));
Group([ (1,2,3,4), (5,6,7,8), (9,10,11,12) ])
gap> B:=Group((1,2,3,4)^2*(5,6,7,8)^2,(5,6,7,8)^2*(9,10,11,12)^2,
(9,10,11,12)^2*(1,2,3,4)^2);
Group([ (1,3)(2,4)(5,7)(6,8), (5,7)(6,8)(9,11)(10,12), (1,3)(2,4)
(9,11)(10,12) ])
gap> N:=FactorGroup(A,B);
Group([ f1, f2, f3, f4 ])
gap> NaturalHomomorphismByNormalSubgroup(A,B);
[ (1,2,3,4), (5,6,7,8), (9,10,11,12) ] -> [ f3, f2, f1 ]
The last command returns the homomorphism from A to another group whose
kernel is B: in other words, gives the mapping of A onto A/B. The tree generators
(1,2,3,4),(5,6,7,8),(9,10,11,12) in A are mapped to f3,f2,f1 in A/B.
gap> N.1;
f1
gap> N.2;
f2
gap> N.3;
69
f3
gap> N.4;
f4
There is a newcomer f4; we must identify it. For this purpose, it is enough to
compute the multiplication table of for generators.
gap> NV:=[N.1,N.2,N.3,N.4];
[ f1, f2, f3, f4 ]
gap> TransposedMat([NV])*[NV];
[ [ f4, f1*f2, f1*f3, f1*f4 ], [ f1*f2, f4, f2*f3, f2*f4 ],
[ f1*f3, f2*f3, f4, f3*f4 ],
[ f1*f4, f2*f4, f3*f4, <identity> of ... ] ]
We can see the relation f1*f1=f2*f2=f3*f3=f4, which means that the element f4
corresponds to elements (1,2,3,4)^2, (5,6,7,8)^2, (9,10,11,12)^2, viz. the
fractional transformation (1/2,1/2,1/2).
The size of the fractional group is given as
gap> Size(C);
16
Concerning this, the volume of the parallelepiped spanned by τi is
gap> AbsoluteValue(Determinant(A));
1/16
Thus, in the unit cubic cell, the translations by τi generates 16 discrete points, the
number of which is equal to the order of F/I.
70
It is sufficient to define the automorphisms in N induced by G only for the generators of G: the candidates of which are two matrices, those of No.36 and No.48 in
the list of symmetry operations.
0 0 1
0 0 −1
M36 := 1 0 0 , M48 := 0 −1 0
0 −1 0
1 0 0
Let us check the operations of these two matrices on τi .
gap> A:=[[1,1,1],[-1,-1,1],[-1,1,-1]]/4;
[ [ 1/4, 1/4, 1/4 ], [ -1/4, -1/4, 1/4 ], [ -1/4, 1/4, -1/4 ] ]
gap> A:=TransposedMat(A);
[ [ 1/4, -1/4, -1/4 ], [ 1/4, -1/4, 1/4 ], [ 1/4, 1/4, -1/4 ] ]
gap> M36:=[[0,0,1],[1,0,0],[0,-1,0]];
[ [ 0, 0, 1 ], [ 1, 0, 0 ], [ 0, -1, 0 ] ]
gap> M48:=[[0,0,-1],[0,-1,0],[1,0,0]];
[ [ 0, 0, -1 ], [ 0, -1, 0 ], [ 1, 0, 0 ] ]
The generators are set to be the transposed forms of above matrices.
gap> TM36:=TransposedMat(M36);
[ [ 0, 1, 0 ], [ 0, 0, -1 ], [ 1, 0, 0 ] ]
gap> TM48:=TransposedMat(M48);
[ [ 0, 0, 1 ], [ 0, -1, 0 ], [ -1, 0, 0 ] ]
gap> (A*TM36*A^-1);
[ [ 0, -1, 0 ], [ 0, 0, -1 ], [ -1, 0, 0 ] ]
gap> (A*TM48*A^-1);
[ [ 0, 1, 0 ], [ 0, 0, 1 ], [ -1, -1, -1 ] ]
71
(65)
The last two computations (such as A · M · A−1 ) give the transformations on the
row vectors (τi ) in A.
The operations by M36 and M48 are given as:
M 36 · τ1 = −τ2
(66)
M 36 · τ2 = −τ3
M 36 · τ3 = −τ1
M 48 · τ1 = τ2
M 48 · τ2 = τ3
M 48 · τ3 = −τ1 − τ2 − τ3
Remark. In the definition of GAP, the semidirect product of a group N with a group
G (acting on N via a homomorphism α from G into the automorphism group of N) is
α
the Cartesian product G × N with the multiplication (g, n)·(h, m) = (gh, nh m), since
in GAP all groups act from the right. On the other hand, the familiar style for the
physicists is {G|n} · {H|m} = {G · H|n + G ∗ m}, as in the crystallography; the point
group acts from the left. However, these two different definitions are equivalent. This
is based on this: the crystallography, the symmetry operation {M |t} is represented by
an augmented matrix, which can take two forms. The one is the matrix of the form
[ M 0 ]
[ t 1 ]
acting from the right on row vectors (x, 1). The another is
[ M t ]
[ 0 1 ]
acting from the left on column vectors (x, 1). These two definitions are convertible by
72
transposition and one should interpret the GAP computations by the former definition. The generators used here, (M36, M48), should be replaced with their transposed
forms.
We can identify τi to three permutations:
τ1 := (9, 10, 11, 12) ≡ f1
(67)
τ2 := (5, 6, 7, 8) ≡ f2
τ3 := (1, 2, 3, 4) ≡ f3
Other identifications will do. But we use the identification consistent with the
result of the “NaturalHomomorphismByNormalSubgroup” command.
It suffices that the automorphism in N induced by G is defined at generators of
groups.
gap> gens:=GeneratorsOfGroup(N);;
gap> CC1:=[gens[2]^-1,gens[3]^-1,gens[1]^-1,gens[4]];;
gap> CC2:=[gens[2],gens[3],(gens[1]*gens[2]*gens[3])^-1,gens[4]];;
gap> hom1:=GroupHomomorphismByImages(N,N,gens,CC1);;
gap> hom2:=GroupHomomorphismByImages(N,N,gens,CC2);;
The command “GroupHomomorphismByImage(G, H, gens, imgs)” defines a homomorphism from the group G to the group H through the mapping of the list “gens”
of the generators of G to the list “imgs” of the images in H. If the definition of the
mapping is irrelevant, the program returns a “fail” message.
gap> CC3:=[gens[3]^-1,gens[1]*gens[3]^-1,gens[2]*gens[3]^-1];;
gap> hom3:=GroupHomomorphismByImages(N,N,[gens[1],gens[2],gens[3]],CC3);
fail
73
Let G be the point group of the minimal unit cell, generated from the matrices
TM36 and TM48. By those preparations, we can make the semidirect product GnN .
gap> au:=AutomorphismGroup(N);;
gap> G:=Group(TM36,TM48);;
gap> GTON:=GroupHomomorphismByImages(G,au,GeneratorsOfGroup(G),[hom1,hom2]);;
gap> GPN:=SemidirectProduct(G,GTON,N);;
gap> irrep:=Irr(SemidirectProduct(G,GTON,N));;
Once the semidirect product G n N is computed, the representation of the point
group in the cubic diamond can be deduced from a suitable subgroup of this. For
this purpose let us see the embedding of G and N into the semidirect product. The
command “Embedding” creates the embedding from G and N into the semidirect
product, and the command “Projection” creates the projection from the semidirect
product to G.
gap> embd1:=Embedding(GPN,1);
CompositionMapping( [ f1, f2, f3, f4, f5 ] -> [ f1, f2, f3, f4, f5 ],
CompositionMapping( Pcgs([ (5,8)(6,7), (1,2)(3,5,7,4,6,8), (3,7,6)(4,8,5),
(1,6)(2,5)(3,7)(4,8), (1,3)(2,4)(5,8)(6,7) ]) -> [ f1, f2, f3, f4, f5 ],
<action isomorphism> ) )
gap> embd2:=Embedding(GPN,2);
[ f1, f2, f3, f4 ] -> [ f6, f7, f8, f9 ]
gap> GeneratorsOfGroup(G);
[ [ [ 0, 1, 0 ], [ 0, 0, -1 ], [ 1, 0, 0 ] ],
[ [ 0, 0, 1 ], [ 0, -1, 0 ], [ -1, 0, 0 ] ] ]
gap> List(GeneratorsOfGroup(G),g->Image(embd1,g));
[ f2, f1*f3*f5 ]
gap> GeneratorsOfGroup(N);
74
[ f1, f2, f3, f4 ]
gap> List(GeneratorsOfGroup(N),g->Image(embd2,g));
[ f6, f7, f8, f9 ]
gap> Projection(GPN);
[ f1, f2, f3, f4, f5, f6, f7, f8, f9 ] ->
[ [ [ 1, 0, 0 ], [ 0, 0, 1 ], [ 0, 1, 0 ] ],
[ [ 0, 1, 0 ], [ 0, 0, -1 ], [ 1, 0, 0 ] ],
[ [ 0, 0, -1 ], [ -1, 0, 0 ], [ 0, 1, 0 ] ],
[ [ -1, 0, 0 ], [ 0, 1, 0 ], [ 0, 0, -1 ] ],
[ [ 1, 0, 0 ], [ 0, -1, 0 ], [ 0, 0, -1 ] ],
[ [ 1, 0, 0 ], [ 0, 1, 0 ], [ 0, 0, 1 ] ],
[ [ 1, 0, 0 ], [ 0, 1, 0 ], [ 0, 0, 1 ] ],
[ [ 1, 0, 0 ], [ 0, 1, 0 ], [ 0, 0, 1 ] ],
[ [ 1, 0, 0 ], [ 0, 1, 0 ], [ 0, 0, 1 ] ] ]
The correspondence of the generators is given in table21.
Elements in G and N
Elements in G n N
TM36(in G)
f2
TM48(in G)
f1*f3*f5
f1(in N) (The image of (9,10,11,12))
f6
f2(in N) (The image of (5,6,7,8))
f7
f3(in N) (The image of (1,2,3,4))
f8
f4(in N) (The image of (1,3)(2,4) or (5,7)(6,8) or (1,3)(2,4))
f9
TABLE 21. The correspondence of the generators in groups G, N and G n N .
The generators of the point group of the cubic diamond structure are these elements:
{m36|τi }
{m48|0}
75
(68)
(τi can be any of three alternatives. The addition of pure fractional translations
{E|τi } is superfluous in this case.)
The corresponding subgroup can be constructed in this way:
Subgroup(GPN,[GPN.2*GPN.6,GPN.1*GPN.3*GPN.5])
or
Subgroup(GPN,[GPN.2*GPN.7,GPN.1*GPN.3*GPN.5])
or
Subgroup(GPN,[GPN.2*GPN.8,GPN.1*GPN.3*GPN.5])
These three definitions generate the same group so that any choice can be used.
(They are the subgroups of GPN, generated by the elements in the square brackets
of the second argument.)
G.
1.
Crystal lattice and Cohomology
A brief introduction to group cohomology
In the modern style, the crystal group can be described by the theory of the
cohomology, which is actually computable by the computer algebra.
At first, we construct the one-cohomology. Let K ∈ GL(n, Z) and define V :=
Rn /Zn . The one-cocycle is defined as
Z 1 (K, V ) := {ρ; k → V |ρ(1) = 0, ρ(gh) = ρ(g) + gρ(h) for all g, h ∈ K}.
(69)
The one-coboundary is defined as
B 1 (K, V ) := {ρ ∈ Z 1 (K, V |ρ(g) = gv − v for some v ∈ V }.
76
(70)
The one-cohomology is defined as their quotient:
H 1 (K, V ) := Z 1 (K, V )/B 1 (K, V )
(71)
We can see the vector system in the space group, which satisfies the condition
τ (gh) ≡ τ (g) + g · τ (h) mod Zn (mod lattice translations)
(72)
is a kind of cocycle. Furthermore, the one-coboundary is interpreted as the shift of
origin; we can observe this from the conjugation operation:
{1|s}−1 {g|t}{1|s} = {g|(gs − s) + t}
(73)
Thus, the vector system could be understood in the context of one-cohomology.
The above definition of the one-cohomology is derived from the general theory of
group cohomology. Let G be a finite group and M be a module. Let C n (G, M ) be
the set functions from Gn = (g1 , g2 , . . . , gn ) to M . The differential (dn : C n (G, M ) →
C n+1 (G, M ) is defined as
dn (f )(g0 , g1 , . . . , gn ) = g0 · f (g1 , . . . , gn )
n
X
+
(−1)j f (g0 , . . . , gj−2 , gj−1 gj , gj+1 , . . . , gn )
(74)
j
+(−1)n+1 f (g0 , g1 , . . . , gn−1 ).
It satisfies
dn ◦ dn−1 = 0.
(75)
The n-cocycle Z i (G, M ) is defined as Ker(dn ), and the n-coboundary is as Im(dn−1 ).
The n-th cohomology group is defined as H i (G, M ) := Z i (G, M )/B i (G, M ).
77
2.
The possible fractional translations with one point group
We make use of following matrices (as explained elsewhere in this article) in order
to define the crystallographic group in the diamond crystal lattice:
gap>M1:=[[0,0,1,1/4],[1,0,0,1/4],[0,-1,0,1/4],[0,0,0,1]];
gap>M2:=[[0,0,-1,0],[0,-1,0,0],[1,0,0,0],[0,0,0,1]];
gap>S:=AffineCrystGroupOnLeft([M1,M2]);
gap>P:=PointGroup(S);
In this definition, the fractional translation is given a priori. With regard to
this, there is a question: what is the possible fractional translation compatible to
the point group of the diamond crystal? Is it uniquely determined? Let us inspect
this problem. For example, take the face-centered lattice. For simplicity in the
following computation, the symmetry operations are represented by the “lattice axes
coordinate”, by means of the vector system of the primitive lattice vectors. For this
new coordinate, the new point group Q is made from the group P by the conjugation
of the generators. (The symbol ^ is used for the conjugation in GAP.)
gap> A:=[[0,1/2,1/2],[1/2,0,1/2],[1/2,1/2,0]];
gap> Q:=P^A;
gap> Q.1;Q.2;
[ [ -1, 0, 0 ], [ 0, 0, -1 ], [ 1, 1, 1 ] ]
[ [ 0, 1, 0 ], [ 0, 0, 1 ], [ -1, -1, -1 ] ]
The generators of the space groups are defined now by means of unknown variables.(For the heuristic purpose, we venture to employ generators different to the
first definition.)
a:=Indeterminate(Rationals,1);
78
b:=Indeterminate(Rationals,2);
c:=Indeterminate(Rationals,3);
N1:=[[-1,0,0,0],[0,0,-1,0],[1,1,1,0],[0,0,0,1]];
N2:=[[0,1,0,a],[0,0,1,b],[-1,-1,-1,c],[0,0,0,1]];
Let us consider multiplications of two elements gi and gj in the point group, which
yields the identity. If gi and gj are represented by the products of the generators, the
multiplication gives a relation of the generators both in the point group and in the
space group. For example, take (Q.2*Q.1)^4=I. This relation is lifted up to the space
group as (N2*N1)^4≡ {I|T }, where T is a translation. The linear part of (N2*N1)^4
is the identity matrix; and the translation part [2*a+2*b,2*a+2*b,-2*a-2*b] must
be one of translation vectors. In general, the translation part in the relation of the
generators in the space group gives three linear equations with respect to unknown
variables “a”,“b”, and “c”. These equations should be equivalent to zero, modulo
translation vectors; viz. they should be integer-valued.
The relations in generators in Q shall be set up now. For this purpose, each elements in the group and its inverse are represented by generators respectively, so that
the product of them gives the relation in generators which yields the unity. In the
following, the symbols x1 and x2 indicate the two generators of the group Q. For simplicity, the trivial relation only representing the unity ( returned as <identity ...>
in the GAP result) is eliminated in the list.
gap> MQ:=Elements(Q);;
gap> rel:=List(MQ,i->Factorization(Q,i)*Factorization(Q,i^-1));;
gap> IDENTITY:=Factorization(Q,One(Q));;
gap> rel:=Filtered(rel,i->i<>IDENTITY);
[ (x2^-1*x1*x2*x1)^2, x1*(x2*x1^-1)^2*x1^-1, (x2*x1^2)^2,
(x2^-1*x1)^2, x1*x2^2*x1*x2^-1*x1*x2*x1^-1, x1^6,
79
(x2*x1*x2^2)^2, (x2^2*x1*x2)^2, (x2*x1*x2^2*x1^2)^2,
(x2^-1*x1)^2, (x1*x2^2*x1^2)^2, x2^4, x2*x1^3*x2^-1*x1^3,
x1*x2^4*x1^-1, x1^-1*(x2^-1*x1)^2*x1, x2^-1*x1*x2^3*x1,
(x2^2*x1^3)^2, (x2*x1^-1)^2, x2^4,
x2^-1*x1^-1*(x2^-1*x1)^2*x1*x2, (x1*x2*x1)^2, (x2^-1*x1)^2,
x1^-1*x2^-1*x1*x2^3*x1^2, x2^-1*x1*x2^3*x1,
x1^-1*(x2^-1*x1)^2*x1, x2^2*x1*x2^-1*x1*x2, (x2*x1)^4,
(x1^2*x2)^2, (x2*(x2*x1)^2)^2, (x2*x1*x2^2*x1)^2,
(x1*x2^2*x1*x2)^2, x2^-1*x1^3*x2*x1^3, x2^2*x1*x2^-1*x1*x2,
(x2*x1^-1)^2, (x2^-1*x1^-1*x2^-1*x1^2)^2 ]
The relations of the point group generators are lifted to the space group. The
matrix relations are stored in the list “relMat”; and the fractional translations are
taken out in the list “relv”. Now, if the fractional translation [a,b,c] is admissible,
the list “relv” must be integer-valued by the substitutions to the indeterminates.
(Apropos of this, we can use symbols of unknowns as “a”,“b”,“c” for the input,
whereas the GAP results are represented by “x_1”,“x_2”,“x_3”. Be careful not to
confuse indeterminates and generators of the group.) The GAP program cannot
invert the symbolically represented matrices N1 and N2, so we must prepare the
inverses for them (N1I and N2I) by the power products.
N1I:=N1^5;
N2I:=N2^3;
relMat:=[ (N2I^1*N1*N2*N1)^2, N1*(N2*N1I^1)^2*N1I^1, (N2*N1^2)^2,
(N2I^1*N1)^2, N1*N2^2*N1*N2I^1*N1*N2*N1I^1, N1^6,
(N2*N1*N2^2)^2, (N2^2*N1*N2)^2, (N2*N1*N2^2*N1^2)^2, (N2I^1*N1)^2,
(N1*N2^2*N1^2)^2, N2^4, N2*N1^3*N2I^1*N1^3,
N1*N2^4*N1I^1, N1I^1*(N2I^1*N1)^2*N1, N2I^1*N1*N2^3*N1,
80
(N2^2*N1^3)^2, (N2*N1I^1)^2, N2^4,
N2I^1*N1I^1*(N2I^1*N1)^2*N1*N2, (N1*N2*N1)^2, (N2I^1*N1)^2,
N1I^1*N2I^1*N1*N2^3*N1^2, N2I^1*N1*N2^3*N1,
N1I^1*(N2I^1*N1)^2*N1, N2^2*N1*N2I^1*N1*N2, (N2*N1)^4,
(N1^2*N2)^2, (N2*(N2*N1)^2)^2, (N2*N1*N2^2*N1)^2,
(N1*N2^2*N1*N2)^2, N2I^1*N1^3*N2*N1^3, N2^2*N1*N2I^1*N1*N2,
(N2*N1I^1)^2, (N2I^1*N1I^1*N2I^1*N1^2)^2 ];
relv:=List(relMat,i->[i[1][4],i[2][4],i[3][4]]);
[ [ 2*x_1+x_2+x_3, -2*x_1-x_2-x_3, -2*x_1-x_2-x_3 ],
[ -2*x_1-x_2-x_3, 0, 2*x_1+x_2+x_3 ], [ -x_2-x_3, 2*x_2, 2*x_3 ],
[ 2*x_1+x_2+x_3, -2*x_1-x_2-x_3, 0 ], [ 0, 2*x_1+x_2+x_3, 0 ],
[ 0, 0, 0 ], [ 0, 0, 2*x_1+x_2+x_3 ],
[ 0, 2*x_1+x_2+x_3, -2*x_1-x_2-x_3 ], [ 0, 2*x_1+4*x_2+2*x_3, 0 ],
[ 2*x_1+x_2+x_3, -2*x_1-x_2-x_3, 0 ],
[ -2*x_1-2*x_2, 2*x_1+2*x_2, 2*x_2+2*x_3 ], [ 0, 0, 0 ],
[ 2*x_1, 2*x_2, 2*x_3 ], [ 0, 0, 0 ],
[ -2*x_1-x_2-x_3, 0, 2*x_1+x_2+x_3 ],
[ 2*x_1+x_2+x_3, -2*x_1-x_2-x_3, 0 ],
[ 2*x_1+2*x_2, 2*x_2+2*x_3, -2*x_1-2*x_2 ],
[ 2*x_1+x_2+x_3, 0, 0 ], [ 0, 0, 0 ], [ 0, -2*x_1-x_2-x_3, 0 ],
[ x_2+x_3, -2*x_3, x_2+x_3 ], [ 2*x_1+x_2+x_3, -2*x_1-x_2-x_3, 0 ]
, [ -2*x_1-x_2-x_3, 0, 2*x_1+x_2+x_3 ],
[ 2*x_1+x_2+x_3, -2*x_1-x_2-x_3, 0 ],
[ -2*x_1-x_2-x_3, 0, 2*x_1+x_2+x_3 ],
[ 0, 2*x_1+x_2+x_3, -2*x_1-x_2-x_3 ],
[ 2*x_1+2*x_2, 2*x_1+2*x_2, -2*x_1-2*x_2 ],
[ -x_2-x_3, -x_2-x_3, 2*x_2 ],
81
[ 4*x_1+4*x_2+2*x_3, -x_2-x_3, -4*x_1-2*x_2 ],
[ 4*x_1+2*x_2, x_2+x_3, x_2+x_3 ],
[ -4*x_1-4*x_2-2*x_3, 4*x_1+2*x_2, x_2+x_3 ],
[ 2*x_1+2*x_2+2*x_3, -2*x_1, -2*x_2 ],
[ 0, 2*x_1+x_2+x_3, -2*x_1-x_2-x_3 ], [ 2*x_1+x_2+x_3, 0, 0 ],
[ 2*x_2+2*x_3, -2*x_1-2*x_3, 2*x_1+2*x_3 ] ]
Since there are duplicates in the linear equations of the fractional translation, the
unique ones should be extracted by
gap> Unique(Flat(relv*a^0));
[ 2*x_1+x_2+x_3, -2*x_1-x_2-x_3, 0, -x_2-x_3, 2*x_2, 2*x_3,
2*x_1+4*x_2+2*x_3, -2*x_1-2*x_2, 2*x_1+2*x_2, 2*x_2+2*x_3, 2*x_1,
x_2+x_3, -2*x_3, 4*x_1+4*x_2+2*x_3, -4*x_1-2*x_2, 4*x_1+2*x_2,
-4*x_1-4*x_2-2*x_3, 2*x_1+2*x_2+2*x_3, -2*x_1, -2*x_2,
-2*x_1-2*x_3, 2*x_1+2*x_3 ]
It is necessary to use an input such as “relv*a^0”, so that all entries in the list
should be polynomials, for the convenience of sorting, by dint of the fact a^0 is the
unity as a polynomial. Now we have following four equations to be integer-valued:
x_2+x_3
2*x_1
2*x_2
2*x_3
In Q/Z, there are four representatives of the fractional translations, as are given
below, with the primitive lattice vectors accompanied to them. (The translation
vectors are computed by the substitution in the list “relv”, from which the basis
vectors are extracted by the GAP built-in command “BaseIntMat(mat)”.)
82
Fractional translation
[ 0, 1/2, 1/2 ]
[ [ 1,
[ [ 1,
[ 1/2, 0, 0 ]
[ 1/2, 1/2, 1/2 ]
[ [ 1,
Translation vectors
0, 0 ], [ 0, 1, 0 ], [ 0, 0, 1 ] ]
0, 0 ], [ 0, 1, 0 ], [ 0, 0, 1 ] ]
1, 1 ], [ 0, 2, 0 ], [ 0, 0, 2 ] ]
TABLE 22. The fractional translations and the translation vectors.
To inspect the validity, by means of GAP built-in command, the possible space
groups with a given point group can be computed. In order to check the computed
fractional translations, it is enough to see the generators of the space group and the
primitive translation vectors.
gap> SS:=SpaceGroupsByPointGroupOnLeft(Q);;
gap> Display(List(SS,GeneratorsOfGroup));
[ [ [ [ -1, 0, 0, 0 ], [ 0, 0, -1, 0 ], [ 1, 1, 1, 0 ], [ 0, 0, 0, 1 ] ],
[ [ 0, 1, 0, 0 ], [ 0, 0, 1, 0 ], [ -1, -1, -1, 0 ], [ 0, 0, 0, 1 ] ],
[ [ 1, 0, 0, 1 ], [ 0, 1, 0, 0 ], [ 0, 0, 1, 0 ], [ 0, 0, 0, 1 ] ],
[ [ 1, 0, 0, 0 ], [ 0, 1, 0, 1 ], [ 0, 0, 1, 0 ], [ 0, 0, 0, 1 ] ],
[ [ 1, 0, 0, 0 ], [ 0, 1, 0, 0 ], [ 0, 0, 1, 1 ], [ 0, 0, 0, 1 ] ] ],
[ [ [ -1, 0, 0, 0 ], [ 0, 0, -1, 0 ], [ 1, 1, 1, 0 ], [ 0, 0, 0, 1 ] ],
[ [ 0, 1, 0, 0 ], [ 0, 0, 1, 1/2 ], [ -1, -1, -1, 1/2 ], [ 0, 0, 0, 1 ] ],
[ [ 1, 0, 0, 1 ], [ 0, 1, 0, 0 ], [ 0, 0, 1, 0 ], [ 0, 0, 0, 1 ] ],
[ [ 1, 0, 0, 0 ], [ 0, 1, 0, 1 ], [ 0, 0, 1, 0 ], [ 0, 0, 0, 1 ] ],
[ [ 1, 0, 0, 0 ], [ 0, 1, 0, 0 ], [ 0, 0, 1, 1 ], [ 0, 0, 0, 1 ] ] ],
[ [ [ -1, 0, 0, 0 ], [ 0, 0, -1, 0 ], [ 1, 1, 1, 0 ], [ 0, 0, 0, 1 ] ],
[ [ 0, 1, 0, 1/2 ], [ 0, 0, 1, 0 ], [ -1, -1, -1, 0 ], [ 0, 0, 0, 1 ] ],
[ [ 1, 0, 0, 1 ], [ 0, 1, 0, 0 ], [ 0, 0, 1, 0 ], [ 0, 0, 0, 1 ] ],
[ [ 1, 0, 0, 0 ], [ 0, 1, 0, 1 ], [ 0, 0, 1, 0 ], [ 0, 0, 0, 1 ] ],
[ [ 1, 0, 0, 0 ], [ 0, 1, 0, 0 ], [ 0, 0, 1, 1 ], [ 0, 0, 0, 1 ] ] ],
[ [ [ -1, 0, 0, 0 ], [ 0, 0, -1, 0 ], [ 1, 1, 1, 0 ], [ 0, 0, 0, 1 ] ],
83
[ [ 0, 1, 0, 1/2 ], [ 0, 0, 1, 1/2 ], [ -1, -1, -1, 1/2 ], [ 0, 0, 0, 1 ] ],
[ [ 1, 0, 0, 1 ], [ 0, 1, 0, 0 ], [ 0, 0, 1, 0 ], [ 0, 0, 0, 1 ] ],
[ [ 1, 0, 0, 0 ], [ 0, 1, 0, 1 ], [ 0, 0, 1, 0 ], [ 0, 0, 0, 1 ] ],
[ [ 1, 0, 0, 0 ], [ 0, 1, 0, 0 ], [ 0, 0, 1, 1 ], [ 0, 0, 0, 1 ] ] ] ]
gap> Display(List(SS,TranslationBasis));
[ [ [ 1, 0, 0 ], [ 0, 1, 0 ], [ 0, 0, 1 ] ],
[ [ 1, 0, 0 ], [ 0, 1, 0 ], [ 0, 0, 1 ] ],
[ [ 1, 0, 0 ], [ 0, 1, 0 ], [ 0, 0, 1 ] ],
[ [ 1, 0, 0 ], [ 0, 1, 0 ], [ 0, 0, 1 ] ] ]
There are four space groups with the point group Q. The last three of the GAP
result is the same fractional translation in Table 22. However, their translational
vectors are different: the latter has larger translational vectors. This is because the
GAP computation implicitly includes the primitive translations of the unit cube in
generators of the space groups.
Readers may question about the relationship between the fractional translation
computed now and those of the definition of the space group given in the first part of
this section. The generators defined earlier are rewritten in the lattice axes coordinate
as:
gap> MM1:=[[-1,0,0,1/4],[0,0,-1,1/4],[1,1,1,1/4],[0,0,0,1]];
gap> MM2:=[[0,1,0,0],[0,0,1,0],[-1,-1,-1,0],[0,0,0,1]];
gap> S2:=AffineCrystGroupOnLeft([MM1,MM2]);
We can apply the shift of the origin to this space group by means of the conjugation
{E| − s}{R|t}{E|s} = {R|t + (R − I) · s}. If the matrix R − I is normal, we can annihilate the fractional translation. Let us apply such a conjugation to the generators.
(The command M^T is the abbreviation of the conjugation T−1 · M · T.)
84
gap> T:=[[1,0,0,1/8],[0,1,0,-3/8],[0,0,1,5/8],[0,0,0,1]];
gap> MM1^T;
[ [ -1, 0, 0, 0 ], [ 0, 0, -1, 0 ], [ 1, 1, 1, 0 ], [ 0, 0, 0, 1 ] ]
gap> MM2^T;
[ [ 0, 1, 0, -1/2 ], [ 0, 0, 1, 1 ], [ -1, -1, -1, -1 ], [ 0, 0, 0, 1 ] ]
Thus we can see that the shift of the origin generates the space group, which is equivalent to that composed by matrices N1 and N2 with [a,b,c]=[ -1/2, 1, -1 ]∼
=[1/2,0,0].
Incidentally, let us generate other representation of generators:
N1:=[[-1,0,0,a],[0,0,-1,b],[1,1,1,c],[0,0,0,1]];
N2:=[[0,1,0,0],[0,0,1,0],[-1,-1,-1,0],[0,0,0,1]];
The linear equations for the fractional translation are given as
[ 0, 4*x_3, -2*x_2-2*x_3, -2*x_2+2*x_3, -x_2+x_3, x_2-x_3,
2*x_2+2*x_3, -4*x_1-4*x_2-4*x_3, -2*x_1-2*x_3, -2*x_1+2*x_3,
-2*x_1-4*x_2-2*x_3, -2*x_1-2*x_2, -2*x_1-2*x_2-4*x_3,
-2*x_1-x_2-x_3, -2*x_1+2*x_2, 2*x_1-2*x_3, 2*x_1+2*x_3,
2*x_1-2*x_2, 2*x_1+x_2+x_3, 2*x_1+2*x_2+4*x_3,
2*x_1+4*x_2+2*x_3, 4*x_1+4*x_2+4*x_3 ]
From them, seven possible vectors of [a,b,c] (=[x_1,x_2,x_3]) are obtained as
[ 0, 1/2, 1/2 ], [ 1/4, 1/4, 1/4 ], [ 1/4, 3/4, 3/4 ],
[ 1/2, 0, 0 ], [ 1/2, 1/2, 1/2 ], [ 3/4, 1/4, 1/4 ],
[ 3/4, 3/4, 3/4 ] ]
We can see one of the fractional translations is [1/4,1/4,1/4].
Whereas the GAP computation goes as
85
gap> Q2:=Group(Q.1,Q.2);
Group([ [ [ -1, 0, 0 ], [ 0, 0, -1 ], [ 1, 1, 1 ] ],
[ [ 0, 1, 0 ], [ 0, 0, 1 ], [ -1, -1, -1 ] ] ])
gap> Q2=Q;
true
gap> SS2:=SpaceGroupsByPointGroupOnLeft(Q2);
[ <matrix group with 5 generators>, <matrix group with 5 generators>
, <matrix group with 5 generators>,
<matrix group with 5 generators> ]
gap> List(SS2,GeneratorsOfGroup);
[ [ [ [ 0, 1, 0, 0 ], [ 0, 0, 1, 0 ], [ -1, -1, -1, 0 ], [ 0, 0, 0, 1 ] ],
[ [ -1, 0, 0, 0 ], [ 0, 0, -1, 0 ], [ 1, 1, 1, 0 ], [ 0, 0, 0, 1 ] ],
[ [ 1, 0, 0, 1 ], [ 0, 1, 0, 0 ], [ 0, 0, 1, 0 ], [ 0, 0, 0, 1 ] ],
[ [ 1, 0, 0, 0 ], [ 0, 1, 0, 1 ], [ 0, 0, 1, 0 ], [ 0, 0, 0, 1 ] ],
[ [ 1, 0, 0, 0 ], [ 0, 1, 0, 0 ], [ 0, 0, 1, 1 ], [ 0, 0, 0, 1 ] ] ],
[ [ [ 0, 1, 0, 0 ], [ 0, 0, 1, 0 ], [ -1, -1, -1, 0 ], [ 0, 0, 0, 1 ] ],
[ [ -1, 0, 0, 0 ], [ 0, 0, -1, 1/2 ], [ 1, 1, 1, 1/2 ], [ 0, 0, 0, 1 ] ],
[ [ 1, 0, 0, 1 ], [ 0, 1, 0, 0 ], [ 0, 0, 1, 0 ], [ 0, 0, 0, 1 ] ],
[ [ 1, 0, 0, 0 ], [ 0, 1, 0, 1 ], [ 0, 0, 1, 0 ], [ 0, 0, 0, 1 ] ],
[ [ 1, 0, 0, 0 ], [ 0, 1, 0, 0 ], [ 0, 0, 1, 1 ], [ 0, 0, 0, 1 ] ] ],
[ [ [ 0, 1, 0, 0 ], [ 0, 0, 1, 0 ], [ -1, -1, -1, 0 ], [ 0, 0, 0, 1 ] ],
[ [ -1, 0, 0, 1/4 ], [ 0, 0, -1, 1/4 ], [ 1, 1, 1, 1/4 ], [ 0, 0, 0, 1 ] ],
[ [ 1, 0, 0, 1 ], [ 0, 1, 0, 0 ], [ 0, 0, 1, 0 ], [ 0, 0, 0, 1 ] ],
[ [ 1, 0, 0, 0 ], [ 0, 1, 0, 1 ], [ 0, 0, 1, 0 ], [ 0, 0, 0, 1 ] ],
[ [ 1, 0, 0, 0 ], [ 0, 1, 0, 0 ], [ 0, 0, 1, 1 ], [ 0, 0, 0, 1 ] ] ],
[ [ [ 0, 1, 0, 0 ], [ 0, 0, 1, 0 ], [ -1, -1, -1, 0 ], [ 0, 0, 0, 1 ] ],
[ [ -1, 0, 0, 1/4 ], [ 0, 0, -1, 3/4 ], [ 1, 1, 1, 3/4 ], [ 0, 0, 0, 1 ] ],
86
[ [ 1, 0, 0, 1 ], [ 0, 1, 0, 0 ], [ 0, 0, 1, 0 ], [ 0, 0, 0, 1 ] ],
[ [ 1, 0, 0, 0 ], [ 0, 1, 0, 1 ], [ 0, 0, 1, 0 ], [ 0, 0, 0, 1 ] ],
[ [ 1, 0, 0, 0 ], [ 0, 1, 0, 0 ], [ 0, 0, 1, 1 ], [ 0, 0, 0, 1 ] ] ] ]
Four space groups are generated by GAP: non-zero fractional translations are three
in number:
[ 0, 1/2, 1/2 ], [ 1/4, 1/4, 1/4 ], [ 1/4, 3/4, 3/4 ]
Apparently the GAP result is lacking in some of fractional translations computed just
before by means of linear equations. The GAP result should rather be interpreted as
the basis vectors for the possible fractional translations. The other possible fractional
translations can be generated from this basis set.
It is well to remind readers of the point that the setting up of the admissible
fractional translations (which is called “vector system” in the mathematical context)
is nothing other than the computation of the one-cohomology. In the above example,
the fractional translation in the generator N1 is chosen to be zero from the outset.
It is possible to assume the non-zero translation in N1. However, a shift of the
origin can cancel a nonzero fractional translation as we have seen. The shift of the
origin could be caused by the conjugation, which corresponds to the choice of the
representative of the elements in the map of the one-cohomology.
3.
Group extension
The space group of the crystal is determined by the following short exact sequence
(or the group extension):
π
1→K→E→
− G→1
(76)
with the condition that there exists a normal subgroup of N in E which is isomorphic
to K and that E/K is isomorphic to G. The group E is called an extension of G by K.
87
The equivalence classes of extension of E are computed by a concept of the cohomological algebra, i.e. two-cohomology (denoted as H 2 (G, K) = Z 2 (G, K)/B 2 (G, K)).
In order to understand this concept, however, one should get acquainted with the
language of commutative algebra, especially concerning homological algebra, which
would necessitate long explanation. Thus if the reader would get interested in these
topics, he should try advanced mathematical textbooks. However, the viewpoint of
cohomological algebra might be useful in material designing, especially in studying
the algebraic aspects of crystallography. For example, the Bravais lattice in two- or
three-dimensional crystals can be classified by two-cohomology.
Remark. The construction of the two-cohomology is briefly stated here. For every
g ∈ G there are elements eg ∈ E with π(eg ) = g. However eg is not unique: another
element e0g behaves similarly, if e0g (eg )−1 ∈ ker(π) = K. Now we define a map
ϕg : k → eg ke−1
g , as an automorphism in K. The two-cocycle is defined by
f (g, h) = eg eh e−1
gh .
For each g, let us give an element kg ∈ K (as a homomorphism k(∗) from G to K).
Then the two-coboundary is defined as
−1
b(g, h) = kg ϕg (kh )kgh
.
(It will be represented as k(g) + g · k(h) − k(gh) when K is a module.)
Then we have
f (g, h)f (gh, k) = ϕg (f (h, k))f (g, hk)
and for another choice (e0g = kg eg ),
f 0 (g, h) = e0g e0h (e0gh )−1 = b(h, g)f (g, h).
88
In two-cohomology f (∗, ∗) and f 0 (∗, ∗) represent the same class. The elements of
E is uniquely written in the form aeg for some a ∈ K and g ∈ G. The product is
given as (aeg )(beh ) = aϕg (b)f (g, h)egh , which is the usual semidrect product when
f (g, h) = 1 for all g, h.
2
2
For simplicity, consider the two-dimensional
case.
hx, y|x2 =
Let P =
y = (xy) = 1i,
−1 0
1 0
and My =
. The action
which acts on T = Z2 through Mx =
0 1
0 −1
1
−1
x
y
y
and
on T is written as {x e1 = e−1
1 , e2 = e2 , e1 = e1 , e2 = e2 }. (e1 =
0
0
e2 = .
1
Let us extend this group, and, if possible, let us construct space groups which
have this group as their point groups. (This example is taken from the lecture notes
of Webb[36]. For theoretical details, consult with that article and references therein.)
There is a general procedure (Zassenhauss algorithm) to compute two-cohomology,
which is based on the following theorem.
Theorem 1. Let P be a finite group given by the representation
P = hx1 , x2 , ..., xd |r1 , r2 , ..., rt i .
Let n be the dimension of the space, and T be a module. From the relators ri , a
matrix Λ ∈ Mnt,nd (Z) is computed, and the elements Λij are defined as
ri − 1 =
d
X
Λij (xj − 1) (i = 1, ..., t).
j=1
89
Then the two-cohomology is given by
H 2 (P, T ) ∼
= {x ∈ (RT )d |Λ · x ∈ T t }/(Ker Λ + T d ).
(77)
This definition says that the representative elements in two-cohomology are given as
(v1 , · · · , vd ) , the concatenation of the vectors in RT n , (from which the component
in Ker Λ + T d should be factored out), so as to be
v
1
..
Λ . ∈ T t.
vd
(78)
The integer matrix Λ can be converted into its Smith normal form
b1
b2
..
.
bl
0
..
.
(79)
0
with bi 6= 0, dividing bj for i < j. (A matrix A ∈ Mm,n (Z) can be converted into the
diagonal form as this, by means matrix operation P · A · Q with P ∈ GL(m, Z) and
Q ∈ GL(n, Z).)
Then H 2 (P, T ) ∼
= Z/b1 Z ⊕ · · · ⊕ Z/bl Z. Now the relators in P are lifted in the
90
extension of P by T as
r1
r2
= Λ · v (∀v ∈ H 2 (P, T )).
..
.
rt
(80)
The vector systems are given as {xi |vi } (i = 1, ..., d) in the crystallography.
Remark. This algorithm is a refined style of the computation of one-cohomology
given in the previous section. The relation such as xyxy = 1( in G) is lifted to
exyxy = 1E , as the relation in E. The latter is rewritten by two-cocycles
1E = exyxy = f (x, yxy)−1 ϕx (f (y, xy)−1 )ϕx (ϕy (f (x, y)−1 ))ex ey ex ey
It is furthermore rewritten as
ex ey ex ey = ϕx (ϕy (f (x, y)))ϕx (f (y, xy))f (x, yxy).
The right side of this equation lies in K. The formula of Λ̂(xy)2 in the algorithm
represents the translational part in ex ey ex ey by means of those of ex and ey , and it
must be in K(= T ). To find a v such as Λ · v ∈ T is to fix a representative of cocycle
f (∗, ∗) in two-cohomology.
91
In the aforesaid question, the matrix Λ is computed as
2
2
Λx ,x Λx ,y
Λ = Λy2 ,x Λy2 ,y
Λ(xy)2 ,x Λ(xy)2 ,y
0 0 0 0
0
x
+
1
0
0
= 0
y+1 =
0
xy + 1 xyx + x
0
2 0 0
0 2 0
0 0 0
0 0 0
(81)
0 0 0 0
From this, it can be concluded that H 2 (P, T ) = Z/2Z⊕Z/2Z. The homomorphism
representing the two-cohomology, which determines the extension of the group, is
generated from these vectors:
0
0
1e
1/2 0
0
2
2
,
=
,
1
0 1/2
0
e
1
2
0
0
(82)
The linear combination of them (with integer coefficients, of 0 or 1 ) admits four
possible mappings. The
in the extension E are obtained from the entries in
relators
2
Λ̂x
Λ(c1 · e1 + c2 · e2 ) = Λ̂y2 . They are listed as: They generate four different
Λ̂(xy)2
Λ̂x2
Λ̂y2
Λ̂(xy)2
type 1
type 2
type 3
type 4
0 · e1 + 0 · e2 0 · e1 + 1 · e2 0 · e1 + 0 · e2 0 · e1 + 1 · e2
0 · e1 + 0 · e2 0 · e1 + 0 · e2 1 · e1 + 0 · e2 1 · e1 + 1 · e2
0 · e1 + 0 · e2 0 · e1 + 0 · e2 0 · e1 + 0 · e2 0 · e1 + 0 · e2
TABLE 23. Four possible types of relators in the extensions.
92
extensions in table 24. The extensions of type 2 and type 3 are isomorphic, since
they are converted with each other by interchanging x and y, and e1 and e2 .
type
type
type
type
−1
x
y
y
1
x, y, e1 , e2 |x2 = y 2 = (xy)2 = [e1 , e2 ] = 1, x e1 = e−1
1 , e2 = e2 , e1 = e1 , e2 = e2
−1 x
2
2
2
x
y
y
2 x, y, e1 , e2 |x = e2 , y = (xy) = [e1 , e2 ] = 1, e1 = e1 , e2 = e2 , e1 = e1 , e2 = e−1
2
−1
x
y
y
,
e
=
e
,
e
=
e
,
e
=
e
3 x, y, e1 , e2 |y 2 = e1 , x2 = (xy)2 = [e1 , e2 ] = 1, x e1 = e−1
2
2
1
1
2
1
2
−1
4 x, y, e1 , e2 |x2 = e2 , y 2 = e1 , (xy)2 = [e1 , e2 ] = 1, x e1 = e1 , x e2 = e2 , y e1 = e1 , y e2 = e−1
2
TABLE 24. Four types of the extension. e1 and e2 should be assumed to be translations
{E|e1 } and {E|e2 }, rather than vectors.
We can see the relators of the extensions provide the relations which determine
the vector system in the non-symmorphic crystal, with the point group P .(To obtain
the vector system, it is enough to re-express x,y as the augmented matrix forms
of the space groups, with indeterminate fractional translations. The e1 and e2 are
also represented by matrices as translations. The equivalence relations in the entries
of the matrix products in the group multiplications will give the vector systems,
as was demonstrated elsewhere in this article. In that demonstration, the vector
system was also derived from cohomological algebra, not from two-cohomology, but
from one-cohomology.) From this simple example, we can observe some importance
of homological algebra, which provides us the criterion of the existence of the nonsymmorphic crystal, in accordance with the non-triviality of two-cohomology. On
the other hand, if the two-cohomology is trivial, or of zero-mapping, the extension
of the group simply becomes a semidirect product, which is always a symmorphic
crystal. In this case of zero two-cohomology, the exact sequence splits, in the sense
that there is a homomorphism s : G → H such as π ◦ s = idG . It can be proved that
an extension (or, the short exact sequence as above) is split if and only if the group
H is the semidirect product of N and G. (We should remember this: elsewhere in this
article, the point group of the non-symmorphic crystal is constructed by means of the
semidirect product. In that computation, the point group of the non-symmorphic
93
crystal is not directly obtained as a semidirect product but extracted as one of its
subgroups. So there is no contradiction against the above statement.)
Now the relationship is established among the symmorphic properties of crystals,
the splitting extension of a group, and the two-cohomology, as is shown in the table
below.
type
type
type
type
1 {Mx |0}, {My |0}, {E|e1 }, {E|e2 }
2 {Mx | 12 e2 }, {My |0}, {E|e1 }, {E|e2 }
3 {Mx |0}, {My | 21 e1 }, {E|e1 }, {E|e2 }
4 {Mx | 12 e2 }, {My | 21 e1 }, {E|e1 }, {E|e2 }
TABLE 25. The generators of the four types of the space groups classified by the twocohomology.
In getting through this section, two necessary points for the computations in GAP
are given. The first is how to derive the relators, or the finite representation: the
second is the computation of the Smith Normal form. (The two-cohomology can be
computed by GAP built-in command, for the case of a finite group G and a module
M , over a finite field. Because of this assumption, the GAP built-in command is not
applicable to the above example, in which the module M is Zn .)
A definition is now needed: a group G is polycyclic if there is a series of subgroups
G = C1 > C2 > · · · > Cn > Cn+1 = 1,
where Ci+1 is normal in Ci and Ci /Ci+1 is a cyclic group. The generators of G is
given by a sequence G := (g1 , . . . , gn ) such that Ci = hCi+1 , gi i. As a finite solvable
group, the point group of the crystal is polycyclic, which is composed from the
polycyclic generating system.(The point group of the crystal is always solvable, by
means of a famous theorem of Burnside, which relates the order of the group to the
solvability.) We have defined the point groups by means of matrix generators, or
94
the multiplication table. These groups can be converted into isomorphic polycyclic
groups, and then, polycyclic groups are converted into finitely presented groups with
relators. The conversion goes as follows:
gap> A:=[ [ 0, 0, 1 ], [ 1, 0, 0 ], [ 0, -1, 0 ] ];
gap> B:=[ [ 0, 0, -1 ], [ 0, -1, 0 ], [ 1, 0, 0 ] ];
gap> G:=Group(A,B);
Group([ [ [ 0, 0, 1 ], [ 1, 0, 0 ], [ 0, -1, 0 ] ],
[ [ 0, 0, -1 ], [ 0, -1, 0 ], [ 1, 0, 0 ] ] ])
gap> p:=Pcgs(G);
Pcgs([ [ [ 0, 0, 1 ], [ 0, 1, 0 ], [ -1, 0, 0 ] ],
[ [ 1, 0, 0 ], [ 0, 1, 0 ], [ 0, 0, -1 ] ],
[ [ 0, 0, -1 ], [ -1, 0, 0 ], [ 0, 1, 0 ] ],
[ [ 1, 0, 0 ], [ 0, -1, 0 ], [ 0, 0, -1 ] ],
[ [ -1, 0, 0 ], [ 0, -1, 0 ], [ 0, 0, 1 ] ] ])
gap> iso:=IsomorphismFpGroupByPcgs(p,"g");
CompositionMapping( [ (1,4,6,8)(2,3,5,7), (1,4)(2,3)(5,7)(6,8),
(3,6,7)(4,5,8), (1,7)(2,8)(3,6)(4,5), (1,3)(2,4)(5,8)(6,7) ] ->
[ g1, g2, g3, g4, g5 ], <action isomorphism> )
gap> fp:=Image(iso);
<fp group of size 48 on the generators [ g1, g2, g3, g4, g5 ]>
gap> RelatorsOfFpGroup(fp);
[ g1^2*g5^-1*g4^-1, g2^-1*g1^-1*g2*g1*g5^-1*g4^-1,
g3^-1*g1^-1*g3*g1*g5^-1*g4^-1*g3^-1, g4^-1*g1^-1*g4*g1*g5^-1*g4^-1,
g5^-1*g1^-1*g5*g1*g5^-1*g4^-1, g2^2, g3^-1*g2^-1*g3*g2*g4^-1,
g4^-1*g2^-1*g4*g2, g5^-1*g2^-1*g5*g2, g3^3, g4^-1*g3^-1*g4*g3*g5^-1*g4^-1,
g5^-1*g3^-1*g5*g3*g4^-1, g4^2, g5^-1*g4^-1*g5*g4, g5^2 ]
95
The command “Pcgs(G)” returns the polycyclic generating system of G. (In
this example, the generators are accessible through the list p[1],...p[5].) The
command “IsomorphismOfFpGroupByPcgs(p,”g”)” gives the map which converts
the polycyclic generating system into a finitely represented group. As the image of
this isomorphism, the polycyclic group is given. However, the polycyclic generating
system is not minimal. To generate the minimal one, we should do as this:
gap> iso:=IsomorphismFpGroupByGenerators(G,GeneratorsOfGroup(G));
<composed isomorphism:[ [ [ 0, 0, 1 ], [ 1, 0, 0 ], [ 0, -1, 0 ] ],
[ [ 0, 0, -1 ], [ 0, -1, 0 ], [ 1, 0, 0 ] ] ]->[ F1, F2 ]>
gap> fp:=Image(iso);
<fp group of size 48 on the generators [ F1, F2 ]>
gap> RelatorsOfFpGroup(fp);
[ F2^4, (F1*F2^-1)^2, F1^6, (F1*F2*F1)^2 ]
gap> [G.2^4,(G.1*G.2^-1)^2,G.1^6,(G.1*G.2*G.1)^2];
[ [ [ 1, 0, 0 ], [ 0, 1, 0 ], [ 0, 0, 1 ] ],
[ [ 1, 0, 0 ], [ 0, 1, 0 ], [ 0, 0, 1 ] ],
[ [ 1, 0, 0 ], [ 0, 1, 0 ], [ 0, 0, 1 ] ],
[ [ 1, 0, 0 ], [ 0, 1, 0 ], [ 0, 0, 1 ] ] ]
The smith normal form is also computed by GAP:
gap> M:=[[0,0,0,0],[0,2,0,0],[0,0,0,0],[0,0,2,0],[0,0,0,0],[0,0,0,0]];
gap> SmithNormalFormIntegerMat(M);
[ [ 2, 0, 0, 0 ], [ 0, 2, 0, 0 ], [ 0, 0, 0, 0 ], [ 0, 0, 0, 0 ],
[ 0, 0, 0, 0 ], [ 0, 0, 0, 0 ] ]
96
H.
Remarks to this section
In this section, the technical details which would be important in the actual application of the computational group theory in the field of solid-state physics are
explained. The chosen topics are as follows: the algorithm to determine the symmetry of the crystal; the computation of the compatibility relation, which is useful
in the discussion of the splitting of the energy spectra with the reduction of the
structural symmetry; the concept of the semidirect product, which is applicable to
the preparation of the character table of the supercell structure; the concept of the
group extension, by means of cohomology, which can be utilized to design the crystal lattice. The author hopes that the readers of this article would get familiar with
these technical or advanced topics in order to deepen the knowledge of group theory,
applicable to concrete problems in material science. The small programs presented in
this article can be executed on a desktop PC. They are composed in accordance with
the elementary formulas and definitions in the references and textbooks so that they
should be as readable as possible, thus they are not are always effective in speed, and
some part of them could be replaced by built-in functions in GAP. In the preceding sections of this article, the topics are taken from the solid-state physics. In the
forthcoming sections, however, the application to the quantum chemistry (applicable
to the isolated system, equipped with a certain symmetry, without periodicity) shall
be discussed in a parallel way as the solid state physics, where the importance of
the symmetry in the quantum dynamics of the molecule shall be illustrated in the
different light.
97
VII.
SYMMETRY IN C60
The C60 molecule is characterized by the peculiar structure, that of soccer-ball,
composed of 60 atoms, 90 bonds, and 12 pentagons, and 60 hexagons. The group
theoretical symmetry is of the icosahedral type, which contains 120 symmetric operations and 10 irreducible representations, according to which, the eigenstates are
classified and the origins of degeneracies in energy spectrum are explained. We begin
the construction of the symmetric operations by means of permutations of atoms,
generate the icosahedral group by GAP, compute the irreducible representations, and
compose the projection operators. We execute the electronic structure calculations of
the molecule and analyze the wave-functions from the viewpoint of their irreducible
representation. Furthermore, some concepts of symmetries (to which physicists pays
little notice) are presented also by the analysis of the computer algebra. Throughout
the study, it is necessary that whole subgroups of icosahedral symmetry should be
extracted and classified by equivalence, and we will see that the utilization of the
computer algebra is effective for this purpose.
A.
The initial set-up
The C60 molecule (in Fig. 6) is projected on a plane as in the Fig. 7. The
vertexes are numbered from 1 to 60, where we follow the following rule: the two
atoms numbered as I and I+30 (I=1,...,30) are located in the opposite poles in the
molecular sphere. The pentagons are also numbered from 1 to 12 (with under-bars),
and two pentagons numbered as J and J+6 (J=1,...,6) are opposed with each other.
There are two types of bonds, single and double, having slightly different lengths,
distinguished by broad and broken segments in the figure. The icosahedral group are
generated by three basic operations and represented by the permutations of atoms as
98
FIG. 6. A C60 molecule. The numbering is given on some atoms and pentagons, in
correspondence with the projected images of the molecule, in Fig. 7.
follows. The group is the direct product of the fifth alternating group “A5” (the 5-th
alternating group) and the group of the inversion “I”. The generators are represented
by permutations:
gap> g1:=(1, 14, 20)(2, 15, 16)(3, 11, 17)(4, 12, 18)(5, 13, 19)
(6, 58, 22)(7, 57, 23)( 8, 56, 24)( 9, 60, 25)( 10, 59, 21)
( 31, 44, 50)( 32, 45, 46)( 33, 41, 47)( 34, 42, 48)( 35, 43, 49)
( 36, 28, 52)( 37, 27, 53)( 38, 26, 54)( 39, 30, 55)( 40, 29, 51);
gap> g2:=( 1, 5, 4, 3, 2)( 6,26,21,16,11)( 7,27,22,17,12)( 8,28,23,18,13)
( 9,29,24,19,14)(10,30,25,20,15)(31,35,34,33,32)(36,56,51,46,41)
(37,57,52,47,42)(38,58,53,48,43)(39,59,54,49,44)(40,60,55,50,45);
gap> g3:=( 1, 31)( 2, 32)( 3, 33)( 4, 34)( 5, 35)( 6, 36)( 7, 37)( 8, 38)
( 9, 39)( 10, 40)( 11, 41)( 12, 42)( 13, 43)( 14, 44)( 15, 45)( 16, 46)
( 17, 47)( 18, 48)( 19, 49)( 20, 50)( 21, 51)( 22, 52)( 23, 53)( 24, 54)
99
FIG. 7. The projected image of C60 on a plane. The pentagon in the center of this figure
(indexed as 1) is the one located in the lower half in Fig. 6. (The pentagon at No.7, not
shown in the projected image, is formed by five vertex at 31,32,33,34,35. )
( 25, 55)( 26, 56)( 27, 57)( 28, 58)( 29, 59)( 30, 60);
The elements g1 (the rotation by 120 degrees) and g2 (that by 72 degrees) generate
“A5“ and g3 generates “I”. The permutation, denoted by (a, b, c, d,...,x), should be
interpreted as the replacement of atoms in the following way: the atom located at the
100
site “a” is replaced by that at the site “b”, that at “b” by that at “c”, then in turn and
finally, the last entry “x” by the first “a”. The usage of sixty symbols for the definition
of the group is superfluous, and in fact, twelve symbols (representing the 12 vertexes
of the icosahedron, or the 12 pentagons in the molecule) will do. If it is unnecessary
to use the geometrical image, one can use 7 symbols, since, the icosahedral group is
the direct product of the 5-th alternating group A5 and the inversion I, as the former
requires five symbols “1,2,3,4,5” and the latter two symbols “6,7”. The equivalence
of these different definitions is assured by constructing isomorphisms among them,
which is an easy task for the computer algebra.
Remark. This correspondence (between the group elements and the atoms) is consistent to the operation of the permutation in GAP. For example, the permutation
(3,4,5) operates on the list [1,2,3,4,5] as follows:
gap> ListPerm((3,4,5));
[1,2,4,5,3];
Here the entries 3,4,5 are replaced by 4,5,3, as our interpretation assumes.
The group is constructed as
gap> gp:=Group(g1,g2,g3);
For the later utilization, it is more convenient to generate the group by means of
matrices (which represents the permutations on sixty vertexes). We use the built-in
function “ListPerm()” which computes the action on the permutation on the list
[1,2,...,60].
gap> pl1:=ListPerm(g1);
[ 14, 15, 11, 12, 13, 58, 57, 56, 60, 59, 17, 18, 19, 20, 16, 2, 3, 4, 5, 1,
10, 6, 7, 8, 9, 54, 53, 52, 51, 55, 44, 45, 41, 42, 43, 28, 27, 26, 30, 29,
101
47, 48, 49, 50, 46, 32, 33, 34, 35, 31, 40, 36, 37, 38, 39, 24, 23, 22, 21,
25 ]
gap> pl2:=ListPerm(g2);;
gap> pl3:=ListPerm(g3);;
These three lists are converted into three matrices by means of this user-defined
function. (The computed matrix element h[i][j] is nonzero when the symbol “i” goes
to “j” by permutation given in the list “l”.)
ListToMat:=function(l)
local size,h,i,j;
size:=Size(l);
h:=List([1..size],i->List([1..size],j->0));
for i in [1..size] do
h[i][l[i]]:=1;
od;
return h;
end;
gap> m1:=ListToMat(pl1);;
gap> m2:=ListToMat(pl2);;
gap> m3:=ListToMat(pl3);;
The irreducible representations are computed now.
gap> g:=Group(m1,m2,m3);
gap> irr:=Irr(g);
gap> Size(irr);
10
102
Remark. The computed irreducible representations by the group generated by the
permutations and that by the matrices are listed by different orders (at least, in
the result by GAP of the version 4.7.7.) This will cause some trouble in doing
computations. In this article, to avoid confusion, the groups generated by different
resources are given by different names even if they are isomorphic, and the generators
are presented in each time of the definition.
The characters for each element can be listed as numeral tables.
gap> elm:=Elements(g);
gap> A1:=List(elm,g->g^irr[1]);;
gap> A2:=List(elm,g->g^irr[2]);;
...............................
gap> A10:=List(elm,g->g^irr[10]);;
The notation g^irr[X] means the computation of the character value of the element
“g” by the X-th representation in this context.
The characters of the irreducible representations are orthogonal and normalized.
gap> A1*A1;
120
gap> A2*A2;
120
gap> A1*A2;
0
The virtual character is decomposed into the direct sum.
gap> (4*A1+2*A2)*TransposedMat([A1,A2,A3,A4,A5,A6,A7,A8,A9,A10])/Size(elm);
[4, 2, 0, 0, 0, 0, 0, 0, 0, 0]
103
I.1
I.2
I.3
I.4
I.5
I.6
I.7
I.8
I.9
I.10
I.Tr
CL1
1
1
1
3
3
3
3
4
4
5
5
60
CL2
12
1
-1
-E(5)-E(5)^4
-E(5)^2-E(5)^3
E(5)^2+E(5)^3
E(5)+E(5)^4
-1
1
0
0
0
CL3
12
1
1
-E(5)^2-E(5)^3
-E(5)-E(5)^4
-E(5)-E(5)^4
-E(5)^2-E(5)^3
-1
-1
0
0
0
CL4
20
1
-1
0
0
0
0
1
-1
-1
1
0
CL5
20
1
1
0
0
0
0
1
1
-1
-1
0
CL6
15
1
-1
-1
-1
1
1
0
0
1
-1
4
CL7
12
1
-1
-E(5)^2-E(5)^3
-E(5)-E(5)^4
E(5)+E(5)^4
E(5)^2+E(5)^3
-1
1
0
0
0
CL8
1
1
-1
3
3
-3
-3
4
-4
5
-5
0
CL9
CL10
12
15
1
1
1
1
-E(5)-E(5)^4
-1
-E(5)^2-E(5)^3 -1
-E(5)^2-E(5)^3 -1
-E(5)-E(5)^4
-1
-1
0
-1
0
0
1
0
1
0
0
TABLE 26. The table of characters of the irreducible representation of the icosahedral
group. Each row shows the characters of 10 conjugacy classes (CL1,...,CL10).
There are 10 conjugacy classes and 10 irreducible representations. They are shown
in the tables 26 and 27. In table 26, the second row shows the number of elements in
each conjugacy classes (CL1,...,CL10), and the ten rows bellow are characters for each
of the ten irreducible representations (I.1,...,I.10). The last row (I.Tr) shows the trace
of the group elements as matrices, which is a virtual character (not irreducible). The
√
notation E(5) means a complex number, the fifth root of the unity exp(2 −1π/5),
thus the characters are real-valued. Simply the computed result by GAP is tabulated
here, as it is. The projector on the p-th irreducible representation is computed by
the formula
P (p) =
l(p) X (p)∗
χ (T ) · O(T )
|G| T ∈G
by summing up the elements T of the group G, where l(p) is the dimension of the
p-th irreducible representation, g the order of the group, χ(p) the character, O(T )
the operation of the group (which would be represented as matrices and applied to
the vector representation of the wave function. The irreducible representation of the
104
The representatives
CL1
<identity ...>
x1^-1*x2^-2*x3
CL2
CL3
x2^-1*x1*x2*x1
x3*x2*x1*x2^2
CL4
CL5 x2^-2*x1*x2*x1 (also including x1)
CL6
x3*x2^2*x1*x2^2
CL7
x3*x2^2
CL8
x3
CL9 x1^-1*x2^-2 (also including x2)
CL10
x1*x2*x1
TABLE 27. The representatives in the conjugacy classes of the irreducible representations
of the icosahedral group. The generators of the groups are represented by symbols x1...xN,
as are numbered by GAP. In this case x1=m1,x2=m2,x3=m3.
wave function is determined numerically, by the condition that the application of the
projector should keep the wave function as it is, or delete it.)
B.
The analysis of the eigenstates
Now that we have the irreducible representations and the symmetry operations,
we can construct the projection operators.
At first, we assume a model of C60 , where, in each vertex, one s or π orbital is
located. The inter-atomic interactions are limited to those among three different
neighboring sites, as are signified by single and double bonds. We permit two different strengths of interactions in the Fock matrix H, according to two types of the
bonding. We neglect the overlap matrix elements between different sites, the overlapping matrix S being the unity. A computation is executed by the above assumption,
with the values of Hsingle =-2 (on the single bonds in the pentagons), Hdouble =-1 (on
the double bonds in the hexagons).
Before the electronic structure computation, we can deduce a prediction from
105
the group theory. The group elements given by the matrices which represent the
replacement of the 60 vertexes provide us the trace representation I.Tr. The virtual
representation is decomposed by the irreducible representations (as the sum of row
vectors in the table):
I.Tr = 1×I.1+0×I.2+1×I.3+1×I.4+2×I.5+2×I.6+2×I.7+2×I.8+3×I.9+2×I.10.
This relation tells us that the 60-dimensional vector space (on which the matrix group
acts) is divided into the ten irreducible components, given as the ten irreducible
representations of the group. The multiplicity of X-th component is equal to the
coefficient to “I.X”. One of these components, belonging to “I.2”, is a null space in
this case. The wave functions, computed as the eigenvectors of the Fock matrix, are
also subject to this division of the vector space.
Let us verify this. The eigenstates and the corresponding irreducible representations are shown in the following list. The abbreviations as “I.”,“D.”,“E” denote
the index of the irreducible representation, the degeneracy of the energy, and the
energies. The values of the energies are listed in the bracket. If the bracket is blank,
the eigenstates in correspondence with the irreducible representations do not exist.
The number of different eigenvalues in each irreducible representation is equal to the
multiplicity in the above relation between I.Tr and I.1,..., I.10.
I.1
D.1 E.[ -5.0000 ]
I.2
D.1 E.[ None ]
I.3
D.3 E.[ 4.2361 ]
I.4
D.3 E.[ -0.2361 ]
I.5
D.3 E.[ -4.6542, -0.5819 ]
I.6
D.3 E.[ -3.6765, 2.9126 ]
I.7
D.4 E.[ -1.4495, 3.4495 ]
106
I.8
D.4 E.[ -2.1623, 4.1623 ]
I.9
D.5 E.[ -4.0642, -1.6946, 2.7588 ]
I.10 D.5 E.[ -1.0000, 3.0000 ]
For a realistic quantum chemistry computation of C60 , the similar analysis is possible. For simplicity, one can make use of the amplitudes of 1s orbitals, because
those orbitals are transformed among themselves, without mixing with other types,
2s,2p,3s,3p. Except for one special case where the amplitudes of 1s orbital are null,
the correspondence between the eigenstates and the irreducible representations can
be well established. The result is listed in the following list. (The last entry of the list
I.0 is for the case of the failures in the identification.) Even in the case of the failure
of the analysis, the degeneracy is single, and one can conclude the corresponding
irreducible representation is I.1 or I.2, that of 1-dimension. The numbers of the different eigenvalues in the each irreducible representations are predictable by means of
the decomposition of the trace representation of the character of the transformation
matrices of the 1s,2s,2px,2py,2pz,3s,3px,3py,3pz orbitals. (The p orbitals are transformed by the same matrices prepared for the vibrational mode of the molecule, as
will be discussed later.)
I.1
D.1
E.[
7 different eigenvalues ]
I.2
D.1
E.[ ? ]
I.3
D.3
E.[ 11 different eigenvalues ]
I.4
D.3
E.[ 11 different eigenvalues ]
I.5
D.3
E.[ 16 different eigenvalues ]
I.6
D.3
E.[ 16 different eigenvalues ]
I.7
D.4
E.[ 24 different eigenvalues ]
I.8
D.4
E.[ 24 different eigenvalues ]
I.9
D.5
E.[ 25 different eigenvalues ]
107
I.10 D.5
E.[ 20 different eigenvalues ]
I.0
E.[
C.
D.1
2 different eigenvalues ]
The analysis again, against the failure in the identification
As the degeneracy of the energy is one, one can see the levels denoted as I.0 in
the previous section belong to the representation I.1 or I.2. The characters of the
one-dimensional representation in table 26 are shown here again (in table 28).
CL1 CL2 CL3 CL4 CL5 CL6 CL7 CL8 CL9 CL10
I.1 1
1
1
1
1
1
1
1
1
1
I.2 1 -1 1 -1 1 -1 -1 -1 1
1
TABLE 28. The one-dimensional irreducible representations.
To distinguish the representation I.1 and I.2 with ease, one may apply the element
in the conjugacy class 8 (the inversion). By the inversion, one vertex moves to the
opposite site with respect to the center of the geometry, and the coefficients of porbitals change signs. The amplitudes of the wave-functions (in the two atoms located
in the antipodes positions) in I.0 are typically obtained as in table 29.
CA 1S
2S
2PX
2PY
2PZ
CB 1S
2S
2PX
2PY
2PZ
0
0
X
Y
Z
0
0
X
Y
Z
TABLE 29. The amplitudes of the wave-function (coefficients for each orbital) in two
opposing atoms.
108
The inversion changes the sign of non-zero coefficients. We must recall that the
wave-functions are basis functions of some irreducible representations, and the traces
of their transformation matrices are the characters. Since the traces are -1 in the
above cases, the wave-functions belong to the representation I.2. The reason of the
failure in the analysis by the projection operators applied on s orbitals are fundamental one: the projection operator of I.2 over s orbitals is exactly the zero-matrix,
which gives rise to exact zero amplitudes in s orbitals (i.e. the null space.) The
reason of the existence of the orbitals of the I.2 is that the actual C60 molecule has
the symmetry of the inversion and the carbon atom has p-orbitals.
D.
The perturbation and the symmetry
We often have to compute matrix elements in the perturbation theory such as
hψi |Vp |ψj i = [Ii |Ip |Ij ] = [Ii |Ip ∗ Ij ].
(83)
The product of the perturbation potential and the wave function belongs to the
direct product of the representation, which is decomposed into the direct sum
Ip ∗ Ij =
X
Ap,j
k Ik .
(84)
k
We can determine beforehand whether this matrix element would be zero or not, only
from the group theoretical consideration: if the two irreducible representations in the
inner product are the same, the matrix element will be non-zero. In the system with
the high symmetry, the group theoretical constraint is so strong that it annihilates
many matrix elements. However, the symmetry is reduced, there is some possibility
where the matrix element will become non-zero. Indeed there is no symmetry, there
109
is no group theoretical constraint which annihilates the matrix elements. And if
one uses the group theory, one can see the least reduction of the symmetry where
the matrix element will become non-zero. The irreducible representation of group g
ramifies in the subgroups as follows.
(g,s)
(g,s)
(g,s)
Igp = Ci,1 Is1 + Ci,2 Is2 + · · · + Ci,M IsM .
(85)
One can find the subgroup where the non-zero matrix element will exist, by tracing
the sequence of subgroups (in the subgroup lattice) and checking the possible combinations of the ramified irreducible representations. In this stage, the concrete style
of the structural deformation is not discussed yet.
One of the important perturbation potentials is that of optical transition, such as
E ·r. Now we determine the irreducible representation to which the perturbation E ·r
(or the spatial vector r) belongs. We now construct the icosahedral group from the
rotation matrices in three-dimensional Euclidean space, acting on the twelve vertexes
of one icosahedron.
The twelve vertexes in one icosahedron are given by
1
[ 0, 0, 1 ]
2
[ 2/5*E(5)-2/5*E(5)^2-2/5*E(5)^3+2/5*E(5)^4, 0,
1/5*E(5)-1/5*E(5)^2-1/5*E(5)^3+1/5*E(5)^4 ]
3
[ -3/5*E(5)-2/5*E(5)^2-2/5*E(5)^3-3/5*E(5)^4,
-1/5*E(20)+1/5*E(20)^9+2/5*E(20)^13-2/5*E(20)^17,
1/5*E(5)-1/5*E(5)^2-1/5*E(5)^3+1/5*E(5)^4 ]
4
[ 2/5*E(5)+3/5*E(5)^2+3/5*E(5)^3+2/5*E(5)^4,
-2/5*E(20)+2/5*E(20)^9-1/5*E(20)^13+1/5*E(20)^17,
1/5*E(5)-1/5*E(5)^2-1/5*E(5)^3+1/5*E(5)^4 ]
5
[ 2/5*E(5)+3/5*E(5)^2+3/5*E(5)^3+2/5*E(5)^4,
110
2/5*E(20)-2/5*E(20)^9+1/5*E(20)^13-1/5*E(20)^17,
1/5*E(5)-1/5*E(5)^2-1/5*E(5)^3+1/5*E(5)^4 ],
6
[ -3/5*E(5)-2/5*E(5)^2-2/5*E(5)^3-3/5*E(5)^4,
1/5*E(20)-1/5*E(20)^9-2/5*E(20)^13+2/5*E(20)^17,
1/5*E(5)-1/5*E(5)^2-1/5*E(5)^3+1/5*E(5)^4 ]
7
[ 0, 0, -1 ]
8
[ -2/5*E(5)+2/5*E(5)^2+2/5*E(5)^3-2/5*E(5)^4, 0,
-1/5*E(5)+1/5*E(5)^2+1/5*E(5)^3-1/5*E(5)^4 ]
9
[ 3/5*E(5)+2/5*E(5)^2+2/5*E(5)^3+3/5*E(5)^4,
1/5*E(20)-1/5*E(20)^9-2/5*E(20)^13+2/5*E(20)^17,
-1/5*E(5)+1/5*E(5)^2+1/5*E(5)^3-1/5*E(5)^4 ]
10
[ -2/5*E(5)-3/5*E(5)^2-3/5*E(5)^3-2/5*E(5)^4,
2/5*E(20)-2/5*E(20)^9+1/5*E(20)^13-1/5*E(20)^17,
-1/5*E(5)+1/5*E(5)^2+1/5*E(5)^3-1/5*E(5)^4 ]
11
[ -2/5*E(5)-3/5*E(5)^2-3/5*E(5)^3-2/5*E(5)^4,
-2/5*E(20)+2/5*E(20)^9-1/5*E(20)^13+1/5*E(20)^17,
-1/5*E(5)+1/5*E(5)^2+1/5*E(5)^3-1/5*E(5)^4 ]
12
[ 3/5*E(5)+2/5*E(5)^2+2/5*E(5)^3+3/5*E(5)^4,
-1/5*E(20)+1/5*E(20)^9+2/5*E(20)^13-2/5*E(20)^17,
-1/5*E(5)+1/5*E(5)^2+1/5*E(5)^3-1/5*E(5)^4 ]
The three generators of the group are as follows:
gap> f1:=(1,2,3)(4,6,11)(5,10,12)(7,8,9);;
gap> f2:=(2,6,5,4,3)(8,12,11,10,9);;
gap> f3:=(1,7)(2,8)(3,9)(4,10)(5,11)(6,12);;
These permutations are also represented by the rotation matrices as
111
gap> f1:=[ [ 3/10*E(5)+1/5*E(5)^2+1/5*E(5)^3+3/10*E(5)^4,
1/10*E(20)-1/10*E(20)^9-1/5*E(20)^13+1/5*E(20)^17,
2/5*E(5)-2/5*E(5)^2-2/5*E(5)^3+2/5*E(5)^4 ],
[ -1/2*E(20)+1/2*E(20)^9, -1/2*E(5)-1/2*E(5)^4, 0 ],
[ -3/5*E(5)-2/5*E(5)^2-2/5*E(5)^3-3/5*E(5)^4,
-1/5*E(20)+1/5*E(20)^9+2/5*E(20)^13-2/5*E(20)^17,
1/5*E(5)-1/5*E(5)^2-1/5*E(5)^3+1/5*E(5)^4 ] ];
gap> f2:=[ [ 1/2*E(5)+1/2*E(5)^4, -1/2*E(20)+1/2*E(20)^9, 0 ],
[ 1/2*E(20)-1/2*E(20)^9, 1/2*E(5)+1/2*E(5)^4, 0 ], [ 0, 0, 1 ] ];;
gap> f3:=[ [ -1, 0, 0 ], [ 0, -1, 0 ], [ 0, 0, -1 ] ];;
N.1
N.2
N.3
N.4
N.5
N.6
N.7
N.8
N.9
N.10
Dimension
1
1
3
3
3
3
4
4
5
5
[
[
[
[
[
[
[
[
[
[
[
[
[
[
[
[
[
[
[
[
0,
0,
0,
0,
1,
0,
0,
0,
0,
0,
0,
0,
0,
0,
0,
0,
0,
0,
0,
0,
0
0
0
0
0
0
0
0
0
0
],
],
],
],
],
],
],
],
],
],
Operator
[ 0, 0, 0 ],
[ 0, 0, 0 ],
[ 0, 0, 0 ],
[ 0, 0, 0 ],
[ 0, 1, 0 ],
[ 0, 0, 0 ],
[ 0, 0, 0 ],
[ 0, 0, 0 ],
[ 0, 0, 0 ],
[ 0, 0, 0 ],
[
[
[
[
[
[
[
[
[
[
0,
0,
0,
0,
0,
0,
0,
0,
0,
0,
0,
0,
0,
0,
0,
0,
0,
0,
0,
0,
0
0
0
0
1
0
0
0
0
0
]
]
]
]
]
]
]
]
]
]
]
]
]
]
]
]
]
]
]
]
The characters of [f1,f2,f3]
[ 1, 1, 1 ]
[ 1, 1, -1 ]
[ 0,-E(5)^2-E(5)^3, 3 ]
(I.4 in table 26)
[ 0,-E(5)-E(5)^4, 3 ]
(I.3 in 26)
[ 0,-E(5)^2-E(5)^3, -3 ]
[ 0,-E(5)-E(5)^4, -3 ]
[ 1, -1, 4 ]
[ 1, -1, -4 ]
[ -1, 0, 5 ]
[ -1, 0, -5 ]
TABLE 30. The projection operators. Be careful that the computation by GAP returns the representations (N.1,...,N.10) in the slightly different order from those in table
26(I.1,...,I.10).
The ten projection operators are computed as in table 30. The operators, except
that of the fifth representation, are exactly zero. The symmetry operations on one
vector, which is subject to the icosahedral symmetry, are only representable in the
fifth irreducible representation, and the perturbation term E · r is represented by
the fifth representation. (The fifth representation can be distinguished from others,
112
as such having characters [0,-E(5)^2-E(5)^3,-3] for three generators. Especially,
the rotation of 72 degrees should have the character value -E(5)^2-E(5)^3. The
rotation by 72 degrees and its inverse (f2 and f2^-1) have the same character value,
because they are conjugate.)
Remark. In order to see whether the elements x and y are conjugate in group G,
try the command.
gap> RepresentativeAction(G,x,y);
E.
The possible deformation in the reduced symmetry
We can construct molecular geometry which is admissible in each subgroup. If the
vertexes are dislocated as V(i) → V(i)+d(i), the set of d(i) should satisfy a certain
geometrical constraint in the symmetry operation in the subgroup. We should take
note of this : since the icosahedral group which includes 120 operations and acts on
60 vertexes in one C60 , there are two operations, both of which equally transfers one
atomic site into same, but another site. If the dislocations of atoms are still subject
to the icosahedral symmetry, there is a constraint. When the atom at the site V[i]
moves to V[j] by the operation R, it must be
R · V[i] = V[j],
and
R · (V[i] + d[i]) = V[j] + d[j],
i.e.
R · d[i] = d[j].
If there are two possible transformations R and Q, it must be
R · d[i] = Q · d[i].
113
If we chose T=R−1 Q, then T·d[i]=d[i], and the possible dislocation is the solution of
the equation (T − E) · d[i] = 0 (E: the unit matrix). In the case of the icosahedral
symmetry in C60 , such a T is the reflection at the plain in which one double bond
and its antipodes are located.
In general, the dislocations d[i] on all of the vertex are generated as follows. For
example, take V[1], and find the operation R, Q, T as above, and then find a possible
solution d of (T − E) · d = 0 as the generator. The application of the 120 operations
on V[1]+d shall generate 60 sites, which are represented by V[i]+d[i], and the set
of d[i] is obtained. (This is equivalent to the projection on the trivial irreducible
representation, in which the basis set is invariant under the group action.) In a
subgroup, the symmetric operations do not act transitively, and the “orbit” of one
vertex by the symmetry operations (which covers all of the vertexes in C60 by the
icosahedral symmetry) splits into several subsets, as we will see later. For each
subset, we construct independent d, the generator of the dislocation in each subset.
If the subgroups contain the reflection, the dislocations on the atoms have similar
constraints to d (being fixed on one of the reflectional planes) as in the case of
the icosahedral symmetry. On the contrary, when the reflection is missing in the
subgroup, we can choose arbitrary generator of dislocations in each sub-orbits. One
should note that the center of mass will not be kept since d[1]+d[2]+...+d[60] is not
equal to 0 in general.
Concerning this topic, we can classify the vibrational mode of C60 according to
the group theory. The procedure is to form the direct product of the virtual representation with respect to the replacement of atoms and the irreducible representation
of the symmetry operations on the space vector, and then to decompose it into the
direct sum. One will find that the vibrational mode in the icosahedral symmetry
includes the duplicated trivial irreducible representation, to which the degrees of the
freedom is two. The geometrical interpretation of this is that the possible distortion
114
which keeps the icosahedral symmetry is confined in one plane as discussed above.
F.
The reduction of the symmetry, from the viewpoint of the orbit in the
permutation
The full symmetry of C60 is the icosahedral group, generated by “g1”, “g2”, “g3”
in the previous section, and it has 164 subgroups. All of the subgroups are computed
as
gap> gr:=Group(g1,g2,g3);
gap> AS:=AllSubgroups(gr);;
The list of the subgroups is given in the descending order of the sizes of the
subgroups, and the largest of them (No.164) is the icosahedral group.
gap> irrA:=Irr(AS[164]);
Take a subgroup No.162. This group is of the order 24, and the irreducible
representation is also computed.
gap> irrB:=Irr(AS[162]);;
The irreducible representations of these groups are given in tables 31,32,33, and
34.
The branching relation from the group of No.164 to that of No.162 is computed
now. The characters of the elements of the group of No.162 are computed and listed
in two ways: the one is by the irreducible representations of the group of No.164,
and another by those of the group of No.162. The representations in the first list are
decomposed as the summation of the entries in the second list, and the result is the
branching relation.
115
No.1
No.2
No.3
No.4
No.5
No.6
No.7
No.8
No.9
No.10
CL1
1
1
3
3
3
3
4
4
5
5
CL2
1
-1
-1
-1
1
1
0
0
1
-1
CL3
1
1
-E(5)-E(5)^4
-E(5)^2-E(5)^3
-E(5)-E(5)^4
-E(5)^2-E(5)^3
-1
-1
0
0
CL4
1
1
-E(5)^2-E(5)^3
-E(5)-E(5)^4
-E(5)^2-E(5)^3
-E(5)-E(5)^4
-1
-1
0
0
CL5
1
1
-1
-1
-1
-1
0
0
1
1
CL6
1
-1
-E(5)^2-E(5)^3
-E(5)-E(5)^4
E(5)^2+E(5)^3
E(5)+E(5)^4
-1
1
0
0
CL7
1
1
0
0
0
0
1
1
-1
-1
CL8
1
-1
0
0
0
0
1
-1
-1
1
CL9
1
-1
3
3
-3
-3
4
-4
5
-5
CL10
1
-1
-E(5)-E(5)^4
-E(5)^2-E(5)^3
E(5)+E(5)^4
E(5)^2+E(5)^3
-1
1
0
0
TABLE 31. The character table of the group AS[164].
CL1
CL2
CL3
CL4
CL5
CL6
CL7
CL8
CL9
CL10
Representative
<identity ...>
x1^-1*x2^-2*x3*x1*x2*x1
x2^-1
x2^-2
x2^-2*x1*x2
x2^2*x1*x2^2*x1^-1*x3
x2^2*x1*x2
x2^-1*x1^-1*x2^-2*x1^-1*x3
x3
x2^-1*x3
Size
1
15
12
12
15
12
20 (including g1)
20
1
12
TABLE 32. The conjugacy classes of the group AS[164].
gap> LA:=List(irrA,a->List(Elements(AS[162]),g->g^a));
[ [ 1, 1, 1, 1, 1, 1, 1, 1, 1, 1, 1, 1, 1, 1, 1, 1, 1, 1, 1, 1, 1, 1, 1, 1 ],
.............................................................................
.............................................................................
[ 5, -1, 1, -1, -1, 1, -1, 1, 1, -1, 1, -1, -5, 1, -1, 1, 1, -1, 1, -1, -1,
1, -1, 1 ] ]
gap> LB:=List(irrB,a->List(Elements(AS[162]),g->g^a));
[ [ 1, 1, 1, 1, 1, 1, 1, 1, 1, 1, 1, 1, 1, 1, 1, 1, 1, 1, 1, 1, 1, 1, 1, 1 ],
116
No.1
No.2
No.3
No.4
No.5
No.6
No.7
No.8
CL1
1
1
1
1
1
1
3
3
CL2
1
-1
-1
-1
1
1
1
-1
CL3
1
-1
-E(3)
-E(3)^2
E(3)^2
E(3)
0
0
CL4
1
1
E(3)
E(3)^2
E(3)^2
E(3)
0
0
CL5
1
1
E(3)^2
E(3)
E(3)
E(3)^2
0
0
CL6
1
-1
-E(3)^2
-E(3)
E(3)
E(3)^2
0
0
CL7
1
1
1
1
1
1
-1
-1
CL8
1
-1
-1
-1
1
1
-3
3
TABLE 33. The character table of the group AS[162].
CL1
CL2
CL3
CL4
CL5
CL6
CL7
CL8
Representative
<identity ...>
x2^-1*x1^-1*x2^-2*x3*x1*x2*x1
x1^-1*x2^-2*x1^-1*x3
x2^-1*x1*x2
x2^-2*x1
x1*x2^2*x1^-1*x2^-1*x3
x2*x1^-1
x3
Size
1
3
4
4
4
3
3
1
TABLE 34. The conjugacy classes of the group AS[162].
.............................................................................
.............................................................................
0 ], [ 3, -1, 0, 0, 0, 0, -1, -1, 0, 0, 0, 0, 3, -1, 0, 0, 0, 0, -1,
-1, 0, 0, 0, 0 ] ]
gap> LA*TransposedMat(LB)/Size(AS[162]);
The result returns as the matrix in the bellow: the irreducible representations are
ramified as
No.164 -> No.162
[ [ 1, 0, 0, 0, 0, 0, 0, 0 ],
1
-> 1
[ 0, 1, 0, 0, 0, 0, 0, 0 ],
2
-> 2
117
[ 0, 0, 0, 0, 0, 0, 0, 1 ],
3
-> 8
[ 0, 0, 0, 0, 0, 0, 0, 1 ],
4
-> 8
[ 0, 0, 0, 0, 0, 0, 1, 0 ],
5
-> 7
[ 0, 0, 0, 0, 0, 0, 1, 0 ],
6
-> 7
[ 1, 0, 0, 0, 0, 0, 0, 1 ],
7
-> 1 + 8
[ 0, 1, 0, 0, 0, 0, 1, 0 ],
8
-> 2 + 7
[ 0, 0, 0, 0, 1, 1, 0, 1 ],
9
-> 5 + 6 + 8
[ 0, 0, 1, 1, 0, 0, 1, 0 ] ]
10
-> 3 + 4 + 7
One can see the change in the geometry by checking the “orbit” of the vertex “1”
and the bond “[1,6]”. In the icosahedral symmetry, the vertex is transferred to other
sixty vertexes transitively, and this double bond is transferred to all other thirty
double bonds.
gap> List(Elements(AS[164]),g->OnPoints(1,g));
[ 1, 1, 2, 2, 3, 3, 4, 4, 5, 5, 6, 6, 7, 7, 8, 8, 9, 9, 10, 10, 11, 11, 12,
12, 13, 13, 14, 14, 15, 15, 16, 16, 17, 17, 18, 18, 19, 19, 20, 20, 21, 21,
22, 22, 23, 23, 24, 24, 25, 25, 26, 26, 27, 27, 28, 28, 29, 29, 30, 30, 31,
31, 32, 32, 33, 33, 34, 34, 35, 35, 36, 36, 37, 37, 38, 38, 39, 39, 40, 40,
41, 41, 42, 42, 43, 43, 44, 44, 45, 45, 46, 46, 47, 47, 48, 48, 49, 49, 50,
50, 51, 51, 52, 52, 53, 53, 54, 54, 55, 55, 56, 56, 57, 57, 58, 58, 59, 59,
60, 60 ]
gap> BONDA:=Unique(List(Elements(AS[164]),g->OnSets([1,6],g)));
[ [ 1, 6 ], [ 2, 11 ], [ 3, 16 ], [ 4, 21 ], [ 5, 26 ], [ 7, 30 ], [ 8, 49 ],
[ 9, 53 ], [ 10, 12 ], [ 13, 54 ], [ 14, 58 ], [ 15, 17 ], [ 18, 59 ],
[ 19, 38 ], [ 20, 22 ], [ 23, 39 ], [ 24, 43 ], [ 25, 27 ], [ 28, 44 ],
[ 29, 48 ], [ 31, 36 ], [ 32, 41 ], [ 33, 46 ], [ 34, 51 ], [ 35, 56 ],
[ 37, 60 ], [ 40, 42 ], [ 45, 47 ], [ 50, 52 ], [ 55, 57 ] ]
118
In the symmetry of the subgroup No.162, in contrast, the vertex “1” is transferred
only to 24 other vertexes. The orbit of single vertexes (which covers the whole system
in the case of the subgroup 164) splits into three subsets. Similarly the bond “[1,6]”
is not transferred transitively, only to 24 of double bonds. The remaining six double
bonds are transferred among themselves. This situation can be interpreted that
one of the double bonds (one of these connecting one pentagon with five neighbors)
slightly changes the length, while those of remaining four bonds are unchanged.
gap> List(Elements(AS[162]),g->OnPoints(1,g));
[ 1, 5, 9, 10, 11, 15, 18, 19, 21, 22, 27, 28, 31, 35, 39, 40, 41, 45, 48,
49, 51, 52, 57, 58 ]
gap> List(Elements(AS[162]),g->OnPoints(2,g));
[ 2, 4, 8, 6, 12, 14, 17, 20, 25, 23, 26, 29, 32, 34, 38, 36, 42, 44, 47, 50,
55, 53, 56, 59 ]
gap> List(Elements(AS[162]),g->OnPoints(3,g));
[ 3, 3, 7, 7, 13, 13, 16, 16, 24, 24, 30, 30, 33, 33, 37, 37, 43, 43, 46, 46,
54, 54, 60, 60 ]
gap> BONDB:=Unique(List(Elements(AS[162]),g->OnSets([1,6],g)));
gap> Sort(BONDB);
gap> BONDB;
[ [ 1, 6 ], [ 2, 11 ], [ 4, 21 ], [ 5, 26 ], [ 8, 49 ], [ 9, 53 ],
[ 10, 12 ], [ 14, 58 ], [ 15, 17 ], [ 18, 59 ], [ 19, 38 ], [ 20, 22 ],
[ 23, 39 ], [ 25, 27 ], [ 28, 44 ], [ 29, 48 ], [ 31, 36 ], [ 32, 41 ],
[ 34, 51 ], [ 35, 56 ], [ 40, 42 ], [ 45, 47 ], [ 50, 52 ], [ 55, 57 ] ]
The splitting of the energy degeneracy with the symmetry reduction is verified
numerically, for example, by a model computation, in which the matrix elements of
these two split orbits of double bonds take slightly different values, in correspon119
dence with the change in the symmetry. In this computation, the elements of the
Fock matrix which belong to the one of the split orbit (including six bonds, the complementary part of BONDB in BONDA) are enhanced slightly, multiplied by 1.005.
The energy spectra at HOMO and LUMO are computed as follows.(The positions of
HOMO and LUMO are ambiguous in this case. We presently assume that the half of
the 60 eigenstates are occupied by 60 electrons. The left and right columns show the
spectra of the full and the reduced symmetries. The HOMO and LUMO belong to
the irreducible representations of No.10 and No.5 of the character table of the group
AS[164], and the splitting of the spectra is deduced from the above relation.)
HOMO : Split from one fivefold degeneracy
to one threefold degeneracy and two single degeneracies.
26
-1.0000000
->
-1.0000021
27
-1.0000000
->
-1.0000021
28
-1.0000000
->
-1.0000021
29
-1.0000000
->
-1.0000000
30
-1.0000000
->
-1.0000000
In this case, two single states still have the same energy levels. This is due to
the fact that the two single states belong to two of the one-dimensional irreducible
representation which are complex-conjugates with each other. As the eigenvalue
problem is the real-valued one, these two eigenvalues should be equal owing to the
invariance by the complex conjugation acting on it. It could also be interpreted as
of the time-reversal symmetry in the context of the quantum physics.
LUMO: Without splitting.
31 -0.5818861
->
-0.5810887
32 -0.5818861
->
-0.5810887
33 -0.5818861
->
-0.5810887
120
The symmetry reduction from the icosahedral group to one of the subgroups of
order 24 can be realized clearly, as supposed in this computation. On the other hand,
the difference between the icosahedral group and the maximal subgroup AS[163] (the
5-th alternating group: A5) is more subtle. The splitting of the “orbits” by the group
action does not occur in the transposition of the bonds between nearest or secondnearest neighbors. The length of the single or double bonds is also unique in A5
symmetry, and the change in the symmetry could be seen in the “orbit” of the bonds
among more distant neighbors. Take the bond of [1,11].
gap> lg:=List(Elements(AS[164]),i->OnSets([1,11],i));
[ [ 1, 11 ], [ 1, 26 ], [ 2, 6 ], [ 2, 16 ], [ 3, 11 ], [ 3, 21 ], [ 4, 16 ],
[ 4, 26 ], [ 5, 6 ], [ 5, 21 ], [ 6, 30 ], [ 6, 12 ], [ 1, 7 ], [ 7, 49 ],
[ 8, 30 ], [ 8, 53 ], [ 9, 49 ], [ 9, 12 ], [ 1, 10 ], [ 10, 53 ],
[ 10, 11 ], [ 11, 17 ], [ 2, 12 ], [ 12, 54 ], [ 10, 13 ], [ 13, 58 ],
[ 14, 54 ], [ 14, 17 ], [ 2, 15 ], [ 15, 58 ], [ 15, 16 ], [ 16, 22 ],
[ 3, 17 ], [ 17, 59 ], [ 15, 18 ], [ 18, 38 ], [ 19, 59 ], [ 19, 22 ],
[ 3, 20 ], [ 20, 38 ], [ 20, 21 ], [ 21, 27 ], [ 4, 22 ], [ 22, 39 ],
[ 20, 23 ], [ 23, 43 ], [ 24, 39 ], [ 24, 27 ], [ 4, 25 ], [ 25, 43 ],
[ 25, 26 ], [ 7, 26 ], [ 5, 27 ], [ 27, 44 ], [ 25, 28 ], [ 28, 48 ],
[ 29, 44 ], [ 7, 29 ], [ 5, 30 ], [ 30, 48 ], [ 31, 41 ], [ 31, 56 ],
[ 32, 36 ], [ 32, 46 ], [ 33, 41 ], [ 33, 51 ], [ 34, 46 ], [ 34, 56 ],
[ 35, 36 ], [ 35, 51 ], [ 36, 60 ], [ 36, 42 ], [ 31, 37 ], [ 19, 37 ],
[ 38, 60 ], [ 23, 38 ], [ 19, 39 ], [ 39, 42 ], [ 31, 40 ], [ 23, 40 ],
[ 40, 41 ], [ 41, 47 ], [ 32, 42 ], [ 24, 42 ], [ 40, 43 ], [ 28, 43 ],
[ 24, 44 ], [ 44, 47 ], [ 32, 45 ], [ 28, 45 ], [ 45, 46 ], [ 46, 52 ],
[ 33, 47 ], [ 29, 47 ], [ 45, 48 ], [ 8, 48 ], [ 29, 49 ], [ 49, 52 ],
[ 33, 50 ], [ 8, 50 ], [ 50, 51 ], [ 51, 57 ], [ 34, 52 ], [ 9, 52 ],
121
[ 50, 53 ], [ 13, 53 ], [ 9, 54 ], [ 54, 57 ], [ 34, 55 ], [ 13, 55 ],
[ 55, 56 ], [ 37, 56 ], [ 35, 57 ], [ 14, 57 ], [ 55, 58 ], [ 18, 58 ],
[ 14, 59 ], [ 37, 59 ], [ 35, 60 ], [ 18, 60 ] ]
On the other hand, the “orbit” by A5 includes only 60 bonds.
gap> la5:=List(Elements(AS[163]),i->OnSets([1,11],i));
[ [ 1, 11 ], [ 2, 16 ], [ 3, 21 ], [ 4, 26 ], [ 5, 6 ], [ 6, 30 ], [ 7, 49 ],
[ 8, 53 ], [ 9, 12 ], [ 1, 10 ], [ 10, 11 ], [ 12, 54 ], [ 13, 58 ],
[ 14, 17 ], [ 2, 15 ], [ 15, 16 ], [ 17, 59 ], [ 18, 38 ], [ 19, 22 ],
[ 3, 20 ], [ 20, 21 ], [ 22, 39 ], [ 23, 43 ], [ 24, 27 ], [ 4, 25 ],
[ 25, 26 ], [ 27, 44 ], [ 28, 48 ], [ 7, 29 ], [ 5, 30 ], [ 31, 56 ],
[ 32, 36 ], [ 33, 41 ], [ 34, 46 ], [ 35, 51 ], [ 36, 42 ], [ 31, 37 ],
[ 38, 60 ], [ 19, 39 ], [ 23, 40 ], [ 41, 47 ], [ 32, 42 ], [ 40, 43 ],
[ 24, 44 ], [ 28, 45 ], [ 46, 52 ], [ 33, 47 ], [ 45, 48 ], [ 29, 49 ],
[ 8, 50 ], [ 51, 57 ], [ 34, 52 ], [ 50, 53 ], [ 9, 54 ], [ 13, 55 ],
[ 37, 56 ], [ 35, 57 ], [ 55, 58 ], [ 14, 59 ], [ 18, 60 ] ]
The remaining 60 bonds are located in the separated orbit now.
gap> la5_2:=List(Elements(AS[163]),i->OnSets([2,6],i));
[ [ 2, 6 ], [ 3, 11 ], [ 4, 16 ], [ 5, 21 ], [ 1, 26 ], [ 1, 7 ], [ 8, 30 ],
[ 9, 49 ], [ 10, 53 ], [ 6, 12 ], [ 2, 12 ], [ 10, 13 ], [ 14, 54 ],
[ 15, 58 ], [ 11, 17 ], [ 3, 17 ], [ 15, 18 ], [ 19, 59 ], [ 20, 38 ],
[ 16, 22 ], [ 4, 22 ], [ 20, 23 ], [ 24, 39 ], [ 25, 43 ], [ 21, 27 ],
[ 5, 27 ], [ 25, 28 ], [ 29, 44 ], [ 30, 48 ], [ 7, 26 ], [ 35, 36 ],
[ 31, 41 ], [ 32, 46 ], [ 33, 51 ], [ 34, 56 ], [ 31, 40 ], [ 36, 60 ],
[ 19, 37 ], [ 23, 38 ], [ 39, 42 ], [ 32, 45 ], [ 40, 41 ], [ 24, 42 ],
[ 28, 43 ], [ 44, 47 ], [ 33, 50 ], [ 45, 46 ], [ 29, 47 ], [ 8, 48 ],
122
[ 49, 52 ], [ 34, 55 ], [ 50, 51 ], [ 9, 52 ], [ 13, 53 ], [ 54, 57 ],
[ 35, 60 ], [ 55, 56 ], [ 14, 57 ], [ 18, 58 ], [ 37, 59 ] ]
The bonds, equivalent to “[1,11]”, with the unique length in the icosahedral symmetry, are permitted to have two different lengths in the A5 symmetry. The deformation from the icosahedral symmetry to that of A5 is realized not in the bond
structures between the nearest or the second-nearest neighbors, but those between
remoter atomic sites. By such deformation into A5 symmetry, the two pentagons,
located in the antipodes positions, can rotate themselves in the opposite directions
with each other.
G.
The symmetry in the eigenstates (hidden one)
The irreducible representations of the same dimension show a certain similarity
in the icosahedral group. For example, in table 26, the representations of I.3 and I.4
or I.5 and I.6 seem to be interchangeable, if we replace E(5), E(5)^2, E(5)^3, E(5)^4
among themselves. This suggests a certain kind of affinity of the basis functions
of the wave functions. This is the symmetry in the context of the Galois theory,
especially of the cyclotomic field, and manifest itself in the secular equation of the
energy spectrum.
I.3
I.4
I.5
I.6
CL1
3
3
3
3
CL2
-E(5)-E(5)^4
-E(5)^2-E(5)^3
E(5)^2+E(5)^3
E(5)+E(5)^4
CL3
-E(5)^2-E(5)^3
-E(5)-E(5)^4
-E(5)-E(5)^4
-E(5)^2-E(5)^3
CL4
0
0
0
0
CL5
0
0
0
0
CL6
-1
-1
1
1
CL7
-E(5)^2-E(5)^3
-E(5)-E(5)^4
E(5)+E(5)^4
E(5)^2+E(5)^3
CL8
3
3
-3
-3
CL9
CL10
-E(5)-E(5)^4
-1
-E(5)^2-E(5)^3 -1
-E(5)^2-E(5)^3 -1
-E(5)-E(5)^4
-1
TABLE 35. The three dimensional irreducible representations in the icosahedral group.
Let us start again at the model computation of the electronic structure of C60 . The
eigenvalue problem is solved by one of the functions in the GAP package. By this,
123
the secular equation is decomposed into polynomials with coefficients over the specified field, and the “generalized” eigenstates for these polynomials are obtained.(The
eigenvalues are computed as the set of polynomials, each of which has a part of
eigenvalues as the roots. The eigenstates of the matrix H are the solutions of the
matrix equation p(H)·x=0, where p(t) is one of these polynomials.)
Let “hh” be the Fock matrix, used in the model computation of the electronic
structure of C60 , with the same parameter as in the previous section. The generators
of the group g1,g2,g3 are the same as defined previously. The eigenvalue problem is
solved now. The last part of the following computation is to extract the basis set of
the eigenspaces. (The secular equation is representable as the product of polynomials
equipped the parameter for the transfer integrals. See the appendix H.)
Hgen:=function(GP,a)
local i,h,tuples;
h:=List([1..60],i->List([1..60],j->0));
tuples:=Orbit(GP,a,OnTuples);
for i in tuples
do
h[i[1]][i[2]]:=1;
od;
return h;
end;
This function sets non-zero matrix elements hij , when the symmetry of the molecule
demand the bond between site “i” and “j”. To do this, the bond defined in the
list “a”, in the argument of this function, is transferred to equivalent bonds in the
symmetry of the group “GP”.
gap> G:=Group(g1,g2,g3);
124
No. Polynomial
1
2
3
4
5
6
7
8
Eigenspace
dimension
x_1-3
5
x_1+1
5
x_1+5
1
x_1^2-4*x_1-1
6
x_1^2-2*x_1-9
8
x_1^2-2*x_1-5
8
x_1^3+3*x_1^2-9*x_1-19
15
x_1^4+6*x_1^3-4*x_1^2-54*x_1-29 12
Degeneracy
of one eigenvalue
5
5
1
3
4
4
3
3
Irreducible
representation
I.10
I.10
I.1
I.3,I.4
I.8
I.7
I.9
I.5,I.6
TABLE 36. The classification of the solutions of the generalized eigenvalue problem. The
irreducible representations are those presented in table 26.
gap> h1:=Hgen(G,[1,2]);
gap> h2:=Hgen(G,[1,6]);
gap> hh:=-2*h1-h2;
gap> ge:=GeneralizedEigenvalues(Rationals,hh);
[ x_1-3,
x_1+1,
x_1^2-2*x_1-5,
x_1+5,
x_1^2-4*x_1-1,
x_1^2-2*x_1-9,
x_1^3+3*x_1^2-9*x_1-19,
x_1^4+6*x_1^3-4*x_1^2-54*x_1-29 ]
gap> gs:=GeneralizedEigenspaces(Rationals,hh);;
gap> gsbas:=List(gs,i->Basis(i));;
The eight polynomials of the eigenvalues, decomposed up to the field of rational
number are given as the table 36, and the basis vectors in the eigenspace are allotted
into the irreducible representations.
The representations of I.3 and I.4 or I.5 and I.6 belong to the same eigenspace,
specified by the same polynomial.
The automorphism of the cyclotomic field
(E(5)+E(5)^4→E(5)^2 + E(5)^3), which forms a group, interchanges these pairs
of the irreducible representation, and consequently, also interchanges the vector solutions x of the equation p(H)·x=0. From the viewpoint of the Galois theory, Let
125
us review the situation. Take the last polynomial of the four degrees. The 8-th
polynomial in the above table has C4 as the Galois group, and the four roots are
transported with each other by the operation of this group( as 1 → 2 → 3 → 4 → 1,
by a certain sequential order of the roots.)
Remark. The Galois group of the polynomial is computed by GAP.
gap> ge[8];
x_1^4+6*x_1^3-4*x_1^2-54*x_1-29
gap> GaloisType(ge[8]);
1
gap> TransitiveGroup(4,1);
C(4) = 4
The command Galoistype(f ) computes the Galois Group of the polynomial f, and
returns an integer i if the Galois group is isomorphic to TransitiveGroup(n, i), where
n is the degree of f.
Let us represent this Galois group concretely as the operation of the eigenvectors.
The polynomial splits in four polynomials of one degree in the 15-th cyclotomic
field Q15 . The generalized eigenvalue problem is solved again in this field, for which
CF(15) is the notation in GAP.
gap> ge2:=GeneralizedEigenvalues(CF(15),hh);
[ x_1-3, x_1+1, x_1+5, x_1+(E(5)+3*E(5)^2+3*E(5)^3+E(5)^4),
x_1+(3*E(5)+E(5)^2+E(5)^3+3*E(5)^4),
x_1+(E(15)+E(15)^2+E(15)^4+3*E(15)^7+3*E(15)^8+E(15)^11+E(15)^13+E(15)^14),
x_1+(E(15)+E(15)^2+3*E(15)^4+E(15)^7+E(15)^8+3*E(15)^11+E(15)^13+E(15)^14),
x_1+(E(15)+3*E(15)^2+E(15)^4+E(15)^7+E(15)^8+E(15)^11+3*E(15)^13+E(15)^14),
x_1+(3*E(15)+E(15)^2+E(15)^4+E(15)^7+E(15)^8+E(15)^11+E(15)^13+3*E(15)^14),
126
No. Polynomial
1
2
3
4
5
6
x_1-3
x_1+1
x_1+5
x_1+(E(5)+3*E(5)^2+3*E(5)^3+E(5)^4)
x_1+(3*E(5)+E(5)^2+E(5)^3+3*E(5)^4)
x_1+(E(15)+E(15)^2+E(15)^4+3*E(15)^7
+3*E(15)^8+E(15)^11+E(15)^13+E(15)^14)
7 x_1+(E(15)+E(15)^2+3*E(15)^4+E(15)^7
+E(15)^8+3*E(15)^11+E(15)^13+E(15)^14)
8 x_1+(E(15)+3*E(15)^2+E(15)^4+E(15)^7
+E(15)^8+E(15)^11+3*E(15)^13+E(15)^14)
9 x_1+(3*E(15)+E(15)^2+E(15)^4+E(15)^7
+E(15)^8+E(15)^11+E(15)^13+3*E(15)^14)
10 x_1^2-2*x_1-9
11 x_1^2-2*x_1-5
12 x_1^3+3*x_1^2-9*x_1-19
Eigenspace
dimension
5
5
1
3
3
3
Irreducible
representation
10
10
1
3
4
6
3
5
3
6
3
5
8
8
15
8
7
9
TABLE 37. The classification of the solutions of the generalized eigenvalue problem in
Q15 .
x_1^2-2*x_1-9, x_1^2-2*x_1-5, x_1^3+3*x_1^2-9*x_1-19 ]
gap> gs2:=GeneralizedEigenspaces(CF(15),hh);;
gap> gsbas2:=List(gs2,i->Basis(i));;
Remark. In fact, the splitting field of the polynomial x_1^4+6*x_1^3-4*x_1^2-54*x_1-29
is not Q15 ( of the extension of eight degrees to Q), but Q(a) (of the extension of four
degrees to Q), the latter of which is obtained by adding one of the solution “a” of
this polynomial to Q. It is equal to Q(E(15)+E(15)^14) (denoted as NF(15,[1,14])
in GAP). To be exact, one should discuss in this field.
The Galois group of Q15 is computed as.
gap> GaloisGroup(CF(15));
127
<group with 2 generators>
gap> g:=GaloisGroup(CF(15));
<group with 2 generators>
gap> g.1;
ANFAutomorphism( CF(15), 11 )
gap> g.2;
ANFAutomorphism( CF(15), 7 )
The generator g.1 is the map, replacing E(15)→ E(15)^11, of order 2, and g.2,
E(15)→ E(15)^7, of order 4. The latter generator, in table 37, transports the 6th eigenspace to the 7-th, the 7-th to the 8-th, in turn, and the 10-th to the 6-th,
generating the group C4 ( the forth cyclic group.), realizing the Galois group of
the polynomial x_1^4+6*x_1^3-4*x_1^2-54*x_1-29 as the following computation
verifies.
gap> ge2[6]*ge2[7]*ge2[8]*ge2[9]=ge[8];
true
gap> for j in [1..3] do;
> Print(RootsOfPolynomial(ge2[6])[1]^g.2=RootsOfPolynomial(ge2[7])[1]);
> Print(RootsOfPolynomial(ge2[7])[1]^g.2=RootsOfPolynomial(ge2[8])[1]);
> Print(RootsOfPolynomial(ge2[8])[1]^g.2=RootsOfPolynomial(ge2[9])[1]);
> Print(RootsOfPolynomial(ge2[9])[1]^g.2=RootsOfPolynomial(ge2[6])[1]);
>od;
truetruetruetruetruetruetruetruetruetruetruetrue
gap> for j in [1..3] do;
> Print(List(gsbas2[6][j],x->x^g.2)=gsbas2[7][j]);
128
> Print(List(gsbas2[7][j],x->x^g.2)=gsbas2[8][j]);
> Print(List(gsbas2[8][j],x->x^g.2)=gsbas2[9][j]);
> Print(List(gsbas2[9][j],x->x^g.2)=gsbas2[6][j]);
>od;
truetruetruetruetruetruetruetruetruetruetruetrue
The first part of the above computation shows that the polynomial ge[8] (the 8-th
solution in the field of rational numbers) splits into ge2[6],...,ge2[9], (the four solutions, from the 6-th to 9-th, in Q15 ). The second part verifies the C4 group property
in the transportation of the eigenvalues (the root of ge2[6],...,ge2[9]) , and the
third part verifies it in the transportation of the eigenvectors, gsbas2[6],...,gsbas2[9],
each of which includes three basis (gsbas2[X][1],gsbas2[X][2],gsbas2[X][3]).
The irreducible representations are transported as the following sequence by g.2,
I.5 → I.6 → I.5 → I.6. In this sequence, the same types of the irreducible representations appear two times, but due to the duplicity of these representations in
the whole eigenspace (as we have seen the decomposition of the trace representation I.Tr in the previous section) they are two distinct subspaces which belong to
the same type of the irreducible representation. Therefore the Galois group has
the order 4, running through the four distinct subspaces. The replacement of the
eigenspaces by this group is given in table38, from which one could extract the Galois
correspondence between sub-fields and subgroup.
This example may be regarded as a hidden symmetry in the sense that the wave
functions are transformed into each other by the operations of a group, which is
different from that of the structural symmetry of the molecule.
We can approach the hidden symmetry in another direction: there is an automorphism on the icosahedral group which interchanges these two irreducible representations. For simplicity, we reduce the problem in A5. (The icosahedral group is
129
<identity ...> 1 2 3 4 5 6 7
x2*x1
1235496
x2^2
1234589
x2
1235478
x2^-1*x1
1235478
x1
1234589
x2^-1
1235496
x2^2*x1
1234567
8
7
6
9
9
6
7
8
9
8
7
6
6
7
8
9
10
10
10
10
10
10
10
10
11
11
11
11
11
11
11
11
12
12
12
12
12
12
12
12
TABLE 38. The replacement of 12 eigenspaces by the automorphism in CF(15). The x1
and x2 are the generators of the Galois group, given as g.1 and g.2 in the article.
constructed from this by the direct product with the inversion.) The generators of
A5 are g1 and g2, (given as the permutations in the previous section.)
gap> G:=Group(g1,g2);;
gap> Irr(G);;
The character table is as follows (written briefly by a list in GAP).
[[ 1, 1, 1, 1, 1 ],
[ 3, -E(5)-E(5)^4, -E(5)^2-E(5)^3, -1, 0 ],
[ 3, -E(5)^2-E(5)^3, -E(5)-E(5)^4, -1, 0 ],
[ 4, -1, -1, 0, 1 ],
[ 5, 0, 0, 1, -1 ] ]
The conjugacy classes are computed now. (As the list of the classes is given as
the permutations, it is transformed to the representations by generators, to make the
list as brief as possible. The symbols x1 and x2 represent g1 and g2 respectively.)
gap> A:=List(ConjugacyClasses(G),Elements);;
gap> B:=List(A,i->List(i,j->Factorization(G,j)));
[ [ <identity ...> ],
130
[ x2^-1, x2, x2*x1*x2, x1*x2, x2^2*x1, x2*x1, x2^-1*x1^-1, x2^-2*x1^-1,
x2^-1*x1^-1*x2^-1, x1^-1*x2^-1, x1*x2^2, x1^-1*x2^-2 ],
[ x2^-2, x2^2, x2^2*x1*x2^2*x1^-1, x1*x2^2*x1^-1, (x1*x2^2)^2,
(x2^-1*x1^-1*x2^-1)^2, (x2*x1*x2)^2, x2^-1*x1^-1*x2^-2*x1*x2,
(x2^2*x1)^2, x1^-1*x2^-2*x1, (x2*x1)^2, x2^-1*x1*x2*x1 ],
[ x2^-2*x1*x2, x2^-1*x1, x2*x1^-1, x2^2*x1^-1*x2^-1, x2^2*x1*x2^2,
x1^-1*x2^-2*x1*x2*x1, x1*x2*(x2*x1)^2, (x2*x1*x2)^2*x1,
(x2^2*x1)^2*x2*x1, x2^-1*x1^-1*x2^-2*x1*x2*x1, x2*x1*x2^2*x1^-1,
(x2*x1*x2)^2*x2, (x2^2*x1)^2*x2, x2^-1*x1^-1*x2^-2*x1, x1*x2*x1 ],
[ x2^2*x1*x2, x2^-1*x1*x2, x2^-2*x1, x1, x2^2*x1^-1, x1^-1,
x2^-2*x1^-1*x2^-1, x2*x1^-1*x2^-1, x2^-1*x1^-1*x2^-2, x2*x1*x2^2,
x2^-1*x1^-1*x2^-2*x1^-1, x1^-1*x2^-2*x1^-1, x1^-1*x2^-2*x1^-1*x2^-1,
x1*x2^2*x1^-1*x2^-1, x1*x2^2*x1*x2, x1^-1*x2^-2*x1*x2, x1*x2^2*x1,
x2*x1*x2^2*x1, x2*(x2*x1)^2, x2^-2*x1*x2*x1 ] ]
The group of all automorphism is computed as
gap> AUALL:=AutomorphismGroup(G);
Take first of the generator, AUALL.1.
gap> AUALL.1;
[ (1,14,20)(2,15,16)(3,11,17)(4,12,18)(5,13,19)(6,58,22)(7,57,23)
(8,56,24)(9,60,25)(10,59,21)(26,54,38)(27,53,37)(28,52,36)(29,51,40)
(30,55,39)(31,44,50)(32,45,46)(33,41,47)(34,42,48)(35,43,49),
(1,2,3,4,5)(6,11,16,21,26)(7,12,17,22,27)(8,13,18,23,28)(9,14,19,24,29)
(10,15,20,25,30)(31,32,33,34,35)(36,41,46,51,56)(37,42,47,52,57)
(38,43,48,53,58)(39,44,49,54,59)(40,45,50,55,60) ] ->
[ (1,57,45)(2,56,44)(3,60,43)(4,59,42)(5,58,41)(6,55,47)(7,54,46)(8,53,50)
131
(9,52,49)(10,51,48)(11,35,28)(12,34,29)(13,33,30)(14,32,26)
(15,31,27)(16,37,24)(17,36,25)(18,40,21)(19,39,22)(20,38,23),
(1,59,9,16,55)(2,58,10,17,54)(3,57,6,18,53)(4,56,7,19,52)
(5,60,8,20,51)(11,14,12,15,13)(21,35,30,38,50)(22,34,26,37,49)
(23,33,27,36,48)(24,32,28,40,47)(25,31,29,39,46)(41,44,42,45,43) ]
This automorphism maps g1 and g2 in the following way.
gap> Factorization(G,Image(AUALL.1,g1));
x2*(x2*x1)^2
gap> Factorization(G,Image(AUALL.1,g2));
x1^-1*x2^-2*x1
The two conjugacy classes (of the second and of the third) are interchanged with
each other, while the other classes are unchanged. This causes the exchange of the
character values and that of the two irreducible representation (of No.2 and No.3).
Let us verify this. The image of the automorphism on the elements of the conjugacy
class 2 is computed and sorted as
gap> imgCl2:=List(A[2], a ->Image(AUALL.1,a));;
gap> SortedList(List(imgCl2,i->Factorization(G,i)));
[ x2^-2, x2^2, x1^-1*x2^-2*x1, x1*x2^2*x1^-1, x2^-1*x1*x2*x1, (x2*x1)^2,
(x1*x2^2)^2, (x2^-1*x1^-1*x2^-1)^2, x2^-1*x1^-1*x2^-2*x1*x2, (x2*x1*x2)^2,
x2^2*x1*x2^2*x1^-1, (x2^2*x1)^2 ]
The sorted list of the conjugacy class 3 is:
gap> SortedList(B[3]);
[ x2^-2, x2^2, x1^-1*x2^-2*x1, x1*x2^2*x1^-1, x2^-1*x1*x2*x1, (x2*x1)^2,
(x1*x2^2)^2, (x2^-1*x1^-1*x2^-1)^2, x2^-1*x1^-1*x2^-2*x1*x2, (x2*x1*x2)^2,
x2^2*x1*x2^2*x1^-1, (x2^2*x1)^2 ]
132
These two lists are identical. As the automorphism AUALL.1 has the order 2, the
conjugacy classes 2 and 3, or the irreducible representation (of No.2 and No.3), are
interchanged with each other by this automorphism.
As the each element of the group A5 has a correspondence with one of the vertexes
of the C60 molecule, this automorphism can be interpreted as a geometrical operation,
that of the replacement of vertex. But this replacement of atoms does not conserve
the bonding structure between nearest neighbors. This means that this replacement
of atoms is not the symmetry operation on the C60 .
H.
The super-symmetries
In this section, we will see that we can construct a symmetry higher than the
icosahedral one. The icosahedral group is the direct product of A5 and I. If one of
the components, the group A5, is extended to a larger group, the direct product is a
super-group of the icosahedral one. The bonding structure of the C60 is representable
by the pairs of the sequential numbers of the two vertexes, and it can be interpreted
as the “orbit” of these pairs by the action of the icosahedral group. Then the orbit
of the pairs of vertex, generated by the action of this new larger group, provides us
with the bonding structure in the higher symmetry. The guiding principle is based
the following observation: in the previous section, we have seen the existence of a
certain group automorphism which interchanges two of the irreducible representation.
If a group, including A5 and this group automorphism, could be constructed, the
two three-dimensional irreducible representations will be merged into one. And if
a molecule obeys this higher symmetry, the electronic structure will show different
features, such as sixfold degeneracies. Now, concerning the extension of a group, the
concept of the semi-direct product (with the automorphism group) is usable. For this
purpose, we should construct a group composed from all of the automorphisms on
133
the group A5. Taking this automorphism group or its subgroups, we can construct
the semi-direct product with the group A5. The necessary computation is executed
in the following way. As the full automorphism group of A5 is too large, we take a
small subgroup of this (a subgroup generated by one of the generators of this group).
gap> G:=Group(g1,g2);
gap> AUALL:=AutomorphismGroup(G);
gap> p:=SemidirectProduct(Group(AUALL.1),G);
The newly created group “p” is larger than A5 (the group “G”), and contains A5
as a subgroup. The three generators are given as
gap> p.1;
(3,16,22)(4,17,18)(5,13,19)(6,14,20)(7,15,21)(8,60,24)(9,59,25)(10,58,26)
(11,62,27)(12,61,23)(28,56,40)(29,55,39)(30,54,38)(31,53,42)(32,57,41)
(33,46,52)(34,47,48)(35,43,49)(36,44,50)(37,45,51)
gap> p.2;
(3,7,6,5,4)(8,28,23,18,13)(9,29,24,19,14)(10,30,25,20,15)(11,31,26,21,16)(12,
32,27,22,17)(33,37,36,35,34)(38,58,53,48,43)(39,59,54,49,44)(40,60,55,50,
45)(41,61,56,51,46)(42,62,57,52,47)
gap> p.3;
(1,2)(4,61)(5,11)(6,18)(7,57)(8,21)(9,28)(10,42)(12,46)(13,54)(14,39)
(15,49)(16,59)(17,44)(19,51)(20,26)(22,47)(23,58)(25,34)(27,45)(29,36)
(31,55)(32,43)(33,50)(37,38)(40,48)(41,60)(52,53)
The first and the second generators are inherited from A5, and the third is the new
one. They are represented by 62 symbols. The symbols from 3 to 62 correspond to
the 60 vertexes in one C60 . The extension of the group introduces extra symbols “1”
and “2”. We can give a geometrical interpretation to this group in the following way.
134
From the 62 symbols, we can construct a bonding system which contains 120 atoms,
located in the double C60 molecules. The atoms will be indexed as [i,j] (i=3,..,62,
and j=1,2) by the duplication of single C60 . This structure (C120 ) is regulated by the
symmetry of the new extended group. The operation p.1 and p.2 are the rotations
by 120 and 72 degrees. The operation p.3 interchanges 60 vertexes between two C60
shells. To see this, we have only to compute the orbit of the vertex at [I, J] by
these operations. The permutation (1,2), applied to the latter element of the pairs,
represents the exchange of vertexes between the two C60 .
By means of a function defined below, the bonding structure in this new symmetry
can be set up in matrices (of 120 × 120 dimension). If one bond is bridged between
the vertex [i[1],i[2]] and [i[3],i[4]], the corresponding element in the networking matrix
is set to be the unity.
listforh:=function(A)
local h,i1,i2,i3,i4,k,l,i;
h:=List([1..120],i->List([1..120],j->0));
for i in A do
i1:=i[1];
i2:=i[2];
i3:=i[3];
i4:=i[4];
k:=(i1-2)+60*(i2-1);
l:=(i3-2)+60*(i4-1);
h[k][l]:=1;
h[l][k]:=1;
od;
return h;
135
end;
The orbits of the double bond [3,1,4,1] (read as from [3,1] to [4,1]), the single
bond [3,1,8,1], and the bond between the first and the second C60 ( such as [3,1,4,2],
a possible example, taken par hazard) are computed into the lists “A1”,“B1”,“C1”
as bellow. The bonding structures generated from them are represented by three
matrices “a1”,“b1”,“c1”.
gap> A1:=Orbit(p,[3,1,4,1],OnTuples);
[ [ 3, 1, 4, 1 ], [ 16, 1, 17, 1 ], [ 4, 1, 5, 1 ], [ 3, 2, 61, 2 ],
[ 22, 1, 18, 1 ], [ 21, 1, 22, 1 ], [ 59, 2, 44, 2 ], [ 17, 1, 13, 1 ],
[ 5, 1, 6, 1 ], [ 61, 2, 11, 2 ], [ 16, 2, 23, 2 ], [ 4, 2, 41, 2 ],
...(omitted hereafter)...
gap> B1:=Orbit(p,[3,1,8,1],OnTuples);;
gap> C1:=Orbit(p,[3,1,4,2],OnTuples);;
gap> a1:=listforh(A1);;
gap> b1:=listforh(B1);;
gap> c1:=listforh(C1);;
In the C60 , the symmetry operation of the icosahedral symmetry includes the
inversion. The inversion itself is not included in the new group, but we can furnish
this operation with first sixty vertexes (from the third to the sixty-second) in the
same manner as the icosahedral case. And the inversion in the second 60 vertexes can
be written in the following way. The second sixty vertex are transported from the
first sixty, by the operation p.3. If the inversions on the first and second 60 vertexes
are denoted as I1, I2, they are related by the conjugation I2 := (p.3) · I1 · (p.3)−1
(symbolically written here). The inversion on the total system is written by their
136
union as
I1 0
0 I2
where the each entry is assumed to be 60×60-dimensional matrices.
The destinations of the vertexes by the inversion “g3” in one C60 shell are stored
in the list bellow.
gap> ListInv:=List([1..60],i->OnPoints(i,g3));
The destinations of the vertexes in the first C60 shell to the second one, by the
operation “p.3”, is stored in the list bellow.
gap> mconjg:=List([3..62],i->OnPoints(i,p.3))-2;
(In the last computation, we must extract 2 from the result, due to the numbering
in the atoms, which are numbered from 3 to 62.)
The inversions in the first and the second C60 shell are composed now (as matrices
“invrs”, “invrs2”, and by uniting them, the inversion in the full system is computed,
as in the matrix “ivfull”.
gap> invrs:=List([1..60],i->List([1..60],j->0));
gap> mc:=List([1..60],i->List([1..60],j->0));
gap> for i in [1..60] do
invrs[i][ListInv[i]]:=1;
mc[i][mconjg[i]]:=1;
od;
gap> invrs2:=mc*invrs*mc^-1;
gap> ivfull:=List([1..120],i->List([1..120],j->0));
gap> for i in [1..60] do
137
for j in [1..60] do
ivfull[i][j]:=invrs[i][j];
ivfull[i+60][j+60]:=invrs2[i][j];
od;
od;
The generators of the group “p”, as the symmetry operations of the 120 atoms,
are also represented by matrices. At first, we prepare the lists to represent the
permutation.
gap> l1:=List([1..2],j->List([3..62],i->(OnPoints(i,p.1)+60*(j-1))));
gap> l1:=Flat(l1);
gap> l2:=List([1..2],j->List([3..62],i->(OnPoints(i,p.2)+60*(j-1))));
gap> l2:=Flat(l2);
gap> l3:=List([1..2],j->List([3..62],i->(OnPoints(i,p.3)+60*(OnPoints(j,p.3)-1))));
gap> l3:=Flat(l3);
The permutations are transformed to the matrix forms by this function.
ListToMat:=function(l)
local size,h,i,j;
size:=Size(l);
h:=List([1..size],i->List([1..size],j->0));
for i in [1..size] do
h[i][l[i]]:=1;
od;
return h;
end;
138
gap> P1:=ListToMat(l1-2);
gap> P2:=ListToMat(l2-2);
gap> P3:=ListToMat(l3-2);
We must examine the congruence of the computed data, by checking whether
the matrices representing the bonding structure are subject to the symmetry of the
extended group. For this purpose, we can compute the commutators between these
matrices (“a1”,“b1”,“c1”) and the generators of this group (P1, P2, P3 and ivfull):
the results should be zero. The matrices “a1” and “b1” pass the test, but “c1”
not. As the matrix “c1” is not conserved by the matrix of the inversion “ivfull”
(or not commutable), the symmetrization is necessary for the construction of the
Fock matrix. One of the Fock matrices is given as, fixing the values of the transfer
integrals,
gap> hh:=-2*a1-b1-1/2*(c1+ivfull*c1*ivfull);
The generalized eigenvalue problems is solved up to in the field of rational numbers.
gap> gev:=GeneralisedEigenvalues(Rationals,hh);
[ x_1-3, x_1+1, x_1+3, x_1+7, x_1^2-4*x_1-13/4, x_1^2-4*x_1-2,
x_1^2-2*x_1-81/4, x_1^2-2*x_1-9/4, x_1^2-12, x_1^2-21/4,
x_1^3+5*x_1^2-4*x_1-24, x_1^4+6*x_1^3-6*x_1^2-60*x_1-36 ]
gap> ges:=GeneralisedEigenspaces(Rationals,hh);
[ <vector space over Rationals, with 10 generators>,
<vector space over Rationals, with 15 generators>,
<vector space over Rationals, with 1 generators>,
<vector space over Rationals, with 1 generators>,
<vector space over Rationals, with 8 generators>,
139
<vector space over Rationals, with 12 generators>,
<vector space over Rationals, with 8 generators>,
<vector space over Rationals, with 8 generators>,
<vector space over Rationals, with 10 generators>,
<vector space over Rationals, with 8 generators>,
<vector space over Rationals, with 15 generators>,
<vector space over Rationals, with 24 generators> ]
The group is composed and the irreducible representation is computed now.
gap> irrg:=Irr(Group(P1,P2,P3,ivfull));
[ Character( CharacterTable( <matrix group of size 240 with 4 generators> ),
[ 1, 1, 1, 1, 1, 1, 1, 1, 1, 1, 1, 1, 1, 1 ] ),
Character( CharacterTable( <matrix group of size 240 with 4 generators> ),
[ 1, -1, -1, -1, -1, -1, 1, 1, 1, 1, 1, 1, 1, -1 ] ),
Character( CharacterTable( <matrix group of size 240 with 4 generators> ),
[ 1, -1, -1, -1, 1, 1, 1, -1, 1, -1, 1, -1, -1, 1 ] ),
Character( CharacterTable( <matrix group of size 240 with 4 generators> ),
[ 1, 1, 1, 1, -1, -1, 1, -1, 1, -1, 1, -1, -1, -1 ] ),
Character( CharacterTable( <matrix group of size 240 with 4 generators> ),
[ 4, -2, 1, 0, 1, -2, 0, -1, -1, 1, 1, 0, 4, 0 ] ),
Character( CharacterTable( <matrix group of size 240 with 4 generators> ),
[ 4, 2, -1, 0, -1, 2, 0, -1, -1, 1, 1, 0, 4, 0 ] ),
Character( CharacterTable( <matrix group of size 240 with 4 generators> ),
[ 4, -2, 1, 0, -1, 2, 0, 1, -1, -1, 1, 0, -4, 0 ] ),
Character( CharacterTable( <matrix group of size 240 with 4 generators> ),
[ 4, 2, -1, 0, 1, -2, 0, 1, -1, -1, 1, 0, -4, 0 ] ),
Character( CharacterTable( <matrix group of size 240 with 4 generators> ),
140
[ 5, 1, 1, -1, 1, 1, 1, 0, 0, -1, -1, 1, 5, -1 ] ),
Character( CharacterTable( <matrix group of size 240 with 4 generators> ),
[ 5, -1, -1, 1, -1, -1, 1, 0, 0, -1, -1, 1, 5, 1 ] ),
Character( CharacterTable( <matrix group of size 240 with 4 generators> ),
[ 5, 1, 1, -1, -1, -1, 1, 0, 0, 1, -1, -1, -5, 1 ] ),
Character( CharacterTable( <matrix group of size 240 with 4 generators> ),
[ 5, -1, -1, 1, 1, 1, 1, 0, 0, 1, -1, -1, -5, -1 ] ),
Character( CharacterTable( <matrix group of size 240 with 4 generators> ),
[ 6, 0, 0, 0, 0, 0, -2, 1, 1, 0, 0, -2, 6, 0 ] ),
Character( CharacterTable( <matrix group of size 240 with 4 generators> ),
[ 6, 0, 0, 0, 0, 0, -2, -1, 1, 0, 0, 2, -6, 0 ] ) ]
Solution
x_1-3
x_1+1
x_1+3
x_1+7
x_1^2-4*x_1-13/4
x_1^2-4*x_1-2
x_1^2-2*x_1-81/4
x_1^2-2*x_1-9/4
x_1^2-12
x_1^2-21/4
x_1^3+5*x_1^2-4*x_1-24
x_1^4+6*x_1^3-6*x_1^2-60*x_1-36
Representation
[ 11, 12 ]
[ 10, 11, 12 ]
[2]
[1]
[6]
[ 13 ]
[8]
[7]
[ 10 ]
[5]
[9]
[ 14 ]
Dimension
5
5
1
1
4
6
4
4
5
4
5
6
TABLE 39. The correspondence between the solutions of the generalized eigenvalue problem and the irreducible representations. The left column shows the polynomial, the center
the corresponding irreducible representation by sequential numbers, and the right the dimension of the irreducible representation.
The four eigenvalues, as the solutions of “x_1^4+6*x_1^3-6*x_1^2-60*x_1-36”,
composed by the <vector space over Rationals, with 24 generators>, are of
141
the sixfold degeneracies, allotted to the 14th irreducible representation, which we
have expected to see in this structure. Although this example is rather an imaginary one, it suggests us the importance of the concept of “group extension” in the
artificial design of the electronic structure in molecular systems of high symmetries.
Usually, we argue the electronic structure from the viewpoint of the splitting of the
degeneracies in accordance with the reduction of the symmetry, but we will be able
to proceed in the inverse direction by actually composing the “super-group”.
I.
Remarks to this section
In this section, with the viewpoint of the group theory, we analyze the electronic
properties of C60 . The group theoretical properties of the icosahedral symmetry of
C60 are computed by a computer algebra program package GAP, and the obtained
results are utilized to the analysis. We started in a traditional way in order to classify
the eigenstates into irreducible representations. However, the possible applications of
the computer algebra are not limited to this sort of the rather classical analysis, since
we can construct the group and extract the group theoretical properties more freely
and adroitly by the computer algebra than by the utilization of the group theoretical
data given in some references. We can construct the all of the subgroups of the
icosahedral symmetry. By means of the “orbitals” of the movement of the vertex by
the symmetry operations, the correspondence between the subgroup and the possible
deformation of the molecule can be established. The change in the energy spectra
and the existence of the zero or nonzero transition moments could be discussed
qualitatively, without massive computation. Furthermore, through the construction
of the semi-direct group of the icosahedral group, the structure equipped with a
symmetry higher than that of the icosahedral symmetry, can be generated. The
interest with relation to the “molecular design” is as follows: the higher symmetry
142
in the semidirect product merges some of the irreducible representations into one
and gives rise to a higher degeneracy of the energy spectra. These examples suggest
the importance of the group theoretical viewpoint in the material design, re-armored
(or revived) by the modern computational tools, developed as “computer algebra”
among mathematicians.
VIII.
ANALYSIS OF VIBRATIONAL MODE IN C60
In this section, the computation of the vibrational modes of the C60 by means
of the computer algebra is demonstrated. The C60 molecule has the unique symmetry, comparable to one soccer-ball (see Fig. 6). The symmetry of the molecule is
described by the icosahedral group, and this group is represented by the permutation of 60 vertexes or the group of rotation matrices operating on them. Owing to
this high symmetry, the molecule shows peculiar electronic and dynamic properties,
both of scientific and industrial interest, and the vibrational mode might be one of
the principal phenomena in governing the quantum dynamics in it. As a tool of
the computation, we utilize the software GAP, developed in the field of the pure
mathematics[12]. This software can construct the symmetry group and compute the
irreducible representations and other group-theoretical properties and enable us to
put into practice the application in the material science, possibly in a modernized
way. The computation presented hereafter will be useful to the working physicists
and the students, who want to deepen the understanding of the group theory that is
given in the textbooks[2, 4, 5, 8, 9].
143
A.
The vibrational modes of C60
The atoms and the pentagons in the molecule are provided with sequential numbers as in Fig. 7 in the preceding section. The symmetry operations on the spatial
vectors on the vertexes of C60 are made by the direct products of the two types
of matrices. The first type Ri represents the replacement of 60 vertexes and the
second type Gi the rotation and the reflection of the spatial vectors. The symmetry
operations on the spatial vectors placed into 60 vertexes are represented as this:
Oi = Rif ⊗ Gi ,
(86)
where Ri and Gi should represent the equivalent operation, and the superscript f
is the group automorphism in the group R. The matrices generates the icosahedral
group, and the trace of Oi provides us with the trace representation χtr (Oi )(not irreducible). The decomposition of the trace representation by means of the irreducible
representations indicates the distinct modes of the vibration, and the projection operators composed by Oi can generate the basis vectors for them. And to obtain the
vibrational mode, the degrees of the freedoms of the translation and the rotation of
the molecule should be removed:
χtr − χT rans − χRot .
(87)
The presence of the automorphism f is inevitable. Consider the operation r =
(1, 2, 3, 4, 5) · · · (· · · ). This operation, as the rotation G, acts on the dislocated ver144
texes (V ([j] → V [j] + d[j]):
G(V [1] + d[1]) = V [2] + G · d[1]
G(V [2] + d[2]) = V [3] + G · d[2]
G(V [3] + d[3]) = V [4] + G · d[3]
G(V [4] + d[4]) = V [5] + G · d[4]
G(V [5] + d[5]) = V [1] + G · d[5]
····················· .
The operation on d[i] should be represented as
0 0 0 0 G ···
G
0
0
0
..
.
d[1]
d[2]
0 0 0 0 ···
G 0 0 0 · · · d[3]
d[4]
0 G 0 0 ···
0 0 G 0 · · · d[5]
.. .. .. ..
..
. . . .
.
so that the rotation to d[j] should be located at the correct position, while the
permutation r is represented in the form which is the transposition of that operation:
0 1 0 0 0 ···
0
0
0
1
..
.
0 1 0 0 ···
0 0 1 0 ···
0 0 0 1 ···
0 0 0 0 ···
.. .. .. ..
. . . .
145
It happens that the permutation r should act on d[j] as its transposition, or its
inverse. And the product of the operations is given by (R1 , G1 ) ◦ (R2 , G2 ) = (R2 ·
R1 , G1 · G2 ), with inverse directions for Ri and Gi . This observation shows the
inadequacy of the simple definition of the direct product Ri ⊗ Gi as the symmetry
operation. However, the naive redefinition, such as Ri−1 ⊗ Gi for all i, does not
work well. Now we can make use the fact the conjugation (A → M · A · M−1 ) is an
automorphism which preserves the sequential order of the group multiplication; we
should find the elements which inverts the generators of the group by the conjugation
in order to define the exact symmetry operations, the products of which go from
right to left both for Rif and Gi . Indeed this can be done in the final stage of the
computation. So, before stepping into it, let us execute the naive computation (with
the direct product Ri ⊗ Gj ). Concerning the dimension counting of the basis set,
such a computation does not fail.
The icosahedral group is defined in several ways. We prepare the twelve points,
located at the centers of the pentagons in the molecule (or the vertexes in the icosahedron) as follows. E(5) is the fifth root of the unity, exp(2πi/5), and the symbolically
presented data are all real. (The data and results are given by the expressions in
GAP.)
1
[0,0,1],
2
[2/5*E(5)-2/5*E(5)^2-2/5*E(5)^3+2/5*E(5)^4,
0,
1/5*E(5)-1/5*E(5)^2-1/5*E(5)^3+1/5*E(5)^4],
3
[-3/5*E(5)-2/5*E(5)^2-2/5*E(5)^3-3/5*E(5)^4,
1/5*E(20)-1/5*E(20)^9-2/5*E(20)^13+2/5*E(20)^17,
1/5*E(5)-1/5*E(5)^2-1/5*E(5)^3+1/5*E(5)^4],
4
[2/5*E(5)+3/5*E(5)^2+3/5*E(5)^3+2/5*E(5)^4,
146
2/5*E(20)-2/5*E(20)^9+1/5*E(20)^13-1/5*E(20)^17,
1/5*E(5)-1/5*E(5)^2-1/5*E(5)^3+1/5*E(5)^4],
5
[2/5*E(5)+3/5*E(5)^2+3/5*E(5)^3+2/5*E(5)^4,
-2/5*E(20)+2/5*E(20)^9-1/5*E(20)^13+1/5*E(20)^17,
1/5*E(5)-1/5*E(5)^2-1/5*E(5)^3+1/5*E(5)^4],
6
[-3/5*E(5)-2/5*E(5)^2-2/5*E(5)^3-3/5*E(5)^4,
-1/5*E(20)+1/5*E(20)^9+2/5*E(20)^13-2/5*E(20)^17,
1/5*E(5)-1/5*E(5)^2-1/5*E(5)^3+1/5*E(5)^4],
7
[0,0,-1],
8
[-2/5*E(5)+2/5*E(5)^2+2/5*E(5)^3-2/5*E(5)^4,
0,
-1/5*E(5)+1/5*E(5)^2+1/5*E(5)^3-1/5*E(5)^4],
9
[3/5*E(5)+2/5*E(5)^2+2/5*E(5)^3+3/5*E(5)^4,
-1/5*E(20)+1/5*E(20)^9+2/5*E(20)^13-2/5*E(20)^17,
-1/5*E(5)+1/5*E(5)^2+1/5*E(5)^3-1/5*E(5)^4],
10 [-2/5*E(5)-3/5*E(5)^2-3/5*E(5)^3-2/5*E(5)^4,
-2/5*E(20)+2/5*E(20)^9-1/5*E(20)^13+1/5*E(20)^17,
-1/5*E(5)+1/5*E(5)^2+1/5*E(5)^3-1/5*E(5)^4],
11
[-2/5*E(5)-3/5*E(5)^2-3/5*E(5)^3-2/5*E(5)^4,
2/5*E(20)-2/5*E(20)^9+1/5*E(20)^13-1/5*E(20)^17,
-1/5*E(5)+1/5*E(5)^2+1/5*E(5)^3-1/5*E(5)^4],
12
[3/5*E(5)+2/5*E(5)^2+2/5*E(5)^3+3/5*E(5)^4,
1/5*E(20)-1/5*E(20)^9-2/5*E(20)^13+2/5*E(20)^17,
-1/5*E(5)+1/5*E(5)^2+1/5*E(5)^3-1/5*E(5)^4]
These twelve points are transported among themselves by these three operations
(g1,g2,g3). The three operations (g1,g2,g3) induces the three permutations (f1,f2,f3)
147
on the 12 vertexes of the icosahedron, and(r1,r2,r3) on 60 vertexes in the C60 .
gap> g1:=[ [ 3/10*E(5)+1/5*E(5)^2+1/5*E(5)^3+3/10*E(5)^4,
1/10*E(20)-1/10*E(20)^9-1/5*E(20)^13+1/5*E(20)^17,
2/5*E(5)-2/5*E(5)^2-2/5*E(5)^3+2/5*E(5)^4 ],
[ -1/2*E(20)+1/2*E(20)^9, -1/2*E(5)-1/2*E(5)^4, 0 ],
[ -3/5*E(5)-2/5*E(5)^2-2/5*E(5)^3-3/5*E(5)^4,
-1/5*E(20)+1/5*E(20)^9+2/5*E(20)^13-2/5*E(20)^17,
1/5*E(5)-1/5*E(5)^2-1/5*E(5)^3+1/5*E(5)^4 ] ];
gap> g2:=[ [ 1/2*E(5)+1/2*E(5)^4, 1/2*E(20)-1/2*E(20)^9, 0 ],
[ -1/2*E(20)+1/2*E(20)^9, 1/2*E(5)+1/2*E(5)^4, 0 ], [ 0, 0, 1 ] ];
gap> g3:=[ [ -1, 0, 0 ], [ 0, -1, 0 ], [ 0, 0, -1 ] ];
gap> G:=Group(g1,g2,g3);
gap> f1:=(1,2,3)(4,6,11)(5,10,12)(7,8,9);
gap> f2:=(2,3,4,5,6)(8,9,10,11,12);
gap> f3:=(1,7)(2,8)(3,9)(4,10)(5,11)(6,12);
gap> F:=Group(f1,f2,f3);
gap> r1:=(
(
1, 14, 20)(
6, 58, 22)(
2, 15, 16)(
7, 57, 23)(
3, 11, 17)(
8, 56, 24)(
4, 12, 18)(
5, 13, 19)
9, 60, 25)( 10, 59, 21)
( 31, 44, 50)( 32, 45, 46)( 33, 41, 47)( 34, 42, 48)( 35, 43, 49)
( 36, 28, 52)( 37, 27, 53)( 38, 26, 54)( 39, 30, 55)( 40, 29, 51);
gap> r2:=(
1,
2,
3,
4,
5)(
6, 11, 16, 21, 26)(
148
7, 12, 17, 22, 27)
(
8, 13, 18, 23, 28)(
9, 14, 19, 24, 29)( 10, 15, 20, 25, 30)( 31, 32, 33, 34, 35)
( 36, 41, 46, 51, 56)( 37, 42, 47, 52, 57)( 38, 43, 48, 53, 58)( 39, 44, 49, 54, 59)
( 40, 45, 50, 55, 60);
gap> r3:=(
(
1, 31)(
2, 32)(
3, 33)(
4, 34)(
5, 35)(
6, 36)(
7, 37)(
8, 38)
9, 39)( 10, 40)( 11, 41)( 12, 42)( 13, 43)( 14, 44)( 15, 45)( 16, 46)( 17, 47)
( 18, 48)( 19, 49)( 20, 50)( 21, 51)( 22, 52)( 23, 53)( 24, 54)( 25, 55)( 26, 56)
( 27, 57)( 28, 58)( 29, 59)( 30, 60);
gap> R:=Group(r1,r2,r3);
gap> Irr(R);
The command “Irr(R)” returns the character table of R, a part of which is listed
in table 40:
Rep
N.1
N.2
N.3
N.4
N.5
N.6
N.7
N.8
N.9
N.10
Dim
1
1
3
3
3
3
4
4
5
5
[f1,f2,f3]
[ 1, 1, 1 ]
[ 1, 1, -1 ]
[ 0,-E(5)-E(5)^4, 3 ]
[ 0,-E(5)^2-E(5)^3, 3 ]
[ 0,-E(5)-E(5)^4, -3 ]
[ 0,-E(5)^2-E(5)^3, -3 ]
[ 1, -1, 4 ]
[ 1, -1, -4 ]
[ -1, 0, 5 ]
[ -1, 0, -5 ]
Ag
Au
F2g
F1g
F2u
F1u
Hg
Hu
Gg
Gu
TABLE 40. The characters of the generators in ten irreducible representations of the
icosahedral group.
We use the group R as the standard. The isomorphism from R to G is set up at
the three generators (from [R.1,R.2,R.3] to [G.1,G.2,G.3].)
gap> hom:=GroupHomomorphismByImages(R,G,[R.1,R.2,R.3],[G.1,G.2,G.3]);
149
It is necessary to judge the irreducible representations to which the translation belongs. The projectors (applicable to three-dimensional spatial vectors) are computed
now for this purpose. They are defined as
P (p) =
l(p) X (p)∗
χ (T ) · O(T )
|G| T ∈G
by summing up the elements T of the group G, where l(p) is the dimension of the
p-th irreducible representation, |G| the order of the group, χ(p) the character, O(T )
the operation of the group. The gap script is given as:
Projopr:=function(irr,Em,hom)
#
# irr: Irreducible representation of the group.
# Em: Elements in the group.
# hom: Homomorphism of the permutation group to the matrix group
#
return List(irr,i->Sum(List(Em,j->j^i*Image(hom,j)))/Size(Em)*i[1]);
end;
In this function, the elements “j” (in the list “Em”) in the given group are mapped
to another group by the homomorphism “hom”. The characters are computed in the
first group (with the irreducible representation “irrep”) and multiplied to the images
of “j” in the second group. The summation over them produces the projector. In
this definition, the projectors of the matrix group G are composed through the
isomorphism from the permutation group R to G, not directly from G. This is a
precaution to avoid a possible confusion, since the character tables of R and G,
computed by GAP, might be presented in the slightly different orders, even if these
two groups are isomorphic. The result is given in table 41. The non-zero one exists
150
only in the representation N.6, so we conclude the spatial translation belongs to this
representation.
N.1
N.2
N.3
N.4
N.5
N.6
N.7
N.8
N.9
N.10
[
[
[
[
[
[
[
[
[
[
[
[
[
[
[
[
[
[
[
[
0,
0,
0,
0,
0,
1,
0,
0,
0,
0,
0,
0,
0,
0,
0,
0,
0,
0,
0,
0,
0
0
0
0
0
0
0
0
0
0
],
],
],
],
],
],
],
],
],
],
Operator
[ 0, 0, 0 ],
[ 0, 0, 0 ],
[ 0, 0, 0 ],
[ 0, 0, 0 ],
[ 0, 0, 0 ],
[ 0, 1, 0 ],
[ 0, 0, 0 ],
[ 0, 0, 0 ],
[ 0, 0, 0 ],
[ 0, 0, 0 ],
[
[
[
[
[
[
[
[
[
[
0,
0,
0,
0,
0,
0,
0,
0,
0,
0,
0,
0,
0,
0,
0,
0,
0,
0,
0,
0,
0
0
0
0
0
1
0
0
0
0
]
]
]
]
]
]
]
]
]
]
]
]
]
]
]
]
]
]
]
]
TABLE 41. The projection operators for the spatial vectors.
.
There is another necessity to judge the representation of the spatial rotations.
The group operation T on the rotation matrix M is defined as the product from both
sides, T · M · Tt . This operation should be treated with care. This product is also
a rotation matrix, but the summation of such products not. Thus the projection
operator defined as above will not generate the proper rotation matrix if it is naively
applied to the rotation matrices. Concerning this, the exponential map is useful,
which is defined as
M = exp(A),
T · M · Tt = exp(T · A · Tt ).
The pre-image A of the rotation M is the skew-symmetric matrix. The projection
operator should be applied to the linear space of these pre-images. The gap computation for the determination of the irreducible representation is now presented. The
151
projectors are computed by the formula
P (p) (A) =
l(p) X (p)∗
χ (T )T · A · T t
|G| T ∈G
by means of the function
ProjoprR:=function(irr,Em,MM,hom)
return List(irr,i->Sum(List(Em,j->j^i*Image(hom,j)
*MM*TransposedMat(Image(hom,j))))/Size(Em)*i[1]);
end;
The computation is done now.
gap> x:=Indeterminate(Rationals,"x");
gap> y:=Indeterminate(Rationals,"y");
gap> z:=Indeterminate(Rationals,"z");
gap> MM:=[[0*x,x,y],[-x,0*x,z],[-y,-z,0*x]];
gap> ProjoprR(Irr(R),Elements(R),MM,hom);
Remark. The expressions such as 0 and 0*x signify different objects. The former is
the zero as a number, and the latter the zero monomial. If one promiscuously uses
0 and 0*x, it might cause some errors.
The non-zero projector is the fourth one. The spatial rotations belong to the
fourth representation (N.4).
Remark. In the basis vectors of the vibrational mode computed hereafter, those which
152
N.1
N.2
N.3
N.4
N.5
N.6
N.7
N.8
N.9
N.10
Operator
[ [ 0, 0, 0 ], [ 0, 0, 0 ], [ 0, 0, 0 ] ]
[ [ 0, 0, 0 ], [ 0, 0, 0 ], [ 0, 0, 0 ] ]
[ [ 0, 0, 0 ], [ 0, 0, 0 ], [ 0, 0, 0 ] ]
[ [ 0, x, y ], [-x, 0, z ], [-y, -z, 0 ] ]
[ [ 0, 0, 0 ], [ 0, 0, 0 ], [ 0, 0, 0 ] ]
[ [ 0, 0, 0 ], [ 0, 0, 0 ], [ 0, 0, 0 ] ]
[ [ 0, 0, 0 ], [ 0, 0, 0 ], [ 0, 0, 0 ] ]
[ [ 0, 0, 0 ], [ 0, 0, 0 ], [ 0, 0, 0 ] ]
[ [ 0, 0, 0 ], [ 0, 0, 0 ], [ 0, 0, 0 ] ]
[ [ 0, 0, 0 ], [ 0, 0, 0 ], [ 0, 0, 0 ] ]
TABLE 42. The projection operators for the rotations.
.
belong to the fourth irreducible representation, include the elements such as
Ii
0
0
.
..
0 0 ···
Ii 0 · · ·
·v
0 Ii · · ·
.. ..
. .
where v are the basis vector in the trivial irreducible representation (invariant by the
symmetry operation), and Ii (i = 1, 2, 3) are
0 0 0
0 0 1
0 1 0
I1 = −1 0 0 , I2 = 0 0 0 , I3 = 0 0 1 .
0 −1 0
−1 0 0
0 0 0
These vectors represent the degrees of freedom of the rotation. They are set of 60
of three-dimensional vectors, defined on each vertex, and located in the tangent plane
to the sphere where the 60 vertexes are placed.
Now we construct the symmetry operation (Oi = Ri ⊗ Gi ) on the spatial vectors
153
on the 60 vertexes. The following function converts a permutation Ri to an integer
matrix which represents the replacement of vertexes.
ListToMat:=function(l)
local size,h,i,j;
size:=Size(l);
h:=List([1..size],i->List([1..size],j->0));
for i in [1..size] do
h[i][l[i]]:=1;
od;
return h;
end;
The following function constructs the direct product of Ri and Gi . The isomorphism “iso” from the group R to G convert an element “rr” (in R) to the equivalence
in G exactly. In the generated matrix, the non-zero entries in Ri are replaced by Gi
(the rotation), as this.
0
1
0
·
·
·
0
0 Gi 0 · · ·
1 0 0 ··· 0
Gi 0 0 · · ·
Ri = .
→
R
⊗
G
=
i
i
..
.
.
.
0 0 0 ··· 1
0 0 0 ···
0
0
..
.
Gi
LargeMatrix:=function(rr,iso)
local l,rmat,rrimg,i,i1,i2,n,m;
l:=ListPerm(rr,60);
rmat:=List([1..Size(l)*3],i->List([1..Size(l)*3],j->0));
rrimg:=Image(iso,rr);
for i in [1..Size(l)]
154
do
i1:=i-1;i2:=l[i]-1;
for n in [1..3]
do
for m in [1..3]
do
rmat[3*i1+n][3*i2+m]:=rrimg[n][m];
od;
od;
od;
return rmat;
end;
This function at first prepares a 180×180 matrix, in order to represent the symmetry
operation on the vectors at the vertex of C60 . The symmetry operation of the group
element “rr” is transformed by the isomorphism “iso” to Gi and placed in the large
matrix, by means of the list “l” which stores the data of the permutation Ri .
The projector function, composed from the direct product, is defined as:
ProjoprE:=function(irr,Em,hom)
#
# irr:Irreducible representation.
# Em :Elements of the group R.
# hom:Isomorphism from the group R to the group G.
#
return List(irr,i->Sum(List(Em,j->j^i*LargeMatrix(j,hom)))/Size(Em)*i[1]);
end;
We can make now the basis set of the mode of deformation. The projectors are
155
placed on the list of ten matrices (of dimension 180) (PRJE[1]...PRJE[10]), and each
of them does the projection on each irreducible representations (N.1...N.10). The
basis vectors for the each vibrational mode are generated from the column vectors of
the corresponding projector. In the computation, the list “BASSIZE” gives us the
dimensions of each mode.
gap> PRJE:=ProjoprE(Irr(R),Elements(R),hom);;
gap> BAS:=List(PRJE,i->BaseMat(TransposedMat(i)));;
gap> BASSIZE:=List(BAS,Size);
[ 2, 1, 12, 12, 15, 15, 24, 24, 40, 35 ]
gap> BAS;
[ [ [ 1, E(20)-E(20)^9+E(20)^13-E(20)^17, 0, 2*E(5)+E(5)^2+E(5)^3+2*E(5)^4,
-E(20)^13+E(20)^17, 0, -2*E(5)-2*E(5)^4, 0, 0,
2*E(5)+E(5)^2+E(5)^3+2*E(5)^4, E(20)^13-E(20)^17, 0, 1,
-E(20)+E(20)^9-E(20)^13+E(20)^17, 0,
-1/5*E(5)+1/5*E(5)^2+1/5*E(5)^3-1/5*E(5)^4,
1/5*E(20)-1/5*E(20)^9+3/5*E(20)^13-3/5*E(20)^17,
12/5*E(5)+8/5*E(5)^2+8/5*E(5)^3+12/5*E(5)^4,
-6/5*E(5)-4/5*E(5)^2-4/5*E(5)^3-6/5*E(5)^4,
...........................................
...........................................
Remark. One might demand numerical basis as well as symbolic one. The latter can
be easily modified into quasi-Fortran form. For example, by replacements of strings,
a formula such as
f:=-1/5*E(5)+1/5*E(5)^2+1/5*E(5)^3-1/5*E(5)^4;
is rewritten as
156
FUNCTION E(I)
INTEGER I
COMPLEX*16 E
REAL*8 TWOPI
TWOPI=8.0D0*ATAN(1.0D0)
E=DCMPLX(COS(TWOPI/I),SIN(TWOPI/I))
RETURN
END
SUBROUTINE(f)
COMPLEX*16 f,E
f=-1./5*E(5)+1./5*E(5)**2+1./5*E(5)**3-1./5*E(5)**4
RETURN
END
At present projectors “PRJE[i]” are defined by the “bare” forms, using the direct
product Ri ⊗ Gi . and they should be modified to the correct ones, such as Mf ·
f
PRJE[i] · M−1
f , composed by the direct product Ri ⊗ Gi , where f is a conjugation
which inverts the generators of the permutation group R. There are two candidates
for such automorphisms.
gap>RPC:=Filtered(Elements(R),x->List([R.1,R.2,R.3],i->i^x)=[R.1^-1,R.2^-1,R.3^-1]);
[ (1,4)(2,3)(6,21)(7,25)(8,24)(9,23)(10,22)(11,16)(12,20)(13,19)(14,18)(15,
17)(27,30)(28,29)(31,34)(32,33)(36,51)(37,55)(38,54)(39,53)(40,52)(41,
46)(42,50)(43,49)(44,48)(45,47)(57,60)(58,59),
(1,34)(2,33)(3,32)(4,31)(5,35)(6,51)(7,55)(8,54)(9,53)(10,52)(11,46)(12,
50)(13,49)(14,48)(15,47)(16,41)(17,45)(18,44)(19,43)(20,42)(21,36)(22,
40)(23,39)(24,38)(25,37)(26,56)(27,60)(28,59)(29,58)(30,57) ]
157
gap>
RPC[1]*r3=RPC[2];
true
The conjugation is computed as a matrix through this function.
ConjugMatrix:=function(rr)
local l,rmat,vmat,i,i1,i2,n,m;
l:=ListPerm(rr,60);
vmat:=List([1..Size(l)*3],i->List([1..Size(l)*3],j->0));
for i in [1..Size(l)]
do
i1:=i-1;i2:=l[i]-1;
for n in [1..3]
do
vmat[3*i1+n][3*i2+n]:=1;
od;
od;
return vmat;
end;
gap> PCM:=ConjugMatrix(RPC[1]);
gap> PRJEC:=List(PRJE,i->i^PCM);;
gap> BASC:=List(PRJEC,i->BaseMat(TransposedMat(i)));;
gap> BASCSIZE:=List(BASC,Size);
[ 2, 1, 12, 12, 15, 15, 24, 24, 40, 35 ]
In fact, in order to obtain the correct vibrational modes, it is enough to multiply the
matrix “PCM” for the conjugation to the basis vectors “BAS” computed by the bare
158
projectors. If we neglect such corrections, it happens that the “breathing mode”
would not belong to the trivial representation.
To check the validity, the characters of χT r and those by the irreducible representations are computed in the two lists “CTRC” and “CIRR”. The irreducible
representation of χV ec is located at the sixth component of CIRR, and that of χRot
is at the fourth.
CTRC:=List(Elements(R),i->Trace(ListToMat(ListPerm(i,60))));
CIRR:=List(Irr(R),r->List(Elements(R),i->i^r));
The χT r × χV ec is computed and decomposed.
gap> List([1..Size(R)],i->CTRC[i]*CIRR[6][i])*TransposedMat(CIRR)/Size(R);
[ 2, 1, 4, 4, 5, 5, 6, 6, 8, 7 ]
gap> List([1..10],i->BASSIZE[i]/Irr(R)[i][1]);
[ 2, 1, 4, 4, 5, 5, 6, 6, 8, 7 ]
These results are the multiplicities to each irreducible representation (in theory and
in computation) and shows the validity of the computation. The direct product of
the characters (i.e. the vibrational mode plus extra terms) is decomposed as
χtr ×χV ec = 2·N.1+1·N.2+4·N.3+4·N.4+5·N.5+5·N.6+6·N.7+6·N.8+8·N.9+7·N.10.
B.
One technical problem
The modes of vibrations are decomposed into irreducible representations, some
of which appear with multiplicities. Take one of the basis vectors, computed in the
previous section, and apply the symmetry operations on it: for example, one of those
which belong to the tenth irreducible representation. In the following computation
159
the two lists “EMR” and “AR” contains the group elements in R and the whole of
the symmetry operation, “vs” is the set of the generated vectors, “vsb”, the basis set
to it, and the command RankMat(...) computes the dimension of the vector space.
gap> EMR:=Elements(R) ;;
gap> AR:=List(Elements(R),j->LargeMatrix(j,hom));;
gap> vs:=List(AR,x->x*BAS[10][1]);
gab> vsb:=BaseMat(vs);
gap> RankMat(vs);
25
The chosen vector generates the vector space of dimension 25 by the symmetry
operation.(The basis vectors generated in this way are “vertex-wise” orthogonal to
those in the trivial representation: the basis vectors are the set of 60 small threedimensional vectors, and the small vectors of the former and the latter are orthogonal
at each vertex. The vectors, “vertex-wise” orthogonal to those in the trivial representation, make the subspace that includes several copies of the proper irreducible
space.) But the tenth irreducible representation is of dimension 5. The generated
vectors span the copies of this irreducible space, not confined to one of them. For
the practical purpose, we should remove the degrees of freedoms which correspond
to those of the translation and the rotation. And it is enough to the reset the dislocated origin or the disorientated axes at the proper positions when we make the
structures altered by vibrations. Thus the technical problem of this kind might not
be of importance. However, we can divide the multiplied irreducible spaces, so that
the each component is proper minimal one, separated with each other, kept invariant
under the symmetry operation. In this section, two kinds of such constructions are
demonstrated.
(p)
The first is to use the projectors composed from matrix representations Dkl (OT ),
160
not from characters, as was explained in the previous section. The formula is given
again:
(p)
Pkl =
lp X (p)∗
D (OT ) · OT .
|G| T ∈G kl
The basis vectors are transformed by the matrix representation, forming the subspace
which has the proper dimension of the irreducible representation. We can make the
(p)
plural irreducible spaces from the column vectors of the projector one of the Pkk ,
the diagonal part of the projectors. Unfortunately, even in this construction, the
generated irreducible spaces overlap with each other, not being separated. Thus we
should make the separation, projecting out overlapping subspaces. At first the matrix
representation is computed in the following way. (The GAP, in each run, returns the
different results concerning the matrix representation, since we can determine them
up to the isomorphism, but not furthermore; there are plural possibilities.)
gap> irrR:=Irr(R);;
gap> irrRM:=IrreducibleRepresentations(R);;
gap> EMR:=Elements(R);;
gap> AM:=List(EMR,j->LargeMatrix(j,hom));;
(p)
The projector Pij is defined as a function
PRJR:=function(i,j,p)
local C,REP,dim;
REP:=irrRM[p];
dim:=Trace(()^REP);
C:=List([1..Size(AM)],k->AsList(EMR[k]^REP)[i][j]*AM[k]);
return ComplexConjugate(Sum(C))/Size(AM)*dim;
end;
161
(10)
Take the P1,1 component of the projector. The basis vectors are generated from
column vectors of this. As the GAP software defines the vector set by row vectors,
The basis sets are computed from the transposed matrix. The basis set of the vector
(10)
(10)
space is listed in “vtb”. The projectors P1,1 , ..., P5,1 are stored now for the afterward
usage.(As for the matrix representations, see Appendix I.)
gap> vt:=TransposedMat(PRJR(1,1,10));;
gap> vtb:=BaseMat(vt);;
gap> PRJRS:=List([1..5],i->PRJR(i,1,10));;
The validity of the transformation of the partner functions is checked now.
gap> PRJR(2,2,10)*PRJR(2,1,10)*vtb[1]=PRJR(2,1,10)*vtb[1];
true
This function at first puts p-th irreducible matrix representation in “REP”. (The irreducible matrix representations are already stored in “irrRM”.) For a group element
“x”, the matrix representation D(p) (x) is computed by “x∧ REP”. The dimension of
the irreducible representation “dim” is computed from the trace of the matrix representation of the unit of the group ( represented by “( )”). Each group element “x”
(p)
(stored in the list “EMR”) is transformed to Dij (x), and multiplied to the equivalent matrix operation on the spatial vectors on the vertex of C60 (stored in the list
“AM”). The summation over them, through the complex conjugation, makes the
projector.
Each of the basis vector x generates the five-dimensional, irreducible vector spaces
(10)
(10)
under the symmetry operation, and the vector space is composed of P1,1 ·x, ..., P5,1 ·
x. In the following computations, the irreducible vector space for each of basis vector
is stored in the list “zz” (in the seven component zz[1],...,zz[7]). However, these
162
irreducible spaces are overlapped with each other. The computations of the inner
products show this.
gap> zz:=List(vtb,y->List(List([1..5],i->PRJRS[i]*y)));;
gap> List(zz,a->List(zz,b->RankMat(a*TransposedMat(b))));
[ [ 5, 0, 5, 0, 0, 0, 0 ], [ 0, 5, 0, 0, 0, 0, 0 ], [ 5, 0, 5, 5, 0, 0, 0 ],
[ 0, 0, 5, 5, 0, 5, 0 ], [ 0, 0, 0, 0, 5, 0, 0 ], [ 0, 0, 0, 5, 0, 5, 5 ],
[ 0, 0, 0, 0, 0, 5, 5 ] ]
We now project out the overlapping between irreducible subspaces. We start from
the one of the irreducible space, say zz[1], and construct the orthogonal space to it.
(10)
One of the basis vectors in the latter space, if the projector P1,1 is applied to it,
(10)
belongs to the irreducible space invariant by P1,1 . This vector is chosen as the
generator of another irreducible space. Similarly, we construct the orthogonal space
to the set of the irreducible spaces obtained until now, from this, we extract the
generator of another irreducible space and proceed so on.
gap> B:=ShallowCopy(zz[1]);;
gap> D:=[ShallowCopy(zz[1])];;
gap> for n in [2..Size(zz)]
do
OV:=BaseOrthogonalSpaceMat(B);;
xx:=BaseMat(List(OV,x->PRJRS[1]*x))[1];
vv:=List(List([1..5],i->PRJRS[i]*xx));
Append(B,vv);
Append(D,[vv]);
od;
We generated seven separated irreducible spaces; let us check this (by computing
the inner products of vectors):
163
gap> List(D,a->List(D,b->RankMat(a*TransposedMat(b))));;
[ [ 5, 0, 0, 0, 0, 0, 0 ], [ 0, 5, 0, 0, 0, 0, 0 ], [ 0, 0, 5, 0, 0, 0, 0 ],
[ 0, 0, 0, 5, 0, 0, 0 ], [ 0, 0, 0, 0, 5, 0, 0 ], [ 0, 0, 0, 0, 0, 5, 0 ],
[ 0, 0, 0, 0, 0, 0, 5 ] ]
The vector space generated by the symmetry operations from one of the separated
irreducible spaces should be orthogonal to others; let us check this (using the list
“AM” of symmetry operations prepared in the preceding computation):
gap> OO:=List(AM,o->o*D[1][1]);;
gap> List(D,d->RankMat(OO*TransposedMat(d)));
[ 5, 0, 0, 0, 0, 0, 0 ]
gap> OO:=List(AM,o->o*D[2][1]);;
gap> List(D,d->RankMat(OO*TransposedMat(d)));
[ 0, 5, 0, 0, 0, 0, 0 ]
. . . etc . . .
We can see that all is done well.
The second method is to divide the multiplied irreducible spaces W (as computed
in the previous section, by the projectors defined by the characters) into proper
minimal components, invariant under the symmetry operation. This method utilizes
the coset decomposition of the group G by the subgroup H, G/H. In general, one
vector v1 in W, under the operation of the elements gi in G, generates the set of
vectors, the total number of which is equal to the order of G. Now let us take
another vector w1 , also in W, generated by the operation of the elements hi in H,
P
such as w1 = i hi · v1 . Under the symmetry operation of G, it generates the set
of vectors, the total number of which is equal to the order of G/H. The number of
the generated vectors decreased in the latter case. We might expect the dimension
of the basis vectors of the generated vectors would also be diminished in the latter
164
case. And if w1 would generate the vector space of the proper dimension of the
irreducible representation, we adopt it, as the generator of the separated, proper,
irreducible space. (If not, with another chose of H, this step is repeated.) Then we
project out the vector space generated by w1 from W. We take another vector v2
from W, make an H-invariant vector w2 with the suitable H (if necessary, H should
be chosen again) to generate the another proper irreducible space. We proceed so
on until W is exhausted. This method will be applicable to the present problem in
order to separate each irreducible component, even if it involves the trial-and-error,
in the choice of the suitable subgroup H. At the start, the vector v1 is taken from the
multiplied irreducible spaces. The candidate vector wi is generated by the projection
of vi into the trivial irreducible representation of the subgroup H. (Indeed, it could
also be prepared by the projection of v into other irreducible representations of H,
for the purpose of diminishing the number of the generated vectors by the symmetry
operation.)
In order to execute this computation, the function “Cutout” (given in appendix
J) is prepared. The function cuts out the subspace (as minimal as possible) from the
given vector space, making use of the coset decomposition of R by the subgroups H.
In projecting out the proper irreducible component from the multiplied copies, we
employ slightly different step from the above description. Let W be the initial starting
place, that is, the multiplied components of the irreducible representations, from
which the separated proper irreducible components W1 , ..., Wi are computed. The
next candidate of the generator vi+1 is not taken directly from the complementary
space (W \ (W1 ∪ W2 ∪ · · · ∪ Wi )); instead, in the whole vector space S (of 180
dimension in this case), the orthogonal space to W1 ∪ W2 ∪ · · · ∪ Wi , biz, S \ (W1 ∪
W2 ∪ · · · ∪ Wi ) is constructed. The basis set in the latter space is projected into the
irreducible representation under the consideration, and the candidate vi+1 is taken
from this projected space (the renewed W). This treatment is to quicken the symbolic
165
computation.
In our problem, the latter method works well, too. We can separate out seven
distinct irreducible spaces. In the following computation the function “Cutout”
tries to divide out the vector space “BAS[10]” with respect to the tenth irreducible
representation. In this vector space seven copies of irreducible spaces exist. The
divided subspaces are given in list “H1”. Each subspace is of dimension 5 and
orthogonal to others. The symmetry operation (given in list “AM”, in the preceding
computation) to each vector in each subspace generates the 5 dimensional vector
space, and the generated vector space exits exactly in the same subspace of that
vector. (As for the other irreducible representations, it is successful, with the suitable
choice of the coset decomposition. For example, the subgroups of order 5 operate well
at the first, third, fourth, fifth, sixth, ninth and tenth irreducible representations;
that of order 8 does at the seventh; that of order 4 does at the eighth. )
gap> H1:=Cutout(BAS[10],PRJE[10],R,hom,5,5);;
gap> List(H1,a->List(H1,b->RankMat(a*TransposedMat(b))));
[ [ 5, 0, 0, 0, 0, 0, 0 ], [ 0, 5, 0, 0, 0, 0, 0 ], [ 0, 0, 5, 0, 0, 0, 0 ],
[ 0, 0, 0, 5, 0, 0, 0 ], [ 0, 0, 0, 0, 5, 0, 0 ], [ 0, 0, 0, 0, 0, 5, 0 ],
[ 0, 0, 0, 0, 0, 0, 5 ] ]
gap> OO:=List(AM,o->o*H1[1][1]);;
gap> List(H1,d->RankMat(OO*TransposedMat(d)));
[ 5, 0, 0, 0, 0, 0, 0 ]
gap> OO:=List(AM,o->o*H1[2][1]);;
gap> List(H1,d->RankMat(OO*TransposedMat(d)));
[ 0, 5, 0, 0, 0, 0, 0 ]
. . . etc . . .
A warning is given: in the above discussion, the symmetry operations are defined
166
in the bare forms, using the direct product Ri ⊗ Gi . To be precise one should use
Rif ⊗ Gi , even if the argument on the dimension counting on the vector space is not
affected. To obtain the proper vector system one should dutifully operate Mf ( the
matrix prepared in the previous section for the conjugation) to any vector v, such as
Mf · v.
C.
Remark to this section
In this section, the computation of the vibrational mode of C60 by means of the
computer algebra system GAP is demonstrated. The computation can be executed
by a standard desktop personal computer. Though the topic in the present article
is limited to the C60 molecule, the way of the computation shall be applicable to
other systems, such as crystal. The small programs in the article are composed as
brief as possible and as comprehensible as possible but without considerations of
the optimization, such as memory usage, or the avoidance of the re-computation of
same objects. So there is the possibility to improve the efficiency, although they
are computable even by the standard desktop personal computer. The computed
basis of the vibrational modes are obtained as the symbolic formulas, and they are
convertible to the numerical program functions simply by the rewriting of strings.
Some of the formulas might be so lengthy that the generated program functions might
be beyond the capacity of the compiler. However, the computations presented in this
article can be done by the floating-point computation, started from the numerical
representation of the rotation matrices, with a certain toleration to the numerical
error under the threshold.
I must add this: there are various studies on the electronic structure and the
deformation of C60 in the quantitative electronic structure computations (not of
the group theoretical view) such as in the references[37–39], or in the empirical
167
model (of the group theoretical view)[13]. The computations in this article, of this
and the previous section, only classifies geometrically possible deformations with
respect to irreducible representations and generates the basis set of deformations in
the symbolic formulas. The obtained results, however, could be applied to more
quantitative studies as those.
IX.
FINAL REMARKS
The theme of this article is how to simulate the property of the material by
means of electronic structure calculations with the aid of computer algebra. The
collaboration of the electronic structure calculation and the computer algebra is not
novel one; rather being regarded as the modernization of the union of the quantum
dynamics and the group theory, the success of which was brilliantly displayed in the
former half of last century. In this article, the treated topics in the group theory are
rather classical ones concerning the character theory. However, there can be various
applications of the myriad topics of mathematics. As for the combinatorics, the Polya
counting theory is a typical example; it is applicable to determine the number of
configurations, equivalent by some symmetry operations, for example, to determine
the total number of possible molecular formations indicated by a single chemical
formula[40]. This type of computations in combinatorics can also be executed by the
computer algebra. As for the application of combinatorics in quantum physics, a sort
of extension of Polya counting theorem can be useful in computing the spin weight in
molecules[41–43]. In a recent study, it is shown that the framework of the electronic
structure could be reconstructed by means of a concept in the algebraic geometry,
so-called Groebner basis theory, through the utilization of symbolic computation[44].
In my article, I have omitted several important application of the group theory
and the algebra. One of this is the treatment of magnetic or color groups. Also, I
168
omitted the treatment of Lie algebra, important in the spin system or the study of
angular momentum. As for the former, a package of color groups is implemented
in GAP[12]. As for the latter, the Lie algebra is computable by GAP, although it
is not of the digested style, suitable for the practical usage of physicists. (Some of
the implemented GAP packages are intended to study the model of spin systems,
proposed in the statistical- or mathematical physics.) I must admit that I have
not yet developed my skills to compute them through GAP. But I believe that the
readers if they are acquainted with GAP through my article, can construct their own
application in these topics.
In the present circumstances of the material physics, indeed, there is an awkward
inclination in which one solely relies on computer powers; only by trial and errors,
or, only by varying the input data, one tries to execute “material design”, without
any effective guiding principle. We must admit that the computational packages
have developed so complex and so vast that a common researcher, if not being an
expert, would be obliged to utilize them as black boxes, without knowing the detail
of the programs; fortunately, those packages are so powerful and effective that one
can obtain reliable results with ease. The blind reliance on these packages, however,
should be avoided; the mere enumeration of results of computations, with the slight
modification of input data, will be the waste of computational resources. Certainly,
by gathering and comparing computed results we could surmise vague tendencies,
but may not go anymore, if we do not know the physical or mathematical origin.
Instead, it is necessary for us to have another utensil, which will compensate for
such a shortcoming; the group theory, or its modernization, the computer algebra,
will be of great use for this purpose. With the aid of computer algebra, advanced
concepts in group theory are easily accessible and applicable, so that we can execute
systematic research of material design from the mathematical first principles, as the
group theoretical view illuminates for us the mathematical principles lurking behind
169
observable phenomena. I hope that the computer algebra will bring forth the blessed
reunion of physics and mathematics in the research of materials; the available tool
for us is not limited to the classics such as Courant-Hilbert[45], although they had
been powerful in the early stages of the development of quantum dynamics. The
mathematics and physics have progressed greatly since then; the forefront of these
developments seems to be so esoteric that non-experts feel dismayed when urged to
put it into practice. Nevertheless, one should not be timid; for the sake of ardent
researchers, the computer algebra will be a most suitable guide for their efforts, and
there they will find a most speedily passable bridge over the ever-deepening chasm
between physics and mathematics.
Appendix A: Symmetry operations in the diamond structure (real space)
The symmetry operations in the diamond structure (in Cartesian coordinates)
are given in the table. The operation on the atomic coordinate is given by x →
R · x + T 1 · a1 + T 2 · a2 + T 3 · a3 , where (a1 , a2 , a3 ) are primitive translations.There are
two types of fractional translations, according to the location of the origin of the unit
cell. The first, denoted as (T1,T2,T3) corresponds to the case of atomic positions
of ±(1/8 · a1 + 1/8 · a3 + 1/8 · a3 ):the second, (T1(*),T2(*),T3(*)) corresponds to
positions of (0 · a1 + 0 · a3 + 0 · a3 ) and (1/4 · a1 + 1/4 · a3 + 1/4 · a3 ). Although the
latter case is more conventional, the former has a merit such that the coefficients of
the plane-wave-expression of the wavefunctions are real-valued.
R11 R12 R13 R21 R22 R23 R31 R32 R33 T1 T2 T3 T1(*) T2(*) T3(*)
1
1
0
0
0
1
0
0
0
2
1
0
0
0
-1
0
0
3
-1
0
0
0
1
0
0
0
0
0
0
0
0
0
-1 0.5
0
0
0
0
0
0
-1
0 0.5
0
0
0
0
170
1
4
-1
0
0
0
-1
0
0
0
1
0
0 0.5
0
0
0
5
0
1
0
0
0
1
1
0
0
0
0
0
0
0
0
6
0
1
0
0
0
-1
-1
0
0 0.5
0
0
0
0
0
7
0
-1
0
0
0
1
-1
0
0
0 0.5
0
0
0
0
8
0
-1
0
0
0
-1
1
0
0
0
0 0.5
0
0
0
9
0
0
1
1
0
0
0
1
0
0
0
0
0
0
0
10
0
0
1
-1
0
0
0
-1
0 0.5
0
0
0
0
0
11
0
0
-1
1
0
0
0
-1
0
0 0.5
0
0
0
0
12
0
0
-1
-1
0
0
0
1
0
0
0 0.5
0
0
0
13
0
-1
0
-1
0
0
0
0
-1
0
0
0
0.25
0.25
0.25
14
0
-1
0
1
0
0
0
0
1 -0.5
0
0
0.25
0.25
0.25
15
0
1
0
-1
0
0
0
0
1
0 -0.5
0
0.25
0.25
0.25
16
0
1
0
1
0
0
0
0
-1
0
0 -0.5
0.25
0.25
0.25
17
-1
0
0
0
0
-1
0
-1
0
0
0
0
0.25
0.25
0.25
18
-1
0
0
0
0
1
0
1
0 -0.5
0
0
0.25
0.25
0.25
19
1
0
0
0
0
-1
0
1
0
0 -0.5
0
0.25
0.25
0.25
20
1
0
0
0
0
1
0
-1
0
0
0 -0.5
0.25
0.25
0.25
21
0
0
-1
0
-1
0
-1
0
0
0
0
0
0.25
0.25
0.25
22
0
0
-1
0
1
0
1
0
0 -0.5
0
0
0.25
0.25
0.25
23
0
0
1
0
-1
0
1
0
0
0 -0.5
0
0.25
0.25
0.25
24
0
0
1
0
1
0
-1
0
0
0
0 -0.5
0.25
0.25
0.25
25
-1
0
0
0
-1
0
0
0
-1
0
0
0
0.25
0.25
0.25
26
-1
0
0
0
1
0
0
0
1 -0.5
0
0
0.25
0.25
0.25
27
1
0
0
0
-1
0
0
0
1
0 -0.5
0
0.25
0.25
0.25
28
1
0
0
0
1
0
0
0
-1
0 -0.5
0.25
0.25
0.25
171
0
29
0
-1
0
0
0
-1
-1
0
0
0
0
0
0.25
0.25
0.25
30
0
-1
0
0
0
1
1
0
0 -0.5
0
0
0.25
0.25
0.25
31
0
1
0
0
0
-1
1
0
0
0 -0.5
0
0.25
0.25
0.25
32
0
1
0
0
0
1
-1
0
0
0
0 -0.5
0.25
0.25
0.25
33
0
0
-1
-1
0
0
0
-1
0
0
0
0
0.25
0.25
0.25
34
0
0
-1
1
0
0
0
1
0 -0.5
0
0
0.25
0.25
0.25
35
0
0
1
-1
0
0
0
1
0
0 -0.5
0
0.25
0.25
0.25
36
0
0
1
1
0
0
0
-1
0
0
0 -0.5
0.25
0.25
0.25
37
0
1
0
1
0
0
0
0
1
0
0
0
0
0
0
38
0
1
0
-1
0
0
0
0
-1 0.5
0
0
0
0
0
39
0
-1
0
1
0
0
0
0
-1
0 0.5
0
0
0
0
40
0
-1
0
-1
0
0
0
0
1
0
0 0.5
0
0
0
41
1
0
0
0
0
1
0
1
0
0
0
0
0
0
0
42
1
0
0
0
0
-1
0
-1
0 0.5
0
0
0
0
0
43
-1
0
0
0
0
1
0
-1
0
0 0.5
0
0
0
0
44
-1
0
0
0
0
-1
0
1
0
0
0 0.5
0
0
0
45
0
0
1
0
1
0
1
0
0
0
0
0
0
0
0
46
0
0
1
0
-1
0
-1
0
0 0.5
0
0
0
0
0
47
0
0
-1
0
1
0
-1
0
0
0 0.5
0
0
0
0
48
0
0
-1
0
-1
0
1
0
0
0
0 0.5
0
0
0
TABLE 43: Symmetry operations in the point group in the
diamond unit cell. If fractional translations are set to be zero,
these operations are applicable to the cubic lattice.
172
Appendix B: Symmetry operations in the hexagonal lattice (real space)
The symmetry operations in graphite (in Cartesian coordinates)are given in the
table:
R11
1
R12 R13
1
0
√
0
1
0
0
0
1
0
0
3/2 0.5
0
0
0
1
0
0 -0.5
3/2 -0.5
0
0
0
1
0
0
-1
0
0
0
1
0
0 -0.5
-0.5
0
0
0
1
0
0
0.5
0
0
0
1
0
0 -0.5
0.5
0
0
0
-1
0
0 -0.5
-0.5
0
0
0
-1
0
0
-1
0
0
0
-1
0
0 -0.5
3/2 -0.5
0
0
0
-1
0
0
3/2 0.5
0
0
0
-1
0
0 -0.5
1
0
0
0
-1
0
0
0 -1
√
0 - 3/2 -0.5
√
0 - 3/2 0.5
0
0
0
-1
0
0 -0.5
0
0
0
-1
0
0
0
0
0
-1
0
0 -0.5
0
1
0
0
0
-1
0
0
3/2 0.5
0
0
0
-1
0
0 -0.5
3/2 -0.5
0
0
0
-1
0
0
0
3/2 -0.5
0
0
0
1
0
0
0
3/2 0.5
0
0
0
1
0
0 -0.5
0
0
0
1
0
0
2 0.5 - 3/2
√
3 -0.5 - 3/2
0
4
0
0
3/2
0
3/2
0
7 -0.5 - 3/2
√
8 0.5 - 3/2
0
9
0
-1
5 -0.5
6 0.5
√
√
√
1
10 0.5
11 -0.5
12
-1
13
-1
14 -0.5
15 0.5
16
0
√
√
√
√
1
3/2
0
0
0
3/2
0
0
0
0
0
3/2
3/2
0
√
17 0.5 - 3/2
√
18 -0.5 - 3/2
√
19 0.5
3/2
√
20 -0.5
3/2
0
21
0
-1
0
R21 R22 R23 R31 R32 R33 T1 T2 T3
0
0
0
0
√
√
0
√
- 3/2
√
- 3/2
√
- 3/2
√
- 3/2
0
√
√
0
0
√
√
√
√
0
1
173
0
0
0
0
0
0
0
0
0
√
22 -0.5 - 3/2
√
23 0.5 - 3/2
√
0 - 3/2 0.5
√
0 - 3/2 -0.5
24
0
1
0
0
-1
0
0
0
1
0
0 -0.5
0
0
0
1
0
0
0
0
0
1
0
0 -0.5
0
TABLE 44: The symmetry operations in graphite in real
space.
Appendix C: Symmetry operations in the diamond structure (wave-number
space)
The symmetry operations in the diamond structure, as the rotation matrices in
the reciprocal spaces are given in the table:
R’11 R’12 R’13 R’21 R’22 R’23 R’31 R’32 R’33 T1 T2 T3
1
1
0
0
0
1
0
0
0
1
0
0
0
2
-1
0
0
-1
0
1
-1
1
0 0.5
0
0
3
0
-1
1
0
-1
0
1
-1
0
0 0.5
0
4
0
1
-1
1
0
-1
0
0
-1
0
0 0.5
5
0
1
0
0
0
1
1
0
0
0
0
0
6
0
-1
0
1
-1
0
0
-1
1 0.5
0
0
7
1
0
-1
0
0
-1
0
1
-1
0 0.5
0
8
-1
0
1
-1
1
0
-1
0
0
0
0 0.5
9
0
0
1
1
0
0
0
1
0
0
0
0
10
0
0
-1
0
1
-1
1
0
-1 0.5
0
0
11
-1
1
0
-1
0
0
-1
0
1
0 0.5
0
12
1
-1
0
0
-1
1
0
-1
0
0
0 0.5
13
0
-1
0
-1
0
0
0
0
-1
0
0
174
0
14
0
1
0
0
1
-1
-1
1
15
1
0
-1
1
0
0
1
-1
16
-1
0
1
0
-1
1
0
17
-1
0
0
0
0
-1
18
1
0
0
1
-1
19
0
-1
1
0
20
0
1
-1
21
0
0
22
0
23
0
0
0
0 -0.5
0
0
1
0
0 -0.5
0
-1
0
0
0
0
0
1
0
-1 -0.5
0
0
0
1
-1
0
1
0 -0.5
0
-1
1
0
0
1
0
0
0 -0.5
-1
0
-1
0
-1
0
0
0
0
0
0
1
-1
0
1
0
-1
1 -0.5
0
0
-1
1
0
0
1
0
0
1
-1
0 -0.5
0
24
1
-1
0
1
0
-1
1
0
0
0
0 -0.5
25
-1
0
0
0
-1
0
0
0
-1
0
0
0
26
1
0
0
1
0
-1
1
-1
0 -0.5
0
0
27
0
1
-1
0
1
0
-1
1
0
0 -0.5
0
28
0
-1
1
-1
0
1
0
0
1
0
0 -0.5
29
0
-1
0
0
0
-1
-1
0
0
0
0
0
30
0
1
0
-1
1
0
0
1
-1 -0.5
0
0
31
-1
0
1
0
0
1
0
-1
1
0 -0.5
0
32
1
0
-1
1
-1
0
1
0
0
0
0 -0.5
33
0
0
-1
-1
0
0
0
-1
0
0
0
0
34
0
0
1
0
-1
1
-1
0
1 -0.5
0
0
35
1
-1
0
1
0
0
1
0
-1
0 -0.5
0
36
-1
1
0
0
1
-1
0
1
0
0
0 -0.5
37
0
1
0
1
0
0
0
0
1
0
0
0
38
0
-1
0
0
-1
1
1
-1
0 0.5
0
0
175
0 -0.5
39
-1
0
1
-1
0
0
-1
1
0
0 0.5
0
40
1
0
-1
0
1
-1
0
0
-1
0
0 0.5
41
1
0
0
0
0
1
0
1
0
0
0
0
42
-1
0
0
-1
1
0
-1
0
1 0.5
0
0
43
0
1
-1
0
0
-1
1
0
-1
0 0.5
0
44
0
-1
1
1
-1
0
0
-1
0
0
0 0.5
45
0
0
1
0
1
0
1
0
0
0
0
0
46
0
0
-1
1
0
-1
0
1
-1 0.5
0
0
47
1
-1
0
0
-1
0
0
-1
1
0 0.5
0
48
-1
1
0
-1
0
1
-1
0
0
0
0 0.5
TABLE 45: Symmetry operations in the diamond crystal in
the reciprocal space.
Appendix D: Symmetry operations in the hexagonal lattice (wave-number
space)
The crystal axes are chosen to be a1 = (1, 0, 0), a2 = (−1/2,
p
(3)/2, 0), a3 =
(0, 0, c). Then The symmetry operations as the operations in the wave-number space
are given as:
R’11 R’12 R’13 R’21 R’22 R’23 R’31 R’32 R’33 T1 T2 T3
1
1
0
0
0
1
0
0
0
1
0
0
2
0
-1
0
1
1
0
0
0
1
0
0 -0.5
3
-1
-1
0
1
0
0
0
0
1
0
0
4
-1
0
0
0
-1
0
0
0
1
0
0 -0.5
5
0
1
0
-1
-1
0
0
0
1
0
0
176
0
0
0
6
1
1
0
-1
0
0
0
0
1
0
0 -0.5
7
-1
-1
0
0
1
0
0
0
-1
0
0 -0.5
8
0
-1
0
-1
0
0
0
0
-1
0
0
9
1
0
0
-1
-1
0
0
0
-1
0
0 -0.5
10
1
1
0
0
-1
0
0
0
-1
0
0
11
0
1
0
1
0
0
0
0
-1
0
0 -0.5
12
-1
0
0
1
1
0
0
0
-1
0
0
13
-1
0
0
0
-1
0
0
0
-1
0
0 -0.5
14
0
1
0
-1
-1
0
0
0
-1
0
0
15
1
1
0
-1
0
0
0
0
-1
0
0 -0.5
16
1
0
0
0
1
0
0
0
-1
0
0
17
0
-1
0
1
1
0
0
0
-1
0
0 -0.5
18
-1
-1
0
1
0
0
0
0
-1
0
0
0
19
1
1
0
0
-1
0
0
0
1
0
0
0
20
0
1
0
1
0
0
0
0
1
0
0 -0.5
21
-1
0
0
1
1
0
0
0
1
0
0
22
-1
-1
0
0
1
0
0
0
1
0
0 -0.5
23
0
-1
0
-1
0
0
0
0
1
0
0
24
1
0
0
-1
-1
0
0
0
1
0
0 -0.5
TABLE 46: The symmetry operations in graphite in reciprocal space.
177
0
0
0
0
0
0
0
Appendix E: The computation of Wyckoff positions by GAP
Computations for space groups can be executed by means of “Cryst” package in
GAP. Though the manual of this package is provided, the description is too curt and
the computed result is too unfriendly for beginners to utilize. So in this section, the
usage of this package is illustrated.
The crystallographic group is defined at first (in this example, with minimal generators):
gap> M1:=[[0,0,1,0],[1,0,0,0],[0,-1,0,0],[1/4,1/4,1/4,1]];
gap> M2:=[[0,0,-1,0],[0,-1,0,0],[1,0,0,0],[0,0,0,1]];
gap> S:=AffineCrystGroup([M1,M2]);
The point group is obtained by:
gap> P:=PointGroup(S);
The Wyckoff positions are computed now:
gap> W:=WyckoffPositions(S);
[ < Wyckoff position, point group 4, translation := [ 1/8, 3/8, 7/8 ],
basis := [
] >
, < Wyckoff position, point group 4, translation := [ 1/8, 3/8, 3/8 ],
basis := [
] >
, < Wyckoff position, point group 5, translation := [ 1/4, 1/4, 3/4 ],
basis := [
] >
, < Wyckoff position, point group 5, translation := [ 1/4, 1/4, 1/4 ],
basis := [
] >
, < Wyckoff position, point group 3, translation := [ 0, 0, 0 ],
178
basis := [ [ 1, -1, -1 ] ] >
, < Wyckoff position, point group 6, translation := [ 1/4, 1/4, 1/2 ],
basis := [ [ 0, 0, 1 ] ] >
, < Wyckoff position, point group 7, translation := [ 0, 1/4, 7/8 ],
basis := [ [ 1/2, 1/2, 0 ] ] >
, < Wyckoff position, point group 2, translation := [ 0, 0, 0 ],
basis := [ [ 1/2, -1/2, 0 ], [ 0, 0, 1 ] ] >
, < Wyckoff position, point group 1, translation := [ 0, 0, 0 ],
basis := [ [ 1/2, 0, 1/2 ], [ 0, 1/2, 1/2 ], [ 0, 0, 1 ] ] >
]
The returned result is the list of the generating set of the Wyckoff positions. The
expression such as
, < Wyckoff position, point group 3, translation := [ 0, 0, 0 ],
basis := [ [ 1, -1, -1 ] ] >
indicates the information of a generator for the Wyckoff positions, which is located
at the segment [0,0,0]+t[1,-1,-1], where t is arbitrary real number inasmuch that the
segment is confined in the unit cell. Now we get nine generators. The computed
result, therefore, should be read as,
[ 1/8, 3/8, 7/8 ]
[ 1/8, 3/8, 3/8 ]
[ 1/4, 1/4, 3/4 ]
[ 1/4, 1/4, 1/4 ]
[ 0, 0, 0 ]+t[ 1, -1, -1 ]
[ 1/4, 1/4, 1/2 ]+t[ 0, 0, 1 ]
[ 0, 1/4, 7/8 ]+t[ 1/2, 1/2, 0 ]
179
[ 0, 0, 0 ]+t1[ 1/2, -1/2, 0 ]+t2[ 0, 0, 1 ]
[ 0, 0, 0 ]+t1[ 1/2, 0, 1/2 ]+t2[ 0, 1/2, 1/2 ]+t3[ 0, 0, 1 ]
The first four are single points, but the remaining five are one-dimensional segments, two-dimensional planes, or, a three-dimensional shape. (The last of them is
the arbitrary position in the unit cell.) Let us pay attention to the fourth of them. It
is the position of one carbon atom in the minimal diamond unit cell; which is moved
to [0,0,0] (the position of another Carbon atom) by means of symmetry operations.
In order to obtain whole of the Wyckoff positions, the equivalent coordinate points
(orbits by the symmetry operations) can be evaluated in the following way.
The total numbers of the equivalent coordinate points are given by
gap> List(WyckoffPositions(S),x->Size(WyckoffOrbit(x)));
[ 4, 4, 2, 2, 8, 12, 24, 24, 48 ]
For each generating points, the equivalents coordinates can be computed separately:
gap> WyckoffOrbit(WyckoffPositions(S)[1]);
[ < Wyckoff position, point group 4, translation := [ 1/8, 3/8, 7/8 ],
basis := [
] >
, < Wyckoff position, point group 4, translation := [ 3/8, 1/8, 7/8 ],
basis := [
] >
, < Wyckoff position, point group 4, translation := [ 3/8, 3/8, 5/8 ],
basis := [
] >
, < Wyckoff position, point group 4, translation := [ 1/8, 1/8, 5/8 ],
basis := [
] >
]
gap> WyckoffOrbit(WyckoffPositions(S)[2]);
180
[ < Wyckoff position, point group 4, translation := [ 1/8, 3/8, 3/8 ],
basis := [
] >
, < Wyckoff position, point group 4, translation := [ 3/8, 1/8, 3/8 ],
basis := [
] >
, < Wyckoff position, point group 4, translation := [ 3/8, 3/8, 1/8 ],
basis := [
] >
, < Wyckoff position, point group 4, translation := [ 1/8, 1/8, 1/8 ],
basis := [
] >
]
gap> ...................................................................
gap> WyckoffOrbit(WyckoffPositions(S)[8]);
[ < Wyckoff position, point group 2, translation := [ 0, 0, 0 ],
basis := [ [ 1/2, -1/2, 0 ], [ 0, 0, 1 ] ] >
, < Wyckoff position, point group 2, translation := [ 1/4, 1/4, 1/4 ],
basis := [ [ -1/2, 0, 1/2 ], [ 0, -1, 0 ] ] >
, < Wyckoff position, point group 2, translation := [ 0, 0, 0 ],
..................................................................
..................................................................
]
Appendix F: Subgroup lattice
In this article, the sequence of crystalline symmetry reduction is displayed by
means of the mathematical concept of the lattice of subgroups, in which the inclusion
relations among them are illustrated as a graph. The GAP computation can be done
as follows:
gap> M1:=[[0,1,0],[0,0,1],[1,0,0]];
181
[ [ 0, 1, 0 ], [ 0, 0, 1 ], [ 1, 0, 0 ] ]
gap> M2:=[[0,1,0],[-1,0,0],[0,0,1]];
[ [ 0, 1, 0 ], [ -1, 0, 0 ], [ 0, 0, 1 ] ]
gap> G:=Group(M1,M2);
Group([ [ [ 0, 1, 0 ], [ 0, 0, 1 ], [ 1, 0, 0 ] ],
[ [ 0, 1, 0 ], [ -1, 0, 0 ], [ 0, 0, 1 ] ] ])
gap> Size(G);
24
gap> l:=LatticeSubgroups(G);
<subgroup lattice of Group(
[ [ [ 0, 1, 0 ], [ 0, 0, 1 ], [ 1, 0, 0 ] ],
[ [ 0, 1, 0 ], [ -1, 0, 0 ], [ 0, 0, 1 ] ] ]), 11 classes,
30 subgroups>
gap> ConjugacyClassesSubgroups(l);
[ Group([],[ [ 1, 0, 0 ], [ 0, 1, 0 ], [ 0, 0, 1 ] ])^G,
Group([ [ [ 1, 0, 0 ], [ 0, -1, 0 ], [ 0, 0, -1 ] ] ])^G,
Group([ [ [ -1, 0, 0 ], [ 0, 0, -1 ], [ 0, -1, 0 ] ] ])^G,
Group([ [ [ 0, 1, 0 ], [ 0, 0, 1 ], [ 1, 0, 0 ] ] ])^G,
Group([ [ [ -1, 0, 0 ], [ 0, 1, 0 ], [ 0, 0, -1 ] ],
[ [ 1, 0, 0 ], [ 0, -1, 0 ], [ 0, 0, -1 ] ] ])^G,
Group([ [ [ -1, 0, 0 ], [ 0, 0, -1 ], [ 0, -1, 0 ] ],
[ [ 1, 0, 0 ], [ 0, -1, 0 ], [ 0, 0, -1 ] ] ])^G,
Group([ [ [ 1, 0, 0 ], [ 0, 0, 1 ], [ 0, -1, 0 ] ],
[ [ 1, 0, 0 ], [ 0, -1, 0 ], [ 0, 0, -1 ] ] ])^G,
Group([ [ [ -1, 0, 0 ], [ 0, 0, -1 ], [ 0, -1, 0 ] ],
[ [ 0, 1, 0 ], [ 0, 0, 1 ], [ 1, 0, 0 ] ] ])^G,
Group([ [ [ -1, 0, 0 ], [ 0, 1, 0 ], [ 0, 0, -1 ] ],
182
[ [ 1, 0, 0 ], [ 0, -1, 0 ], [ 0, 0, -1 ] ],
[ [ -1, 0, 0 ], [ 0, 0, -1 ], [ 0, -1, 0 ] ] ])^G,
Group([ [ [ -1, 0, 0 ], [ 0, 1, 0 ], [ 0, 0, -1 ] ],
[ [ 1, 0, 0 ], [ 0, -1, 0 ], [ 0, 0, -1 ] ],
[ [ 0, 1, 0 ], [ 0, 0, 1 ], [ 1, 0, 0 ] ] ])^G,
Group([ [ [ -1, 0, 0 ], [ 0, 1, 0 ], [ 0, 0, -1 ] ],
[ [ 1, 0, 0 ], [ 0, -1, 0 ], [ 0, 0, -1 ] ],
[ [ 0, 1, 0 ], [ 0, 0, 1 ], [ 1, 0, 0 ] ],
[ [ -1, 0, 0 ], [ 0, 0, -1 ], [ 0, -1, 0 ] ] ])^G ]
gap> DotFileLatticeSubgroups(l,"paradigm.dot");
The last command “DotfileLatticeSubgroups” is used to generate an output file
for the visualization.
Appendix G: A short program to compute the semidirect product of groups
In the main article, the characters of C3v = Group((1, 2, 3), (1, 2)) ≡ S3 are
computed by the semidirect product between C3 = Group((1, 2, 3)) and I =
Group((1, 2)). The computations are tedious and complicated; a naive implementation by means of GAP programming language could be given as follows:
# The definition of the groups G and A is done.
# The irreducible representations and the lists
# of group elements are prepared.
G:=Group((1,2));
A:=Group((1,2,3));
repG:=Irr(G);
183
repA:=Irr(A);
Gmem:=Elements(G);
Amem:=Elements(A);
#
# The definition of the automorphism on A induced by G
# is given in list "au".
# The elements of G and A are indexed
#
#
#
G[1],G[2],...
and
A[1],A[2],....
# The elements in A are mapped as
#
#
A[i]^G[j]=A[au[i,j]];
#
au:=[[1,1],[2,3],[3,2]];
#
This list now means the following automorphism:
#
x(in A)^E(in G)
: E -> E, a -> a,
#
x(in A)^g(in_G)
: E -> E, a -> a^-1, a^-1 -> a
#
Here a=(1,2,3),
#
and the elements in A are listed [e, a, a^-1].
#
The following is the list of elements
#
in the semidirect product [A,G],
#
given by the Cartesian products of sequential numbers.
a^-1 -> a^-1
#
Elist:=[[1,1],[1,2],[2,2],[3,2],[2,1],[3,1]];
184
GAction:=function(a,g)
#
# A function to compute the automorphism
(a^g)
# on an element "a" in A
# induced by an element "g" in G.
# This function returns the group element,
# according to the list "au".
#
local apos,gpos,ha;
apos:=Position(Amem,a);
gpos:=Position(Gmem,g);
ha:=au[apos][gpos];
return Amem[ha];
end;
InvarianceCHK:=function(iA,g,au)
#
# This function checks whether the character x in A
# (that by the iA-th representation) is
# fixed by the automorphism induced by the element g in G.
#
# iA: index to the irreducible representation of A
# au: the automorphism on A by G, defined above.
#
local ol,tl;
#
185
# The list of characters, before the automorphism.
ol:=List(Amem,x->x^repA[iA]);
# The list of characters, after the automorphism.
tl:=List(Amem,x->Amem[au[Position(Amem,x)][Position(Gmem,g)]]);
tl:=List(tl,x->x^repA[iA]);
return ol=tl;
end;
GetGx:=function(iA,au)
#
# Gx is the set of elements in G,
# such that the character x in A
# (that by the iA-th representation)
# is fixed by the automorphism induced
# by the elements in Gx.
#
# iA: the index to the irreducible representation of A
# au: the automorphism on A by G, defined above.
#
#
local Gx;
Gx:=Filtered(G,g->InvarianceCHK(iA,g,au));
return Group(Gx);
end;
ProSumma:=function(a,g,Gx,iA,iG)
# This is a subordinate function to compute the character
186
# of the elements (a,g),
# when the representations in group A and G,
# and Gx (defined above) are specified.
#
# iA: index to the irreducible representations of A
# iG: index to the irreducible representations of G
#
local k,sum,l1,l2,l3,l4,repGx;
l1:=List(Gmem,x->x*g*x^-1);
l2:=List(Gmem,x->GAction(a,x));
l3:=List(l2,x->x^repA[iA]);
repGx:=Irr(Gx);
sum:=0;
for k in [1..Size(l1)]
do
if l1[k] in Gx then
l4:=l1[k]^repGx[iG];
sum:=sum+l3[k]*l4;
fi;
od;
sum:=sum/Size(Gx);
return sum;
end;
ProChr:=function(au,iA)
#
# au
: the automorphism, defined in the list as above.
187
# iA
: the index to the irreducible representations in A
#
# This function computes the characters in the semidirect product
# (for each element in Elist, defined above).
#
local Gx,repGx,index,retTBL;
Gx:=GetGx(iA,au);
repGx:=Irr(Gx);
retTBL:=[];
for index in [1..Size(repGx)]
do;
Append(retTBL,
[List(Elist,x->ProSumma(Amem[x[1]],Gmem[x[2]],Gx,iA,index))]
) ;
od;
return retTBL;
end;
gap> for iA in [1..Size(repA)]
do;
Display(ProChr(au,iA));
od;
[ [
1,
1,
1,
1,
1,
1 ],
[
1,
-1,
-1,
-1,
1,
1 ] ]
[ [
2,
0,
0,
0,
-1,
-1 ] ]
[ [
2,
0,
0,
0,
-1,
-1 ] ]
188
Appendix H: The analytical representation of the energy spectra of C60 .
In the model of C60 in section VII B, if we chose the transfer integral as Hsingle =1 (on the single bonds in the pentagons) and Hdouble =-y (on the double bonds in
the hexagons), the secular equation is decomposed into the polynomial with some
multiplicity in each representation, as in table 47. The irreducible representations
I.1,...,I.10 are the same as given in the main article, at table 26. The energy spectra
of in the main article can be constructed from this table, setting E = 2x and y=-1/2.
I.1
I.2
I.3
I.4
I.5
I.6
I.7
I.8
I.9
I.10
Polynomial
[ x+y+2 ]
[ ]
[ x-y+(E(5)^2+E(5)^3) ]
[ x-y+(E(5)+E(5)^4) ]
[ x^2-y^2+(-E(5)-2*E(5)^2-2*E(5)^3-E(5)^4)*x
+(-E(5)-E(5)^4)*y+(2*E(5)+2*E(5)^4) ]
[ x^2-y^2+(-2*E(5)-E(5)^2-E(5)^3-2*E(5)^4)*x
+(-E(5)^2-E(5)^3)*y+(2*E(5)^2+2*E(5)^3) ]
[ x^2-y^2-x-1 ]
[ x^2-y^2-x-2*y-1 ]
[ x^3+x^2*y-x*y^2-y^3+x^2+2*x*y+y^2-3*x-y-2 ]
[ x^2-y^2-x+y-1 ]
multiplicity
1
3
3
3
3
4
4
5
5
TABLE 47. The decomposition of the secular equation.
Appendix I: Comment on the irreducible representations in the symmetry
group of C60
In section VIII we have computed the irreducible matrix representations of the
symmetry group of C60 . Since the tenth irreducible matrix representation is a real
representation, the generated bases are real and we can work in the real number.
Meanwhile, the four-dimensional irreducible matrix representations (computed by
189
GAP) are complex and the generated basis are complex. However, the complex
irreducible representation, in this case, can be remade into real one. The complexvalued matrix representation of a generator (Mi ) can be replaced by a real matrix:
Re(Mi ) −Im(Mi )
Im(Mi ) Re(Mi )
.
This representation is reducible, including two copies of the irreducible representation, but as the matrix group, it is isomorphic to the original group. We extract
the basis set from the columns of the projector of this matrix group and divide it
into two independent irreducible vector spaces by means of the procedure explained
thereafter in the article. Each of the separated vector space provides us the realvalued basis set and the real-valued matrix representation. One set of such real
matrix representations (for the generators r1 and r2) is given as follows:
r1:=[[1,
3/2*E(5)+1/2*E(5)^2+1/2*E(5)^3+3/2*E(5)^4,
-3*E(5)-7*E(5)^2-7*E(5)^3-3*E(5)^4,
-1081/2131*E(5)-1797/2131*E(5)^2-1797/2131*E(5)^3-1081/2131*E(5)^4],
[0,
-1/2*E(5)-E(5)^2-E(5)^3-1/2*E(5)^4,
-5*E(5)-14*E(5)^2-14*E(5)^3-5*E(5)^4,
-1681/2131*E(5)-3510/2131*E(5)^2-3510/2131*E(5)^3-1681/2131*E(5)^4],
[0,
-2*E(5)-3/2*E(5)^2-3/2*E(5)^3-2*E(5)^4,
-2*E(5)-7*E(5)^2-7*E(5)^3-2*E(5)^4,
-600/2131*E(5)-1713/2131*E(5)^2-1713/2131*E(5)^3-600/2131*E(5)^4],
[0,
190
113/44*E(5)+92/11*E(5)^2+92/11*E(5)^3+113/44*E(5)^4,
623/22*E(5)+807/11*E(5)^2+807/11*E(5)^3+623/22*E(5)^4,
5/2*E(5)+8*E(5)^2+8*E(5)^3+5/2*E(5)^4]];
r2:=[[-4/11*E(5)-3/11*E(5)^2-3/11*E(5)^3-4/11*E(5)^4,
7/22*E(5)+19/22*E(5)^2+19/22*E(5)^3+7/22*E(5)^4,
36/11*E(5)+104/11*E(5)^2+104/11*E(5)^3+36/11*E(5)^4,
383/2131*E(5)+2127/2131*E(5)^2+2127/2131*E(5)^3+383/2131*E(5)^4],
[12/11*E(5)-2/11*E(5)^2-2/11*E(5)^3+12/11*E(5)^4,
-16/11*E(5)-1/11*E(5)^2-1/11*E(5)^3-16/11*E(5)^4,
57/11*E(5)+128/11*E(5)^2+128/11*E(5)^3+57/11*E(5)^4,
-1516/2131*E(5)+1596/2131*E(5)^2+1596/2131*E(5)^3-1516/2131*E(5)^4],
[32/11*E(5)+13/11*E(5)^2+13/11*E(5)^3+32/11*E(5)^4,
-39/11*E(5)-31/22*E(5)^2-31/22*E(5)^3-39/11*E(5)^4,
53/11*E(5)+92/11*E(5)^2+92/11*E(5)^3+53/11*E(5)^4,
-1784/2131*E(5)+703/2131*E(5)^2+703/2131*E(5)^3-1784/2131*E(5)^4],
[0,0,
-623/22*E(5)-807/11*E(5)^2-807/11*E(5)^3-623/22*E(5)^4,
-2*E(5)-7*E(5)^2-7*E(5)^3-2*E(5)^4]];
We can use them in common in the two four-dimensional representations, with r3 :=
±I (the four dimensional unit matrix).
Appendix J: The GAP script in the computation of deformation in C60
As is pointed in the main article, some of the modes of deformation in C60 computed by the projectors includes several copies of proper irreducible representations.
To decompose them into separated irreducible spaces, the following function is pre191
pared. The function tries to find the irreducible subspace of dimension “REPSIZ”
in the operation of the group “R”. The function cuts out the subspace (as minimal
as possible) from the given vector space, making use of the coset decomposition of
R by the subgroups H of order “NMULTI”. In projecting out the proper irreducible
component from the multiplied copies, we employ these steps. Let W be the initial
starting place, that is, the multiplied components of the irreducible representations,
from which the separated proper irreducible components W1 , ..., Wi are computed.
The next candidate of the generator vi+1 is not taken directly from the complementary space (W \ (W1 ∪ W2 ∪ · · · ∪ Wi )); instead, in the whole vector space S
(of 180 dimension in this case), the orthogonal space to W1 ∪ W2 ∪ · · · ∪ Wi , biz,
S \ (W1 ∪ W2 ∪ · · · ∪ Wi ) is constructed. The basis set in the latter space is projected
into the irreducible representation under the consideration, and the candidate vi+1
is taken from this projected space (the renewed W). This treatment is to quicken the
symbolic computation.
gopr:=function(O,v,i,hom)
#
# When i=1, This function applies the symmetry operations of Group O
# to a vector v. And the generated vectors are summed up.
# (In fact, this function is the operation of the projector to the vector v.)
#
local irr;
irr:=Irr(O);
return Sum(List(Elements(O),j->j^irr[i]*LargeMatrix(j,hom)*v));
end;
Cutout:=function(BAS,PROJ,R,hom,NMULTI,REPSIZ)
192
#
# BAS
: THE VECTOR SPACE INCLUDING THE COPIES OF IRREDUCIBLE SPACES
# PROJ
: THE PROJECTORS
# R
: THE GROUP IN THE PROBLEM.
# hom
: THE HOMOMORPHISM FROM <R> TO THE MATRIX GROUP.
# NMULTI : THE SIZE OF THE SUBGROUP TO FORM THE COSET.
# REPSIZ : THE SIZE OF THE IRREDUCIBLE REPRESENTATION.
#
local C,D,BASN,EA,O,l,P,p,B,OV,SI,AS,OREP,ww,OT,ith,ifind;
C:=[];
D:=[];
BASN:=BAS;
EA:=List(Elements(R),j->LargeMatrix(j,hom));;
#
# The subgroups with the order of "NMULTI" are extracted.
#
AS:=AllSubgroups(R);
while (BASN<>[])
do
Print("\n",Size(BASN),"\n");
O:=Filtered(AS,g->Order(g)=NMULTI);;
Print(O);
#
# The subgroups which generate the vector space of
# the correct dimension of the irreducible representation
# are chosen.
#
193
if (O=[]) then
O:=[Group(One(R))];
fi;
l:=List([1..Size(O)],i->0);
ifind:=0;
ith:=1;
while (ifind=0 and ith<=Size(O))
do
ww:=gopr(O[ith],BASN[1],1,hom);
l[ith]:=RankMat(List(EA,g->g*ww));
Print(l);
if (l[ith]=REPSIZ) then
ifind:=1;
OT:=O[ith];
fi;
ith:=ith+1;
od;
if (ifind=0) then
Print("CHANGE THE SUPPOSITION!\n");
return 0;
fi;
#
#
By means of the chosen subgroup, the irreducible
#
vector space "B" is generated.
#
It is stored in the vector space "D".
#
B:=BaseMat(List(EA,g->g*gopr(OT,BASN[1],1,hom)));;
194
Append(C,[B]);
Append(D,B);
#
# The orthogonal space "OV" to the vector space "D" is prepared.
#
OV:=BaseOrthogonalSpaceMat(D);
#
# The vector space "BASN" to be processed is renewed now,
# by the projection "PROJ" on the vectors in "OV".
#
BASN:=BaseMat(List(OV,o->PROJ*o));
#
# Instead, it is possible that "BASN" is renewed
# as the intersection between "BASN" and "OV".
#
#
SI:=SumIntersectionMat(BASN,OV);
#
BASN:=SI[2];
#
od;
#
# Return C, the separated subspaces.
#
return C;
end;
195
[1] H. Weyl, The Theory of Groups and Quantum Mechanics,(Dover,1950).
[2] T. Inui, Y. Tanabe and Y. Onodera, Group Theory and Its applications in Physics
(Springer-Verlag,1990).
[3] M. Tinkham, Group Theory and Quantum Dynamics (Dover Books on Chemistry,2003).
[4] M. S. Dresselhaus, G. Dresselhaus, and A. Jorio, Group Theory: Application to the
Physics of Condensed Matter, (Springer-Verlag, 2008).
[5] W. Burnside, Theory of groups of finite order,(Cambridge University Press,1897).[Now
available at Project Gutenberg, <http://www.gutenberg.org/ebooks/40395>.]
[6] J. W. Thomas, Abstract Algebra: Theory and applications, <http://abstract.ups.
edu/>
[7] T. Y. Lam, Representations of finite groups: a hundred years, Notices of the AMS
(American Mathematical Society) 45(3,4):361–372 (Part I),465–474(Part II) (1998).
[8] Jean-Pierre Serre, Linear Representations of Finite Groups, (Springer-Verlag,1977).
[9] W. Fulton, J. Harris, Representation theory. A first course, Graduate Texts in Mathematics, Readings in Mathematics 129, (Springer-Verlag,1991).
[10] D. Holt, B. Eick and E. O’Brien,Handbook of Computational Group Theory, (Chapman
& Hall/CRC Press,2005)
[11] A. Hulpke, Note On Computational Group Theory, <http://www.math.colostate.
edu/~hulpke/CGT/cgtnotes.pdf>.
[12] “GAP - Groups, Algorithms, Programming - a System for Computational Discrete
Algebra”, <http://www.gap-system.org/>.
[13] J. Mooij, The vibrational spectrum of Buckminsterfullerene - An application of symmetry reduction and computer algebra, Master’s Thesis in Mathematics, Katholieke
196
Universiteit Nijmegen (2003).
[14] Michael El-Batanouny and F. Wooten, Symmetry and condensed matter physics: a
computational approach, Cambridge University Press, 2008.
[15] D. Stauffer, F.W.Hehl, N.Ito, V.Winkelmann, J.G.Zabolitzky, Computer simulation
and of principle or computer algebra: lectures for beginners, (Springer Science & Business Media, 2012).
[16] W.Hergert, M. Däne, “Group theoretical investigations of photonic Band Structures”,
Phys. Stat. Sol. (a) 197, No. 3, 620–634 (2003).
[17] W.Hergert, M. Däne, and D. Ködderitzsch, Symmetry Properties of Electronic and
Photonic Band Structures, Computational Materials Science, volume 642 of the series
Lecture Notes in Physics, pp 103–125, (Springer Science & Business Media, 2004).
[18] J.Thijssen,Computational physics, (Cambridge university press, 2007).
[19] H.Sakiyama, K.Waki, “Conformational Analysis of Hexagon’s-Thiamine Nickel(II)
Complex on the Basis of Computational Group Theory and Density Functional Theory”, J. Computed. Chem. Jpn., Vol. 13, No. 4, pp. 223–228 (2014).
[20] E. Rykhlinskaya, “A Computer-algebraic Approach to the Study of the Symmetry
Properties of Molecules and Clusters”. Doctoral thesis, Kassel University, 2006.
[21] S.Fritzshche, “Application of Point-Group Symmetries in Chemistry and Physics: A
Computer-Algebraic Approach”, International Journal of Quantum Chemistry, Vol
106, 98–129 (2006).
[22] M. P. Barnett,“Computer algebra in the life sciences”,Communications in Computer
Algebra, Vol 36, No.4, 5–32, 2002.
[23] M. I. Aroyo, J. M. Perez-Mato, D. Orobengoa, E. Tasci, G. de la Flor, A. Kirov, “Crystallography online: Bilbao Crystallographic Server”, Bulg. Chem. Commun. 43(2)
183-197 (2011).
[24] M. I. Aroyo, J. M. Perez-Mato, C. Capillas, E. Kroumova, S. Ivantchev, G.
197
Madariaga, A. Kirov and H. Wondratschek “Bilbao Crystallographic Server I:
Databases and crystallographic computing programs” Z. Krist. 221, 1, 15–27 (2006).
doi:10.1524/zkri.2006.221.1.15.
[25] M. I. Aroyo, A. Kirov, C. Capillas, J. M. Perez-Mato and H. Wondratschek “Bilbao
Crystallographic Server II: Representations of crystallographic point groups and space
groups” Acta Cryst. A62, 115–128 (2006). doi:10.1107/S0108767305040286.
[26] B. Kostant, “The Graph of the Truncated Icosahedron and the Last Letter of Galois”,
Notice of AMS, Volume 42, No. 9, 959–968, (1995).
[27] F.R.K.Chung, B. Kostant, B., S. Sternberg, “Groups and the buckyball”, in Lie Theory and geometry, Birkhauser, Boston, pp 97–126,(1994).
[28] B. Eick and B. Souvignier, “Algorithms for crystallographic groups”, International
Journal of Quantum Chemistry 106, 316–343 (2006).
[29] J. D. Dixon, “High speed computation of group characters”, Numer. Math. 10,446–
450, 1099, (1967).
[30] G. J. A. Schneider,“ Dixon’s character table algorithm revisited”, J. Symbolic Comput.9(5-6):601–606, (1990).
[31] A. Hulpke, “Zur Berechnung von Charaktertafeln”, Diplomarbeit, Lehrstuhl D für
Mathematik, Rheinisch Westfälische Technische Hochschule, (1993).
[32] C. Herring, “Character tables for two space groups”, J. Franklin, Inst. 233, 525–543
(1942).
[33] R. W. Wyckoff, The Analytical Expression of the Results of the theory of Space Groups,
Carnegie Institute: Washington, DC,(1922).
[34] O. N. Nyrasov and A. J. Freeman, “Electronic band structure of indium tin oxide and
criteria for transparent conducting behavior”, Phys.Rev.B 64,233111-1–3(2001).
[35] P. Etingof, O. Golberg, S. Hensel, T. Liu, A. Schwendner, D. Vaintrob, and E.
Yudovina, Introduction to representation theory, <http://math.mit.edu/~etingof/
198
replect.pdf>, p76–77.
[36] P. J. Webb, “An Introduction to the Cohomology of Groups”, <http://www.math.
umn.edu/~webb/oldteaching/Year2010-11/8246CohomologyNotes.pdf>.
[37] R.C.Haddon, L.E.Brus. K.Raghavachari,“Electronic structure and binding in icosahedral C60 , Chem. Phys. Lett, 125, issues 5–6, 459–464, (1986).
[38] W. H. Green, Jr, S. M. Gorun, G. Fitzgerald, P. W. Fowler, A. Ceulemans and B.
Titeca, “Electronic Structures and Geometries of C60 Anions via Density Functional
Calculations”, J.Phys.Chem, 100, 14892–14898, (1996).
[39] V. de Coulon, J.L.Martins, F.Reuse, “Electronic structure of neutral and charged C60
clusters”, Phys. Rev. B 45, No. 23, 13671–13675, (1992).
[40] G.
und
Polya,
“Kombinatorische
chemische
Verbindungen”,
Anzahlbestimmungen
Acta
Mathematica
für
Gruppen,
68(1):
Graphen
145–254(1937),
doi:10.1007/BF02546665.
[41] R. Schmied, K. K. Lehmann, “Computer-generated character tables and nuclear spin
statistical weights: Application to benzene dimer and methane dimer”,Journal of
Molecular Spectroscopy 226 201–202 (2004), DOI:10.1016/j.jms.2004.04.003.
[42] D. M. Jonas,“Spin statistics: An error in Landau and Lifschitz’ Quantum Mechanics”,
J.Chem.Phys.90(10),5563–5565 (1989).
[43] L. D. Landau and E. M. Lifschitz, Quantum Mechanics, 3rd ed. Sec.105 (Pergamon,1977).
[44] A. Kikuchi, “An approach to first principles electronic structure computation by
symbolic-numeric computation”, QScience Connect 2013:14 (2013), <http://dx.doi.
org/10.5339/connect.2013.14>.
[45] R. Courant and D. Hilbert, Method of mathematical physics, Vol.1–2,(WileyVCH,1989).
199
SUPPLEMENT: MATRIX GENERATORS OF THE POINT GROUP
In this supplement I give the matrix generators, in order to define the point group.
They are written in the GAP language.
#
# THE MATRIX IN THE POINT GROUP OF CUBIC LATTICE.
#
# THE OPERATIONS IN THE REAL SPACE.
#
CMR:=[[[1,0,0],[0,1,0],[0,0,1]],[[1,0,0],[0,-1,0],[0,0,-1]],
[[-1,0,0],[0,1,0],[0,0,-1]],[[-1,0,0],[0,-1,0],[0,0,1]],
[[0,1,0],[0,0,1],[1,0,0]],[[0,1,0],[0,0,-1],[-1,0,0]],
[[0,-1,0],[0,0,1],[-1,0,0]],[[0,-1,0],[0,0,-1],[1,0,0]],
[[0,0,1],[1,0,0],[0,1,0]],[[0,0,1],[-1,0,0],[0,-1,0]],
[[0,0,-1],[1,0,0],[0,-1,0]],[[0,0,-1],[-1,0,0],[0,1,0]],
[[0,-1,0],[-1,0,0],[0,0,-1]],[[0,-1,0],[1,0,0],[0,0,1]],
[[0,1,0],[-1,0,0],[0,0,1]],[[0,1,0],[1,0,0],[0,0,-1]],
[[-1,0,0],[0,0,-1],[0,-1,0]],[[-1,0,0],[0,0,1],[0,1,0]],
[[1,0,0],[0,0,-1],[0,1,0]],[[1,0,0],[0,0,1],[0,-1,0]],
[[0,0,-1],[0,-1,0],[-1,0,0]],[[0,0,-1],[0,1,0],[1,0,0]],
[[0,0,1],[0,-1,0],[1,0,0]],[[0,0,1],[0,1,0],[-1,0,0]],
[[-1,0,0],[0,-1,0],[0,0,-1]],[[-1,0,0],[0,1,0],[0,0,1]],
[[1,0,0],[0,-1,0],[0,0,1]],[[1,0,0],[0,1,0],[0,0,-1]],
[[0,-1,0],[0,0,-1],[-1,0,0]],[[0,-1,0],[0,0,1],[1,0,0]],
[[0,1,0],[0,0,-1],[1,0,0]],[[0,1,0],[0,0,1],[-1,0,0]],
[[0,0,-1],[-1,0,0],[0,-1,0]],[[0,0,-1],[1,0,0],[0,1,0]],
[[0,0,1],[-1,0,0],[0,1,0]],[[0,0,1],[1,0,0],[0,-1,0]],
200
[[0,1,0],[1,0,0],[0,0,1]],[[0,1,0],[-1,0,0],[0,0,-1]],
[[0,-1,0],[1,0,0],[0,0,-1]],[[0,-1,0],[-1,0,0],[0,0,1]],
[[1,0,0],[0,0,1],[0,1,0]],[[1,0,0],[0,0,-1],[0,-1,0]],
[[-1,0,0],[0,0,1],[0,-1,0]],[[-1,0,0],[0,0,-1],[0,1,0]],
[[0,0,1],[0,1,0],[1,0,0]],[[0,0,1],[0,-1,0],[-1,0,0]],
[[0,0,-1],[0,1,0],[-1,0,0]],[[0,0,-1],[0,-1,0],[1,0,0]]];;
#
# THE OPERATIONS IN THE RECIPROCAL SPACE.
#
CMT:=[[[1,0,0],[0,1,0],[0,0,1]],[[-1,0,0],[-1,0,1],[-1,1,0]],
[[0,-1,1],[0,-1,0],[1,-1,0]],[[0,1,-1],[1,0,-1],[0,0,-1]],
[[0,1,0],[0,0,1],[1,0,0]],[[0,-1,0],[1,-1,0],[0,-1,1]],
[[1,0,-1],[0,0,-1],[0,1,-1]],[[-1,0,1],[-1,1,0],[-1,0,0]],
[[0,0,1],[1,0,0],[0,1,0]],[[0,0,-1],[0,1,-1],[1,0,-1]],
[[-1,1,0],[-1,0,0],[-1,0,1]],[[1,-1,0],[0,-1,1],[0,-1,0]],
[[0,-1,0],[-1,0,0],[0,0,-1]],[[0,1,0],[0,1,-1],[-1,1,0]],
[[1,0,-1],[1,0,0],[1,-1,0]],[[-1,0,1],[0,-1,1],[0,0,1]],
[[-1,0,0],[0,0,-1],[0,-1,0]],[[1,0,0],[1,-1,0],[1,0,-1]],
[[0,-1,1],[0,0,1],[-1,0,1]],[[0,1,-1],[-1,1,0],[0,1,0]],
[[0,0,-1],[0,-1,0],[-1,0,0]],[[0,0,1],[-1,0,1],[0,-1,1]],
[[-1,1,0],[0,1,0],[0,1,-1]],[[1,-1,0],[1,0,-1],[1,0,0]],
[[-1,0,0],[0,-1,0],[0,0,-1]],[[1,0,0],[1,0,-1],[1,-1,0]],
[[0,1,-1],[0,1,0],[-1,1,0]],[[0,-1,1],[-1,0,1],[0,0,1]],
[[0,-1,0],[0,0,-1],[-1,0,0]],[[0,1,0],[-1,1,0],[0,1,-1]],
[[-1,0,1],[0,0,1],[0,-1,1]],[[1,0,-1],[1,-1,0],[1,0,0]],
[[0,0,-1],[-1,0,0],[0,-1,0]],[[0,0,1],[0,-1,1],[-1,0,1]],
[[1,-1,0],[1,0,0],[1,0,-1]],[[-1,1,0],[0,1,-1],[0,1,0]],
201
[[0,1,0],[1,0,0],[0,0,1]],[[0,-1,0],[0,-1,1],[1,-1,0]],
[[-1,0,1],[-1,0,0],[-1,1,0]],[[1,0,-1],[0,1,-1],[0,0,-1]],
[[1,0,0],[0,0,1],[0,1,0]],[[-1,0,0],[-1,1,0],[-1,0,1]],
[[0,1,-1],[0,0,-1],[1,0,-1]],[[0,-1,1],[1,-1,0],[0,-1,0]],
[[0,0,1],[0,1,0],[1,0,0]],[[0,0,-1],[1,0,-1],[0,1,-1]],
[[1,-1,0],[0,-1,0],[0,-1,1]],[[-1,1,0],[-1,0,1],[-1,0,0]]];;
#
# THE PRIMITIVE TRANSLATIONS.
#
A:=[[0,1,1],[1,0,1],[1,1,0]]/2;;
#
# TRANSFORMATION FROM MR TO MT;
#
CMT=List(CMR,m->TransposedMat(A^-1*TransposedMat(m)*A));
#
# THE MULTIPLICATION TABLE.
#
TBL:=List(CMR,i->List(CMR,j->Position(CMR,i*j)));
#
#
# THE MATRIX IN THE POINT GROUP OF HEXAGONAL LATTICE.
#
# THE PRIMITIVE TRANSLATIONS HA:=[A1,A2,A3];
#
[ Sqrt(3)/2=-1/2*E(12)^7+1/2*E(12)^11 ]
#
HA:=[[1,-1/2,0],[0,-1/2*E(12)^7+1/2*E(12)^11,0],[0,0,1]];;
#
202
# THE OPERATIONS IN THE REAL SPACE.
#
HMR:=[[[1,0,0],[0,1,0],[0,0,1]],
[[1/2,1/2*E(12)^7-1/2*E(12)^11,0],[-1/2*E(12)^7+1/2*E(12)^11,1/2,0],[0,0,1]],
[[-1/2,1/2*E(12)^7-1/2*E(12)^11,0],[-1/2*E(12)^7+1/2*E(12)^11,-1/2,0],[0,0,1]],
[[-1,0,0],[0,-1,0],[0,0,1]],
[[-1/2,-1/2*E(12)^7+1/2*E(12)^11,0],[1/2*E(12)^7-1/2*E(12)^11,-1/2,0],[0,0,1]],
[[1/2,-1/2*E(12)^7+1/2*E(12)^11,0],[1/2*E(12)^7-1/2*E(12)^11,1/2,0],[0,0,1]],
[[-1/2,1/2*E(12)^7-1/2*E(12)^11,0],[1/2*E(12)^7-1/2*E(12)^11,1/2,0],[0,0,-1]],
[[1/2,1/2*E(12)^7-1/2*E(12)^11,0],[1/2*E(12)^7-1/2*E(12)^11,-1/2,0],[0,0,-1]],
[[1,0,0],[0,-1,0],[0,0,-1]],
[[1/2,-1/2*E(12)^7+1/2*E(12)^11,0],[-1/2*E(12)^7+1/2*E(12)^11,-1/2,0],[0,0,-1]],
[[-1/2,-1/2*E(12)^7+1/2*E(12)^11,0],[-1/2*E(12)^7+1/2*E(12)^11,1/2,0],[0,0,-1]],
[[-1,0,0],[0,1,0],[0,0,-1]],[[-1,0,0],[0,-1,0],[0,0,-1]],
[[-1/2,-1/2*E(12)^7+1/2*E(12)^11,0],[1/2*E(12)^7-1/2*E(12)^11,-1/2,0],[0,0,-1]],
[[1/2,-1/2*E(12)^7+1/2*E(12)^11,0],[1/2*E(12)^7-1/2*E(12)^11,1/2,0],[0,0,-1]],
[[1,0,0],[0,1,0],[0,0,-1]],
[[1/2,1/2*E(12)^7-1/2*E(12)^11,0],[-1/2*E(12)^7+1/2*E(12)^11,1/2,0],[0,0,-1]],
[[-1/2,1/2*E(12)^7-1/2*E(12)^11,0],[-1/2*E(12)^7+1/2*E(12)^11,-1/2,0],[0,0,-1]],
[[1/2,-1/2*E(12)^7+1/2*E(12)^11,0],[-1/2*E(12)^7+1/2*E(12)^11,-1/2,0],[0,0,1]],
[[-1/2,-1/2*E(12)^7+1/2*E(12)^11,0],[-1/2*E(12)^7+1/2*E(12)^11,1/2,0],[0,0,1]],
[[-1,0,0],[0,1,0],[0,0,1]],
[[-1/2,1/2*E(12)^7-1/2*E(12)^11,0],[1/2*E(12)^7-1/2*E(12)^11,1/2,0],[0,0,1]],
[[1/2,1/2*E(12)^7-1/2*E(12)^11,0],[1/2*E(12)^7-1/2*E(12)^11,-1/2,0],[0,0,1]],
[[1,0,0],[0,-1,0],[0,0,1]]];;
#
#
203
# THE OPERATIONS IN THE RECIPROCAL SPACE.
#
HMT:=[[[1,0,0],[0,1,0],[0,0,1]],[[0,-1,0],[1,1,0],[0,0,1]],
[[-1,-1,0],[1,0,0],[0,0,1]],[[-1,0,0],[0,-1,0],[0,0,1]],
[[0,1,0],[-1,-1,0],[0,0,1]],[[1,1,0],[-1,0,0],[0,0,1]],
[[-1,-1,0],[0,1,0],[0,0,-1]],[[0,-1,0],[-1,0,0],[0,0,-1]],
[[1,0,0],[-1,-1,0],[0,0,-1]],[[1,1,0],[0,-1,0],[0,0,-1]],
[[0,1,0],[1,0,0],[0,0,-1]],[[-1,0,0],[1,1,0],[0,0,-1]],
[[-1,0,0],[0,-1,0],[0,0,-1]],[[0,1,0],[-1,-1,0],[0,0,-1]],
[[1,1,0],[-1,0,0],[0,0,-1]],[[1,0,0],[0,1,0],[0,0,-1]],
[[0,-1,0],[1,1,0],[0,0,-1]],[[-1,-1,0],[1,0,0],[0,0,-1]],
[[1,1,0],[0,-1,0],[0,0,1]],[[0,1,0],[1,0,0],[0,0,1]],
[[-1,0,0],[1,1,0],[0,0,1]],[[-1,-1,0],[0,1,0],[0,0,1]],
[[0,-1,0],[-1,0,0],[0,0,1]],[[1,0,0],[-1,-1,0],[0,0,1]]];;
#
#
# THE OPERATIONS (ROTATIONS AND TRANSLATIONS),
# AS ARE GIVEN IN THE TABLES IN THE ARTICLE.
#
# CUBIC LATTICE.(DIAMOND)
#
OPRCUB:=[[1,1,0,0,0,1,0,0,0,1,0,0,0,0,0,0],
[2,1,0,0,0,-1,0,0,0,-1,1/2,0,0,0,0,0],
[3,-1,0,0,0,1,0,0,0,-1,0,1/2,0,0,0,0],
[4,-1,0,0,0,-1,0,0,0,1,0,0,1/2,0,0,0],
[5,0,1,0,0,0,1,1,0,0,0,0,0,0,0,0],
[6,0,1,0,0,0,-1,-1,0,0,1/2,0,0,0,0,0],
204
[7,0,-1,0,0,0,1,-1,0,0,0,1/2,0,0,0,0],
[8,0,-1,0,0,0,-1,1,0,0,0,0,1/2,0,0,0],
[9,0,0,1,1,0,0,0,1,0,0,0,0,0,0,0],
[10,0,0,1,-1,0,0,0,-1,0,1/2,0,0,0,0,0],
[11,0,0,-1,1,0,0,0,-1,0,0,1/2,0,0,0,0],
[12,0,0,-1,-1,0,0,0,1,0,0,0,1/2,0,0,0],
[13,0,-1,0,-1,0,0,0,0,-1,0,0,0,1/4,1/4,1/4],
[14,0,-1,0,1,0,0,0,0,1,-1/2,0,0,1/4,1/4,1/4],
[15,0,1,0,-1,0,0,0,0,1,0,-1/2,0,1/4,1/4,1/4],
[16,0,1,0,1,0,0,0,0,-1,0,0,-1/2,1/4,1/4,1/4],
[17,-1,0,0,0,0,-1,0,-1,0,0,0,0,1/4,1/4,1/4],
[18,-1,0,0,0,0,1,0,1,0,-1/2,0,0,1/4,1/4,1/4],
[19,1,0,0,0,0,-1,0,1,0,0,-1/2,0,1/4,1/4,1/4],
[20,1,0,0,0,0,1,0,-1,0,0,0,-1/2,1/4,1/4,1/4],
[21,0,0,-1,0,-1,0,-1,0,0,0,0,0,1/4,1/4,1/4],
[22,0,0,-1,0,1,0,1,0,0,-1/2,0,0,1/4,1/4,1/4],
[23,0,0,1,0,-1,0,1,0,0,0,-1/2,0,1/4,1/4,1/4],
[24,0,0,1,0,1,0,-1,0,0,0,0,-1/2,1/4,1/4,1/4],
[25,-1,0,0,0,-1,0,0,0,-1,0,0,0,1/4,1/4,1/4],
[26,-1,0,0,0,1,0,0,0,1,-1/2,0,0,1/4,1/4,1/4],
[27,1,0,0,0,-1,0,0,0,1,0,-1/2,0,1/4,1/4,1/4],
[28,1,0,0,0,1,0,0,0,-1,0,0,-1/2,1/4,1/4,1/4],
[29,0,-1,0,0,0,-1,-1,0,0,0,0,0,1/4,1/4,1/4],
[30,0,-1,0,0,0,1,1,0,0,-1/2,0,0,1/4,1/4,1/4],
[31,0,1,0,0,0,-1,1,0,0,0,-1/2,0,1/4,1/4,1/4],
[32,0,1,0,0,0,1,-1,0,0,0,0,-1/2,1/4,1/4,1/4],
[33,0,0,-1,-1,0,0,0,-1,0,0,0,0,1/4,1/4,1/4],
205
[34,0,0,-1,1,0,0,0,1,0,-1/2,0,0,1/4,1/4,1/4],
[35,0,0,1,-1,0,0,0,1,0,0,-1/2,0,1/4,1/4,1/4],
[36,0,0,1,1,0,0,0,-1,0,0,0,-1/2,1/4,1/4,1/4],
[37,0,1,0,1,0,0,0,0,1,0,0,0,0,0,0],
[38,0,1,0,-1,0,0,0,0,-1,1/2,0,0,0,0,0],
[39,0,-1,0,1,0,0,0,0,-1,0,1/2,0,0,0,0],
[40,0,-1,0,-1,0,0,0,0,1,0,0,1/2,0,0,0],
[41,1,0,0,0,0,1,0,1,0,0,0,0,0,0,0],
[42,1,0,0,0,0,-1,0,-1,0,1/2,0,0,0,0,0],
[43,-1,0,0,0,0,1,0,-1,0,0,1/2,0,0,0,0],
[44,-1,0,0,0,0,-1,0,1,0,0,0,1/2,0,0,0],
[45,0,0,1,0,1,0,1,0,0,0,0,0,0,0,0],
[46,0,0,1,0,-1,0,-1,0,0,1/2,0,0,0,0,0],
[47,0,0,-1,0,1,0,-1,0,0,0,1/2,0,0,0,0],
[48,0,0,-1,0,-1,0,1,0,0,0,0,1/2,0,0,0]];;
#
#
# HEXAGONAL LATTICE.
#
OPRHEXA:=[[1,1,0,0,0,1,0,0,0,1,0,0,0],
[2,1/2,-Sqrt(3)/2,0,Sqrt(3)/2,1/2,0,0,0,1,0,0,-1/2],
[3,-1/2,-Sqrt(3)/2,0,Sqrt(3)/2,-1/2,0,0,0,1,0,0,0],
[4,-1,0,0,0,-1,0,0,0,1,0,0,-1/2],
[5,-1/2,Sqrt(3)/2,0,-Sqrt(3)/2,-1/2,0,0,0,1,0,0,0],
[6,1/2,Sqrt(3)/2,0,-Sqrt(3)/2,1/2,0,0,0,1,0,0,-1/2],
[7,-1/2,-Sqrt(3)/2,0,-Sqrt(3)/2,1/2,0,0,0,-1,0,0,-1/2],
[8,1/2,-Sqrt(3)/2,0,-Sqrt(3)/2,-1/2,0,0,0,-1,0,0,0],
206
[9,1,0,0,0,-1,0,0,0,-1,0,0,-1/2],
[10,1/2,Sqrt(3)/2,0,Sqrt(3)/2,-1/2,0,0,0,-1,0,0,0],
[11,-1/2,Sqrt(3)/2,0,Sqrt(3)/2,1/2,0,0,0,-1,0,0,-1/2],
[12,-1,0,0,0,1,0,0,0,-1,0,0,0],
[13,-1,0,0,0,-1,0,0,0,-1,0,0,-1/2],
[14,-1/2,Sqrt(3)/2,0,-Sqrt(3)/2,-1/2,0,0,0,-1,0,0,0],
[15,1/2,Sqrt(3)/2,0,-Sqrt(3)/2,1/2,0,0,0,-1,0,0,-1/2],
[16,1,0,0,0,1,0,0,0,-1,0,0,0],
[17,1/2,-Sqrt(3)/2,0,Sqrt(3)/2,1/2,0,0,0,-1,0,0,-1/2],
[18,-1/2,-Sqrt(3)/2,0,Sqrt(3)/2,-1/2,0,0,0,-1,0,0,0],
[19,1/2,Sqrt(3)/2,0,Sqrt(3)/2,-1/2,0,0,0,1,0,0,0],
[20,-1/2,Sqrt(3)/2,0,Sqrt(3)/2,1/2,0,0,0,1,0,0,-1/2],
[21,-1,0,0,0,1,0,0,0,1,0,0,0],
[22,-1/2,-Sqrt(3)/2,0,-Sqrt(3)/2,1/2,0,0,0,1,0,0,-1/2],
[23,1/2,-Sqrt(3)/2,0,-Sqrt(3)/2,-1/2,0,0,0,1,0,0,0],
[24,1,0,0,0,-1,0,0,0,1,0,0,-1/2]];;
SUPPLEMENT:SOME FUNCTIONS USED IN THE COMPUTATIONS
AT ”SYMMETRY IN C60 ”
#
# This file contains some functions used in the computations
# in VII. "SYMMETRY IN C60".
#
listtoh:=function(A)
207
local h,i1,i2,i3,i4,k,l,i;
h:=List([1..120],i->List([1..120],j->0));
for i in A do
i1:=i[1];
i2:=i[2];
i3:=i[3];
i4:=i[4];
k:=(i1-2)+60*(i2-1);
l:=(i3-2)+60*(i4-1);
h[k][l]:=1;
h[l][k]:=1;
od;
return h;
end;
#
# We construct an extended group by the direct product
# between A5 and one of the automorphism group.
#
# The new group contains 120 elements, represented by 62 symbols.
# The symbols from the 3rd to the 62nd are descendants from A5.
# The symbols of the first and the second are the newcomers
# in the extension.
#
# From the 62 symbols, we construct a bonding system
# which contains 120 atoms, regulated by the symmetry of the
# new group. (C120)
208
# The atoms will be indexed as (i,j) (i=3,..,62, and j=1,2)
# by the duplication of single C60.
#
#
# The generators of the group. (A5 and I)
#
g1:=(
(
1, 14, 20)(
6, 58, 22)(
2, 15, 16)(
7, 57, 23)(
3, 11, 17)(
8, 56, 24)(
4, 12, 18)(
5, 13, 19)
9, 60, 25)( 10, 59, 21)
( 31, 44, 50)( 32, 45, 46)( 33, 41, 47)( 34, 42, 48)( 35, 43, 49)
( 36, 28, 52)( 37, 27, 53)( 38, 26, 54)( 39, 30, 55)( 40, 29, 51);
g2:=(
(
1,
2,
3,
4,
8, 13, 18, 23, 28)(
5)(
6, 11, 16, 21, 26)(
7, 12, 17, 22, 27)
9, 14, 19, 24, 29)( 10, 15, 20, 25, 30)
( 31, 32, 33, 34, 35)( 36, 41, 46, 51, 56)( 37, 42, 47, 52, 57)
( 38, 43, 48, 53, 58)( 39, 44, 49, 54, 59)( 40, 45, 50, 55, 60);
g3:=(
(
1, 31)(
8, 38)(
2, 32)(
3, 33)(
4, 34)(
5, 35)(
6, 36)(
7, 37)
9, 39)( 10, 40)( 11, 41)( 12, 42)( 13, 43)( 14, 44)
( 15, 45)( 16, 46)( 17, 47)( 18, 48)( 19, 49)( 20, 50)( 21, 51)
( 22, 52)( 23, 53)( 24, 54)( 25, 55)( 26, 56)( 27, 57)( 28, 58)
( 29, 59)( 30, 60);
# I define g2 in this way in the article.
g2:=g2^-1;
#
209
G:=Group(g1,g2);
AUALL:=AutomorphismGroup(G);
p:=SemidirectProduct(Group(AUALL.1),G);
N:=Image(Embedding(p,2));
#
# The orbits of the double bond [(3,1),(4,1)]
# and the single bonds [3,1,8,1]
# and the bond between the first and the second C60
# are computed as A1,B1,C1
#
# The bonding structures from them are represented by
# three matrices a1,b1,c1.
#
A1:=Orbit(p,[3,1,4,1],OnTuples);
B1:=Orbit(p,[3,1,8,1],OnTuples);
C1:=Orbit(p,[3,1,4,2],OnTuples);
a1:=listtoh(A1);
b1:=listtoh(B1);
c1:=listtoh(C1);
#
# The symmetry operations of the 120 atoms are written as:
#
l1:=List([1..2],j->List([3..62],i->(OnPoints(i,p.1)+60*(j-1))));
l1:=Flat(l1);
210
l2:=List([1..2],j->List([3..62],i->(OnPoints(i,p.2)+60*(j-1))));
l2:=Flat(l2);
l3:=List([1..2],j->List([3..62],i->(OnPoints(i,p.3)+60*(OnPoints(j,p.3)-1))));
l3:=Flat(l3);
#
#
inv1:=List([1..60],i->OnPoints(i,g3)+2);
inv2:=inv1+60;
inv:=Flat([inv1,inv2]);
ListToMat:=function(l)
local size,h,i,j;
size:=Size(l);
h:=List([1..size],i->List([1..size],j->0));
for i in [1..size] do
h[i][l[i]]:=1;
od;
return h;
end;
#
# Now generators (p.1,p.2,p.3) are represented by matrix P1,P2,P3.
#
P1:=ListToMat(l1-2);
P2:=ListToMat(l2-2);
211
P3:=ListToMat(l3-2);
#
# In C60, of the icosahedral symmetry, the symmetry operation also includes
# the inversion. The inversion itself is not includes in the new group,
# but we can write the operation of this operation on first 60 vertexes
# (from the 3rd to the 62nd)
# The inversion in the second 60 vertexes can be written in the following way.
# The second 60 vertexes are transmuted from the first 60, by the operation p.3.
#
# If the inversions on the first and second 60 vertexes are denoted as I1,I2,
# they are related by the conjugation I2:= (p.3)*I1*(p.3)^(-1).
# The inversion on the total system is written as
# [[ I1,
#
[
0]
0, I2]].
#
#
# ListInv: the operation of "g3" in the first shell.
ListInv:=List([1..60],i->OnPoints(i,g3));
# mconjg : the operation of "p.3" from the first to the second shell.
mconjg:=List([3..62],i->OnPoints(i,p.3))-2;
#
# These operations are put into matrix, "invrs", "mc".
#
invrs:=List([1..60],i->List([1..60],j->0));
212
mc:=List([1..60],i->List([1..60],j->0));
for i in [1..60] do
invrs[i][ListInv[i]]:=1;
mc[i][mconjg[i]]:=1;
od;
invrs2:=mc*invrs*mc;
# "invrs": the inversion in the first shell.
# "invrs2": the inversion in the second shell.
#
# Set the inversions in two shells in a matrix "ivfull".
#
ivfull:=List([1..120],i->List([1..120],j->0));
for i in [1..60] do
for j in [1..60] do
ivfull[i][j]:=invrs[i][j];
ivfull[i+60][j+60]:=invrs2[i][j];
od;
od;
#
# As the matrix c1 are not conserved by the inversion "ivfull",
# (as ivfull*c1*ivfull^-1 is not equal to c1)
# it is modified to have inversion symmetry.
#
hh:=-2*a1-b1-1/2*(c1+ivfull*c1*ivfull);
213
#
# The generalized eigenvalue problems are solved in
# the field of rational numbers.
#
gev:=GeneralisedEigenvalues(Rationals,hh);
ges:=GeneralisedEigenspaces(Rationals,hh);
#
# The irreducible representation is computed.
#
# "irrg" is the group of the symmetry in the composed structure.
# "irrg2" is the subgroup.
#
irrg:=Irr(Group(P1,P2,P3,ivfull));
irrg2:=Irr(Group(P1,P2,ivfull));
Em:=Elements(Group(P1,P2,P3,ivfull));;
Em2:=Elements(Group(P1,P2,ivfull));;
Projopr:=function(irr,Em)
#
# The projector, by means of characters.
#
return List(irr,i->Sum(List(Em,j->j^i*j))/Size(Em)*i[1]);
end;
prjs:=Projopr(irrg,Em);;
214
prjs2:=Projopr(irrg2,Em2);;
#
# The basis vectors are computed and allotted to the irreducible representation.
#
bases:=List(ges,i->Basis(i));
for b in bases do;Display(List(b,j->List( prjs,p->(p*j)^2 ))); od;
for b in bases do;Display(List(b,j->List(prjs2,p->(p*j)^2 )));od;
SUPPLEMENT:SOME FUNCTIONS USED IN THE COMPUTATIONS
AT ”ANALISYS OF VIBRATIONAL MODE IN C60 ”
#
# THIS FILE CONTAINS SMALL FUNCTIONS FOR THE COMPUTATION CONCERNING C60.
# (VIII."ANALYSIS OF VIBRATIONAL MODE IN C60")
#
cc:=(E(5)+E(5)^4)/2;
ss:=(E(5)-E(5)^4)/2/E(4);
cc10:=(E(10)+E(10)^4)/2;
ss10:=(E(10)-E(10)^4)/2/E(4);
x:=Indeterminate(Rationals,"x");
z:=Indeterminate(Rationals,"z");
y:=Indeterminate(Rationals,"y");
215
#
# MATRIX OF ROTATION OF 72 DEGREES (A GENERATOR OF GROUP A5).
#
M72D:=[[cc,-ss,0],[ss,cc,0],[0,0,1]];
#
#
# HEAFTER, VERY TRIVIAL COMPUTATIONS ARE DONE
# IN ORDER TO GIVE 12 VERTEX ON ICOSAHEDRON.
#
v:=[[0,0,0],[0,0,0],[0,0,0]];
v[1]:=[0*x,0*x,x^0];
v[2]:=[x,0*x,z];
v[3]:=M72D^-1*v[2];
v21:=v[2]-v[1];
v32:=v[3]-v[2];
L21:=v21*v21;
L32:=v32*v32;
B:=[L21-L32,v[2]*v[2]-1];
GB:=GroebnerBasis(B,MonomialLexOrdering());
RTPS:=RootsOfPolynomial(GB[3]);
zval:=RTPS[2];
xval:=Sqrt(1-zval*zval);
vval:=List(v,i->List(i,j->Value(j,[x,z],[xval,zval])));
216
#
# NOW 12 VERTEX ARE PREPARED.
#
sommets:=List([1..12],i->0);
sommets[1]:=vval[1];
sommets[2]:=
vval[2];
sommets[3]:=M72D^-1*vval[2];
sommets[4]:=M72D^-2*vval[2];
sommets[5]:=M72D^-3*vval[2];
sommets[6]:=M72D^-4*vval[2];
sommets[7]:=-vval[1];
sommets[8]:=-
vval[2];
sommets[9]:=-M72D^-1*vval[2];
sommets[10]:=-M72D^-2*vval[2];
sommets[11]:=-M72D^-3*vval[2];
sommets[12]:=-M72D^-4*vval[2];
#
Tv:=TransposedMat(vval);
Tv2:=TransposedMat([vval[2],vval[3],vval[1]]);
Inv:=[[-1,0,0],[0,-1,0],[0,0,-1]];
#
# THE THREE VECTORS, STORED IN THE COLUMNS in "Tv" AND "Tv2"
# ARE RELATED BY THE ROTATION, WHICH IS OTHER GENERATOR OF A5.
#
# THE GROUP (A5 x I) , DEFINED IN 3D-EUCLID SPACE.(THE GROUP G)
217
#
# THE GENERATORS OF THIS GROUP (G.1,G.2,G.3) ARE (g1,g2,g3) IN THE ARTICLE.
#
G:=Group(Tv2*Tv^-1,M72D^-1,Inv);
irrg:=Irr(G);
elm:=Elements(G);
#
# THE PROJECTORS, EXPLANED IN THE ARTICLE.
#
Projopr:=function(irr,Em,hom)
#
return List(irr,i->Sum(List(Em,j->j^i*Image(hom,j)))/Size(Em)*i[1]);
end;
ProjoprR:=function(irr,Em,MM,hom)
#
return
List(irr,i->Sum(List(Em,j->j^i*Image(hom,j)*MM*TransposedMat(Image(hom,j))))
/Size(Em)*i[1]);
end;
#
# THE DEFINITION OF GROUP (A5 x I) BY THE PERMUTATION OF 12 VERTEX IN ICOSAHEDRON
# (GROUP F)
218
#
f1:=(1,2,3)(4,6,11)(5,10,12)(7,8,9);
f2:=(2,3,4,5,6)(8,9,10,11,12);
f3:=(1,7)(2,8)(3,9)(4,10)(5,11)(6,12);
F:=Group(f1,f2,f3);
#
# THE DEFINITION OF GROUP (A5 x I) BY THE PERMUTATION OF 60 VERTEX IN C60
# (GROUP R)
#
r1:=(
(
1, 14, 20)(
6, 58, 22)(
2, 15, 16)(
7, 57, 23)(
3, 11, 17)(
8, 56, 24)(
4, 12, 18)(
5, 13, 19)
9, 60, 25)( 10, 59, 21)
( 31, 44, 50)( 32, 45, 46)( 33, 41, 47)( 34, 42, 48)( 35, 43, 49)
( 36, 28, 52)( 37, 27, 53)( 38, 26, 54)( 39, 30, 55)( 40, 29, 51);
r2:=(
(
1,
2,
3,
4,
8, 13, 18, 23, 28)(
5)(
6, 11, 16, 21, 26)(
7, 12, 17, 22, 27)
9, 14, 19, 24, 29)( 10, 15, 20, 25, 30)
( 31, 32, 33, 34, 35)( 36, 41, 46, 51, 56)( 37, 42, 47, 52, 57)
( 38, 43, 48, 53, 58)( 39, 44, 49, 54, 59)( 40, 45, 50, 55, 60);
r3:=(
(
1, 31)(
8, 38)(
2, 32)(
3, 33)(
4, 34)(
5, 35)(
6, 36)(
7, 37)
9, 39)( 10, 40)( 11, 41)( 12, 42)( 13, 43)( 14, 44)
( 15, 45)( 16, 46)( 17, 47)( 18, 48)( 19, 49)( 20, 50)( 21, 51)
( 22, 52)( 23, 53)( 24, 54)( 25, 55)( 26, 56)( 27, 57)( 28, 58)
( 29, 59)( 30, 60);
R:=Group(r1,r2,r3);
#
# THE ISOMORPHISM BETWEEN GROUPS
219
#
iso:=IsomorphismGroups(F,R);
iso2:=IsomorphismGroups(F,G);
hom:=GroupHomomorphismByImages(F,G,[F.1,F.2,F.3],[G.1,G.2,G.3]);
hom2:=GroupHomomorphismByImages(R,G,[R.1,R.2,R.3],[G.1,G.2,G.3]);
Display(List(Irr(R),r->List([R.1,R.2,R.3],i->i^r)));
MM:=[[0*x,x,y],[-x,0*x,z],[-y,-z,0*x]];
PRJ:=Projopr(Irr(R),Elements(R),hom2);
PRJR:=ProjoprR(Irr(R),Elements(R),MM,hom2);
#
# THE FUNCTIONS DEFINED AND EXPLAINED IN THE ARTICLE.
#
#
ListToMat:=function(l)
local size,h,i,j;
size:=Size(l);
h:=List([1..size],i->List([1..size],j->0));
for i in [1..size] do
h[i][l[i]]:=1;
od;
return h;
end;
LargeMatrix:=function(rr,iso)
220
local l,rmat,rrimg,i,i1,i2,n,m;
l:=ListPerm(rr,60);
rmat:=List([1..Size(l)*3],i->List([1..Size(l)*3],j->0));
rrimg:=Image(iso,rr);
for i in [1..Size(l)]
do
i1:=i-1;i2:=l[i]-1;
for n in [1..3]
do
for m in [1..3]
do
rmat[3*i1+n][3*i2+m]:=rrimg[n][m];
od;
od;
od;
return rmat;
end;
#
# THE DEFINITION OF PROJECTOR, DEFINED IN THE ARTICLE
# IN ORDER TO COMPUTE THE MODE OF DEFORMATION IN C60.
#
ProjoprE:=function(irr,Em,hom2)
#
local A,B;
#A:=List(Em,j->LargeMatrix(j,hom2));
221
#B:=List(irr,i->List(Em,j->j^i));
return List(irr,i->Sum(List(Em,j->j^i*LargeMatrix(j,hom2)))/Size(Em)*i[1]);
end;
PRJE:=ProjoprE(Irr(R),Elements(R),hom2);
BAS:=List(PRJE,i->BaseMat(TransposedMat(i)));
BASSIZE:=List(BAS,Size);
CROT:=List(Elements(R),i->Trace(ListToMat(ListPerm(i,60))));
CIRR:=List(Irr(R),r->List(Elements(R),i->i^r));
# In the trace representation the basis set composed from column vectors
# of the projector may include the multiplicated spaces of
# the representation. The basis vectors might be transmuted
# with each other transitively, not confined in the one of
# the component of the multiplicated irreducible spaces.
# The basis set of the decomposion into the irreduciple
# representations and the conjugacy classes are prepared.
cng:=ConjugacyClasses(R);;
rcng:=List(cng,Representative);;
rcng2:=List(rcng,j->LargeMatrix(j,hom2));;
charact:=function(g,B)
222
return Trace(B*g*TransposedMat(B)*(B*TransposedMat(B))^-1);
end;
gopr:=function(O,v,i,hom)
local irr;
irr:=Irr(O);
return Sum(List(Elements(O),j->j^irr[i]*LargeMatrix(j,hom)*v));
end;
Cutout:=function(BAS,PROJ,R,hom,NMULTI,REPSIZ)
#
# BAS
: THE VECTOR SPACE INCLUDING THE COPIES OF IRREDUCIBLE SPACES
# PROJ
: THE PROJECTORS
# R
: THE GROUP IN THE PROBLEM.
# hom
: THE HOMOMORPHISM FROM <R> TO THE MATRIX GROUP.
# NMULTI : THE SIZE OF THE SUBGROUP TO FORM THE COSET.
# REPSIZ : THE SIZE OF THE IRREDUCIBLE REPRESENTATION.
#
local C,D,BASN,EA,O,l,P,p,B,OV,SI,AS,OREP,ww,OT,ith,ifind;
C:=[];
D:=[];
BASN:=BAS;
EA:=List(Elements(R),j->LargeMatrix(j,hom));;
#
# The subgroups with the order of "NMULTI" are extracted.
#
223
AS:=AllSubgroups(R);
while (BASN<>[])
do
Print("\n",Size(BASN),"\n");
O:=Filtered(AS,g->Order(g)=NMULTI);;
Print(O);
#
# The subgroups which generate the vector space of
# the correct dimension of the irreducible representation
# are chosen.
#
if (O=[]) then
O:=[Group(One(R))];
fi;
l:=List([1..Size(O)],i->0);
ifind:=0;
ith:=1;
while (ifind=0 and ith<=Size(O))
do
ww:=gopr(O[ith],BASN[1],1,hom);
l[ith]:=RankMat(List(EA,g->g*ww));
Print(l);
if (l[ith]=REPSIZ) then
ifind:=1;
OT:=O[ith];
fi;
ith:=ith+1;
224
od;
if (ifind=0) then
Print("CHANGE THE SUPPOSITION!\n");
return 0;
fi;
#
#
By means of the chosen subgroup, the irreducible
#
vector space "B" is generated.
#
It is stored in the vector space "D".
#
B:=BaseMat(List(EA,g->g*gopr(OT,BASN[1],1,hom)));;
Append(C,[B]);
Append(D,B);
#
# The orthogonal space "OV" to the vector space "D" is prepared.
#
OV:=BaseOrthogonalSpaceMat(D);
#
# The vector space "BASN" to be processed is renewed now,
# by the projection "PROJ" on the vectors in "OV".
#
BASN:=BaseMat(List(OV,o->PROJ*o));
#
# Instead, it is possible that "BASN" is renewed
# as the intersection between "BASN" and "OV".
#
#
SI:=SumIntersectionMat(BASN,OV);
225
#
BASN:=SI[2];
#
od;
#
# Return C, the separated subspaces.
#
return C;
end;
H1:=Cutout(BAS[9],PRJE[9],R,hom2,5,5);
#
# TO MAKE THE MATRIX NEEDED IN THE COMPUTATION OF MATRIX CONJUGATION
# EXPLAINED IN THE ARTICLE.
#
ConjugMatrix:=function(rr)
local l,rmat,vmat,i,i1,i2,n,m;
l:=ListPerm(rr,60);
vmat:=List([1..Size(l)*3],i->List([1..Size(l)*3],j->0));
for i in [1..Size(l)]
do
i1:=i-1;i2:=l[i]-1;
for n in [1..3]
do
vmat[3*i1+n][3*i2+n]:=1;
od;
226
od;
return vmat;
end;
227
| 5cs.CE
|
arXiv:1706.03122v1 [cs.AI] 9 Jun 2017
Off The Beaten Lane:
AI Challenges In MOBAs Beyond Player Control
Michael Cook
Adam Summerville
Simon Colton
Games Academy
Falmouth University
Expressive Intelligence Studio
UC Santa Cruz
Games Academy
Falmouth University
Abstract
MOBAs represent a huge segment of online gaming and are
growing as both an eSport and a casual genre. The natural
starting point for AI researchers interested in MOBAs is to
develop an AI to play the game better than a human - but
MOBAs have many more challenges besides adversarial AI.
In this paper we introduce the reader to the wider context of
MOBA culture, propose a range of challenges faced by the
community today, and posit concrete AI projects that can be
undertaken to begin solving them.
Introduction
Multiplayer Online Battle Arenas, a clumsy phrase shortened
to MOBA, describes a growing genre of videogames typically
designed as highly competitive eSports. Many of them are
part of a family tree that traces back to Warcraft 3 mods such
as DOTA Allstars, and now form a contingent of the most
popular and most player games of today, including League of
Legends, DOTA 2 and Heroes Of The Storm.
MOBAs are huge and rapidly expanding in every sense in terms of their cultural impact on games, in terms of their
financial impact on the industry, and in terms of their impact
on how the games community is spending its time. As a
result, they are attracting increasing interest from researchers,
both those interested in building technological systems into
the game, and those interested in studying the many different
stakeholder groups in the community.
In an age of deep learning and AIs making headlines, the
obvious target for AI researchers to study would be the implementation of bots that can play the game at human or
superhuman level. Google DeepMind have already stated
that their next objective will be to develop a system that can
play Starcraft 2, a competitive videogame with similarities
to MOBAs at least culturally if not mechanically. We believe
that the development of player-competitive AI represents just
one small challenge offered by the complex world of MOBA
playing, but that many of these challenges are hard for researchers to access because of the vast knowledge barriers
that exist when trying to understand the genre or what makes
it interesting.
In this paper we try to pick apart interesting challenges
in and around the MOBA genre. Some of them touch upon
deeply technical problems within the game, others relate to
ways that AI can assist in the emerging activities that happen
around MOBA games, especially relating to the professional
eSports scene. We do our best to present these challenges with
the minimum of context for the reader, so that researchers
whose background may not be in strategy games or MOBAs
specifically can hopefully see why these problems are interesting and hopefully motivate more diverse research in the
area.
The remainder of the paper is organised as follows: in
Background we provide a minimal introduction to MOBAs,
primarily focusing on one specific game, DOTA 2 due to the
particular experience of the authors. We discuss the game
itself, the professional scene, and surrounding culture. In
Challenges we outline some important challenges we have
identified throughout the MOBA genre, and provide specific
ideas about projects that could be conducted to help work towards solving or assisting humans in tackling these problems.
In Existing & Related Work we discuss some of the work already done in the MOBA genre, as well as highlighting some
work beyond MOBAs that could contribute to the challenges
in this paper.
Background - DOTA 2
DOTA 2 is a complicated game with a large amount of intersecting systems, lists of specific knowledge and historical
trends. For the purposes of this paper we will be focusing
on a light coverage of key themes and ideas, sufficient to
understand the challenges we present in the next section. Additionally, although DOTA 2 is a game about exceptions and
special cases, we will make generalisations in this section in
order to simplify the high-level description of the game.
Gameplay - Drafting
A single game typically has two phases – a drafting phase
where players choose heroes to play, and a gameplay phase
where the main game takes place. The nature of the draft
phase depends on the game mode. In this paper we will primarily discuss Captain’s Mode, a game mode used in professional play but less popular with casual players. In Captain’s
Mode each player nominates a captain who chooses heroes
for their team. The two captains go through a fixed order of
picking heroes (which adds them to their team) and banning
heroes (which stops either team from picking them). A hero
can only be picked once. Figure 1 shows a screenshot from
Figure 1: The drafting phase from a match between Team
Empire and OG. Picked heroes appear in large portraits, while
bans are listed on the left in smaller images. Team Empire
are about to pick their fourth hero.
a professional game’s drafting phase. Drafting is complex
enough to warrant its own paper. Indeed, professional players often turn up to tournaments with ‘drafting bibles’ filled
with pages of notes about enemy team strategies, their own
prepared ideas for drafts, and historical information about
synergies and counter-strategies.
There are 112 heroes in the game in total, and each one
has a particular combination of skills and strengths, some
of which may overlap or synergise with other heroes. For
the purposes of this paper, we will simplify heroes into two
kinds: support heroes and core heroes. Support heroes are
strong in the early parts of a game, typically because they
have damaging spells or spells with special effects (such as
a stun, which temporarily stops a hero from acting). Core
heroes have properties which make them better later in the
game, often because they either need to buy a particular
item to become effective, or because they scale up faster
than other heroes. For example, one hero is able to attack
up to five targets simultaneously. This means that an item
which provides increased damage is potentially five times
as valuable on that hero. The number of cores selected for
a team, and how much time they need to spend collecting
resources, affects the kind of strategies available to a team
during the game itself.
Gameplay - Match
The primary objective in a game of DOTA 2 is to destroy the
opposing team’s Ancient, a structure protected in the centre
of each team’s base. Figure 2 shows an annotated overheap
drawing of the entire DOTA 2 map1 with the Ancients represented as circles in opposite corners. Connecting the two
bases are three pathways called lanes, two around the edges
of the map and one through the centre. Periodically, NPC
creatures called creeps spawn from both bases and travel
along these lanes, attacking each other when they meet.
Towers are also placed along these lanes, indicated by
squares in Figure 2. Towers are strong structures which at1
This map is from an earlier version of the game but is used here
as the key details remain the same.
Figure 2: An annotated map from a version of DOTA 2.
Arrows indicate the directions creeps move along the three
lanes. Squares indicate the location of towers, while the large
circle is each team’s Ancient.
tack nearby enemy units, provide vision for their team, and
help define the borders of a team’s territory. Towers are invulnerable to damage unless they are the outermost tower in
a lane, and the Ancient is invulnerable to damage unless one
of the three towers at the end of a lane are destroyed. Thus,
destroying towers in lanes is an important marker of overall
progress in the game, and many strategies exist to achieve
this, including some which intentionally avoid fighting enemy
heroes in favour of destroying buildings.
Games are colloquially broken up into three phases. The
first phase is the early game, in which core players are focusing on obtaining gold and experience points. During this
phase support players may protect their team’s cores, or they
may attempt to move around the map and attack the opposing
team’s cores. The second phase is the mid-game, where some
cores have reached a point where they can begin to dictate the
game’s direction. This might involve grouping with support
heroes to destroy towers, or moving as a larger group to fight
enemy heroes. Some cores may continue to obtain gold and
experience if they need more than other cores. The final phase
is the late game. At this point heroes have generally stopped
focusing on resources and have important items and skills.
This phase of the game is much less predictable as players
tend to have less experience of it (all matches have an early
game, but only some last long enough to enter the late game).
This phase involves a larger amount of movement, fighting
and strategic decision-making.
Professional Play
In 2011 Valve announced DOTA 2 by organising an global
tournament called The International, in which sixteen profes-
sional DOTA 1 teams were invited to compete for a share of
$1.6m, the largest prize pool in eSports at the time. In 2015
The International’s fifth incarnation had a prize pool of over
$18m. In the intervening years the professional scene has
grown to become both an aspiration for young players and
a big business for sponsors and organisers. Most of the top
teams are run by independent organisations, have multiple
sponsors, and employ managers, coaches, dedicated analysts,
PR and various other staff.
Most DOTA 2 tournaments, from the smallest amateur
regionals to the highest-tier multi-million viewer events, are
broadcast online to watch for free, with casters providing
commentary, analysis and statistics on the matches. This is
common across most eSports, including Heartstone, Starcraft 2 and other MOBAs like League of Legends. A common
structure for an eSports broadcast has two distinct panels
of people: an analysis desk which discusses games during
downtime between matches; and a caster desk which describes and comments on the action during a game. Larger
broadcasts may also employ statistics analysts who listen to
commentary and provide relevant statistics about past events
or current trends, as well as dedicated observers responsible
for controlling the in-game camera based on commentator
interest.
Challenges
In this section we describe challenges besides AI character
control that we believe will become important and fruitful
areas of research in the next few years. In each case we
provide a little additional background where necessary, state
the nature of the problem and how it is currently tackled (if at
all) and then propose possible routes for AI research to take.
Commentary
Commentary is broken into three distinct areas, each with
their own unique challenges - draft analysis, play-by-play
and hypecasting.
Draft Analysis During the draft the analysis desk typically
discusses the heroes being picked and banned, the wider
context of the game and how it reflects the current metagame,
and how they expect the game itself to play out when the
draft is complete. A draft timer limits the amount of time
teams can spend picking heroes, but teams tend to maximise
their use of this time, meaning drafts are often slow-paced
and thus are a good area for researchers to focus on initially
for commentary generation and assistance.
Good predictive models for which heroes are likely to
be picked or banned are fundamental to AI draft analysis.
Machine learning is likely to perform well here, but the problem is complicated by the relative lack of data on professional matches. Towards the end of a particular patch cycle
(discussed later) there may be a few thousand matches to
pull from at most. More specific predictions may have much
smaller datasets – data on a specific team, for example, may
include only a few dozen matches. Hybrid techniques may be
needed to provide useful analyses at different levels of detail,
employing decision trees or linear regressions for smaller
data pools.
A secondary problem for good draft analysis, which also
applies to statistics provided during the game, is selection.
A vast quantity of data is available for DOTA 2, from the
heroes drafted to the exact time at which a player bought an
item in a specific game. Facts of all shapes and sizes can be
discovered – records set by players, trends or preferences of
certain captains, historical stories and expected performances
– but selecting which are most appropriate for the current
discussion is a separate challenge. This is partly a natural
language processing problem, monitoring the conversation
between analysts to dynamically rank facts and statistics for
relevance, but models of ‘interestingness’ are also important.
Subjective interestingness, and notions of ‘actionability’ and
‘unexpectedness’ from the knowledge discovery domain (Silberschatz and Tuzhilin 1995), may help inform models for
statistic selection.
Play-By-Play Commentating a game involves solving several different problems simultaneously and in real time. First
and foremost, commentators must identify the most salient
events happening in the game and describe them succinctly.
During moments where many heroes are fighting together
the challenge comes from identifying the most important
features of the fight – which player actions are having the
most impact, which actions are causing important changes
in who has advantage. A major challenge here is ranking
the actions of ten players simultaneously and being able to
predict the next few seconds of action (since it takes time
to commentate on an action, during which time more events
have taken place).
When large fights are not happening, the choice of what to
discuss becomes broader. Commentators may talk about the
emerging themes in the game – what strategies the teams are
employing, how those strategies are faring, or the individual
performances of a particular player. They may also identify
what teams are planning to do next, based on the behaviour of
players currently. In periods of downtime commentators may
also discuss the game in the wider context of the tournament
or series it is in – how it reflects the trends elsewhere in
the tournament, or how the winner of this match might fare
against the other teams they are likely to face next.
Hypecasting & Wordplay It’s tempting to consider commentary as being purely concerned with conveying the facts
of the game to the viewer, but a major part of MOBA commentary is in providing a sense of playfulness and excitement
to proceedings. For some casters this is achieved through a
sense of power and intensity in their delivery – Tobi Dawson
is famous for his high-energy commentary and emotive descriptions of play. Other casters employ colourful language,
phrasing and inventive wordplay to entertain viewers. Below is an excerpt from the grand final of The International
in 2012, where David Gorman is addressing his co-caster,
David ‘Lumi’ Zhang.
There’s a freight train running down the tracks and it’s
about to hit a car, and let me tell you Lumi, the car gives
way – not the freight train.
Inventing clever descriptions for events and coining names
for teams or players links in well with existing research into
metaphor, analogy and humour in computational creativity.
Work in (Ritchie et al. 2007) lays out ways to create humour
from linguistic collisions of concepts, and (Veale and Alnajjar
2015) describes systems for making perceptual connections
between the real world and linguistic constructions. This
shows that we can do more with language than simply state
what is happening on screen, and research into AI casters
should endeavour to go beyond this and be as engaging and
innovative with language as human casters are.
Camera Control
Automated Camera Control In the early days of eSports
commentators would broadcast their screen directly to viewers, and simultaneously describe what was happening as they
tried to smoothly direct the camera’s view of proceedings.
Today, the bigger tournaments and studios hire dedicated
observers whose only job is to control the camera and focus
on the most relevant action. Camera control is much harder
in eSports compared to traditional sports as games often lack
a single point of focus (such as a ball). Deciding what to
show and how to show it is a difficult task. Tools that can
help automate or assist in this process not only help high-end
broadcasters, but also help amateur broadcasters and casual
spectators a chance to focus on the game itself rather than
controlling a camera.
There is also a healthy body of research relating to camera
control for other game genres – in (Yannakakis, Martı́nez,
and Jhala 2010) the authors specifically consider the affective
qualities of automated camera control, which is highly relevant for framing different kinds of action in a hectic MOBA
game, while in (Burelli and Preuss 2014) the authors look
at camera control as a multi-objective optimisation problem
which is particularly appropriate for the dynamic action in
spectator eSports.
Highlighting An emerging problem for eSports design in
general is a desire for succinct highlighting of matches. This
is something that game designers are showing interest in
not only at the professional level, highlighting important
moments from top matches, but at the casual level where
individual games may be able to produce their own highlight reels. DOTA 2 can be asked to automatically provide
highlights of any game replay file, although the process is
rudimentary and has not been improved in years.
Highlighting is a mix of problems from camera control
and commentary, with an additional pressure to present the
results in an engaging and exciting way. The most interesting and significant moments from a game must be extracted,
and framed in a way that best displays the action at hand.
While many AI techniques could be used here, one promising possibility would be to leverage Twitch chat as a form
of supervised learning. Twitch, a popular streaming service
used to broadcast most professional DOTA 2 live, has a prolific (and, we should stress, often toxic) chat community. At
times of intense action, humour, surprise or skill chatters
often simultaneously send certain emotes in large quantities.
These emote surges can be used to automatically label events
in certain professional replays2 , which could then be cross2
Inspired by http://www.skip2.tv which implemented an emote
Figure 3: A heatmap for a single player’s movement during
the first ten minutes of a DOTA 2 match.
referenced with replay data to train systems to recognise what
causes such a strong reaction. These models could then be
reapplied to ordinary games as an automated highlighter.
Vision
An observer ward is a cheap but limited-quantity item that
can be placed on the map to provide vision in a 360 circle,
obscured by certain map features like trees or cliffs. Observer
wards are invisible, but can be detected and destroyed by certain items. If not destroyed, they expire naturally after seven
minutes. Because observer wards are limited in number but
destroyable, a tension exists between placing wards in good
places while avoiding becoming predictable. Professional
teams often study how certain players place wards in order
to predict their behaviour in future games.
This is an example of a tightly defined system within
DOTA 2 that is ripe for analysis and optimisation. Publicly
available replay data for any DOTA 2 match, including professional games, includes data about hero movement and
ward placement, allowing a complete model to be built for
which players were visible at what times, and what wards
are commonly placed. Figure 3 shows a heatmap of a single
player’s movement in the early game phase of a match. A
tool to analyse the warding habits of a team to predict placement in an upcoming game, or to analyse common player
movements and suggest ward placements for optimal coverage, would be useful for all players from new learners to
professional teams.
Existing research into modelling player vision may inform
further work here. For example, in (Tremblay, Torres, and
Verbrugge 2014) the authors analyse risk in traversing stealth
game levels. The authors have published a number of papers
counter to identify exciting moments in Twitch streams
on related topics in stealth, which relate well to challenges
related to vision and adversarial player movement in MOBAs.
Patch Changes
Like many videogames, MOBAs are frequently patched to fix
bugs, rebalance game features and add new content. Because
games like DOTA 2 are highly competitive, these patches
also serve as a way to adjust the metagame by improving
(or ‘buffing’) heroes, items and spells which are not used
often, and weakening (or ‘nerfing’) heroes, items and spells
which are overused or too powerful. Besides targeting specific
entities in the game, patches can also adjust the way the game
is played on a larger scale by targeting certain strategies or
styles of play, adjusting the map geometry, or changing the
order of drafting.
New patches are significant moments in the history of the
game. A major patch changes hundreds of game elements,
from the cost of an item to the number of seconds a skill is
on cooldown for after use. The first matches and tournaments
played professionally after a new patch typically involve a lot
of speculation and experimentation, and previously dominant
teams and players can suddenly find themselves scrambling
to compete as new strategies are stumbled upon or hypotheses
are tested.
Patches also represent a serious challenge for any AI system that relies on an understanding of the metagame (including player bot AI, commentary AI, draft analysis and
more). The game fundamentally changes overnight in numerous ways, and existing archives of play data (and systems
trained on them) no longer represent how the game currently
plays. Two major challenges stand out in this area:
Assessing Patch Impact In the days and weeks following
a patch there is a scramble to understand what the overall effect the patch will have on the metagame. This understanding
is vital for commentators to discuss in games, for teams to
gain an edge on competitors, and for designers to assess the
efficacy of their changes. Sometimes changes have obvious
effects – if the cost of an item increases, that item will be
harder to purchase. Other effects can be subtle and may take
weeks to emerge – a buff to a particular item leads to a hero
becoming stronger, thus picked more often, which in turn
increases the value of a second hero whose primary purpose
is to counter them. Being able to predict these chains or identify trends and strong strategies ahead of time is extremely
valuable.
We believe that building a simple forward model for DOTA
2’s combat may help assess the impact of small changes.
Combining such a model with simple AI agents, replay data
from the previous patch can be resimulated with the addition
of one or more changes from the new patch, having AI agents
take over at the point where the simulation diverges from the
past data (for example, a patch change causes someone to live
where they had died in the replay data, because the damage
they took was reduced). From this, we can assess simple
surface-level impacts from the patch such as changes to the
efficacy of items or skills. While this is unlikely to provide a
deep assessment of a patch’s impact on a hero, it may provide
an indication on a micro-scale about what impact a patch is
likely to have.
Predicting Patch Content Towards the end of a patch cycle the metagame often becomes stagnant – highly-valued
heroes are regularly banned and picked, many teams use the
same strategies, and play approaches a state of equilibrium.
A new patch will rectify this by rebalancing the game as
described above, forcing strategies to be re-evaluated and a
long process of experimentation and discovery to take place.
A system that can assess the current metagame and predict
or suggest patch changes is valuable both as a design tool
(for designers, in suggesting balance changes or helping alter
popular strategies) and as a competitive tool for teams to
predict what new trends may emerge in the next patch, in
order to prepare for them.
AI learning approaches might be able to infer potential
patch changes based on trends in previous patches (heroes are
often buffed or nerfed in successive patches until a particular
effect is achieved). We also believe the use of a forward
model, as suggested above, might be able to predict possible
changes based on items, heroes and strategies which are
overused or have statistically abnormal winrates. Analogies
may also be made between previous patch changes and their
root causes. If an item was previously rebalanced after having
a high winrate or being bought too often, items following
similar patterns of usage might be candidates for similar
rebalancings in a future patch.
Inventing Techniques & Discovering Exploits
Many of the crucial mechanical systems that are now built
into DOTA 2 and many other MOBAs originated as bugs,
exploits or emergent behaviour found by players in the game.
For example, in DOTA 2 players earn gold by performing the
killing blow on a non-hero creature. An unintentional feature
of DOTA’s original implementation in Warcraft 3 allowed
players to attack their own creeps to kill them. Players used
this feature to invent the notion of ‘denying’, since attacking
their own creeps stops their opponents from gaining gold
(they received no gold themselves from doing this). This is
now a fundamental part of contesting players for resources
in DOTA 2, and recognised in the game through statistical
tracking and its inclusion in tutorials.
Such discoveries still happen today, as patches introduce
new items, skills, systems and interactions. Discovering new
mechanics has benefits for many groups: it gives players and
teams a temporary edge, and allows developers to fix true
exploits and adjust interesting ones (denying was balanced
multiple times as it became an official mechanic, while other
games like League of Legends removed it entirely). An AI
system that can curiously explore a MOBA’s game systems
to uncover new interactions and beneficial effects would be
highly valued.
This challenge bears similarities to automated playtesting
research (Zook, Fruchter, and Riedl 2014) – in essence, many
of these interesting discovered systems can be regarded as
bugs or unintentional side effects of the game’s intentionallydesigned systems. The special case we are considering here
is whether the discovered exploits provide some kind of competitive advantage or inspire a new strategy in playing the
game, perhaps similar to objective-driven mechanic discovery in platformers (Cook et al. 2013). The most significant
challenge here is in identifying progress or utility in a discovered system. Often these are highly innovative and creative in
their application, which is the reason why they take so long
to find despite hundreds of thousands of games taking place
every day. Independently curious agents (Saunders 2002)
playing the game as part of automated bot matches, ignoring
objectives and instead seeking novel game interactions might
yield interesting results.
Abuse
Many MOBAs are designed to be highly competitive environments which emphasise player skill, improvement and
ultimately mastery. The framing of these games as sports is
further enhanced by the tight links with the professional scene
in which the most popular players are put forward as role
models, as well as a relentless stream of statistics and records
tracking the performance of players and their ranking among
their peers. This atmosphere, combined with DOTA 2’s high
barrier to entry and intimidating learning curve, makes the
community a pressure cooker of negative emotions. Abusive
players are a major problem in MOBAs.
The only attempt made to curb this behaviour is a playerrun reports system which relies upon the playerbase to inform
Valve when a player is acting abusively. However, this system
is more often used to report players who are perceived to be
playing badly, itself a form of abuse. Riot Games, developer
of the MOBA League of Legends, have reportedly3 applied
machine learning techniques to automatically detect abusive
language, although more complex problems like sarcasm,
passive-aggressiveness, or abusive in-game behaviour (to intentionally sabotage another player or their own team) remain
difficult to detect.
Existing research has looked into textual abuse in games
(Kwak and Blackburn 2015), although much work remains
to be done. The task of detecting abusive behaviour, however,
including bots, scripts and intentional ability abuse is still
a relatively open problem. Progress in this area benefits all
games, but the richness and volume of data focused around
the same structured activities makes MOBAs an appealing
place to start in this case.
Related Work
Existing work investigating DOTA 2 as a domain for AI research primarily focus on the task of building AI agents to
play the game. This is an extremely appealing challenge
for adversarial AI research – playing the game requires
both micro-scale decision-making with fast reactions, and
macro-scale strategic planning. It also has an aspect lacking
in many of the recent fashionable AI challenge domains other players. DOTA 2 is fundamentally a game about team
co-operation, cohesion and communication, and this adds
a much less-examined, more complex dimension to the AI
problem.
The research questions which have received the most attention work with readily-available data that can be extracted
3
http://tinyurl.com/riotresearch
from public replays, such as analyses of the heroes making up
a team. In (Pobiedina et al. 2013) the authors analyse whether
a more balanced distribution of roles within a team improves
a team’s chance of winning, a fundamental part of selecting
heroes that works well together. (Drachen et al. 2014) takes
a different approach, assessing the position of heroes and
how they change throughout the game as an indicator of team
performance (high movement and team grouping can be an
indicator of success in some metagames). Another study in
(Yang, Harrison, and Roberts 2014) focuses on the relationship between hero roles and how these develop throughout
different phases of the game.
Existing research of this kind may be repurposed to tackle
some of the challenges outlined in this paper – for example,
understanding the phases of gameplay and the expected behaviour from different roles, as shown in (Yang, Harrison,
and Roberts 2014), can provide a useful baseline of knowledge for commentators, provided it can change to keep up
with patches and the metagame (something which, in general,
existing work does not address). Similarly, research outside
the AI domain but focused on MOBAs, such as the analysis
of pedagogical professional streams in (Georgen, Duncan,
and Cook 2015), will help inform the creation of AI tools
that teach and relay information to players and spectators.
Conclusions
In this paper we outlined a diverse range of challenges offered
by the MOBA genre, using DOTA 2 as an illustration. They
show that the genre has more to offer than simply a harder
adversarial AI problem, with emerging challenges that pose
problems for machine learning, natural language processing
and generation, computational creativity, player modelling,
and more. We proposed concrete projects and potential avenues for them, and discussed existing work that might relate
to these challenges. MOBAs are a Frankensteinian genre, a
patchwork of ideas and systems from RPGs, action games,
sports, roguelikes, boardgames and more. As researchers
come to tackle the problems posted by the genre, it will be
crucial to pull in ideas, systems and inspiration from a wide
sampling of research areas.
MOBAs are time-consuming games to learn, taking hundreds of hours to gain an average understanding of the game’s
main principles. One of the many ways in which this affects
the game is that it makes it very difficult for researchers to
come to this area as an outsider and apply their knowledge
and expertise to it. Yet the genre contains within it a multitude of exciting research opportunities, vast stores of publicly
available data, and professionals eager to work with new
technology and ideas. Both the game’s community and our
research community must find ways to make these challenges
amenable to people and make the genre a more accessible
proposition for those who are not already experienced with
it.
References
[Burelli and Preuss 2014] Burelli, P., and Preuss, M. 2014.
Applications of Evolutionary Computation: 17th European
Conference, EvoApplications 2014, Granada, Spain, April
23-25, 2014, Revised Selected Papers. Berlin, Heidelberg:
Springer Berlin Heidelberg. chapter Automatic Camera Control: A Dynamic Multi-Objective Perspective, 361–373.
[Cook et al. 2013] Cook, M.; Colton, S.; Raad, A.; and Gow,
J. 2013. Applications of Evolutionary Computation: 16th European Conference, EvoApplications 2013, Vienna, Austria,
April 3-5, 2013. Proceedings. Berlin, Heidelberg: Springer
Berlin Heidelberg. chapter Mechanic Miner: ReflectionDriven Game Mechanic Discovery and Level Design, 284–
293.
[Drachen et al. 2014] Drachen, A.; Yancey, M.; Maquire, J.;
Chu, D.; Wang, Y.; Mahlmann, T.; Shubert, M.; and Klabjan,
D. 2014. Skill-Based Differences in Spatio-Temporal Team
Behaviour in Defence of The Ancients 2 (DotA 2). United
States: IEEE. 1–8.
[Georgen, Duncan, and Cook 2015] Georgen, C.; Duncan, S.;
and Cook, L. 2015. From lurking to participatory spectatorship: Understanding the affordances of the dota 2 noob
stream. In Proceedings of Computer-Supported Collaborative Learning.
[Kwak and Blackburn 2015] Kwak, H., and Blackburn, J.
2015. Social Informatics: SocInfo 2014 International Workshops. Cham: Springer International Publishing. chapter
Linguistic Analysis of Toxic Behavior in an Online Video
Game, 209–217.
[Pobiedina et al. 2013] Pobiedina, N.; Neidhardt, J.; del Carmen Calatrava Moreno, M.; Grad-Gyenge, L.; and Werthner,
H. 2013. On successful team formation: Statistical analysis
of a multiplayer online game. In CBI, 55–62. IEEE Computer
Society.
[Ritchie et al. 2007] Ritchie, G.; Manurung, R.; Pain, H.;
Waller, A.; Black, R.; and O’Mara, D. 2007. A practical application of computational humour. In Proceedings of the 4th
International Joint Workshop on Computational Creativity,
91–98. London, UK: Goldsmiths, University of London.
[Saunders 2002] Saunders, R. 2002. Curious design agents
and artificial creativity. phd, University of Sydney.
[Silberschatz and Tuzhilin 1995] Silberschatz, A., and
Tuzhilin, A. 1995. On subjective measures of interestingness
in knowledge discovery. In KDD, volume 95, 275–281.
[Tremblay, Torres, and Verbrugge 2014] Tremblay, J.; Torres,
P. A.; and Verbrugge, C. 2014. Measuring risk in stealth
games. In FDG’14: Proceedings of the 9th International
Conference on Foundations of Digital Games.
[Veale and Alnajjar 2015] Veale, T., and Alnajjar, K. 2015.
Unweaving the lexical rainbow: Grounding linguistic creativity in perceptual semantics. In Proceedings of the Sixth International Conference on Computational Creativity (ICCC
2015), 63–70. Park City, Utah: Brigham Young University.
[Yang, Harrison, and Roberts 2014] Yang, P.; Harrison, B.;
and Roberts, D. L. 2014. Identifying patterns in combat
that are predictive of success in moba games. In Proceedings
of the Foundations of Digital Games Conference.
[Yannakakis, Martı́nez, and Jhala 2010] Yannakakis, G. N.;
Martı́nez, H. P.; and Jhala, A. 2010. Towards affective
camera control in games. User Modeling and User-Adapted
Interaction 20(4):313–340.
[Zook, Fruchter, and Riedl 2014] Zook, A.; Fruchter, E.; and
Riedl, M. O. 2014. Automatic playtesting for game parameter tuning via active learning. In Proceedings of the
9th International Conference on the Foundations of Digital
Games.
| 2cs.AI
|
The Distance Standard Deviation
arXiv:1705.05777v1 [math.ST] 16 May 2017
Dominic Edelmann,∗ Donald Richards,† and Daniel Vogel‡
May 17, 2017
Abstract
The distance standard deviation, which arises in distance correlation analysis of multivariate data, is studied as a measure of spread. New representations
for the distance standard deviation are obtained in terms of Gini’s mean difference and in terms of the moments of spacings of order statistics. Inequalities for
the distance variance are derived, proving that the distance standard deviation is
bounded above by the classical standard deviation and by Gini’s mean difference.
Further, it is shown that the distance standard deviation satisfies the axiomatic
properties of a measure of spread. Explicit closed-form expressions for the distance variance are obtained for a broad class of parametric distributions. The
asymptotic distribution of the sample distance variance is derived.
Key words and phrases. characteristic function; distance correlation coefficient; distance
variance; Gini’s mean difference; measure of spread; dispersive ordering; stochastic
ordering; U-statistic; order statistic; sample spacing; asymptotic efficiency.
2010 Mathematics Subject Classification. Primary: 60E15, 62H20; Secondary: 60E05,
60E10.
1
Introduction
In recent years, the topic of distance correlation has been prominent in statistical analyses of dependence between multivariate data sets. The concept of distance correlation
was defined in the one-dimensional setting by Feuerverger [7] and subsequently in the
multivariate case by Székely, et al. [25, 26], and those authors applied distance correlation methods to testing independence between random variables and vectors.
∗
German Cancer Research Center, Im Neuenheimer Feld 280, 69120 Heidelberg, Germany.
Department of Statistics, Pennsylvania State University, University Park, PA 16802, U.S.A.
‡
Institute for Complex Systems and Mathematical Biology, University of Aberdeen, Aberdeen AB24
3UE, U.K.
∗
Corresponding author; e-mail address: [email protected]
†
1
2
Edelmann, Richards, and Vogel
Since the appearance of [25, 26], enormous interest in the theory and applications
of distance correlation has arisen. We refer to the articles [22, 27, 28] on statistical
inference; [8, 9, 14, 33] on time series; [4, 5, 6] on affinely invariant distance correlation
and connections with singular integrals; [19] on metric spaces; and [23] on machine
learning. Distance correlation methods have also been applied to assessing familial
relationships [17], and to detecting associations in large astrophysical databases [20, 21].
For z ∈ C, denote by |z| the modulus of z. For any positive integer p and s, x ∈ Rp ,
we denote by hs, xi the standard Euclidean inner product on Rp and by ksk = hs, si1/2
the standard Euclidean norm. Further, we define the constant
cp =
π (p+1)/2
.
Γ (p + 1)/2
For jointly distributed random vectors X ∈ Rp and Y ∈ Rq , let
√
fX,Y (s, t) = E exp −1(hs, Xi + ht, Y i) ,
s ∈ Rp , t ∈ Rq , be the joint characteristic function of (X, Y ) and let fX (s) = fX,Y (s, 0)
and fY (t) = fX,Y (0, t) be the corresponding marginal characteristic functions. The
distance covariance between X and Y is defined as the nonnegative square root of
Z
1
ds dt
2
2
V (X, Y ) =
fX,Y (s, t) − fX (s)fY (t)
;
(1.1)
p+1
cp cq Rp+q
ksk ktkq+1
the distance variance is defined as
Z
ds dt
1
2
2
2
;
fX (s + t) − fX (s)fX (t)
V (X) := V (X, X) = 2
p+1
cp R2p
ksk ktkp+1
(1.2)
and we define the distance standard deviation V(X) as the nonnegative square root of
V 2 (X). The distance correlation coefficient is defined as
V(X, Y )
R(X, Y ) = p
V(X)V(Y )
(1.3)
as long as V(X), V(Y ) 6= 0, and R(X, Y ) is defined to be zero otherwise.
The distance correlation coefficient, unlike the Pearson correlation coefficient, characterizes independence: R(X, Y ) = 0 if and only if X and Y are mutually independent.
Moreover, 0 ≤ R(X, Y ) ≤ 1; and for one-dimensional random variables X, Y ∈ R,
R(X, Y ) = 1 if and only if Y is a linear function of X. The empirical distance correlation possesses a remarkably simple expression ([7], [25, Theorem 1]), and efficient
algorithms for computing it are now available [13].
The objective of this paper is to study the distance standard deviation V(X). Since
distance standard deviation terms appear in the denominator of the distance correlation
coefficient (1.3) then properties of V(X) are crucial to understanding fully the nature
3
The Distance Standard Deviation
of R(X, Y ). Now that R(X, Y ) has been shown to be superior in some instances to
classical measures of correlation or dependence, there arises the issue of whether V(X)
constitutes a measure of spread suitable for situations in which the classical standard
deviation cannot be applied.
As V(X) is possibly a measure of spread, we should compare it to other such measures. Indeed, suppose that E(kXk2 ) < ∞, and let X, X 0 , and X 00 be independent and
identically distributed (i.i.d.); then, by [25, Remark 3],
V 2 (X) = E(kX − X 0 k2 ) + (EkX − X 0 k)2 − 2E(kX − X 0 k · kX − X 00 k),
(1.4)
The second term on the right-hand side of (1.4) is reminiscent of the Gini mean difference [10, 31], which is defined for real-valued random variables Y as
∆(Y ) := E|Y − Y 0 |,
(1.5)
where Y and Y 0 are i.i.d. Furthermore, if X ∈ R then one-half the first summand in
(1.4) equals σ 2 (X), the variance of X:
1
1
E(|X − X 0 |2 ) = E(X 2 − 2XX 0 + X 02 ) = E(X 2 ) − E(X)E(X 0 ) ≡ σ 2 (X).
2
2
Let X and Y be real-valued random variables with cumulative distribution functions
F and G, respectively. Further, let F −1 and G−1 be the right-continuous inverses of F
and G, respectively. Following [24, Definition 2.B.1], we say that X is smaller than Y
in the dispersive ordering, denoted by X ≤ disp Y , if for all 0 < α ≤ β < 1,
F −1 (β) − F −1 (α) ≤ G−1 (β) − G−1 (α).
(1.6)
According to [2], a measure of spread is a functional τ (X) satisfying the axioms:
(C1) τ (X) ≥ 0,
(C2) τ (a + bX) = |b| τ (X) for all a, b ∈ R, and
(C3) τ (X) ≤ τ (Y ) if X ≤ disp Y .
The distance standard deviation V(X) obviously satisfies (C1). Moreover, Székely, et
al. [25, Theorem 4] prove that:
1. If V(X) = 0 then X = E[X], amost surely,
2. V(a + bX) = |b| V(X) for all a, b ∈ R, and
3. V(X + Y ) ≤ V(X) + V(Y ) if X and Y are independent.
In particular, V(X) satisfies the dilation property (C2). In Section 5, we will show that
V(X) satisfies condition (C3), proving that V(X) is a measure of spread in the sense
4
Edelmann, Richards, and Vogel
of [2]. However, we will also derive some stark differences between V(X), on the one
hand, and the standard deviation and Gini’s mean difference, on the other hand.
The paper is organized as follows. In Section 2, we derive inequalities between
the summands in the distance variance representation (1.4). For real-valued random
variables, we will prove that V(X) is bounded above by Gini’s mean difference and by
the classical standard deviation. In Section 3, we show that the representation (1.4)
can be simplified further, revealing relationships between V(X) and the moments of
spacings of order statistics. Section 4 provides closed-form expressions for the distance
variance for numerous parametric distributions. In Section 5, we show that V(X) is a
measure of spread in the sense of [2]; moreover, we point out some important differences
between V(X), the standard deviation, and Gini’s mean difference. Section 6 studies
the properties of the sample distance variance.
2
Inequalities between the distance variance, the
variance, and Gini’s mean difference
The integral representation in equation (1.2) of the distance variance V 2 (X) generally
is not suitable for practical purposes. Székely, et al. [25, 26] derived an alternative
representation; they show that if the random vector X ∈ Rp satisfies EkXk2 < ∞ and
if X, X 0 , and X 00 are i.i.d. then
V 2 (X) = T1 (X) + T2 (X) − 2 T3 (X),
(2.1)
where
T1 (X) = E(kX − X 0 k2 ),
T2 (X) = (EkX − X 0 k)2 ,
(2.2)
and
T3 (X) = E kX − X 0 k · kX − X 00 k ,
(2.3)
Corresponding to the representation (2.1), a sample version of V 2 (X) then is given by
Vn2 (X) = T1,n (X) + T2,n (X) − 2 T3,n (X),
where
n
n
1 XX
kXi − Xj k2 ,
T1,n (X) = 2
n i=1 j=1
n X
n
1 X
2
T2,n (X) =
kX
−
X
k
,
i
j
n2 i=1 j=1
(2.4)
(2.5)
5
The Distance Standard Deviation
and
n
n
n
1 XXX
T3,n (X) = 3
kXi − Xj k · kXi − Xk k.
n i=1 j=1 k=1
(2.6)
We remark that the version (2.4) is biased; indeed, throughout the paper, we work with
biased sample versions to avoid dealing with numerous complicated, but unessential,
constants in the ensuing results. In any case, an unbiased sample version can be defined
in a similar fashion; see, e.g., [27]).
In the following we will study inequalities between the summands showing up in
equations (2.1) and (2.4). In the one-dimensional case, these inequalities will lead to
crucial results concerning the relationships between the distance standard deviation,
Gini’s mean difference and the standard deviation.
Lemma 2.1. Let X = (X (1) , . . . , X (p) )t ∈ Rp be a random vector. Moreover let
X = (X1 , . . . , Xn ) denote a random sample from X and let T1 (X), T2 (X), T3 (X),
and T1,n (X), T2,n (X), T3,n (X) be defined as in equations (2.1)-(2.6). Then
T2,n (X) ≤ T3,n (X) ≤ T1,n (X),
T1,n (X) ≤ 2T3,n (X).
(2.7)
Further, if EkXk2 < ∞ then
T2 (X) ≤ T3 (X) ≤ T1 (X),
T1 (X) ≤ 2T3 (X).
(2.8)
Proof. First note that
n
n
n
1 XXX
kXi − Xj k · kXi − Xk k
T3,n (X) = 3
n i=1 j=1 k=1
n
n
2
1 XX
= 3
kXi − Xj k .
n i=1 j=1
P
P
By the Cauchy-Schwarz inequality, ( ni=1 ai )2 ≤ n ni=1 a2i for all a1 , . . . , an ∈ R; applying this inequality to the sums which define T1,n , T2,n and T3,n , we obtain
n
T2,n (X) =
≤
n
2
1 XX
kX
−
X
k
i
j
n4 i=1 j=1
n
n
2
n XX
kX
−
X
k
= T3,n (X)
i
j
n4 i=1 j=1
and
n
n
2
1 XX
T3,n (X) = 3
kXi − Xj k
n i=1 j=1
n
n
n XX
≤ 3
kXi − Xj k2 = T1,n (X).
n i=1 j=1
6
Edelmann, Richards, and Vogel
The second assertion in (2.7) follows by the triangle inequality:
n
n
1 XX
T1,n (X) = 2
kXi − Xj k2
n i=1 j=1
n
n
n
1 XXX
= 3
kXi − Xj k · kXi − Xk + Xk − Xj k
n i=1 j=1 k=1
n
n
n
1 XXX
kXi − Xj k kXi − Xk k + kXk − Xj k
≤ 3
n i=1 j=1 k=1
= 2 T3,n (X).
The corresponding inequalities (2.8) for the population measures follow from the
strong consistency of the respective sample measures. Alternatively they can be derived
by applying Jensen’s inequality and the triangle inequality, respectively.
Using the inequalities in Lemma 2.1, we can derive upper bounds for the distance
variance in terms of the variance of the components X (1) , . . . , X (p) and the Gini mean
difference of the vector X.
Theorem 2.2. Let X = (X (1) , . . . , X (p) )t ∈ Rp be a random vector with EkXk < ∞,
0
0
and let X 0 = (X (1) , . . . , X (p) )t denote an i.i.d. copy of X. Then
2
V (X) ≤
p
X
σ 2 (X (i) ),
i=1
and
V 2 (X) ≤ (EkX − X 0 k)2 .
Proof. To prove the first assertion, we note that
V 2 (X) = lim T1,n (X) + T2,n (X) − 2T3,n (X)
n→∞
≤ lim T2,n (X)
n→∞
= (EkX − X 0 k)2 ,
where the inequality follows by Lemma 2.1.
To extablish the second inequality we can assume, without loss of generality, that
7
The Distance Standard Deviation
EkXk2 < ∞. Then
T1 (X) = EkX − X 0 k2
p
X
=E
(X (i) − X 0(i) )2
i=1
p
h
i2
X
E (X (i) − EX (i) ) + (EX (i) − X 0(i) )
=
i=1
=2
p
X
σ 2 (X (i) ).
i=1
Applying Lemma 2.1 yields
V 2 (X) = T1 (X) + T2 (X) − 2T3 (X)
≤ T1 (X) − T3 (X)
≤ 21 T1 (X)
p
X
=
σ 2 (X (i) ).
i=1
The proof now is complete.
In the one-dimensional case, Theorem 2.2 implies that the distance variance is
bounded above by the variance and the squared Gini mean difference.
Corollary 2.3. Let X be a real-valued random variable with EkXk < ∞. Then,
V 2 (X) ≤ σ 2 (X),
V 2 (X) ≤ ∆2 (X).
Let us note further that for X ∈ R, the inequality T2 (X) ≤ T1 (X) can be sharpened.
Proposition 2.4. Let X be a real-valued random variable with E(|X|2 ) < ∞. Then,
T2 (X) ≤ 32 T1 (X).
Proof. By [31, p. 25],
1 ≥ [Cor(X, F (X))]2 =
Cov2 (X, F (X))
.
σ 2 (X) σ 2 (F (X))
(2.9)
By [30, equation (2.3)], Cov(X, F (X)) = ∆(X)/4; also, since F (X) is uniformly distributed on the interval [0, 1] then Var(F (X)) = 1/12. By the definition of the Gini
mean difference (1.5) and by (2.2), ∆2 (X) = T2 (X) and σ 2 (X) = T1 (X)/2. Therefore,
it follows from (2.9) that
12 ∆2 (X)
3 T2 (X)
1≥
=
,
2
16 σ (X)
2 T1 (X)
8
Edelmann, Richards, and Vogel
and the proof now is complete.
Interestingly, Gini’s mean difference and the distance standard deviation coincide
for distributions whose mass is concentrated on two points.
Theorem 2.5. Let X be Bernoulli distributed with parameter p. Then
V 2 (X) = ∆2 (X) = 4p2 (1 − p)2 .
Conversely, if X is a non-trivial random variable for which V 2 (X) = ∆2 (X) then the
distribution of X is concentrated on two points.
Proof. It is straightforward from (2.1) to verify that, for a Bernoulli distributed
random variable X, ∆(X) = 2 σ 2 (X) = 2 T3 (X) = 2 p(1 − p). Hence, by (2.1),
V 2 (X) = 2 σ 2 (X) + ∆2 (X) − 2 T3 (X) = 4 p2 (1 − p)2 .
Conversely, if X is a non-trivial random variable for which V 2 (X) = ∆2 (X) then
the conclusion that the distribution of X is concentrated on two points follows from
Theorem 3.1.
For the Bernoulli distribution with p = 12 , Theorem 2.5 implies immediately that
V 2 (X), σ 2 (X), and ∆2 (X) attain the same value, namely, 1/4. Hence, applying Corollary 2.3 and the dilation property V(aX) = |a|V(X) in (C2), we obtain
Corollary 2.6. Let X denote the set of all real-valued random variables and let c > 0.
Then
max{V 2 (X) : σ 2 (X) = c} = max{V 2 (X) : ∆2 (X) = c} = c,
X∈X
X∈X
and both maxima are attained by Z = 2 c1/2 Y , where Y is Bernoulli distributed with
parameter p = 21 .
This result answers a question raised by Gabor Székely (private communication,
November 23, 2015).
We remark, that the second implication of Theorem 2.2 as well as Theorem 2.5 also
follow directly from the result for the generalized distance variance in [19, Proposition
2.3]. However, the proof presented here provides a different and more elementary
approach to these findings.
3
New representations for the distance variance
The representation of V given in (2.1), although more applicable than the expression
given in equation (1.2), still has the drawback that it is undefined for random vectors
9
The Distance Standard Deviation
with infinite second moments. This problem can be circumvented by considering the
representation
V 2 (X) = ∆2 (X) + W (X),
(3.1)
where
h
i
W (X) = E kX − X 0 k · kX − X 0 k − 2 kX − X 00 k .
In the one-dimensional case, the representation (3.1) can be further simplified using
the concept of order statistics.
Theorem 3.1. Let X be a real-valued random variable with E|X| < ∞, and let X, X 0 ,
and X 00 be i.i.d. copies of X. If X1:3 ≤ X2:3 ≤ X3:3 are the order statistics of the triple
(X, X 0 , X 00 ) then
V 2 (X) = ∆2 (X) − 34 E[(X2:3 − X1:3 ) (X3:3 − X2:3 )]
= ∆2 (X) − 8 E[(X − X 0 )+ (X 00 − X)+ ],
(3.2)
(3.3)
where t+ = max(t, 0), t ∈ R.
Proof. We first prove the theorem for the case in which X is continuous. In this case,
we apply the Law of Total Expectation and use the independence of the ranks and the
order statistics [29, Lemma 13.1] to obtain
W (X)
h
i
0
0
00
= E |X − X | |X − X | − 2 |X − X |
=
3
X
h
i
E |X − X 0 | |X − X 0 | − 2|X − X 00 | (rX , rX 0 , rX 00 ) = (k, k 0 , k 00 )
k,k0 ,k00 =1
k,k0 ,k00 are pairwise distinct
× P (rX , rX 0 , rX 00 ) = (k, k 0 , k 00 ) .
10
Edelmann, Richards, and Vogel
Using the symmetry of X, X 0 , and X 00 , it follows that
1
W (X) =
6
=
1
6
3
X
h
i
E |Xk:3 − Xk0 :3 | |Xk:3 − Xk0 :3 | − 2 |Xk:3 − Xk00 :3 |
k,k0 ,k00 =1
k,k0 ,k00 are pairwise distinct
3
h
X
i
h
i
E |Xk:3 − Xk0 :3 |2 − E |Xk:3 − Xk0 :3 | · |Xk:3 − Xk0 :3 | .
k,k0 ,k00 =1
k,k0 ,k00 are pairwise distinct
Evaluating the first summand in the latter equation yields
1
6
3
X
E |Xk:3 − Xk0 :3 |2
k,k0 ,k00 =1
k,k0 ,k00 are pairwise distinct
=
1
E (X1:3 − X2:3 )2 + E (X1:3 − X3:3 )2 + E (X2:3 − X3:3 )2 .
3
Proceeding analogously with the second summand and simplifying the outcome, we
obtain
4
W (X) = − E (X2:3 − X1:3 ) (X3:3 − X2:3 ) .
3
This proves (3.2) in the continuous case.
For the case of general random variables, we now apply the method of quantile
transformations. Let U be uniformly distributed on the interval [0, 1] and let U , U 0 ,
and U 00 be i.i.d.. Further, let F denote the cumulative distribution function of X.
With F −1 (p) = inf{x : F (x) ≥ p} denoting the right-continuous inverse of F , we define
X̃ = F −1 (Ũ ), X̃ 0 = F −1 (Ũ 0 ), and X˜00 = F −1 (U˜00 ). By [29, Theorem 21.1], the random
variables X̃, X̃ 0 , and X˜00 are i.i.d. copies of X and
W (X)
h
i
= E |X̃ − X̃ 0 | · |X̃ − X̃ 0 | − 2 |X̃ − X˜00 |
=
3
X
h
i
E |X̃ − X̃ 0 | · |X̃ − X̃ 0 | − 2 |X̃ − X˜00 | (rU , rU 0 , rU 00 ) = (k, k 0 , k 00 )
k,k0 ,k00 =1
k,k0 ,k00 are pairwise distinct
× P (rU , rU 0 , rU 00 ) = (k, k 0 , k 00 )
11
The Distance Standard Deviation
1
=
6
3
X
h
i
E |Xk:3 − Xk0 :3 | · |Xk:3 − Xk0 :3 | − 2 |Xk:3 − Xk00 :3 |
k,k0 ,k00 =1
k,k0 ,k00 are pairwise distinct
4
= − E[(X2:3 − X1:3 ) (X3:3 − X2:3 )].
3
The second representation for W (X), (3.3), now follows by a combinatorial symmetry
argument from the first representation.
In the continuous case with finite second moment, equation (3.3) is equivalent to
E(|X − X 0 | · |X 00 − X 0 |) = σ 2 (X) + 4J(X),
where
Z
∞
Z
x
Z
(3.4)
∞
(x − y) (z − x)f (z) f (y)f (x)dz dy dx.
J(X) =
x=−∞
y=−∞
z=x
Formula (3.4) is essentially the key result in the classical paper by Lomnicki [18], who
also gave a simple expression for the variance of the empirical Gini mean difference,
X
2
b n (X) =
∆
|Xi − Xj |.
(3.5)
n (n − 1) 1≤i<j≤n
Indeed, it is shown in [18] that
1
b n (X) =
4 (n − 1) σ 2 (X) + 16 (n − 2)J(X) − 2 (2n − 3)∆2 (X) . (3.6)
Var ∆
n (n − 1)
We note two consequences of Theorem 3.1 and equation (3.6). First, Theorem 3.1
implies that the decomposition (3.6) holds in an analogous way for the non-continuous
case. Second, for distributions with finite second moment, calculating the distance
b n and vice versa. These considerations imply that the
variance yields the variance of ∆
b
b
asymptotic variance ASV (∆(X))
= limn→∞ n Var( ∆(X)
) can be expressed alternatively as
b
ASV (∆(X))
= 4 σ 2 (X) − 2 V 2 (X) − 2 ∆2 (X).
(3.7)
For a random sample X1 , . . . , Xn of real-valued random variables, the difference
between successive order statistics, Di:n := Xi+1:n − Xi:n , i = 1, . . . , n − 1, is called
the ith spacing of X = (X1 , . . . , Xn ). Jones and Balakrishnan [15] (see also [30, 31])
studied closed-form expression for the moments of spacings and showed that
ZZ
2
2
σ (X) = E(D1:2 ) = 2
F (x)(1 − F (y))dx dy
(3.8)
−∞<x<y<∞
and
Z
∞
F (x)(1 − F (x))dx.
∆(X) = E(D1:2 ) = 2
(3.9)
−∞
By applying results in [15], we obtain an analogous representation for the distance
variance.
12
Edelmann, Richards, and Vogel
Theorem 3.2. Let X be a real-valued variable with E(|X|) < ∞ and let X, X 0 , X 00 ,
and X 000 be i.i.d. Then,
ZZ
2
V (X) = 8
F 2 (x)(1 − F (y))2 dx dy
(3.10)
−∞<x<y<∞
where X1:4 ≤ X2:4 ≤ X3:4
2
= E[(X3:4 − X2:4 )2 ],
(3.11)
3
≤ X4:4 denote the order statistics of (X, X 0 , X 00 , X 000 ).
Proof. By equation (3.9), we obtain
h Z ∞
i2
2
∆ (X) = 2
F (x) (1 − F (x))dx
−∞
Z ∞Z ∞
=4
F (x) [1 − F (x)] F (y) [1 − F (y)] dx dy
−∞ −∞
ZZ
F (x) [1 − F (x)] F (y) [1 − F (y)] dx dy.
=8
−∞<x<y<∞
Moreover, by [15, equation (3.5)]
E[(X2:3 − X1:3 ) (X3:3 − X2:3 )]
ZZ
F (x) [F (y) − F (x)] [1 − F (y)] dx dy.
=8
−∞<x<y<∞
Hence,
V 2 (X) = ∆2 (X) − 34 E[(X(2) − X(1) ) (X(3) − X(2) )]
ZZ
[F (x)]2 [1 − F (y)]2 dx dy,
=8
−∞<x<y<∞
which proves (3.10).
Finally, the formula (3.11) follows from (3.10) and from [15, equation (3.4)].
Theorem 3.2 now yields for the distance variance a new sample version which is
distinct from Vn2 (X), as follows.
Corollary 3.3. Let X be a real-valued variable with E(|X|) < ∞ and let X =
(X1 , . . . , Xn ) be a random sample from X. Then, a strongly consistent sample version for V 2 (X) is
−2 X
n−1
2
2
n
2
Un (X) =
min(i, j) n − max(i, j) Di:n Dj:n ,
(3.12)
2
i,j=1
where Dk:n = Xk+1:n − Xk:n denotes the kth sample spacing of X, 1 ≤ k ≤ n − 1.
The Distance Standard Deviation
13
Proof. Let h : R4 7→ R be the symmetric kernel defined by
2
h(X1 , . . . , X4 ) = (X3:4 − X2:4 )2 ,
3
where X1:4 ≤ X2:4 ≤ X3:4 ≤ X4:4 are the order statistics of X1 , . . . , X4 . By Theorem
3.2, we have E[h(X1 , . . . , X4 )] < ∞. Hence, by Hoeffding [12],
Ubn2 (X)
−1
2 n
=
3 4
1≤i
X
h(Xi1 , . . . , Xi4 )
1 <i2 <i3 <i4 ≤n
is a strongly consistent estimator for V 2 (X). Using a straightforward combinatorial
calculation, we obtain
−1 X
2
n
2
Ubn (X) =
(i − 1) (n − j)(Xj:n − Xi:n )2 .
3 4
1≤i<j≤n
On inserting the definition of the spacings, the latter equation reduces to
−1 X
n
2
2
(i − 1) (n − j) (Di:n + · · · + Dj−1:n )2
Ubn (X) =
3 4
1≤i<j≤n
−1 X
j−1
X
2 n
(i − 1) (n − j)
Dk:n Dl:n .
≡
3 4
1≤i<j≤n
k,l=i
Interchanging the above summations, we obtain
−1 X
min(k,l)
n−1
X
2 n
2
b
Un (X) =
Dk:n Dl:n
3 4
i=1
k,l=1
n
X
(i − 1) (n − j)
j=max(k,l)+1
−1 X
n−1
1 n
Dk:n Dl:n min(k, l) min(k, l) − 1
=
6 4
k,l=1
× n − max(k, l) n − max(k, l) − 1 ,
where the latter equality follows from the fact that
Pk
i=1
i = k(k − 1)/2. Since
−1
1 n
4
=
,
6 4
n (n − 1) (n − 2) (n − 3)
then we deduce that Un2 (X) = Ubn2 (X) + o(1). This completes the proof.
Denoting the vector of spacings by D = (D1:n , . . . , Dn−1:n ), we can write the
quadratic form in (3.12) as
Un2 (X) = Dt V D,
14
Edelmann, Richards, and Vogel
Figure 3.1: Illustration of (from left to right) the sample distance variance
b 2 , and the sample variance
Un2 , the squared sample Gini mean difference ∆
σ
b2 via their respective quadratic form matrices V , G, and S for sample size
n = 1, 000. The coordinate (i, j) corresponds to the (i, j)th entry of the
corresponding matrix, and the size of the corresponding matrix element is
specified via color code (see legend).
where the (i, j)th element of the matrix V is
−2
2
2
n
Vi,j =
min(i, j) n − max(i, j)
2
Both the squared sample Gini mean difference and the sample variance
n
σ
bn2 (X)
X
1
(Xi − X n )2
:=
n (n − 1) i=1
can also be expressed as quadratic forms in the spacings vector D; specifically,
b 2n (X) = Dt G D,
∆
σ
bn2 (X) = Dt S D,
where the elements of G and S are given by
−2
n
Gi,j =
i j (n − i) (n − j)
2
and
Si,j
−1
1 n
min(i, j) (n − max(i, j)).
=
2 2
Hence, comparing Un2 , ∆2n , and σn2 is equivalent to comparing the matrices V , G
and S. We use this fact to graphically illustrate differing features of V, ∆, and σ by
plotting the values of the underlying matrices; see Figure 3.1.
Moreover, these quadratic form representations lead to the rediscovery of results
from Section 2. For example, since V and G have the same diagonal entries then it
The Distance Standard Deviation
15
follows that V and ∆ are equal for Bernoulli-distributed random variables. Also, if n
is even then the elements Vn/2,n/2 , Gn/2,n/2 , and Sn/2,n/2 all coincide, representing the
fact that the underlying measures coincide for the Bernoulli distribution with p = 12 .
Finally, since Vij ≤ Gij and Vij ≤ Sij for all i, j then we obtain an alternative proof of
Corollary 2.3.
It is also remarkable that V is twice the second Hadamard power of S and that V
and S both are positive definite, while G is positive semidefinite with rank 1. Finally,
we mention that there are numerous other statistics which can be written as quadratic
forms or square-roots of quadratic forms in the spacings, e.g., the Greenwood statistic,
the range, and the interquartile range.
4
Closed form expressions for the distance variance
of some well-known distributions
Exploiting the different representations of the distance variance derived in the preceding
sections, we can now state the distance variance of many well-known distributions. In
the following result, we use the standard notation 1 F1 and 2 F1 for the classical confluent
and Gaussian hypergeometric functions.
Theorem 4.1. 1. Let X be Bernoulli distributed with parameter p. Then V 2 (X) =
4 p2 (1 − p)2 .
2. Let X be normally distributed with mean µ and variance σ 2 . Then
2
V (X) = 4
1 − √3
π
1 2
+
σ .
3
3. Let X be uniformly distributed on the interval [a, b]. Then V 2 (X) = 2(b − a)2 /45.
4. Let X be Laplace-distributed with density function, fX (x) = (2α)−1
exp(−|x − µ|/α), x ∈ R, α > 0, µ ∈ R. Then V 2 (X) = 7α2 /12.
5. Let X be Pareto-distributed with parameters α > 1 and xm > 0, and density function
fX (x) = αxαm x−(α+1) , x ≥ xm . Then,
4α2 x2m
V (X) =
(α − 1) (2α − 1)2 (3α − 2)
2
6. Let X be exponentially distributed with parameter λ > 0 and density function
fX (x) = λ exp(−λ x), x ≥ 0. Then, V 2 (X) = (3λ2 )−1 .
16
Edelmann, Richards, and Vogel
7. Let X be Gamma-distributed with shape parameter α > 0 and scale parameter 1.
Then
∞
X
V 2 (X) = 22(2−2 α)
Aj,k (α)2 ,
j,k=1
where
1/2
(α)j (α)k
Aj,k (α) = 2
j! k!
Γ(2α + j + k − 1)
×
2 F1 (−j − k + 2, 1 − α − j; 2 − 2α − j − k; 2) .
Γ(α + j) Γ(α + k)
−j−k
8. Let X be Poisson-distributed with parameter λ > 0. Then
∞
X
4j+k−1 j+k 2
λ Ajk ,
V (X) =
j!
k!
j,k=1
2
where
Ajk
1
=
(j − 1)!
b(j−k)/2c
j−k
(−1)l ( 21 )l ( 21 )j−l−1 1 F1 (j − l − 12 ; j; −4λ).
2l
X
l=0
9. Let X be negative binomially distributed with parameters c and β. Then
∞
X
(β)j (β)k
V (X) = (1 − c)
(1 + c2 )−2β−2j 22k cj+k A2jk ,
j!
k!
j,k=1
2
4β
where
Ajk
l
j−k
∞
X
X
j−k
j−k
2c
(β + j) − l
l1
l2
=
(−c) (−1) (|l1 − l2 |)!
l
l
l!
1 + c2
1
2
l ,l =0
l=0
1 2
|l1 −l2 |
×
X
m=0
(−2)m
2k+m−1 ( 12 )k+m−1
(m)|l1 −l2 |
(|l1 − l2 | − m)! (2m)! (k + m − 1)!
× 2 F1 (−l, k + m − 21 ; k + m; 2).
10. Let X = (X1 , . . . , Xp ) be a multivariate normally distributed random vector with
mean µ = (µ1 , . . . , µp ) and identity covariance matrix Ip := diag(1, . . . , 1). Then
c2p−1
V (X) = 4π 2
cp
2
"
#
Γ( 12 p) Γ( 21 p + 1)
1
1 1
1
1
2 − 2 2 F1 − 2 , − 2 ; 2 p; 4 + 1 .
Γ 2 (p + 1)
The Distance Standard Deviation
17
Proof. 1. See Theorem 2.5.
2. See the proof of Theorem 7 in [25] or [4, p. 14].
3. and 4. These follow directly from Theorem 3.1 and the results in Table 3 in [10].
5. and 6. These results follow directly from the representation (2.1) and [32, equations (4.2) and (4.4)].
7., 8., and 9. See [6, Propositions 5.6, 5.7, and 5.8].
10. See [4, Corollary 3.3].
By equations (3.6) and (3.7), we can also derive expresssions for the variance and
asymptotic variance of the sample Gini mean difference for the distributions 1.- 9. in
Theorem 4.1. To the best of our knowledge, these expressions are novel for the Gamma,
Poisson, and negative binomial distributions.
5
The distance standard deviation as a measure of
spread
In this section, we show that the distance standard deviation V(X) satisfies the criteria
(C1)-(C3) stated in Section 1 and therefore is an axiomatic measure of spread in the
sense of [2]. Moreover, we point out some differences and commonalities between V, ∆
and σ. First, we state some additional preliminaries about stochastic orders.
Definition 5.1 ([24], Section 1.A.1). A random variable X is said to be stochastically
smaller than a random variable Y , or X is smaller than Y in the stochastic ordering,
written X ≤st Y , if P(X > u) ≤ P(Y > u) for all u ∈ R.
Proposition 5.2 ([24], Section 1.A.1). A necessary and sufficient condition that X ≤st
Y is that
E[φ(X)] ≤ E[φ(Y )]
(5.1)
for all increasing functions φ for which these expectations exist.
Another important ordering of random variables is the dispersive order, ≤ disp , which
was stated earlier at (1.6) in the introduction. Bartoszewicz [1] proved the following
result.
Proposition 5.3 ([1], Proposition 3). Let (X1 , . . . , Xn ) and (Y1 , . . . , Yn ) be random
samples from the random variables X and Y , respectively, and let Dj = Xj+1:n − Xj:n
and Ej = Yj+1:n − Yj:n , j = 1, . . . , n − 1 denote the corresponding sample spacings. If
X ≤ disp Y then Dj:n ≤ st Ej:n for all j = 1, . . . , n − 1.
Applying this result to the representation of the distance variance derived in Theorem 3.2, we conclude that the distance standard deviation V is indeed a measure of
spread in the sense of [2].
18
Edelmann, Richards, and Vogel
Theorem 5.4. If X ≤ disp Y then V(X) ≤ V(Y ).
Proof. Let us consider i.i.d. replicates (X, Y ), (X 0 , Y 0 ), (X 00 , Y 00 ), and (X 000 , Y 000 ).
Moreover, let X1:4 ≤ X2:4 ≤ X3:4 ≤ X4:4 and Y1:4 ≤ Y2:4 ≤ Y3:4 ≤ Y4:4 denote the
respective order statistics. By Proposition 5.3,
(X3:4 − X2:4 ) ≤ st (Y3:4 − Y2:4 ).
Applying equation (5.1) and Theorem 3.2 concludes the proof.
Using similar arguments, we can show that the result of Theorem 5.4 holds analogously for the standard deviation and Gini’s mean difference; see also [16].
Theorem 5.5 ([24], Theorem 3.B.7). The random variable X satisfies the property
X ≤ disp X + Y for any random variable Y which is independent of X
if and only if X has a log-concave density.
Applying Theorem 5.5, we obtain the following corollary of Theorem 5.4.
Corollary 5.6. Let X be a random variable with a log-concave density. Then
V(X + Y ) ≥ V(X)
for any random variable Y independent of X.
In particular if X and Y are independently distributed, continuous, random variables with log-concave densities, then
V(X + Y ) ≥ max(V(X), V(Y )).
(5.2)
It is well known, both for the standard deviation and for Gini’s mean difference,
that analogous assertions hold without any restrictions on the distributions of X and
Y . In particular, for any pair of independent random variables X and Y with existing
first or second moments, respectively, there holds
σ 2 (X + Y ) = σ 2 (X) + σ 2 (Y ) ≥ max(σ 2 (X), σ 2 (Y )).
(5.3)
Also, letting X 0 and Y 0 denote i.i.d. copies of X and Y , respectively, we have
∆(X + Y ) = E[max(|X − X 0 |, |Y − Y 0 |)] ≥ max(∆(X), ∆(Y )).
(5.4)
However we now show that this property does not hold generally for the distance
standard deviation, V, thereby answering a second question raised by Gabor Székely
(private communication, November 23, 2015).
19
The Distance Standard Deviation
Example 5.7. Let X be Bernoulli distributed with parameter p = 21 and let Y be
uniformly distributed on the interval [0, 1] and independent of X. Then V(X) > V(X +
Y ).
Proof. By a straightforward calculation using (2.1), we obtain
V 2 (X + Y ) = T1 (X + Y ) + T2 (X + Y ) − 2 T3 (X + Y )
2 4 14
8
= + −
= .
3 9 15
45
However, by Theorem 2.5, V 2 (X) = 1/4 > V 2 (X + Y ).
Other common properties of the classical standard deviation and Gini’s mean difference concerns differences and sums of independent random variables. From the
representations of σ 2 (X + Y ) and ∆(X + Y ) given in (5.3) and (5.4), we see that
∆(X + Y ) = ∆(X − Y ),
σ(X + Y ) = σ(X − Y )
for any independent random variables X and Y for which these expressions exist.
On the other hand, these properties do not hold in general for the distance standard
deviation.
Example 5.8. Let X and Y be independently Bernoulli distributed with parameter
p 6= 21 . Then V(X + Y ) > V(X − Y ).
Proof. By a straightforward calculation using (2.1), we obtain
V 2 (X + Y ) = 8 (p − p2 )2 2 (p − p2 )2 − 6 (p − p2 ) + 2
and
V 2 (X − Y ) = 8 (p − p2 )2 2 (p − p2 )2 − 2 (p − p2 ) + 1 .
Hence,
V 2 (X + Y ) − V 2 (X − Y ) = 8 (p − p2 )2 (1 − 2p)2 ,
and this difference obviously is positive for p 6= 21 .
However, an analogous property holds when either of the two variables has a symmetric distribution.
Theorem 5.9. Let X and Y be independent real random variables with E|X +Y | < ∞.
Then V 2 (X + Y ) = V 2 (X − Y ) if either X or Y is symmetric about µ.
Proof. Since V(X − Y ) = V(Y − X) then we can assume, without loss of generality,
that Y is symmetric. Moreover since V 2 (X + µ) = V 2 (X) then we can assume that the
point of symmetry is at 0.
20
Edelmann, Richards, and Vogel
By equation (1.2),
V 2 (X − Y )
Z
1
2 dsdt
fX−Y (s + t) − fX (s)f−Y (t)
= 2
π R2
|s|2 |t|2
Z
1
2 dsdt
= 2
fX (s + t) f−Y (s + t) − fX (s)f−Y (s)fX (t)f−Y (t)
π R2
|s|2 |t|2
Z
1
2 dsdt
= 2
fX (s + t) fY (s + t) − fX (s)fY (s)fX (t)fY (t)
π R2
|s|2 |t|2
= V 2 (X + Y ),
where the third equality follows from the fact that Y and −Y have the same distribution.
6
The distance standard deviation as an estimator
In this section, we investigate the properties of the sample distance variance Vn2 (X) and
the sample distance standard deviation Vn (X) as estimators and derive their asymptotic distributions. For these purposes, we employ the representation (3.1), viz.,
V 2 (X) = W (X) + ∆2 (X)
where
W (X) = E [kX − Y k (kX − Y k − 2 kX − Zk)] ,
and ∆(X) denotes the population value of Gini’s mean difference. Throughout this
section, X, Y , Z are i.i.d. p-variate random vectors with distribution F . For a sample
of i.i.d. random vectors X = (X1 , . . . , Xn )t , each with distribution F , we define the
corresponding empirical quantities,
n
n
1 XX
∆n (X) = 2
kXi − Xj k
n i=1 j=1
and
n
n
n
1 XXX
kXi − Xj k (kXi − Xj k − 2kXi − Xk k) .
Wn (X) = 3
n i=1 j=1 k=1
Note that
Wn (X) = T1,n (X) − 2 T3,n (X),
cf. (2.4), and
Vn2 (X) = Wn (X) + ∆2n (X).
The Distance Standard Deviation
21
Further, it is straightforward to verify that
EWn (X) =
(n − 1)(n − 2)
W (X).
n2
(6.1)
b n (X), the unbiased
The statistic ∆n (X) does not wear a hat to distinguish it from ∆
version of the sample Gini difference, defined for univariate observations in (3.5) and
which we extend to multivariate observations by replacing the absolute value | · | by the
Euclidean norm k · k; thus,
∆2n (X) =
(n − 1)2 b 2
∆n (X).
n2
b n (X), we define
Similar to ∆
cn (X) =
W
≡
1
n(n − 1)(n − 2)
X
kXi − Xj k (kXi − Xj k − 2kXi − Xk k)
1≤i,j,k≤n
i6=j,j6=k,k6=i
n2
Wn (X).
(n − 1)(n − 2)
cn (X) is an unbiased sample version of W (X).
By (6.1), W
bn (X) = [W
cn (X) + ∆
b 2n (X)]1/2 is an alternative to Vn (X) as an empirical
Also, V
bn (X) is based on the
version of the distance standard deviation V(X). Although V
cn (X) and ∆
b n (X), the estimator V
bn (X) itself has a larger bias
unbiased estimators W
than Vn (X). The results of Table 2 below indicate that Vn (X) is to be preferred
bn (X) as an estimator of V(X) because it exhibits smaller finite-sample bias and
over V
smaller variance for scenarios considered in our simulations.
b 2 (X) have the same asymptotic distribution. In order
Nevertheless, Vn2 (X) and V
n
to establish that result, we define for x ∈ Rp ,
ψ1 (x) = Ekx − Y k2 ,
ψ3 (x) = E(kx − Y k · kY − Zk),
ψ2 (x) = E(kx − Y k · kx − Zk),
ψ4 (x) = Ekx − Y k,
and, with T1 (X), T2 (X), and T3 (X) as defined in (2.2)-(2.3), we also define
2
2
m11 = 4 E ψ1 (X) − ψ2 (X) − 2ψ3 (X) − 4 T1 (X) − 3T3 (X) ,
m12 = 4 E ψ4 (X) ψ1 (X) − ψ2 (X) − 2ψ3 (X) − 4T2 (X) T1 (X) − 3T3 (X) ,
2
m22 = 4 Eψ42 (X) − 4 T2 (X) .
(6.2)
and let
γ = m11 + 4m12 ∆(X) + 4m22 ∆2 (X).
We now provide in the following result the asymptotic distribution of Vn2 (X) and
bn2 (X).
V
22
Edelmann, Richards, and Vogel
Theorem 6.1. Suppose that E(kXk4 ) < ∞. Then, as n → ∞,
d
√
n Vn2 (X) − V 2 (X) −→ N (0, γ)
(6.3)
b 2 (X).
and the same result holds for V
n
bn (X) = W
cn (X), ∆
b n (X) t , which has exProof. Consider the bivariate statistic B
pected value B(X) = (W (X), ∆(X))t . Define the functions K, L : Rp × Rp × Rp → R
such that
K(x, y, z) = kx−yk(kx−yk−2kx−zk) + ky−zk(ky−zk−2ky−xk)
+ kz −xk(kz −xk−2kz −yk)
and
L(x, y, z) = (kx − yk + ky − zk + kz − xk),
bn (X) can be written as a U-statistic with
(x, y, z) ∈ Rp × Rp × Rp . Then the statistic B
the bivariate, permutation-symmetric kernel of order three, h : Rp × Rp × Rp → R2 ,
where
1 K(x, y, z)
h(x, y, z) =
,
3 L(x, y, z)
(x, y, z) ∈ Rp ×Rp ×Rp . Define the function h1 : Rp → R2 , where h1 (x) = Eh(x, Y, Z)−
B(X); then, h1 is the linear part in the Hoeffding decomposition of the kernel h, and
we calculate that
2 ψ1 (x) − ψ2 (x) − 2ψ3 (x) − T1 (X) + 3T3 (X)
h1 (x) =
,
ψ4 (x) − T2 (X)
3
x ∈ Rp . Since E(kXk)4 < ∞ then E[(h(X, Y, Z))2 ] < ∞; therefore, we deduce from a
classical result of Hoeffding [11, Theorem 7.1] that
d
√
bn (X) − B(X) −→
n B
N2 0, 9 Eh1 (X)h1 (X)t .
Denote the symmetric 2 × 2 matrix 9 Eh1 (X)h1 (X)t by M = (mij )i,j=1,2 , where the
elements m11 , m12 , and m22 are given in (6.2). Define g : R2 → R by g(x, y) = x + y 2 ;
b 2 (X) = g B
bn (X) . Since ∇h(x, y) = (1, 2y)t then, by applying the Delta
then V
n
d
√ b2
2
Method, we obtain n V
n (X) − V (X) −→ N (0, γ).
In the case of Vn2 (X), we need only to apply the formulas Wn (X) = (n − 1)(n −
cn (X)/n2 and ∆2 (X) = (n−1)2 ∆
b 2 (X)/n2 to deduce that V 2 (X)−V
b 2 (X) = o(n−1 ).
2)W
n
n
n
n
2
Then it follows by the Delta Method that Vn (X) has the same asymptotic distribution
as Vn2 (X), as given in (6.3).
The asymptotic distribution of the sample distance standard deviation Vn (X) now
follows from Theorem 6.1 by the Delta Method:
23
The Distance Standard Deviation
Distribution, F
N (0, 1)
L(0, 1)
t5
t3
ARE(Vn ; F )
0.784
0.952
0.992
0.965
ARE(b
σn ; F )
1
0.8
0.4
0
ARE(dbn ; F )
b n; F )
ARE(∆
0.876
1
0.941
0.681
0.978
0.964
0.859
0.524
Table 1: Asymptotic relative efficiencies with respect to the respective
maximum likelihood estimators of the distance standard deviation Vn ,
the standard deviation σ
bn , the mean devation dbn , and Gini’s mean
b n at the normal distribution, the Laplace distribution,
difference ∆
and the tν -distributions with ν = 5 and ν = 3.
Corollary 6.2. Under the conditions of Theorem 6.1, we have
√
d
n(Vn (X) − V(X)) −→ N 0, γ/4V 2 (X) ,
bn (X).
and the same result holds for V
In the following, we study the empirical distance standard deviation Vn (X) as an
√
estimator of spread in the univariate case. For any n-consistent and asymptotically normal estimator sn (X), we define its asymptotic variance ASV (sn (X); F ) at
√
the distribution F to be the variance of the limiting distribution of n(sn (X) − s),
as n → ∞, where sn (X) is evaluated at an i.i.d. sequence drawn from F and s de(1)
notes the corresponding population value of sn (X) at F . Any estimators sn (X) and
(2)
sn (X) which estimate possibly different population values s1 and s2 , respectively, at
a given distribution F , and which obey the dilation property (C2) in Section 1, can
be compared efficiency-wise by standardizing them through their respective population
(1)
(2)
values. We define the asymptotic relative efficiency of sn (X) with respect to sn (X)
at the population distribution F as
(1)
ARE
(2)
s(1)
n (X), sn (X); F
=
ASV (sn (X); F )/s21
(2)
ASV (sn (X); F )/s22
.
Even in the univariate case with normally distributed data, the integrals underlying
the parameters m11 , m12 , and m22 in (6.2) do not admit straightforward analytical
expressions. Nevertheless, by means of numerical integration, we can obtain values for
the asymptotic variance of Vn (X) for given population distributions and thus deduce
properties of the efficiency of Vn (X) in relation to other widely-used estimators of scale.
In Table 1, we provide the asymptotic relative efficiency of the distance standard deviation with respect to the respective maximum likelihood estimator at the
normal distribution, the Laplace distribution, and tν distributions with ν = 5 and
24
Edelmann, Richards, and Vogel
10
0.658
0.665
0.276
0.298
Sample size,
50
0.640
0.639
0.255
0.259
Distribution, F
n
500
0.634
0.634
0.255
0.255
∞
0.633
0.633
0.256
0.256
N (0, 1)
E(Vn )
bn )
E(V
nVar(Vn )
bn )
nVar(V
5
0.663
0.701
0.297
0.359
L(0, 1)
E(Vn )
bn )
E(V
nVar(Vn )
bn )
nVar(V
0.888
0.942
0.955
1.136
0.861
0.864
0.836
0.858
0.790
0.785
0.668
0.663
0.767
0.766
0.605
0.604
0.764
0.764
0.613
0.613
t5
E(Vn )
bn )
E(V
nVar(Vn )
bn )
nVar(V
0.818
0.866
0.761
0.931
0.799
0.804
0.632
0.655
0.744
0.741
0.474
0.471
0.727
0.727
0.432
0.432
0.725
0.725
0.424
0.424
t3
E(Vn )
bn )
E(V
1.003
1.074
5.762
8.420
0.960
0.967
2.001
2.177
0.861
0.855
1.089
1.067
0.817
0.816
0.777
0.774
0.810
0.810
0.680
0.680
nVar(Vn )
bn )
nVar(V
Table 2: Simulated finite-sample values of the mean and the variance of the distance standard deviation Vn (X) for n = 5, 10, 50, 500
bn (X) refers to the
compared to asymptotic values (last column); V
cn (X) and ∆
b n (X); 10 000
version based on the unbiased estimates W
repetitions.
ν = 3. The maximum likelihood estimator of scale in the location-scale family generated by the N (0, 1), or standard normal, distribution is the standard deviation. In
the Laplace model, the analogous estimator of scale is the mean deviation dbn (X) =
P
n−1 ni=1 |Xi − mn (X)|, where mn (X) denotes the sample median of X. In the case
of the tν -distribution, the maximum likelihood estimator of the scale parameter does
generally not admit an explicit representation.
In Table 1, the asymptotic efficiencies of the distance standard deviation are compared with those of the standard deviation σ
bn (X), the mean deviation dbn (X), and
b n (X). The asymptotic variance of the maximum likelihood
Gini’s mean difference ∆
estimator of the scale parameter for the tν -distribution is (ν + 3)/2ν. The population
values and asymptotic variances of the other estimators mentioned at the respective
distributions are given by Gerstenberger and Vogel [10, Tables 2 and 3].
The Distance Standard Deviation
25
While the distance standard deviation has moderate efficiency at normality, it turns
out to be asymptotically very efficient in the case of heavier-tailed populations. For the
t5 - and t3 -distributions, the distance standard deviation outperforms its three competitors considered here and moreover is very close to the respective maximum likelihood
estimator.
In Table 2, we complement our asymptotic analysis by finite-sample simulations. For
sample sizes n = 5, 10, 50, and 500 and the same population distributions as above, the
(simulated) expectations and variances (based on 10,000 observations) of the empirical
distance standard deviation Vn (X) = [Wn (X) + ∆2n (X)]1/2 and the alternative version
bn (X) = [W
cn (X) + ∆
b 2n (X)]1/2 are given along with their respective asymptotic values.
V
b n (X) are
The corresponding values for the competing estimators σ
bn (X), dbn (X), and ∆
also provided by Gerstenberger and Vogel [10]. Table 2 indicates that Vn (X) indeed is
bn (X) in terms of bias as well as variance.
to be preferred over V
References
[1] Bartoszewicz, J. (1986). Dispersive ordering and the total time on test transformation. Statist. Probab. Lett., 4, 285–288.
[2] Bickel, P. J., and Lehmann, E. L. (2012). Descriptive statistics for nonparametric
models, IV. Spread. In: Selected Works of E. L. Lehmann (J. Rojo, editor),
519–526. Springer, New York.
[3] David, H. A., and Nagaraja, H. N. (1981). Order Statistics. Wiley, New York.
[4] Dueck, J., Edelmann, D., Gneiting, T., and Richards, D. (2014). The affinely
invariant distance correlation. Bernoulli, 20, 2305–2330.
[5] Dueck, J., Edelmann, D., and Richards, D. (2015). A generalization of an integral
arising in the theory of distance correlation. Statist. Probab. Lett., 97, 116–119.
[6] Dueck, J., Edelmann, D., and Richards, D. (2017). Distance correlation coefficients
for Lancaster distributions, J. Multivariate Anal., 154, 19–39.
[7] Feuerverger, A. (1993). A consistent test for bivariate dependence. Int. Stat. Rev.,
61, 419–433.
[8] Fiedler, J. (2016). Distances, Gegenbauer expansions, curls, and dimples: On dependence measures for random fields. Doctoral dissertation, University of Heidelberg.
[9] Fokianos, K., and Pitsillou, M. (2016). Consistent testing for pairwise dependence
in time series. Technometrics, to appear.
26
Edelmann, Richards, and Vogel
[10] Gerstenberger, C., and Vogel, D. (2015). On the efficiency of Gini’s mean difference.
Stat. Methods. Appt., 24, 569–596.
[11] Hoeffding, W. (1948). A class of statistics with asymptotically normal distribution.
Ann. Math. Stat., 19, 293–325.
[12] Hoeffding, W. (1961). The strong law of large numbers for U -statistics. Institute of
Statistics, Mimeograph Series No. 302. University of North Carolina, Chapel Hill,
NC.
[13] Huo, X., and Székely, G. J. (2016). Fast computing for distance covariance. Technometrics, 58, 435–447.
[14] Jentsch, C., Leucht, A., Meyer, M., and Beering, C. (2016). Empirical characteristic functions-based estimation and distance correlation for locally stationary
processes. Preprint, Mannheim University.
[15] Jones, M. C., and Balakrishnan, N. (2002). How are moments and moments of
spacings related to distribution functions? J. Stat. Plan. Inference, 103, 377–390.
[16] Kochar, S. C. (1999) On stochastic orderings between distributions and their sample spacings. Statist. Probab. Lett., 42, 345–352.
[17] Kong, J., Klein, B. E. K., Klein, R., and Wahba, G. (2012). Using distance correlation and SS-ANOVA to access associations of familial relationships, lifestyle
factors, diseases, and mortality. Proc. Natl. Acad. Sci., 109, 20352–20357.
[18] Lomnicki, Z. A. (1952). The standard error of Gini’s mean difference. Ann. Math.
Stat., 23, 635-637.
[19] Lyons, R. (2013). Distance covariance in metric spaces. Ann. Probobab, 41, 3284–
3305.
[20] Martı́nez-Gómez, E., Richards, M. T., and Richards, D. St. P. (2014). Distance
correlation methods for discovering associations in large astrophysical databases.
Astrophys. J., 781, 39 (11 pp.).
[21] Richards, M. T., Richards, D. St. P., and Martı́nez-Gómez, E. (2014). Interpreting
the distance correlation results for the COMBO-17 survey. Ann. Appl. Stat., 784,
L34 (5 pp.).
[22] Rizzo, M. L., and Székely, G. J. (2010). DISCO analysis: A nonparametric extension of analysis of variance. Ann. Appl. Stat., 4, 1034–1055.
The Distance Standard Deviation
27
[23] Sejdinovic, D., Sriperumbudur, B., Gretton, A., and Fukumizu, K. (2013). Equivalence of distance-based and RKHS-based statistics in hypothesis testing. Ann.
Stat., 41, 2263–2291.
[24] Shaked, M., and Shanthikumar, G. (2007) Stochastic Orders. Springer, New York.
[25] Székely, G. J., Rizzo, M. L., and Bakirov, N. K. (2007). Measuring and testing
independence by correlation of distances. Ann. Stat., 35, 2769–2794.
[26] Székely, G. J., and Rizzo, M. (2009). Brownian distance covariance. Ann. Appl.
Stat., 3, 1236–1265.
[27] Székely, G. J., and Rizzo, M. (2013). The distance correlation t-test of independence in high dimension. J. Multivariate Anal., 117, 193–213.
[28] Székely, G. J., and Rizzo, M. (2014). Partial distance correlation with methods for
dissimilarities. Ann. Stat., 42, 2382–2412.
[29] Van der Vaart, A. W. (2000) Asymptotic Statistics, Volume 3. Cambridge University Press, New York.
[30] Yitzhaki, S. (2003). Gini’s mean difference: A superior measure of variability for
non-normal distributions. Metron, 61, 285-316.
[31] Yitzhaki, S., and Schechtman, E. (2012). The Gini Methodology: A Primer on a
Statistical Methodology. Springer, New York.
[32] Zenga, M., Polisicchio, M., and Greselin, F. (2004). The variance of Gini’s mean
difference and its estimators. Statistica, 64, 455-475.
[33] Zhou, Z. (2012). Measuring nonlinear dependence in time-series, a distance correlation approach. J. Time Series Anal., 33, 438–457.
| 10math.ST
|
A Review of Intelligent Practices for Irrigation
Prediction
Krupakar Hans
SSN College Of Engineering
Chennai, India
E-mail: [email protected]
Akshay Jayakumar
SSN College Of Engineering
Chennai, India
E-mail: [email protected]
Dhivya G
SSN College Of Engineering
Chennai, India
E-mail: [email protected]
ABSTRACT:
Population growth and increasing droughts are creating unprecedented strain on the continued availability of
water resources. Since irrigation is a major consumer of fresh water, wastage of resources in this sector could have
strong consequences. To address this issue, irrigation water management and prediction techniques need to be
employed effectively and should be able to account for the variabilities present in the environment. The different
techniques surveyed in this paper can be classified into two categories: computational and statistical.
Computational methods deal with scientific correlations between physical parameters whereas statistical methods
involve specific prediction algorithms that can be used to automate the process of irrigation water prediction.
These algorithms interpret semantic relationships between the various parameters of temperature, pressure,
evapotranspiration etc. and store them as numerical precomputed entities specific to the conditions and the area
used as the data for the training corpus used to train it. We focus on reviewing the computational methods used to
determine Evapotranspiration and its implications. We compare the efficiencies of different data mining and
machine learning methods implemented in this area, such as Logistic Regression, Decision Tress Classifier, SysFor,
Support Vector Machine(SVM), Fuzzy Logic techniques, Artifical Neural Networks(ANNs) and various hybrids of
Genetic Algorithms (GA) applied to irrigation prediction. We also recommend a possible technique for the same
based on its superior results in other such time series analysis tasks.
INTRODUCTION:
Water scarcity is becoming a major issue throughout the world today, which in turn has increased the
threat of a major food crisis. Therefore, there needs to be an efficient method to utilize the available resources
judiciously. As agriculture is one of the largest water consuming sectors, managing irrigation levels can play an
important part in saving water for other purposes. This is also due to the fact that 25% of the water taken for
agriculture is wasted due to poor management1,2. In order to maximize water productivity and improve water
management, application of statistical methods have become significant.
As far as India is concerned, agriculture is the largest contributor to the GDP. As of 2010-11, the
agricultural sector has contributed 14.2 percent of the GDP. Also, India's agricultural exports account for 1.4%
of the world trade in agriculture3. Therefore, there is an increasing need to develop a prediction system effective
management and utilization of soil nutrients and water resources. These systems based on spatial database on
agriculture can improve agricultural management in India4,5.
The earliest methods to address this issue involved computational models. In these methods, the focus
lay on correlation between physical factors alone. Since these models relied on empirical scientific equations,
they are only suitable in perfectly ideal conditions. They do not take into account the uncertainties and
variations from ideal conditions that were used in determination of their implications that are found in real time
data6–8.
Over the last 15 years, there have been several applications of various data mining and prediction
models in this area that have made coming up with an irrigation management and prediction system possible.
These techniques are powerful, explorative and can perform intensive analysis on the given data. Some of the
techniques used include Logistic Regression, Decision Tree Classifiers, SysFor, Support Vector Machines,
Fuzzy Logic Classifiers and Artifical Neural Networks(ANNs).
The models described for irrigation prediction system are generally developed and trained using a
large amount of historical data (training data) of entities or features that would influence the amount of water
that is required in exactitude or at least inside of a coherent range. Once built, the models are used on real time
data not used in training i.e. to make a prediction of the amount of irrigation necessary to sustain the crop’s
healthy growth. The nature of these models would be such that even though the algorithm allows for
generalization to randomized test data, the predictions are most correspondent specifically to the area and
conditions used in the training data, thereby trying to account for the otherwise invisible variations specific to
that land and surroundings1,9–11.
The objective of this paper is to study and review the computational various data mining techniques in
the field of hydrology and to compare the efficiencies of the predictions made in the same. Section 2 of this
paper looks at all the parameters or sources of features used in order to facilitate prediction of irrigation water
requirement of crops. Section 3 of this paper reviews the computational parameter evapotranspiration and its
implications and the different data mining techniques and the data analysis areas used in the field so far. Section
4 compares the accuracy of prediction of the different techniques discussed in Section 3. Section 5 describes a
novel RNN LSTM model that is proposed in this paper that can be used to obtain better results, a richer
contextualization and better semantic correlations found in the parameters used in real-time data12–17. Section 6
concludes the paper.
SOURCES INFLUENCING IRRIGATION DEMAND:
The sources of parameters that affect the process of crop irrigation usually tend to vary according to
the terrain that agriculture is practiced. While some generic ones apply to all kinds of lands, we are going to
restrict the discussion about the sources of feature vectors to terrains at ground level18. This list of features is by
no means exhaustive and there are more and more parameters that can be considered with the advent of the new
age Internet of Things sensors that can track various things that were previously not possible to track19,20. The
sources that make irrigation water demand prediction possible can be broadly categorized into meteorological
factors, crop input and agricultural parameters.
1. Meteorological Factors
Meteorological factors are important in determining the water requirement in crop irrigation because
there are areas of the world, like in India, where the primary source of water is rainfall. Combined with this,
there are also other factors that have the potential to affect the amount of water required for the efficient
management and yield of crops. One of the main factors in this respect is the temperature parameters: maximum
temperature, minimum temperature and average temperature. The effect on the amount of sunshine that a region
requires is undeniably crucial to the plant’s water needs as too much heat would cause lesser water availability
to crops in general21.
The speed of winds also tends to affect the amount of irrigation required, albeit not as much as
temperature does. It is generally a practice to use a metric like Crop Water Stress Index (CWSI) to determine
how much water is required for a sustained and healthy growth of plants. CWSI is the relative transpiration rate
occurring from a plant by means of the temperature of the plant and the measure of humidity of the air in terms
of vapour pressure deficit. It has been found that during the calculation of CWSI, predictions made during
conditions of low wind speed tend to be in excess of the actual values and vice versa22.
Other factors affecting crop irrigation requirement are rainfall, solar radiation intensity and duration,
precipitation etc. The effects of these factors on crop irrigation demand prediction are quite straightforward as
these factors affect the surface water availability and to an extent, the amount of groundwater present
directly23,24.
2. Crop Input Factors
One of the most important criteria that determines the irrigation demand is the soil type that is used to
grow the particular crops. Major types of soil used for various crops tend to have different effects as the water
retention capacity of the soil heavily depends on the composition of the soil. The difference can be seen clearly
when the water requirement of various crops that grow on various types of soil are compared together25. One
other factor considered include the type of crop itself. The behavior of water retention and requirement of every
crop varies as the water content of each and every part of the plant (shoots, flowers, leaves, fruits etc.) varies 26.
Another parameter that is important is the soil moisture stress, which is a metric used to determine the extent of
water that is not present in the soil27.
Also, there are several other crop data surface coefficients that change with time that influence the
water requirement of plants. But the use of these parameters is something that is left up to the programmer
because these parameters are found out using relationships of the other feature vectors used in the mining
algorithm. It is generally not advisable to do so as precomputed values might push the model to be prone to the
errors in these parameters that would occur6,28–32.
3. Agricultural Factors
Some other obvious factors include the time series parameters that determine the amount required in a
situation that is affected by the history of occurrence. One such parameter is the actual amount of water that is
used in the time steps that have occurred before the one being trained for. Some other information like amount
of time remaining till the harvest are also considered sometimes so as to enforce pressurized irrigation behavior
that the mining algorithm can learn from the best patterns possible as well33.
DESCRIPTION OF METHODS:
The methods described in the following sections include ones that have been used for irrigation water
prediction across all the parts of the world34. These methods have been used in specific parts of the world where
electronically collected data about the conditions and the parameters is occurring. In the techniques discussed
below, Evapotranspiration is the only one that is a computational method. All of the other methods discussed
fall into the category of data mining and prediction algorithms based on historical context. Some of the
comparative analyses specified in this paper comes from the work by Khan et al. from the comparative study he
performed on the various data mining algorithms in terms of their performance of prediction by using the
Coleambally Irrigation Area (CIA) in Australia35.
1. Evapotranspiration
One of the really popular computational methods used to semantically predict the amount of water
required is called EvapoTranspiration. EvapoTranspiration (ET) is a semantic measure of the total water loss
that occurs due to evaporation and transpiration which specifies the amount of water essential for growth of
plants. It is expressed as a depth (usually in inches). ET is based on a number of factors that can include: local
temperature, precipitation, cloud cover, solar radiation, and the type of plants you are growing. The rate
expresses the amount of water lost from a cropped surface in units of water depth in a unit area of land taken in
the same units as the depth. The time unit can be an hour, day, decade, month or even an entire growing period
or year. The parameter is usually compared with a Reference EvapoTranspiration (ET0) that is used to study the
demand of water requirement independent of crop type and variations in agricultural practices used. ET0 is used
to compute the ET on a hypothetical grass reference crop with specific characteristics6,36,37.
Since this method involves modeling a scientific equation, there is no learning phase and the
parameters for the day are simply substituted into the equation in order to calculate the score. Since empirical
equations are considered, this method performs poorly on real-time scenarios due to uncertainties that are not
taken into account in these equations.
The Evapotranspiration measure is derived from the FAO Penman-Monteith equation is computed as:
where Lv is the volumetric latent heat of vaporization i.e. the energy required per water volume vaporized, E is
the mass water evapotranspiration rate in g / sec m2, ETo is the water volume evapotranspired in mm / sec, Δ is
the rate of change of saturation specific humidity with air temperature in Pa / K, Rn is the net irradiance in W /
m2 which is the external source of energy flux, G is the ground heat flux which is a little difficult to measure, in
W / m2, Cp is the Specific heat capacity of air in J / kg K, ρa is the dry air density in kg / m3, δe is the specific
humidity in Pa ga is the atmospheric conductance in m / s, gs is surface conductance in m / s, γ is the
Psychrometric constant which is close to 66 Pa / K38.
2. Logistic Regression
Logistic regression is a very effective binary classifier that can be used to try to predict the water
requirement as a time series data. Logistic regression can be Simple or Multivariate. Simple logistic regression
is used to predict binary values. In this case, given the amount of water to be used, the model can predict
whether the specification is adequate or not. This is a naive approach and predicted with a rather low accuracy.
Multivariate Logistic Regression takes into account various features to provide decisions39,40. This can
be seen as the simplest feed-forward neural network with one hidden layer and therefore one weight. In this
model, each class specifies the range of water required for irrigation. However, this model does not provide
crop-specific decisions and in no way maps the temporal dependencies35.
Logistic regression is modeled as:
where X is the set of feature vectors of size n, Yi is the predicted class label, W and B are the weights and biases
respectively that would be learned by means of the gradient descent algorithm41.
3. Decision Tree Classifier
A decision tree deduces logical relationships between the class labels in the form of a hierarchical tree.
The patterns explored by this algorithm is known as logic rules. Therefore, from the classifying attributes,
classifier attributes are deduced42. Decision trees are made of nodes and leaves where the nodes represent the
classifying attributes and the leaves represent the classifier attributes. These logical rules are framed with the
use of information gain based on the C4.5 algorithm43.
The model built using Decision Tree Classifiers has been found to provide greater accuracy than
logistic regression. As with the problem of logistic regression, it is difficult to provide crop-specific predictions
and model the temporal dependencies using this model44.
Fig 1. Decision Tree Classifier
4.SysFor
SysFor stands for Systematically developed Forests of multiple tress. This algorithm is a modification
of the Decision Tree Classifier. The principle behind SysFor (Systematic Forests) is that better knowledge can
be gained through extraction of better knowledge. This algorithm consists of several different steps. First, the
‘goodness’ of attributes and the split points in the dataset are determined using the gain ratio. The gain ratio is
computed as:
where Info refers to the entropy and is computed as the probability of value being every unique element in the
dataset in each and every feature. The gain ratio Gain is computed as the difference in entropy before the split
and after the split occurred. This ratio is used to find the split points and the best features45.
The next step is the specification of the number of trees used for the forest by the user. Each good
attribute found is made the root attribute of a distinct tree. The first tree's root attribute is used to logically split
the feature selection and prediction process horizontally and join the trees created in each split46. These steps are
repeated till the user-defined number of trees are created in the forest. From this forest, the required classifier
attributes are selected based on the logic rules of each tree47,48.
The model used with this algorithm provides the best results with crop-specific predictions as found out in Khan
et al35. This model has been found to be superior to even ANN models for the CIA land area, even though these
models are not designed to be used in time series prediction.
5.Support Vector Machine(SVM)
An SVM tries to find the best amongst all the linear classifiers that can be possible between any 2
classes, over several high dimensional planes of data. This is achieved by construction of a hyperplane per
dimension i.e. a set of hyperplanes in a higher dimensional space like the problem of crop irrigation demand
prediction. The class labels used are identified in such a way that the distance between the hyperplanes used to
find the best possible linear classifier is maximized. This ensures maximum differentiability of the class labels
involved. The margin can be of two types- soft margin and hard margin. In soft margin, certain number of data
points in each class are allowed to disobey the margins, which is not possible in hard margin SVMs. The
margins are found using non-linear functions like sigmoid, tanh, gaussian function etc. Furthermore, if the
number of features turns out to be very less, special Kernel functions can be used that form substituted
combinations of the already existing features in a non linear space like sigmoid, tanh, exponential etc. A
common practice used with these SVMs is the use of hierarchical microclustering to accommodate for huge
datasets which all cannot be considered at once because of the complexities of computation of the centroid of a
large number of data points. 49,50 The SVM classifier optimizes:
where J is the cost function optimized based on the distance between the hyperplanes with respect to the linear
classifier and yi is the output class label for item i.
The model developed with this algorithm has found to show surprisingly low accuracy. Also, the
training period of this model is large. While this model does provide predictions that are crop-specific, the lack
of ability to map time series correlations is also a problem.
6. Fuzzy Systems:
A. Remote Sensing and Crop Models
Jones et al., in 2000, came up with an integrated module that combined remote sensing and crop models
for decision support crop management. The model takes the remote sensed data, corrects and calibrates the data,
and interprets this data, making it ready for training. Now, this data is sent to the crop and decision models for
them to compute a fuzzy composite modeling scheme. The entire model is supplemented with ancillary data i.e.
various other dependent parameters like soil type, rainfall etc. The farm management decision system involved
3 main decision-making criteria: profitability, environment and sustainability. A simple fuzzy composite
programming to try to estimate various parameters relevant to the agricultural domain. It is a farm management
system on the whole more than a targeted water level predictor51.
B. Fuzzy Logic Crop Water Stress Index (FL-CWSI)
Around the same time, Al-Faraj et al. built a rule-based Fuzzy Logic Crop Water Stress Index (FLCWSI) which returns an index between 0 and 1 corresponding to the metric CWSI using fuzzy logic techniques.
This index was usually a result of sigmoid, Gaussian, absolute difference or a variety of other membership
functions used in fuzzy logic52. It eliminated the necessity of using various other factors like aerodynamic
resistance etc. These models were found to perform better than the rule based models used for irrigation
prediction53.
C. ANFIS
By 2007, a new model was used for prediction of daily irrigation water demand using Adaptive NeuroFuzzy Inference Systems (ANFIS) technique. This first order Sugeno fuzzy model has 2 inputs and one output.
The inputs are carried through 5 layers before an output prediction is reached. The first layer passes the input
through a membership function that transforms input and convert it into a fuzzy set range. The second layer
provides a product of the two membership function results on the two inputs. The third layer normalizes the
influence of various rules by taking a ratio of the firing rate. The fourth layer introduces various parameters that
will be used to predict accurate values The learning of these parameters usually takes place by means of a time
series algorithm like ARMA, AR, ARIMA etc. The fifth layer computes a weighted average of all the input
signals out of the 4th layer54. These models use the backpropagation algorithm through the 5 steps in reverse to
facilitate the training process. They have a much better accuracy predicting irrigation requirement when
compared to the Auto Regression and the ARMA models when Root Mean Squared (RMS) and Mean Absolute
Percentage Error (MAPE)55.
Fig 2. ANFIS System with 2 inputs and 1 output
D. ARIMA
In 2009, Wang et al. published a paper surveying the various kernel-based and other AI algorithms that
were used for water flow prediction. Also, the analysis by Landeras et al. in 2009 revealed that the the ARIMA
model and the ANN model, when used for calculation of evapotranspiration, gave a 6-8% reduction in the error
percentage. Also, the work by Wang et al. showed that AR, ARMA, ANFIS models gave much better results
when compared to the ANN models in the domain of water flow prediction56. ARIMA models are generalized
class of models that can predict a stationary time series. Sometimes, stationarity detrending by means of taking
the dth differential and modeling that as a stationary series, non-linear transformations etc. are carried out to
ensure correct predictive results. A stationary series is constant in graph throughout the time series i.e. its
central tendencies tend to remain constant inside of a range. This also acts as a disadvantage to this model as
stationarity is a hard constraint to enforce on the dynamic real time data57–59.
ARIMA(p,d,q) is given by:
where Xt is the predicted output at time step t, µ is the mean over the entire dataset, the summation from
1 to p is the autoregressive model that learns predictions based on a weighted sum of a predefined number of
time step values, the summation from 1 to q is the moving average model that maps the error in prediction of
every time step, ξt is the error parameter that is permitted. The parameter d specifies the number of times the
given input data needs to be differentiated in order to achieve a data distribution that is stationary, where the
differentiation is carried out at each level by means of finding the difference between the timesteps. Of these
parameters, ξt, Φ and γ are the learnable parameters of the model60.
E. Fuzzy inference to single factor prediction
By 2010, a new model that made use of fuzzy inference to single factor prediction method used to
predict the amount of irrigation required. First, pre-recorded examples are taken and discretized into a universe
U. Then a reasoning model of single factor prediction characterized by (e-f)-fold by e-degree compound fuzzy
implicative propositional logic. The excessive dimensionality problem is taken care of using logic primitives
that work to reduce the feature space. The inference rules present are taken and conformed to a required size set
to ensure a rule based solution61.
F. Hybrid Model
There have been a lot of hybrid models that have been used for time series prediction effectively that
involve various combinations of ANNs, fuzzy logic and Genetic Algorithms in various other domains61–65. In
2010, Paulido-Calvo et al. came up with a hybrid model of combining CNNs with fuzzy membership function,
optimized by a genetic optimization algorithm to test the efficacy of irrigation demand prediction. He found that
the selection of the best out of 5 ANN models in triangular partitions and subjecting that through a fuzzy
membership function, optimized by genetic optimization algorithms using IF-THEN rules gave the best results.
The application of fuzzy algorithms was quickly followed by defuzzification algorithms to produce the required
output. The genetic operations of reproduction, crossover and mutation were defined to facilitate the genetic
optimization. This system beat the accuracy of the conventional ANN system1.
G. Inexact rough interval fuzzy linear programming
Lu et al. tried to use inexact rough interval fuzzy linear programming method to generate conjunctive
water allocation strategies. The rough interval method embodies dual-uncertain parameters quite robustly. The
results from this model were comparable to the existing state-of-the-art even though the execution time of the
algorithm was much better than the contemporary model whose performance is similar to this one. One of the
main advantages in using this model is its ability to predict good results in both normal and special system
conditions, making it a more generalized solution that can be achieved by selective targeting of features66.
H. Decision Support System
In 2014, Mousa et al. developed a fuzzy based decision support system that evaluated evapotranspiration
value using the computational formula before being used in training to predict the amount of irrigation required
by means of use of the decision tree algorithm. The model took input from sensors of wind speed, temperature,
humidity etc and also given as input were the irrigation strategies used and the equipment used to see if these
factor into the prediction methodology. The irrigation strategy planned and equipment information gave the
model good performance and considerably better speed of execution. This model gave accurate and fast results
when compared to some of the other models like AR, ARMA etc.67
7. ANNs
Artificial Neural Networks are those architectures in which the input is morphed into a plane where a
linear separation or regression is made possible by means of a non-linearity to warp the input into a plane where
a straight line is enough. Various hyper-parameters used in an ANN would be hidden layer size, number of
hidden layers, batch size, activation function, learning rate, optimization function, initialization strategy and
normalization. The model is trained by means of various training algorithms like BackPropagation, radial basis
function etc. The complexity of what the model can model is usually proportional to the number of layers in the
architecture68–70.
where a is the activation function used, which is usually any number of nonlinearities like sigmoid, tanh etc.
The model is trained as a composition of activation functions that occurs the same number of times as the
number of layers to give the final output. While these models operate in the nonlinear space and have much
more scope for accommodating generic input, these models don’t have the capabilities to map the time series
dependencies found in irrigation demand prediction71.
Around 2001, the performance of artificial neural network architectures was compared with most of the
other models at the time like ARMA, AR, SVMs etc. It was found that ANNs performed consistently worse
than the SMT and the ANFIS models56. But by 2003, ANNs were tweaked efficiently enough to become better
in performance than the other models like ANFIS, ARMA etc.72 Hardaha et al. did extensive work on tweaking
the various parameters of the ANN to give an exploratory understanding on the dependence of each hyperparameter to the accuracy of the model. It was found that training based on the radial basis function gave the
best results for prediction of water requirement of wheat crops73.
A. Artificial Neuro-Genetic Networks
A hybrid ANN model that incorporates a genetic algorithm on top of an ANN to improve the accuracy
of the model was proposed in 2009. This model used the Artificial Neuro-Genetic Networks (ANGN) to predict
the irrigation requirement of the crops. Initially, the collection of ANNs of size N are taken to be the entire
population. A first generation would contain an initialized collection of ANNs all with their own hyperparameters. The output from the initial population is sent into the ANN chosen is then used to train the
particular input. Now the accuracies of the different models are checked to see which ANGN selection would
result in the best performance. To optimize the entire process, a pareto front is added before the output is
specified. Objective evaluation functions F1 and F2 are first calculated over the entire corpus and are further
used in the ANGN selection process74.
A study by Khan et al. in 2013 compares different AI models and their accuracies when it comes to
irrigation prediction. It was found that of all the models, the model with the 3-fold cross validation multiple
decision trees SysFor model gave the best overall results. But the actual water content required by the crop was
accurately predicted by ANNs really accurately. The difference in error percentage between ANNs and SysFor
was almost 20%. Thus, it was concluded that SysFor, ANN and decision tree techniques are the most suitable
for the task of irrigation prediction35.
COMPARISION OF PERFORMANCES
Model Used
Dataset / Study area
Time Step
Accuracy (%)
Reference
Logistic Regression
Decision Tree
Classifier
SysFor
Support Vector
Machines
Remote Sensing and
Crop Models
FL-CWSI
Coleambally, Australia
Coleambally, Australia
1 day
1 day
56
74
Khan et al.35
Khan et al.35
Coleambally, Australia
Coleambally, Australia
1 day
1 day
78
64
Khan et al.35
Khan et al.35
MAC, University of Arizona
58 to 78
Jones et al.51
CWSI
Comparison Only
MAPE
comparison Only
ET Comparison
Only
Value
Comparison Only
Al-Faraj et al.53
ANFIS
Chania in Crete, Greece
Variable (Rule
Based)
Index – no
time series info
1 day
ARIMA
Álava, Basque County, Spain
1 week
Fuzzy inference to
Single Factor
Prediction
Hybrid ANN, Fuzzy,
GA
Inexact rough
interval fuzzy
Decision Support
System
ANNs
Shangqiu, Henan, China
1 year
Cordoba province, Spain
1 day
Custom dataset
Custom
Baghdad, Iraq (Simulation)
1 month
Bhakra Canal system,
Rajasthan, India
Bembézar Irrigation District
(Spain)
1 month
ANGN
Horticulture Dept., UNL
1 day
Atsalakis et al.55
Landeras et al.59
Chen et al.61
79.73 (in terms of
20.27 SEP %)
Comparison Only
Pulido-Calvo et
al.1
Lu et al.66
ET Comparison
Only
r2 Comparison
Only
87.37 (in terms of
12.63 SEP %)
Mousa et al.67
Hardaha et al.73
Perea et al.74
Table 1 Comparison of Models
Simply using fuzzy system architectures will help capture the randomness in the representation and the
fading of hard boundaries at inflection points, but will not do much about the sequential nature of the data.
ARIMA models take into context the temporal dependencies but only work for univariate regression.
Multivariate regression is essential for irrigation prediction. Even the ANFIS architecture is not designed to
capture time series dependencies. ANNs are very good at prediction based on training, but lack the capability of
inferring semantic meaning from the sequential flow. SVMs and other kernel based methods are designed to be
trained on normal data and will not be able to capture the essential sequential information. ANGN also fails to
map temporal dependencies when reading features and predicting outputs.
PROPOSED MODEL
The above mentioned fallacies have left us in an unfortunate situation of choosing amongst a pool of
ordinary choices and simply settling for whatever accuracy we can obtain. Hence, we propose a novel
methodology of using a sequence learning based recurrent neural network (RNN) model that uses the LSTM
activation function to model for irrigation requirement so that it doesn’t have memory problems on long input
streams. The input historical context will be stored inside the context vector which will be passed on through
the mapping of the output through time in the network75–78.
where a is the activation function, Xt is the input at time step t, W, Wh and B are learnable parameters of the
model and h refers to the hidden layer output that occurs during every time step t.
The underlying architecture of the should contain three to five hidden LSTM or GRU layers, of size
n, that are all back-propagated during training. The model uses BackPropagation Through Time (BPTT)
algorithm for training and tweaking of the layer weights. The model will typically contain thousands of weight
vectors all trained for each sequence input. The input will be in the form of features like temperature, rainfall,
humidity etc. determined using real time sensors. A fixed context needs to be assigned for how much of the
historic context is required for prediction. Every time step will involve processing all the features of that
particular time step13,79,80.
A typical LSTM circuit is described by:
where i is the input gate, f the forget gate, o the output gate, c the cell gate and h the hidden layer output. The
interactions in the cell state by means of which the mapping of the backpropagated gradients occurs not
separately but alongwith a cell state interaction that ensures that the gradient never reaches zero81.
Fig 3. RNN LSTM Model Proposed
The advantages of using such an architecture are plentiful. The RNN is robust enough to map the
intrinsic variations in observations that occur due to various factors like global warming, faulty equipment etc.
Uncertainty will be automatically detected and ignored by the RNN. Also, RNNs are designed to incorporate
sequence information inside of the hidden layer vector as context. The downsides to this model are the
computational feasibility and the availability of hardware (GPUs) powerful enough to train a big model in less
time, and the requirement of a huge corpus for training, which is why it was not possible to implement this
model. Nevertheless, the runtime of the model will be really fast because testing is merely a few multiplications.
Another advantage of this model is its efficiency in storage. Deep learning models only occupy very small
amounts of space (in hundreds of MB) when compared to the traditional AI algorithms (a few GB)77.
Because of all these factors, and the recent success of RNN sequence to sequence architectures, the
application of this architecture should improve the current SOTA by quite a bit.
REFERENCES
1.
Pulido-Calvo I, Gutierrez-Estrada JC. Improved irrigation water demand forecasting using a softcomputing hybrid model. Biosyst Eng 2009; 102: 202–218.
2.
Sivakumar MVK, Stigter CJ, Rijks D. Foreword. Agric For Meteorol 2000; 103: 1–2.
3.
Bhalla GS, Singh G. Economic liberalisation and Indian agriculture: a district-level study. Book, SAGE
Publications India, 2012.
4.
Cain JD, Jinapala K, Makin IW, et al. Participatory decision support for agricultural management. A case
study from Sri Lanka. Agric Syst 2003; 76: 457–482.
5.
Singh V, Singh UC. Assessment of groundwater quality of parts of Gwalior (India) for agricultural
purposes. Indian J Sci Technol 2008; 1: 1–5.
6.
Allen RG, Pereira LS, Raes D, et al. Crop evapotranspiration-Guidelines for computing crop water
requirements-FAO Irrigation and drainage paper 56. FAO, Rome 1998; 300: D05109.
7.
Smith M. CROPWAT: A computer program for irrigation planning and management. Book, Food &
Agriculture Org., 1992.
8.
Wu IP. Energy gradient line approach for direct hydraulic calculation in drip irrigation design. Irrig Sci
1992; 13: 21–29.
9.
Alvisi S, Franchini M, Marinelli A. A short-term, pattern-based model for water-demand forecasting. J
Hydroinformatics 2007; 9: 39–50.
10.
Zhou SL, McMahon TA, Walton A, et al. Forecasting operational demand for an urban water supply
zone. J Hydrol 2002; 259: 189–202.
11.
Patel H, Patel D. A Comparative Study on Various Data Mining Algorithms with Special Reference to
Crop Yield Prediction. Indian J Sci Technol; 9.
12.
Gers FA, Eck D, Schmidhuber J. Applying LSTM to time series predictable through time-window
approaches. In: International Conference on Artificial Neural Networks. Inproceedings, 2001, pp. 669–
676.
13.
Gers FA, Schmidhuber J, Cummins F. Learning to forget: Continual prediction with LSTM. Neural
Comput 2000; 12: 2451–2471.
14.
Graves A, Jaitly N, Mohamed A. Hybrid speech recognition with deep bidirectional LSTM. In:
Automatic Speech Recognition and Understanding (ASRU), 2013 IEEE Workshop on. Inproceedings,
2013, pp. 273–278.
15.
Graves A, Schmidhuber J. Framewise phoneme classification with bidirectional LSTM and other neural
network architectures. Neural Networks 2005; 18: 602–610.
16.
Sundermeyer M, Oparin I, Gauvain J-L, et al. Comparison of feedforward and recurrent neural network
language models. In: 2013 IEEE International Conference on Acoustics, Speech and Signal Processing.
Inproceedings, 2013, pp. 8430–8434.
17.
Sak H, Senior A, Beaufays F. Long short-term memory based recurrent neural network architectures for
large vocabulary speech recognition. arXiv Prepr arXiv14021128.
18.
Hanks RJ, Hill RW, others. Modeling crop responses to irrigation in relation to soils, climate and
salinity. Book, International Irrigation Information Center., 1980.
19.
Murthy ASR, Sudheer Y, Mounika K, et al. Cloud Technology on Agriculture using Sensors. Indian J
Sci Technol; 9.
20.
Kumar MK, Ravi KS. Automation of Irrigation System based on Wi-Fi Technology and IOT. Indian J
Sci Technol; 9.
21.
Schlenker W, Roberts MJ. Nonlinear temperature effects indicate severe damages to US crop yields
under climate change. Proc Natl Acad Sci 2009; 106: 15594–15598.
22.
O’Toole JC, Hatfield JL. Effect of wind on the crop water stress index derived by infrared thermometry.
Agron J 1983; 75: 811–817.
23.
Fritschen LJ. Net and solar radiation relations over irrigated field crops. Agric Meteorol 1967; 4: 55–62.
24.
Fischer RA. Number of kernels in wheat crops and the influence of solar radiation and temperature. J
Agric Sci 1985; 105: 447–461.
25.
Tolk JA, Howell TA, Evett SR. Effect of mulch, irrigation, and soil type on water use and yield of maize.
Soil Tillage Res 1999; 50: 137–147.
26.
Rao AVMS, Chandran MAS. Crop water requirements. Agrometeorol Data Collect Anal Manag 1977;
51.
27.
Slatyer RO. The influence of progressive increases in total soil moisture stress on transpiration, growth,
and internal water relationships of plants. Aust J Biol Sci 1957; 10: 320–336.
28.
Loew A, Ludwig R, Mauser W. Derivation of surface soil moisture from ENVISAT ASAR wide swath
and image mode data in agricultural areas. IEEE Trans Geosci Remote Sens 2006; 44: 889–899.
29.
Renard KG, Foster GR, Weesies GA, et al. Predicting soil erosion by water: a guide to conservation
planning with the Revised Universal Soil Loss Equation (RUSLE). Book, US Government Printing Office
Washington, DC, 1997.
30.
Moran MS, Clarke TR, Inoue Y, et al. Estimating crop water deficit using the relation between surfaceair temperature and spectral vegetation index. Remote Sens Environ 1994; 49: 246–263.
31.
Allen RG. Using the FAO-56 dual crop coefficient method over an irrigated region as part of an
evapotranspiration intercomparison study. J Hydrol 2000; 229: 27–41.
32.
Jensen ME, Wright JL, Pratt BJ. Estimating soil moisture depletion from climate, crop and soil data.
Trans ASAE 1971; 14: 954–959.
33.
Phocaides A. Technical handbook on pressurized irrigation techniques.
34.
Igbadun HE. Irrigation Scheduling Impact Assessment MODel (ISIAMOD): A decision tool for
irrigation scheduling. Indian J Sci Technol 2012; 5: 3090–3099.
35.
Khan M.A., Islam M.Z. HM. Evaluating the Performance of Several Data Mining Methods for Predicting
Irrigation Water Requirement. In: Zhao Y. LJKPJ, Christen P (eds) Data Mining and Analytics 2012
(AusDM 2012). Inproceedings, Sydney, Australia: ACS, pp. 199–208.
36.
Brouwer C, Heibloem M. Irrigation water management: irrigation water needs. Train Man; 3.
37.
Jensen ME, Burman RD, Allen RG. Evapotranspiration and irrigation water requirements. Inproceedings,
1990.
38.
Zotarelli L, Dukes MD, Romero CC, et al. Step by step calculation of the Penman-Monteith
Evapotranspiration (FAO-56 Method). Inst Food Agric Sci Univ Florida.
39.
Christensen R. Linear models for multivariate, time series, and spatial data. Book, Springer Science &
Business Media, 1991.
40.
Christensen R. Log-linear models and logistic regression. Book, Springer Science & Business Media,
2006.
41.
Bottou L. Large-scale machine learning with stochastic gradient descent. In: Proceedings of
COMPSTAT’2010. Incollection, Springer, 2010, pp. 177–186.
42.
Islam MZ. EXPLORE: A Novel Decision Tree Classification Algorithm. In: MacKinnon LM (ed) Data
Security and Security Data: 27th British National Conference on Databases, BNCOD 27, Dundee, UK,
June 29 - July 1, 2010. Revised Selected Papers. Inbook, Berlin, Heidelberg: Springer Berlin Heidelberg,
pp. 55–71.
43.
John GH, Kohavi R, Pfleger K, et al. Irrelevant features and the subset selection problem. In: Machine
learning: proceedings of the eleventh international conference. Inproceedings, 1994, pp. 121–129.
44.
Quinlan JR. C4.5: Programs for Machine Learning. Book, San Francisco, CA, USA: Morgan Kaufmann
Publishers Inc., 1993.
45.
Quinlan JR. Induction of decision trees. Mach Learn 1986; 1: 81–106.
46.
Islam MZ, Giggins H. Knowledge Discovery through SysFor - a Systematically Developed Forest of
Multiple Decision Trees. In: Vamplew P. SAOK-LCP, Kennedy PJ (eds) Australasian Data Mining
Conference (AusDM 11). Inproceedings, Ballarat, Australia: ACS, pp. 195–204.
47.
Islam MZ. EXPLORE: a novel decision tree classification algorithm. In: British National Conference on
Databases. Inproceedings, 2010, pp. 55–71.
48.
Mitchell TM. Machine learning. 1997. Burr Ridge, McGraw Hill 1997; 45: 37.
49.
Burges CJC. A tutorial on support vector machines for pattern recognition. Data Min Knowl Discov
1998; 2: 121–167.
50.
Vapnik VN. The Nature of Statistical Learning Theory. Book, New York, NY, USA: Springer-Verlag
New York, Inc., 1995.
51.
Jones D, Barnes EM. Fuzzy composite programming to combine remote sensing and crop models for
decision support in precision crop management. Agric Syst 2000; 65: 137–158.
52.
Zadeh LA. Fuzzy sets. Inf Control 1965; 8: 338–353.
53.
Al-Faraj A, Meyer GE, Horst GL. A crop water stress index for tall fescue (Festuca arundinacea Schreb.)
irrigation decision-making—a traditional method. Comput Electron Agric 2001; 31: 107–124.
54.
Jang J-S. ANFIS: adaptive-network-based fuzzy inference system. IEEE Trans Syst Man Cybern 1993;
23: 665–685.
55.
Atsalakis G, Minoudaki C, others. Daily irrigation water demand prediction using Adaptive Neuro-Fuzzy
Inferences Systems (ANFIS). Inproceedings.
56.
Wang W-C, Chau K-W, Cheng C-T, et al. A comparison of performance of several artificial intelligence
methods for forecasting monthly discharge time series. J Hydrol 2009; 374: 294–306.
57.
Box GEP, Pierce DA. Distribution of residual autocorrelations in autoregressive-integrated moving
average time series models. J Am Stat Assoc 1970; 65: 1509–1526.
58.
Valipour M, Banihabib ME, Behbahani SMR. Comparison of the ARMA, ARIMA, and the
autoregressive artificial neural network models in forecasting the monthly inflow of Dez dam reservoir. J
Hydrol 2013; 476: 433–441.
59.
Landeras G, Ortiz-Barredo A, López JJ. Forecasting weekly evapotranspiration with ARIMA and
artificial neural network models. J Irrig Drain Eng 2009; 135: 323–334.
60.
Wei WW-S. Time series analysis. Book, Addison-Wesley publ Reading, 1994.
61.
Chen Q, Jing L, Gao Y. Study on Applying Fuzzy Inference to Single Factor Prediction Method for
Precipitation Irrigation Requirement Forecast. In: 2010 2nd International Workshop on Intelligent
Systems and Applications. Inproceedings, 2010.
62.
Hassan MR, Nath B, Kirley M. A fusion model of HMM, ANN and GA for stock market forecasting.
Expert Syst Appl 2007; 33: 171–180.
63.
Mukerji A, Chatterjee C, Raghuwanshi NS. Flood forecasting using ANN, neuro-fuzzy, and neuro-GA
models. J Hydrol Eng 2009; 14: 647–652.
64.
Ruan D, Wang PP. Intelligent hybrid systems: fuzzy logic, neural networks, and genetic algorithms.
Book, Springer, 1997.
65.
Kavousi-Fard A. A new fuzzy-based feature selection and hybrid TLA--ANN modelling for short-term
load forecasting. J Exp Theor Artif Intell 2013; 25: 543–557.
66.
Lu H, Huang G, He L. An inexact rough-interval fuzzy linear programming method for generating
conjunctive water-allocation strategies to agricultural irrigation systems. Appl Math Model 2011; 35:
4330–4340.
67.
Mousa AK, Croock MS, Abdullah MN. Fuzzy based Decision Support Model for Irrigation System
Management. Int J Comput Appl; 104.
68.
Yegnanarayana B. Artificial neural networks. Book, PHI Learning Pvt. Ltd., 2009.
69.
Raman H, Sunilkumar N. Multivariate modelling of water resources time series using artificial neural
networks. Hydrol Sci J 1995; 40: 145–163.
70.
Zhang GP. Time series forecasting using a hybrid ARIMA and neural network model. Neurocomputing
2003; 50: 159–175.
71.
Zhang GP, Patuwo BE, Hu MY. A simulation study of artificial neural networks for nonlinear time-series
forecasting. Comput Oper Res 2001; 28: 381–396.
72.
Pulido-Calvo I, Roldán J, López-Luque R, et al. Demand forecasting for irrigation water distribution
systems. J Irrig Drain Eng 2003; 129: 422–431.
73.
Hardaha MK, Chouhan SS, Ambast SK. APPLICATION OF ARTIFICIAL NEURAL NETWORK IN
PREDICTION OF RESPONSE OF FARMERS’WATER MANAGEMENT DECISIONS ON WHEAT
YIELD. J Indian Water Resour Soc; 32.
74.
Perea RG, Poyato EC, Montesinos P, et al. Irrigation Demand Forecasting Using Artificial NeuroGenetic Networks. Water Resour Manag 2015; 29: 5551–5567.
75.
Medsker LR, Jain LC. Recurrent neural networks. Des Appl.
76.
Funahashi K, Nakamura Y. Approximation of dynamical systems by continuous time recurrent neural
networks. Neural networks 1993; 6: 801–806.
77.
Connor JT, Martin RD, Atlas LE. Recurrent neural networks and robust time series prediction. IEEE
Trans neural networks 1994; 5: 240–254.
78.
Dorffner G. Neural networks for time series processing. In: Neural Network World. Inproceedings, 1996.
79.
Werbos PJ. Backpropagation through time: what it does and how to do it. Proc IEEE 1990; 78: 1550–
1560.
80.
Mikolov T, Karafiat M, Burget L, et al. Recurrent Neural Network based Language Model. Interspeech
2010; 1045–1048.
81.
Hochreiter S, Schmidhuber J. LSTM can solve hard long time lag problems. Adv Neural Inf Process Syst
1997; 473–479.
| 9cs.NE
|
International Journal of Managing Public Sector Information and Communication Technologies (IJMPICT)
Vol. 3, No. 2, December 2012
A NEW APPROACH IN DYNAMIC TRAVELING
SALESMAN PROBLEM: A HYBRID OF ANT COLONY
OPTIMIZATION AND DESCENDING GRADIENT
Farhad Soleimanian Gharehchopogh1, Isa Maleki2, Seyyed Reza Khaze3
1,2,3
Computer Engineering Department, Science and Research Branch, Islamic Azad
University, West Azerbaijan, Iran
{bonab.farhad1, maleki.misa2, khaze.reza3}@gmail.com
ABSTRACT
Nowadays swarm intelligence-based algorithms are being used widely to optimize the dynamic traveling
salesman problem (DTSP). In this paper, we have used mixed method of Ant Colony Optimization (AOC)
and gradient descent to optimize DTSP which differs with ACO algorithm in evaporation rate and
innovative data. This approach prevents premature convergence and scape from local optimum spots and
also makes it possible to find better solutions for algorithm. In this paper, we’re going to offer gradient
descent and ACO algorithm which in comparison to some former methods it shows that algorithm has
significantly improved routes optimization.
KEYWORDS
DTSP, ACO, Gradient Descent, Combinational Methods.
1. INTRODUCTION
Nowadays method of efficiently-solving large number of problems known as NP-Hard
problems is one of the most difficult combinational optimization issues. Solving such problems
means to find the best solution among a large set of solutions to the problem. DTSP is one of
the most famous combinational optimization problems. This problem initially introduced by
Psaraftis in 1988 [1]. In recent decades many Heuristic methods have been suggested for
solving DTSP’s and these methods are more flexible in comparison to the classical and
traditional methods. Some heuristic methods such as combinational algorithms [2], ACO
algorithms [3, 4, and 5] and evolutionary computation [6, 7] can be mentioned in this field.
Swarm intelligence methods, are tools to find optimal answers or solutions. These methods take
advantage of search and cooperation to look for an optimization problem’s search space. So the
more one algorithm controls these parameters, more near-optimal solutions finds to the
problem. ACO algorithm refers to a class of meta-heuristic algorithms that have adapted from
the social behaviour of ants in nature. These algorithms’ aim is to find optimal solutions in
reasonable time. Ants are autonomous factors in nature and when they work together, their
performance is really smart [8].
The optimization-problems’ aim is to find optimal or near-optimal points. In some functions in
addition to public minimum there are some other local minimums.in that case it’s not possible
to easily find the public optimal point. In such problems, we can use combinational algorithms
in search space to find optimal solution for problem. Gradient descent is an optimization
algorithm to find minimum points in function [9, 10]. Gradient descent refers to optimizing
processes and adjustment of input variables affect the output of the process in them.
Mathematical methods use a collection of efficient and necessary circumstances that are true in
optimization problem’s answer. Presence or absence of limitations in the optimization is
essential in these approaches. In optimization problems without limitations such as gradient
methods, the aim is to find optimal or near-optimal solution to an objective function [11, 12].
DOI : 10.5121/ijmpict.2012.3201
1
International Journal of Managing Public Sector Information and Communication Technologies (IJMPICT)
Vol. 3, No. 2, December 2012
Beside swarm intelligence methods that are known as smart optimization methods, evolutionary
computation or meta-heuristic methods, using mathematical methods, depending on the
application may be faster. Gradient descent algorithm can process the searching in complex
combinational problems space to find near-optimal solutions with high performance by creating
an appropriate balance in ACO algorithms.
The overall structure of our paper is organized as follows: We have organized the general
structure of this paper as following: In section second introduce literature review; in section
third we introduce descent gradient algorithm; in section fourth of this paper we’ll introduce
ACO; in section five we’ll explain the proposed algorithm, in section sixth we’ll discuss about
valuation and results of the proposed method and finally in section seventh we will draw some
conclusions from this paper.
2. LITERATURE REVIEW
For more accurate optimization results, we should use pre-heuristic methods (such as ACO
algorithm, Genetic Algorithm, etc.) and to optimize possible solutions, we must use
computational methods (e.g. gradient descending method). As DTSP has lots of local
minimums, Computational methods, start from the basic hypothetical point to find first and
closest near-optimal solution, and if it find the solution, defines it as optimal point; while metaheuristic methods Randomly in high repetitions are looking for the best solution and by jump
out of point to another trying to find the best and most efficient point and provide optimal
solutions. In recent years meta-heuristic methods have been considered to find near-optimal
solutions to combinational optimization problems.
F.S.Gharehchopogh et al. in [2] has used ACO algorithm and genetic algorithm combinational
algorithms to solve DTSP. They have used genetic algorithm in his paper to optimize paths
obtained from ACO algorithm in search space. In this paper chromosome genes in genetic
algorithm are paths obtained from ACO algorithm. M.Guntsch in [3] has solved DTSP using
ACO algorithm. He has used probability of motion and update rules changes to optimization. As
gradient descent is local and only has high potential to find minimum and maximum points, it
offers probability of better and more acceptable answers to find optimal solutions to the DTSP
in comparison to the meta-heuristic methods.
The increasing usage of mathematical optimization techniques has led many researchers to
focus on developing and optimize problems using mathematical operations. Since there is no
unique approach to solve all the optimization problems, various computational methods for
solving optimization problems have emerged. Gradient descent algorithm is the basis of many
optimization algorithms and it’s being used to learn artificial neural networks and minimize the
network errors to adjust the network parameters [13]. In [14] X.Wang has utilized gradient
descent usages. He has used gradient descent in order to solve complex integrals. D.Kumar &
Y.Kumar in [15] have used optimization methods. In their journal they have they have studied
the optimization methods that are gained from the nature. Among these methods and ways they
have marked gradient descent as an optimization method to find the max or min of the
problems.
3. GRADIENT DESCENT ALGORITHM
There is a group of gradient algorithms in which the process of searching for gradient is used
with some changes, to find the optimal answer. The continuous answers which are found in
continuous repeats are similar to the main optimal answer. So, we stop when the experimental
answers are gotten similar to the optimal answer much enough [16, 17].
Since there is no limitation in the problem it seems that moving in the same direction as gradient
could be a practical way [17]. If the purpose of the problem is to minimize the f(x) we have to
move in the opposite direction as the gradient. In the other words, we should use equation (1) in
order to achieve the answer [16].
2
International Journal of Managing Public Sector Information and Communication Technologies (IJMPICT)
Vol. 3, No. 2, December 2012
xn+1 = xn − t n ∇f ( xn )
(1)
In the above equation (1) the function has been begun from an arbitrary point to find the
optimal answer and the function also performs minimizing in the same direction as gradient. In
the equation (1), t is a non-negative scalar which minimizes the function in the same direction as
gradient. The new gradient is created by the previous one in each repeat.
When x is decreased we move from the current experimental answer to the next answer. Thus,
in order to generalize a searching method for a variable, firstly we should specify the movement
direction of the gradient using the partial derivatives. As we assume that f(x) is derivable, in
each point x will have a gradient vector which is represented in the equation (2) as ∇f(x') [16].
∇f ( x′) = (
∂f ∂f
∂f
,
,...,
)
∂x1 ∂x2
∂xn
x = x′
(2)
The reason that gradient is important is that we can find minimal changes in function in x point
as we move in the same direction as gradient.
Pseudo code for the algorithm of the gradient descent is represented in the following:
Choose randomly
x0
While ||f ( xn +1 ) – f ( xn ) > Ɛ|| do
γ n (generally 1/n)
Compute x n +1 = xn - t n ∇f ( xn )
Choose a decreasing
End while
Do some random restarts
Return the lowest couple xn , f ( xn ) found.
Figure 1. Pseudo Code for the Algorithm of the Gradient Descent
As the purpose of the problem is to find the optimal answer, if we move in the same direction as
gradient, we will achieve the answer much sooner [19]. The method of gradient is to move
towards the optimized answer in a zigzag direction instead of a straight direction.
An appropriate reaction is to start the movement from the current answer in a distinct direction
and not to stop as the f(x) is decreasing. When f(x) stops decreasing we stop too and the next
answer is ready there. So, the gradient is calculated again to specify the direction of the next
movement [19]. In this reaction each repeat contains the changing of the current answer of the
X. The repeats of the gradient will continue until the gradient equals zero, if the distance of ε is
considered acceptable. According to equation (3) we stopped [16].
∂f
≤ε
∂x n
n = 1,2,..., n
(3)
ε is the distance of the gradient. It can be assumed either short or long. If the distance is short,
the convergence rate is low. If this distance is small, the convergence rate is low. Gradient with
increasing distance until convergence is guaranteed, the convergence rate increases.
4. ANT COLONY OPTIMIZATION (ACO)
ACO algorithm, suggested by Dorigo in 1996 [20, 21], is a branch of artificial intelligence
which is known as collective intelligence group. The movement probability of the ant K from
the city I to the city j in the time t is suggested based on the equation (4). In this statement ηi,j is
3
International Journal of Managing Public Sector Information and Communication Technologies (IJMPICT)
Vol. 3, No. 2, December 2012
the purview and equals 1/di,j(the more nearby cities are more likely to be chosen and τi,j is the
amount of pheromone on the mane in the time t.
Allowed k is the group of cities that the ant hasn't met up until now but can possibly meet them
at the next step. α & β are the parameters of the impact of the pheromone laid on the main and
the distance impact, respectively.
[τ ij (t )]α ⋅ [η ij ]β
if k ∈ allowed k
α
β
pijk (t ) = ∑ [τ ij (t )] ⋅ [η ij ]
j∈allowed k
0
otherwise
(4)
The update rule of the laid pheromone on the mane is formed based on the equation (5).
τ ij (t + n ) = (1 − ρ ) × τ ij (t ) + ∆τ ij
(5)
In the equation (5), 1-ρ specifies the evaporation rate of the pheromone between the time t and (t
+ n). ρ Is assumed as the parameter of the pheromone evaporation in the range that: (0<ρ <1).
The much the amount of the ρ is, the much the speed of evaporation increases.
m
∆τ ij = ∑ ∆τ ijk
(6)
k =1
∆τ ijk Represents the amount of pheromone of the best tour and it is calculated according to the
equation (7).
Q Lk if k ant uses edge (i, j) at time (t, t + n)
∆τ ijk =
otherwise
0
(7)
In the equation (7), Q is a constant parameter which is dependent on the problem and Lk is the
length of the path which is travelled by the ant K. Thus, the initial level of pheromone and
evaporation rate, an important influence on the optimization process and is required to be
determined in an appropriate way.
5. PROPOSAL ALGORITHM
A large number of optimization methods are called collective optimization methods. In these
methods algorithm starts finding solutions with some numbers of population. Then the
algorithm is repeated among variables until there appears a proper answer. This process will go
on until there appears an optimized answer for the problem. There are a lot's of ways in order to
optimize the objective function such as numerical methods or random methods. Normally, the
computational methods jump over the objective function and move toward the lower slope.
Combinatorial problems are often longer and cannot be solved using computational v. the
answer that is achieved on thee algorithms may be very far from the optimal answer.
In this proposed method, we use a combined method (which includes both ACO algorithm and
gradient descent method) in order to optimize The DTSP. In this way, we have the high speed of
gradient descent methods and also are able to search for the overall minima. The ACO
algorithm like other collective intelligence algorithms has two major disadvantages. The first
disadvantage is the premature convergence. Other disadvantages of the ACO algorithm are the
impact of parameters of this algorithm on the problems. If the parameters are not adjusted
properly, the ACO algorithm hades towards local optimization and the algorithm performance
decreases. In order to avoid disadvantages of premature convergence and the dependence of
algorithm parameters on the problems, we have applied some changes in the ACO algorithm.
By these changes, we try to solve DTSP’s by increasing the variety of solutions on how to
4
International Journal of Managing Public Sector Information and Communication Technologies (IJMPICT)
Vol. 3, No. 2, December 2012
update the pheromone. Math is used as a tool for finding optimization of many algorithms. Grad
search method is one of the mathematical optimization techniques. The main idea of this
method is that the optimization problem (here is finding the optimal path) begins by choosing
any primary answer. Then all adjacent neighbours with the primary point are evaluated by
objective function. Then between answers to neighbours and the primary point, the answer is
the most acceptable is replaced to selected primary point and the algorithm is repeated until a
point is found that no more than your neighbor is not acceptable algorithm stops. Fig. 2 shows
the schema of the proposed algorithm.
Figure 2. Flowchart of the Proposed Algorithm.
Thus, by taking the descending gradient movement, ACO have been modified in order to reach
the optimal solution can be achieved more quickly. For one problem, there are several possible
answers that to compare and choose the best optimal solution, a function called the objective
function is defined. Select this function is dependent on the nature of the problem. In the DTSP,
path length is a common goal of optimization. Pseudo code of proposed algorithm based on
optimization algorithm of ACO algorithm and descending gradient includes the Fig. 3:
1. Initialize parameters
2. Loop
Each ant is positioned on a starting city
Repeat
State transition rule
Local pheromone updating
3. Gradient descent
Start with a point random
Repeat
Determine a descent direction
Choose a step
Update point
Until stopping criterion is satisfied
Until all ants have built a complete solution
Global pheromone updating
Until End condition is reached.
Figure 3. Pseudo Code of Proposed Algorithm
5
International Journal of Managing Public Sector Information and Communication Technologies (IJMPICT)
Vol. 3, No. 2, December 2012
To obtain the more optimal routes, descending grad algorithm is combined with the
pheromone’s updating law. Pheromone’s updating law when a more optimal answer achieved
for algorithm compared to previous repetitions, increased in optimized paths of ant pheromone.
When a better answer than previous repetitions is obtained it is possible that in the
neighbourhood of this answer, exist another best answer that can be find out the answer with
further searches in the neighbourhood of current answer.
As shown in Fig. 4, the algorithm works based on removing edge of the tour and reconnect it in
another way. Assuming the city, is the third city, ants meet all paths connected to the third city
according to descending gradient. Whatever length of the path is less, the steep of grad is more
in that path, and the path is chosen. It should be noted that there are several ways to make the
tour again but only the condition will be accepted that in the constraints of problem (see each
city only once) applies and new tour obtained better amount than previous answer. Remove of
edge and reconnect it in succession will continue as far as to not found any new improver
movement for algorithm.
Figure 4. How to Select Cities
According to what can be seen from figure this six samples city have been obtained as a primary
tour. This output of the desirable or undesirable to reduce the path length using the gradient is
descending for the next tour.
To increase the efficiency of the suggested algorithm mode of pheromone updating can be
combined with equation (8).
xn+1 = xn − t n ∇f ( xn )
(8)
By putting the amount of descending gradient in the right side of the equation (3) the following
equation is obtained.
τ ij (t + n ) = (1 − ρ )× τ ij (t ) + ∆τ ij + xn
(10)
In the above equation for select the best paths, the upper limit of pheromone is obtained.
A meta-innovative algorithm cannot reach the optimum quality of the final answer, if the
parameters are not set correctly. In order to solve this problem, to compute the overall optimum,
we should distinguish the problem-dependent parameters to algorithms is converge faster to the
optimal answer. In this regard, we use descending grad for set pheromone In ACO algorithm.
When we are faced with a large number of independent variables, determination the appropriate
combination of these parameters to presence in problem and in order to estimate problem
dependent variable, is importance.
Descending gradient is a firm computational tool for selecting best combination variables of
problem, that can done by ascending and descending selection approach. In descending method,
the amount that has the greatest impact, is added to the model. In fact in descending gradient
6
International Journal of Managing Public Sector Information and Communication Technologies (IJMPICT)
Vol. 3, No. 2, December 2012
algorithm the value that has the greatest impact is added to subject and value is less effective
than the model is removed.
6. EVALUATION AND RESULTS
To evaluate the proposed algorithm with other algorithms, the number of cities considered to be
30 and results is estimated for 100 times iterations of the algorithm. One of the important
parameters in the design of hybrid algorithm optimization and the resulting is path length.
Descending gradient algorithm performance is dependent on the initial point and it does not
have the ability to search all areas of problem. Thus by using ACO, that searches large space,
descending grad algorithm is capable to approach the original optimal point. Parameter amounts
for the proposed algorithm results are shown in Table 1.
In the proposed algorithm, there are several parameters that affect the performance, in the Table
1, by using the parameter (α) the amount of spilled pheromone on edge applied. Parameter β
determine the importance of relative distance to pheromone, in the selection of the next city.
Parameter (rho) weakened (evaporation) of pheromone, and the parameter (t) minimizing the
function in direction of descending gradient.
Table 1. Value of Parameters in the Proposed Algorithm
Parameter Name
Value
α
1
β
5
rho
0.1
t
0.4
Experimental results show that the use of descending gradient to select path, makes a significant
improvement in the efficiency and the convergence rate of the algorithm and cause the close
answers to optimal.
In the proposed algorithm for increase variety in the answer, after ants converging to a path, we
second initialized that. This means that the value of each ant is equal to reduction of path length
that ant has wended in the previous repetition and the best length achieved by other ants.
Table 2, shows comparing proposed method with other algorithms. As the results show the
proposed algorithm is able to achieve a better response than the other algorithms.
Table 2. Results obtained of the proposed algorithm and other algorithm
Algorithms
ACO [2]
GA [2]
Hybrid
Algorithm [2]
Proposed
Algorithm
Worst
Average
Best
Solution Solution Solution
385
340
368
464
349
826
384
340
358
365
340
347
Diagram Fig. 5, comparing the proposed algorithm with other algorithms shows. As Fig. 5
shows the performance of proposed algorithm is better.
7
International Journal of Managing Public Sector Information and Communication Technologies (IJMPICT)
Vol. 3, No. 2, December 2012
Figure 5. compared performance of the proposed algorithm with other algorithms
In Fig. 6, the impact of parameter (t) with 10 times performances in the proposed algorithm
process is shown. According to Fig. 6, it is better for be optimized paths, reduce the amount of
(t). With increase the amount of (t), the convergence rate has slowed down.
Figure 6. The impact of parameter (t)
Increasing the amount of (t) reduce the steep of descending grad. Because the possibility of
production new parts increases and searching becomes more extensive. So, reducing the amount
of (t) intensified the optimization process.
7. CONCLUSIONS
In this paper, we have described a new technique for optimizing the DTSP. In problem solving
of DTSP meta-innovative methods are used mainly. In this paper a combination method from
analytical techniques, using descending grad and collective intelligence methods for this
problem has been proposed. Algorithm Compound of ACO algorithm and descending grad
algorithm causes when caught in a local minimum, with initialized pheromone as descending
grad to algorithm we cause to escape from a local minimum. The results also show that answer
of the proposed method is better than the quite meta-innovative methods.
8
International Journal of Managing Public Sector Information and Communication Technologies (IJMPICT)
Vol. 3, No. 2, December 2012
REFERENCES
[1] H. N. Psaraftis,”Dynamic vehicle routing problems”. In: Golden, B.L., Assad, A.A. (eds.) Vehicle
Routing: Methods and Studies, Elsevier, Amsterdam, pp. 223–248, 1988.
[2] F.S. Gharehchopogh, I.Maleki, And M.Farahmandian, "New Approach for Solving Dynamic
[3]
[4]
[5]
[6]
[7]
[8]
[9]
[10]
[11]
[12]
[13]
[14]
[15]
[16]
[17]
[18]
[19]
[20]
[21]
Traveling Salesman Problem with Hybrid Genetic Algorithms and Ant Colony Optimization",
International Journal of Computer Applications (IJCA), vol:53, No.1, pp. 39-44, September 2012.
M.Guntsch and M.Middendorf. “Pheromone Modification Strategies for Ant Algorithms Applied to
Dynamic TSP”. In Applications of Evolutionary Computing, Lecture Notes in Computer Science,
Vol.2037, pp. 213-220, 2001.
M. Guntsch, M. Middendorf and H. Schemck. “An Ant Colony Optimization Approach to Dynamic
TSP”. Proceedings of the GECCO 2001. San Francisco, USA, 7-11 July, 2001, Morgan Kaufmann,
pp.860-867, 2001.
C. J. Eyckelhof, M. Snoek, “Ant systems for a dynamic TSP”. In: Proceedings of the 3rd
international workshop on ant algorithms, pp. 88–99, 2002.
Z.C.Huang, X.L.Hu, S.D.Chen, “Dynamic Traveling Salesman Problem based on Evolutionary
Computation”, In Congress on Evolutionary Computation, IEEE Press, pp. 1283-1288, 2001.
L. Kang, A. Zhou, B. McKay, Y. Li, Z. Kang, “Benchmarking Algorithms for Dynamic Travelling
Salesman Problems”. In: Proceedings of the Congress on Evolutionary Computation, Portland,
Oregon ,2004.
M. Dorigo, V. Maniezzo, and A. Colorni, “The ant system: Optimization by a colony of cooperating
agents”, IEEE Transactions on System, Man, and Cybernetics, Part B, Vol. 26, pp. 29-41, 1996.
G. V. Deklaits, A. Ravindran, K.M. Rogesdell, “Engineering optimization Methods and
Application”, a Wiley-Inter science Publication, 1983.
A.V.Fiacco, G.P.McCormick, “Nonlinear programming: Sequential Unconstrained Minimization
Techniques”, SIAM, Philadelphia, 1990.
G.P.McCormick, “Nonlinear Programming- Theory, Algorithms and Applications”, Wiley, New
York, 1983.
E. Polak: “Computational Methods in Optimization”, Academic Press, New York, 1971.
R.Battiti, “First and second order methods for learning: Between steepest descent and Newton's
method,” Neural Computation, Vol. 4, No. 2, pp. 141-166, 1992.
X. Wang. “Method of Steepest Descent and its Applications”, In IEEE Microwave and Wireless
Components Letters, ISSN: 15311309, Vol. 12, pp. 24-26, 2008.
Y. Kumar, D. Kumar, "Parametric Analysis of Nature Inspired Optimization Techniques",
International Journal of Computer Applications (IJCA), Vol. 32, No.3, pp. 42-49, October 2011.
F. S. Hillier, G. J. Lieberman: “Introduction to operations research”, McGraw-Hill, 2001.
A. P. Engelbrecht: “Computational intelligence”, 2nd ed, Wiley, 2007.
H. B. Curry, “The method of steepest descent for non-linear minimization problems”. Quart. Appl.
Math., vol.2 ,pp.258-261.1944.
E.K.P. Chong, S.H. Zak: “An Introduction to Optimization”, Wiley, New York, 1996.
M. Dorigo, V. Maniezzo, and A. Colorni, “The ant system: Optimization by a colony of cooperating
agents”, IEEE Transactions on System, Man, and Cybernetics, Part B, Vol. 26, pp. 29-41, 1996.
M. Dorigo and L. M. Gambardella, “The colony system: A cooperative learning approach to the
traveling salesman problem”, IEEE Transactions on Evolutionary Computation, Vol. 1, No.1,
April, 1997.
9
| 9cs.NE
|
AN ANALYSIS OF GENE EXPRESSION DATA USING PENALIZED
FUZZY C-MEANS APPROACH
P.K. Nizar Banu1 H. Hannah Inbarani2
1
Department of Computer Applications, B. S. Abdur Rahman University, Chennai, Tamilnadu,
India
Email: [email protected]
2
Department of Computer Science, Periyar University, Salem, Tamilnadu, India
Email: [email protected]
ABSTRACT
With the rapid advances of microarray technologies, large amounts of high-dimensional gene
expression data are being generated, which poses significant computational challenges. A first
step towards addressing this challenge is the use of clustering techniques, which is essential in
the data mining process to reveal natural structures and identify interesting patterns in the
underlying data. A robust gene expression clustering approach to minimize undesirable
clustering is proposed. In this paper, Penalized Fuzzy C-Means (PFCM) Clustering algorithm is
described and compared with the most representative off-line clustering techniques:
K-Means Clustering, Rough K-Means Clustering and Fuzzy C-Means clustering. These
techniques are implemented and tested for a Brain Tumor gene expression Dataset. Analysis of
the performance of the proposed approach is presented through qualitative validation
experiments. From experimental results, it can be observed that Penalized Fuzzy C-Means
algorithm shows a much higher usability than the other projected clustering algorithms used in
our comparison study. Significant and promising clustering results are presented using Brain
Tumor Gene expression dataset. Thus patterns seen in genome-wide expression experiments can
be interpreted as indications of the status of cellular processes. In these clustering results, we find
that Penalized Fuzzy C-Means algorithm provides useful information as an aid to diagnosis in oncology.
Keywords: Clustering, Microarray, Gene Expression, Brain Tumor, Fuzzy Clustering, FCM,
PFCM
produced. Searching for meaningful
information patterns and dependencies
among genes, in order to provide a basis for
hypothesis testing, typically includes the
initial step of a natural basis for organizing
gene expression data to group genes together
with similar patterns of expression. The field
of gene expression data analysis has grown
in the past few years from being purely datacentric
to
integrative,
aiming
at
complementing microarray analysis with
data and knowledge from diverse available
sources.
Advances
in
microarray
technologies have made it possible to
measure the expression profiles of thousands
1.INTRODUCTION
Gene expression is the fundamental
link between genotype and phenotype in a
species, with microarray technologies
facilitating the measurement of thousands of
Gene expression values under tightly
controlled conditions, e.g. (i) from a
particular point in the cell cycle, (ii) after an
interval response to some environmental
change, (iii) from RNA, isolated from a
tissue
exhibiting
certain
phenotypic
characteristics and so on (Kerr, et. al., 2008).
A problem inherent in the use of microarray
technologies is the huge amount of data
1
of genes in parallel under varying
experimental conditions. Due to the large
number of genes and complex gene
regulation networks, clustering is a useful
exploratory technique for analyzing these
data. It divides data of interest into a small
number of relatively homogeneous groups
or clusters. Clustering is a popular data
mining technique for various applications.
One of the reasons for its popularity is the
ability to work on datasets with minimum or
on a priori knowledge. This makes
clustering practical for real world
applications. We can view the expression
levels of different genes as attributes of the
samples, or the samples as the attributes of
different genes. Clustering can be performed
on genes or samples (Michael et. al. 1998).
This paper introduces the application of
Penalized Fuzzy C-Means algorithm to
cluster Brain Tumor genes.
expression
profiles
are
hierarchical
clustering algorithms (Michael et. al. 1998),
self-organizing maps (Paul, 1999) and
K-Means clustering algorithms (Tavazoie,
et. al., 1999). Hierarchical algorithms merge
genes with the most similar expression
profiles iteratively in a bottom-up manner.
Self-organizing
maps
and
K-means
algorithms partition genes into userspecified k optimal clusters. Other full space
clustering algorithms applied on gene
expression data include Bayesian network
(Friedman et. al. 2000) and neural network.
A robust image segmentation method that
combines the watershed segmentation and
penalized fuzzy Hopfield neural network
algorithms to minimize over-segmentation is
described in (Kuo et. al. 2006). (Brehelin et.
al., 2008) evaluates the stability of clusters
derived from hierarchical clustering by
taking repeated measurements. A wide
variety of clustering algorithms are available
for clustering gene expression data (Bezdek,
1981). They are mainly classified as
Partitioning methods, hierarchical methods,
Density based methods, Model based
methods, Graph Theoretic methods, soft
computing methods etc.
This paper is organized as follows.
In Section 2, Research background for
clustering Gene Expression patterns is
discussed. In Section 3, Methodology for
preparing Gene Expression patterns is
presented. Section 4 presents a method for
extracting highly suppressed and expressed
genes based on Penalized Fuzzy C-Means
algorithm. The Experimental results are
discussed in Section 5. Finally, section 6
concludes this paper by enumerating the
merits of the proposed approaches.
Multiple expression measurements
are commonly recorded as a real-valued
matrix, with row objects corresponding to
gene expression measurements over a
number of experiments and columns
corresponding to the pattern of expression of
all genes for a given microarray experiment.
Each entry xij, is the measured expression of
gene i in experiment j. Dimensionality of a
gene refers to the number of expression
values recorded for it. A gene/gene-profile x
is a single data item (row) consisting of d
measurements, x = (x1, x2, . . . , xd). An
experiment/sample y is a single microarray
experiment corresponding to a single
column in the gene expression matrix,
y = (x1, x2, …, xn)T where n is the number of
2. RESEARCH BACKGROUND
2.1. Clustering
Clustering genes, groups similar
genes into the same cluster based on a
proximity measure. Genes in the same
cluster have similar expression patterns. One
of the characteristics of gene expression data
is that it is meaningful to cluster both genes
and samples. The most commonly applied
full space clustering algorithms on gene
2
genes in the dataset. Clustering is considered
an interesting approach for finding
similarities in data and putting similar data
into groups. Initial step in the analysis of
gene expression data is the detection of
groups of genes that exhibit similar
expression patterns. In gene expression,
elements are usually genes and the vector of
each gene is its expression pattern. Patterns
that are similar are allocated in the same
cluster, while the patterns that differ
significantly are put in different clusters.
Gene expression data are usually of high
dimensions and relatively small samples,
which results in the main difficulty for the
application of clustering algorithms.
Clustering the microarray matrix can be
achieved in two ways:
(i) Genes can form a group which
show similar expression across
conditions,
(ii) Samples can form a group which
shows similar gene expression
across all genes.
This gives rise to global clustering, where a
gene or sample is grouped across all
dimensions. Additionally, the clustering can
be complete or partial. A complete
clustering assigns each gene to a cluster,
whereas a partial clustering does not. Partial
clustering tends to be more suited to gene
expression, as the dataset often contains
irrelevant genes or samples. Clearly this
allows:
(i) Noisy genes to be left out, with
correspondingly less impact on the
outcome and
(ii) Genes belonging to no cluster omitting a large number of
irrelevant contributions.
Microarrays measure expression for the
entire genome in one experiment, but genes
may change expression, independent of the
experimental condition. Forced inclusion in
well-defined but inappropriate groups may
impact the final structures found for the
data. Partial clustering avoids the situation
where an interesting sub-group in a cluster is
hidden through forcing membership of
unrelated genes (Kerr, et. al., 2008).
2.2. Categories of Gene Expression data
clustering
Methods of clustering can be
categorized as Hard Clustering or Soft
Clustering. Hard Clustering requires each
gene to belong to a single cluster, whereas
Soft
Clustering
permit
genes
to
simultaneously be members of numerous
clusters. Hard Clustering tells whether a
gene belongs to a cluster or not. Whereas in
Soft Clustering, with membership values,
every gene belongs to each cluster with a
membership weight between 0 (doesn’t
belong) and 1 (belongs). Clustering
algorithms which permit genes to belong to
more than one cluster are more applicable to
Gene expression. Gene expression data has
certain special characteristics and is a
challenging research problem
A modern working definition of a
gene is "a locatable region of genomic
sequence, corresponding to a unit of
inheritance, which is associated with
regulatory regions, transcribed regions, and
or other functional sequence regions".
Currently, a typical microarray experiment
contains 103 to 104 genes, and this number is
expected to reach to the order of 106.
However, the number of samples involved
in a microarray experiment is generally less
than 100. One of the characteristics of gene
expression data is that it is meaningful to
cluster both genes and samples. On one
hand, co-expressed genes can be grouped
into clusters based on their expression
patterns (Ben-Dor, et. al., 1999&Michael et.
al. 1998). In such gene-based clustering, the
genes are treated as the objects, while the
samples are the features. On the other hand,
the samples can be partitioned into
homogeneous groups. Each group may
3
correspond to some particular macroscopic
phenotype, such as clinical syndromes or
cancer types (Golub et. al., 1999). Such
sample-based clustering considers the
samples as the objects and the genes as the
features. The distinction of gene-based
clustering and sample-based clustering is
based on different characteristics of
clustering tasks for gene expression data.
Some clustering algorithms, such as KMeans and hierarchical approaches, can be
used both to group genes and to partition
samples.
Hard clustering algorithms like
K-Means and k-medoid place a restriction
that a data object can belong precisely to
only one cluster during clustering process.
This can be too restrictive while clustering
high dimensional data like Gene Expression
data because genes have a property of
getting expressed in multiple conditions.
Fuzzy set clustering like Fuzzy C-Means,
Penalized Fuzzy C-Means allows data
objects to belong to multiple clusters based
on the degree of membership.
to remember that when a gene expression
profile is analyzed in a given sample, it is
just a snapshot in time and space.
2.4 Clustering Techniques
K-Means Algorithm
The K-means method aims to minimize
the sum of squared distances between all
points and the cluster centre. This procedure
includes the steps, as described by Tou and
Gonzalez (Tou et. al. 1974).
Rough Set Theory
Rough set theory introduced by
Pawlak (Pawlak, 1982) deals with
uncertainty and vagueness. Rough set theory
became popular among scientists around the
world due to its fundamental importance in
the field of artificial intelligence and
cognitive sciences. Similar to fuzzy set
theory it is not an alternative to classical set
theory but it is embedded in it. Rough set
theory can be viewed as a specific
implementation of Frege’s idea of
vagueness, i.e., imprecision in this approach
is expressed by a boundary region of a set,
and not by a partial membership, like in
fuzzy set.
2.3. Analysis of Gene Expression data
Gene expression data is usually
represented by a matrix, with rows
corresponding to genes, and columns
corresponding to conditions, experiments or
time points. The content of the matrix is the
expression levels of each gene under each
condition. Those levels may be absolute,
relative or otherwise normalized. Each
column contains the results obtained from a
single array in a particular condition, and is
called the profile of that condition. Each row
vector is the expression pattern of a
particular gene across all the conditions.
Analyzing gene expression data is a
process by which a gene's information is
converted into the structures and functions
of a cell. Thousands of different mRNAs are
present in a given cell; together they make
up the transcriptional profile. It is important
Rough Clustering
A rough cluster is defined in a
similar manner to a rough set that is with
lower and upper approximation. The lower
approximation of a rough cluster contains
genes that only belong to that cluster. The
upper approximation of a rough cluster
contains genes in the cluster which are also
members of other clusters (SushmitaMitra,
2004 & SushmitaMitra, 2006). To use the
theory of rough sets in clustering, the value
set (Va) need to be ordered. This allows a
measure of the distance between each object
to be defined. Distance is a form of
similarity, which is a relaxing of the strict
4
requirement of indiscernibility outlined in
canonical rough sets theory, and allows the
inclusion of genes that are similar rather
than identical. Clusters of genes are then
formed on the basis of their distance from
each other. An important distinction
between rough clustering and traditional
clustering approaches is that, with rough
clustering, an object can belong to more than
one cluster.
3. METHODOLOGY
Cluster analysis, is an important tool in gene
expression
data
analysis.
For
experimentation, we used a set of gene
expression data that contains a series of gene
expression measurements of the transcript
(mRNA) levels of brain tumor gene. In
clustering gene expression data, the genes
are treated as objects and the samples are
treated as attributes. Gene pattern extraction
consists of 4 steps.
i)
Data Preparation
ii)
Data Normalization
iii)
Clustering
iv)
Pattern analysis
Fuzzy Clustering
Cluster analysis is a method of
grouping data with similar characteristics
into larger units of analysis. First in (Zadeh,
1965) fuzzy set theory that gave rise to the
concept of partial membership, based on a
membership function, fuzziness was
articulated and has received increasing
attention. Fuzzy clustering which produces
overlapping cluster partitions has been
widely studied and applied in various areas.
In fuzzy clustering, the Fuzzy C-Means
(FCM) clustering algorithm is the best
known and most powerful methods used in
cluster analysis (Bezdek, 1981). In (Yu et.
al., 2007), a general theoretical method to
evaluate the performance of fuzzy clustering
algorithm is proposed. The Fuzzy integrated
model is accurate than rough integrated
model and conventional integrated model
(Banu et. al., 2011). Fuzzy clustering
approach captures the uncertainty that
prevails in gene expression and becomes
more suitable for tumor prediction. One of
the important parameters in the FCM is the
weighting exponent m. When m is close to
one, the FCM approaches the hard C-Means
algorithm. When m approaches infinity, the
only solution of the FCM will be the mass
center of the data set. Therefore, the
weighting exponent m plays an important
role in the FCM algorithm.
3.1. Data Preparation
We represent the gene expression
ˆ {m |i 1,2,...,n , j 1,2,...,n},
data as ng by ns matrix: M
i, j
g
s
There are ns columns, one for each sample
and ng rows, one for each gene. One row of
genes is also called a gene vector, denoted
. Thus a gene
as g i m , m ,..., m
vector contains the values of a particular
attribute for all samples.
i ,1
i,2
i ,n
s
3.2. Data Normalization
Data sometimes need to be
transformed before being used. For example;
attributes may be measured using different
scales, such as centimeters and kilograms. In
instances where the range of values differ
widely from attribute to attribute, these
differing attribute scales can dominate the
results of the cluster analysis. It is therefore
common to normalize the data so that all
attributes are on the same scale. The
following are two common approaches for
data normalization of each gene vector:
m i, j
m i, j m i
m i, j
5
,
mi
m i, j m i
i
or
,
where
mi
ns
j 1
ns
m i, j
m
ns
, i
j 1
The membership weighting system reduces
noise effects, as a low membership grade is
less important in centroid calculation.
PFCM algorithm helps in identifying
hidden pattern and providing enhanced
understanding of the functional genomics in
a better way.
mi
2
i, j
ns 1
m i, j denotes the normalized value
and
for gene vector i of sample j, mi,j represents
the original value for gene i of sample j, ns is
the number of samples, m i is the mean of
the values for gene vector i over all samples,
and i is the standard deviation of the ith
gene vector.
Fuzzy Clustering permit genes to
belong to more than one cluster, is more
applicable to Gene Expression. Noisy genes
are unlikely to be members of several
clusters and genes with similar change in
expression for a set of samples are involved
in several biological functions and groups
should not be co-active under all conditions.
This gives rise to high inconsistency in the
gene groups and some overlap between
them. The boundary of a cluster is usually
fuzzy for three reasons:
The Brain Tumor gene expression
data is used for our experiments. This data
set is publically available in Broad Institute
web site. The various cluster validation
techniques namely Root Mean Square Error
(RMSE), Mean Absolute Error (MAE) and
Xie-Beni (XB) validity index are used to
validate the clusters obtained after applying
the clustering algorithms.
i.
4. PROPOSED PPROACH:
PENALIZED FUZZY C-MEANS
ii.
Penalized Fuzzy C-Means (PFCM)
algorithm for clustering gene expression
data is introduced in this paper, which
modified Fuzzy C-Means (FCM) algorithm
to produce more meaningful fuzzy clusters.
Genes are assigned a membership degree to
a cluster indicating its percentage
association with that cluster. The two
algorithms differ in the weighting scheme
used for the contribution of a gene to the
mean of the cluster. FCM membership
values for a gene are divided among clusters
in proportion to similarity with that clusters
mean. The contribution of each gene to the
mean of a cluster is weighted, based on its
membership grade. Membership values are
adjusted iteratively until the variance of the
system falls below a threshold. These
calculations require the specification of a
degree of fuzziness parameter which is
problem specific (Dembele et. al., 2003).
iii.
The gene expression dataset might
be noisy and incomplete
The similarity measurement between
genes is continuous and there is no
clear cutoff value for group
membership
A gene might behave similarly to
gene1 under a set of samples and
behave similarly to another gene2
under another set of samples.
Therefore, there is a great need for a
fuzzy clustering method, which
produces clusters in which genes can
belong to a cluster partially and to
multiple clusters at the same time
with different membership degrees.
The main objective of using this method is
to minimize the objective function value so
that the highly suppressed genes and highly
expressed genes are clustered separately and
also it helps to diagnose at an early stage of
tumor formation.
6
Based on the numerical results PFCM is
more accurate than FCM.
The steps of the PFCM algorithm are given
as follows:
Step 1: Initialize the cluster centroids
wj (2 j c),v(v 0),
fuzzification parameter,
m (1 m ),
and the value 0. Gives a fuzzy
C-partition ( 0 ) and t=1.
4.1. Penalized Fuzzy C-Means Algorithm
Another
strategy
for
fuzzy
clustering, called the penalized Fuzzy CMeans (PFCM) algorithm, with the addition
of a penalty term was proposed by Yang
(Yang, 1993 &Yang, 1994). Yang made the
fuzzy extension of the Classification
Maximum Likelihood (CML) procedure in
conjunction with fuzzy C-partitions and
called it a class of fuzzy CML procedures.
The idea of penalization is also important in
statistical learning. Combining the CML
procedure and penalty idea, Yang (Yang,
1993) added a penalty term to the FCM
objective function JFCM and then extended
the FCM to the so-called Penalized FCM
(PFCM) which produces more meaningful
and effective results than the FCM
algorithm. Thus, the PFCM objective
function is defined as follows:
Step 2: Calculate (j t ) , w(jt ) with ( t 1) using
Eqs. (1) and (2).
Step 3: Calculate the membership matrix
( t ) [ui , j ]with (jt ) , w(jt )
using Eq. (3)
c n
1c n m
2 1
JPFCM ui, j || xi wj || vuim, j lnj
2 j1 i1
2 j1 i1
1 c n m
JFCM vui, j lnj
2 j1 i1
(
5. EXPERIMENTAL RESULTS
where j is a proportional constant of
class j and v (≥ 0) is a constant. The penalty
term 1 v u ln is added to the
c
n
j1
i1
m
i, j
2
The effectiveness of the algorithms
based on cluster validity measure is
demonstrated in this section.
j
objective function, when v=0, JPFCM is equal
to JFCM.
αj, wj and
ui,j are defined as
n
u im, j
j c i 1 n
, j 1, 2,..., c
m
ui, j
(1)
j 1 i 1
wj
n
1
n
u
u
m
i, j
xi
5.1. Data Source
A set of gene expression data that
contains a series of gene expression
measurements of the transcript (mRNA)
levels of brain Tumor gene is used in this
paper to analyze the efficiency of the
proposed approach.
The brain tumor dataset is taken
from
the
Broad-Institute
website
(http://www.broadinstitute.org/cgibin/cancer/datasets.cgi). The dataset is titled
as “Gene Expression-Based Classification
and Outcome Prediction of Central Nervous
(2)
m i 1
i, j
i 1
ui, j
Step 4: Compute max | (t 1) ( t ) | .
If , and , t t 1 go to Step 2;
otherwise go to Step 5.
(
Step 5: Find the results1for the final class
centroids.
)
c || x w || 2 v ln
i
j
j
l 1 || x w || 2 v ln
i
l
l
1 /( m 1)
1 /( m 1)
1
(3)
7
System Embryonal Tumors”. The Datasets
Consists of three types of Brain Tumors data
namely Medulloblastoma classic and
desmoplastic, Multiple Brain tumors,
Medulloblastoma treatment outcome. Each
dataset contains nearly 7000 genes with 42
samples.
In order to evaluate the proposed
algorithm, we applied it to the Brain Tumor
gene expression data taken from Broad
Institute by taking 7129, 5000, 3000, and
1000 genes for all samples with various
numbers of clusters.
above mentioned validity measures and the
effectiveness of the proposed algorithm is
well understood. We tested our method for
the Brain Tumour gene expression dataset to
cluster the highly suppressed and highly
expressed genes and are depicted for various
dataset sizes. It is observed that for each set
of genes taken, the value of validity
measures for the proposed algorithm is
lower than the value of validity measures for
other algorithms and it is graphically
illustrated from Figure 1 to Figure 12.
Among
these
clustering
algorithms
Penalized Fuzzy C-Means produces better
results in identification of differences
between data sets. This helps to correlate the
samples according to the level of gene
expression.
5.2. Comparative Analysis
A comparative study of the
performance of K-Means (Tou et. al. 1974),
Rough K-Means (Peters, 2006) and Fuzzy
C-Means (Peters, 2006) is made with
Penalized Fuzzy C-Means algorithm (Shen
et. al, 2006).
In terms of MAE, Penalized Fuzzy
C-Means shows superior performance and
K-Means and rough K-Means exhibits better
performance than Fuzzy C-Means.
Cluster Validation
In this paper, Root Mean Square
Error, Mean Absolute Error and Xie-Beni
validity index are used to validate the
clusters obtained after applying the
clustering algorithms. To assess the quality
of our method, we need an objective
external criterion. In order to validate our
clustering results, we employed Root Mean
Square Error (RMSE), Mean Absolute Error
(MAE) (Pablo de Castro et. al., 2007).
Xie-Beni validity index has also been
chosen as the cluster validity measure
because it has been shown to be able to
detect the correct number of clusters in
several experiments (Pal et. al. 1995).
Also, With respect to RMSE and
Xie-Beni Index Penalized Fuzzy C-Means
produces greater performance than the other
algorithms.
The comparative results based on
Root Mean Square Error, Mean Absolute
Error and Xie-Beni validity measure for all
the Gene expression data clustering
algorithms for various Brain Tumor data sets
taken with different number of clusters are
enumerated in Table 1.
Figure 1 to Figure 3 shows the
comparative analysis of various approaches
for 7129 genes taking K as 7, 5 and 3
respectively. It can be observed from the
figures,
Penalized
Fuzzy
C-Means
outperforms other approaches Fuzzy CMeans, Rough K-Means and K-Means.
K Means, Rough K-Means, Fuzzy C-Means
and Penalized Fuzzy C-Means clustering
algorithm are applied and analysed for Brain
Tumour genes.
Table.1 shows the experimental
results that are obtained by applying the
8
Data Set Size=7129
k=7
Value of Valididty Measures
Value of Valididty Measures
1.2
1
0.8
0.6
0.4
0.2
0
FCMeans
RKMeans
KMeans
MAE
RMSE
XB
PFCMeans
1.2
1
0.8
0.6
0.4
0.2
0
Value of Valididty Measures
Value of Valididty Measures
FCMeans
RKMeans
KMeans
XB
PFCMeans
1.2
1
0.8
0.6
0.4
0.2
0
FCMeans
0.4
RKMeans
Value of Valididty Measures
Value of Valididty Measures
0.6
KMeans
MAE
RMSE
XB
PFCMeans
FCMeans
RKMeans
KMeans
RMSE
XB
PFCMeans
Figure5: Validity Measure for Data set size = 5000, k=5
Data Set Size=7129
k=3
0.2
0
XB
Validity Measures
Figure 2: Validity Measure for Data set size =7129, k=5
0.8
RMSE
Data Set Size=5000
k=5
MAE
Validity Measures
1
KMeans
Figure4: Validity Measure for Data set size =5000, k=7
Data Set Size=7129
k=5
RMSE
RKMeans
Validity Measures
Figure 1: Validity Measure for Data set size = 7129, k=7
MAE
FCMeans
MAE
Validity Measures
1.2
1
0.8
0.6
0.4
0.2
0
Data Set Size=5000
k=7
PFCMeans
1
0.8
Data Set Size=5000
k=3
0.6
0.4
FCMeans
0.2
KMeans
RKMeans
0
MAE
RMSE
XB
PFCMeans
Validity Measures
Validity Measures
Figure 3: Validity Measure for Data set size = 7129, k=3
Figure6: Validity Measure for Data set size =5000, k=3
Figure 4 to Figure 6 shows the
comparative analysis of various approaches
for 5000 genes taking K as 7, 5 and 3
respectively. The experimental result shows
that the Penalized Fuzzy C-Means
outperforms other approaches Fuzzy CMeans, Rough K-Means and K-Means.
Figure 7 to Figure 9 shows the
comparative analysis of various approaches
for 3000 genes taking K as 7, 5 and 3
respectively. The experimental result shows
that the Penalized Fuzzy C-Means gives
better results than the other algorithms.
9
Data Set Size=3000
k=7
3
Value of Valididty
Measures
Value of Valididty Measures
4
FCMeans
2
RKMeans
1
KMeans
0
MAE
RMSE
XB
PFCMeans
3
2.5
2
1.5
1
0.5
0
Value of Valididty
Measures
Value of Valididty Measures
FCMeans
RKMeans
0.5
KMeans
0
MAE
RMSE
XB
PFCMeans
4
3
PFCMeans
Data Set Size=1000
k=5
FCMeans
RKMeans
1
KMeans
0
MAE
RMSE
XB
PFCMeans
Validity Measures
Figure8: Validity Measure for Data set size =3000, k=5
Figure 11: Validity Measure for Dataset Size =1000, k=5
Data Set Size=3000
k=3
Value of Valididty Measures
Value of Valididty
Measures
XB
2
Validity Measures
1.2
1
0.8
0.6
0.4
0.2
0
KMeans
Figure10: Validity Measure for Data set size =1000,
k=7
Data Set Size=3000
k=5
1
RKMeans
Validity Measures
Figure 7: Validity Measure for Data set size = 3000, k=7
1.5
FCMeans
MAE RMSE
Validity Measures
2
Data Set Size=1000
k=7
FCMeans
RKMeans
KMeans
PFCMeans
Validity Measures
1.5
Data Set Size=1000
k=3
1
FCMeans
RKMeans
0.5
KMeans
0
MAE
RMSE
XB
PFCMeans
Validity Measures
Figure12: Validity Measure for Data set size =1000, k=3
Figure 9: Validity Measure for Data set size = 3000, k=3
Figure 10 to Figure 12 shows the
comparative analysis of various approaches
for 1000 genes taking K as 7, 5 and 3
respectively. The experimental result shows
that the Penalized Fuzzy C-Means performs
better than the other algorithms
10
Table 1: Performance Analysis of K-Means, Rough K-Means, Fuzzy C-Means and Penalized Fuzzy C-Means
No. of
No. of
Clusters
Genes
7
5
3
7
7129
5000
3000
1000
Root Mean
Mean
Square
Absolute
Error
Error
K-Means
0.0019
0.0044
0.2804
Rough K-Means
0.0511
0.0874
0.0020
Fuzzy C-Means
1.0438
0.8627
0.9548
Penalized Fuzzy C-Means
0.0043
0.0009
0.0001
K-Means
0.0034
0.0074
0.4798
Rough K-Means
0.0624
0.4654
0.0581
Fuzzy C-Means
1.0336
0.7504
0.0184
Penalized Fuzzy C-Means
0.0255
0.0015
0.0002
K-Means
0.0065
0.0150
0.1240
Rough K-Means
0.0336
0.0082
0.0203
Fuzzy C-Means
0.9925
0.6331
0.0624
Penalized Fuzzy C-Means
0.0079
0.0002
0.0001
K-Means
0.0225
0.0470
0.4994
Rough K-Means
0.1204
1.0548
0.1079
Fuzzy C-Means
1.8542
2.7108
0.1107
Penalized Fuzzy C-Means
0.0056
0.0015
0.0001
Clustering Algorithms
Xie-Beni
Index
Table .2 Parameter setting and other issues
Clustering
Algorithm
Cluster
Membership
Input
Proximity
Measure
K-Means
Binary
Starting Prototype,
Stopping Threshold, K
Pair wise
Distance
Rough K-Means
Rough
Membership
Starting Prototype,
Stopping Threshold, K
Pair wise
Distance
Fuzzy C-Means
Fuzzy
Membership
Degree of fuzziness,
Starting Prototypes,
Stopping Threshold, K
Pair wise
Distance
Penalized Fuzzy
C-Means
Improved Fuzzy
Membership
Fuzzification Parameter,
K
Pair wise
Distance
11
Other Issues
Very Sensitive to Input
parameters and order of
Input
Imprecision in Gene
Expression data can be
captured
Careful interpretations
of membership values.
Sensitive to Input
parameters and order of
Input
Quality of membership
is increased by
introducing Penalty
term
Pattern Analysis
The parameter setting and other
issues of the clustering approaches which
are discussed are given in Table.2.
For each gene, red indicates a high
level of expression (highly expressed Genes)
relative to the mean; green indicates a low
level of expression (highly suppressed
Genes) relative to the mean.
Transcriptional Initiation is the most
important mode for control of gene
expression
level.
Suppressed
gene
expression level may be stimulated by gene
therapy (i.e.) promoter insertion or up
regulation of suppressed gene and radiation
therapy (i.e.) more amount of radiation may
cause sudden mutation in gene, due to this
sudden change the gene expression may be
in high level. This helps to correlate the
samples according to the level of gene
expression. When precise functions for over
or under expressed genes are determined,
new avenues for intervention strategies may
emerge. These studies are in their infancy;
however,
the
improved
technology
employed here shows reasonable promise as
our understanding of these deadly tumours
increases. Therefore, future treatment
decisions based on the expression profile of
a primary tumour is a rational approach
towards preventing the outgrowth of
metastases.
6. CONCLUSION
Cluster
analysis
applied
to
microarray measurements aims to highlight
meaningful patterns for gene expression.
The goal of gene clustering is to identify the
important gene markers. Gene expression
data are the representation of nonlinear
interactions among genes and environmental
factors. Brain Tumor is so deadly because,
it is not usually diagnosed until it has
reached an advance stage. Early detection
can help prolong or save lives, but clinicians
currently have no specific and sensitive
method. Computing analysis of these data is
expected to gain knowledge of gene
functions and disease mechanisms. We used
a set of gene expression data that contains a
series of gene expression measurements of
the transcript (mRNA) levels of Brain
Tumor gene. Highly expressed genes and
suppressed genes are identified and
clustered
using
various
clustering
techniques. The following methods such as
Post-Translational Modification, Small
RNAs and Control of Transcript Levels,
Translational Initiation, Transcript Stability,
RNA Transport, Transcript Processing and
Modification, Epigenetic Control and
Transcriptional Initiation can be used to
reduce the tumor level.
The empirical results also reveal the
importance of using Penalized Fuzzy CMeans (PFCM) clustering methods for
clustering
more
meaningful
highly
suppressed and highly expressed genes from
the gene expression data set. Meanwhile,
evaluation metrics Root Mean Square Error,
Figure 13: Clusters of Brain Tumor Genes
Figure.13 represents the Clusters of tumor
genes by Penalized Fuzzy C-Means using all
genes exhibiting variation across the data
set.
12
Mean Absolute Error and Xie-Beni Index is
adopted to assess the quality of clusters, and
the experimental results have shown that the
proposed approach is capable of effectively
discovering gene expression patterns and
revealing the underlying relationships
among genes as well. This clustering
approach can be applied for any gene
expression dataset and tumor growth can be
predicted easily.
[8]
[9]
REFERENCES
[1]
[2]
[3]
[4]
[5]
[6]
[7]
A. Ben-Dor, R. Shamir and Z.Yakhini,
“Clustering gene expression patterns”,
Journal of Computational Biology,
Vol. 6, No. (3/4), pp. 281–297, 1999.
J.C. Bezdek, “Fuzzy mathematics in
pattern
classification,
Ph.D.
Dissertation, Applied Mathematics”,
Cornell University, Ithac, New York,
1973.
J.C. Bezdek, “Pattern Recognition with
Fuzzy
Objective
Function
Algorithms”, Plenum Press, New York,
1981.
P. A.D.de Castro, O. F. de França H.
M. Ferreira and F. J. Von Zuben,
“Applying Biclustering to Perform
Collaborative
Filtering”,
In
Proceedings
of
the
Seventh
International Conference on Intelligent
Systems Design and Applications, pp.
421-426, 2007.
D. Dembele and P. Kastner. “Fuzzy cmeans
method
for
clustering
microarray
data”.
Bioinformatics
(Oxford, England), Vol. 19, No.8,
pp.973-980, May 2003.
J.C. Dunn, “A fuzzy relative of the
ISODATA process and its use in
detecting compact and well- separated
clusters”, Journal of Cybernetics, Vol.
3, pp. 32-57, 1974.
B. Eisen Michael, T. Spellman Paul,
O. Brown Patrick and Botstein David.
[10]
[11]
[12]
[13]
[14]
13
“Cluster analysis and display of
genome-wide expression patterns”.
Proc. Natl. Acad. Sci. USA, Vol. 95,
No. 25, pp. 14863–14868, Dec. 1998.
N. Friedman, M. Linial, I. Nachman,
and D. Peter. “Using Bayesian network
to analyze expression data”, Journal of
Computational Biology, Vol. 7, pp.
601-620, Aug. 2000.
T.R. Golub, D. K. Slonim, P. Tamayo,
C. Huard, M. Gassenbeek, J.P.
Mesirov, H. Coller, M. L. Loh, J.R.
Downing, M. A. Caligiuri, D. D.
Bloomfield,
and
E.S.
Lander,
“Molecular classification of cancer:
Class discovery and class prediction by
gene expression monitoring”. Science,
Vol. 286. No. 15, pp. 531–537,
October 1999.
Jian Yu and Miin-shen Yang, “A
Generalized
Fuzzy
Clustering
Regularization Model with optimality
Tests
and
Model
Complexity
Analysis”, IEEE Transactions on
Fuzzy Systems, Vol. 15, No. 5, pp.
904-915, 2007.
G. Kerr, H. J. Ruskin, M. Crane, P.
Doolan, “Techniques for Clustering
Gene Expression data”, Computers in
Biology and Medicine, Vol.36 No. 3,
pp 283-293, March2008.
Laurent Brehelin, Olivier Gascue and
Olivier Martin, “Using repeated
measurements to validate hierarchical
gene clusters”, Bioinformatics, Vol.
24, No. 5, pp. 682-688, 2008.
Miin-Shen Yang, Wen-Liang Hung
and Chia-Hsuan Chang, “A penalized
fuzzy clustering algorithm”, In.
Proceedings of the 6th WSEAS
International Conference on Applied
Computer Science, Spain, Dec. 2006.
N.R. Pal and J.C. Bezdek, “On cluster
validity for the fuzzy c-means model”,
IEEE Trans. Fuzzy Systems, Vol. 3, pp.
370-379, 1995.
[15] T. Paul, “Interpreting patterns of gene
expression with self-organizing maps:
Methods
and
application
to
hematopoietic differentiation”. In
Proc. Natl. Acad. Sci. USA, Vol. 96,
pp. 2907–2912, 1999.
[16] P. K. Nizar Banu, H. Hannah Inbarani,
“Analysis of Click Stream Patterns
using Soft Biclustering Approaches”,
International Journal of Information
Technologies and Systems Approach
(IJITSA), Vol. 4, No. 1, 2011.
[17] Z. Pawlak, “Rough sets”, International
Journal of Computer and Information
Sciences, Vol.2, pp. 341–356, 1982.
[18] G. Peters, “Some refinements of rough
kmeans
clustering”,
Pattern
Recognition, Vol. 39, pp. 1481–1491,
2006.
[19] SushmitaMitra, “An evolutionary
rough partitive clustering”, Pattern
Recognition Letters Vol. 25, pp.1439–
1449, 2004.
[20] SushmitaMitra,
“Rough
Fuzzy
Colloborative
Clustering”,
IEEE
transactions on Systems, Man and
Cybernetics, Vol. 36, No. 4, August
2006.
[21] S. Tavazoie, J. D. Hughes, M. J.
Campbell, R. J. Cho, and G. M.
Church. “Systematic determination of
genetic network architecture”. In
Nature Genetics, Vol. 22, pp. 281–
285, 1999.
[22] J. T. Tou and R. C. Gonzalez, Pattern
Recognition Principles, Massachusetts:
Addision-Wesley, 1974.
[23] Wen-Feng Kuo, Chi-Yuan Lin and
Yung-Nien Sun, “Region Similarity
Relationship between watershed and
Penalzed Fuzzy Hopfield Neural
Network Algorithms for Brain Image
Segmentation”, International Journal
of pattern Recognition and Artificial
Intelligence, Vol. 22. No. 7, pp: 14031425, 2008.
[24] M. S. Yang, “On a class of fuzzy
classification maximum likelihood
procedures”, Fuzzy Sets Syst. Vol. 57,
No. 3, pp.365-375, 1993.
[25] M. S. Yang and C. F. Su, “On
parameter estimation for normal
mixtures based on fuzzy clustering
algorithms”, Fuzzy Sets Syst. Vol. 68,
No. 1, pp. 13-28, 1994.
[26] L.A. Zadeh, “Fuzzy sets”, Inform.
Control, Vol.8, pp. 338-353, 1965.
14
| 5cs.CE
|
HOMOCLINICALLY EXPANSIVE ACTIONS AND A GARDEN OF
EDEN THEOREM FOR HARMONIC MODELS
arXiv:1803.03541v1 [math.DS] 8 Mar 2018
TULLIO CECCHERINI-SILBERSTEIN, MICHEL COORNAERT, AND HANFENG LI
Abstract. Let Γ be a countable Abelian group and f ∈ Z[Γ], where Z[Γ] denotes the
integral group ring of Γ. Consider the Pontryagin dual Xf of the cyclic Z[Γ]-module
Z[Γ]/Z[Γ]f and suppose that f is weakly expansive (e.g., f is invertible in ℓ1 (Γ), or,
when Γ is not virtually Z or Z2 , f is well-balanced) and that Xf is connected. We prove
that if τ : Xf → Xf is a Γ-equivariant continuous map, then τ is surjective if and only
if the restriction of τ to each Γ-homoclinicity class is injective. We also show that this
−1
equivalence remains valid in the case when Γ = Zd and f ∈ Z[Γ] = Z[u1 , u−1
1 , . . . , ud , ud ]
is an irreducible atoral polynomial such that its zero-set Z(f ) is contained in the image
of the intersection of [0, 1]d and a finite union of hyperplanes in Rd under the quotient
map Rd → Td (e.g., when d ≥ 2 such that Z(f ) is finite). These two results are analogues
of the classical Garden of Eden theorem of Moore and Myhill for cellular automata with
finite alphabet over Γ.
Contents
1. Introduction
2. Background material and preliminaries
2.1. Group actions
2.2. Convolution
2.3. Pontryagin duality
2.4. Algebraic dynamical systems
2.5. Finitely presented algebraic dynamical systems
2.6. The homoclinic group
2.7. Connectedness of the phase space
3. Weak forms of expansivity for algebraic actions
3.1. p-expansive algebraic actions and p-homoclinic groups
3.2. Homoclinically expansive actions
3.3. Principal algebraic actions associated with weakly expansive polynomials
3.4. Expansive principal algebraic actions
3.5. Harmonic models
4. Topological rigidity
2
6
6
6
8
8
9
10
11
12
12
13
17
19
19
21
Date: March 12, 2018.
2010 Mathematics Subject Classification. 37D20, 37C29, 54H20, 37A45, 22D32, 22D45.
Key words and phrases. Garden of Eden theorem, finitely presented algebraic action, principal algebraic
action, harmonic model, homoclinic group, Pontryagin duality, topological rigidity, Moore property, Myhill
property, expansive action, homoclinic expansive action, weakly expansive polynomial.
1
2
TULLIO CECCHERINI-SILBERSTEIN, MICHEL COORNAERT, AND HANFENG LI
4.1. Affine maps
4.2. Topological rigidity
5. Proof of Theorem 1.2
6. Atoral polynomials and proof of Theorem 1.5
7. Concluding remarks
7.1. Surjunctivity
7.2. Counterexamples for mixing algebraic dynamical systems
7.3. p-pre-injectivity and the p-Moore and p-Myhill properties
References
21
21
23
23
25
25
26
26
27
1. Introduction
Consider a dynamical system (X, α), consisting of a compact metrizable space X, called
the phase space, equipped with a continuous action α of a countable group Γ. Let d be
a metric on X that is compatible with the topology. Two points x, y ∈ X are said to be
homoclinic if limγ→∞ d(γx, γy) = 0, i.e., for every ε > 0, there exists a finite set F ⊂ Γ
such that d(γx, γy) < ε for all γ ∈ Γ \ F . Homoclinicity is an equivalence relation on X.
By compactness of X, this relation does not depend on the choice of the compatible metric
d. A map with source set X is called pre-injective (with respect to α) if its restriction to
each homoclinicity class is injective.
An endomorphism of the dynamical system (X, α) is a continuous map τ : X → X that
is Γ-equivariant (i.e., τ (γx) = γτ (x) for all γ ∈ Γ and x ∈ X).
The original Garden of Eden theorem is a statement in symbolic dynamics that characterizes surjective endomorphisms of shift systems with finite alphabet. To be more
specific, let us fix a compact metrizable space A, called the alphabet. Given a countable
group Γ, the shift over the group Γ with alphabet A is the dynamical system (AΓ , σ), where
AΓ = {x : Γ → A} is equipped with the product topology and σ is the action defined by
γx(γ ′ ) := x(γ −1 γ ′ ) for all x ∈ AΓ and γ, γ ′ ∈ Γ. The Garden of Eden theorem states
that, under the hypotheses that the group Γ is amenable and the alphabet A is finite, an
endomorphism of (AΓ , σ) is surjective if and only if it is pre-injective. It was first proved
for Γ = Zd by Moore and Myhill in the early 1960s. Actually, the implication surjective
=⇒ pre-injective was first proved by Moore in [27] while the converse implication was
established shortly after by Myhill in [29]. The Garden of Eden theorem was subsequently
extended to finitely generated groups of subexponential growth by Machı̀ and Mignosi [25]
and finally to all countable amenable groups by Machı̀, Scarabotti, and the first author
in [8].
Let us say that the dynamical system (X, α) has the Moore property if every surjective
endomorphism of (X, α) is pre-injective and that it has the Myhill property if every preinjective endomorphism of (X, α) is surjective. We say that the dynamical system (X, α)
has the Moore-Myhill property, or that it satisfies the Garden of Eden theorem, if it has
both the Moore and the Myhill properties.
GARDEN OF EDEN THEOREM FOR HARMONIC MODELS
3
The goal of the present paper is to establish a version of the Garden of Eden theorem
for principal algebraic dynamical systems associated with weakly expansive polynomials
over countable Abelian groups and with connected phase space. By an algebraic dynamical
system, we mean a dynamical system of the form (X, α), where X is a compact metrizable
Abelian group and α is an action of a countable group Γ on X by continuous group
automorphisms. Note that, in this case, the set ∆(X, α) = {x ∈ X : x is homoclinic
to 0X } ⊂ X, where 0X is the identity element of X, is a subgroup of X, called the
homoclinic group and two points x, y ∈ X are homoclinic if and only if x − y ∈ ∆(X, α),
that is, they belong to the same coset of ∆(X, α) in X. By Pontryagin duality, algebraic
dynamical systems with acting group Γ are in one-to-one correspondence with countable
left Z[Γ]-modules. Here Z[Γ] denotes the integral group ring of Γ. This correspondence has
been intensively studied in the last decades and revealed fascinating connections between
commutative algebra, number theory, harmonic analysis, ergodic theory, and dynamical
systems (see in particular the monograph [31] and the survey [20]).
Let f ∈ Z[Γ] and consider the cyclic left Z[Γ]-module Mf := Z[Γ]/Z[Γ]f obtained by
quotienting the ring Z[Γ] by the principal left ideal generated by f . The algebraic dynamical
system associated by Pontryagin duality with Mf is denoted by (Xf , αf ) and is called the
principal algebraic dynamical system associated with f .
We denote by C0 (Γ) the real vector space of all functions g : Γ → R vanishing at infinity
(i.e., for every ε > 0 there exists a finite subset Ω ⊂ Γ such that |gγ | ≤ ε for all γ ∈ Γ \ Ω).
Moreover, for f ∈ Z[Γ] and g ∈ C0 (Γ) we denote by f g ∈ C0 (Γ) their convolution product
(see Subsection 2.2).
Definition 1.1. A polynomial f ∈ Z[Γ] is said to be weakly expansive provided:
(we-1) ∀g ∈ C0 (Γ), f g = 0 ⇒ g = 0;
(we-2) ∃ω ∈ C0 (Γ) such that f ω = 1Γ .
Our first result is the following.
Theorem 1.2 (Garden of Eden theorem for algebraic actions associated with weakly
expansive polynomials). Let Γ be a countable Abelian group and f ∈ Z[Γ]. Suppose that f
is weakly expansive and that Xf is connected. Then the dynamical system (Xf , αf ) has the
Moore-Myhill property.
There are two main ingredients in our proof of Theorem 1.2. The first one, Corollary
4.4, is a rigidity result (a generalization of [1, Corollary 1]) for algebraic dynamical systems
associated with weakly expansive polynomials and with connected phase space. We use
it to prove that, under the above conditions, every endomorphism of (Xf , αf ) is affine
with linear part of the form x 7→ rx for some r ∈ Z[Γ]. The second one, Theorem 3.9, a
generalization of [19, Lemma 4.5]), asserts that, if (Xf , αf ) is weakly expansive, then its
homoclinic group ∆(Xf , αf ), equipped with the induced action of Γ, is dense in Xf and
isomorphic, as a Z[Γ]-module, to Z[Γ]/Z[Γ]f ∗ , where f ∗ ∈ Z[Γ] is defined by (f ∗ )γ := fγ −1
for all γ ∈ Γ.
Recall that a dynamical system (X, α) is called expansive if there exists a constant ε0 > 0
such that, for every pair of distinct points x, y ∈ X, there exists an element γ ∈ Γ such
4
TULLIO CECCHERINI-SILBERSTEIN, MICHEL COORNAERT, AND HANFENG LI
that d(γx, γy) > ε0 . Such a constant ε0 is called an expansivity constant for (X, α, d). The
fact that (X, α) is expansive or not does not depend on the choice of the metric d. For
instance, the shift system (AΓ , σ) is expansive for every countable group Γ whenever the
alphabet A is finite.
Let f ∈ Z[Γ] and suppose that the associated principal algebraic dynamical system
(Xf , αf ) is expansive. In Corollary 3.12 we show that f is weakly expansive, and from
Theorem 1.2 we thus deduce the following:
Corollary 1.3. Let Γ be a countable Abelian group and f ∈ Z[Γ]. Suppose that the
principal algebraic dynamical system (Xf , αf ) associated with f is expansive and that Xf
is connected. Then the dynamical system (Xf , αf ) has the Moore-Myhill property.
This result was obtained by the first two named authors in [7] and, shortly after, as
a particular case of a much more general Garden of Eden theorem for expansive actions
(where Γ is amenable and no connectedness of the phase space is assumed) proved by the
third author in [18].
Recall that a polynomial f ∈ R[Γ] is said to be well-balanced (cf. [4, Definition 1.2]) if:
P
(wb-1) γ∈Γ fγ = 0,
(wb-2) fγ ≤ 0 for all γ ∈ Γ \ {1Γ },
(wb-3) fγ = fγ −1 for all γ ∈ Γ (i.e., f is self-adjoint),
(wb-4) and supp(f ) := {γ ∈ Γ : fγ 6= 0}, the support of f , generates Γ.
If f ∈ Z[Γ] is well-balanced, the associated dynamical system (Xf , Γ) is called a harmonic
P
−1
−1
d
model. For Γ = Zd and f = 2d − di=1 (ui + u−1
i ) ∈ Z[u1 , u1 , . . . , ud , ud ] = Z[Z ], the
corresponding harmonic model shares interesting measure theoretic and entropic properties
with other different models in mathematical physics, probability theory, and dynamical
systems such as the Abelian sandpile model, spanning trees, and the dimer models [32, 4].
Since a well-balanced polynomial f ∈ Z[Γ], with Γ infinite countable not virtually Z or Z2 ,
is weakly expansive (cf. Proposition 3.14), from Theorem 1.2 we deduce:
Corollary 1.4 (Garden of Eden theorem for harmonic models). Let Γ be an infinite countable Abelian group which is not virtually Z or Z2 (e.g. Γ = Zd , with d ≥ 3). Suppose that
f ∈ Z[Γ] is well-balanced and that Xf is connected. Then the dynamical system (Xf , αf )
has the Moore-Myhill property.
Let (X, α) be an algebraic dynamical system and 1 ≤ p ≤ ∞. In [10] (see also Subsection
3.1), the notion of p-homoclinic group (denoted ∆p (X, α) ⊂ X) associated with (X, α), was
introduced. We then say that a map τ : X → X is p-pre-injective if the restriction of τ to
each coset of the p-homoclinic group ∆p (X, α) is injective. Note that ∆∞ (X, α) = ∆(X, α)
so that ∞-pre-injectivity is the same thing as pre-injectivity.
If Γ = Zd , then any polynomial f ∈ R[Γ] may be regarded, by duality, as a function
b = Td . We denote by Z(f ) := {(t1 , t2 , . . . , td ) ∈ Td : f (t1 , t2 , . . . , td ) = 0} its zeroon Γ
set. Recall that an irreducible polynomial f is atoral [22, Definition 2.1] if there is some
g ∈ Z[Γ] such that g 6∈ Z[Γ]f and Z(f ) ⊂ Z(g).
We are now in position to state the following:
GARDEN OF EDEN THEOREM FOR HARMONIC MODELS
5
Theorem 1.5 (A Garden of Eden theorem for irreducible atoral polynomials). Let f ∈
Z[Zd ] be an irreducible atoral polynomial such that Z(f ) is contained in the image of the
intersection of [0, 1]d and a finite union of hyperplanes in Rd under the natural quotient map
Rd → Td (e.g., when d ≥ 2 such that Z(f ) is finite). Let τ : Xf → Xf be a Γ-equivariant
continuous map. Then the following conditions are equivalent:
(a) τ is surjective,
(b) τ is pre-injective,
(c) τ is p-pre-injective for all 1 ≤ p ≤ ∞,
(d) τ is 1-pre-injective.
In particular, (Xf , αf ) satisfies the Moore-Myhill property.
Our motivation for the present work originated from a sentence of Gromov [13, p. 195]
suggesting that the Garden of Eden theorem could be extended to dynamical systems with
a suitable hyperbolic flavor other than shifts and subshifts. A first step in that direction
was made in [6], where the two first named authors proved that all Anosov diffeomorphisms
on tori generate Z-actions with the Moore-Myhill property, and another one in [5], where
sufficient conditions for expansive actions of countable amenable groups to have the Myhill
property were presented. Finally, in [18] the third named author proved the very general
Garden of Eden theorem for expansive actions of amenable groups we alluded to above.
The paper is organized as follows. Section 2 introduces notation and collects background
material on Pontryagin duality and algebraic dynamical systems. In particular, in Subsection 2.6 we present a characterization of the homoclinic group. In Section 3 we consider
several weak forms of expansivity: first we recall from [10] the notions of p-expansivity
(1 ≤ p ≤ ∞) and p-homoclinicity for general algebraic actions. Then we introduce and
study the notion of homoclinically expansive action: when the algebraic dynamical system
is finitely generated, homoclinic expansivity is a stronger condition than p-expansivity for
1 ≤ p < ∞ (Proposition 3.5), and a characterization is derived (Theorem 3.6). Then, we
study principal algebraic actions associated with weakly expansive polynomials with an
emphasis on the corresponding homoclinic group. In Subsection 3.4 we then show that
polynomials yielding principal algebraic expansive actions are weakly expansive and, in
Subsection 3.5, we prove that well-balanced polynomials, with not virtually Z or Z2 infinite countable group, are weakly expansive as well. In Section 4, we discuss topological
rigidity of algebraic dynamical systems associated with weakly expansive polynomials. The
proof of Theorem 1.2 is then given in Section 5. In the following section we discuss the
notion of atorality for irreducible polynomials in Z[Zd ], we present a few examples, and give
the proof of Theorem 1.5. In the last section, we collect some final remarks. In particular,
we exhibit some examples showing that Theorem 1.2 becomes false if weak expansivity of
f ∈ Z[Γ] is replaced by the weaker hypothesis that the homoclinic group ∆(Xf , αf ) is dense
in Xf , or that the dynamical system (Xf , αf ) is mixing. We also introduce and discuss
the notions of p-pre-injectivity, p-Moore, and p-Myhill properties for algebraic actions, and
prove some variations on the Garden of Eden theorem in this framework.
Acknowledgments. Hanfeng Li was partially supported by NSF and NSFC grants.
6
TULLIO CECCHERINI-SILBERSTEIN, MICHEL COORNAERT, AND HANFENG LI
2. Background material and preliminaries
2.1. Group actions. Let Γ be a countable group. We use multiplicative notation for the
group operation in Γ and denote by 1Γ its identity element.
An action of Γ on a set X is a map α : Γ × X → X such that α(1Γ, x) = x and
α(γ1 , α(γ2, x)) = α(γ1 γ2 , x) for all γ1 , γ2 ∈ Γ and x ∈ X. In the sequel, to simplify, we
shall write γx instead of α(γ, x), if there is no risk of confusion.
If α is an action of Γ on a set X, we denote by Fix(X, α) the set of points of X that are
fixed by α, i.e., the set of points x ∈ X such that γx = x for all γ ∈ Γ.
If Γ acts on two sets X and Y , a map τ : X → Y is said to be Γ-equivariant if one has
τ (γx) = γτ (x) for all γ ∈ Γ and x ∈ X.
2.2. Convolution. Let Γ be a countable group. We denote by ℓ∞ (Γ) the vector space
consisting of all formal series
X
f=
fγ γ,
γ∈Γ
with coefficients fγ ∈ R for all γ ∈ Γ and
kf k∞ := sup |fγ | < ∞.
γ∈Γ
For 1 ≤ p < ∞ we denote by ℓp (Γ) the vector subspace of ℓ∞ (Γ) consisting of all f ∈ ℓ∞ (Γ)
such that
! p1
X
kf kp :=
|fγ |p
< ∞.
γ∈Γ
1
Note that ℓ (Γ) ⊂ ℓ (Γ) ⊂ ℓ (Γ) ⊂ ℓ (Γ) for all 1 < p < q < ∞. When f ∈ ℓ∞ (Γ) and
g ∈ ℓ1 (Γ) (resp. f ∈ ℓ1 (Γ) and g ∈ ℓ∞ (Γ)) we define the convolution product f g ∈ ℓ∞ (Γ)
by setting
X
X
fγδ−1 gδ
fγ1 gγ2 =
(2.1)
(f g)γ :=
p
q
∞
γ1 ,γ2 ∈Γ:
γ1 γ2 =γ
δ∈Γ
for all γ ∈ Γ. Note that kf gk∞ ≤ kf k∞ · kgk1 (resp. kf gk∞ ≤ kf k1 · kgk∞ ). We have the
associativity rule
(2.2)
(f g)h = f (gh) for all f ∈ ℓ∞ (Γ), g, h ∈ ℓ1 (Γ) (resp. f, g ∈ ℓ1 (Γ), h ∈ ℓ∞ (Γ)).
We denote by R[Γ] = {f ∈ ℓ∞ (Γ) : fγ = 0 for all but finitely many γ ∈ Γ} and
by Z[Γ] = {f ∈ R[Γ] : fγ ∈ Z for all γ ∈ Γ} the real and, respectively, the integral
group ring of Γ. Observe that the convolution product extends the group operation on
Γ ⊂ Z[Γ] ⊂ R[Γ].
Note also that, as a Z-module, Z[Γ] is free with base Γ.
±1
If we take Γ = Zd , then Z[Γ] is the Laurent polynomial ring Rd := Z[u±1
1 , . . . , ud ] on d
commuting indeterminates u1 , . . . , ud .
GARDEN OF EDEN THEOREM FOR HARMONIC MODELS
7
Recall that C0 (Γ) denotes the vector space consisting of all functions f : Γ → R vanishing
at infinity: we express this condition by writing limγ→∞ f (γ) = 0. We then have the
inclusions
Γ ⊂ Z[Γ] ⊂ R[Γ] ⊂ ℓ1 (Γ) ⊂ ℓp (Γ) ⊂ C0 (Γ) ⊂ ℓ∞ (Γ),
(2.3)
for all 1 ≤ p < ∞. Moreover, there is a natural involution f 7→ f ∗ on ℓ∞ (Γ) defined by
(f ∗ )γ := fγ −1
(2.4)
for all f ∈ ℓ∞ (Γ) and γ ∈ Γ. Observe that every set in (2.3) is ∗-invariant and that
(2.5)
(f g)∗ = g ∗ f ∗ for all f ∈ ℓ∞ (Γ) and g ∈ ℓ1 (Γ) (resp. f ∈ ℓ1 (Γ) and g ∈ ℓ∞ (Γ)).
The normed space (ℓ1 (Γ), k · k1 ) is a unital Banach *-algebra for the convolution product
and the involution. The unity element of ℓ1 (Γ) is 1Γ .
Lemma 2.1. Let f, h ∈ R[Γ] and g ∈ ℓ∞ (Γ). Then (f g)h = f (gh).
Proof. By linearity we can reduce to the case when f = γ and h = δ with γ, δ ∈ Γ. For
η ∈ Γ we have
((γg)δ)η = (γg)ηδ−1 = gγ −1 (ηδ−1 ) =∗ g(γ −1 η)δ−1 = (gδ)γ −1 η = (γ(gδ))η ,
where =∗ follows from the associative rule in Γ.
Let now k, n ∈ N and denote by Matn,k (Z[Γ]) := {(aij )1≤i≤n : aij ∈ Z[Γ]} the space
1≤j≤k
of all n-by-k matrices with coefficients in the group ring Z[Γ]. We identify Z[Γ]k (resp.
Z[Γ]n ) and Mat1,k (Z[Γ]) (resp. Mat1,n (Z[Γ])) so that if g = (g 1, g 2 , . . . , g n) ∈
Z[Γ]n and
P
A = (aij )1≤i≤n ∈ Matn,k (Z[Γ]), the element gA ∈ Z[Γ]k is defined by (gA)j := ni=1 g i aij ∈
1≤j≤k
Z[Γ], that is,
(2.6)
(gA)jγ
=
n X
X
i
gγη
aij
η−1
i=1 η∈Γ
for all j = 1, 2, . . . , k and γ ∈ Γ. Given A = (aij )1≤i≤n ∈ Matn,k (Z[Γ]), we define
(2.7)
A∗ = (aji )
∗
1≤j≤k
1≤i≤k
1≤j≤n
∈ Matk,n (Z[Γ]).
Note that when k = n = 1 and A = f ∈ Z[Γ], then (2.7) reduces to (2.4). Also, we set
kAk∞ = sup sup |aij
γ|
1≤i≤n γ∈Γ
1≤j≤k
and
kAk1 =
X
1≤i≤n
1≤j≤k
kaij k1 =
X X
1≤i≤n γ∈Γ
1≤j≤k
|aij
γ |.
8
TULLIO CECCHERINI-SILBERSTEIN, MICHEL COORNAERT, AND HANFENG LI
2.3. Pontryagin duality. Let us briefly review some basic facts and results regarding
Pontryagin duality. For more details and complete proofs, the reader is refered to [28].
Let X be an LCA group, i.e., a locally compact, Hausdorff, Abelian topological group.
A continuous group morphism from X into the circle T := R/Z is called a character of X.
b of all characters of X, equipped with pointwise multiplication and the topology
The set X
of uniform convergence on compact sets, is also an LCA group, called the character group
or Pontryagin dual of X.
b × X → T, given by hχ, xi = χ(x) for all x ∈ X and χ ∈ X
b
The natural map h·, ·i : X
bb
is bilinear and non-degenerate. Moreover, the evaluation map ι : X → X,
defined by
bb
ι(x)(χ) := hχ, xi, is a topological group isomorphism from X onto its bidual X.
This
bb
canonical isomorphism is used to identify X with X.
The space X is compact (resp. discrete, resp. metrizable, resp. σ-compact) if and only
b is discrete (resp. compact, resp. σ-compact, resp. metrizable). In particular, X is
if X
b is discrete and countable. When X is compact,
compact and metrizable if and only if X
b is a torsion-free group (i.e., a group with no non-trivial
X is connected if and only if X
elements of finite order).
If X is an LCA group and Y a closed subgroup of X, then X/Y is an LCA group whose
Pontryagin dual is canonically isomorphic, as a topological group, to the closed subgroup
b defined by
Y ⊥ of X
b : hχ, yi = 0 for all y ∈ Y }.
Y ⊥ := {χ ∈ X
Let X, Y be LCA groups and ϕ : X → Y a continuous group morphism. The map
b defined by ϕ(χ)
:= χ ◦ ϕ for all χ ∈ Yb is a continuous group morphism,
ϕ
b : Yb → X,
b
bb = ϕ. If
called the dual of ϕ. If we identify X and Y with their respective biduals, then ϕ
ϕ is surjective, then ϕ
b is injective. Conversely, if ϕ is injective, then ϕ
b has dense image
(and it may happen that ϕ
b is not surjective; however, if ϕ is both injective and open,
then ϕ
b is surjective). As a consequence, if X and Y are either both compact or both
discrete, then ϕ is injective (resp. surjective) if and only if ϕ
b is surjective (resp. injective)
[28, Proposition 30].
Let X be an LCA group and suppose that there is a countable group Γ acting continuously on X by group automorphisms. By linearity, this action induces a left Z[Γ]-module
b by continuous group automorphisms,
structure on X. There is a dual action of Γ on X
defined by
b
γχ(x) := χ(γ −1 x) for all γ ∈ Γ, x ∈ X, and χ ∈ X.
b Note that the canonical topoTherefore there is also a left Z[Γ]-module structure on X.
bb
logical group isomorphism ι : X → X
is Γ-equivariant and hence a left Z[Γ]-module isomorphism.
2.4. Algebraic dynamical systems. An algebraic dynamical system is a pair (X, α),
where X is a compact metrizable Abelian topological group and α is an action of a countable
group Γ on X by continuous group automorphisms.
GARDEN OF EDEN THEOREM FOR HARMONIC MODELS
9
As an example, if A is a compact metrizable Abelian topological group (e.g. A = T) and
Γ a countable group, then the system (AΓ , σ), where AΓ = {x : Γ → A} is equipped with
the product topology, and σ is the shift action, defined by
(σ(γ, x))(γ ′ ) := x(γ −1 γ ′ ) for all γ, γ ′ ∈ Γ and x ∈ AΓ ,
is an algebraic dynamical system.
Let (X, α) be an algebraic dynamical system with acting group Γ. As X is compact
b is a discrete countable Abelian group. We have
and metrizable, its Pontryagin dual X
b
seen at the end of the previous subsection that there is a left Z[Γ]-module structure on X
induced by the action of Γ on X. Conversely, if M is a countable left Z[Γ]-module and we
c is a compact metrizable
equip M with its discrete topology, then its Pontryagin dual M
c by continuous group
Abelian group and there is, by duality, an action αM of Γ on M
c
automorphisms, so that (M , αM ) is an algebraic dynamical system. In this way, algebraic
dynamical systems with acting group Γ are in one-to-one correspondence with countable
left Z[Γ]-modules.
2.5. Finitely presented algebraic dynamical systems. Let Γ be a countable group.
One says that an algebraic dynamical system (X, α) with acting group Γ is finitely generated
b equipped with the dual action α
provided its Pontryagin dual X,
b of Γ, is a finitely generated
left Z[Γ]-module.
Recall (cf. [16, Definition 4.25]) that for a unital ring R, a left R-module M is said to be
finitely presented provided there exist k ∈ N and some finitely generated left R-submodule
J ≤ Rk such that M = Rk /J.
Let now k, n ∈ N and A ∈ Matn,k (Z[Γ]). Then Z[Γ]n A is a finitely generated left
Z[Γ]-submodule of Z[Γ]k . We denote by MA := Z[Γ]k /Z[Γ]n A the corresponding finitely
presented left Z[Γ]-module.
d
To simplify notation, let us write XA instead of M
A and αA instead of αMA . The
algebraic dynamical system (XA , αA ) is called the finitely presented algebraic dynamical
system associated with A.
For example, if k = n = 1 and A = f ∈ Z[Γ], then Mf = Z[Γ]/Z[Γ]f , where Z[Γ]f is the
principal left ideal of Z[Γ] generated by f , and the algebraic dynamical system (Xf , αf ) is
called the principal algebraic dynamical system associated with f .
One can regard (XA , αA ) as a group subshift of ((Tk )Γ , σ), i.e., as a closed subgroup of
(Tk )Γ that is invariant under the shift action σ of Γ on (Tk )Γ , in the following way. The
Pontryagin dual of (Tk )Γ is Z[Γ]k with pairing h·, ·i : Z[Γ]k × (Tk )Γ → T given by
hg, xi =
k X
X
gηj xj (η)
j=1 η∈Γ
for all g = (g 1 , g 2, . . . , g k ) ∈ Z[Γ]k and x = (x1 , x2 , . . . , xk ) ∈ (TΓ )k = (Tk )Γ .
10
TULLIO CECCHERINI-SILBERSTEIN, MICHEL COORNAERT, AND HANFENG LI
Therefore, writing A = (aij )1≤i≤n and Ai = (ai1 , ai2 , . . . , aik ) ∈ Z[Γ]k for i = 1, 2, . . . , n,
1≤j≤k
we have
k /Z[Γ]n A
XA = Z[Γ]\
= (Z[Γ]n A)⊥
= {x ∈ (Tk )Γ : hg, xi = 0 for all g ∈ Z[Γ]n A}
=∗ {x ∈ (Tk )Γ : hγAi , xi = 0 for all γ ∈ Γ and i = 1, 2, . . . , n}
k Γ
= {x ∈ (T ) :
k X
X
aij
η xj (γη) = 0 for all γ ∈ Γ and i = 1, 2, . . . , n}
j=1 η∈Γ
k Γ
= {x ∈ (T ) :
k X
X
xj (γη)(aij )∗η−1 = 0 for all γ ∈ Γ and i = 1, 2, . . . , n},
j=1 η∈Γ
where =∗ follows by taking g := γ(0, . . . , 0, 1Γ, 0, . . . , 0)A = γAi ∈ Z[Γ]k . In other words,
(2.8)
XA = {x ∈ (Tk )Γ : xA∗ = 0(Tn )Γ },
with the action αA of Γ on XA ⊂ (Tk )Γ obtained by restricting to XA the shift action σ.
In particular, if A = f ∈ Z[Γ], then (2.8) becomes
(2.9)
Xf = {x ∈ TΓ : xf ∗ = 0TΓ }.
Consider the surjective map π : ℓ∞ (Γ)k → (Tk )Γ defined by π(g)(γ)i = gγi mod 1 for
all g = (g 1, g 2 , . . . , g k ) ∈ ℓ∞ (Γ)k , γ ∈ Γ, and i = 1, 2, . . . , k. Denote by ℓ∞ (Γ, Z)k the set
consisting of all g ∈ ℓ∞ (Γ)k such that gγi ∈ Z for all γ ∈ Γ and i = 1, 2, . . . , k.
Proposition 2.2. Let Γ be a countable group. Let x ∈ (Tk )Γ and g ∈ ℓ∞ (Γ)k such that
π(g) = x. With the above notation, the following conditions are equivalent:
(a) x ∈ XA ;
(b) gA∗ ∈ ℓ∞ (Γ, Z)n .
Proof. This follows immediately from (2.8) and the equality π(gA∗ ) = π(g)A∗ = xA∗ .
2.6. The homoclinic group. Let (X, α) be an algebraic dynamical system with acting
group Γ. The set of points in X that are homoclinic to 0X with respect to α is a Z[Γ]submodule ∆(X, α) ⊂ X, which is called the homoclinic group of (X, α) (cf. [19], [20]).
Note that x ∈ ∆(X, α) if and only if limγ→∞ γx = 0X . We can choose a compatible metric
d on X that is translation-invariant so that
d(γx, γy) = d(γx − γy, 0X ) = d(γ(x − y), 0X )
for all x, y ∈ X and γ ∈ Γ. We deduce that x and y are homoclinic if and only if
x − y ∈ ∆(X, α).
Lemma 2.3. Let Γ be a countable group, k ∈ N, and let x ∈ (Tk )Γ . The following
conditions are equivalent.
GARDEN OF EDEN THEOREM FOR HARMONIC MODELS
11
(a) x ∈ ∆((Tk )Γ , σ);
(b) limγ→∞ x(γ) = 0Tk ;
(c) there exists g ∈ C0 (Γ)k such that x = π(g).
Proof. Suppose (a). Then limγ→∞ γ −1 x = 0(Tk )Γ . As x(γ) = γ −1 x(1Γ ), this implies
lim x(γ) = lim γ −1 x(1Γ ) = 0Tk .
γ→∞
γ→∞
This shows (a) =⇒ (b).
Conversely, suppose (b). Let W be a neighborhood of 0(Tk )Γ in (Tk )Γ and let us show
that there exists a finite subset Ω ⊂ Γ such that
(2.10)
γx ∈ W for all γ ∈ Γ \ Ω.
By definition of the product topology, we can find a neighborhood V of 0Tk in Tk and a
finite subset Ω1 ⊂ Γ such that W contains all y ∈ (Tk )Γ that satisfy
y(ω1) ∈ V for all ω1 ∈ Ω1 .
On the other hand, since limγ→∞ x(γ) = 0Tk , we can find a finite subset Ω2 ⊂ Γ such that
(2.11)
x(γ) ∈ V for all γ ∈ Γ \ Ω2 .
Take Ω := Ω1 Ω−1
2 ⊂ Γ and suppose that γ ∈ Γ \ Ω. Then for every ω1 ∈ Ω1 , we have that
γ −1 ω1 ∈ Γ \ Ω2 and hence
γx(ω1 ) = x(γ −1 ω1 ) ∈ V
by (2.11). This implies that γx ∈ W . Thus (2.10) is satisfied. This proves (b) =⇒ (a).
The fact that (c) implies (b) is an immediate consequence of the continuity of the quotient
map Rk → Rk /Zk = Tk . Conversely, if we assume (b), then the unique g ∈ ℓ∞ (Γ)k such
that −1/2 ≤ gγj < 1/2 and xj (γ) = gγj mod 1 for all γ ∈ Γ and j = 1, 2, . . . , k, clearly
satisfies (c).
2.7. Connectedness of the phase space. A non-zero element f ∈ Z[Γ] is called primitive if there is no integer n ≥ 2 that divides all coefficients of f . Every nonzero element
f ∈ Z[Γ] can be uniquely written in the form f = mf0 with m a positive integer and f0
primitive. The integer m is called the content of f . For principal algebraic dynamical
systems with elementary amenable acting group we have the following criterion for connectedness of the phase space. Recall (cf. for instance [9]) that the class of elementary
amenable groups is the smallest class of groups containing all finite groups and all Abelian
groups that is closed under the operations of taking subgroups, quotiens, extensions, and
direct limits.
Proposition 2.4. Let Γ be a countable torsion-free elementary amenable group (e.g. Γ =
Zd ). Let f ∈ Z[Γ] with f 6= 0. Then the following conditions are equivalent:
(a) Xf is connected;
(b) f is primitive.
12
TULLIO CECCHERINI-SILBERSTEIN, MICHEL COORNAERT, AND HANFENG LI
Proof. (a) ⇒ (b): Suppose that f is not primitive. Then f = mg for some integer m ≥ 2
and g ∈ Z[Γ]. By [15, Theorem 1.4] the ring Q[Γ] is a domain. If g = hf for some h ∈ Z[Γ],
then m1 f = hf in Q[Γ], and hence h = m1 1Γ , which is a contradiction. Thus g + Z[Γ]f is a
nonzero element of Z[Γ]/Z[Γ]f , while m(g + Z[Γ]f ) = 0. Therefore g + Z[Γ]f is a nonzero
torsion element of Mf , and hence Xf is not connected.
(b) ⇒ (a): Suppose that f is primitive and that Xf is not connected. Then Mf has
torsion, so that there exists a ∈ Mf with finite order n ≥ 2. Replacing a by some integral
multiple of a, we may assume that n is a prime number p. Write a = g + Z[Γ]f for some
g ∈ Z[Γ]. Then pg = hf for some h ∈ Z[Γ]. Denote by ψ the natural ring morphism
Z[Γ] → (Z/pZ)[Γ] obtained by reducing coefficients modulo p. Then ψ(h)ψ(f ) = 0. By
[15, Theorem 1.4], the ring (Z/pZ)[Γ] is a domain. Since f is primitive, ψ(f ) 6= 0. Thus
ψ(h) = 0, i.e. h = pw for some w ∈ Z[Γ]. Then pg = hf = pwf , and hence g = wf ∈ Z[Γ].
This means that a = 0, which is a contradiction.
3. Weak forms of expansivity for algebraic actions
In this section we present and study weak forms of expansivity for algebraic actions.
This applies in particular to the harmonic models introduced in [32] (see also [4]).
3.1. p-expansive algebraic actions and p-homoclinic groups. In this section we review the notions of p-expansive algebraic actions and of p-homoclinic groups introduced
by Chung and the third named author in [10, Sections 4 and 5].
Let Γ be a countable group acting by automorphisms of a compact metrizable Abelian
group X. Let also 1 ≤ p ≤ ∞.
b define the function Ψ′ on Γ by setting
For x ∈ X and χ ∈ X,
x,χ
(3.1)
Ψ′x,χ (γ) = e2πihγx,χi − 1,
for all γ ∈ Γ.
One then says that the algebraic dynamical system (X, α) is p-expansive provided there
b and ε > 0 such that 0X is the only point x ∈ X satisfying
exists a finite subset W ⊂ X
X
(3.2)
kΨ′x,χkp < ε.
χ∈W
The following collects the main properties of p-expansivity.
Theorem 3.1 ([10, Proposition 4.3 and Theorem 4.11]). Let (X, Γ, α) be an algebraic
dynamical system. Let 1 ≤ p ≤ ∞. Then the following hold:
(1) If α is p-expansive, then it is q-expansive for all 1 ≤ q ≤ p.
b is a finitely generated left Z[Γ]-module.
(2) If α is p-expansive, then X
b generating X
b as a left Z[Γ](3) If α is p-expansive, then for any finite subset W ⊂ X
module, there exists ε > 0 such that 0X is the only point x ∈ X satisfying (3.2).
(4) α is ∞-expansive if and only if it is expansive.
GARDEN OF EDEN THEOREM FOR HARMONIC MODELS
13
(5) Let k, n ∈ N and A ∈ Matn,k (Z[Γ]). Then αA is p-expansive if an only if the R[Γ]morphism ℓp (Γ)k → ℓp (Γ)n sending g to gA∗ is injective. Moreover, if in addition, Γ
is amenable, then the following conditions are equivalent:
(a) the topological entropy of (XA , αA ) is finite;
(b) αA is 1-expansive;
(c) αA is 2-expansive;
(d) the Z[Γ]-morphism Z[Γ]k → Z[Γ]n sending g to gA∗ is injective;
(e) the R[Γ]-morphism R[Γ]k → R[Γ]n sending g to gA∗ is injective.
We now recall the definitions of a p-homoclinic point and of the p-homoclinic group (cf.
[10, Section 5]). Let Γ be a countable group acting by automorphisms of the compact
metrizable Abelian group X and let 1 ≤ p < ∞. One says that a point x ∈ X is pb Let then ∆p (X, α) denote the set of
homoclinic provided that Ψ′x,χ ∈ ℓp (Γ) for all χ ∈ X.
all p-homoclinic points of X. This is called the p-homoclinic group (cf. Theorem 3.2.(2))
of the algebraic dynamical system (X, α). Also one sets ∆∞ (X, α) := ∆(X, α). Note that
for p = 1, the set ∆1 (X, α) was studied in [32] and [21]. Here below we collect some basic
properties of the p-homoclinic groups.
Theorem 3.2 ([10, Proposition 5.2, Lemma 5.3, and Lemma 5.4]). Let Γ be a countable
group acting by automorphisms of the compact metrizable Abelian group X and let 1 ≤ p ≤
∞. Then the following hold:
One has ∆p (X, α) ⊂ ∆q (X, α) for all p ≤ q ≤ ∞.
∆p (X, α) is a Γ-invariant subgroup of X.
If α is p-expansive, then ∆p (X, α) is countable.
If Z[Γ] is left Noetherian and α is p-expansive, then ∆p (X, α) is a finitely generated
left Z[Γ]-module.
(5) Assume that p < ∞ and let k, n ∈ N and A ∈ Matn,k (Z[Γ]). If αA is p-expansive,
then ∆p (XA , αA ) is isomorphic to a Z[Γ]-submodule of Z[Γ]n /Z[Γ]k A∗ . If, in addition,
the R[Γ]-morphism ℓp (Γ)k → ℓp (Γ)n sending g to gA∗ is invertible, then ∆p (XA , αA ) is
isomorphic to Z[Γ]n /Z[Γ]k A∗ .
(1)
(2)
(3)
(4)
3.2. Homoclinically expansive actions. In this section we introduce and study a new
form of weak expansivity for dynamical systems.
Definition 3.3. Let (X, α) be a dynamical system with acting group Γ. One says that
the action is homoclinically expansive if there exists a constant ε0 > 0 such that, for each
pair of distinct homoclinic points x, y ∈ X, there exists an element γ ∈ Γ such that
d(γx, γy) > ε0 , where d is any compatible metric on X. Such a constant ε0 is then called
a homoclinic-expansivity constant for (X, α, d).
Note that the fact that (X, α) is homoclinically expansive is in fact independent of
the choice of the metric d by compactness of X. A pseudometric d˜ on X is said to be
˜
dynamically-generating if for all distinct x, y ∈ X there is γ ∈ Γ such that d(γx,
γy) > 0
(cf. [14, Definition 9.35]). Now, given a dynamically-generating continuous pseudometric
14
TULLIO CECCHERINI-SILBERSTEIN, MICHEL COORNAERT, AND HANFENG LI
d˜ on X, we can define a compatible metric d on X by setting
∞
X
1 ˜
d(γn x, γn y),
d(x, y) =
n
2
n=0
where γ0 = 1Γ , γ1 , . . . is an enumeration of the elements of Γ. Then
˜
˜
sup d(γx,
γy) ≤ sup d(γx, γy) ≤ 2 sup d(γx,
γy)
γ∈Γ
γ∈Γ
γ∈Γ
for all x, y ∈ X. Thus in Definition 3.3 we may take d to be any dynamically-generating
continuous pseudometric on X.
In the following, we study homoclinic expansivity for algebraic actions. Let (X, α) be
an algebraic dynamical system with acting group Γ.
For any t ∈ R, we set |t + Z| := minm∈Z |t + m|. More generally, for k ∈ N and
t = (t1 , t2 , . . . , tk ) ∈ Rk we set
|t + Zk | := max |tj + Z|.
(3.3)
1≤j≤k
b define a function Ψx,χ on Γ by setting
Given x ∈ X and χ ∈ X,
(3.4)
Ψx,χ (γ) = |hγx, χi|,
for all γ ∈ Γ.
As a comparison between (3.1) and (3.4), note that there is some constant C > 0 such
that
C|t| ≤ |e2πit − 1| ≤ C −1 |t|
for all t ∈ [−1/2, 1/2]. It follows that, for all 1 ≤ p ≤ ∞,
CkΨx,χ kp ≤ kΨ′x,χ kp ≤ C −1 kΨx,χ kp .
It is easy to see that for any x ∈ X, one has that x ∈ ∆(X, α) if and only if Ψx,χ ∈ C0 (Γ)
b
for all χ ∈ X.
Proposition 3.4. Let (X, α) be an algebraic dynamical system with acting group Γ. Then
the following conditions are equivalent:
(a) the action α is homoclinically expansive;
b and ε > 0 such that 0X is the only point x in ∆(X, α)
(b) there exist a finite subset W of X
satisfying
X
(3.5)
kΨx,χ k∞ < ε.
χ∈W
b such that h(χ) > 0 for every χ ∈ X.
b Consider the
Proof. Take a function h ∈ ℓ1 (X)
compatible metric ρ on X defined by
X
X
ρ(x, y) =
h(χ)|hx, χi − hy, χi| =
h(χ)|hx − y, χi|.
b
χ∈X
b
χ∈X
GARDEN OF EDEN THEOREM FOR HARMONIC MODELS
15
Assume that (b) holds with W and ε and let us set ε0 := ε minχ∈W h(χ)/|W |. Let (x, y)
be a homoclinic pair of X with supγ∈Γ ρ(γx, γy) < ε0 . Note that
X
sup ρ(γx, γy) = sup
h(χ)|hγx − γy, χi|
γ∈Γ
γ∈Γ
= sup
γ∈Γ
b
χ∈X
X
h(χ)Ψx−y,χ (γ)
b
χ∈X
≥ sup(min h(χ))
γ∈Γ χ∈W
≥
X
Ψx−y,χ (γ)
χ∈W
minχ∈W h(χ) X
kΨx−y,χ k∞ .
|W |
χ∈W
P
Thus
χ∈W kΨx−y,χ k∞ < ε, and hence x = y. This shows that ε0 is a homoclinicexpansivity constant for (X, α), and (b)⇒(a) follows.
Now assume that (a) holds and let ε0 > 0 be a homoclinic-expansivity constant for
P
b such that
(X, α). Take a finite subset W of X
h(χ) < ε0 /2 and set ε :=
b
χ∈X\W
P
ε0 /(2 maxχ∈W h(χ)). Let x ∈ ∆(X, α) with χ∈W kΨx,χ k∞ < ε. Then
X
sup ρ(γx, γ0X ) = sup
h(χ)Ψx,χ (γ)
γ∈Γ
γ∈Γ
b
χ∈X
≤ ε0 /2 + sup
γ∈Γ
X
h(χ)Ψx,χ (γ)
χ∈W
≤ ε0 /2 + (max h(χ))
χ∈W
< ε0 ,
X
kΨx,χ k∞
χ∈W
and hence x = 0X . This shows that (a)⇒(b).
Proposition 3.5. Let (X, Γ, α) be an algebraic dynamical system. Suppose that α is homob is finitely generated (as a left Z[Γ]-module). Then the following
clinically expansive and X
hold:
(1) α is p-expansive for all 1 ≤ p < ∞;
b generating X
b as a left Z[Γ]-module, there exists ε > 0
(2) for any finite subset W of X
P
such that 0X is the only point x in ∆(X, α) satisfying χ∈W kΨx,χ k∞ < ε.
b and u, v ∈ Z[Γ] we have
Proof. Note that for any x ∈ X, χ, χ′ ∈ X,
(3.6)
kΨx,uχ+vχ′ kq ≤ kuk1 kΨx,χ kq + kvk1 kΨx,χ′ kq
for all 1 ≤ q ≤ ∞.
(1). Since α is homoclinically expansive, by virtue of Proposition 3.4 we can find a finite
b and ε > 0 satisfying (3.5). Enlarging W if necessary, we may assume that
subset W of X
b as a left Z[Γ]-module. Let 1 ≤ p < ∞. If x ∈ X and P
W generates X
χ∈W kΨx,χ kp < ε,
16
TULLIO CECCHERINI-SILBERSTEIN, MICHEL COORNAERT, AND HANFENG LI
P
then from (3.6) we know that x ∈ ∆p (X, α) ⊂ ∆(X, α), and using
χ∈W kΨx,χ k∞ ≤
P
kΨ
k
<
ε,
we
conclude
that
x
=
0
.
Therefore
α
is
p-expansive.
x,χ p
X
χ∈W
(2) follows from Proposition 3.4 and (3.6).
For finitely presented (e.g. principal) algebraic actions we have the following characterization of homoclinic expansivity (compare with Theorem 3.1.(5)).
Theorem 3.6. Let Γ be a countable group. Let k, n ∈ N and A ∈ Matn,k (Z[Γ]). Then the
following conditions are equivalent:
(a) (XA , αA ) is homoclinically expansive;
(b) the linear map C0 (Γ)k → C0 (Γ)n sending g to gA∗ is injective.
Proof. For x ∈ XA consider the function Φx ∈ ℓ∞ (Γ) defined by Φx (γ) := |xγ | for all γ ∈ Γ,
where | · | is as in (3.3).
Assume that (b) fails. Then gA∗ = 0 for some nonzero g ∈ C0 (Γ)k . As a consequence,
for all λ ∈ R one also has λgA∗ = 0 and hence π(λg) ∈ ∆(XA , αA ), by Proposition 2.2
cA denote the image of the canonical basis of Z[Γ]k under
and Lemma 2.3. Let W ⊂ X
c
the quotient map Z[Γ]k → Z[Γ]k /Z[Γ]n A. Note that W then
Pgenerates XA as a left Z[Γ]module. When λ → 0, one has kΦπ(λg) k∞ → 0, and hence χ∈W kΨπ(λg),χ k∞ → 0. Since
g 6= 0, when |λ| is sufficiently small and nonzero, π(λg) 6= 0XA . From Proposition 3.5.(2),
we deduce that (XA , αA ) is not homoclinically expansive. This shows (a)⇒(b).
Now assume that (b) holds so that, in particular, A 6= 0. Let d be a translationinvariant compatible metric on XA . Then there is some ε0 > 0 such that for any x ∈ XA
with d(x, 0XA ) ≤ ε0 , one has |x1Γ | < 1/(2kAk1 ). Let x, y ∈ XA be two homoclinic points
with maxγ∈Γ d(γx, γy) ≤ ε0 . Then x − y ∈ ∆(XA , αA ), and for any γ ∈ Γ we have
d(γ(x − y), 0XA ) ≤ ε0 , and hence |(x − y)γ −1 | = |(γ(x − y))1Γ | < 1/(2kAk1 ). Let g be the
unique element of ℓ∞ (Γ)k satisfying kgk∞ ≤ 1/(2kAk1 ) and π(g) = x − y. Since x − y is
in ∆(XA , αA ), we have g ∈ C0 (Γ)k . It follows that kgA∗k∞ ≤ kgk∞ kAk1 ≤ 1/2 and, by
Proposition 2.2, gA∗ ∈ ℓ∞ (Γ, Z). Therefore, gA∗ = 0. By (b) we have g = 0, and hence
x = y. This shows that ε0 is a homoclinic expansivity constant for (XA , αA ), and (b)⇒(a)
follows as well.
Note that when n = k = 1 and A = f ∈ Z[Γ], condition (b) in Theorem 3.6 (and
therefore homoclinic expansivity of αf ) is equivalent to (we-1) in Definition 1.1.
Proposition 3.7. Let Γ be a countable group. Let k, n ∈ N and A ∈ Matn,k (Z[Γ]) and
suppose that (XA , αA ) is homoclinically expansive. Then ∆(XA , αA ) is isomorphic to a left
Z[Γ]-submodule of Z[Γ]n /Z[Γ]k A∗ .
Proof. For each x ∈ ∆(XA , αA ), take x̃ ∈ C0 (Γ)k with π(x̃) = x (cf. Lemma 2.3). Then,
by virtue of Proposition 2.2, x̃A∗ ∈ ℓ∞ (Γ, Z)n ∩ C0 (Γ)n = Z[Γ]n . If we choose another
x̃′ ∈ C0 (Γ)k with π(x̃′ ) = x, then x̃−x̃′ ∈ ℓ∞ (Γ, Z)k ∩C0 (Γ)k = Z[Γ]k , and hence x̃A∗ −x̃′ A∗ ∈
Z[Γ]k A∗ . Thus the map ϕ : ∆(XA , αA ) → Z[Γ]n /Z[Γ]k A∗ sending x to x̃A∗ + Z[Γ]k A∗ is
well defined. Clearly, ϕ is a left Z[Γ]-module morphism. Let x ∈ ker(ϕ). Then x̃A∗ = gA∗ ,
for some g ∈ Z[Γ]k . By virtue of Theorem 3.6, we have x̃ = g and hence x = 0XA . Thus ϕ
is injective.
GARDEN OF EDEN THEOREM FOR HARMONIC MODELS
17
Examples 3.8. Here below we describe some examples of principal algebraic actions and
discuss their homoclinic expansivivity.
(1) Suppose that γ ∈ Γ has infinite order and denote by Γ′ ∼
= Z the subgroup of Γ it
generates. It follows from [23, Theorem 5.1] that for any nonzero f ∈ C[Γ′ ] and any
nonzero g ∈ C0 (Γ′ ), one has f g 6= 0. Using the right-coset decomposition of Γ, it follows
that for any nonzero f ∈ Z[Γ′ ] and any nonzero g ∈ C0 (Γ), one has f g 6= 0. This shows
that the associated principal algebraic action αf is homoclinically expansive. Note that
if f := 1Γ − γ ∈ Z[Γ′ ], then f is not invertible in ℓ1 (Γ) and hence, by [11, Theorem 3.2]
(cf. Theorem 3.11 below), αf is not expansive. (cf. [10, Example 4.6].)
(2) Let Γ = Zd and let f ∈ R[Γ]. Note that f may be regarded, by duality, as a function
b = Td . We denote by Z(f ) ⊂ Td its zero set. It follows from [24, Theorem 2.2]
on Γ
that Z(f ) is contained in the image of the intersection of [0, 1]d and a finite union of
hyperplanes in Rd under the natural quotient map P : Rd → Td if and only if f g 6= 0 for
all nonzero g ∈ C0 (Γ). From Theorem 3.6 we thus deduce that the principal algebraic
action associated with f ∈ Z[Γ] is homoclinically expansive if and only if Z(f ) is
contained in the image of the intersection of [0, 1]d and a finite union of hyperplanes
in Rd under P . This is the case, for instance, if Z(f ) is finite. (cf. [10, Example 4.9].)
(3) Suppose that Γ contains two elements γ, γ ′ ∈ Γ that generate a non-Abelian free
subsemigroup. Consider the polynomial f := ±3 · 1Γ − (1Γ + γ − γ 2 )γ ′ ∈ Z[Γ]. It follows
from an argument similar to that in [19, Example 7.2] that the associated principal
algebraic action is homoclinically expansive (though not expansive by [17, Example
A.1] and [11, Theorem 3.2]). (cf. [10, Example 4.10].)
(4) In [10, Example 4.7] it is shown that for Γ = Zd , d ≥ 2, the element h = 2d − 1 −
Pd
−1
−1
−1
2d
d
j=1 (uj + uj ) ∈ Z[u1 , u1 , · · · , ud , ud ] = Z[Z ] satisfies that, for any d−1 < p ≤ +∞,
the corresponding principal algebraic action αh is not p-expansive. It follows from
Proposition 3.5.(1) that αh is not homoclinically expansive either.
3.3. Principal algebraic actions associated with weakly expansive polynomials.
Theorem 3.9. Let Γ be a countable group and suppose that f ∈ Z[Γ] is weakly expansive.
Then the following hold:
(1) The element ω ∈ C0 (Γ) satisfying (we-2) in Definition 1.1 is unique and, moreover,
ωf = 1Γ .
(2) ∆(Xf , αf ) is dense in Xf .
(3) ∆(Xf , αf ) is isomorphic to Z[Γ]/Z[Γ]f ∗ as a left Z[Γ]-module.
Proof. Let ω, ω ′ ∈ C0 (Γ) satisfying (we-2). Then f ω = 1Γ = f ω ′ yields f (ω − ω ′ ) = 0 and
condition (we-1) in Definition 1.1 infers ω = ω ′. This proves uniqueness of ω. Moreover,
from Lemma 2.1 and (we-2) we deduce
f (ωf ) = (f ω)f = 1Γ f = f = f 1Γ .
Thus, f (ωf − 1Γ ) = 0, and, again by (we-1), we get ωf = 1Γ . This shows (1).
18
TULLIO CECCHERINI-SILBERSTEIN, MICHEL COORNAERT, AND HANFENG LI
In order to prove (2), let now a ∈ Z[Γ]/Z[Γ]f with ha, π(γω ∗ )i = 0 for all γ ∈ Γ (where
π : ℓ∞ (Γ) → TΓ is as in Subsection 2.5). Write a = g + Z[Γ]f for some g ∈ Z[Γ]. Then
0 = ha, π(γω ∗ )i
X
=
(γω ∗ )δ gδ + Z
δ∈Γ
=
X
(γω ∗ )δ (g ∗ )δ−1 + Z
δ∈Γ
= (γω ∗ g ∗ )1Γ + Z
= (ω ∗ g ∗ )γ −1 + Z
= (gω)γ + Z
for all γ ∈ Γ. Thus h := gω lies in ℓ∞ (Γ, Z). Since ω ∈ C0 (Γ) and g ∈ Z[Γ], one has
gω ∈ C0 (Γ). Therefore h ∈ ℓ∞ (Γ, Z) ∩ C0 (Γ) = Z[Γ]. Using Lemma 2.1 and (1) it follows
that
hf = (gω)f = g(ωf ) = g1Γ = g,
and hence g = hf ∈ Z[Γ]f , which means that a = 0. By Lemma 2.3 we have {π(γω ∗ ) : γ ∈
Γ} ⊂ ∆(Xf , αf ) and by Pontryagin duality we conclude that ∆(Xf , αf ) is dense in Xf .
We are only left to prove (3). In the proof of Proposition (3.7) (here we take n = k = 1
and A = f ) we have defined an injective left Z[Γ]-module morphism ϕ : ∆(Xf , αf ) →
Z[Γ]/Z[Γ]f ∗ sending x to x̃f ∗ + Z[Γ]f ∗ . Let us show that ϕ is surjective. Let h ∈ Z[Γ].
Using (we-2) and Lemma 2.1 we deduce that
ϕ(hπ(ω ∗ )) = ϕ(π(hω ∗)) = (hω ∗ )f ∗ + Z[Γ]f ∗ = h(ω ∗ f ∗ ) + Z[Γ]f ∗
= h(f ω)∗ + Z[Γ]f ∗ = h1Γ + Z[Γ]f ∗ = h + Z[Γ]f ∗ .
This shows that ϕ is surjective. Therefore ϕ is indeed an isomorphism.
It follows from the proof of Theorem 3.9.(3) and the notation therein that the cyclic
Z[Γ]-module ∆(Xf , αf ) is generated by the element x∆ := π(ω ∗) ∈ ∆(Xf , αf ), called the
fundamental homoclinic point of (Xf , αf ) (cf. [19]).
Let (X, α) be an algebraic dynamical system with an infinite acting group Γ and denote
by µ the Haar probability measure on X. Let r ∈ N with r ≥ 2. One says that (X, α) is
mixing of order r if, for all measurable subsets B1 , B2 , . . . , Br ⊂ X, one has
(3.7)
µ(γ1 B1 ∩ γ2 B2 ∩ · · · ∩ γr Br ) −→ µ(B1 )µ(B2 ) · · · µ(Br )
as γi−1 γj → ∞ in Γ for all 1 ≤ i < j ≤ r. If (X, α) is mixing of order r = 2, one simply
says that (X, α) is mixing. Note that every mixing algebraic dynamical system is ergodic.
If (X, α) is mixing of order r for all r ≥ 2, one then says that (X, α) is mixing of all orders.
Observe that if A is a compact metrizable Abelian group and Γ is any infinite countable
group, then the Γ-shift (AΓ , σ) is mixing of all orders since (3.7) is trivially satisfied when
the Bi s are cylinders, for all r ≥ 2.
GARDEN OF EDEN THEOREM FOR HARMONIC MODELS
19
It follows from [4, Proposition 4.6] that an algebraic dynamical system (X, α) admitting
a dense homoclinic group is mixing of all orders. From Theorem 3.9.(2) we then deduce
the following:
Corollary 3.10. Let Γ be an infinite countable group and suppose that f ∈ Z[Γ] is weakly
expansive. Then the associated algebraic dynamical system (Xf , αf ) is mixing of all orders.
3.4. Expansive principal algebraic actions. The following result is due to Deninger
and Schmidt [11, Theorem 3.2] (see also [20, Theorem 5.1]).
Theorem 3.11. Let Γ be a countable group and f ∈ Z[Γ]. Then the following conditions
are equivalent:
(a) the dynamical system (Xf , αf ) is expansive;
(b) f is invertible in ℓ1 (Γ).
For other characterizations of expansivity for algebraic dynamical systems we refer to
[30, 31] (for Γ = Zd , d ∈ N), [26] (for Γ Abelian), [12] (for (X, α) finitely presented), [2]
(for X connected and finite-dimensional), and [10, Theorem 3.1].
i.e., there exists an element γ0 ∈ Γ such that |fγ0 | >
PAs observed in [20], if f is lopsided,
1
γ6=γ0 |fγ |, then f is invertible in ℓ (Γ). On the other hand, there are f ∈ Z[Γ] invertible in
1
ℓ (Γ) that are not lopsided. For instance, if we take Γ = Z, then the polynomial u2 −u−1 ∈
Z[Γ] = Z[u, u−1] is not lopsided although it is invertible in ℓ1 (Γ) (the associated principal
algebraic dynamical system is conjugate to the Z-system generated by Arnold’s cat map
(x1 , x2 ) 7→ (x2 , x1 + x2 ) on the 2-dimensional torus T2 , see e.g. [31, Example 2.18.(2)]).
The following result justifies our terminology for weakly expansive polynomials.
Corollary 3.12. Let Γ be a countable group and f ∈ Z[Γ]. Suppose that the dynamical
system (Xf , αf ) is expansive. Then f is weakly expansive.
Proof. Expansivity of (Xf , αf ) implies, by Theorem 3.11, that f is invertible in ℓ1 (Γ). Then
ω := f −1 ∈ ℓ1 (Γ) ⊂ C0 (Γ) yields (we-2) in Definition 1.1.
Let now g ∈ C0 (Γ) and suppose that f g = 0. Then, recalling (2.2), we deduce that
0 = ω0 = ω(f g) = (ωf )g = 1Γ g = g,
and (we-1) follows as well.
3.5. Harmonic models. Let f ∈ Z[Γ] be well-balanced. It follows from (2.9) that x ∈ TΓ
belongs to Xf if and only if x satisfies the harmonicity equation (mod 1)
X
fη x(γη) = 0,
η∈Γ
for all γ ∈ Γ. This explains the terminology. Note that for Γ = Zd , the polynomial
−1
f ∈ Z[Γ] = Z[u1 , u−1
1 , . . . , ud , ud ] given by
f = 2d −
d
X
i=1
(ui + u−1
i )
20
TULLIO CECCHERINI-SILBERSTEIN, MICHEL COORNAERT, AND HANFENG LI
is well-balanced and the corresponding harmonicity equation is the discrete analogue of
the Laplace equation (cf. Eq. (4.5) in [32]).
Lemma 3.13. Let Γ be a countable infinite group and f ∈ R[Γ]. Suppose that f is wellbalanced. Then the map g 7→ f g from C0 (Γ) to C0 (Γ) is injective. In particular, harmonic
models are homoclinically expansive.
Proof. Set
(3.8)
µ := 1Γ −
1
f ∈ R[Γ].
f1Γ
Then µ is a probability measure on Γ which is symmetric and its support S := supp(µ)
generates Γ as a semigroup, by (wb-1), (wb-3), and (wb-4), respectively.
In order to show (we-1) we apply the maximum principle. Let g ∈ C0 (Γ) and suppose
that f g = 0, equivalently, µg = g. Set M := maxδ∈Γ |gδ | and observe that A := {γ ∈ Γ :
|gγ | = M} is non-empty. Moreover, if γ ∈ A one has, using the triangle inequality and the
properties of µ we alluded to above,
X
X
Mµδ = M,
M = |gγ | = |(µg)γ | ≤
µδ−1 |gδγ | ≤
δ∈S
δ∈S
forcing |gδγ | = M for all δ ∈ S. This shows that SA ⊂ A. A recursive argument immediately shows that S n A ⊂ A for all n ∈ N. Since S generates Γ as a semigroup, we get that
A = Γ. In other words, |g| is a constant function. As g ∈ C0 (Γ), we conclude that g = 0.
The last statement follows immediately after Theorem 3.6.
In the arguments preceding Lemma 4.8 in [4] it is shown that if Γ is a countable infinite
group which is not virtually Z or Z2 and f ∈ Z[Γ] is well-balanced, then
(3.9)
∞
1 X k
ω :=
µ ∈ C0 (Γ),
f1Γ k=0
where µ is as in (3.8), satisfies that f ω = 1Γ , so that (we-2) holds. Combining this with
Lemma 3.13, we deduce:
Proposition 3.14. Let Γ be a countable infinite group Γ that is not virtually Z or Z2 .
Then every balanced polynomial f ∈ Z[Γ] is weakly expansive.
Remark 3.15. It is a well known fact in the theory of Markov chains (cf. for instance
[36, Definition 1.14]) that the sum in (3.9) expressing ω is (pointwise) convergent if and
only if the random walk on Γ associated with µ is transient (i.e., given any finite subset
Ω ⊂ Γ, there exists t(Ω) ∈ N such that, with probability one, the position x(t) ∈ Γ of the
random walker on Γ at time t ≥ t(Ω) satisfies that x(t) ∈ Γ \ Ω) and it is a deep result
of Varopoulos (cf. [33, 34] and [36, Theorem 3.24]) that this is the case exactly if Γ is not
virtually Z or Z2 .
GARDEN OF EDEN THEOREM FOR HARMONIC MODELS
21
4. Topological rigidity
4.1. Affine maps. Let X and Y be two topological Abelian groups. A map τ : Y → X
is called affine if there is a continuous group morphism λ : Y → X and an element t ∈ X
such that τ (y) = λ(y) + t for all y ∈ Y . Note that λ and t are then uniquely determined
by τ since they must satisfy t = λ(0Y ) and λ(y) = τ (y) − t for all y ∈ Y . One says that λ
and t are respectively the linear part and the translational part of the affine map τ .
The following two obvious criteria will be useful in the sequel.
Proposition 4.1. Let (X, α) be an algebraic dynamical system and let τ : X → X be an
affine map with linear part λ : X → X. Then the following conditions are equivalent:
(a) τ is pre-injective;
(b) λ is pre-injective;
(c) Ker(λ) ∩ ∆(X, α) = {0X }.
Proposition 4.2. Let (X, α) be an algebraic dynamical system and let τ : X → X be an
affine map with linear part λ : X → X and translational part t ∈ X. Then the following
conditions are equivalent:
(a) τ is Γ-equivariant;
(b) λ is Γ-equivariant and t ∈ Fix(X, α).
4.2. Topological rigidity. One says that the algebraic dynamical system (X, α) is topologically rigid if every endomorphism τ : X → X of (X, α) is affine.
Before stating our rigidity results, let us introduce some notation. Let L(X) denote
b → R equipped with the topology of
the real vector space of all group homomorphisms X
pointwise convergence. Note that Γ acts on L(X) by setting [γψ](χ) := ψ(γ −1 χ) for all
b Moreover, the map E : L(X) → X defined by E(ψ)(χ) := ψ(χ) + Z
ψ ∈ L(X) and χ ∈ X.
b is a continuous (Γ-equivariant) group homomorphism.
for all ψ ∈ L(X) and χ ∈ X
Theorem 4.3. Let (X, α) and (Y, β) be algebraic dynamical systems with acting group Γ.
Suppose that Y is connected and that the homoclinic group ∆(Y, β) is dense in Y . Also
suppose that (X, α) is homoclinically expansive. Then every Γ-equivariant continuous map
Y → X is affine.
Proof. Let τ : Y → X be a Γ-equivariant continuous map. By Bhattacharya’s extension of
van Kampen theorem [1, Theorem 1], there are a Γ-equivariant affine map λ : Y → X and
a Γ-equivariant continuous map Φ : Y → L(X) such that Φ(0Y ) = 0 and τ = λ + E ◦ Φ.
(Note that in the statement of [1, Theorem 1], X is assumed to be connected as well;
however, in its proof, this condition is never used.) Thus it suffices to show that Φ = 0.
b For any γ ∈ Γ we have
Let y ∈ ∆(Y, β) and χ ∈ X.
[Φ(γy)](χ) = [γΦ(y)](χ) = [Φ(y)](γ −1 χ).
22
TULLIO CECCHERINI-SILBERSTEIN, MICHEL COORNAERT, AND HANFENG LI
When γ → ∞, we have Φ(γy) → Φ(0Y ) = 0, and hence [Φ(y)](γ −1 χ) → 0. For any t ∈ R,
we have
ΨE(tΦ(y)),χ (γ) = |hγE(tΦ(y)), χi| = |hE(tΦ(y)), γ −1χi|
≤ |[tΦ(y)](γ −1 χ)| = |t| · |[Φ(y)](γ −1χ)|
for all γ ∈ Γ, and hence ΨE(tΦ(y)),χ ∈ C0 (Γ). Therefore E(tΨ(y)) ∈ ∆(X, α). Since α is
b and ε > 0
homoclinically expansive, by Proposition 3.4 there exist a finite subset W of X
such that 0X is the only point x in ∆(X, α) satisfying
X
kΨx,χ k∞ < ε.
χ∈W
Set C :=
P
χ∈W
supγ∈Γ |[Φ(y)](γ −1χ)| < ∞. Then
X
kΨE(tΦ(y)),χ k∞ ≤ |t|C.
χ∈W
Thus for all t ∈ R with |t| < ε/C we have E(tΦ(y)) = 0X , which means that tΦ(y) takes
integer values. It follows that Φ(y) = 0. Since ∆(Y, β) is dense in Y and Φ is continuous,
we conclude that Φ = 0 as desired.
Corollary 4.4. Let Γ be a countable group. Suppose that f ∈ Z[Γ] satisfies that (Xf , αf )
is homoclinically expansive and ∆(Xf , αf ) is dense in Xf (e.g., that f is weakly expansive),
and the phase space Xf is connected. Then (Xf , αf ) is topologically rigid.
If, in addition, Γ is Abelian, then for a map τ : Xf → Xf the following conditions are
equivalent:
(a) τ is an endomorphism of the dynamical system (Xf , αf );
(b) there exist r ∈ Z[Γ] and t ∈ Fix(Xf , αf ) such that τ (x) = rx + t for all x ∈ Xf .
Proof. The first statement follows immediately from Theorem 4.3 after taking X = Y = Xf
(if f is weakly expansive, recall Theorem 3.6 and Theorem 3.9.(2)).
Suppose now that Γ is Abelian and let τ : Xf → Xf be a map. For each r ∈ Z[Γ], the
self-mapping of Xf given by x 7→ rx is Γ-equivariant since Z[Γ] is commutative. Therefore,
the fact that (b) implies (a) follows from Proposition 4.2.
To prove the converse, suppose that τ is an endomorphism of the dynamical system
(Xf , αf ). It follows from the first statement that τ is affine. Therefore, by using Proposition 4.2, there exist a continuous Z[Γ]-module morphism λ : Xf → Xf and t ∈ Fix(Xf , αf )
cf =
such that τ (x) = λ(x) + t for all x ∈ Xf . As the ring Z[Γ] is commutative and X
cf .
Z[Γ]/Z[Γ]f is a cyclic Z[Γ]-module, there is s ∈ Z[Γ] such that b
λ(χ) = sχ for all χ ∈ X
b
b setting r := s∗ ∈ Z[Γ] it follows that λ(x) = s∗ x = rx, and hence τ (x) = rx+t,
Since λ = λ,
for all x ∈ Xf .
GARDEN OF EDEN THEOREM FOR HARMONIC MODELS
23
5. Proof of Theorem 1.2
Proof of Theorem 1.2. Let Γ be a countable Abelian group and let f ∈ Z[Γ] be a weakly expansive polynomial such that Xf is connected. Also let τ be an endomorphism of (Xf , αf ),
i.e., a Γ-equivariant continuous map τ : Xf → Xf . We want to show that τ is surjective if
and only if it is pre-injective.
By Corollary 4.4, there exists r ∈ Z[Γ] such that τ is affine with linear part λ : Xf → Xf
given by λ(x) = rx for all x ∈ Xf . Clearly τ is surjective if and only if λ is. As Xf is
compact, we know that surjectivity of λ is equivalent to injectivity of its Pontryagin dual
b: X
b
cf → X
cf . Now we observe that λ(χ)
cf .
λ
= r ∗ χ for all χ ∈ X
cf = Z[Γ]/Z[Γ]f → Z[Γ]/Z[Γ]f ∗ defined by
As Γ is Abelian, the natural mapping Φ : X
Φ(g + Z[Γ]f ) = g ∗ + Z[Γ]f ∗
for all g ∈ Z[Γ] is a group isomorphism. Note that
Φ(g1 (g2 + Z[Γ]f )) = Φ(g1 g2 + Z[Γ]f ) = (g1 g2 )∗ + Z[Γ]f ∗ = g1∗ g2∗ + Z[Γ]f ∗ = g1∗ Φ(g2 + Z[Γ]f )
cf → ∆(Xf , αf ) the composition
for all g1 , g2 ∈ Z[Γ]. As a consequence, denoting by κ : X
∗
of Φ with the Z[Γ]-module isomorphism Z[Γ]/Z[Γ]f → ∆(Xf , αf ) in Theorem 3.9.(3), we
have the commuting diagram
cf
X
κy
χ7→r ∗ χ
−−−−→
x7→rx
cf
X
κ
y
∆(Xf , αf ) −−−→ ∆(Xf , αf ).
b is equivalent to injectivity of the group endomorphism
We deduce that injectivity of λ
µ of ∆(Xf , αf ) defined by µ(x) := rx for all x ∈ ∆(Xf , αf ). As µ is the restriction of λ to
∆(Xf , αf ), we conclude that surjectivity of τ is equivalent to pre-injectivity of τ , by using
Proposition 4.1.
From the proof of Theorem 1.2 and Proposition 3.7 we immediately deduce the following:
Corollary 5.1. Let Γ be a countable Abelian group and let f ∈ Z[Γ]. Suppose that (Xf , αf )
is homoclinically expansive, ∆(Xf , αf ) is dense in Xf , and that Xf is connected. Then
(Xf , αf ) has the Moore property.
6. Atoral polynomials and proof of Theorem 1.5
Lind, Schmidt, and Verbistkiy [22, Theorem 3.2] gave the following geometric-dynamical
characterization of atorality for irreducible polynomials in Z[Zd ].
Theorem 6.1. Let Γ = Zd and suppose that f ∈ Z[Γ] is irreducible. Then the following
conditions are equivalent:
(a) ∆1 (Xf , αf ) 6= {0Xf };
(b) ∆1 (Xf , αf ) is dense in Xf ;
24
TULLIO CECCHERINI-SILBERSTEIN, MICHEL COORNAERT, AND HANFENG LI
(c) f is atoral in the sense that there is some g ∈ Z[Γ] such that g 6∈ Z[Γ]f and Z(f ) ⊂
Z(g);
(d) dim Z(f ) ≤ d − 2.
The meaning of dim(·) in Theorem 6.1.(d) is explained in [22, page 1063]; in particular,
one has dim(∅) := −∞. Also remark that, if d = 1, an irreducible polynomial f ∈ Z[Z] =
Z[u1 , u−1
1 ] is atoral if and only if Z(f ) = ∅ and this, in turn, is equivalent to (Xf , αf ) being
expansive (cf. [19, Lemmma 2.1.(1)]).
Examples 6.2. Here below, we present some examples of irreducible atoral polynomials
f ∈ Z[Zd ], mainly from [21, Section 3] and [22, Section 4]. We can then apply Theorem
1.5 and deduce that the corresponding algebraic dynamical systems (Xf , αf ) satisfy the
Garden of Eden theorem.
(1) Let d = 1 and f (u) = u2 − u − 1 ∈ Z[u, u−1] = Z[Z]. Then f is irreducible and,
since Z(f ) = ∅, atoral. The associated principal algebraic dynamical system
, αf )
(Xf
0 1
is conjugated to the hyperbolic dynamical system (T2 , ϕA ), where A =
∈
1 1
GL(2, Z) is Arnold’s cat and ϕA : T2 → T2 is the associated hyperbolic automorphism
of the two-dimensional torus. Note that (Xf , αf ) is expansive so that we can deduce
the Moore-Myhill property also from Corollary 1.3 (this is a particular case of the
Garden of Eden theorem for Anosov diffeomorphisms on tori [5], we alluded to in the
Introduction).
−1
2
(2) Let d = 2 and f (u1, u2 ) = 2 − u1 − u2 ∈ Z[u1 , u−1
1 , u2 , u2 ] = Z[Z ]. Then Z(f ) =
{(1, 1)}, and so f is atoral. Thus, by Example 3.8.(2), (Xf , αf ) is homoclinically
expansive. Also, it follows from [21, §5] and [22, Example 4.3]) that there is some ω in
C0 (Z2 ) with f ω = ωf = 1Γ . As a consequence, f is weakly expansive (though not wellbalanced). Moreover, f is also primitive, so that, by Proposition 2.4, Xf is connected.
Applying Theorem 1.2, we obtain an alternative proof of the fact that (Xf , αf ) has the
Moore-Myhill property.
−1
(3) Let d = 2, and consider the harmonic model f (u1, u2 ) = 4 − u1 − u−1
1 − u2 − u2 ∈
−1
−1
Z[u1 , u1 , u2 , u2 ] = Z[Z2 ]. One has Z(f ) = {(1, 1)}. Thus f is atoral and (Xf , αf )
satisfies the Garden of Eden theorem, by virtue of Theorem 1.5. (Note that we cannot
apply Theorem 1.2.)
−1
2
(4) Let d = 2, and f (u1 , u2) = 1 + u1 + u2 ∈ Z[u1 , u−1
1 , u2 , u2 ] = Z[Z ]. Then Z(f ) =
{(ω, ω 2), (ω 2, ω)}, where ω = exp(2πi/3). The algebraic dynamical system (Xf , αf ) is
called the connected Ledrappier shift.
−1
2
(5) Let d = 2, and f (u1, u2 ) = 2 − u21 + u2 − u1u2 ∈ Z[u1 , u−1
1 , u2 , u2 ] = Z[Z ]. One has
Z(f ) = {(ξ, η), (ξ, η)}, where ξ, η are algebraic numbers.
−1
2
(6) Let d = 2, and f (u1 , u2) = 2 − u31 + u2 − u1 u2 − u21 u2 ∈ Z[u1 , u−1
1 , u2 , u2 ] = Z[Z ]. Here
Z(f ) = {(1, 1), (i, ζ), (−i, ζ), (ξ, η), (ξ, η)}, where ζ, ξ, η are algebraic numbers.
−1
−1
3
(7) Let d = 3 and f (u1 , u2, u3 ) = 1 + u1 + u2 + u3 ∈ Z[u1 , u−1
1 , u2 , u2 , u3 , u3 ] = Z[Z ]. The
corresponding zero-set Z(f ) is the union of three circles, namely, {(−1, s, −s) : s ∈ T},
{(s, −1, −s) : s ∈ T}, and {(s, −s, −1) : s ∈ T}. Hence, f is atoral.
GARDEN OF EDEN THEOREM FOR HARMONIC MODELS
25
−1
−1
(8) Let d = 3 and f (u1, u2 , u3 ) = 3 + 3u1 − 3u31 + u41 − u2 − u3 ∈ Z[u1 , u−1
1 , u2 , u2 , u3 , u3 ] =
Z[Z3 ]. One has has Z(f ) = {(η, η, η), (η, η, η)}, where η is an algebraic integer.
We are now in position to prove Theorem 1.5.
Proof of Theorem 1.5. We first note that, by Example 3.8.(2), αf is homoclinically expansive. Moreover, by Theorem 6.1, ∆1 (Xf , αf ) and therefore ∆(Xf , αf ) are dense in Xf .
Finally, since f is irreducible it is primitive and hence (cf. Proposition 2.4) Xf is connected.
Thus we may apply Corollary 4.4 and deduce that (Xf , αf ) is topologically rigid.
(a)⇒(b) is given by Corollary 5.1.
(b)⇒(c) follows from ∆p (Xf , αf ) ⊂ ∆(Xf , αf ) for all 1 ≤ p ≤ ∞ (cf. Theorem 3.2.(1)).
(c)⇒(d) this is trivial.
(d)⇒(a). Let τ : Xf → Xf be a Γ-equivariant 1-pre-injective continuous map. By topological rigidity, there exists r ∈ Z[Γ] and t ∈ Xf such that the group endomorphism λ
of Xf defined by setting λ(x) := rx for all x ∈ Xf satisfies that τ (x) = λ(x) + t for all
b: X
b
cf → X
cf is given by λ(χ)
cf . Set
x ∈ Xf . Then the dual map λ
= r ∗ χ for all χ ∈ X
mf = Z[Γ] ∩ (f ℓ1 (Γ)), which is an ideal of Z[Γ]. For each g ∈ mf , one has g = f vg
for a unique vg ∈ ℓ1 (Γ). Then x(g) := π(vg∗ ) lies in ∆1 (Xf , αf ), where π is, as usual,
the projection map ℓ∞ (Γ) → TΓ . The map g + Z[Γ]f 7→ x(g) is clearly a group isomorphism from mf /Z[Γ]f onto ∆1 (Xf , αf ) (cf. [22, Corollary 3.3]). Take g ∈ mf \ Z[Γ]f .
cf to
Since f is irreducible, the group homomorphism h + Z[Γ]f 7→ hg + Z[Γ]f from X
cf → ∆1 (Xf , αf ) defined by
mf /Z[Γ]f is injective. Thus the group homomorphism κ : X
(hg)
κ(h + Z[Γ]f ) := x
is injective. We have the commutative diagram:
cf
X
κy
χ7→r ∗ χ
−−−−→
x7→rx
cf
X
yκ
∆1 (Xf , αf ) −−−→ ∆1 (Xf , αf ).
Since τ is 1-pre-injective, so is λ, that is, λ is injective on ∆1 (Xf , αf ). Hence the dual
b is injective. By Pontryaging duality, this is equivalent to λ being surjective. It
map λ
follows that τ is surjective as well.
7. Concluding remarks
7.1. Surjunctivity. A dynamical system (X, α) is called surjunctive if every injective
endomorphism of (X, α) is surjective (and hence a homeomorphism). As injectivity implies
pre-injectivity, we deduce from Theorem 1.2 that if Γ is a countable Abelian group, f ∈ Z[Γ]
is weakly expansive and Xf is connected, then the dynamical system (Xf , αf ) is surjunctive.
Actually, in the case Γ = Zd , this last result is a particular case of Theorem 1.5 in [3]
c, αM ) is
which asserts that if Γ = Zd and M is a finitely generated Z[Γ]-module, then (M
surjunctive.
26
TULLIO CECCHERINI-SILBERSTEIN, MICHEL COORNAERT, AND HANFENG LI
7.2. Counterexamples for mixing algebraic dynamical systems. Recall (cf. Corollary 3.10) that if Γ is infinite and f ∈ Z[Γ] is weakly expansive, then the associated algebraic
dynamical system (Xf , αf ) is mixing of all orders.
The examples below show that Theorem 1.2 becomes false if the hypothesis that f ∈
Z[Γ] is weakly expansive is replaced by the weaker hypotheses that the homoclinic group
∆(Xf , αf ) is dense and that (Xf , αf ) is mixing.
Example 7.1. Let T = R/Z, Γ = Zd , d ≥ 1, and consider the Γ-shift (TΓ , σ) (this is
(Xf , αf ) for f = 0 ∈ Z[Γ]). Then the endomorphism τ of (TΓ , σ) defined by τ (x)(γ) =
2x(γ) for all x ∈ TΓ and γ ∈ Γ, is clearly surjective. However, τ is not pre-injective since
the configuration y ∈ TΓ , defined by y(γ) = 1/2 mod 1 if γ = 1Γ and 0 otherwise, is a
non-trivial element in the homoclinic group of (TΓ , σ) and satisfies τ (y) = τ (0TΓ ) = 0TΓ .
It follows that (TΓ , σ) does not have the Moore property.
Example 7.2. Let Γ = Z and consider the polynomial
f = 1 − 2u1 + u21 − 2u31 + u41 ∈ Z[u1 , u−1
1 ] = Z[Γ].
The associated algebraic dynamical system (Xf , αf ) is conjugate to the system (T4 , β),
where β is the action of Z on T4 generated by the companion matrix of f . It is mixing
since f is not divisible by a cyclotomic polynomial (cf. [31, Theorem 6.5.(2)]). On the other
hand, f has four distinct roots in C, two on the unit circle, one inside and one outside.
As f is irreducible over Q, it follows that the homoclinic group ∆(Xf , αf ) is reduced to
0 (cf. [19, Example 3.4]). The trivial endomorphism of (Xf , αf ), that maps each x ∈ Xf
to 0, is pre-injective but not surjective. Consequently, (Xf , αf ) does not have the Myhill
property. However, (Xf , αf ) has the Moore property since each homoclinicity class of
(Xf , αf ) is reduced to a single point, so that every map with source set Xf is pre-injective.
Note that (Xf , αf ) = (T4 , β) is topologically rigid since every mixing toral automorphism
is topologically rigid by a result of Walters [35].
7.3. p-pre-injectivity and the p-Moore and p-Myhill properties. In Subsection 3.1
we have recalled from [10] the notions and the main properties of p-expansivity and of
p-homoclinic group (denoted ∆p (X, α)) for an algebraic dynamical system (X, α). In the
Introduction we have also defined the notion of a p-pre-injective map τ : X → X.
Note that, if τ is a group endomorphism of X, then (cf. Proposition 4.1):
(7.1)
τ is p-pre-injective if and only if ker(τ ) ∩ ∆p (X, α) = {0X }.
We shall then say that the algebraic dynamical system (X, α) has the p-Moore property
if every surjective endomorphism of (X, α) is p-pre-injective and that it has the p-Myhill
property if every p-pre-injective endomorphism of (X, α) is surjective. Note that the ∞Moore property (resp. ∞-Myhill property) is nothing but the Moore property (resp. Myhill
property).
It follows immediately from Theorem 3.2.(1) that if 1 ≤ p ≤ q ≤ ∞ the every qpre-injective map τ : X → X is p-pre-injective, so that every algebraic dynamical system
satisfying the q-Moore property (resp. the p-Myhill property) satisfies the p-Moore property
(resp. the q-Myhill property).
GARDEN OF EDEN THEOREM FOR HARMONIC MODELS
27
From the proof of Theorem 1.2 and from Theorem 3.2.(5) we immediately deduce the
following:
Corollary 7.3. Let Γ be a countable Abelian group and let f ∈ Z[Γ]. Let also 1 ≤ p < ∞.
Suppose that (Xf , αf ) is p-expansive and topologically rigid. Then (Xf , αf ) has the p-Moore
property.
In the following result we relax the commutativity condition for the acting group Γ.
Theorem 7.4. Let (X, α) be a finitely generated algebraic dynamical system. Suppose that
Γ is amenable and that (X, α) is topologically rigid and has finite entropy. Then (X, α)
has the 1-Moore property.
Proof. Let τ : X → X be a continuous Γ-equivariant surjective map. By topological rigidity, we can find a continuous group endomorphism λ : X → X and t ∈ X such that
τ (x) = λ(x) + t for all x ∈ X. Since τ is surjective (resp. injective) if and only if λ is
surjective (resp. injective), it is not restrictive to suppose that t = 0, that is, τ is a group
endomorphism of X. Then the entropy addition formula (cf. [17, Corollary 6.3], see also
[14, Theorem 13.48]) yields (we denote by h(·) topological entropy)
h(ker(τ )) = h(X) − h(X/ ker(τ )) = h(X) − h(τ (X)) = 0,
since τ is surjective and X has finite entropy. For any finitely generated algebraic dynamical
system (Y, β) with acting group Γ, if ∆1 (Y, β) is nontrivial, then h(Y ) > 0 (cf. [10, Theorem
b is
7.3 and Propositions 5.7 and 7.10], see also [14, Theorems 13.23 and 13.35]). Since X
\) is a quotient Z[Γ]-module of X,
\) is
b ker(τ
a finitely generated left Z[Γ]-module and ker(τ
also a finitely generated left Z[Γ]-module. It follows that
∆1 (X, α) ∩ ker(τ ) = ∆1 (ker(τ ), α|ker(τ ) ) = {0X }.
This, together with (7.1), shows that τ is 1-pre-injective.
References
[1] S. Bhattacharya, Orbit equivalence and topological conjugacy of affine actions on compact Abelian
groups, Monatsh. Math. 129 (2000), pp. 89–96.
[2] S. Bhattacharya, Expansiveness of algebraic actions on connected groups, Trans. Am. Math. Soc.
356 (12) (2004), pp. 4687–4700.
[3] S. Bhattacharya, T. Ceccherini-Silberstein, and M. Coornaert, Surjunctivity and topological rigidity of algebraic dynamical systems, arXiv:1702.06201, to appear in Ergodic Theory and
Dynamical Systems.
[4] L. Bowen and H. Li, Harmonic models and spanning forsets of residually finite groups, J. Funct.
Anal. 263 (2012), pp. 1769–1808.
[5] T. Ceccherini-Silberstein and M. Coornaert, Expansive actions of countable amenable groups,
homoclinic pairs, and the Myhill property, Illinois J. Math. 59 (2015), pp. 597–621.
[6] T. Ceccherini-Silberstein and M. Coornaert, A Garden of Eden theorem for Anosov diffeomorphisms on tori, Topology Appl. 212 (2016), pp. 49–56.
[7] T. Ceccherini-Silberstein and M. Coornaert, A Garden of Eden theorem for principal algebraic
actions, arXiv:1706.06548.
28
TULLIO CECCHERINI-SILBERSTEIN, MICHEL COORNAERT, AND HANFENG LI
[8] T. Ceccherini-Silberstein, A. Machı̀, and F. Scarabotti, Amenable groups and cellular automata, Ann. Inst. Fourier (Grenoble) 49 (1999), pp. 673–685.
[9] C. Chou, Elementary amenable groups, Ill. J. Math. 24(3) (1980), pp. 396–407.
[10] N.-Ph. Chung and H. Li, Homoclinic group, IE groups, and expansive algebraic actions, Invent.
math. 199 (2015), pp. 805–858.
[11] C. Deninger and K. Schmidt, Expansive algebraic actions of discrete residually finite amenable
groups and their entropy, Ergodic Theory Dynam. Systems 27 (2007), pp. 769–786.
[12] M. Einsiedler and H. Rindler, Algebraic actions of the discrete Heisenberg group and other nonabelian groups, Aequ. Math. 62(1-2) (2001), pp. 117–135.
[13] M. Gromov, Endomorphisms of symbolic algebraic varieties, J. Eur. Math. Soc. (JEMS) 1 (1999),
pp. 109–197.
[14] D. Kerr and H. Li, Ergodic theory. Independence and dichotomies. Springer Monographs in Mathematics. Springer, Cham, 2016.
[15] P. H. Kropholler, P. A. Linnell, and J. A. Moody, Applications of a new K-theoretic theorem
to soluble group rings, Proc. Amer. Math. Soc. 104 (1988), pp. 675–684.
[16] T. Y. Lam, Lectures on modules and rings. Graduate Texts in Mathematics. Springer, New York
(1999).
[17] H. Li, Compact group automorphisms, addition formulas and Fuglede-Kadison determinants, Ann. of
Math. (2) 176 (2012), pp. 303–347.
[18] H. Li, Garden of Eden and specification, Ergodic Theory Dynam. Systems (to appear).
arXiv:1708.09012.
[19] D. Lind and K. Schmidt, Homoclinic points of algebraic Zd -actions, J. Amer. Math. Soc. 12 (1999),
pp. 953–980.
[20] D. Lind and K. Shmidt, A survey of algebraic actions of the discrete Heisenberg group, Uspekhi
Mat. Nauk 70 (2015), pp. 77–142.
[21] D. Lind, K. Schmidt and E. Verbitskiy, Entropy and growth rate of periodic points of algebraic
Zd -actions. In: Dynamical numbers: interplay between dynamical systems and number theory, ed. S.
Kolyada, Yu. Manin, M. Möller, P. Moree and T. Ward, pp. 195–211, Contemporary Mathematics,
vol. 523. American Mathematical Society, Providence (RI) (2010).
[22] D. Lind, K. Schmidt, and E. Verbitskiy, Homoclinic points, atoral polynomials, and periodic
points of algebraic Zd -actions, Ergodic Theory Dynam. Systems 33 (2013), pp. 1060–1081.
[23] P.A. Linnel, Analytic versions of the zero divisor conjecture. In: Geometry and cohomology in group
thoery (Durham, 1994), Lecture Notes Series, vol. 252, pp. 209–248. London Mathematical Society,
Cambridge University Press, Cambridge (1998).
[24] P.A. Linnel and M.J. Puls, Zero divisors and Lp (G), II, New York J. Math. 7 (2001), pp. 49–58.
[25] A. Machı̀ and F. Mignosi, Garden of Eden configurations for cellular automata on Cayley graphs
of groups, SIAM J. Discrete Math. 6 (1993), pp. 44–56.
[26] R. Miles, Expansive algebraic actions of countable abelian groups, Mon. Math. 147(2) (2006), pp. 155–
164.
[27] E. F. Moore, Machine models of self-reproduction, vol. 14 of Proc. Symp. Appl. Math., American
Mathematical Society, Providence, 1963, pp. 17–34.
[28] S. A. Morris, Pontryagin duality and the structure of locally compact Abelian groups, Cambridge
University Press, Cambridge-New York-Melbourne, 1977. London Mathematical Society Lecture Note
Series, No. 29.
[29] J. Myhill, The converse of Moore’s Garden-of-Eden theorem, Proc. Amer. Math. Soc. 14 (1963),
pp. 685–686.
[30] K. Schmidt, Automorphisms of compact abelian groups and affine varieties, Proc. London Math.
Soc. (3) 61 (1990), pp. 480–496.
[31] K. Schmidt, Dynamical systems of algebraic origin, vol. 128 of Progress in Mathematics, Birkhäuser
Verlag, Basel, 1995.
GARDEN OF EDEN THEOREM FOR HARMONIC MODELS
29
[32] K. Schmidt and E. Verbitskiy, Abelian sandpiles and the harmonic model, Comm. Math. Phys.
292 (2009), pp. 721–759.
[33] N.Th. Varopoulos, Long range estimates for Markov chains, Bull. Sci. Math. (2) 109 (1985),
pp. 225–252.
[34] N.Th. Varopoulos, L. Saloff-Coste, and Th. Coulhon, Analysis and geometry on groups,
Cambridge Tracts in Math., vol. 100, Cambridge University Press, Cambridge, 1992.
[35] P. Walters, Topological conjugacy of affine transformations of tori, Trans. Amer. Math. Soc. 131
(1968), pp. 40–50.
[36] W. Woess, Random walks on infinite graphs and groups, Cambridge Tracts in Math., vol. 138,
Cambridge University Press, Cambridge, 2000.
Dipartimento di Ingegneria, Università del Sannio, 82100 Benevento, Italy
E-mail address: [email protected]
Université de Strasbourg, CNRS, IRMA UMR 7501, F-67000 Strasbourg, France
E-mail address: [email protected]
Department of Mathematics, Chongqing University, Chongqing 401331, China
Department of Mathematics, SUNY at Buffalo, Buffalo, NY 14260-2900, USA
E-mail address: [email protected]
| 4math.GR
|
Conditional Generation and Snapshot Learning in
Neural Dialogue Systems
arXiv:1606.03352v1 [cs.CL] 10 Jun 2016
Tsung-Hsien Wen, Milica Gašić, Nikola Mrkšić, Lina M. Rojas-Barahona,
Pei-Hao Su, Stefan Ultes, David Vandyke, Steve Young
Cambridge University Engineering Department,
Trumpington Street, Cambridge, CB2 1PZ, UK
{thw28,mg436,nm480,lmr46,phs26,su259,djv27,sjy11}@cam.ac.uk
Abstract
Recently a variety of LSTM-based conditional language models (LM) have been applied across a range of language generation
tasks. In this work we study various model architectures and different ways to represent and
aggregate the source information in an endto-end neural dialogue system framework. A
method called snapshot learning is also proposed to facilitate learning from supervised
sequential signals by applying a companion
cross-entropy objective function to the conditioning vector. The experimental and analytical results demonstrate firstly that competition occurs between the conditioning vector
and the LM, and the differing architectures
provide different trade-offs between the two.
Secondly, the discriminative power and transparency of the conditioning vector is key to
providing both model interpretability and better performance. Thirdly, snapshot learning
leads to consistent performance improvements
independent of which architecture is used.
1
Introduction
Recurrent Neural Network (RNN)-based conditional language models (LM) have been shown to
be very effective in solving a number of real world
problems, such as machine translation (MT) (Cho
et al., 2014) and image caption generation (Karpathy and Fei-Fei, 2015). Recently, RNNs were applied to task of generating sentences from an explicit semantic representation (Wen et al., 2015a).
Attention-based methods (Mei et al., 2016) and
Long Short-term Memory (LSTM)-like (Hochreiter
and Schmidhuber, 1997) gating mechanisms (Wen et
al., 2015b) have both been studied to improve generation quality. Although it is now clear that LSTMbased conditional LMs can generate plausible natural language, less effort has been put in comparing the different model architectures. Furthermore,
conditional generation models are typically tested
on relatively straightforward tasks conditioned on
a single source (e.g. a sentence or an image) and
where the goal is to optimise a single metric (e.g.
BLEU). In this work, we study the use of conditional LSTMs in the generation component of neural network (NN)-based dialogue systems which depend on multiple conditioning sources and optimising multiple metrics.
Neural conversational agents (Vinyals and Le,
2015; Shang et al., 2015) are direct extensions of
the sequence-to-sequence model (Sutskever et al.,
2014) in which a conversation is cast as a source to
target transduction problem. However, these models are still far from real world applications because they lack any capability for supporting domain
specific tasks, for example, being able to interact
with databases (Sukhbaatar et al., 2015; Yin et al.,
2016) and aggregate useful information into their responses. Recent work by Wen et al. (2016a), however, proposed an end-to-end trainable neural dialogue system that can assist users to complete specific tasks. Their system used both distributed and
symbolic representations to capture user intents, and
these collectively condition a NN language generator to generate system responses. Due to the diversity of the conditioning information sources, the best
way to represent and combine them is non-trivial.
In Wen et al. (2016a), the objective function for
learning the dialogue policy and language generator
depends solely on the likelihood of the output sentences. However, this sequential supervision signal
may not be informative enough to learn a good conditioning vector representation resulting in a generation process which is dominated by the LM. This
can often lead to inappropriate system outputs.
In this paper, we therefore also investigate the use
of snapshot learning which attempts to mitigate this
problem by heuristically applying companion supervision signals to a subset of the conditioning vector.
This idea is similar to deeply supervised nets (Lee
et al., 2015) in which the final cost from the output layer is optimised together with the companion
signals generated from each intermediary layer. We
have found that snapshot learning offers several benefits: (1) it consistently improves performance; (2) it
learns discriminative and robust feature representations and alleviates the vanishing gradient problem;
(3) it appears to learn transparent and interpretable
subspaces of the conditioning vector.
2
Related Work
Machine learning approaches to task-oriented dialogue system design have cast the problem as
a partially observable Markov Decision Process
(POMDP) (Young et al., 2013) with the aim of
using reinforcement learning (RL) to train dialogue policies online through interactions with real
users (Gašić et al., 2013). In order to make RL
tractable, the state and action space must be carefully designed (Young et al., 2010) and the understanding (Henderson et al., 2014; Mrkšić et al.,
2015) and generation (Wen et al., 2015b; Wen et al.,
2016b) modules were assumed available or trained
standalone on supervised corpora. Due to the underlying hand-coded semantic representation (Traum,
1999), the conversation is far from natural and the
comprehension capability is limited. This motivates
the use of neural networks to model dialogues from
end to end as a conditional generation problem.
Interest in generating natural language using NNs
can be attributed to the success of RNN LMs for
large vocabulary speech recognition (Mikolov et al.,
2010; Mikolov et al., 2011). Sutskever et al. (2011)
showed that plausible sentences can be obtained by
sampling characters one by one from the output
layer of an RNN. By conditioning an LSTM on a
sequence of characters, Graves (2013) showed that
machines can synthesise handwriting indistinguishable from that of a human. Later on, this conditional
generation idea has been tried in several research
fields, for example, generating image captions by
conditioning an RNN on a convolutional neural network (CNN) output (Karpathy and Fei-Fei, 2015;
Xu et al., 2015); translating a source to a target language by conditioning a decoder LSTM on top of an
encoder LSTM (Cho et al., 2014; Bahdanau et al.,
2015); or generating natural language by conditioning on a symbolic semantic representation (Wen et
al., 2015b; Mei et al., 2016). Among all these methods, attention-based mechanisms (Bahdanau et al.,
2015; Hermann et al., 2015; Ling et al., 2016) have
been shown to be very effective improving performance using a dynamic source aggregation strategy.
To model dialogue as conditional generation,
a sequence-to-sequence learning (Sutskever et al.,
2014) framework has been adopted. Vinyals and Le
(2015) trained the same model on several conversation datasets and showed that the model can generate plausible conversations. However, Serban et al.
(2015b) discovered that the majority of the generated responses are generic due to the maximum likelihood criterion, which was latter addressed by Li
et al. (2016a) using a maximum mutual information
decoding strategy. Furthermore, the lack of a consistent system persona was also studied in Li et al.
(2016b). Despite its demonstrated potential, a major barrier for this line of research is data collection.
Many works (Lowe et al., 2015; Serban et al., 2015a;
Dodge et al., 2016) have investigated conversation
datasets for developing chat bot or QA-like general
purpose conversation agents. However, collecting
data to develop goal oriented dialogue systems that
can help users to complete a task in a specific domain remains difficult. In a recent work by Wen et
al. (2016a), this problem was addressed by designing an online, parallel version of Wizard-of-Oz data
collection (Kelley, 1984) which allows large scale
and cheap in-domain conversation data to be collected using Amazon Mechanical Turk. An NNbased dialogue model was also proposed to learn
from the collected dataset and was shown to be able
to assist human subjects to complete specific tasks.
Snapshot learning can be viewed as a special form
of weak supervision (also known as distant- or self
supervision) (Craven and Kumlien, 1999; Snow et
al., 2004), in which supervision signals are heuristically labelled by matching unlabelled corpora with
entities or attributes in a structured database. It has
been widely applied to relation extraction (Mintz et
al., 2009) and information extraction (Hoffmann et
al., 2011) in which facts from a knowledge base (e.g.
Freebase) were used as objectives to train classifiers.
Recently, self supervision was also used in memory networks (Hill et al., 2016) to improve the discriminative power of memory attention. Conceptually, snapshot learning is related to curriculum learning (Bengio et al., 2009). Instead of learning easier examples before difficult ones, snapshot learning
creates an easier target for each example. In practice, snapshot learning is similar to deeply supervised nets (Lee et al., 2015) in which companion objectives are generated from intermediary layers and
optimised altogether with the output objective.
3
Neural Dialogue System
The testbed for this work is a neural network-based
task-oriented dialogue system proposed by Wen et
al. (2016a). The model casts dialogue as a source
to target sequence transduction problem (modelled
by a sequence-to-sequence architecture (Sutskever
et al., 2014)) augmented with the dialogue history (modelled by a belief tracker (Henderson et
al., 2014)) and the current database search outcome
(modelled by a database operator). The model consists of both encoder and decoder modules. The details of each module are given below.
3.1
Encoder Module
At each turn t, the goal of the encoder is to produce
a distributed representation of the system action mt ,
which is then used to condition a decoder to generate the next system response in skeletal form1 . It
consists of four submodules: intent network, belief
tracker, database operator, and policy network.
Intent Network The intent network takes a sequence of tokens1 and converts it into a sentence embedding representing the user intent using an LSTM
1
Delexicalisation: slots and values are replaced by generic
tokens (e.g. keywords like Chinese food are replaced by
[v.food] [s.food] to allow weight sharing.
network. The hidden layer of the LSTM at the last
encoding step zt is taken as the representation. As
mentioned in Wen et al. (2016a), this representation
can be viewed as a distributed version of the speech
act (Traum, 1999) used in traditional systems.
Belief Trackers In addition to the intent network,
the neural dialogue system uses a set of slot-based
belief trackers (Henderson et al., 2014; Mrkšić et al.,
2015) to track user requests. By taking each user input as new evidence, the task of a belief tracker is
to maintain a multinomial distribution p over values
v ∈ Vs for each informable slot s2 , and a binary
distribution for each requestable slot2 . These probability distributions pst are called belief states of the
system. The belief states pst , together with the intent
vector zt , can be viewed as the system’s comprehension of the user requests up to turn t.
Database Operator Based on the belief states pst ,
a DB query is formed by taking the union of the
maximum values of each informable slot. A vector
xt representing different degrees of matching in the
DB (no match, 1 match, ... or more than 5 matches)
is produced by counting the number of matched entities and expressing it as a 6-bin 1-hot encoding. If xt
is not zero, an associated entity pointer is maintained
which identifies one of the matching DB entities selected at random. The entity pointer is updated if the
current entity no longer matches the search criteria;
otherwise it stays the same.
Policy Network Based on the vectors zt , pst , and
xt from the above three modules, the policy network
combines them into a single action vector mt by a
three-way matrix transformation,
mt = tanh(Wzm zt + Wxm xt +
s
s
s∈G Wpm pt )
P
(1)
s , and W
where matrices Wzm , Wpm
xm are parameters and G is the domain ontology.
3.2
Decoder Module
Conditioned on the system action vector mt provided by the encoder module, the decoder module uses a conditional LSTM LM to generate the
required system output token by token in skeletal
form1 . The final system response can then be formed
2
Informable slots are slots that users can use to constrain the
search, such as food type or price range; Requestable slots are
slots that users can ask a value for, such as phone number. This
information is specified in the domain ontology.
(a) Language model type LSTM
(b) Memory type LSTM
(c) Hybrid type LSTM
Figure 1: Three different conditional generation architectures.
by substituting the actual values of the database entries into the skeletal sentence structure.
3.2.1 Conditional Generation Network
In this paper we study and analyse three different
variants of LSTM-based conditional generation architectures:
Language Model Type The most straightforward
way to condition the LSTM network on additional
source information is to concatenate the conditioning vector mt together with the input word embedding wj and previous hidden layer hj−1 ,
ij
sigmoid
fj sigmoid
=
oj sigmoid
tanh
ĉj
mt
W4n,3n wj
hj−1
ĉj = tanh Wc (wj ⊕ hj−1 )
cj = fj
cj−1 + ij
hj = o j
cj−1 + ij
hj = oj
ij
sigmoid
fj sigmoid
oj = sigmoid
sigmoid
rj
ĉj
tanh(cj )
where index j is the generation step, n is the hidden
layer size, ij , fj , oj ∈ [0, 1]n are input, forget, and
output gates respectively, ĉj and cj are proposed cell
value and true cell value at step j, and W4n,3n are
the model parameters. The model is shown in Figure 1a. Since it does not differ significantly from the
original LSTM, we call it the language model type
(lm) conditional generation network.
Memory Type The memory type (mem) conditional generation network was introduced by Wen et
al. (2015b), shown in Figure 1b, in which the conditioning vector mt is governed by a standalone reading gate rj . This reading gate decides how much information should be read from the conditioning vector and directly writes it into the memory cell cj ,
ij
sigmoid
mt
fj sigmoid
=
W4n,3n wj
oj sigmoid
hj−1
rj
sigmoid
ĉj + rj
mt
tanh(cj )
where Wc is another weight matrix to learn. The
idea behind this is that the model isolates the conditioning vector from the LM so that the model has
more flexibility to learn to trade off between the two.
Hybrid Type Continuing with the same idea as the
memory type network, a complete separation of conditioning vector and LM (except for the gate controlling the signals) is provided by the hybrid type
network shown in Figure 1c,
cj = f j
mt
W4n,3n wj
hj−1
ĉj = tanh Wc (wj ⊕ hj−1 )
cj = f j
hj = o j
cj−1 + ij
tanh(cj ) + rj
ĉj
mt
This model was motivated by the fact that long-term
dependency is not needed for the conditioning vector because we apply this information at every step j
anyway. The decoupling of the conditioning vector
and the LM is attractive because it leads to better interpretability of the results and provides the potential
to learn a better conditioning vector and LM.
3.2.2
Attention and Belief Representation
Attention An attention-based mechanism provides
an effective approach for aggregating multiple information sources for prediction tasks. Like Wen et al.
(2016a), we explore the use of an attention mechanism to combine the tracker belief states in which
the policy network in Equation 1 is modified as
mjt = tanh(Wzm zt + Wxm xt +
j
s
s
s∈G αs Wpm pt )
P
where the attention weights αsj are calculated by,
αsj = softmax r| tanh Wr · (vt ⊕ pst ⊕ wjt ⊕ htj−1 )
where vt = zt + xt and matrix Wr and vector r are
parameters to learn.
Belief Representation The effect of different belief state representations on the end performance are
also studied. For user informable slots, the full belief
state pst is the original state containing all categorical values; the summary belief state contains only
three components: the summed value of all categorical probabilities, the probability that the user said
they “don’t care” about this slot and the probability that the slot has not been mentioned. For user
requestable slots, on the other hand, the full belief
state is the same as the summary belief state because
the slot values are binary rather than categorical.
3.3
Snapshot Learning
Learning conditional generation models from sequential supervision signals can be difficult, because
it requires the model to learn both long-term word
dependencies and potentially distant source encoding functions. To mitigate this difficulty, we introduce a novel method called snapshot learning
to create a vector of binary labels Υjt ∈ [0, 1]d ,
d < dim(mjt ) as the snapshot of the remaining part
of the output sentence Tt,j:|Tt | from generation step
j. Each element of the snapshot vector is an indicator function of a certain event that will happen in the
future, which can be obtained either from the system response or dialogue context at training time. A
companion cross entropy error is then computed to
force a subset of the conditioning vector m̂jt ⊂ mjt
to be close to the snapshot vector,
Lss (·) = −
P P
t
j
j
j
E[H(Υt , m̂t )]
(2)
where H(·) is the cross entropy function, Υjt and m̂jt
are elements of vectors Υjt and m̂jt , respectively. In
order to make the tanh activations of m̂jt compatible with the 0-1 snapshot labels, we squeeze each
Figure 2: The idea of snapshot learning. The snapshot vector was trained with additional supervisions
on a set of indicator functions heuristically labelled
using the system response.
value of m̂jt by adding 1 and dividing by 2 before
computing the cost.
The indicator functions we use in this work have
two forms: (1) whether a particular slot value (e.g.,
[v.food]1 ) is going to occur, and (2) whether the system has offered a venue3 , as shown in Figure 2. The
offer label in the snapshot is produced by checking
the delexicalised name token ([v.name]) in the entire dialogue. If it has occurred, every label in subsequent turns is labelled with 1. Otherwise it is labelled with 0. To create snapshot targets for a particular slot value, the output sentence is matched with
the corresponding delexicalised token turn by turn,
per generation step. At each generation step, the target is labelled with 0 if that delexicalised token has
been generated; otherwise it is set to 1. However, for
the models without attention, the targets per turn are
set to the same because the condition vector will not
be able to learn the dynamically changing behaviour
without attention.
4
Experiments
Dataset The dataset used in this work was collected in the Wizard-of-Oz online data collection described by Wen et al. (2016a), in which the task of
the system is to assist users to find a restaurant in
Cambridge, UK area. There are three informable
slots (food, pricerange, area) that users can use to
constrain the search and six requestable slots (address, phone, postcode plus the three informable
3
Details of the specific application used in this study are
given in Section 4 below.
Architecture
Belief
Success(%)
SlotMatch(%)
T5-BLEU
T1-BLEU
Belief state representation
lm
full
72.6 / 74.5
lm
summary 74.5 / 76.5
52.1 / 60.3
57.4 / 61.2
0.207 / 0.229
0.221 / 0.231
0.216 / 0.238
0.227 / 0.240
Conditional architecture
lm
summary
mem
summary
hybrid
summary
74.5 / 76.5
75.5 / 77.5
76.1 / 79.2
57.4 / 61.2
59.2 / 61.3
52.4 / 60.6
0.221 / 0.231
0.222 / 0.232
0.202 / 0.228
0.227 / 0.240
0.231 / 0.243
0.212 / 0.237
Attention-based model
lm
summary
mem
summary
hybrid
summary
79.4 / 78.2
76.5 / 80.2
79.0 / 81.8
60.6 / 60.2
57.4 / 61.0
56.2 / 60.5
0.228 / 0.231
0.220 / 0.229
0.214 / 0.227
0.239 / 0.241
0.228 / 0.239
0.224 / 0.240
Table 1: Performance comparison of different model architectures, belief state representations, and snapshot
learning. The numbers to the left and right of the / sign are learning without and with snapshot, respectively.
The model with the best performance on a particular metric (column) is shown in bold face.
slots) that the user can ask a value for once a restaurant has been offered. There are 676 dialogues in the
dataset (including both finished and unfinished dialogues) and approximately 2750 turns in total. The
database contains 99 unique restaurants.
Training The training procedure was divided into
two stages. Firstly, the belief tracker parameters
θb were pre-trained using cross entropy errors between tracker labels and predictions. Having fixed
the tracker parameters, the remaining parts of the
model θ\b are trained using the cross entropy errors
from the generation network LM,
L(θ\b ) = −
P P
t
j
H(yjt , ptj ) + λLss (·)
(3)
where yjt and ptj are output token targets and predictions respectively, at turn t of output step j, Lss (·)
is the snapshot cost from Equation 2, and λ is the
tradeoff parameter in which we set to 1 for all models trained with snapshot learning. We treated each
dialogue as a batch and used stochastic gradient descent with a small l2 regularisation term to train the
model. The collected corpus was partitioned into
a training, validation, and testing sets in the ratio
3:1:1. Early stopping was implemented based on the
validation set considering only LM log-likelihoods.
Gradient clipping was set to 1. The hidden layer
sizes were set to 50, and the weights were randomly
initialised between -0.3 and 0.3 including word embeddings. The vocabulary size is around 500 for
both input and output, in which rare words and
words that can be delexicalised have been removed.
Decoding In order to compare models trained with
different recipes rather than decoding strategies, we
decode all the trained models with the average log
probability of tokens in the sentence. We applied
beam search with a beamwidth equal to 10, the
search stops when an end-of-sentence token is generated. In order to consider language variability, we
ran decoding until 5 candidates were obtained and
performed evaluation on them.
Metrics We compared models trained with different recipes by performing a corpus-based evaluation
in which the model is used to predict each system
response in the held-out test set. Three evaluation
metrics were used: BLEU score (on top-1 and top5 candidates) (Papineni et al., 2002), slot matching
rate and objective task success rate (Su et al., 2015).
The dialogue is marked as successful if both: (1)
the offered entity matches the task that was specified to the user, and (2) the system answered all the
associated information requests (e.g. what is the address?) from the user. The slot matching rate is the
percentage of delexicalised tokens (e.g. [s.food] and
[v.area]1 ) appear in the candidate also appear in the
reference. We computed the BLEU scores on the
skeletal sentence forms before substituting with the
actual entity values. All the results were averaged
over 10 random initialised networks.
Results Table 1 shows the evaluation results. The
numbers to the left and right of each table cell are the
same model trained w/o and w/ snapshot learning.
(a) Hybrid LSTM w/o snapshot learning
(b) Hybrid LSTM w/ snapshot learning
Figure 3: Learned attention heat maps over trackers. The first three columns in each figure are informable
slot trackers and the rest are requestable slot trackers. The generation model is the hybrid type LSTM.
The first observation is that snapshot learning consistently improves on most metrics regardless of the
model architecture. This is especially true for BLEU
scores. We think this may be attributed to the more
discriminative conditioning vector learned through
the snapshot method, which makes the learning of
the conditional LM easier.
In the first block belief state representation, we
compare the effect of two different belief representations. As can be seen, using a succinct representation is better (summary>full) because the identity of each categorical value in the belief state does
not help when the generation decisions are done in
skeletal form. In fact, the full belief state representation may encourage the model to learn incorrect coadaptation among features when the data is scarce.
In the conditional architecture block, we compare the three different conditional generation architectures as described in section 3.2.1. This result
shows that the language model type (lm) and memory type (mem) networks perform better in terms of
BLEU score and slot matching rate, while the hybrid
type (hybrid) networks achieve higher task success.
This is probably due to the degree of separation between the LM and conditioning vector: a coupling
approach (lm, mem) sacrifices the conditioning vector but learns a better LM and higher BLEU; while
a complete separation (hybrid) learns a better conditioning vector and offers a higher task success.
Lastly, in the attention-based model block we
train the three architectures with the attention mech-
Model
hybrid, full
hybrid, summary
+ att.
ij
fj
rj /oj
0.567
0.539
0.540
0.502
0.540
0.559
0.405
0.428
0.459
Table 2: Average activation of gates on test set.
anism and compare them again. Firstly, the characteristics of the three models we observed above
also hold for attention-based models. Secondly, we
found that the attention mechanism improves all
the three architectures on task success rate but not
BLEU scores. This is probably due to the limitations of using n-gram based metrics like BLEU to
evaluate the generation quality (Stent et al., 2005).
5
Model Analysis
Gate Activations We first studied the average activation of each individual gate in the models by averaging them when running generation on the test
set. We analysed the hybrid models because their
reading gate to output gate activation ratio (rj /oj )
shows clear tradeoff between the LM and the conditioning vector components. As can be seen in Table 2, we found that the average forget gate activations (fj ) and the ratio of the reading gate to the output gate activation (rj /oj ) have strong correlations
to performance: a better performance (row 3>row
2>row 1) seems to come from models that can learn
a longer word dependency (higher forget gate ft activations) and a better conditioning vector (therefore
(a)
(b)
Figure 4: Two example responses generated from the hybrid model trained with snapshot and attention.
Each line represents a neuron that detects a particular snapshot event.
higher reading to output gate ratio rj /oj ).
6
Learned Attention
We have visualised the
learned attention heat map of models trained with
and without snapshot learning in Figure 3. The attention is on both the informable slot trackers (first
three columns) and the requestable slot trackers (the
other columns). We found that the model trained
with snapshot learning (Figure 3b) seems to produce a more accurate and discriminative attention
heat map comparing to the one trained without it
(Figure 3a). This may contribute to the better performance achieved by the snapshot learning approach.
This paper has investigated different conditional
generation architectures and a novel method called
snapshot learning to improve response generation in
a neural dialogue system framework. The results
showed three major findings. Firstly, although the
hybrid type model did not rank highest on all metrics, it is nevertheless preferred because it achieved
the highest task success and also it provided more interpretable results. Secondly, snapshot learning provided gains on virtually all metrics regardless of the
architecture used. The analysis suggested that the
benefit of snapshot learning mainly comes from the
more discriminative and robust subspace representation learned from the heuristically labelled companion signals, which in turn facilitates optimisation
of the final target objective. Lastly, the results suggested that by making a complex system more interpretable at different levels not only helps our understanding but also leads to the highest success rates.
However, there is still much work left to do. This
work focused on conditional generation architectures and snapshot learning in the scenario of generating dialogue responses. It would be very helpful
if the same comparison could be conducted in other
application domains such as machine translation or
image caption generation so that a wider view of the
effectiveness of these approaches can be assessed.
Snapshot Neurons As mentioned earlier, snapshot learning forces a subspace of the conditioning vector m̂jt to become discriminative and interpretable. Two example generated sentences together
with the snapshot neuron activations are shown in
Figure 4. As can be seen, when generating words
one by one, the neuron activations were changing
to detect different events they were assigned by the
snapshot training signals: e.g. in Figure 4b the light
blue and orange neurons switched their domination
role when the token [v.address] was generated; the
offered neuron is in a high activation state in Figure 4b because the system was offering a venue,
while in Figure 4a it is not activated because the system was still helping the user to find a venue. More
examples can be found in the Appendix.
Conclusion and Future Work
Acknowledgments
Tsung-Hsien Wen and David Vandyke are supported
by Toshiba Research Europe Ltd, Cambridge Research Laboratory.
References
Dzmitry Bahdanau, Kyunghyun Cho, and Yoshua Bengio. 2015. Neural machine translation by jointly
learning to align and translate. In ICLR.
Yoshua Bengio, Jérôme Louradour, Ronan Collobert, and
Jason Weston. 2009. Curriculum learning. In ICML.
Kyunghyun Cho, Bart van Merrienboer, Çaglar Gülçehre,
Fethi Bougares, Holger Schwenk, and Yoshua Bengio.
2014. Learning phrase representations using RNN
encoder-decoder for statistical machine translation. In
EMNLP.
Mark Craven and Johan Kumlien. 1999. Constructing
biological knowledge bases by extracting information
from text sources. In ISMB.
Jesse Dodge, Andreea Gane, Xiang Zhang, Antoine Bordes, Sumit Chopra, Alexander Miller, Arthur Szlam,
and Jason Weston. 2016. Evaluating prerequisite qualities for learning end-to-end dialog systems.
ICLR.
Milica Gašić, Catherine Breslin, Matthew Henderson,
Dongho Kim, Martin Szummer, Blaise Thomson, Pirros Tsiakoulis, and Steve Young. 2013. On-line policy
optimisation of bayesian spoken dialogue systems via
human interaction. In ICASSP.
Alex Graves. 2013. Generating sequences with recurrent
neural networks. arXiv preprint:1308.0850.
Matthew Henderson, Blaise Thomson, and Steve Young.
2014. Word-based dialog state tracking with recurrent
neural networks. In SIGdial.
Karl Moritz Hermann, Tomás Kociský, Edward Grefenstette, Lasse Espeholt, Will Kay, Mustafa Suleyman,
and Phil Blunsom. 2015. Teaching machines to read
and comprehend. In NIPS.
Felix Hill, Antoine Bordes, Sumit Chopra, and Jason Weston. 2016. The goldilocks principle: Reading children’s books with explicit memory representations. In
ICLR.
Sepp Hochreiter and Jürgen Schmidhuber. 1997. Long
short-term memory. Neural Computation.
Raphael Hoffmann, Congle Zhang, Xiao Ling, Luke
Zettlemoyer, and Daniel S. Weld. 2011. Knowledgebased weak supervision for information extraction of
overlapping relations. In ACL.
Andrej Karpathy and Li Fei-Fei. 2015. Deep visualsemantic alignments for generating image descriptions. In CVPR.
John F. Kelley. 1984. An iterative design methodology
for user-friendly natural language office information
applications. ACM Transaction on Information Systems.
Chen-Yu Lee, Saining Xie, Patrick Gallagher, Zhengyou
Zhang, and Zhuowen Tu. 2015. Deeply-supervised
nets. In AISTATS.
Jiwei Li, Michel Galley, Chris Brockett, Jianfeng Gao,
and Bill Dolan. 2016a. A diversity-promoting objective function for neural conversation models. In
NAACL-HLT.
Jiwei Li, Michel Galley, Chris Brockett, Jianfeng Gao,
and Bill Dolan. 2016b. A persona-based neural conversation model. arXiv perprint:1603.06155.
Wang Ling, Edward Grefenstette, Karl Moritz Hermann,
Tomás Kociský, Andrew Senior, Fumin Wang, and
Phil Blunsom. 2016. Latent predictor networks for
code generation. arXiv preprint:1603.06744.
Ryan Lowe, Nissan Pow, Iulian Serban, and Joelle
Pineau. 2015. The ubuntu dialogue corpus: A large
dataset for research in unstructured multi-turn dialogue systems. In SIGdial.
Hongyuan Mei, Mohit Bansal, and Matthew R. Walter.
2016. What to talk about and how? selective generation using lstms with coarse-to-fine alignment. In
NAACL.
Tomáš Mikolov, Martin Karafit, Lukáš Burget, Jan
Černocký, and Sanjeev Khudanpur. 2010. Recurrent
neural network based language model. In InterSpeech.
Tomáš Mikolov, Stefan Kombrink, Lukáš Burget, Jan H.
Černocký, and Sanjeev Khudanpur. 2011. Extensions of recurrent neural network language model. In
ICASSP.
Mike Mintz, Steven Bills, Rion Snow, and Dan Jurafsky.
2009. Distant supervision for relation extraction without labeled data. In ACL.
Nikola Mrkšić, Diarmuid Ó Séaghdha, Blaise Thomson,
Milica Gašić, Pei-Hao Su, David Vandyke, TsungHsien Wen, and Steve Young. 2015. Multi-domain
Dialog State Tracking using Recurrent Neural Networks. In ACL.
Kishore Papineni, Salim Roukos, Todd Ward, and WeiJing Zhu. 2002. Bleu: a method for automatic evaluation of machine translation. In ACL.
Iulian Vlad Serban, Ryan Lowe, Laurent Charlin, and
Joelle Pineau. 2015a. A survey of available corpora for building data-driven dialogue systems. arXiv
preprint:1512.05742.
Iulian Vlad Serban, Alessandro Sordoni, Yoshua Bengio,
Aaron C. Courville, and Joelle Pineau. 2015b. Hierarchical neural network generative models for movie
dialogues. In AAAI.
Lifeng Shang, Zhengdong Lu, and Hang Li. 2015. Neural responding machine for short-text conversation. In
ACL.
Rion Snow, Daniel Jurafsky, and Andrew Y. Ng. 2004.
Learning syntactic patterns for automatic hypernym
discovery. In NIPS.
Amanda Stent, Matthew Marge, and Mohit Singhai.
2005. Evaluating evaluation methods for generation
in the presence of variation. In CICLing 2005.
Pei-Hao Su, David Vandyke, Milica Gasic, Dongho Kim,
Nikola Mrksic, Tsung-Hsien Wen, and Steve J. Young.
2015. Learning from real users: Rating dialogue success with neural networks for reinforcement learning
in spoken dialogue systems. In Interspeech.
Sainbayar Sukhbaatar, Arthur Szlam, Jason Weston, and
Rob Fergus. 2015. End-to-end memory networks. In
NIPS.
Ilya Sutskever, James Martens, and Geoffrey E. Hinton.
2011. Generating text with recurrent neural networks.
In ICML.
Ilya Sutskever, Oriol Vinyals, and Quoc V. Le. 2014.
Sequence to sequence learning with neural networks.
In NIPS.
David R. Traum, 1999. Foundations of Rational Agency,
chapter Speech Acts for Dialogue Agents. Springer.
Oriol Vinyals and Quoc V. Le. 2015. A neural conversational model. In ICML Deep Learning Workshop.
Tsung-Hsien Wen, Milica Gašić, Dongho Kim, Nikola
Mrkšić, Pei-Hao Su, David Vandyke, and Steve
Young. 2015a. Stochastic language generation in dialogue using recurrent neural networks with convolutional sentence reranking. In SIGdial.
Tsung-Hsien Wen, Milica Gašić, Nikola Mrkšić, Pei-Hao
Su, David Vandyke, and Steve Young. 2015b. Semantically conditioned lstm-based natural language generation for spoken dialogue systems. In EMNLP.
Tsung-Hsien Wen, Milica Gašić, Nikola Mrkšić, Pei-Hao
Su, Stefan Ultes, David Vandyke, and Steve Young.
2016a. A network-based end-to-end trainable taskoriented dialogue system. arXiv preprint:1604.04562.
Tsung-Hsien Wen, Milica Gašić, Nikola Mrkšić, Pei-Hao
Su, David Vandyke, and Steve Young. 2016b. Multidomain neural network language generation for spoken dialogue systems. In NAACL-HLT.
Kelvin Xu, Jimmy Ba, Ryan Kiros, Kyunghyun Cho,
Aaron C. Courville, Ruslan Salakhutdinov, Richard S.
Zemel, and Yoshua Bengio. 2015. Show, attend and
tell: Neural image caption generation with visual attention. In ICML.
Pengcheng Yin, Zhengdong Lu, Hang Li, and Ben Kao.
2016. Neural enquirer: Learning to query tables. In
IJCAI.
Steve Young, Milica Gašić, Simon Keizer, François
Mairesse, Jost Schatzmann, Blaise Thomson, and Kai
Yu. 2010. The hidden information state model: A
practical framework for pomdp-based spoken dialogue
management. Computer, Speech and Language.
Steve Young, Milica Gašić, Blaise Thomson, and Jason D. Williams. 2013. Pomdp-based statistical spoken dialog systems: A review. Proceedings of IEEE.
Appendix: More snapshot neuron visualisation
Figure 5: More example visualisation of snapshot neurons and different generated responses.
| 9cs.NE
|
Large deformation and post-failure simulations of segmental retaining walls using
mesh-free method (SPH)
Simulations de grandes déformations et post-rupture des murs de soutènement segmentaires
utilisant la méthode des mailles-libres (SPH)
Bui H.H., Kodikara J.A, Pathegama R., Bouazza A., Haque A.
Department of Civil Engineering, Monash University, Melbourne, Vic 3800, Australia.
ABSTRACT: Numerical methods are extremely useful in gaining insights into the behaviour of reinforced soil retaining walls.
However, traditional numerical approaches such as limit equilibrium or finite element methods are unable to simulate large
deformation and post-failure behaviour of soils and retaining wall blocks in the reinforced soil retaining walls system. To overcome
this limitation, a novel numerical approach is developed aiming to predict accurately the large deformation and post-failure behaviour
of soil and segmental wall blocks. Herein, soil is modelled using an elasto-plastic constitutive model, while segmental wall blocks are
assumed rigid with full degrees of freedom. A soft contact model is proposed to simulate the interaction between soil-block and
block-block. A two dimensional experiment of reinforced soil retaining walls collapse was conducted to verify the numerical results.
It is shown that the proposed method can simulate satisfactory post-failure behaviour of segmental wall blocks in reinforced soil
retaining wall systems. The comparison showed that the proposed method can provide satisfactory agreement with experiments.
RÉSUMÉ : Les méthodes numériques sont extrêmement utiles pour obtenir un aperçu du comportement des murs de soutènement en
sol renforcé. Cependant, les approches numériques traditionnelles tels que l'équilibre limite ou méthodes d'éléments finis sont
incapables de simuler les déformations importantes et le comportement post-rupture des sols et les blocs de béton des murs de
soutènement segmentaires. Pour contourner cette limitation, une nouvelle approche numérique est présentée dans cet article. Le sol
est modélisé à l'aide d'un modèle élasto-plastique, tandis que les blocs segmentaires muraux sont supposés être rigides avec une degré
de liberté total. Un modèle de contact souple a été développé pour modéliser l'interaction entre le sol-bloc et bloc-bloc. Un modèle
expérimentale en deux dimensions d’un effondrement d’un mur de soutènement renforcé a été réalisé pour vérifier les résultats
numériques. Il est montré que la méthode de simulation proposée permet de simuler le comportement post-rupture des blocs de mur
segmentaires des murs de soutènement renforcé. La comparaison a montré que la méthode proposée peut fournir un accord
satisfaisant avec les résultats expérimentaux.
KEYWORDS: retaining wall modelling, segmental walls, large deformation and failure, mesh-free, SPH.
1
modelled by spring and dashpot systems (Cundall & Strack,
1979). The main advantages of this approach are that it can
handle large deformation and failure problems; and the concept
is relatively simple and easy to implement in a computer code.
Thus, DEM could be considered an ideal method to simulate the
full degrees-of-freedom motion of the retaining wall blocks in
SRW systems. However, to model soil behaviour, DEM suffers
from low accuracy because suitable parameters for the contact
model are difficult to determine. The discontinuous deformation
analysis (DDA) method proposed by Shi et al. (1998) has also
been applied to geotechnical applications, but is mainly used for
rock engineering, etc. Other continuum based mesh-free
methods such as the mesh-free Galerkin element method (EFG),
material point method (MPM), particle in cell method (PIC),
etc., could be also applied to simulate large deformation of soil.
However, these methods are quite time consuming and
complicated to implement into a computer code as they consist
of both interpolation points and the background mesh. On the
other hand, the smoothed particle hydrodynamics (SPH) method,
originally proposed by Gingold & Monaghan (1977), has been
recently developed for solving large deformation and postfailure behaviour of geomaterials (Bui et al. 2007-2012; Pastor
et al. 2009, Blanc et al. 2012) and represents a powerful way to
understand and quantify the failure mechanisms of soil in such
challenging problems. In this paper, taking into consideration
this unique advantage, SPH is further extended to simulate large
deformation and post-failure of SRW systems.
INTRODUCTION
In recent years, segmental retaining walls (SRWs) have
received great attentions for their low material cost, short
construction period, ease of construction, and aesthetic
appearance. They have been used as an effective method to
stabilize cuts and fills adjacent to highways, and embankments,
amongst many other applications. Because of the flexible
structural materials used (no mortar, or concrete footing), SRWs
can tolerate minor ground movement and settlement without
causing damage or cracks. In addition, dry stacked SRW
construction allows free draining of water through the wall face,
thereby reducing hydrostatic pressure build-up behind the wall.
Thus far, several analytical and numerical approaches have
been developed to assist SWR design. Among these techniques,
the finite element method (FEM) has been frequently applied to
investigate stability and settlement of segmental retaining wall
systems. FEM has also been used to simulate seismic loadinduced large deformation of SRW systems. However, because
of the mesh-based nature, FEM suffers from mesh entangling
when dealing with large deformation problems, even when the
updated Lagrangian method is adopted. Re-meshing may help
to resolve this problem but the procedure is quite complicated.
It is also worth mentioning that the free rotation motion of
retaining wall blocks in SWR systems could not be modeled by
FEMs. As an alternative for such computational complications,
it is attractive to develop mesh-free methods. So far, the most
popular mesh-free method applied in geotechnical engineering
is the discrete element method (DEM) which tracks the motion
of a large number of particles, with inter-particle contacts
687
Proceedings of the 18th International Conference on Soil Mechanics and Geotechnical Engineering, Paris 2013
2
SIMULATION APPROACHES
2.1
In the SPH method, motion of a continuum is modeled using a
set of moving particles (interpolation points); each assigned a
constant mass and “carries” field variables at the corresponding
location. The continuous fields and their spatial derivatives are
taken to be interpolated from the surrounding particles by a
weighted summation, in which the weights diminish with
distance according to an assumed kernel function. Details of the
interpolation procedure and its application to soil can be found
in Bui et al. (2008). The motion of a continuum can be
described through the following equation,
σ g f ext
u
where I1 and J2 are the first and second invariants of the
stress tensor; and and kc are Drucker-Prager constants that
are calculated from the Coulomb material constants c (cohesion)
and (internal friction). In plane strain, the Drucker-Prager
constants are computed by,
9 12 tan
kc
3c
(6)
9 12 tan 2
(7)
g p I 1 J 2 constant
where is a dilatancy factor that can be related to the
dilatancy angle in a fashion similar to that between and
friction angle . Substituting Equation (7) into Equation (4) in
association with the consistency condition, that is the stress
state must be always located on the yield surface f during the
plastic loading, the stress-strain relation of the current soil
model at particle a can be written as,
d a
2Ga ea K aa a
dt
a 3K a a (Ga / J 2 a ) sa
(2)
(8)
where e is the deviatoric strain-rate tensor; s is the
deviatoric shear stress tensor; and is the rate of change of
plastic multiplier of particle a, which in SPH is specified by
a
3 a K aa (Ga / J 2 a )sa a
(9)
9 a K a a Ga
where the strain-rate tensor is computed by
a
1
u u
2
(10)
a
When considering a large deformation problem, a stress rate
that is invariant with respect to rigid-body rotation must be
employed for the constitutive relations. In the current study, the
Jaumann stress rate is adopted:
a a a a a a
̂
where
(3)
where De is the elastic constitutive tensor; is the strain rate
tensor; and
is its plastic component. An additive
decomposition of the strain rate tensor has been assumed into
elastic and plastic components. The plastic component can be
calculated using the plastic flow rule,
g p
ε p
σ
and
2
The non-associated plastic flow rule specifies the plastic
potential function by,
(1)
where and denote Cartesian components x, y, z with the
Einstein convention applied to repeated indices; a indicates the
particle under consideration; a and b are the densities of
particles a and b respectively; N is the number of “neighbouring
particles”, i.e. those in the support domain of particle a; mb is
the mass of particle b; C is the stabilization term employed to
remove the stress fluctuation and tensile instability (Bui et al.,
2011); W is the kernel function which is taken to be the cubic
Spline function (Monaghan & Lattanzio 2005); and fext a is the
unit external force acting on particle a.
The stress tensor of soil particles in Equation (2) can be
computed using any soil constitutive model developed in the
literature. For the purpose of soil-structure interaction, the
Drucker-Prager model has been chosen with non-associated
flow rule, which was implemented in the SPH framework by
Bui et al. (2008) and shown to be a useful soil model for
simulating large deformation and post-failure behaviour of
aluminum rods used in the current paper as model ground. The
stress-strain relation of this soil model is given by,
σ D e : (ε ε p )
tan
where u is the displacement; a raised dot denotes the time
derivative; is the density; is the total stress tensor, where
negative is assumed for compression; g is the acceleration due
to gravity; and fext is the additional external force(s). The total
stress tensor of soil is normally composed of the effective stress
(´) and the pore-water pressure (pw), and follows Terzaghi’s
concept of effective stress. Because the pore-water pressure is
not considered, the total stress tensor and the effective stress are
identical throughout this paper and can be computed using a
constitutive model.
In the SPH framework, Equation (1) is often discretized
using the following form,
N
a Wab g a f ext
ua mb a2 b2 Cab
a
b 1
b
a
(5)
f I 1 J 2 k c 0
Soil modelling in SPH framework
(4)
where is the rate of change of plastic multiplier, and gp is
the plastic potential function. The plastic deformation occurs
only if the stress state reaches the yield surface. Accordingly,
plastic deformation will occur only if the following yield
criterion is satisfied,
688
a
(11)
is the spin-rate tensor computed by
1
u u
2
(12)
a
As a result, the final form of the stress-strain relationship for
the current soil model is modified to
d a
a a a a 2Ga ea K a a a
dt
a 3K a a (G / J 2 ) a s a
(13)
Equations (2) and (13) are finally integrated using Leapfrog
algorithm to describe the motion of soil medium. Validation of
this soil model with SPH has been extensively documented in
Technical Committee 103 / Comité technique 103
the literature (Bui et al. 2008-2012), and reader can refer to
these references for further details on the validation process.
Rigid body motion of retaining wall blocks
2.2
The segmental retaining wall simulated in this paper is
comprised of individual rectangular blocks; each is assumed as
a rigid body and has complete degrees-of-freedom. The motion
of the block can be determined by specifying the motion of the
central mass and the rotation about its mass central. The
equation of motion of the central mass is given as follows,
M
dV
F
dt
(14)
where M is the central mass, V is the velocity vector of the
central mass, F is total force vector acting on the body. The
equation of rotation about the central mass is,
d
I
T
dt
(15)
where I is the inertial moment, is the angular velocity
which is perpendicular to the plane of the motion, and T is the
total torque about the central mass.
In the computation, the rectangular block is represented by
the set of boundary particles that are equi-spaced around the
boundary. Denoting the force vector acting on each boundary
particle i located on the moving block is fi, Equations (14) and
(15) can be rewritten, respectively, as follows,
dV
fi
dt
i
d
I
k (ri R) f i
dt
(16)
M
dri
V ( ri R )
dt
(18)
n d ai ( ha hi ) / 2
(21)
cn 2 mai K ai
(22)
k cs vais
f asi ai s
0
In this paper, a soft contact model based on a concept of the
spring and dash-pot system is proposed to model the interaction
between soil and retaining wall blocks and between blocks.
Accordingly, the radial force acting between two particles can
be calculated using the following equation,
(19)
where K is the radial stiffness; n is the allowable
overlapping distance between two particles; cn is the radial
damping coefficient; vn is the relative radial velocity vector
between particle a and particle i; ha and hi are the initial
distance (so-called smoothing length in SPH) between soil
particles and between boundary particles, respectively; and dai is
the distance between two particles. The stiffness, overlapping
689
( ha hi ) 2d ai
(ha hi ) 2d ai
(23)
where kai is the shear stiffness; s is the relative displacement
between the two particles in the shear direction; cs is the shear
damping coefficient; vs is the relative shear velocity vector
between particle a and particle i. These unknown variables can
be calculated using the following relationships,
k ai 4Geq heq n
(24)
s vais dt
(25)
cn 2 mai k ai
(26)
where Geq is the equivalent shear modulus. The current shear
force must satisfy Coulomb’s friction law which implies that the
maximum shear force must not exceed the maximum resisting
force,
s n
f a i
s
(27)
Finally, these forces are converted to the conventional
coordinate system and added to Equations (2), (16) and (17) to
progress the motion of soil and rigid bodies.
3
Contact force model
(ha hi ) 2d ai
( ha hi ) 2d ai
(20)
f asi
The force fi acting on a boundary particle on the rigid body is
due to the surrounding soil particles or boundary particles that
belong to different rigid bodies. This force can be calculated
using a suitable contact model which will be described in the
next section.
K ai n3 / 2 cn vakn
f ani
0
K ai Eeff heff / 3
where Eeq and heq are equivalent Young’s modulus and
smoothing length, respectively. The contact force in the shear
direction which is perpendicular to the radial direction can be
calculated in the same manner,
(17)
where ri and R are vector coordinates of boundary particle
and central mass, respectively. The rigid body boundary
particles move as a part of the rigid body, thus the change on
position of boundary particle i is given by,
2.3
distance and damping coefficient can be calculated using the
following relationships,
OUTLINE OF MODEL TEST
Two-dimensional experiments of retaining wall collapse were
conducted to validate the SPH numerical results. Figure 1
shows a schematic diagram of the two-dimensional
experimental setup and Figure 2 shows the initial setup
condition of the model ground and retaining wall blocks in the
experiment. The size of the model ground is 15cm in height and
50cm in width at the base. Aluminum rods of 5cm in length,
having diameters of 1.5 to 3mm and mixed with the ratio 3:2 in
weights, are used as the model ground. The total unit weight of
the model ground is 23kN/m3. The retaining block is 3.2cm in
width, 2.5cm in height, and 5cm in length, which is
manufactured from aluminum (Young’s modulus of 69GPa and
unit weight of 26.5kN/m3). In the experiment, the segmental
retaining wall system was constructed by successively placing
one block on the top of the other with an overlapping of 1.9cm,
followed by filling the model ground at each layer. To visualize
the failure pattern of the model ground, square grids
(2.52.5cm) were drawn on the specimen. The experiments
were initiated by quickly removing the stopper stand and digital
photos were taken to record the failure process as well as the
final configuration of the retaining wall system after collapse.
Proceedings of the 18th International Conference on Soil Mechanics and Geotechnical Engineering, Paris 2013
Model ground
(Aluminium bars)
Retaining blocks
a) Experimental results
Stopper barrier
5cm
Figure 1. Schematic diagram of 2D experimental model.
Model ground
(Aluminium bars)
Wall blocks after collapse
6
Stopper stand
5 4
2 1
3
b) Computed results
5cm
Retaining blocks
Wall blocks after collapse
6
5 4
3
2 1
5cm
6
5
4
3
Figure 3. Comparison between experiment and SPH simulation.
2
1
5
Figure 2. Initial setting condition of the model test.
In addition, tests were conducted to measure static friction
coefficients. It was found that the static friction () between
retaining wall blocks is 0.62, between wall block and the
bottom wall boundary is 0.60, and between the retaining wall
block and model ground is 0.56. The experiment series were
conducted starting from one block and then gradually increasing
the number of blocks in the retaining wall system until the
retaining wall collapsed. It was observed that the retaining wall
system collapsed when reaching 6 blocks. Accordingly, a
numerical model consisting of 6 retaining wall blocks was
conducted for the benchmark study.
4
RESULT AND DISCUSSION
The model test shown in Figure 2 was simulated using 11,304
SPH particles arranged in a rectangular lattice with an initial
separation of 0.25cm. Rigid blocks were created by placing
boundary particles uniformly around the boundary at a constant
distance. In order to simulate the smooth surface, half of particle
spacing was chosen for the rigid body boundary particles.
Model ground parameters including elastic modulus E =1.5MPa,
Poisson’s ratio = 0.3, friction angle = 19.8o, dilatant angle
= 0o, and cohesion c = 0kPa were taken similar to those used in
Bui et al. (2008). The unit weight of the ground model is s =
23kN/m3. The friction coefficients between the rigid blocks,
between the block and the bases of the wall boundary, and
between the block and the ground model were taken to be
similar to those measured in the experiment as mentioned in
Section 3. The boundary conditions for the model ground are
restrained with a free-slip boundary at the lateral boundaries and
fixed in both directions at the base.
Figure 3 shows the comparison between the experiment and
the computation. It can be seen that, because the complete
degrees-of-freedom of the rigid body was taken into
consideration, the computed result could predict fairly well the
behaviour of all rigid blocks observed in the experiment after
the SWR system collapsed. The final run-out distance of Block
No.1 in the simulation is 67.5cm from the left-most solid
boundary. This result is in very good agreement with that
observed in experiment (68cm). It is suggested that the
proposed soft contact model could be applied to simulate the
soil-structure interaction in the SRW system. However, further
refinement of the contact model should be considered to provide
more accurate prediction of the retaining wall blocks and the
model ground as the simulation results showed a slight overprediction of the failure zone observed in the experiment.
690
CONCLUSION AND FUTURE WORKS
This paper presented a novel numerical approach for simulation
of large deformation and post-failure of segmental retaining
wall. It was shown that the proposed method provides good
agreement with the experimental results. The most significant
advantage of the new method is that the complete degrees-offreedom of the retaining wall blocks, which could not be
simulated using traditional numerical approaches, can now be
simulated in the proposed numerical framework. Large
deformation and post-failure behaviour of geomaterials can also
be readily simulated. To broaden the application of the proposed
numerical approach, further implementations such as coupling
with geo-grid reinforcement, modelling seismic earthquake
loading, and bonding between blocks should be considered in
the future. Full extension to three-dimension code would yield
significant benefits to gain further insights into the mechanisms
of SWR.
6
REFERENCES
Bui H.H, Sako K, Fukagawa R. 2007. Numerical simulation of soilwater interaction using smoothed particle hydrodynamics (SPH)
method. Journal of Terramechanics, 44, 339-346.
Bui H.H, Fukagawa R, Sako K. and Ohno S. 2008. Lagrangian meshfree particle method (SPH) for large deformation and post-failure of
geomaterials using elastic-plastic soil constitutive model,
International Journal for Numerical and Analytical Methods in
Geomechanics, 32, 1537-1570.
Bui H.H, Fukagawa R, and Sako K. 2011. Slope stability analysis and
discontinuous slope failure simulation by elasto-plastic smoothed
particle hydrodynamics (SPH). Géotechnique, 61, 565-574.
Bui H.H. and Fukagawa R. 2012. An improvement of SPH for saturated
soils and its application to investigate the mechanisms of
embankment failure: Case of hydrostatic pore-water pressure,
International Journal for Numerical and Analytical Methods in
Geomechanics. doi: 10.1002/nag.1084.
Cundall P.A, Struck ODL. 1979. A discrete numerical model for
granular assemblies. Géotechnique, 29, 47-65.
Gingold R.A., Monaghan J.J. 1977. Smoothed particle hydrodynamics:
Theory and application to non-spherical stars. Mon. Not. Roy.
Astron. Soc., 181: 375-389.
Monaghan J.J. 2003. SPH elastic dynamics. Computer Methods in
Applied Mechanics and Engineering, 190, 6641-6662.
Monaghan J.J, Lattanzio J.C. 1985. A refined particle method for
astrophysical problems. Astronomic & Astrophysics, 149,135-143.
Pastor M., Haddad B., Sorbino G., Cuomo S., Drempetic V. 2009. A
depth-integrated, coupled SPH model for flow-like landslides and
related phenomena. International Journal for Numerical and
Analytical Methods in Geomechanics, 33, 143-172.
Shi G.H. 1988. Discontinuous deformation analysis: a new numerical
model for the static and dynamics of block systems. Ph.D. thesis,
University of California, Berkeley.
| 5cs.CE
|
1
Dynamic Modularity Approach to
Adaptive Inner/Outer Loop Control of
Robotic Systems
arXiv:1603.05557v4 [cs.SY] 7 Jan 2017
Hanlei Wang, Wei Ren, Chien Chern Cheah, Yongchun Xie, and
Shangke Lyu
Abstract—Modern applications of robotics typically involve a robot
control system with an inner PI (proportional-integral) or PID
(proportional-integral-derivative) control loop and an outer user-specified
control loop. The existing outer loop controllers, however, do not take
into consideration the dynamic effects of robots and their effectiveness
relies on the ad hoc assumption that the inner PI or PID control loop
is fast enough, and other torque-based control algorithms cannot be
implemented in robotics with closed architecture. This paper investigates
the adaptive control of robotic systems with an inner/outer loop structure,
taking into full account the effects of the dynamics and the system
uncertainties, and both the task-space control and joint-space control
are considered. We propose a dynamic modularity approach to resolve
this issue, and a class of adaptive outer loop control schemes is proposed
and their role is to dynamically generate the joint velocity (or position)
command for the low-level joint servoing loop. Without relying on the ad
hoc assumption that the joint servoing is fast enough or the modification
of the low-level joint controller structure, we rigorously show that the
proposed outer loop controllers can ensure the stability and convergence
of the closed-loop system. We also propose the outer loop versions of
several standard joint-space direct/composite adaptive controllers for
rigid or flexible-joint robots, and a promising conclusion may be that
most torque-based adaptive controllers for robots can be designed to fit
the inner/outer loop structure by using the new definition of the joint
velocity (or position) command. Simulation results are provided to show
the performance of various adaptive outer loop controllers, using a threeDOF (degree-of-freedom) manipulator, and experiment results using the
UR10 robotic system are also presented.
Index Terms—Inner/outer loop, adaptive control, dynamic modularity
approach, robotics.
I. I NTRODUCTION
User-friendliness is an important aspect of modern automatic
machines, especially if they are expected to do extensive work in
cooperation with human beings. The control systems for modern
robotic systems, unfortunately, have not yet reach this expectation
though numerous control algorithms have been developed over the
past several decades. To serve this purpose, the controlled robotic
system might have to be reliable, robust, and flexible to satisfy the
user’s needs. As we take a deep look at the development of computers
(for instance, the recent hybrid computers—Surface Pro 4, iPad Pro,
etc.), the module design plays a vital role in promoting their success
in our everyday life. The relationship between the operating systems
(e.g., the Windows System) and the application programs further
validates the desirability of this module design philosophy.
Historically, the applications of robotics have undergone the following phases:
1) Traditional industrial applications—factory automation in
a structured environment with a simple joint-space PID
(proportional-integral-derivative) inner loop control as well as
inverse kinematics;
H. Wang and Y. Xie are with the Science and Technology on
Space Intelligent Control Laboratory, Beijing Institute of Control Engineering, Beijing 100094, China (e-mail: [email protected]; [email protected]).
W. Ren is with the Department of Electrical and Computer Engineering,
University of California, Riverside, CA 92521 USA (e-mail: [email protected]).
C. C. Cheah and S. Lyu are with the School of Electrical and Electronic
Engineering, Nanyang Technological University, Singapore 639798, Republic
of Singapore (e-mail: [email protected]; [email protected]).
2) Modern applications—beyond factory automation in an unstructured environment with sensory feedback in the task space,
e.g., (outer loop) visual servoing and task-space control, which,
however, either cannot be implemented in robotics with closed
architecture or rely on the ad hoc assumption that the combination of the inner and outer loops is stable and the effect of the
dynamics can be neglected. This ad hoc assumption typically
(approximately) holds only on the occasion that the given task
is slow enough.
The inner/outer loop structure of robotic systems (e.g., most commercial robotic systems) enjoys certain module design flavor and
has some desirable properties, e.g., it is beneficial for generating a
high joint stiffness by employing a fast inner joint servoing while
it is not reliable as directly specifying the control torque due to the
limitation of the communication (generally required for exerting a
coupling control action) reliability. For this reason, the gap between
the study of advanced robot control theory and practical applications
is longstanding. In the academic field, most advanced controller
designs are torque-based and typically require an open torque control
loop. In practical applications, the much more reliable and robust
velocity control mode is adopted. It seems necessary here to recall the
standard inner/outer loop structure of most commercial (industrial)
robotic systems: 1) an outer (kinematic) loop using a centralized
computer with enough computing power; 2) an inner dynamic loop
that consists of n independent processors equipped at each joint in
a decentralized way (i.e., each processor only receives/sends signals
from/to its local sensors, actuator, and the centralized computer). The
inner dynamic loop usually evolves at a much faster sampling than
the outer (kinematic) loop.
The attempts that aim to address the control of this kind of robotic
systems in the task space occur in, e.g., [1], [2], [3], [4]. However,
these controllers cannot ensure the tracking error convergence without
significantly modifying the low-level PI (proportional-integral) controller to be a more complex one. One may note that these control
algorithms are all based on the standard resolved motion rate control
proposed by [5] to design the joint velocity command. Then what
actually prevents the application of advanced robot controllers? Let
us first retrospect the realization of the standard computed torque
controller that has been discussed in [6, p. 209, p. 210] (see also the
trajectory precorrection based on the computed torque feedforward
in [7]). The specific procedure in [6] is to modify the terms in the
feedforward action that may involve coupling by replacing certain
signals (positions or velocities) with their desired values. In this
way, no communication between the joint processors at the dynamic
servoing loop is required since the desired values of all the joints are
stored in each joint’s computer a priori (which would, on the other
hand, mean that this algorithm is not flexible in the case that the
desired trajectory is subjected to changes). The overall impression
is that for the sake of reducing computational burden, this scheme
performs the feedforward at the joint control loop in a relatively limited manner and with many nonlinear terms being neglected. Another
important well-recognized reason is that most commercial/industrial
robotic systems do not have an open torque control loop (see, e.g.,
[8]). These two factors give rise to the awkward situation of the
modern torque-based robot control algorithms (e.g., the adaptive
algorithms in [9], [10], [11], [12], [13], [14], [15] and the robust
algorithms in [16], [17]), i.e., it is hard to apply these algorithms to
robotic systems with an inner/outer loop structure. The precorrection
scheme in [7], by producing a trajectory correction term based on the
inverse manipulator dynamics and then adding it to the desired joint
trajectory, improves the performance of the industrial robotic system
without modifying the low-level controller structure. The main proofs
2
of the performance improvement, however, are by the experimental
results rather than by the rigorous analysis.
In this paper, we rigorously address this issue in the context
of adaptive task-space/joint-space control for robotic systems with
an embedded low-level PI joint velocity controller (or PID joint
position controller) and with uncertain dynamics (and kinematics).
The use of PI velocity controller or PID position controller in most
industrial/commercial robotic systems is well recognized (see, e.g.,
[18], [19]). Our main purpose here is to develop a class of adaptive
outer loop controllers that can ensure the stability and convergence
of the robotic systems with the dynamic effect being taken into full
account and without modifying the embedded inner PI or PID control
loop. The application of the current adaptive (or robust) task-space
regulation/tracking algorithms (e.g., [20], [21], [12], [13], [15], [22],
[23], [14], [24], [25]) to robotic systems with an inner/outer loop
structure, for a long period, relies on the ad hoc assumption that the
inner joint servoing loop is fast enough or the modification of the
inner joint controller structure. A preliminary version of the paper
was presented in [26] where the simultaneous opening of the position
and velocity commands is required. We here extend this preliminary
result to address the case that only the position (or velocity) command
is designable (which is considered to be much more common in most
industrial/commercial robots), and to additionally consider the case
of composite adaptation and joint flexibility as well as include the
experimental results.
We first propose two adaptive task-space regulation controllers
that rule out the fundamental limitations of the existing results, by
dynamically incorporating an adaptively scaled dynamic compensation that exploits the physically independent nature of the lowlevel controller structure. The first controller, by introducing an
adaptive filter, avoids the task-space velocity measurement, and the
second one avoids the use of the task-space velocity by using an
observer [motivated by the one in [27] with a modified feedback
gain (which depends on the estimated Jacobian matrix) to achieve
feedback separation]. Both of the adaptive controllers are qualified
outer loop control schemes that can be applied to robotic systems
with an unmodifiable joint servoing controller (PI velocity or PID
position controller) (e.g., most industrial/commercial robots), taking
into account the dynamic loop of the robotic systems. From a
robot control perspective, most existing kinematic algorithms are
not mathematically rigorous in that either the effects of the inner
joint control loop are not considered (e.g., [28], [29], [30]), or the
low-level joint servoing controller is assumed to be strong enough
to ensure (yet cannot rigorously guarantee due to the absence of
dynamic compensation action in the low-level dynamic loop) the
square-integrability and boundedess of the velocity tracking error
(e.g., [31], [32]), or even the modification of the low-level controller
structure (e.g., [1], [2], [3]). These limitations are mainly caused
by the inner/outer loop structure and closed controller architecture.
The proposed outer loop controller here, by dynamically incorporating adaptively scaled dynamic compensation action and adaptive
transpose Jacobian feedback, ensures the singularity-robust stability
and convergence of the task-space position error without relying on
any modification of the low-level PI/PID controller structure. Due to
the independence of the design of the outer loop controller and that
of the low-level PI/PID controller and the injection of the dynamic
compensation, the proposed design approach is referred to as dynamic
modularity approach.
We then show that the observer-based task-space regulation scheme
can be extended to the case of task-space tracking. The obtained
control scheme has an interesting feature that the inverse of the
estimated Jacobian matrix is used for introducing feedforward and
its transpose is used for introducing feedback, in contrast to most
.LQHPDWLF
SODQQHU
qc qc
3,YHORFLW\ 3,'
SRVLWLRQ FRQWUROOHU
0DQLSXODWRU
G\QDPLFV
,QQHU ORRS
q
q
2XWHU ORRS
x
0DQLSXODWRU
NLQHPDWLFV
Fig. 1. Inner/outer loop control (qc is the position command, q̇c is the
velocity command, and q, q̇, and x are the joint position, joint velocity, and
task-space position, respectively).
existing task-space algorithms that only employ the inverse of the
Jacobian matrix to exert feedforward and feedback actions (see, e.g.,
[9], [33], [21], [14]), and the benefit of this lies in two folds: 1) it
yields the feedback separation of the kinematic and dynamic loops,
thus reducing the activity of the dynamic compensation action; 2) it
is reducible in the sense that once the desired task-space velocity
becomes zero, the tracking control law reduces to the regulation
control law without involving the inverse of the estimated Jacobian
matrix.
Finally, we illustrate how the adaptively scaled dynamic compensation enables several typical adaptive robot controllers for jointspace trajectory tracking to be applicable to robotic systems with an
inner/outer loop structure (e.g., most industrial/commercial robotic
systems). In particular, the proposed dynamic modularity approach is
further shaped to incorporate the composite adaptation for improving
performance and to address the issue of joint flexibility. By these
additional examples, it seems hopeful that most adaptive dynamic
controllers for robots in the literature with the use of adaptively
scaled dynamic compensation and new definition of the joint velocity
(or position) command would be rendered to be qualified outer loop
schemes. Another potential favorable point may be the reduction of
the cost of the laboratory experimental research on advanced robot
control theory in that there no longer needs to develop a specific
manipulator with an open torque control loop (usually requiring
extensive efforts and time) and any commercial robot (cost-efficient
due to the large-scale production) can be directly used.
II. P RELIMINARIES
A. Background and Motivation
Inner/outer loop control is typical in modern applications of robotic
systems (see Fig. 1), and generally the inner loop is designed by the
robot production company and closed and only the outer loop is open
to the user. The user can specify the position or velocity command
within the outer loop based on measurements in joint space [and
task space (e.g., image space)], which is then sent to the inner loop
as a reference signal. The main benefit of adopting an inner/outer
loop structure may be that the inner loop (due to simplicity) can
be operated at a high sampling rate and thus a high stiffness can be
maintained, and that the outer loop can be operated at a low sampling
rate allowing relatively complicated communication and sensing. In
addition, inner/outer loop structure does help to realize the relative
independence of the manipulator production company and the users,
and to promote the large-scale production of manipulators due to the
invariance of the inner control loop.
Historically, most theoretical results on adaptive manipulator control are presented in the context that the joint torque is directly
designable rather than in the framework of the inner/outer loop. This
results in the longstanding gap between the study of advanced robot
control theory and the applications of robots. The existing results
(e.g., [1], [2], [3], [7], [28], [29], [30]) are either ad hoc, e.g., effective
3
under the assumption that the joint servoing is fast enough, relying
on the modification of the low-level controller structure which is
unmodifiable in practice, or not theoretically rigorous.
Our main purpose is to develop a class of adaptive outer loop
controllers using the task-space and joint-space sensory measurement
(instead of obtaining the task-space information based on the kinematics as in Fig. 1 since the kinematics is unknown) to take place of
the kinematic planner in Fig. 1 so that the stability and convergence
of the closed-loop robotic systems with uncertain dynamics (and
kinematics) can rigorously be ensured, and to finally approach the
goal of “modularity”.
B. Manipulator Kinematics and Dynamics
Consider an n-DOF (degree-of-freedom) manipulator actuated by
permanent magnet DC motors. Let x ∈ Rm be the position of the
end-effector in the task space and it is relevant to the joint position
by the following nonlinear mapping [34], [35]
x = f (q)
(1)
where q ∈ Rn denotes the joint position and f : Rn → Rm is the
mapping from joint space to task space. We here assume that n ≥ m,
i.e., the manipulator can either be nonredundant or redundant.
Differentiating (1) with respect to time yields the relation between
the task-space velocity ẋ and joint velocity q̇ [34], [35]
ẋ = J(q)q̇
III. A DAPTIVE I NNER /O UTER L OOP C ONTROL
Kinematic control typically appears in the context of inner/outer
loop control of robot manipulators and its focus is on the design of the
joint velocity (or position) command. Historically, the effectiveness
of kinematic control stands on the relatively strong assumption that
the inner control (PI velocity control or PID position control) loop
is fast enough so that the dynamic effects of the inner loop can be
neglected. Here, we present a dynamic modularity approach to ensure
the convergence of the task-space position error, without relying on
this ad hoc assumption or the modification of the inner control loop.
Let xd ∈ Rm denote the desired task-space position. For the
regulation problem, it is set as constant; for the tracking problem,
it can be set as time-varying and in this case, we assume that xd , ẋd ,
and ẍd are all bounded.
A. Filter-Based Adaptive Regulation Control
Consider the case that the inner control loop employs a PI velocity
control, and the PI gains of the inner PI velocity control loop (the
case of PID position control is discussed in Sec. III-D) are supposed
to be KP and KI (diagonal and positive definite). Then, the PI action
can be written as
u = −KP (q̇ − q̇c ) − KI (q − qc ) .
To avoid the task-space velocity measurement, we introduce the
following adaptive passive filter
(2)
ẏ = −K1 y + K1 JˆT (q)∆x,
m×n
where J(q) ∈ R
is the Jacobian matrix. If the kinematic
parameters are unknown, the task-space position/velocity can no
longer be derived by the direct kinematics given above. The typical
practice in this case is to employ certain task-space sensors (e.g., a
camera) to obtain the task-space position information.
The dynamics of the manipulator can be written as [35]
M (q)q̈ + C(q, q̇)q̇ + B q̇ + g(q) = Ku
(3)
where M (q) ∈ Rn×n is the inertia matrix, C(q, q̇) ∈ Rn×n is the
Coriolis and centrifugal matrix, B ∈ Rn×n is a constant diagonal
positive definite matrix, g(q) ∈ Rn is the gravitational torque, u ∈
Rn is the armature voltage, and K ∈ Rn×n is a constant diagonal
positive definite matrix.
Four basic properties associated with (2) and (3) that shall be useful
for the controller design and stability analysis are listed as follows.
Property 1 ([21]): The kinematics (2) depends linearly on a
constant kinematic parameter vector ak , which gives rise to
J(q)ψ = Yk (q, ψ)ak
M (q)ζ̇ + C(q, q̇)ζ + Bζ + g(q) = Yd (q, q̇, ζ, ζ̇)ad
(5)
where ζ ∈ Rn is a differentiable vector, ζ̇ is the time derivative of
ζ, and Yd (q, q̇, ζ, ζ̇) is the dynamic regressor matrix.
(7)
upon which, we define the joint reference velocity as
q̇r = −K2 y
(8)
where ∆x = x − xd , K1 and K2 are diagonal positive definite
matrices, and Jˆ(q) is the estimate of the Jacobian matrix J(q) and
is obtained by replacing ak in J(q) with its estimate âk . Define a
sliding vector
s = q̇ − q̇r .
(9)
Substituting (8) into (2) and using Property 1 gives
ˆ q̇r + J(q)
ˆ q̇
ẋ = − Jˆ(q)K2 y − [Jˆ(q) − J(q)]q̇ − J(q)
ˆ
ˆ
= − J (q)K2 y − Yk (q, q̇)∆ak + J (q)s.
(10)
where ∆ak = âk − ak .
We define the joint velocity command by the following dynamic
system
(4)
where ψ ∈ Rn is a vector and Yk (q, ψ) is the kinematic regressor
matrix.
Property 2 ([36], [35]): The inertia matrix M (q) is symmetric and
uniformly positive definite.
Property 3 ([36], [35]): The matrix C(q, q̇) can be appropriately
defined such that the matrix Ṁ (q) − 2C(q, q̇) is skew-symmetric.
Property 4 ([36], [35]): The dynamics (3) depends linearly on a
constant dynamic parameter vector ad , which yields
(6)
with
q̇c + K̂I qc =q̇r + K̂I qr + diag[ŵi , i = 1, . . . , n]
h
i
× −αJˆT (q)∆x + Yd (q, q̇, q̇r , q̈r )âd
qr =qr (0) +
Z
(11)
t
q̇r (σ)dσ
(12)
0
where qr (0) can be chosen as an arbitrary constant vector, α is a
positive design constant, âd is the estimate of ad , ŵi denotes the
scale weight, i = 1, . . . , n, diag[ŵi , i = 1, . . . , n][−αJˆT (q)∆x +
Yd (q, q̇, q̇r , q̈r )âd ] denotes the adaptively scaled dynamic compensation action, and K̂I denotes the estimate of KI = KP−1 KI , which
can be expressed as
K̂I = diag[ŵI ]
(13)
4
with ŵI being an n-dimensional vector. The adaptation laws for âk ,
ŵ = [ŵ1 , . . . , ŵn ]T , âd , and ŵI are given as
â˙ k = Γk YkT (q, q̇)∆x
ŵ˙ = −Λdiag[−αJˆT (q)∆x + Yd (q, q̇, q̇r , q̈r )âd ]s
â˙ d = −Γd YdT (q, q̇, q̇r , q̈r )s
ŵ˙ I = ΛI diag[qc − qr ]s
(14)
(15)
(16)
(17)
where Λ and ΛI are diagonal positive definite matrices, and Γk and
Γd are symmetric positive definite matrices.
ˆ
Theorem 1: Suppose that K̂I is uniformly positive definite and J(q)
has full row rank. Then, the adaptive outer loop controller given by
(11), (12), (14), (15), (16), and (17) for the robotic system (2) and (3)
under the inner PI controller (6) ensures the stability of the system
and convergence of the task-space position error, i.e., ∆x → 0 as
t → ∞.
Proof: Substituting (6) and (11) into the manipulator dynamics (3)
and taking into account Property 4 gives
M (q)ṡ + C(q, q̇)s
3)
V̇ = −αy T K2 y − sT (K ∗ + B)s ≤ 0.
(20)
Rt
Then we obtain that y ∈ L2 ∩ L∞ , s ∈ L2 ∩ L∞ , 0 s(σ)dσ ∈ L∞
∆x ∈ L∞ , ŵ ∈ L∞ , âk ∈ L∞ , âd ∈ L∞ , and ŵI ∈ L∞ . From
(7), we obtain that ẏ ∈ L∞ and thus y is uniformly continuous.
From the properties of square-integrable and uniformly continuous
functions [37, p. 232], we obtain that y → 0 as t → ∞. From (8),
we know that q̇r ∈ L∞ and thus q̇ ∈ L∞ . From (14), we obtain
ˆ˙
that â˙ k ∈ L∞ , giving rise to the boundedness of J(q).
From (2), we
obtain that ẋ ∈ L∞ , and we then obtain that ÿ ∈ L∞ based on (7).
This means that ẏ is uniformly continuous and thus ẏ → 0 as t → ∞
according to Barbalat’s Lemma [36]. From (7), we then obtain that
JˆT (q)∆x → 0 as t → ∞. This means that ∆x → 0 as t → ∞
ˆ
since J(q)
has full row rank. Furthermore, the result that ẏ ∈ L∞
yields the conclusion that q̈r ∈ L∞ . Rewrite (11) as
(q̇c − q̇r ) + K̂I (qc − qr )
h
i
= diag[ŵi , i = 1, . . . , n] −αJˆT (q)∆x + Yd (q, q̇, q̇r , q̈r )âd ,
and it can be directly shown that qc − qr ∈ L∞ and q̇c − q̇r ∈ L∞
since K̂I is uniformly positive definite. We then obtain that q̇c ∈ L∞ .
From (18), we obtain that ṡ ∈ L∞ using Property 2, and further that
− K ∗ ∆KI qc − Yd (q, q̇, q̇r , q̈r )ad
q̈ ∈ L∞ . Hence s is uniformly continuous, yielding the result that
= − Bs − K ∗ (q̇ + KI q) + K ∗ q̇r + K̂I qr
s → 0 as t → ∞ according to the properties of square-integrable
h
i and uniformly continuous functions [37, p. 232]. This immediately
∗
T
+ K diag[ŵi , i = 1, . . . , n] −αJˆ (q)∆x + Yd (q, q̇, q̇r , q̈r )âd
gives the conclusion that q̇ → 0 as t → ∞ since q̇r = −K2 y → 0
as t → ∞.
− K ∗ ∆KI qc − Yd (q, q̇, q̇r , q̈r )ad
∗
∗
Remark
1:
The
introduction
of
the
adaptive
scales
ŵ
and
ŵ
I
= − Bs − K (q̇ + KI q) + K (q̇r + KI qr )
h
i is to accommodate the uncertain diagonal matrices K ∗ = KKP
+ K ∗ diag[ŵi , i = 1, . . . , n] −αJˆT (q)∆x + Yd (q, q̇, q̇r , q̈r )âd
and KP−1 KI , and their uncertainty comes from both the actuator
∗
model and low-level controller design (generally performed by the
− K ∆KI (qc − qr ) − Yd (q, q̇, q̇r , q̈r )ad
Z t
robot production company). The part due to the actuator model (i.e.,
K) is inherently uncertain and may possibly be subjected to slow
q̇r (σ)dσ + δ0
= − (K ∗ + B)s − KKI q − q(0) −
0
variation. The uncertainty of the part due to the low-level controller
+ K ∗ diag[ŵi , i = 1, . . . , n][−αJˆT (q)∆x + Yd (q, q̇, q̇r , q̈r )âd ]
design (i.e., KP and KI ) is a business strategy for protecting the
intellectual property right and thus it is and will be impossible to be
− Yd (q, q̇, q̇r , q̈r )ad
Z t
disclosed thoroughly in the short run. In addition, the adaptive scale
s(σ)dσ + δ0
= − (K ∗ + B)s − αJˆT (q)∆x − KKI
ŵ used here is computationally efficient and free of computational
0
singularity since it does not involve the computation of inverse of
∗
∗−1
+ K (diag[ŵi − ki , i = 1, . . . , n])
an estimated quantity (which, however, would be encountered if we
× [−αJˆT (q)∆x + Yd (q, q̇, q̇r , q̈r )âd ]
directly estimate the matrix K ∗ ). Similar techniques for handling the
∗
uncertainty of the diagonal torque-constant matrix (which describes
− K ∆KI (qc − qr ) + Yd (q, q̇, q̇r , q̈r )∆ad
Z t
the relation between the torque and current) appear in the context
s(σ)dσ + δ0
= − (K ∗ + B)s − αJˆT (q)∆x − KKI
of adaptive control for rigid-link electrically-driven robots or robots
0
with actuator uncertainty, yet with an open controller structure (i.e.,
+ diag[−αJˆT (q)∆x + Yd (q, q̇, q̇r , q̈r )âd ]K ∗ ∆w
the voltage can be directly specified by the user) [38], [11].
Remark 2: The uniform positive definiteness of K̂P can be con− K ∗ diag[qc − qr ]∆wI + Yd (q, q̇, q̇r , q̈r )∆ad
(18)
veniently
ensured by using the projection algorithms [39], and the
where δ0 = q(0) − qr (0) is a constant vector, K ∗ = KKP = full row rank of Jˆ(q) can be ensured by the assumption of being
∗
∗
diag[kii
, i = 1, . . . , n] with kii
, i = 1, . . . , n being positive away from the singular configuration and the use of the projection
∗−1
∗−1 T
, . . . , ŵn − knn
] , algorithms [21], [13].
constants, ∆ad = âd − ad , ∆w = [ŵ1 − k11
and ∆wI = ŵI − wI with the entries of wI being from the diagonal
Remark 3: The adaptive filter (7) with K1 JˆT (q)∆x as the input
entries of KI . Consider the following Lyapunov function candidate
can be considered as an extension of [40] to address the avoidance of
1
task-space velocity measurement in the context of adaptive task-space
1 T
1
T
−1
T −1
V =α ∆x ∆x + y K2 K1 y + ∆ak Γk ∆ak
control. Clearly, the joint velocity command given by (11) no longer
2
2
2
Z t
T
Z t
involves the task-space velocity measurement since q̈r = −K2 ẏ and
1
1
+ sT M (q)s +
s(σ)dσ + δ0 ẏ given by (7) does not involve the task-space velocity.
s(σ)dσ + δ0 KKI
2
2
= − Bs − K ∗ (q̇ + KI q) + K ∗ q̇c + K̂I qc
0
0
1
1
1
∗
∆wIT Λ−1
+ ∆wT Λ−1 K ∗ ∆w + ∆aTd Γ−1
I K ∆wI
d ∆ad +
2
2
2
(19)
whose derivative with respect to time along the trajectories of (18),
(7), (10), (14), (15), (16), and (17) can be written as (using Property
B. Observer-Based Adaptive Regulation Control
The task-space observer is given as
ẋo = Jˆ(q)q̇r − β Jˆ(q)JˆT (q)∆xo
(21)
5
where xo ∈ Rm denotes the observed quantity of x, ∆xo = xo − x,
and β is a positive design constant, and the joint reference velocity
q̇r is now defined as
ˆT
q̇r = −J (q) [γ(xo − xd )]
(22)
where γ is a positive design constant. The observer (21) is motivated
ˆ JˆT (q) for the purpose of
by [27] yet with a new feedback gain β J(q)
achieving feedback separation, and the idea behind is that the actual
joint velocity finally approaches the joint reference velocity q̇r and
ˆ q̇r would approach the estimated task-space velocity Jˆ(q)q̇.
thus J(q)
The desirable point is that the observer no longer depends on the joint
velocity and is thus not influenced by the noise of the velocity signal.
The joint velocity command is defined as
q̇c + K̂I qc =q̇r + K̂I qr + diag[ŵi , i = 1, . . . , n]
× Yd (q, q̇, q̇r , q̈r )âd
(23)
which no longer needs the scaled dynamic compensation action
diag[ŵi , i = 1, . . . , n][−αJˆT (q)∆x]. The adaptation laws for âk
and ŵ are given as
â˙ k = Γk YkT (q, q̇)(∆x − ∆xo )
ŵ˙ = −Λdiag[Yd (q, q̇, q̇r , q̈r )âd ]s
(24)
ˆ JˆT (q)∆xo + Yk (q, q̇)∆ak − J(q)s.
ˆ
∆ẋo = −β J(q)
(26)
ˆ JˆT (q) [γ(xo − xd )] + Yk (q, q̇)∆ak .
Jˆ(q)s =ẋ + J(q)
(27)
ˆ
By premultiplying s = q̇ − q̇r with J(q)
and using Property 1, we
obtain
Therefore, at the outer loop, we obtain
ˆ
∆ẋo = − β Jˆ(q)JˆT (q)∆xo + Yk (q, q̇)∆ak − J(q)s
(28)
T
ẋ = − Jˆ(q)Jˆ (q) [γ(xo − xd )] − Yk (q, q̇)∆ak + Jˆ(q)s. (29)
We are presently ready to formulate the following theorem.
Theorem 2: Suppose that K̂I is uniformly positive definite and
ˆ
J(q)
has full row rank, and let the controller parameters β and γ be
chosen such that
β > 4γ/9.
(30)
The adaptive outer loop controller given by (23), (12), (21), (24), (25),
(16), and (17) with q̇r being given as (22) for the robotic system (2)
and (3) under the inner PI controller (6) ensures the stability of the
system and convergence of the task-space position error, i.e., ∆x → 0
as t → ∞.
Proof: Substituting (6), (23), and (12) into (3) and using Property
4 gives
Z t
s(σ)dσ + δ0
M (q)ṡ + C(q, q̇)s = − (K ∗ + B)s − KKI
0
+ diag[Yd (q, q̇, q̇r , q̈r )âd ]K ∗ ∆w
+ Yd (q, q̇, q̇r , q̈r )∆ad .
Consider the Lyapunov-like function candidate
T
Z t
1
1
V ∗ = sT M (q)s +
s(σ)dσ + δ0 KKI
2
2 0
Z t
1
s(σ)dσ + δ0 + ∆wT Λ−1 K ∗ ∆w
×
2
0
1
1
∗
+ ∆aTd Γ−1
∆wIT Λ−1
I K ∆wI
d ∆ad +
2
2
V̇ ∗ = −sT (K ∗ + B)s ≤ 0
(31)
(32)
(33)
where we have used RProperty 3. This directly gives the conclusion
t
that s ∈ L2 ∩ L∞ , 0 s(r)dr ∈ L∞ , ŵ ∈ L∞ , âd ∈ L∞ , and
ŵ
R tI T∈ L∞ . Then, there exists a positive constant ℓM such that
s (σ)s(σ)dσ ≤ ℓM , ∀t ≥ 0. Let us now consider the following
0
quasi-Lyapunov function candidate
1
1
1
V ∗∗ = ∆xTo ∆xo + ∆xT ∆x + ∆aTk Γ−1
k ∆ak
2
2
h
Z t 2
i
1
1
+
+
ℓM −
sT (σ)s(σ)dσ
β
γ
0
(34)
with the choice of the last term following the typical practice (see,
e.g., [41, p. 118]), and the derivative of V ∗∗ along the trajectories of
(28), (29), and (24) can be written as
V̇ ∗∗ = − β∆xTo Jˆ(q)JˆT (q)∆xo
ˆ
ˆ JˆT (q)∆xo
− ∆xTo J(q)s
− γ∆xT J(q)
ˆ JˆT (q)∆x + ∆xT Jˆ(q)s −
− γ∆xT J(q)
(25)
The adaptation laws for âd and ŵI remain the same as (16) and (17),
respectively.
Combining (21) and (2) and using Property 1 yields
− K ∗ diag[qc − qr ]∆wI
whose derivative with respect to time along the trajectories of (31),
(25), (16), and (17) can be written as
1
1
+
β
γ
sT s.
(35)
Using the following results derived from the standard basic inequalities
β
ˆ JˆT (q)∆xo + 1 sT s
ˆ
(36)
∆xTo J(q)s
≤ ∆xTo J(q)
4
β
1
γ
ˆ
(37)
∆xT J(q)s
≤ ∆xT Jˆ(q)JˆT (q)∆x + sT s
4
γ
we obtain from (35) that
3β
ˆ JˆT (q)∆xo
∆xTo J(q)
4
3γ
ˆ JˆT (q)∆x
− γ∆xT Jˆ(q)JˆT (q)∆xo −
∆xT J(q)
4
T
T
T
Jˆ (q)∆xo
(3β/4)In (γ/2)In
Jˆ (q)∆xo
≤ − ˆT
≤0
T
(γ/2)In (3γ/4)In
J (q)∆x
Jˆ (q)∆x
{z
}
|
V̇ ∗∗ ≤ −
Q
(38)
due to the positive definiteness of Q in the case that β > 4γ/9,
where In is the n × n identity matrix. Then using similar procedures
as in the proof of Theorem 1, we can show the stability of the
ˆ
system [regardless of the estimated Jacobian matrix J(q)]
and the
convergence of ∆x and ∆xo .
Remark 4: The existing adaptive (or robust) task-space regulation
algorithms either assume the exact knowledge of the gravitational
torques [21], [12], or require the careful choice of the controller
parameters [20], [42], [22], or encounter the overparametrization
problem [22]. Furthermore, many adaptive visual tracking schemes
(e.g., [21], [24], [14], [25]) have also been proposed, yet the necessity
of investigating task-space regulation algorithms is due to the consideration that the choice of specific controllers should take into account
the properties of the specific tasks. It is well accepted that given a
specific task, the control law should be as (computationally) simple
as possible; while the adaptive tracking controllers can also achieve
the regulation of the task-space position to the desired one (constant),
it is not cost-effective to rely on such kind of complexity (usually
involves the inverse of the estimated Jacobian and the singularity
issues) for regulation tasks. The two adaptive regulation controllers
presented here rule out the limitations of the above results, and in
addition the proposed controllers can be applied to robotic systems
with an inner/outer loop structure (e.g., most industrial/commercial
6
robotic systems) that have an unmodifiable joint servoing controller
but admit the design of the joint velocity (or position) command,
benefiting from the dynamic feedback design and the use of the
adaptively scaled dynamic compensation action.
need to make some modifications. We take the filter-based adaptive
regulation control in Sec. III-A as an illustrating example and the
other controllers can be similarly formulated. Specifically, we define
two quantities below
C. Extension to Task-Space Tracking
q̇r∗ =q̇r − Kc (q − qr )
In the case of task-space tracking, a feedforward action needs to be
introduced in the definition of the joint reference velocity. Specifically
we define q̇r as
q̇r = JˆT (q)[Jˆ(q)JˆT (q)]−1 ẋd −γ JˆT (q)(xo − xd )
|
{z
}|
{z
}
feeforward
=q̈r − Kc (q̇ − q̇r )
(44)
with qr being generated by
ˆ
where the use of the generalized inverse of J(q)
follows the typical
practice. The definition given by (39) extends the one in [31]
to address the case of no task-space velocity measurement. The
ˆ
interesting point here is that both the inverse and transpose of J(q)
are incorporated, and one is for introducing a feedforward action
and the other for introducing a feedback action. This is in contrast
to most existing task-space control algorithms that rely on the use
of the inverse of the (estimated) Jacobian matrix to exert both the
feedforward and feedback actions (see, e.g., [9], [33], [14]). As can
be clearly observed, once the desired task-space velocity becomes
zero, q̇r in (39) reduces to the one defined by (22), and this means
that the regulation and tracking cases are unified.
Theorem 3: Suppose that K̂I is uniformly positive definite and
ˆ
J(q)
has full row rank, and let the controller parameters β and γ be
chosen such that
β > 4γ/9.
(40)
The adaptive outer loop controller given by (23), (12), (21), (24),
(25), (16), and (17) with q̇r being given as (39) for the robotic system
(2) and (3) under the inner PI controller (6) ensures the stability of
the system and convergence of the task-space tracking errors, i.e.,
∆x → 0 and ∆ẋ → 0 as t → ∞.
The proof of Theorem 3 can be straightforwardly completed based
on that of Theorem 2, and the major difference lies in the fact that
equation (29) for the case of regulation problem now becomes
ˆ J (q) [γ(xo − xd )] − Yk (q, q̇)∆ak + Jˆ(q)s.
∆ẋ = −J(q)
(43)
(39)
feedback
ˆT
q̈r∗
q̇r = −Kc qr − K2 y + Kc q
−1
where Kc is a diagonal positive definite matrix. Let KP = KD
KP
−1
and KI = KD KI and denote by K̂P and K̂I the estimate of KP
and that of KI , respectively, which are specifically written as
K̂P =diag[ŵP ]
D. Task-Space Adaptive Control With an Inner PID Position Controller
We here investigate another case that the low-level controller takes
the PID position control action, i.e.,
Z t
u = −KD (q̇ − q̇c ) − KP (q − qc ) − KI
[q(σ) − qc (σ)]dσ (42)
0
where KD , KP , and KI are the derivative, proportional, and integral
gains (diagonal and positive definite), respectively. In this case,
to ensure the stability and convergence of the robotic system, we
(46)
K̂I =diag[ŵI ]
(47)
with ŵP and ŵI being n-dimensional vectors. The joint velocity
command for the low-level PID controller is defined as
q̇c + K̂P qc + K̂I
Z
t
0
[qc (σ) − qr (σ)] dσ
=q̇r∗ + K̂P qr + diag[ŵi , i = 1, . . . , n]
h
i
× −αJˆT (q)∆x + Yd (q, q̇, q̇r∗ , q̈r∗ )âd .
(41)
By an analysis of the system given by (28), (41), and (24), we can
derive the stability of the system and convergence of the task-space
tracking errors.
Remark 5: The filter-based adaptive regulation algorithm is computationally simpler in comparison with the observer-based one. But
the main issue of the filter-based algorithm is that it is difficult
to quantitatively evaluate the performance. Here the observer-based
algorithm is extended to cover the case of task-space tracking
by additionally introducing feedforward based on the generalized
inverse of the estimated Jacobian matrix [see (39)]. The extension
of the filter-based algorithm to realize the task-space tracking can be
completed in a similar way.
(45)
(48)
The adaptation laws for ŵ, âd , ŵP , and ŵI are now given as
ŵ˙ = −Λdiag[−αJˆT (q)∆x + Yd (q, q̇, q̇r∗ , q̈r∗ )âd ]ξ
â˙ d = −Γd YdT (q, q̇, q̇r∗ , q̈r∗ )ξ
ŵ˙ P = ΛP diag[qc − qr ]ξ
Z t
ŵ˙ I = ΛI diag
[qc (σ) − qr (σ)]dσ ξ
(49)
(50)
(51)
(52)
0
with
ξ = q̇ −
q̇r∗
= s + Kc
Z
t
s(σ)dσ + δ0 ,
0
(53)
and ΛP and ΛI being diagonal positive definite matrices, and the
adaptation law for âk is still the same as (14).
With these modifications and using the fact that q − qr =
7
Rt
0
inite. The derivative of V can be written as
s(σ)dσ + δ0 , equation (18) becomes
V̇ = − αy T K2 y − sT (K̄ ∗ + B)s
Z
T
Z t
−
s(σ)dσ + δ0 KMKc
M (q)ξ̇ + C(q, q̇)ξ
= − (K̄ ∗ + B)ξ − αJˆT (q)∆x
Z t
− KKP (q − qr ) − KKI
[q(σ) − qr (σ)]dσ
0
t
0
s(σ)dσ + δ0 ≤ 0.
(58)
0
Theorem 4: Suppose that K̂P and K̂I evolve such that the following
+ diag[−αJˆT (q)∆x + Yd (q, q̇, q̇r∗ , q̈r∗ )âd ]K̄ ∗ ∆w
Z t
system
[qc (σ) − qr (σ)]dσ ∆wI
− K̄ ∗ diag[qc − qr ]∆wP − K̄ ∗ diag
z̈ + K̂P ż + K̂I z = 0
(59)
0
+ Yd (q, q̇, q̇r∗ , q̈r∗ )∆ad
= − (K̄ ∗ + B)ξ − αJˆT (q)∆x − (KKP − KKI Kc−1 )(q − qr )
Z t
[q(σ) − qr (σ)]dσ
− KKI Kc−1 q − qr + Kc
0
+ diag[−αJˆT (q)∆x + Yd (q, q̇, q̇r∗ , q̈r∗ )âd ]K̄ ∗ ∆w
Z t
− K̄ ∗ diag[qc − qr ]∆wP − K̄ ∗ diag
[qc (σ) − qr (σ)]dσ ∆wI
0
+ Yd (q, q̇, q̇r∗ , q̈r∗ )∆ad
= − (K̄ ∗ + B)ξ − αJˆT (q)∆x
Z t
s(σ)dσ + δ0
− (KKP − KKI Kc−1 )
0
Z t
−1
ξ(σ)dσ + δ0
− KKI Kc
0
+ diag[−αJˆT (q)∆x + Yd (q, q̇, q̇r∗ , q̈r∗ )âd ]K̄ ∗ ∆w
Z t
[qc (σ) − qr (σ)]dσ ∆wI
− K̄ ∗ diag[qc − qr ]∆wP − K̄ ∗ diag
0
+ Yd (q, q̇, q̇r∗ , q̈r∗ )∆ad
(54)
where K̄ ∗ = KKD . The above equation can further be written as
M (q)ξ̇ + C(q, q̇)ξ
=diag[−αJˆT (q)∆x + Yd (q, q̇, q̇r∗ , q̈r∗ )âd ]K̄ ∗ ∆w
Z t
[qc (σ) − qr (σ)]dσ ∆wI
− K̄ ∗ diag[qc − qr ]∆wP − K̄ ∗ diag
0
+ Yd (q, q̇, q̇r∗ , q̈r∗ )∆ad − αJˆT (q)∆x
h
− K (KD + K −1 B)s
+ [(KD + K −1 B)Kc + (KP − KI Kc−1 )]
Z t
−1
ξ(σ)dσ + δ0 .
− KKI Kc
Z
t
s(σ)dσ + δ0
0
i
(55)
0
Consider the Lyapunov function candidate
1
1
1
V =α ∆xT ∆x + y T K2 K1−1 y + ∆aTk Γ−1
k ∆ak
2
2
2
Z t
T
Z t
1
1 T
ξ(σ)dσ + δ0
ξ(σ)dσ + δ0 KKI
+ ξ M (q)ξ +
2
2 0
0
T
Z t
1
+
s(σ)dσ + δ0 (KM + KKc KD + Kc B)
2
Z t 0
1
1
×
s(σ)dσ + δ0 + ∆wT Λ−1 K̄ ∗ ∆w + ∆aTd Γ−1
d ∆ad
2
2
0
1
1
∗
∗
∆wIT Λ−1
(56)
+ ∆wPT Λ−1
P K̄ ∆wP +
I K̄ ∆wI
2
2
where
M = (KD + K −1 B)Kc + KP − KI Kc−1 .
(57)
By suitably choosing Kc , we can ensure that M is positive semidef-
ˆ
with z ∈ Rn is uniformly exponentially stable and that J(q)
has
full row rank, and choose the matrix Kc such that M given by
(57) is positive semidefinite. Then the adaptive outer loop controller
given by (48), (12), (7), (45), (49), (50), (51), (52), and (14) for the
robotic system (2) and (3) under the inner PID position controller (42)
ensures the stability of the system and convergence of the task-space
position error, i.e., ∆x → 0 as t → ∞.
The proof of Theorem 4 can be completed by following similar
steps as in the proof of Theorem 1.
Remark 6:
1) One key issue in the case of using a low-level PID controller is
the choice of Kc , and obviously large enough Kc can ensure
that M is positive semidefinite. The remaining thing is how to
determine the gain Kc . In practice, since the damping matrix
B may be quite small, we thus neglect it and this implies that
kc,ii (i.e., the i-th diagonal entry of Kc ) should satisfy
kc,ii ≥
kP,ii +
2kI,ii
q
, ∀i = 1, . . . , n
2
kP,ii
+ 4kI,ii kD,ii
(60)
where kD,ii is the i-th diagonal entry of KD , kP,ii is the i-th
diagonal entry of KP , and kI,ii is the i-th diagonal entry of
KI . Here we face the similar situation as the control engineers
of the robot production company. The control engineers are
usually careful about the choice of the integral gain and large
gain may cause instability; a trade-off has to be made between
the attenuation of constant disturbances and the stability margin
of the control system. In practice, kP,ii and kD,ii are possibly/generally chosen to be not less than kI,ii , √and in this
case, we can simply choose kc,ii as kc,ii ≥ ( 5 − 1)/2,
∀i = 1, . . . , n.
2) An important issue in proving Theorem 4 is to clarify the inputoutput properties of the following system [derived from (48)]
Z t
[qc (σ) − qr (σ)]dσ
q̇c − q̇r + K̂P (qc − qr ) + K̂I
0
h
i
=diag[ŵi , i = 1, . . . , n] −αJˆT (q)∆x + Yd (q, q̇, q̇r∗ , q̈r∗ )âd
− Kc (q − qr )
(61)
where the right side is bounded. The main issue can now
be reduced to investigating the stability of the linear timevarying system (59). By the assumption that the system (59)
is uniformly
R t exponentially stable, we can directly obtain from
(61) that 0 [qc (σ)−qr (σ)]dσ, qc −qr , and q̇c − q̇r are bounded.
Then the boundedness of q̈ can be ensured.
On the other hand, by the standard projection algorithms [39],
we can conveniently ensure that K̂P and K̂I are uniformly
positive. In addition, the boundedness of K̂P and K̂I is a
direct consequence of that of the Lyapunov function candidate
given by (56). But even under these two conditions, we still
cannot ensure the uniform exponential stability of (59) [which
is a sufficient condition to ensure the uniform bounded-input
bounded-output stability of (61) according to the standard linear
8
system theory] since the coefficient matrices are time-varying.
In practice, we may slow down the adaptation to KP and KI
(i.e., slow down the variation of K̂P and K̂I ) so that the system
(59) is a slowly time-varying (or quasi-time-invariant) linear
system.
IV. G ENERALIZATIONS AND F URTHER D ISCUSSIONS
Let us now discuss the proposed framework in other closely related
topics concerning adaptive robot control.
A. Direct Adaptation
The first result that we would like to discuss is the well-known
Slotine and Li adaptive controller [9] and the result there is presented
in the context of open joint torque control. If we redefine q̇r in (22)
as
q̇r = q̇d − ᾱ(q − qd )
(62)
with qd ∈ Rn being the desired joint position and ᾱ a positive
design constant, then the adaptive controller given by (23), (12), (15),
(16), and (17) with q̇r being defined by (62) becomes an outer loop
(“applicable”) version of Slotine and Li adaptive controller suitable
for practical robotic systems (without opening the torque control
module) with an unmodifiable inner PI velocity control loop (in the
case of an inner PID position controller, its outer loop version can
be similarly developed by following the steps in Sec. III-D). In this
case of the joint-space position tracking, one can easily show that the
position command qc = qd is also qualified for ensuring the stability
of the robotic system and convergence of the joint tracking errors.
B. Composite Adaptation
The implementation of the standard composition adaptation algorithm given in [43] in the framework of inner/outer controller
structure seems not straightforward and we need to ensure that no
additional number of parameters appear in the filtered dynamic model
(i.e., avoiding the overparameterization). For this purpose, we rewrite
the dynamics (3) as
diag[wi , i = 1, . . . , n][M (q)q̈ + C(q, q̇)q̇ + B q̇ + g(q)]
= KP−1 u = −(q̇ − q̇c ) − KP−1 KI (q − qc )
(63)
and by using the filtering technique in [43], we then have the
following equation without involving joint acceleration measurement
diag[wi , i = 1, . . . , n]Yf (q, q̇, t)ad
= u∗f − diag[wI,i , i = 1, . . . , n]hf
ki∗−1 ,
(64)
−1
wI,i = kP,ii
kI,ii , i = 1, . . . , n, Yf (q, q̇, t) =
λ
λf
∗
f
uf = − p+λf (q̇ − q̇c ), and hf = p+λ
(q − qc )
f
where wi =
λf
Y (q, q̇, q̇, q̈),
p+λf d
with p and λf > 0 being the Laplace variable and the filter parameter,
respectively. Let
û∗f = diag[ŵ]Yf âd + diag[hf ]ŵI
(65)
T
where ŵI is the estimate of wI = [wI,1 , . . . , wI,n ] . Then we define
a prediction error
ef =û∗f − u∗f
=diag[w]Yf ∆ad + diag[Yf âd ]∆w + diag[hf ]∆wI
(66)
and the composite adaptive version of (15), (16), and (17) is given
as
ŵ˙ = − Λ(diag[Yd (q, q̇, q̇r , q̈r )âd ]s + γ0 diag[Yf âd ]ef )
â˙ d = − Γd (YdT (q, q̇, q̇r , q̈r )s + γ0 YfT ef )
ŵ˙ I =ΛI (diag[qc − qr ]s − γ0 diag[hf ]ef )
(67)
(68)
(69)
where γ0 is a positive design constant.
Remark 7: The interesting and also distinguished point here is that
the prediction error ef given by (66) contains the unknown coefficient
matrix diag[w], due to which the regressor matrix is actually only
partially known. This motivates us to wonder whether or not the
stability of the closed-loop robotic system can still be guaranteed
under the composite adaptation here.
Consider the nonnegative function
T −1 ∗
V1 = (1/2)[∆wT Λ−1 K ∗ ∆w + ∆aTd Γ−1
d ∆ad + ∆wI ΛI K ∆wI ]
(70)
whose derivative along (67), (68), and (69) can be written as
V̇1 = − ∆wT K ∗ diag[Yd (q, q̇, q̇r , q̈r )âd ]s − ∆aTd YdT (q, q̇, q̇r , q̈r )s
+ ∆wIT K ∗ diag[qc − qr ]s − γ0 (∆wT K ∗ diag[Yf âd ]ef
+ ∆aTd YfT ef + ∆whT K ∗ diag[hf ]ef )
=−∆wT K ∗ diag[Yd (q, q̇, q̇r , q̈r )âd ]s − ∆aTd YdT (q, q̇, q̇r , q̈r )s
+ ∆wIT K ∗ diag[qc − qr ]s − γ0 eTf K ∗
× (diag[Yf âd ]∆w + diag[w]Yf ∆ad + diag[hf ]∆wI ) (71)
|
{z
}
ef
where the first three terms are used to compensate for the indefinite
terms due to the parametric uncertainty, and in this way, the stability
is ensured. The key point here is to exploit the independent nature
of the joint processors which means that K ∗ is diagonal. The other
versions of composite adaptation, e.g., BGF composite adaptation
and CF composite adaptation (see, e.g., [43]) can also be used so
that smoother parameter adaptation and better tracking performance
can be achieved. Specifically, the CF composite adaptation laws can
be given as
ŵ˙ = −Λ(diag[Yd (q, q̇, q̇r , q̈r )âd ]s + γ0 diag[Yf âd ]ef )
Λ̇ = λ1 (Λ − ΛΛ̄−1 Λ) − γ0 Λ(diag[Yf âd ])2 Λ
â˙ = −Γ (Y T (q, q̇, q̇ , q̈ )s + γ Y T e )
r
r
0 f f
d
d
d
(72)
−1
T
Γ̇
d = λ2 (Γd − Γd Γ̄d Γd ) − γ0 Γd Yf Yf Γd
ŵ˙ I = ΛI (diag[qc − qr ]s − γ0 diag[hf ]ef )
2
Λ̇I = λ3 (ΛI − ΛI Λ̄−1
I ΛI ) − γ0 ΛI (diag[hf ]) ΛI
where λ1 , λ2 , and λ3 are strictly positive forgetting factors (timevarying or constant), and Λ̄, Γ̄d , and Λ̄I denote the upper bounds of
Λ, Γd , and ΛI , respectively. Note that Λ(0) and ΛI (0) are chosen
as diagonal positive definite matrices satisfying 0 < Λ(0) ≤ Λ̄ and
0 < ΛI (0) ≤ Λ̄I , and Γd (0) can be chosen as a symmetric positive
definite matrix satisfying 0 < Γd (0) ≤ Γ̄d , and in this way, it can
be shown that Λ(t) and ΛI (t) are always diagonal, ∀t ≥ 0. The
derivative of the nonnegative function V1 defined by (70) in this case
becomes
V̇1 =−∆wT K ∗ diag[Yd (q, q̇, q̇r , q̈r )âd ]s − ∆aTd YdT (q, q̇, q̇r , q̈r )s
+ ∆wIT K ∗ diag[qc − qr ]s
λ1
λ2
−1
−
∆wT (Λ−1 − Λ̄−1 )K ∗ ∆w −
∆aTd (Γ−1
d − Γ̄d )∆ad
2
2
γ0 T ∗
λ3
−1
∗
∆wIT (Λ−1
e K ef .
(73)
−
I − Λ̄I )K ∆wI −
2
2 f
Obviously, the stability of the system and convergence of the joint
tracking errors can be guaranteed.
C. Adaptive Control of Flexible-Joint Manipulators
The typical result may be the singular-perturbation-based adaptive
control approach [10], [44]. Consider a flexible-joint manipulator
9
governed by [45]
(
M0 (q)q̈ + C(q, q̇)q̇ + B q̇ + g(q) = Ks (θ − q)
Dr θ̈ + Br θ̇ = Ku − Ks (θ − q)
(74)
where θ ∈ Rn is the rotor position, Ks is the constant, diagonal,
and positive definite stiffness matrix, Dr ∈ Rn×n is the rotor inertia
matrix seen from the link side, and Br ∈ Rn×n is the damping
matrix. In this case, the motor velocity command is defined as
q̇c + K̂I qc =q̇r + K̂I qr + diag[ŵi , i = 1, . . . , n]
× Yd (q, q̇, q̇r , q̈r )âd
(75)
D. Further Discussions
with q̇r being defined as
q̇r =q̇d − ᾱ(q − qd ).
(76)
The adaptation laws for ŵ, âd , and ŵI are given as
ŵ˙ = −Λdiag[Yd (q, q̇, q̇r , q̈r )âd ]s
â˙ d = −Γd YdT (q, q̇, q̇r , q̈r )s
ŵ˙ I = ΛI diag[qc − qr ]s
(77)
(78)
(79)
The low-level PI control action in this case of flexible-joint robots
would typically take the form
u = −KP (θ̇ − q̇c ) − KI (θ − qc ) .
(80)
Substituting (80) into the second equation of (74) with some further
manipulations gives
Dr (θ̈ − q̈) + Br (θ̇ − q̇)
= − K ∗ (θ̇ − q̇c ) − KKI (θ − qc )
− Dr q̈ − Br q̇ − Ks (θ − q)
(81)
and the above equation can further be written as
Dr (θ̈ − q̈) + (Br + K ∗ )(θ̇ − q̇) + [(Ks + KKI )Ks−1 ]Ks (θ − q)
= − K ∗ (q̇ − q̇c ) − KKI (q − qc ) − Dr q̈ − Br q̇.
(82)
As the fast dynamics becomes settled, i.e., Ks (θ − q) is quasiconstant, we obtain that
n
Ks (θ − q) = − Ks (Ks + KKI )−1 K ∗ (q̇ − q̇c )
o
+ KKI (q − qc ) + Dr q̈ + Br q̇
(83)
and thus the slow dynamics becomes
[M0 (q) + Ks∗ Dr ]q̈ + C(q, q̇)q̇ + (B + Ks∗ Br )q̇ + g(q)
= − Ks∗ K ∗ [(q̇ − q̇c ) + KI (q − qc )]
Ks∗
−1
even without the need of any modification (in practice, any robot has
certain joint flexibility). In fact, the relative damping suggested in [10]
is naturally included as applying the scaled-dynamic-compensation
versions of most control schemes valid for rigid robots to (flexiblejoint) robots in practice. More remarks in terms of the roles of the
rotor inertias and joint stiffness are presented in the later simulation.
One may also be interested in deriving a composite adaptive
version of the adaptive scheme for flexible-joint manipulators and
it shall be feasible by using similar techniques as those for rigid
manipulators.
(84)
where
= Ks (Ks + KKI ) . The fast/slow-dynamics-based
analysis given above is based on [10], [44], and one can rigorously
obtain the stability and convergence of the system by following
similar arguments as in [10], [44].
Remark 8: Different from the rigid robot case, the actual value
of the scale parameter now satisfies diag[w] = Ks∗ K ∗ . Part of the
low-level integral action in (80) −KI (θ − q) is the same as the
relative position feedback in [44] and its effect is to increase the
joint stiffness, providing the possibility of applying the control to
manipulators with a relatively low joint stiffness [44].
Remark 9: It is interesting to note that the joint velocity and
position commands in the flexible joint case remains the same as
the rigid joint case. This provides a good understanding and more
importantly an effective justification of why most results derived in
the case of rigid robots are generally applicable to (not justified in the
previous literature though) practical robotic systems with inner/output
loop structure (e.g., most industrial/commercial robotic systems),
Here, we take several standard adaptive robot control schemes for
illustrating how the scaled dynamic compensation makes them to
be qualified adaptive outer loop schemes and further the possible
applications to robotic systems with an inner/outer loop structure.
It seems hopeful that most adaptive robot control schemes in the
literature can be reshaped to be adaptive outer loop schemes by
accommodating such modifications.
The inner/outer controller structure basically performs the inner
joint servoing much faster and the outer loop relatively slower, and
thus the scaled dynamic compensation is actually exerted at a quite
lower updating cycle. In the case of fast operating process, this
would result in degrading of the performance and even instability at
certain extreme cases. The main objective of the study and results
presented here is to provide the possibility of exerting dynamic
compensation (feedforward) even in the standard setting of industrial
robotic systems, of course, under the limit of the operation speed.
Once upon a while, direct-drive robots are believed to be promising in
taking over the role of the standard robots using gear reduction in that
direct-drive robots are much efficient and less influenced by friction
and backlash, etc. But this hope advances not so favorably, especially
in applications, and the reasons may perhaps be the following:
• The torque output of direct-drive robots is small and large
torque output would require large and heavy joint motors, which,
however, are significantly constrained by the weight limit of the
manipulator;
• direct-drive robots, as is typically expected, are torque-based,
but torque-based design is relatively risky and not so reliable
since all factors are taken into account at the same time and
in addition the communication constraint presents a limit of the
coupling torque exerting cycle.
These unfavorable factors concerning torque-based design and directdrive robots give rise to the welcome of the inner/outer loop structure
in most practical robotic applications and this may still be going in
the future. A very recent example is Robonaut 2 [46] which uses
harmonic gear transmission instead of the direct-drive configuration,
and the feedforward is shown to be necessary to improve the control
accuracy as well as the system response within the range of the
torque limit. In this specific example, the joint torque control loop
is open just for admitting the injection of feedforward or dynamic
control action, but from a long run and for promoting the large-scale
production and decreasing the cost, the inner/outer loop structure with
the joint control loop sealed may perhaps be more desirable.
V. S IMULATION R ESULTS
A. Task-Space Adaptive Control
Consider a three-DOF manipulator with a tool, as is shown in Fig.
2. Its physical parameters are given in Table I with the labels 1,
2, 3, and E denoting link 1, 2, 3, and the tool, respectively, and
the diagonal rotor inertia matrix (seen from the link side) Dr =
10
<
0.6
=
∆x
regulation errors (m)
;
=
<
;
Fig. 2.
(1)
∆x(2)
0.4
∆x
0.2
(3)
0
−0.2
−0.4
−0.6
−0.8
0
Three-DOF manipulator.
2
4
6
8
10
time (s)
TABLE I
P HYSICAL PARAMETERS OF THE MANIPULATOR
lC,i (m)
0.4320, 0.0720, 0.4320
1.8000
0.9000
0.0054, 0.1620, 0.1620
1.8000
0.9000
0.6000
0.0054, 0.1620, 0.1620
1.8000
0.9000
0.8000
0.0032, 0.0960, 0.0960
1.2000
0.6000
mi (kg)
1
1.6000
2
0.6000
3
E
diag[0.6, 0.3, 0.1]. The diagonal matrix B and K are set as B =
diag[0.20, 0.15, 0.10] and K = diag[60.0, 30.0, 10.0], respectively.
The angle about the axis Z3 between the tool and the third link
is δ = 30 deg. The inner joint servoing loop with a PI velocity
controller (the case of PID position controller is considered in Sec.
V-D) is operated at a high-rate cycle with the sampling period being
0.5 ms, and the outer loop is operated at a low-rate cycle with the
sampling period being 20 ms. The gains of the low-level PI controller
are set as KP = 30.0I3 1 and KI = 15.0I3 , and this means that the
effective PI gains are KP∗ = diag[1800.0, 900.0, 300.0] and KI∗ =
diag[900.0, 450.0, 150.0].
1) Regulation problem: We first perform the simulations of the
system under the filter-based and observer-based task-space regulation schemes. The manipulator starts at the configuration q(0) =
[π/6, π/3, −5π/6]T and the corresponding task-space position is
x(0) = [−0.7500, 1.2990, 0.5196]T . The desired task-space position
is set as xd = [−1.0, 2.0, 0.8]T . In the case of using the filterbased scheme, the controller parameters are set as K1 = 60.0I3 ,
K2 = 2.0I3 , α = 2.0, Γk = 20.0I3 , Λ = 0.001I3 , Γd = 0.006I15 ,
and ΛI = 100.0I3 . The initial values of the parameter estimates
are chosen as âk (0) = [3.0, 5.0, 2.0]T , ŵ(0) = 03 , âd (0) = 015 ,
and ŵI (0) = [1, 1, 1]T . Simulation results are shown in Fig. 3,
Fig. 4, and Fig. 5, which, respectively, give the task-space position
errors, the scale parameter estimates, and the estimate of wI . In the
case of using the observer-based scheme, the controller parameters
β and γ are determined as β = 1.0 and γ = 1.0, which obviously
satisfy the condition (30), and the other controller parameters and
the initial parameter estimates are chosen to be the same as those
of the filter-based scheme. The task-space position error, the scale
parameter estimates, and the estimate of wI are shown in Fig. 6, Fig.
7, and Fig. 8, respectively.
2) Tracking problem: Let us now consider the case of using the
observer-based tracking controller given by (23), (12), (21), (24),
(25), (16), and (17). The desired task-space trajectory is given as
xd = [−1.0500 + 0.3 cos(πt/3), 1.2990 + 0.3 sin(πt/3), 0.5196 +
0.3 sin(πt/3)]T . The initial configuration of the manipulator is set
to be the same as the above. The controller parameters are chosen
as β = 0.8, γ = 0.8, Γk = 200.0I3 , Λ = 6.0I3 , Γd = 27.0I15 ,
1I
ℓ
denotes the ℓ × ℓ identity matrix, ℓ = 2, 3, . . . .
0.03
0.025
parameter estimates
li (m)
i-th body
Task-space position errors (filter-based control).
estimate of w
1
0.02
estimate of w2
estimate of w3
0.015
0.01
0.005
0
0
2
4
6
8
10
time (s)
Fig. 4.
Scale parameter estimates (filter-based control).
4.5
4
parameter estimates
· m2 )
Fig. 3.
estimate of wI,1
3.5
estimate of wI,2
3
estimate of wI,3
2.5
2
1.5
1
0
2
4
6
8
10
time (s)
Fig. 5.
Estimate of wI (filter-based control).
0.6
∆x(1)
(2)
0.4
regulation errors (m)
xx , I yy , I zz (kg
IC,i
C,i C,i
∆x
∆x(3)
0.2
0
−0.2
−0.4
−0.6
−0.8
0
Fig. 6.
5
10
time (s)
15
Task-space position errors (observer-based control).
20
11
0.04
scale parameter estimates
scale parameter estimates
0.015
0.01
estimate of w1
0.005
estimate of w2
0.03
0.02
0.01
estimate of w1
0
estimate of w
estimate of w2
3
0
0
Fig. 7.
5
10
time (s)
15
estimate of w3
−0.01
0
20
Scale parameter estimates (observer-based control).
Fig. 10.
2.5
estimate of wI,1
estimate of w
I,2
estimate of wI,3
2
1.5
5
10
time (s)
15
Fig. 11.
and ΛI = 100.0I3 . The task-space position tracking errors, the scale
parameter estimates, and the estimate of wI are shown in Fig. 9, Fig.
10, and Fig. 11, respectively. The gains are increased, in comparison
with the the case of regulation, and this is feasible since in the
tracking problem here, the desired trajectory starts at the current
position of the manipulator.
B. Joint-Space Adaptive Control
Consider first the outer loop version of the direct adaptive controller proposed by Slotine and Li with the controller parameters
being chosen as ᾱ = 2.0, Λ = 0.5I3 , Γd = 0.5I15 , and
ΛI = 100.0I3 . The desired joint trajectory is set as qd = 36[(1 −
cos πt), sin πt, sin πt]T deg. The initial parameter estimates are
chosen as ŵ(0) = 03 , âd (0) = 015 , and ŵI (0) = [1.0, 1.0, 1.0]T .
2.5
20
estimate of wI,1
estimate of wI,2
estimate of wI,3
2
1.5
1
0
20
Estimate of wI (observer-based control).
5
10
time (s)
15
20
Estimate of wI (observer-based tracking control).
The joint position tracking errors and parameter estimates are shown
in Fig. 12, Fig. 13, and Fig. 14.
We next consider the case of using the composite adaptation in
Sec. IV-B with the controller parameters γ0 and λf being chosen
as γ0 = 0.3 and λf = 1.0 and the other controller parameters
the same as those of the direct adaptive controller. The simulation
results are shown in Fig. 15, Fig. 16, and Fig. 17, and in comparison
with Fig. 12, Fig. 13, and Fig. 14, we see smoother tracking errors
and parameter estimates as well as the improved convergence of the
tracking errors.
C. Joint-Space Adaptive Control Considering Joint Flexibility
The joint stiffness matrix Ks is set as Ks = 106 ×
diag[6.0, 3.0, 1.0] and the diagonal matrix Br is set as Br =
3
0.04
∆x(1)
(2)
2
(3)
1
∆x
0.02
∆x
tracking errors (deg)
tracking errors (m)
15
3
parameter estimates
parameter estimates
Fig. 8.
10
time (s)
Scale parameter estimates (observer-based tracking control).
3
1
0
5
0
−0.02
∆q(1)
∆q(2)
∆q(3)
0
−1
−2
−3
−0.04
−4
−0.06
0
5
10
time (s)
15
−5
0
20
Fig. 9. Task-space position tracking errors (observer-based tracking control).
Fig. 12.
5
10
time (s)
15
20
Joint position tracking errors (direct adaptive controller).
12
2.5
0.5
estimate of w
I,1
2
parameter estimates
parameter estimates
0.4
estimate of w1
0.3
estimate of w
2
estimate of w3
0.2
Fig. 13.
5
10
time (s)
15
2.5
estimate of wI,1
estimate of wI,2
parameter estimates
2
estimate of wI,3
1.5
1
0.5
0
Fig. 14.
5
10
time (s)
15
20
Estimate of wI (direct adaptive controller).
3
(1)
∆q
tracking errors (deg)
2
∆q(2)
∆q(3)
1
0
−1
−2
−3
−4
−5
0
Fig. 15.
5
10
time (s)
15
20
Joint position tracking errors (composite adaptive controller).
0.5
scale parameter estimates
estimate of w1
Fig. 16.
estimate of w2
0.4
estimate of w
3
0.3
0.2
0.1
0
0
1.5
1
5
10
time (s)
15
0
0
20
Scale parameter estimates (direct adaptive controller).
−0.5
0
estimate of wI,3
0.5
0.1
0
0
estimate of wI,2
20
Scale parameter estimates (composite adaptive controller).
Fig. 17.
5
10
time (s)
15
20
Estimate of wI (composite adaptive controller).
diag[0.30, 0.20, 0.15]. The controller parameters are chosen to be
the same as the rigid manipulator case except that the adaptation gain
ΛI is reduced to ΛI = 60.0I3 . The simulation results are shown in
Fig. 18, and we see that the performance is comparable with the rigid
manipulator case. But one may need to be cautious about the choice
of the rotor inertias, and if the rotor inertias are too small [compared
with the manipulator inertia matrix due to the link motion, i.e., M0 (q)
in (74)], it is hard to choose a group of controller parameters that can
stabilize the system. This is understandable as we recall the standard
practice in terms of the design of the motor inertia, i.e., in the case
that the motor inertia is strikingly smaller than the load inertia, the
whole system would be quite difficult to stabilize and the use of
advanced control algorithms does not help much. In particular, we
perform a simulation for the case of reduced joint stiffness matrix,
i.e., setting Ks to be Ks = 104 × diag[6.0, 3.0, 1.0] with the rotor
inertias remaining unchanged. In addition, as the joint becomes more
flexible, it is hard for the manipulator to track a fast time-varying
trajectory, and therefore we slow down the evolution of the desired
trajectory as qd = 36[1 − cos(πt/3), sin(πt/3), sin(πt/3)]T deg.
The joint tracking errors are shown in Fig. 19. But as we reduce
the rotor inertias to, e.g., 50% of the original, it is very difficult to
stabilize the system even with the joint stiffness being unchanged
(relatively high). To illustrate the reason behind this phenomenon,
we calculate the link inertia matrix M0 (q) at q = q(0) = 03 and its
value is
18.9058
M0 (q) = 0
0
0
18.9290
9.4327
0
9.4327 .
5.1205
The three eigenvalues of M0 (q) are 0.3352, 18.9058, and 23.7143.
This means that the maximum load/rotor inertia ratio reaches
23.7143/0.3 ≈ 79 (happening at the second joint) and such a
ratio makes it challenging to stabilize the system. One solution to
this problem, as suggested by the standard results in the design of
the load/motor ratio, is to increase the joint stiffness, and in other
words, larger stiffness allows the specification of larger load/rotor
inertia ratio. Fig. 20 shows the joint tracking errors as the stiffness
is increased to Ks = 108 × diag[6.0, 3.0, 1.0] with the desired
joint position being still the same as the rigid manipulator case.
However, in the simulation, we have to decrease the integration step
size so that the stability of the numerical integration can be ensured
since the degree of stiffness of the system dynamics is increased. In
this particular case (i.e., increased joint stiffness and decreased rotor
inertia), the step size is decreased from 0.5 ms to 0.05 ms.
13
3
4
∆q(1)
∆q(1)
(2)
∆q
∆q(2)
2
∆q(3)
1
tracking errors (deg)
tracking errors (deg)
2
0
−1
−2
−3
∆q(3)
0
−2
−4
−4
−5
0
Fig. 18.
5
10
time (s)
15
−6
0
20
Joint position tracking errors (flexible joint case).
5
10
time (s)
15
20
Fig. 21. Joint position tracking errors (adaptive outer loop controller with
an inner PID position controller).
1
r
0
r
c
q(3)−q(3)
r
c
10
−0.5
qc−qr (deg)
tracking errors (deg)
c
q(2)−q(2)
∆q(1)
−1
∆q(2)
−1.5
∆q(3)
5
0
−2
−2.5
−3
0
Fig. 19.
q(1)−q(1)
15
0.5
−5
5
10
time (s)
15
20
−10
0
5
10
time (s)
15
20
Joint position tracking errors (reduced joint stiffness).
Fig. 22. qc − qr (adaptive outer loop controller with an inner PID position
controller).
D. Joint-Space Adaptive Control With an Inner PID Position Controller
The gains of the inner PID position controller is set as KD =
30.0I3 , KP = 15.0I3 , and KI = 10.0I3 . The controller parameters
for the joint-space adaptive outer loop controller (which can be
developed similarly to the one in Sec. III-D) are chosen as ᾱ = 1.5,
Kc = 0.8, Γd = 0.5I15 , Λ = 0.5I3 , ΛP = 100.0I3 , and ΛI =
50.0I3 . The initial parameter estimates are chosen as âd (0) = 015 ,
ŵ(0) = 03 , ŵP (0) = [1.0, 1.0, 1.0]T , and ŵI (0) = [1.0, 1.0, 1.0]T .
The desired trajectory is set to be the same as the case of an
inner PI velocity controller. The joint position tracking errors are
shown in Fig. 21, which is comparable with the case of an inner PI
velocity controller (Fig. 12). Interestingly, no unstable phenomenon
is observed even with very fast adaptation to KP and KI although
3
∆q(1)
tracking errors (deg)
2
∆q(2)
∆q(3)
1
0
−1
−2
−3
−4
−5
0
5
10
time (s)
15
20
Fig. 20. Joint position tracking errors (increased joint stiffness and decreased
rotor inertias).
it is currently still challenging to rigorously ensure the uniform
exponential stability of (59), and the evolution of qc − qr is plotted
in Fig. 22. The quantity qc − qr (which is apparently bounded based
on the data shown in Fig. 22) characterizes the injected dynamic
compensation of the proposed controller.
E. Further Remarks
One key issue in the above simulations is the choice of controller
parameters including those of the low-level PI controller. Different
from most theoretical results in the literature that design control
laws at the torque level, the system here actually has two loops
with strikingly different updating frequency, namely, the low-rate
outer loop and the high-rate inner joint servoing loop. Due to this
structure, to guarantee the robustness and performance of the whole
system, high gains are specified in the high-rate joint servoing loop
while low gains are specified in the low-rate outer loop. We naturally
produce a system that consists of two loops with two time-scales and
interestingly, it is the system constraint that gives rise to the two-timescale behavior of the closed-loop system that we cannot modify.
Another issue often involved in practice is the computational
efficiency, especially in the case that the number of the DOFs of the
manipulator is very large. The typical solution to this problem is the
recursive implementation of the adaptive controllers—see, e.g., [33],
[47], [48]. The recursive direct adaptive controller as detailed in [33],
[47] has the complexity O(n) and the complexity of the recursive
composite adaptive controller in [48] is O(n2 ), where n denotes the
number of the DOFs of the manipulator. We may also note that all the
complicated computations of the nonlinear and coupling terms take
place in the low-rate outer loop with a powerful computer, and thus
14
0.03
∆x(1)
∆x(2)
tracking errors (m)
0.02
∆x(3)
0.01
0
−0.01
Fig. 23.
Experimental setup with UR10.
the complexity up to O(n) or O(n2 ) is expected to be acceptable.
−0.02
0
Fig. 24.
20
40
60
time (s)
80
100
Position tracking errors (using qc as the joint position command).
VI. E XPERIMENTAL R ESULTS
The controller parameters are chosen as γ = 0.5, Kc = 0.8, Γd =
1.0I16 , Λ = 0.01I3 , ΛP = 0.02I3 , and ΛI = 0.02I3 . The initial
values of the dynamic parameter estimates are set to be zero, i.e.,
âd (0) = 016 . Note that here we expand ad in (5) to include the
parameters associated with the Coulomb friction which can be written
as [36]
fc = Dsgn(q̇)
(86)
where sgn(q̇) = [sgn(q̇1 ), sgn(q̇2 ), sgn(q̇3 )]T and D is a 3 × 3
diagonal positive definite matrix with its diagonal entries unknown.
Correspondingly, the dynamic regressor matrix used in the experiment
is the combination of Yd (q, q̇, q̇r , q̈r ) and diag[sgn(q̇r )]. The initial
values of w, wP , and wI are chosen as ŵ(0) = 03 , ŵP (0) =
[1.8, 1.8, 1.8]T , and ŵI (0) = [1.2, 1.2, 1.2]T , respectively. The
position tracking errors as using the proposed controller are shown in
Fig. 24. For comparison, the position tracking errors as using qr as the
joint position command (i.e., the conventional kinematic controller)
are shown in Fig. 25. The tracking accuracy is apparently improved
by using the proposed dynamic modularity approach.
VII. C ONCLUSION
In this paper, we have proposed a dynamic modularity approach to
adaptive control of robotic systems with an inner/outer loop structure,
and both the task-space and joint-space control are taken into consideration under this framework. The proposed adaptive outer loop
controllers take into full account the system dynamic effects while
∆x(1)
0.06
∆x(2)
∆x(3)
0.04
tracking errors (m)
To further show the practical performance of the proposed adaptive
outer loop controllers, we perform the experimental study using the
UR10 of Universal Robots at Nanyang Technological University (see
Fig. 23). We fix the position of upper three joints of the UR10 robotic
system to be [0, −1.57, 0]T rad, and only the motion of the first
three DOFs is considered, and the low-level inner loop controller is
considered to take the PID position control action. Either the position
or velocity command for the UR10 robot can be designed and it is not
allowed to deign the position and velocity commands simultaneously.
The sampling period of the outer loop is around (not exactly) 0.2 s.
We here only take into consideration the Cartesian-space tracking
control problem with unknown system dynamics and unknown lowlevel PID controller parameters while the kinematics of the system is
accurately known. That is, we use the joint reference velocity defined
by (39) yet with the estimated Jacobian matrix being replaced by the
actual Jacobian matrix and with xo being replaced by x. The desired
trajectory in the Cartesian space is specified as
−0.45 + 0.1 sin(t)
(85)
xd (t) = 0.45 + 0.1 cos(t) .
0.5
0.02
0
−0.02
−0.04
−0.06
0
10
20
time (s)
30
Fig. 25. Position tracking errors (using qr as the joint position command,
i.e., the conventional kinematic controller).
most existing kinematic controllers rely on the ad hoc assumption
of fast enough joint servoing loop or the modification of the lowlevel joint servoing controller to be much more complicated one.
From an application perspective, most existing results cannot ensure
the stability of the system or convergence of the tracking/regulation
error as applied to robotic systems with an inner/outer loop structure
(e.g., most commercial/industrial robotic systems) while the proposed
adaptive outer loop schemes can guarantee the stability and convergence of the system without the need to modify the low-level joint
servoing loop. The goal of the study here is to yield a module robot
control system where the adaptive outer loop is user-defined and the
inner loop is factory-defined and embedded.
It might be worth discussing the roles of feedback separation in
the proposed controllers. Feedback separation is a design objective
introduced in the context of Cartesian-space control and visual
servoing control of robots with uncertain kinematics (see, e.g., [31],
[32]). Feedback separation is initially for generating simple adaptive
kinematic controllers for industrial/commertial robots, and on the
other hand it can potentially reduce the activity of the dynamic compensation action. For instance, the filter-based regulation algorithm
in Sec. III-A actually does not achieve feedback separation and thus
the scaled dynamic compensation action given in (11) involves an
additional term −αJˆT (q)∆x (i.e., a stronger compensation action is
required). In contrast, both the observer-based regulation and tracking
controllers achieve the feedback separation, benefiting from which,
the scaled dynamic compensation action defined in (23) no longer
involves additional terms (of course, the controller structure becomes
more complex since an observer is introduced). From a control
viewpoint, this leads us to reconsider the issue of the cancellation
of indefinite terms in the standard backstepping-based control. The
15
cancellation of indefinite terms can lead to a good form of the
derivative of the Lyapunov function but often gives rise to potentially
decreased robustness and strong coupling between different control
loops. The realization of feedback separation may help avoid the
unfavorable cancellations of indefinite terms.
Furthermore, the proposed approach may possibly be applicable to
other classes of (commercial) mechanical systems (e.g., space robots,
mobile robots, or aerial vehicles) that have a hidden torque/force
control loop yet admit the design of the velocity (or position)
command.
ACKNOWLEDGMENT
The authors would like to thank Dr. Tiantian Jiang, Dr. Yong Hu,
and Dr. Yong Wang for the helpful discussions and comments on
stability issues concerning the linear time-varying systems.
R EFERENCES
[1] M. Aicardi, A. Caiti, G. Cannata, and G. Casalino, “Stability and
robustness analysis of a two layered hierarchical architecture for the
closed loop control of robots in the operational space,” in Proceedings
of the IEEE International Conference on Robotics and Automation,
Nagoya, Japan, 1995, pp. 2771–2778.
[2] J. Roy and L. L. Whitcomb, “Adaptive force control of position/velocity
controlled robots: Theory and experiment,” IEEE Transactions on
Robotics and Automation, vol. 18, no. 2, pp. 121–137, Apr. 2002.
[3] R. Kelly and J. Moreno, “Manipulator motion control in operational
space using joint velocity inner loops,” Automatica, vol. 41, no. 8, pp.
1423–1432, Aug. 2005.
[4] K. Camarillo, R. Campa, V. Santibáñez, and J. Moreno-Valenzuela,
“Stability analysis of the operational space control for industrial robots
using their own joint velocity PI controllers,” Robotica, vol. 26, no. 6,
pp. 729–738, Nov. 2008.
[5] D. E. Whitney, “Resolved motion rate control of manipulators and
human prostheses.” IEEE Transactions on Man-Machine Systems, vol.
MMS-10, no. 2, pp. 47–53, Jun. 1969.
[6] M. W. Spong and M. Vidyasagar, Robot Dynamics and Control. New
York: Wiley, 1989.
[7] M. Grotjahn and B. Heimann, “Model-based feedforward control in
industrial robotics,” The International Journal of Robotics Research,
vol. 21, no. 1, pp. 45–60, Jan. 2002.
[8] F. Sanfilippo, L. I. Hatledal, H. Zhang, M. Fago, and K. Y. Pettersen,
“Controlling Kuka industrial robots: Flexible communication interface
JOpenShowVar,” IEEE Robotics & Automation Magazine, vol. 22, no. 4,
pp. 96–109, Dec. 2015.
[9] J.-J. E. Slotine and W. Li, “On the adaptive control of robot manipulators,” The International Journal of Robotics Research, vol. 6, no. 3, pp.
49–59, Sep. 1987.
[10] M. W. Spong, “Adaptive control of flexible joint manipulators,” Systems
& Control Letters, vol. 13, no. 1, pp. 15–21, Jul. 1989.
[11] C. C. Cheah, C. Liu, and J.-J. E. Slotine, “Adaptive Jacobian tracking
control of robots with uncertainties in kinematic, dynamic and actuator
models,” IEEE Transactions on Automatic Control, vol. 51, no. 6, pp.
1024–1029, Jun. 2006.
[12] Y.-H. Liu, H. Wang, C. Wang, and K. K. Lam, “Uncalibrated visual
servoing of robots using a depth-independent interaction matrix,” IEEE
Transactions on Robotics, vol. 22, no. 4, pp. 804–817, Aug. 2006.
[13] W. E. Dixon, “Adaptive regulation of amplitude limited robot manipulators with uncertain kinematics and dynamics,” IEEE Transactions on
Automatic Control, vol. 52, no. 3, pp. 488–493, Mar. 2007.
[14] H. Wang, “Adaptive visual tracking for robotic systems without imagespace velocity measurement,” Automatica, vol. 55, pp. 294–301, May
2015.
[15] H. Wang, M. Jiang, W. Chen, and Y.-H. Liu, “Visual servoing of robots
with uncalibrated robot and camera parameters,” Mechatronics, vol. 22,
no. 6, pp. 661–668, Sep. 2012.
[16] M. W. Spong, “On the robust control of robot manipulators,” IEEE
Transactions on Automatic Control, vol. 37, no. 11, pp. 1782–1786,
Nov. 1992.
[17] B. Yao and M. Tomizuka, “Smooth robust adaptive sliding mode control
of manipulators with guaranteed transient performance,” Journal of
Dynamic Systems, Measurement, and Control, vol. 118, no. 4, pp. 764–
775, Dec. 1996.
[18] F. Caccavale and P. Chiacchio, “Identification of dynamic parameters and
feedforward control for a conventional industrial manipulator,” Control
Engineering Practice, vol. 2, no. 6, pp. 1039–1050, Dec. 1994.
[19] J. Swevers, W. Verdonck, and J. D. Schutter, “Dynamic model identification for industrial robots,” IEEE Control Systems Magazine, vol. 27,
no. 5, pp. 58–71, Oct. 2007.
[20] C. C. Cheah, M. Hirano, S. Kawamura, and S. Arimoto, “Approximate
Jacobian control for robots with uncertain kinematics and dynamics,”
IEEE Transactions on Robotics and Automation, vol. 19, no. 4, pp. 692–
702, Aug. 2003.
[21] C. C. Cheah, C. Liu, and J.-J. E. Slotine, “Adaptive tracking control for
robots with unknown kinematic and dynamic properties,” The International Journal of Robotics Research, vol. 25, no. 3, pp. 283–296, Mar.
2006.
[22] X. Liang, H. Wang, W. Chen, and Y.-H. Liu, “Uncalibrated image-based
visual servoing of rigid-link electrically driven robotic manipulators,”
Asian Journal of Control, vol. 16, no. 3, pp. 714–728, May 2014.
[23] H. Wang and Y. Xie, “Prediction error based adaptive Jacobian tracking
for free-floating space manipulators,” IEEE Transactions on Aerospace
and Electronic Systems, vol. 48, no. 4, pp. 3207–3221, Oct. 2012.
[24] H. Wang, Y.-H. Liu, and W. Chen, “Uncalibrated visual tracking control
without visual velocity,” IEEE Transactions on Control Systems Technology, vol. 18, no. 6, pp. 1359–1370, Nov. 2010.
[25] A. C. Leite, A. R. L. Zachi, F. Lizarralde, and L. Hsu, “Adaptive
3D visual servoing without image velocity measurement for uncertain
manipulators,” in 18th IFAC World Congress, Milano, Italy, 2011, pp.
14 584–14 589.
[26] H. Wang, W. Ren, C. C. Cheah, and Y. Xie, “Dynamic modularity
approach to adaptive inner/outer loop control of robotic systems,” in
Proceedings of the Chinese Control Conference, Chengdu, China, 2016,
pp. 3249–3255.
[27] H. Wang and Y. Xie, “Observer-based task-space consensus of networked
robotic systems: A separation approach,” in Proceedings of the Chinese
Control Conference, Hangzhou, China, 2015, pp. 7604–7609.
[28] L. E. Weiss, A. C. Sanderson, and C. P. Neuman, “Dynamic sensorbased control of robots with visual feedback,” IEEE Journal of Robotics
and Automation, vol. 3, no. 5, pp. 404–417, Oct. 1987.
[29] B. Siciliano, “A closed-loop inverse kinematic scheme for on-line jointbased robot control,” Robotica, vol. 8, no. 3, pp. 231–243, Jul. 1990.
[30] S. Hutchinson, G. D. Hager, and P. I. Corke, “A tutorial on visual servo
control,” IEEE Transactions on Robotics and Automation, vol. 12, no. 5,
pp. 651–670, Oct. 1996.
[31] H. Wang, “Adaptive control of robot manipulators with uncertain kinematics and dynamics,” IEEE Transactions on Automatic Control, DOI:
10.1109/TAC.2016.2575827.
[32] ——, “Passivity-based adaptive control for visually servoed robotic
systems,” arXiv preprint arXiv:1506.08762, 2015.
[33] G. Niemeyer and J.-J. E. Slotine, “Performance in adaptive manipulator
control,” The International Journal of Robotics Research, vol. 10, no. 2,
pp. 149–161, Apr. 1991.
[34] J. J. Craig, Introduction to Robotics: Mechanics and Control, 3rd ed.
Upper Saddle River, NJ: Prentice-Hall, 2005.
[35] M. W. Spong, S. Hutchinson, and M. Vidyasagar, Robot Modeling and
Control. New York: Wiley, 2006.
[36] J.-J. E. Slotine and W. Li, Applied Nonlinear Control. Englewood
Cliffs, NJ: Prentice-Hall, 1991.
[37] C. A. Desoer and M. Vidyasagar, Feedback Systems: Input-Output
Properties. New York: Academic Press, 1975.
[38] M. M. Bridges, D. M. Dawson, and X. Gao, “Adaptive control of rigidlink electrically-driven robots,” in Proceedings of the IEEE Conference
on Decision and Control, San Antonio, TX, 1993, pp. 159–165.
[39] P. A. Ioannou and J. Sun, Robust Adaptive Control. Englewood Cliffs,
NJ: Prentice-Hall, 1996.
[40] H. Berghuis and H. Nijmeijer, “Global regulation of robots using only
position measurements,” Systems & Control Letters, vol. 21, no. 4, pp.
289–293, Oct. 1993.
[41] R. Lozano, B. Brogliato, O. Egeland, and B. Maschke, Dissipative
Systems Analysis and Control: Theory and Applications. London, U.K.:
Springer-Verlag, 2000.
[42] C. C. Cheah, C. Liu, and J.-J. E. Slotine, “Adaptive Jacobian vision
based control for robots with uncertain depth information,” Automatica,
vol. 46, no. 7, pp. 1228–1233, Jul. 2010.
[43] J.-J. E. Slotine and W. Li, “Composite adaptive control of robot
manipulators,” Automatica, vol. 25, no. 4, pp. 509–519, Jul. 1989.
[44] M. W. Spong, “Adaptive control of flexible joint manipulators: Comments on two papers,” Automatica, vol. 31, no. 4, pp. 585–590, Apr.
1995.
16
[45] ——, “Modeling and control of elastic joint robots,” Journal of Dynamic
Systems, Measurement, and Control, vol. 109, no. 4, pp. 310–318, Dec.
1987.
[46] T. D. Ahlstrom, M. A. Diftler, R. B. Berka, J. M. Badger, S. Yayathi,
A. W. Curtis, and C. A. Joyce, “Robonaut 2 on the International Space
Station: Status update and preparations for IVA mobility,” in AIAA Space
Conference and Exposition, San Diego, CA, 2013, pp. 1–14.
[47] H. Wang, “On the recursive implementation of adaptive control for robot
manipulators,” in Chinese Control Conference, Beijing, China, 2010, pp.
2154–2161.
[48] ——, “Recursive composite adaptation for robot manipulators,” Journal
of Dynamic Systems, Measurement, and Control, vol. 135, no. 2, pp.
021 010–1–021 010–8, Mar. 2013.
| 3cs.SY
|
1
Operational Interpretation of Rényi Information
Measures via Composite Hypothesis Testing
Against Product and Markov Distributions
Marco Tomamichel† , Senior Member, IEEE and Masahito Hayashi‡∗ , Fellow, IEEE
arXiv:1511.04874v3 [cs.IT] 10 Dec 2017
Abstract
We revisit the problem of asymmetric binary hypothesis testing against a composite alternative hypothesis. We
introduce a general framework to treat such problems when the alternative hypothesis adheres to certain axioms. In
this case we find the threshold rate, the optimal error and strong converse exponents (at large deviations from the
threshold) and the second order asymptotics (at small deviations from the threshold). We apply our results to find
operational interpretations of various Rényi information measures. In case the alternative hypothesis is comprised
of bipartite product distributions, we find that the optimal error and strong converse exponents are determined by
variations of Rényi mutual information. In case the alternative hypothesis consists of tripartite distributions satisfying
the Markov property, we find that the optimal exponents are determined by variations of Rényi conditional mutual
information. In either case the relevant notion of Rényi mutual information depends on the precise choice of
the alternative hypothesis. As such, our work also strengthens the view that different definitions of Rényi mutual
information, conditional entropy and conditional mutual information are adequate depending on the context in which
the measures are used.
Index Terms
composite hypothesis testing, error exponent, strong converse exponent, second order, Rényi divergence, mutual
information, conditional entropy, conditional mutual information
I. I NTRODUCTION
Let us first consider simple hypothesis testing. Here the null hypothesis states that a random variable X n follows
the independent and identical (i.i.d.) law P ×n and the alternative hypothesis states that X n follows the i.i.d. law
Q×n , where P and Q are probability mass functions on a finite set X . We write this as follows:
null hypothesis: X n ∼ P ×n ,
alternative hypothesis: X n ∼ Q×n .
(1)
Assume now that our test T n is given as a randomized function from the observed event in X n to {0, 1}.
Here the values 1 and 0 signify that we accept and reject the null hypothesis, respectively. We are particularly
interested in the asymmetric case where two kind of errors are treated differently. The type-I error, given as
αn = P ×n [T n (X n ) = 0], is the probability that the test rejects the null hypothesis when it is true. The type-II
error, given as βn = Q×n [T n (X n ) = 1], is the probability that the test confirms the null hypothesis when the
alternative hypothesis is true. (See Section II-A for formal definitions of these quantities.)
On the one hand, if we impose a constant constraint on the type-I error, namely if we require that αn ≤ ε
for some ε ∈ (0, 1), then the there exists a sequence of tests such that βn goes zero exponentially fast in n. The
exponent is known to be the relative entropy, D(P kQ). This is Stein’s lemma [8] (see also [14], [5]) and we also
call this exponent the threshold rate of the problem. (See Section II-B for definitions of the relevant operational
†
Centre for Quantum Software and Information, University of Technology Sydney, Australia (Email: [email protected])
Graduate School of Mathematics, Nagoya University, Japan (Email: [email protected])
∗
Centre for Quantum Technologies, National University of Singapore (NUS), Singapore
This paper was presented in part at the 2015 International Symposium on Information Theory in Hong Kong, China.
‡
2
quantities.) Further, Yushkevich [64] derived the second order expansion (see also Strassen [54] for a higher order
expansion) of the optimal exponent as
p
√
− log βn = nD(P kQ) − nV (P kQ)Φ−1 (ε) + o( n) ,
(2)
where Φ is the cumulative standard normal distribution function and V is the variance of the logarithmic likelihood
ratio. (See Section II-C for definitions of the relevant information quantities.) All our statements are independent
of the basis choice for the logarithm as long as exp is taken as the inverse of log, unless it is otherwise noted.
On the other hand, if we impose an exponential constraint on the type-II error, namely if we require that
βn ≤ exp(−nR) for some rate R ∈ (0, D(P kQ)), we find that the optimal type-I error decreases exponentially
fast to zero with
1−s
− log αn = n sup
Ds (P kQ) − R + o(n) ,
(3)
s
0<s<1
where Ds (P kQ) is the Rényi relative entropy. This is known as Hoeffding’s bound [34], and the exponent is
called error exponent in the following. Moreover, if the rate R exceeds the threshold rate D(P kQ), the minimum
probability of the second error goes to 1 exponentially fast with
s−1
− log(1 − αn ) = n sup
R − Ds (P kQ) + o(n) .
(4)
s
s>1
This is called strong converse exponent in the following [12], [21].1
These results can partially be extended to the case when the null hypothesis is composite (i.e. when the null
hypothesis is comprised of a set of distributions) as a consequence of Sanov’s theorem [51]. In contrast, our
goal is to extend the above results, in particular Eq. (2)–(4), to the setting where the alternative hypothesis is
composite. More precisely, we want to consider a set Q of distributions on X and the maximal type-II error
βn = maxQ∈Q Q×n [T n (X n ) = 1]. We write the corresponding hypothesis testing problem as follows:
null hypothesis: X n ∼ P ×n ,
alternative hypothesis: X n ∼ Q×n for some Q ∈ Q.
(5)
Sanov’s theorem allows the alternative hypothesis to be the set Q = {Q : D(P kQ) > R} for a given real number
R > 0. More precisely, when the first kind of error probability is restricted to αn ≤ ε, the optimal exponent for βn
is given as − log βn = nR + o(n). Moreover, the Hoeffding bound (3) was extended to certain classes of composite
hypotheses which are composed of i.i.d. distributions [60], [52].
Our Contributions: Our first main result establishes that, if the alternative hypothesis satisfies certain axioms
discussed in Section II-D, the above results, Eq. (2)-(4), hold as stated after we substitute
D(P kQ) → min D(P kQ) = D(P kQ̂),
Q∈Q
V (P kQ) → V (P kQ̂),
Ds (P kQ) → min Ds (P kQ),
Q∈Q
(6)
where Q̂ ∈ Q is the distribution that minimizes the relative entropy. Hence, we generalize Stein’s lemma, Yushkevich’s second order expansion, the Hoeffding bound and the Han-Kobayashi bound to the case of a composite
alternative hypothesis. Moreover, we do not need to restrict the alternative hypothesis to i.i.d. distributions but can
allow permutation invariant or even more general distributions. We formally state all of our results in Section III.
Our second main result, which is an application of the first, is to give an operational interpretation to various
measures of Rényi mutual information, Rényi conditional entropy, and Rényi conditional mutual information. A
complete discussion of this can be found in Section IV, and here we exhibit a few representative examples:
1) Let (X, Y ) be two random variables governed by a joint probability distribution PXY with marginal PX . We
find that the hypothesis testing problem
null hypothesis:
alternative hypothesis:
1
×n
,
(X n , Y n ) ∼ PXY
X n ∼ PX×n independent of Y n ,
The form of (4) is due to Ogawa-Nagaoka [46]. Moreover, Nakagawa-Kanaya [45] first treat the case of large R.
(7)
3
originally proposed by Polyanskiy [47, Sec. II] in the context of channel coding, leads to an operational
interpretation of Sibson’s [53] definition (see also [10, p. 27]) of Rényi mutual information,
(8)
Is↑↓ (X : Y ) = min Ds (PXY kPX × QY ) .
QY
A similar hypothesis testing problem where the alternative hypothesis further requires that X n is uniform
leads to an operational interpretation of Arimoto’s definition of Rényi conditional entropy [3].
2) We further treat the problem of detecting correlations in a collection of random variables. Specifically, consider
a null hypothesis that the random random variables (X1 , X2 , . . . , Xk ) are governed by a specific distribution
PX1 X2 ...Xk and compare this to the alternative hypothesis that these random variables are independent, which
is a natural formulation from the viewpoint of statistics. This can be phrased as the hypothesis testing problem
null hypothesis:
alternative hypothesis:
,
(X1n , X2n , . . . , Xkn ) ∼ PX×n
1 X2 ...Xk
(X1n , X2n , . . . , Xkn ) ∼ QX1 × QX2 × . . . × QXk
×n
for some QXi .
(9)
The quantity D(PX1 X2 ...Xk kPX1 × PX2 × . . . × PXk ) is a measure of correlations for k-partite systems. We
show that it attains operational significance as a threshold rate for the above problem. We also derive error
exponents and strong converse exponents for this problem as long as R is close enough to the threshold rate.
These are determined by the quantities
min
QX1 ,QX2 ,...QXk
Ds (PX1 X2 ...Xk kQX1 × QX2 × . . . × QXk ) .
(10)
Our test depends on the specific distribution PX1 X2 ...Xk , so it is not able to detect arbitrary correlations in
these random variables.
3) Let PXY Z be a joint probability distribution. We find that the hypothesis testing problem
null hypothesis:
alternative hypothesis:
×n
(X n , Y n , Z n ) ∼ PXY
Z ,
(X n , Y n , Z n ) ∼ Q×n
XY Z for some QXZY that is Markov X ↔ Y ↔ Z ,
(11)
yields an operational interpretation for the conditional mutual information, I(X : Z|Y ), as the threshold rate.
Moreover, the error exponents are determined by a certain Rényi conditional mutual information,
Is↓↓↓ (X : Z|Y ) = min Ds (PXY Z kQY × QX|Y × QZ|Y ) .
QXY Z
(12)
However, this definition of Rényi conditional mutual information is by no means the only definition that
attains operational significance. If we vary the problem slightly and only consider alternative hypotheses with
×n
we recover the same threshold rate but different exponents determined by
a fixed marginal (X n , Y n ) ∼ PXY
Is↑↑↓ (X : Z|Y ) = min Ds (PXY Z kPXY × QZ|Y ) .
QZ|Y
(13)
Our results thus yield an operational interpretation for this definition of Rényi conditional mutual information
for all positive s, and we provide a closed form of this quantity in (75). To the best of our knowledge this
definition has not appeared in the literature before. From the operational perspective we have chosen here, it
is a natural extension of Sibson’s definition of Rényi mutual information to the conditional setting, and we
expect it to have other applications in information theory.
On a technical level, our work introduces the concept of a universal distribution and a universal channel. The
purpose of the former is to dominate any i.i.d. product (or permutation invariant) distribution in terms of the relative
entropy and the Rényi relative entropy (cf. Lemmas 7 and 14). More formally, we show that there exists a sequence
of distributions U n on X n such that Ds (P n kQ×n ) ≥ Ds (P n kU n ) + O(log n) for any P n , Q uniformly for all
s ≥ 0. Similarly, the output distribution of the universal channel dominates the output distribution of any memoryless
product (or permutation covariant) channel, whenever both channels are given the same input (cf. Lemma 17). A
similar kind of approximation was discussed perviously by Davisson [13] (see also Clarke and Barron [9]) but they
only showed the approximation (using a Baysian mixture) for the case s = 1 corresponding to the relative entropy.2
2
Their result was extended in a recent paper [29] to the approximation in terms of the Rényi relative entropy, even in the continuous case.
4
We also note that the concept of universal decoding was studied in [17], but in this work the universal decoder
can only dominate finitely many decoders whereas our universal states and channels dominate a continuous set of
states and channels, respectively. Therefore, the methods in References [13], [9], [17] cannot be directly applied to
our analysis. The universal distribution is the classical analogue of the universal state originally introduced in the
quantum setting by one of the authors in [26] and [25]. The latter paper also introduced universal classical-quantum
channels (see also [16] for a fully quantum universal channel).
Another contribution is the axiomatic approach we have taken to the problem. We derive a sufficient condition
for the hypotheses testing problems to derive analogues of Yushkevich’s bound, Hoeffding’s bound and the HanKobayashi bound. Since this approach accepts hypotheses containing non-i.i.d. distributions, we expect it to have
wide applicability.
Outline: The remainder of the paper is structured as follows. In Section II we introduce the axiomatic framework
for composite hypothesis testing that we build on, and also define our information quantities, the Rényi divergences.
In Section III we present our main results and treat some generic examples, with the proofs deferred to the later
sections. The examples discussed in the introduction, which yield an operational interpretation of various notions
of Rényi conditional entropy, mutual information, and conditional mutual information, are then treated in detail in
Section IV. In fact, Section IV can be understood without referring to Section III. The proofs for the Hoeffding
bound are discussed in Section V, the Han-Kobayashi bound follows in Section VI and the second order analysis
of Stein’s lemma is found in Section VII. We conclude our work with a discussion and outlook in Section VIII.
II. A F RAMEWORK
FOR
C OMPOSITE H YPOTHESIS T ESTING
In this section we introduce a general framework for composite hypothesis that encompasses, but is not restricted
to, the examples mentioned in the introduction.
A. Binary Hypothesis Testing with Composite Alternative Hypothesis
We restrict our attention to finite alphabets. Let X be such an alphabet. The set of probability mass functions (in
the following
Poften just called distributions) on X is denoted by P(X ) and comprised of positive valued functions
on X with x∈X P (x) = 1. For P ∈ P(X ) and a random variable X on X , we write X ∼ P to denote that X
is distributed according to the law P . we use X n to denote the n-fold Cartesian product of X and its elements by
vectors xn = (x1 , x2 , . . . , xn )Q
. For any P ∈ P(X ), we use P ×n to denote the identical and independent distribution
×n
n
(i.i.d.) given by P (x ) = ni=1 P (xi ).
We consider hypothesis testing problems with a composite alternative hypothesis of the following form.
Definition 1. A sequence of hypothesis
testing problems with composite alternative hypothesis is determined by
a triple H = X , P, {Qn }n∈N , comprised of a finite set X , a distribution P ∈ P(X ), and a sequence of sets
Qn ⊆ P(X n ) for all n ∈ N. This determines a hypothesis testing problem Hn for each n ∈ N. Namely, we consider
n instances of a random variable X on X and the following two hypotheses.
Hn :
null hypothesis: X n ∼ P ×n ,
alternative hypothesis: X n ∼ Qn for some Qn ∈ Qn .
(14)
We will analyze this problem for sequences of sets Qn that satisfy certain axioms (cf. Section II-D).
For convenience we employ the shorthand notation Q ≡ Q1 and we use Qn to denote the convex hull of
Qn . Consider probabilistic hypothesis tests, given by a function T : X → [0, 1], where T (x) is the probability
of accepting the null hypothesis when observing x. We then define the type-I error probability and type-II error
probability, respectively, as follows:
X
X
α(T ; P ) :=
P (x)(1 − T (x)),
and
β(T ; Q) := sup
Q(x)T (x) .
(15)
Q∈Q x∈X
x∈X
In this work we focus on asymmetric hypothesis testing. In this context it is convenient to define the quantity
α̂(µ; P kQ) as the minimum type-I error probability when the type-II error probability is below a threshold µ ≥ 0,
i.e. we consider the following optimization problem:
α̂(µ; P kQ) := min α(T ; P ) β(T ; Q) ≤ µ ,
for µ ∈ R .
(16)
T
5
Note that µ 7→ α̂(µ; P kQ) is monotonically decreasing and evaluates to 0 for µ ≥ 1. Note that we always have
α̂(µ; P kQ) = α̂(µ; P kQ) .
(17)
Moreover, since the sum over which we take the supremum in (15) is linear in Q and the maximum is thus achieved
on the boundary.
B. Operational Quantities for Asymmetric Hypothesis Testing
Let us now discuss the main operational quantities that we want to investigate.
1) Threshold Rate: We will study the following properties of asymmetric composite hypothesis tests. The first
concept concerns the threshold rate of a sequence of such tests.
Definition 2. Let H = X , P, {Qn }n∈N be a sequence of hypothesis testing problems. We define the threshold rate
of the sequence H as
×n
(18)
Rth (H) := sup R ∈ R : lim sup α̂ exp(−nR); P kQn = 0 .
n→∞
Similarly, we define the strong converse threshold rate of the sequence H as (cf. [22, Def. 4.3.1] and [44])
∗
(H) = inf R ∈ R : lim inf α̂ exp(−nR); P ×n kQn = 1 .
Rth
n→∞
(19)
∗ (H) always holds. Moreover, we note that the threshold rate is always nonnegative because
Clearly Rth (H) ≤ Rth
×n
α̂ exp(−nR); P kQn vanishes for R ≤ 0. The problems we study in this paper are particularly well-behaved
and we will find that the threshold rate and the strong converse threshold rate agree.
2) Error and Strong Converse
Exponents: Moreover, if we choose a rate R below Rth (H) then we will observe
that α̂ exp(−nR); P ×n kQn converges exponentially fast to 0 as n increases. The exponent characterizing this
decrease is called the error exponent of the sequence with regards to R.
Definition 3. Let H = X , P, {Qn }n∈N be a sequence of hypothesis testing problems. For every R > 0, the error
exponent of H with regards to R is defined as
1
eR (H) := lim inf − log α̂ exp(−nR); P ×n kQn ,
(20)
n→∞
n
if this limit exists, or +∞ otherwise.
If we choose a rate R exceeding Rth (H) then we will instead observe that α̂ exp(−nR); P ×n kQn converges
exponentially fast to 1 as n increases. The exponent characterizing this convergence is called the strong converse
exponent of the sequence with regards to R.
Definition 4. Let H = X , P, {Qn }n∈N be a sequence of hypothesis testing problems. For every R > 0, the strong
converse exponent of H with regards to R is defined as
1
scR (H) := lim sup − log 1 − α̂ exp(−nR); P ×n kQn ,
(21)
n
n→∞
if this limit exists, or +∞ otherwise.
C. Information Measures for Asymmetric Hypothesis Testing
All our results will be stated as an equivalence between one of the above-mentioned operational quantities and
an information measures derived from the Rényi divergence. We formally define the Rényi divergence [49] here.
Definition 5. Let P ∈ P(X ) and Q : X → [0, ∞). For s ∈ (0, 1) ∪ (1, ∞), define
X
X
1−s
P (x)s Q(x)1−s ,
P (x)s Q(x) + ε
=
gs (P kQ) := lim
ε→0
x∈X
x∈X
P (x)>0
(22)
6
where the latter expression is finite either if s < 1 or if Q(x) = 0 implies P (x) = 0 for all x ∈ X . Otherwise, we
set gs (P kQ) = +∞. The Rényi divergence of P with regards to Q of order s is then defined as
Ds (P kQ) :=
log gs (P kQ)
s−1
(23)
For s ∈ {0, 1, ∞} the Rényi divergence is defined as the corresponding limit.
See [61] for a comprehensive discussion of its properties, which we summarize here. The Rényi divergence is
lower semi-continuous and diverges to +∞ if the support condition is violated and s > 1 or if s < 1 and P and
Q do not share any support, in which case gs (P kQ) = 0. Our sets Qn may contain elements that violate these
conditions. Nonetheless, sets consistent with our axioms will always contain at least one element that satisfies the
support conditions. One of the most important properties of the latter functional is that gs (·k·) is jointly concave for
s ∈ (0, 1) and jointly convex for s ∈ (1, ∞). Moreover, the function s 7→ log gs (P kQ) is convex and s 7→ Ds (P kQ)
is monotonically increasing. Furthermore, the Kullback-Leibler divergence is given by
X
D(P kQ) := D1 (P kQ) =
P (x) log P (x) − log Q(x) .
(24)
x
Q′
For two positive valued functions Q and
on X , we observe that Q(x) ≤ Q′ (x) for all x ∈ X implies Ds (P kQ) ≥
Ds (P kQ′ ). Furthermore, for any scalar v , we have Ds (P kvQ) = Ds (P kQ) − log v .
To present the second order of Stein’s lemma we need to introduce some additional quantities. The variance of
the logarithmic likelihood ratio is given by
X
2
(25)
V (P kQ) :=
P (x) log P (x) − log Q(x) − D(P kQ) .
x∈X
This variance is proportional to the derivative of the Rényi divergence at s = 1, a consequence of the fact that
the Rényi divergence is proportional to the cumulant generating function of the logarithmic likelihood ratio. More
precisely, the first order Taylor expansion of Ds (P kQ) around s = 1 is given by
Ds (P kQ) = D(P kQ) +
s−1
V (P kQ) + O(s2 ) .
2 log e
(26)
Finally, we define the Rényi divergence of P with regards to a set Q of positive valued functions as
Ds (P kQ) := inf Ds (P kQ) .
Q∈Q
(27)
The minimizer, if it is unique, is defined as
Q̂s := arg min Ds (P kQ) .
(28)
Q∈Q
We define V (P kQ) := V (P kQ̂1 ). Similarly, taking note of the sign of (s − 1), we define
sup gs (P kQ) if s ∈ (0, 1)
Q∈Q
gs (P kQ) := exp (s − 1)Ds (P kQ) =
.
inf gs (P kQ) if s ∈ (1, ∞)
(29)
Q∈Q
D. Axioms for Alternative Hypotheses
Let us fix a probability distribution P ∈ P(X). We present a collection of axioms that the sets {Qn }n∈N must
satisfy in order for our main results to hold. The first axiom ensures that the base set, Q, is convex.
Axiom 1 (convexity). The set Q ⊆ P(X ) is compact convex. Moreover, for all s > 0 the minimizer Q̂s in (28) is
unique and lies in relint(Q).3
The second axiom ensures that i.i.d. products of distributions in Q are elements of Qn .
Axiom 2 (product distributions). The set Qn contains the element Q×n for every Q ∈ Q.
3
The relative interior of a convex set Θ is the set relint(Θ) := {x ∈ Θ : ∀y ∈ Θ ∃λ > 1 s.t. λx + (1 − λ)y ∈ Θ}.
7
As a direct consequence of Axiom 2 and the additivity of the Rényi divergence for product distributions, we find
the following lemma.
Lemma 6 (subadditivity). Assuming Axiom 2, we have Ds (P ×n kQn ) ≤ nDs (P kQ) for all s > 0 and n ∈ N.
The purpose of the next axiom is to ensure that this is in fact an equality.
Axiom 3 (additivity). For all s > 0 and n ∈ N, we have Ds (P ×n kQn ) ≥ nDs (P kQ).
The next axiom concerns the existence of a sequence of universal distributions. Before we state it, let us introduce
some additional notation. We denote the symmetric group of permutations of n elements by Sn . This group has
a natural representation as bistochastic matrices. For every π ∈ Sn , the matrix W n [π] is defined by the relation
P n W n [π](x1 , x2 , . . . , xn ) = P n (xπ(1) , xπ(2) , . . . , xπ(n) ). We say that a probability distribution P n ∈ P(X n ) is
permutation invariant if P n W n [π] = P n holds for all π ∈ Sn . The set of all permutation invariant distributions on
X n is denoted P sym (X n ) and permutation invariant distributions in Qn comprise the subset Qsym
n .
Axiom 4 (universal distribution). There exists a sequence of probability mass functions {U n }n∈N with U n ∈
P sym (X n ) and a polynomial v(n) such that the following relation holds. For all n ∈ N and Q ∈ Qsym
n ,
Q(xn ) ≤ v(n) U n (xn ),
∀ xn ∈ X n ,
and
Ds (P ×n kU n ) ≥ Ds (P ×n kQn ) .
The former condition is rewritten as Dmax (QkU n ) ≤ log v(n) for Q ∈ Qsym
n , where Dmax (QkP ) := log maxx
sym
1 P
n
n
n
Moreover, the set Qn is closed under symmetrization, i.e. if P ∈ Qn then n!
π∈Sn P W [π] ∈ Qn .
(30)
Q(x)
P (x) .
sym
The latter condition in (30) is automatically satisfied if U n ∈ Qn , but this is not necessary. The above axioms
have the following immediate consequence.
Lemma 7. Assume Axioms 3 and 4 hold. Then, for all s > 0
1
lim Ds (P ×n kU n ) = Ds (P kQ) ,
n→∞ n
and the map s 7→ φ(s) := log gs (P kQ) is convex.
(31)
Proof: We first show (31). Additivity implies that n1 Ds (P ×n kU n ) ≥ n1 Ds (P ×n kQn ) = Ds (P kQ). To establish
the other direction, we use Axiom 4. For any Q ∈ Q, we have
1
1
1
Ds (P ×n kQ×n ) + log v(n) = Ds (P kQ) + log v(n)
Ds (P ×n kU n ) ≤
(32)
n
n
n
Hence, minimizing over all such Q we can replace Ds (P kQ) with Ds (P kQ) on the right hand side. Moreover,
using the property that v(n) grows polynomially in n, we find lim supn→∞ n1 Ds (P ×n kU n ) ≤ Ds (P kQ). Finally,
since s 7→ log gs (P kQ) is the point-wise limit of convex functions, it is also convex.
Finally, we note that convexity in Axiom 1 is quite a strong requirement and not satisfied by some of the examples
we consider. Instead of requiring that the set Q is convex, it suffices to assume that there exists a convenient convex
parametrization of the set such that concavity and convexity of gs (P k·) are preserved.
Axiom 5 (convex parametrization, replaces Axiom 1). There exists a compact convex set Θ in a finite-dimensional
vector space, a twice continuously differentiable (C 2 ) function Θ ∋ θ 7→ Qθ ∈ Q, and an open interval (a, b)
containing 1 such that the following holds:
• We have Q = {Qθ : θ ∈ Θ}.
• For all s ∈ (a, b), the minimizer θ̂ s := arg minθ∈Θ Ds (P kQθ ) is unique and lies in relint(Θ).
• The map θ 7→ gs (P kQθ ) has negative definite Hessian at θ̂ s for all s ∈ (a, 1) and positive definite Hessian at
θ̂ s for all s ∈ (1, b). Moreover, the map θ 7→ Ds (P kQθ ) has positive definite Hessian at θ̂ 1 .
Since the map in Axiom 5 is assumed to be C 2 and (s, Q) 7→ Ds (P kQ) is C 2 , Axiom 5 implies that (s, θ) 7→
Ds (P kQθ ) is C 2 as well. Further note that Axiom 1 implies Axiom 5 using the trivial parametrization Θ = Q and
(a, b) = (0, ∞).4 If we assume Axiom 5 instead of Axiom 1 we must also relax the additivity property. Namely,
additivity in Axiom 3 is only required in the interval s ∈ (a, b) and Lemma 7 only holds for s ∈ (a, b).
4
However, the converse argument is not true because for any two points θ and θ′ and λ ∈ (0, 1), the distribution Qλθ+(1−λ)θ′ ∈ Q does
not necessarily equal the distribution λQθ + (1 − λ)Qθ′ ∈ P(X ). That is, the convex combination in the parameter space is different from
the convex combination in P(X ), in general. (Examples will be given in Section III-B.)
8
III. M AIN R ESULTS
AND
E XAMPLES
A. Statement of Main Results
Our first result considers the asymptotic situation where the type-II error probability goes to zero exponentially
with a rate below D(P kQ). In this case, we find that type-I error probability converges to zero exponentially fast,
and the exponent is determined by the Rényi divergence, Ds (P kQ) with s < 1.
To state our result we need the following concept.
Definition 8. Fix P ∈ P(X ) and Q ⊆ P(X ). For any c ≥ 0, the c-critical rate is defined as
Rc := lim sφ′ (s) − φ(s)
s→c
(33)
with φ(s) = (s − 1)Ds (P kQ) as defined in Lemma 7.
The map c 7→ Rc on (a, b) is monotonically increasing (cf. Lemma 20), and furthermore we find R0 = D0 (P kQ)
and R1 = D(P kQ), as well as R∞ ≥ D∞ (P kQ).
Theorem 9. Let H = X , P, {Qn }n∈N be such that Axioms 2–5 are satisfied on (a, 1]. Then, for any R > Ra ,
1−s
Ds (P kQ) − R .
(34)
eR (H) = sup
s
s∈(a,1)
The proof is given in Section V. The case where Qn = {Q×n } are singletons is attributed to Hoeffding [34].
Note that if R ≥ D(P kQ) the right hand side of (34) evaluates to zero, revealing that in this case the error of the
first kind will decay slower than exponential in n. Otherwise the right hand side is always positive.
Our second result considers the case where type-II error probability goes to zero exponentially with a rate
exceeding the mutual information D(P kQ). In this case, we find that type-I error probability converges to 1
exponentially fast, and the exponent is determined by the Rényi divergence Ds (P kQ), with s > 1.
Theorem 10. Let H = X , P, {Qn }n∈N be such that Axioms 2–5 are satisfied on [1, b). Then, for any R < Rb ,
s−1
R − Ds (P kQ) .
(35)
scR (H) = sup
s
s∈(1,b)
The proof is given in Section VI. The case where Qn are singletons is attributed to Csiszár-Longo [12] and
Han-Kobayashi [21]. Note that even in the singleton case, the original results do not apply for R > R∞ . In fact
Nakagawa-Kanaya [45] showed that in this setting the above optimal exponent can be attained only by a randomized
test. We will not further discuss this case.
Again we note that if R ≤ D(P kQ) the right hand side of (35) evaluates to zero, otherwise it is strictly
positive. The threshold rates are thus determined by the above results, and we find the following corollary of
Theorems 9 and 10.
Corollary 11. Let H = X , P, {Qn }n∈N be such that Axioms 2–5 are satisfied with any (a, b) containing 1. Then,
∗
Rth (H) = Rth
(H) = D(P kQ) .
(36)
For completeness, we also investigate the second order behavior, namely we investigate the error of the first kind
√
when the error of the second kind vanishes as exp(−nD(P kQ) − nr). This analysis takes a step beyond Stein’s
lemma and extends Yushkevich’s work for simple alternative hypotheses [64].
Theorem 12. Assume Axioms 2–5 hold for any (a, b) ∋ {1}. Then, for any r ∈ R, we have
!
√ ×n
r
Qn = Φ p
,
lim α̂ exp − nD(P kQ) − nr ; P
n→∞
V (P kQ)
y2
1 R x
where Φ(x) := (2π)− 2 −∞ e− 2 dy and V (P kQ) is defined in the line following (28).
(37)
The proof is given in Section VII. The achievability proof is of a different flavor than previous proofs of the
singleton case and relies on Lévy’s continuity theorem.
9
B. Examples
In the following we will discuss various examples of hypothesis testing problems that can be tackled with the
above framework. The cases we treat here in particular cover the examples in Section IV.
1) Product distributions with a fixed marginal: Let X and Y be two finite sets. Consider a pair of random
variables (X, Y ) ∼ PXY that are governed by a joint probability distribution PXY ∈ P(X × Y). We denote by
PX|Y =y the distribution of X conditioned on the event Y = y . The conditional distribution PX|Y is interpreted as a
stochastic matrix mapping or channel from Y to X . In particular, we write PXY = PY × PX|Y and PX = PY PX|Y .
We assume without loss of generality that PX and PY have full support, i.e. we restrict the sets X and Y to the
support of the marginals PX and PY , respectively. Moreover, let TX ∈ P(X ) be such that TX has full support as
well. Now consider the sets
(38)
Qn = TX×n × QY n : QY n ∈ P(Y n ) .
We emphasize that the distributions QY n are unstructured, in particular they do not have to be n-fold i.i.d. products.
Proposition 13. The sequence of tests X × Y, PXY , {Qn }n∈N with Qn in (38) satisfies Axioms 1–4. Moreover,
this still holds if we restrict Qn to permutation invariant or i.i.d. product distributions.
The proof is given in Appendix A-B and relies on the following Lemma, which might be of independent interest.
sym
n
n }
(X n ) such that, for every n ∈ N
Lemma 14. The exists a sequence of distributions {UX
n n∈N with UX n ∈ P
n (xn ) for all xn ∈ X n .
and QX n ∈ P sym (X n ), we have QX n (xn ) ≤ |Tn (X )| UX
n
In the above lemma we used Tn (X ) to denote the set of X -types of length n. This is the classical analogue of
the universal state originally introduced in the quantum setting [26], [25]. The proof uses the method of types [11],
and the result is only sensible for finite sets.
Proof: The universal distributions are given by
X
1
1
n
n
UX
1{xn is of type λ} ,
(39)
n (x ) =
|Tn (X )| |λ|
λ∈Tn (X )
1
1{xn is of type λ} is the uniform distribution over all sequences of type λ. Every permutation invariant
where |λ|
distribution in QX n ∈ P sym (X ) has to be flat over sequences of the same type. Namely, it has to be of the form
X q(λ)
QX n (xn ) =
1{xn is of type λ}
(40)
|λ|
λ∈Tn (X )
for some distribution q ∈ P(Tn (X )). The desired bound can now be verified easily.
2) General (permutation invariant) product distributions: Consider finite sets X1 , X2 , . . . , Xk and a distribution
PX1 X2 ...Xk Y ∈ P(X1 ×X2 ×. . .×Xk ×Y). Without loss of generality we assume that all the marginals of PX1 X2 ...Xk Y
have full support. Then, consider
Qn = QX1n × QX2n × . . . × QXkn × QY n : QXin ∈ P sym (Xin ), i ∈ [k] and QY n ∈ P(Y n ) , .
(41)
Note that these sets are not convex, so Axiom 1 is certainly violated. Moreover, note that the restriction that the
QXin be permutation invariant is necessary. Without such a restriction, even a correlated null hypothesis lies in the
convex hull of the set of alternative hypothesis since P(X × Y) equals the convex hull of P(X ) × P(Y). Clearly
it is then no longer possible to distinguish these two hypotheses.
Proposition 15. The sequence of tests
X
×
X
×
.
.
.
×
X
×
Y,
P
,
{Q
}
with Qn in (41) satisfy
1
2
X
X
...X
Y
n
n∈
N
k
1
2
k
k
, ∞ . Moreover, this still holds if we restrict Qn to i.i.d. product distributions.
Axioms 2–5 with (a, b) = k+1
The proof is given in Appendix A-C.
10
3) Recovered and other Markov distributions: Let PXY Z ∈ P(X ×Y ×Z) be a joint probability distribution with
marginals PX and PY and PZ . A natural test considers alternative hypothesis comprised of Markov distributions
where only two marginals are fixed. We can see this as the problem of distinguishing a fixed tripartite distribution
PXY Z from the set of distributions that can be “recovered” from its marginal PXY via a probabilistic operation.
We assume without loss of generality that PX (x) and PY (y) and PZ (z) have full support, i.e. we restrict the
sets X , Y , Z to the support of the marginals PX , PY and PZ , respectively. The set of conditional probability
distributions of Z given Y , or channels from Y to Z , is denoted by P(Z|Y). Consider the sets
×n
× QZ n |Y n : QZ n |Y n ∈ P(Z n |Y n ) .
(42)
Qn = PXY
Proposition 16. The sequence of tests X ×Y ×Z, PXY Z , {Qn }n∈N with Qn in (42) satisfy Axioms 1–4. Moreover,
this still holds if we restrict Qn to permutation invariant or i.i.d. product distributions.
This proposition relies on the following lemma, which is of independent interest.
Lemma 17. There exists a sequence of channels, {UYn n |X n }n∈N , where UYn n |X n ∈ P sym (Y n |X n ) such that the
following holds. For every n ∈ N, QY n |X n ∈ P sym (Y n |X n ) and PX n ∈ P(X n ), we have
PX n × QY n |X n (xn , y n ) ≤ |Tn (X × Y)| PX n × UYn n |X n (xn , y n ),
∀xn ∈ X n , y n ∈ Y n .
(43)
Both of the above statements, Proposition 16 and Lemma 17, are proven in Appendix A-D.
Another natural question is to distinguish between a null hypothesis PXY Z and all Markov distributions X ↔
Y ↔ Z , i.e. distributions QXY Z = QY × QX|Y × QZ|Y . Consider the set
Qn = QX n Y n Z n ∈ P sym (X ×Y ×Z)n : QX n Y n Z n = QY n × QX n |Y n × QZ n |Y n .
(44)
Proposition 18. The sequence of tests X × Y × Z, PXY Z , {Qn }n∈N with Qn in (44) satisfy Axioms 2–5 with
(a, b) = ( 23 , ∞). Moreover, this still holds if we restrict Qn to i.i.d. product distributions.
Again note that every distribution is contained in the convex hull of all Markov distributions, and hence some
restrictions on the set are necessary. It is possible to slightly weaken the condition that (X n , Y n , Z n ) is permutation
invariant, but we will not discuss this here. This is shown in Appendix A-E.
IV. O PERATIONAL I NTERPRETATION
OF
R ÉNYI I NFORMATION M EASURES
In this section we present the main application of our results, finding operational interpretations of various
measures of Rényi mutual information, conditional entropy and conditional mutual information.
A. Rényi Mutual Information: Testing Against Independent Distributions
It is well known that the mutual information can be expressed in terms of the Kullback-Leibler divergence in
several ways. Consider two random variables X and Y and a joint distribution PXY ∈ P(X × Y) with marginals
PX and PY . We are interested in the identities
I(X : Y ) = D(PXY kPX × PY )
=
=
min
QY ∈P(Y )
D(PXY kPX × QY )
min
QX ∈P(X), QY ∈P(Y )
D(PXY kQX × QY )
(45)
(46)
(47)
Each of these identities gives rise to a different hypothesis testing problem and a different notion of Rényi mutual
information. In the following we treat these three problems in the above order.
11
1) All marginals fixed: As a warmup consider the following (simple) hypothesis testing problem.
Hmi
↑↑ :
×n
,
null hypothesis: (X n , Y n ) ∼ PXY
alternative hypothesis: X n ∼ PX×n and Y n ∼ PY×n are independent.
(48)
That is, we set Qn = {PX×n × PY×n } for all n in Defintion 1. Stein’s Lemma and its strong converse ensure that
mi
∗
Rth (Hmi
↑↑ ) = Rth (H↑↑ ) = D(PXY kPX × PY ) = I(X : Y ) .
(49)
Moreover, the Hoeffding [34] and Han-Kobyashi [21] bounds give an operational interpretation for the following
Rényi mutual information:
X
X
1
Is↑↑ (X : Y ) := Ds (PXY kPX × PY ) =
PY (y)
PX|Y =y (x)s PX (x)1−s .
log
(50)
s−1
y∈Y
x∈X
As an example, in wire-tap channel coding, this kind of Rényi mutual information is used to express the exponents
of leaked mutual information [27], [23].
2) One marginal fixed: Here we consider a hypothesis test where the alternative hypothesis is comprised of
product distributions where one marginal is fixed. This is the example discussed in Section III-B1 with TX = PX .
This hypothesis test figures prominently when analyzing the converse to various channel coding questions in
the fixed error regime, for example for second-order analysis of the discrete memoryless channels [48], [24] and
beyond [57]. In this context, Polyanskiy [47, Sec. II] raised the following hypothesis testing problem:
×n
,
null hypothesis: (X n , Y n ) ∼ PXY
alternative hypothesis: X n ∼ PX×n independent of Y n .
(51)
This problem has the same threshold, I(X : Y ), but gives an operational interpretation for Sibson’s [53] definition
of Rényi mutual information, which is given by
Is↑↓ (X : Y ) :=
=
=
min
Ds (PXY kPX × QY )
!1
s
X
X
log
PY (y)
PX|Y =y (x)s PX (x)1−s
QY ∈P(Y)
s
s−1
s
E0
s−1
y∈Y
(52)
(53)
x∈X
s−1
, PX ,
s
(54)
where E0 is Gallager’s error exponent function [20].5 The explicit form of the distribution QY that achieves the
minimum is given by Sibson’s identity (cf. Appendix A).
Our results for the optimal error and strong converse exponents then read
1 − s ↑↓
s−1
eR (Hmi
Is (X : Y ) − R and scR (Hmi
R − Is↑↓ (X : Y ) ,
(55)
↑↓ ) = sup
↑↓ ) = sup
s
s
s>1
s∈(0,1)
mi
In the setting of channel coding with constant composition codes of type PX , the exponents eR (Hmi
↑↓ ) and scR (H↑↓ )
are equal to the error exponent [20] and the strong converse exponents [2], respectively. In wire-tap channel coding,
it is used for expressing the exponents of leaked information when the leaked information is measured in terms of
the variational distance [28, Thm. 5].
5
Verdú [62] recently surveyed Sibson’s definition and pointed out its favorable mathematical properties in the case of general alphabets.
12
3) Arbitrary product distributions, permutation invariant: Let us now consider the most general problem,
×n
,
null hypothesis: (X n , Y n ) ∼ PXY
Hmi
↓↓ :
alternative hypothesis: X n and Y n independent, (X n , Y n ) permutation invariant.
(56)
This is covered by the example in Section III-B2. In fact, it is sufficient to require that either X n or Y n be
permutation invariant (cf. Proposition 15).
We find that for R sufficiently close to the threshold I(X : Y ), we have
1 − s ↓↓
s−1
eR (Hmi
sup
Is (X : Y ) − R and scR (Hmi
R − Is↓↓ (X : Y ) ,
(57)
↓↓ ) =
↓↓ ) = sup
s
s
s>1
s∈( 1 ,1)
2
with a different definition of Rényi mutual information,
Is↓↓ (X : Y ) :=
min
QX ∈P(X ), QY ∈P(Y)
Ds (PXY kQX × QY ) .
(58)
Our result gives an operational interpretation for this definition of Rényi mutual information. However, this operational interpretation only applies for s ≥ 12 . In fact, it is unclear if our results can be extended to smaller s and we
do not know of a closed form expression for this quantity. While this work was completed, the properties of this
definition have been independently studied in [37].
B. Rényi Conditional Entropy: Testing Against Uniform and Independent Distribution
The conditional entropy can be expressed in terms of the Kullback-Leiber divergence very similarly to the mutual
information, with the difference that we require one marginal to be fixed to a uniform distribution. This leads to
the following two expressions:
(59)
H(X|Y ) = log |X | − D(PXY kRX × PY )
= log |X | −
min
QY ∈P(Y)
D(PXY kRX × QY ),
(60)
where RX is the uniform distribution over X and, as in the previous section, X and Y are two random variables
governed by a joint distribution PXY ∈ P(X × Y) with marginals PX and PY . Note that the term log |X | can
easily be incorporated in the relative entropy term if we do not require the second argument to be a normalized
probability distribution but instead allow arbitrary positive distributions. Our results extend to this more general
setup but we will restrict to normalized distributions as otherwise the corresponding hypothesis testing problems
are unnatural.
1) Fixed marginal distribution: Consider the following hypothesis testing problem:
Hc↑ :
×n
,
null hypothesis: (X n , Y n ) ∼ PXY
×n
and Y n ∼ PY×n are independent.
alternative hypothesis: X n ∼ RX
(61)
We find that the threshold rate is D(PXY kRX × PY ) = log |X | − H(X|Y ) and the Hoeffding [34] and HanKobyashi [21] bounds establish an operational meaning for the Rényi conditional entropy
X
X
1
Hs↓ (X|Y ) := log |X | − Ds (PXY kRX × PY ) =
PY (y)
PX|Y =y (x)s .
log
(62)
1−s
y∈Y
x∈X
This definition of conditional Rényi entropy is for example used to express the leaked modified mutual information
in the secure random number generation [27, Thm. 2] and [31, Thm. 2].
13
2) Arbitrary marginal distribution: We consider the following problem, in analogy with Section IV-A2:
×n
,
null hypothesis: (X n , Y n ) ∼ PXY
Hc↓ :
×n
independent of Y n .
alternative hypothesis: X n ∼ RX
(63)
This is covered by the example in Section III-B1 with TX the uniform distribution. Again we determine the threshold
log |X | − H(X|Y ) and the error and strong converse exponents given operational significance to Arimoto’s [3]
definition of Rényi conditional entropy,
!1
s
X
X
s
s
.
PY (y)
PX|Y
log
(64)
Hs↑ (X|Y ) := log |X | − min Ds (PXY kRX × QY ) =
=y
1−s
QY ∈Q(Y)
y∈Y
x∈X
The minimum was evaluated using Sibson’s identity (cf. Lemma 29 in Appendix A).
This definition has recently be reviewed in [18] and compares favorably to other definitions of Rényi conditional
entropy that have recently been put forward [55], [35]. For example, it has an operational interpretation determining
the moments of the number of rounds required to guess X from Y [1], and relatedly as an exponent for task encoding
with side information [6]. More precsiely, we have
1−s
s−1
eR (Hc↑↓ ) = sup
log |X | − Hs↑↓ (X : Y ) − R and scR (Hc↑↓ ) = sup
R − log |X | + Hs↑↓ (X : Y ) ,
s
s
s>1
s∈(0,1)
(65)
As a further example, in secure random number extraction, sclog |X |−R (Hc↑↓ ) expresses the error exponent under the
universal composability criterion [28, Thm. 4] and [33, Thm. 30].
C. Rényi Conditional Mutual Information: Testing Against Markov Distributions
Let X, Y and Z be three random variables with a joint distribution PXY Z ∈ P(X × Y × Z). The conditional
mutual information, I(X : Z|Y ), can be seen as a measure of how close the distribution PXY Z is to a Markov
chain X ↔ Y ↔ Z . For example, we can write
(66)
I(X : Z|Y ) = D(PXY Z kPY × PX|Y × PZ|Y )
=
=
min
QZ|Y ∈P(Z|Y )
(67)
D(PXY Z kPY × PX|Y × QZ|Y )
min
QXY Z ∈P(X ×Y×Z)
QXY Z =QX|Y ×QZ|Y ×QZ
(68)
D(PXY Z kQY × QX|Y × QZ|Y ) ,
where the latter optimization is over all distributions satisfying the Markov condition QXY Z = QX|Y × QZ|Y × QZ .
These are only a few of all the possible expressions for the conditional mutual information, but we will focus our
attention on these examples and follow a similar discussion as with the mutual information.
1) All marginals fixed: Again we first consider a simple alternative hypothesis.
Hcmi
↑↑↑ :
×n
null hypothesis: (X n , Y n , Z n ) ∼ PXY
Z,
×n
(69)
, and (Y n , Z n ) ∼ PY×n
alternative hypothesis: X n ↔ Y n ↔ Z n is Markov, (X n , Y n ) ∼ PXY
Z.
This corresponds to the sets Qn := (PY × PX|Y × PZ|Y )×n . The threshold rate for this problem is the
conditional mutual information, I(X : Z|Y ). Furthermore, the Hoeffding [34] and Han-Kobyashi [21] bounds yield
an operational interpretation of the Rényi conditional mutual information given as
Is↑↑↑ (X : Z|Y ) := Ds (PXY Z kPY × PX|Y × PZ|Y )
X
X
1
=
PY (y)
PZ|Y =y (z)
log
s−1
y∈Y
z∈Z
(70)
X
x∈X
PX|Y =y,Z=z (x)s PX|Y =y (x)1−s
!!
.
(71)
In the special case where PXY Z = PY Z × PX|Z , this kind of Rényi mutual information describes the exponent
for leaked mutual information when we employ a superposition code in the wire-tap channel [63, Lem. 16] and [30,
Thm. 20].
14
2) Two marginals fixed, recovery channels: The expression in (67) corresponds to the following problem:
×n
null hypothesis: (X n , Y n , Z n ) ∼ PXY
Z,
Hcmi
↑↑↓ :
×n
.
alternative hypothesis: X n ↔ Y n ↔ Z n is Markov, (X n , Y n ) ∼ PXY
(72)
This problem is discussed in Section III-B3. We show that the threshold for this test is again given by I(X : Z|Y ).
Moreover, the optimal error and strong converse exponents for R close to the threshold are given by
s−1
1 − s ↑↑↓
Is (X : Z|Y ) − R and scR (Hcmi
R − Is↑↑↓ (X : Z|Y ) ,
(73)
eR (Hcmi
↑↑↓ ) = sup
↑↑↓ ) = sup
s
s
s>1
s∈(0,1)
where we have introduced a new definition of the Rényi conditional mutual information. This is given by
Is↑↑↓ (X : Z|Y ) :=
=
min
Ds (PXY Z kPY × PX|Y × QZ|Y )
!1 s
s
X
X
X
log
PY (y)
PZ|Y =y (z)
PX|Y =y,Z=z (x)s PX|Y =y (x)1−s .
QZ|Y ∈P(Z|Y)
1
s−1
y∈Y
z∈Z
(74)
(75)
x∈X
The minimum was evaluated using Sibson’s identity in Lemma 29 for the distribution QZ|Y =y seperately for each
y ∈ Y . (See Appendix A-D for details.) The resulting expression can be regarded as a conditional version of the
ρ
Gallager function by replacing s with ρ−1
. In the special case where PXY Z = PY Z × PX|Z , this quantity is used
in superposition coding to describe the error exponent [36, Sec. II] and the exponent of leaked information [30,
Thm. 22].
3) Arbitrary Markov distribution, permutation invariant: The most general alternative hypothesis that we consider
is comprised of all distributions that have a Markov structure X ↔ Y ↔ Z . More precisely, the following problem:
Hcmi
↓↓↓ :
×n
null hypothesis: (X n , Y n , Z n ) ∼ PXY
Z,
alternative hypothesis: X n ↔ Y n ↔ Z n is Markov, (X n , Y n , Z n ) permutation invariant.
(76)
This is covered in Section III-B3. The threshold is again I(X : Z|Y ) and the optimal error and strong converse
exponents for R close to the threshold are given by
s−1
1 − s ↓↓↓
Is (X : Z|Y ) − R and scR (Hcmi
R − Is↓↓↓ (X : Z|Y ) ,
(77)
sup
eR (Hcmi
↓↓↓ ) = sup
↓↓↓ ) =
s
s
s>1
s∈( 2 ,1)
3
with yet another Rényi conditional mutual information,
Is↓↓↓ (X : Z|Y ) :=
min
QXY Z ∈P(X ×Y×Z)
Ds (PXY Z kQY × QX|Y × QZ|Y ) .
(78)
Note that we only have an operational interpretation of this quantity for s > 23 , and, moreover, we do not know of
a closed form expression.
V. P ROOFS : H OEFFDING B OUND
The proof of Theorem 9 is split into two parts, achievability and optimality, which both rely on different Axioms.
A. Some Properties of φ(s)
Before we state our result, let us introduce some helpful notation. Recall that φ(s) := (s − 1)Ds (P kQ) =
log gs (P kQ) and define
φ̄s0 (s) := (s − 1)Ds (P kQ̂s0 ) = log gs (P kQ̂s0 ) ,
(79)
where Q̂s0 := Qθ̂s0 . Clearly, φ(s0 ) = φ̄s0 (s0 ) by definition of θ̂ s in Axiom 5. An important consequence of this
Axiom is the following lemma, which shows that the first derivative of φ and φ̄ agree as well at s = s0 .
Lemma 19. Assume Axiom 5 holds on (a, b). Then, for s0 ∈ (a, b), we have
d
Ds (P kQ)
ds
s=s0
=
d
Ds (P kQθ̂s0 )
ds
s=s0
.
(80)
15
Moreover, the function φ(s) on (a, b) is continuously differentiable and satisfies φ′ (s0 ) = φ̄′s0 (s0 ).
Proof: Let us define f (s, θ) = gs (P kQθ ). Write θ = (θ1 , . . . , θd ) as a d-dimensional real vector. The point θ̂ s
is determined by the implicit functions
∂
Fi (s, θ) :=
f (s, θ) = 0,
∀i ∈ {1, 2, . . . , d} .
(81)
∂θi
Moreover, the Hessian matrix Hs of θ 7→ f (s, θ) at θ̂ s is given by
(Hs )i,j =
∂2
f (s, θ)
∂θi ∂θj
θ=θ̂ s
(82)
.
The map s 7→ Hs is continuous since (s, θ) 7→ f (s, θ) is C 2 by Axiom 5.
Let us first treat the case s0 > 1. By Axiom 5 we have that Hs is positive definite and thus invertible. Then, the
implicit function theorem yields
∂ θ̂is
∂s
s=s0
=−
d
X
Hs−1
0
j=1
i,j
·
∂Fj (s, θ̂ s0 )
∂s
(83)
s=s0
and in particular s 7→ θ̂ s is continuously differentiable. We further find that
d
f (s, θ̂ s )
ds
s=s0
∂
f (s, θ̂ s0 )
=
∂s
s=s0
+
d
X
i=1
F i (s0 , θ̂ s0 ) ·
∂ θ̂is
∂s
=
s=s0
∂
f (s, θ̂ s0 )
∂s
s=s0
,
(84)
where we used that F i (s0 , θ̂ s0 ) = 0 by definition of θ̂ s0 . This establishes the result for s0 ∈ (1, b). An analogous
argument yields the same result for s0 < 1. For s0 = 1 we instead choose f (s, θ) = Ds (P kQθ ). Again the Hessian
matrix for the derivative with regards to θ is strictly positive definite, and the remainder of the argument proceeds
as before.
Finally, since s 7→ Ds (P kQ) is C 2 for fixed P, Q ∈ P(X) and θ 7→ Qθ as well as s 7→ θ̂ s are continuous, we
deduce that the functions s 7→ Ds (P kQ) and φ(s) = (s − 1)Ds (P kQ) are continuously differentiable.
B. Some Properties of Convex C 1 Functions
In this section let φ̃(s) be a general convex C 1 function on (a, b) ⊆ [0, ∞) and define ψ̃(s) := sφ̃′ (s) − φ̃(s).
Moreover, define R̃c := lims→c ψ̃(s) for every c ∈ [a, b].
Lemma 20. The function ψ̃(s) is continuous and monotonically nondecreasing on (a, b).
Proof: Let a < s0 < s1 < b. By the mean value theorem there exists an s ∈ [s0 , s1 ] such that (s1 − s0)φ̃′ (s) =
φ̃(s1 ) − φ̃(s0 ). Thus,
ψ̃(s1 ) − ψ̃(s0 ) = s1 φ̃′ (s1 ) − s0 φ̃′ (s0 ) − φ̃(s1 ) − φ̃(s0 )
(85)
= s1 φ̃′ (s1 ) − s0 φ̃′ (s0 ) − (s1 − s0 )φ̃′ (s)
= s1 φ̃′ (s1 ) − φ̃′ (s) + s0 φ̃′ (s) − φ̃′ (s0 ) ≥ 0 .
(86)
(87)
The inequality follows from the assumption that φ̃(s) is convex, and φ̃′ (s) thus monotonically increasing.
Lemma 21. Let R ∈ (R̃a , R̃b ). Then, there exists an ŝ ∈ (a, b) such that ψ̃(ŝ) = R and
(ŝ − 1)R − φ̃(ŝ)
(s − 1)R − φ̃(s)
=
s
ŝ
s∈(a,b)
sup
(88)
Proof: By continuity, for every R ∈ (R̃a , R̃b ), there exists (at least one) value ŝ ∈ (a, b) such that ψ̃(ŝ) = R.
φ̃(s)
. This yields
Let us first calculate the derivative of g(s) := (s−1)R−
s
sR − sφ̃′ (s) − (s − 1)R + φ̃(s)
R − ψ̃(s)
=
.
(89)
2
s
s2
Note that the numerator is monotonically decreasing in s due to Lemma 20 and vanishes at s = ŝ. In particular,
we find that g′ (s) ≥ 0 for a < s < ŝ and g′ (s) ≤ 0 for ŝ < s < b. We conclude that ŝ maximizes g(s) on (a, b).
g′ (s) =
16
C. Proof of Achievability
Achievability for Theorem 9 follows from the following statement.
Proposition 22. Assume Axioms 3 and 4 hold with parameter a. Then, we have
1
1−s
×n
Ds (P kQ) − R .
lim inf − log α̂ exp(−nR); P
Qn ≥ sup
n→∞
n
s
s∈(a,1)
(90)
Proof: Note that the expression on the right hand side of (90) is zero if R ≥ D(P kQ) and the inequality thus
holds trivially for that case. We assume that R < D(P kQ) for the remainder of this proof.
Let us fix any s ∈ (a, 1) for the moment. Moreover, let us define the sequence of tests
(
1 if P ×n (xn ) ≥ exp(λn )U n (xn )
Tn (xn ) :=
,
(91)
0 otherwise
where U n is the universal distribution of Axiom 4. We also choose the sequence {λn }n∈N of real numbers as
1
λn =
log v(n) + nR + (s − 1)Ds (P ×n kU n ) .
(92)
s
Axiom 4 implies that Qn is closed under symmetrization. Moreover, the test Tn is permutation invariant. Hence,
for all π ∈ Sn , we have Qn [Tn ] = Qn [Tn Wπn ] = Qn Wπn−1 [T n ] and we can in particular replace Qn with its
symmetrization. This yields
×n n
n
n
n
Q
P
(X
)
≥
exp(λ
)U
(X
)
(93)
β(Tn ; Qn ) = max
n
Qn ∈Qn
= maxsym Qn P ×n (X n ) ≥ exp(λn )U n (X n ) .
(94)
Qn ∈Qn
Next we use the universal distribution in Axiom 4 to further bound
X
β(Tn ; Qn ) ≤ v(n)
U n (xn ) 1 P ×n (xn ) ≥ exp(λn )U n (xn )
(95)
xn ∈X n
≤ v(n) exp(−sλn )
≤ v(n) exp(−sλn )
X
xn ∈X n
X
xn ∈X n
1−s ×n s ×n n
U n (x)
P (x) 1 P (x ) ≥ exp(λn )U n (xn )
1−s ×n s
U n (x)
P (x)
= v(n) exp (−sλn ) exp (s − 1)Ds (P ×n kU n ) .
(96)
(97)
(98)
Hence, the requirement that β(Tn ; Qn ) ≤ exp(−nR) is satisfied by the choice of λn in (92). Note that this statement
can directly be extended to the convex hull due to (17).
Let us now take a closer look at the error of the first kind. Using a similar development as above, we find
(99)
α̂ exp(−nR); P ×n Qn ≤ α(Tn ; P ×n ) = P ×n P ×n (X n ) < exp(λn )U n (X n )
×n
n
(100)
≤ exp (1 − s)λn exp (s − 1)Ds (P kU )
1 − s
= exp
(101)
log v(n) + nR − Ds (P ×n kU n )
s
where we substituted λn from (92) in the last step. Further using the additivity property of Axiom 3, we find that
Ds (P ×n kU n ) ≥ Ds (P ×n kQn ) ≥ nDs (P kQ),
(102)
and thus we arrive at the bound
1−s
nR − nDs (P kQ) + log v(n) .
log α̂ exp(−nR); P ×n Qn ≤
s
Since log v(n) = O(log n), taking the limit n → ∞ yields
1−s
1
lim inf − log α̂ exp(−nR); P ×n Qn ≥
Ds (P kQ) − R .
n→∞
n
s
Finally, since this derivation holds for all s ∈ (a, 1), we established the direct part.
(103)
(104)
17
D. Proof of Optimality
To show optimality, we will directly employ the converse of the Hoeffding bound.
Proposition 23. Assume Axioms 2–5 hold with parameter a. Then, for any R > Ra , we have
1−s
1
Ds (P kQ) − R .
lim sup − log α̂ exp(−nR); P ×n Qn ≤ sup
n
s
n→∞
s∈(a,1)
(105)
Proof: Let us first consider the case R ∈ (Ra , D(P kQ)). We will use the results of Sections V-A and V-B.
Take ŝ ∈ (a, 1) to be the optimizer in Lemma 21 for the functions φ̃ = φ and ψ̃ = ψ on (a, 1). Then we have that
ψ(ŝ) = ŝφ′ (ŝ) − φ(ŝ) = R and
1 − ŝ
1−s
(Ds (P kQ) − R) =
(Dŝ (P kQ) − R)
s
ŝ
s∈(a,1)
sup
(106)
The following consequence of Lemma 19 is crucial. Recall that
φ̄(s) = log gs (P kQθ̂ŝ )
and
ψ̄(s) = sφ̄′ (s) − φ̄(s) ,
(107)
where θ̂ ŝ is the optimal θ at s = ŝ. Note that the function s 7→ φ̄(s) is a C 1 convex function on all of (0, 1) since
gs is evaluated for fixed distributions P and Qθ̂ŝ . Moreover, Lemma 19 implies that φ′ (ŝ) = φ̄′ (ŝ), and hence,
ψ̄(ŝ) = ψ(ŝ) = R. We can thus also apply Lemma 21 for the functions φ̃ = φ̄ and ψ̃ = ψ̄ on (0, 1) and find that
1 − ŝ
1−s
sup
(Dŝ (P kQ) − R) .
(108)
Ds (P kQθ̂ŝ ) − R =
s
ŝ
s∈(0,1)
Next we note that, due to Axiom 2, we have
α̂ exp(−nR); P ×n Qn ≥ α̂ exp(−nR); P ×n Q×n
(109)
for any Q ∈ Q. After applying this for Q = Qθ̂ŝ , we can further apply the converse of the Hoeffding bound (we
take the formulation in [43, Thm. 1]) to the expression on the right-hand side, which yields
1
×n
Qn
(110)
lim sup − log α̂ exp(−nR); P
n
n→∞
1
≤ lim sup − log α̂ exp(−nR); P ×n Q×n
(111)
θ̂ ŝ
n
n→∞
1−s
= sup
(112)
Ds (P kQθ̂ŝ ) − R .
s
s∈(0,1)
which proves the result together with (106) and (108).
If R ≥ D(P kQ) the right hand side of (105) evaluates to zero. Moreover, there exists at least one Q ∈ Q such
that R ≥ D(P kQ). The result then follows from the converse of Hoeffding’s bound since
1
1
lim sup − log α̂ exp(−nR); P ×n Qn ≤ lim sup − log α̂ exp(−nR); P ×n Q×n = 0 .
(113)
n
n
n→∞
n→∞
VI. P ROOFS : S TRONG C ONVERSE E XPONENTS
Again we treat achievability and optimality with separate proofs that rely on different axioms.
18
A. Proof of Achievability
Our proof relies on a variant of the Gärtner-Ellis theorem of large deviation theory (see, e.g., [14, Sec. 2 and
Sec. 3.4] for an overview), which we recall here. Given a sequence of random variables {Zn }n∈N we introduce its
asymptotic cumulant generating function as
1
log E [exp(ntZn )]
,
(114)
ΛZ (t) := lim
n→∞ n
if it exists. For our purposes it is sufficient to use the following variant of the Gärtner-Ellis theorem due to Chen [7,
Thm. 3.6] (see also [40, Lem. A.2] for this exact statement).
Lemma 24. Let us assume that t 7→ ΛZ (t) as defined
in (114) exists and is differentiable in some interval (a, b).
Then, for any z ∈ limtցa Λ′Z (t), limtրb Λ′Z (t) , we have
1
(115)
lim sup − log Pr[Zn ≥ z] ≤ sup {tz − ΛZ (t)} .
n
n→∞
t∈(a,b)
Achievability follows from the following statement.
Proposition 25. Assume Axioms 3–5 hold with parameter b. For any R ∈ (0, Rb ), we have
s−1
1
×n
Qn
≤ sup
lim − log 1 − α̂ exp(−nR); P
R − Ds (P kQ) .
n→∞ n
s
s∈(1,b)
(116)
Proof: Let us first assume that R ∈ (D(P kQ), Rb ). Using Lemma 21 with φ̃ = φ on (1, b), we find the value
ŝ ∈ (1, b) that satisfies
ŝ − 1
s−1
sup
R − Ds (P kQ) =
R − Dŝ (P kQ)
(117)
s
ŝ
s∈(1,b)
and ψ(ŝ) = ŝφ′ (ŝ) − φ(ŝ) = R. We use the same sequence of tests Tn as in (91) and the sequence λn of (92),
substituting ŝ for s. This ensures that β(Tn ; Qn ) ≤ exp(−nR), as shown in the proof of Proposition 22. Moreover,
(118)
1 − α̂ exp(−nR); P ×n Qn ≥ 1 − α(Tn ; P ×n )
= P ×n P ×n (X n ) ≥ exp(λn )U n (X n ) = Pr[Zn ≥ 0] ,
(119)
where we defined the sequence of random variables Zn (X n ) following the law X n ← P ×n and
1
P ×n (xn )
Zn (xn ) =
log n n − λn
n
U (x )
1
nR
P ×n (xn ) log v(n) ŝ − 1
×n
n
=
−
Dŝ (P kU ) −
log n n −
.
n
U (x )
ŝ
ŝ
ŝ
Its asymptotic cumulant generating function then evaluates to
1
log E [exp(n(t − 1)Zn )]
ΛZ (t − 1) = lim
n→∞ n
×n n t−1
P (X )
log v(n) ŝ − 1
R
1
×n
n
log E
+
Dŝ (P kU ) +
− (t − 1)
= lim
n→∞ n
U n (X n )t−1
nŝ
nŝ
ŝ
1
ŝ − 1
R log v(n)
= (t − 1) lim
Dt (P ×n kU n ) −
Dŝ (P ×n kU n ) − −
n→∞ n
nŝ
ŝ
nŝ
t−1
(φ(ŝ) + R)
= φ(t) −
ŝ
= φ(t) − (t − 1)φ′ (ŝ) .
(120)
(121)
(122)
(123)
(124)
(125)
(126)
Here we used the expression for Zn given in (121) and the defintion of the Rényi divergence in the second and
third equality, respectively. To arrive at (125) we evaluated the limit n → ∞ using Lemma 7. The last equality
follows from the relation between ŝ and R given in the line following (117).
19
Note that t 7→ ΛZ (t − 1) is differentiable on (1, b) due to Lemma 19. Moreover, in order to apply Lemma 24
with z = 0, we need to verify the following two inequalities:
d
d
′
′
lim
ΛZ (t − 1) = φ (1) − φ (ŝ) < 0 and
lim
ΛZ (t − 1) = φ′ (b) − φ′ (ŝ) > 0 .
(127)
t→1 dt
t→b dt
We cannot invoke strict convexity of φ(s) to verify the above bounds; instead, note that D(P kQ) < R, and thus
1
ŝ − 1
Dŝ (P kQ) − R
ŝ
ŝ
ŝ − 1
D(P kQ) − Dŝ (P kQ) ≤ 0 .
<
ŝ
To prove the second inequality in (127), we use the fact that Rb > R to show
b−1
1
ŝ − 1
1
Db (P kQ) − R −
Dŝ (P kQ)
φ′ (b) − φ′ (ŝ) = Rb +
b
b
ŝ
ŝ
1
b−1
1
ŝ − 1
> R+
Db (P kQ) − R −
Dŝ (P kQ)
b
b
ŝ
ŝ
b−1
ŝ − 1
=
R − Dŝ (P kQ) −
R − Db (P kQ) ≥ 0 ,
ŝ
b
where the last inequality follows by the definition of ŝ.
We have now verified the conditions of Lemma 24 with z = 0, which yields
1
lim − log Pr[Zn ≥ 0] = sup (t − 1)φ′ (ŝ) − φ(t)
n→∞ n
t∈(1,b)
φ′ (1) − φ′ (ŝ) = D(P kQ) −
= (ŝ − 1)φ′ (ŝ) − φ(ŝ)
ŝ − 1
ŝ − 1
1
′
=
φ(ŝ) =
R − Dŝ (P kQ) .
ŝφ (ŝ) − φ(ŝ) +
ŝ
ŝ − 1
ŝ
(128)
(129)
(130)
(131)
(132)
(133)
(134)
(135)
To evaluate the supremum in (133), we note that the objective function t 7→ (t − 1)φ′ (ŝ) − φ(t) is concave in t and
its derivative vanishes at t = ŝ. This establishes (134). Combining this with (118) concludes the proof.
For R ≤ D(P kQ) the right hand side of (116) evaluates to zero. Since the expression on the left hand side is
clearly monotonically increasing in R we deduce that, for all such R,
1
s−1
×n
R − Ds (P kQ)
= 0.
(136)
lim − log 1 − α̂ exp(−nR); P
Qn
≤
inf
sup
n→∞ n
s
R>D(P kQ) s∈(1,b)
B. Proof of Optimality
Optimality follows as a corollary of Han and Kobayashi’s [21] derivation of the strong converse exponent.
Proposition 26. Assume Axiom 2 holds. For any R ≥ 0, we have
1
s−1
×n
lim inf − log 1 − α̂ exp(−nR); P
Qn
R − Ds (P kQ) .
≥ sup
n→∞
n
s
s>1
(137)
Proof: If R < D(P kQ) the bound holds trivially. Otherwise, analogous to the optimality proof for Theorem 9,
we first fix Q ∈ Q and apply the Han-Kobayashi converse bound [21] (in the form of [46, Ch. VI] and [39,
Thm. IV.9]). This yields
1
×n
Qn
lim inf − log 1 − α̂ exp(−nR); P
(138)
n→∞
n
1
≥ lim inf − log 1 − α̂ exp(−nR); P ×n Q×n
(139)
n→∞
n
s−1
= sup
R − Ds (P Q) .
(140)
s
s>1
As this holds for all Q ∈ Q, we maximize the expression in (140) over Q to arrive at the desired result.
20
VII. P ROOFS : S ECOND O RDER A SYMPTOTICS
OF
S TEIN ’ S L EMMA
In the proofs of this section we assume that log denotes the natural logarithm for ease of presentation. The
following result refines Lemma 7, and is the key ingredient of our proof.
Lemma 27. Assume Axioms 2–5 holds on (a, b) ⊃ {1}. For all t ∈ R,
t2
t
lim √ D1+ √t (P ×n kU n ) − nD(P kQ)
= V (P kQ) .
n
n→∞
n
2
(141)
Proof: By combining Lemma 19 at s0 = 1 with the Taylor expansion in (26), we find D1+s (P kQ) =
D(P kQ) + 2s V (P kQ) + O(s2 ). From (32) we learn that
√
t n
×n
n
V (P kQ) + O(log n) .
(142)
D1+ √t (P kU ) ≤ nD1+ √t (P kQ) + log v(n) = nD(P kQ) +
n
n
2
Furthermore, employing additivity from Axiom 3 yields
√
t n
×n
n
V (P kQ) + O(log n) .
(143)
D1+ √t (P kU ) ≥ nD1+ √t (P kQ) = nD(P kQ) +
n
n
2
Combining (142) and (143)
yields the desired statement.
Now let MX (t) := E exp(tX) denote the moment generating function of a real random variable X . We also
need the following property of moment generating functions, a variant of Lévi’s continuity theorem [41, Thm. 2].
Lemma 28. Let 0 < a < b. If a sequence of random variables {Xn }n∈N satisfies limn→∞ MXn (t) = MX (t) for
some random variable X and all t ∈ (a, b), then limn→∞ Pr[Xn ≤ k] = Pr[X ≤ k] for all k ∈ R.
We prove the direct and converse part of Theorem 12 together.
Proof of Theorem 12: We first show the converse statement. Choosing the optimal distribution Q̂ = Q̂1 ∈ Q
as defined in (28), we find
√
√
α̂ exp(−nD(P kQ) − nr); P ×n Qn ≥ α̂ exp(−nD(P kQ̂) − nr); P ×n Q̂×n ,
(144)
and the limiting statement then follows using [64, Thm. 2] (see also [54, Thm. 1.1]).
√
To show achievability we again rely on the test given in (91) and set λn = nD(P kQ) + nr + log v(n). Then
by Axiom 4 and using the argument leading to (94) to establish the first identity, we have
β(Tn ; Qn ) = maxsym Qn P ×n (X n ) ≥ exp(λn )U n (X n )
(145)
Qn ∈Qn
≤ v(n)U n P ×n (X n ) ≥ exp(λn )U n (X n )
(146)
×n n
×n
n
n
≤ v(n) exp(−λn )P
P (X ) ≥ exp(λn )U (X )
(147)
√
≤ exp − nD(P kQ) − nr .
(148)
Furthermore, we find
√
α(Tn ; P ×n ) = P ×n log P ×n (X n ) − log U n (X n ) < nD(P kQ) + nr + log v(n)
n
= Pr[Yn (X ) < r] ,
where X n ∼ P ×n and we defined the following sequence of random variables as
1
Yn (X n ) := √ log P ×n (X n ) − log U n (X n ) − nD(P kQ) − log v(n) .
n
Lemma 27 then implies that the cumulant generating function converges to
log MY (t) = lim log E exp(tY n )
n→∞
t
√
t
√
− √tn
n
n
n
n
+ ntD(P kQ) − √ log v(n)
= lim log E PX n (x ) U (x )
n→∞
n
t
= lim √ D1+ √t (P ×n kU n ) − nD(P kQ) − log v(n)
n
n→∞
n
2
t
= V (P kQ) .
2
(149)
(150)
(151)
(152)
(153)
(154)
(155)
21
Hence, by Lemma 28, the sequence of random variable {Yn }n converges in distribution to a random variable Y with
cumulant generating function log MY (t), i.e., a Gaussian random variable with zero mean and variance V (P kQ).
In particular, this yields
!
r
lim P ×n [Yn < r] = Pr [Y < r] = Φ p
.
(156)
n→∞
V (P kQ)
√
Since α̂ exp(−nD(P kQ) − nr); P ×n Qn ≤ α(Tn ; P ×n ), this concludes the proof.
VIII. C ONCLUSION
We have introduced a general framework to treat binary hypothesis testing with a composite alternative hypothesis.
In this general framework we show analogues of Stein’s Lemma, Hoeffding’s optimal error exponents and HanKobyashi’s optimal strong converse exponents. We have discussed several concrete examples that lead to operational
interpretations of various Rényi information measures.
The coincidence between our obtained exponents for the hypothesis testing problem in (51) and the corresponding
exponents for channel coding is quite interesting. A similar coincidence has been observed for the case of source
coding with side information and (63). These facts seem to indicate a deep relation between coding and the
composite alternative hypotheses given in (51) and (63). Its further clarification is an interesting future direction of
study, for example one could try to find coding problems that are closely related to (72).
In statistics, the χ2 test is used in an asymptotic setting similar to (9). The test assumes i.i.d. distributions and is
used for the case when both hypotheses are composite (see, e.g., [38]). For small samples and k = 2, Fisher’s exact
test [19] can be used to replace the χ2 test. Recently, the setting of small samples and general k has been studied
using a Gröbner basis approach [15], [50]. In contrast to their formulation, we have not assumed i.i.d. structure for
the independent case; instead, we only require permutation invariance for each random variable. Our result suggests
that we can replace the i.i.d. condition by a permutation invariant condition when testing independence, which can
be expected to have wider applications.
The key ingredient of our derivation is an axiomatic approach based on the universal distribution. Due to its
generality, we can treat many composite hypothesis testing problems without i.i.d. assumption (for the composite
hypothesis), and it will be interesting to explore further examples that fit into our framework. Moreover, because
the universal distribution plays an important role in universal channel coding [24], we can expect that it will play
an important role when analyzing universal protocols for other problems in information theory.
As explained in Section III-B2, we cannot remove the permutation invariance condition in that example, an
essential difference to the example given in Section III-B1. This kind of difference sheds light on the difference
between channel coding and secure random number generation. Originally, for the channel coding, the meta converse
was introduced using simple hypothesis testing [42, Sec. 3] and [48]. Polyanskiy [47, Sec. II] then extended it to
the composite hypothesis testing of the form (7). Although this improvement does not effect the exponents and the
second-order coding rate, it can improve the bound in the finite blocklength regime. Recently, Tyagi-Watanabe [58],
[59] introduced a converse bound for secure random number generation by using simple hypothesis testing between
a true joint distribution and an arbitrary product distribution. Although in their converse bound, we can choose
an arbitrary product distribution as the alternative hypothesis, we cannot replace the alternative hypothesis by a
composite hypothesis composed of all of product distributions. Hence, for secure random number generation, we
cannot extend their bound to a bound based on the composite hypothesis as in [47].
In prior work [32] the present authors have analyzed composite hypothesis testing in the non-commutative
(quantum) regime and found an operational interpretation for various definitions of quantum Rényi mutual information and quantum Rényi conditional entropy [56]. However, the present work is more general than the classical
specialization of that work and requires new techniques. This allows us to deal with more complex composite
alternative hypotheses, and in particular allows for a characterization of Rényi conditional mutual information.
Furthermore, finding appropriate definitions for Rényi conditional mutual information in the non-commutative
setting is an ongoing topic of research [4]. It is possible that an adaption of our analysis to quantum hypothesis
testing will lead to further progress in this direction. However, some caution is advised since already the definitions
of the regular conditional mutual information in (66)–(68) are not equivalent in the quantum case.
22
Acknowledgements: We thank Vincent Y. F. Tan for helpful comments and Christoph Pfister for alterting us to
several typos and small gaps in our presentation. MH is partially supported by a MEXT Grant-in-Aid for Scientific
Research (A) No. 23246071, and by the National Institute of Information and Communication Technology (NICT),
Japan. MT is funded by an University of Sydney Postdoctoral Fellowship and acknowledges support from the ARC
Centre of Excellence for Engineered Quantum Systems (EQUS).
R EFERENCES
[1] E. Arikan. “An Inequality on Guessing and its Application to Sequential Decoding”. IEEE Trans. Inf. Theory 42(1): 99–105 (1996).
[2] S. Arimoto. “On the Converse to the Coding Theorem for Discrete Memoryless Channels”. IEEE Trans. Inf. Theory 19(3): 357–359
(1973).
[3] S. Arimoto. “Information Measures and Capacity of Order Alpha for Discrete Memoryless Channels”. Colloq. Math. Soc. János Bolya
16: 41–52, (1975).
[4] M. Berta, K. Seshadreesan, and M. Wilde. “Rényi Generalizations of the Conditional Quantum Mutual Information”. J. Math. Phys.
56(2): 022205 (2015).
[5] J. A. Bucklew. Large Deviation Techniques in Decision, Simulation, and Estimation. Wiley (1990).
[6] C. Bunte and A. Lapidoth. “Encoding Tasks and Rényi Entropy”. IEEE Trans. Inf. Theory 60(9): 5065–5076 (2014).
[7] P.-N. Chen. “Generalization of Gärtner-Ellis Theorem”. IEEE Trans. Inf. Theory 46(7): 2752–2760, (2000).
[8] H. Chernoff. “Large-Sample Theory: Parametric Case”. Ann. Math. Stat. 27(1): 1–22 (1956).
[9] B. Clarke and A. Barron. “Information-Theoretic Asymptotics of Bayes Methods”. IEEE Trans. Inf. Theory 36(3): 453–471 (1990).
[10] I. Csiszár. “Generalized Cutoff Rates and Renyi’s Information Measures”. IEEE Trans. Inf. Theory 41(1): 26–34 (1995).
[11] I. Csiszár. “The Method of Types”. IEEE Trans. Inf. Theory 44(6): 2505–2523 (1998).
[12] I. Csiszár and G. Longo. “On the Error Exponent for Source Coding and for Testing Simple Statistical Hypotheses”. Stud. Sci. Math.
Hungarica 6: 181–191, (1971).
[13] L. Davisson. “Universal noiseless coding”. IEEE Trans. Inf. Theory 19(6): 783–795 (1973).
[14] A. Dembo and O. Zeitouni. Large Deviations Techniques and Applications. Springer (1998).
[15] P. Diaconis and B. Sturmfels. “Algebraic Algorithms for Sampling from Conditional Distributions”. Ann. Stat. 26(1): 363–397 (1998).
[16] O. Fawzi and R. Renner. “Quantum Conditional Mutual Information and Approximate Markov Chains”. Commun. Math. Phys.
340(2): 575–611 (2015).
[17] M. Feder and A. Lapidoth. “Universal decoding for channels with memory”. IEEE Trans. Inf. Theory 44(5): 1726–1745 (1998).
[18] S. Fehr and S. Berens. “On the Conditional Rényi Entropy”. IEEE Trans. Inf. Theory 60(11): 6801–6810 (2014).
[19] R. A. Fisher. “On the Interpretation of chi-square from Contingency Tables, and the Calculation of P”. J. R. Stat. Soc. 85(1): 87
(1922).
[20] R. G. Gallager. Information Theory and Reliable Communication. Wiley (1968).
[21] T. Han and K. Kobayashi. “The Strong Converse Theorem for Hypothesis Testing”. IEEE Trans. Inf. Theory 35(1): 178–180 (1989).
[22] T. S. Han. Information-Spectrum Methods in Information Theory. Springer (2002).
[23] T. S. Han, H. Endo, and M. Sasaki. “Reliability and Secrecy Functions of the Wiretap Channel Under Cost Constraint”. IEEE Trans.
Inf. Theory 60(11): 6819–6843 (2014).
[24] M. Hayashi. “Information Spectrum Approach to Second-Order Coding Rate in Channel Coding”. IEEE Trans. Inf. Theory
55(11): 4947–4966 (2009).
[25] M. Hayashi. “Universal Approximation of Multi-copy States and Universal Quantum Lossless Data Compression”. Commun. Math.
Phys. 293(1): 171–183 (2009).
[26] M. Hayashi. “Universal Coding for Classical-Quantum Channel”. Commun. Math. Phys. 289(3): 1087–1098 (2009).
[27] M. Hayashi. “Exponential Decreasing Rate of Leaked Information in Universal Random Privacy Amplification”. IEEE Trans. Inf.
Theory 57(6): 3989–4001 (2011).
[28] M. Hayashi. “Tight Exponential Analysis of Universally Composable Privacy Amplification and Its Applications”. IEEE Trans. Inf.
Theory 59(11): 7728–7746 (2013).
[29] M. Hayashi. “Universal Channel Coding for Exponential Family of Channels”. Preprint, arXiv: 1502.02218 (2015).
[30] M. Hayashi and R. Matsumoto. “Secure Multiplex Coding with Dependent and Non-Uniform Multiple Messages”. Preprint,
arXiv: 1202.1332 (2012).
[31] M. Hayashi and V. Y. F. Tan. “Equivocations, Exponents and Second-Order Coding Rates under Various Rényi Information Measures”.
Preprint, arXiv: 1504.02536 (2015).
[32] M. Hayashi and M. Tomamichel. “Correlation Detection and an Operational Interpretation of the Renyi Mutual Information”. Preprint,
arXiv: 1408.6894 (2014).
[33] M. Hayashi and S. Watanabe. “Uniform Random Number Generation from Markov Chains: Non-Asymptotic and Asymptotic Analyses”.
Preprint, arXiv: 1503.04371 (2015).
[34] W. Hoeffding. “Asymptotically Optimal Tests for Multinomial Distributions”. Ann. Math. Stat. 36(2): 369–401, (1965).
[35] M. Iwamoto and J. Shikata. “Information Theoretic Security for Encryption Based on Conditional Rényi Entropies”. In C. Padró,
editor, Proc. ICITS 2013, volume 8317 of LNCS, pages 103–121, (2014).
[36] Y. Kaspi and N. Merhav. “Error Exponents for Broadcast Channels With Degraded Message Sets”. IEEE Trans. Inf. Theory
57(1): 101–123 (2011).
[37] A. Lapidoth and C. Pfister. “Two measures of dependence”. In 2016 IEEE Int. Conf. Sci. Electr. Eng., pages 1–5, (2016).
[38] E. Lehmann and J. P. Romano. Testing Statistical Hypotheses. Springer (2005).
[39] M. Mosonyi and T. Ogawa. “Quantum Hypothesis Testing and the Operational Interpretation of the Quantum Rényi Relative Entropies”.
Commun. Math. Phys. 334(3): 1617–1648 (2015).
23
[40] M. Mosonyi and T. Ogawa. “Two Approaches to Obtain the Strong Converse Exponent of Quantum Hypothesis Testing for General
Sequences of Quantum States”. IEEE Trans. Inf. Theory (2015).
[41] A. Mukherjea, M. Rao, and S. Suen. “A Note on Moment Generating Functions”. Stat. Probab. Lett. 76(11): 1185–1189 (2006).
[42] H. Nagaoka. “Strong Converse Theorems in Quantum Information Theory”. In Proc. ERATO Work. Quantum Inf. Sci. 2001, page 33,
(2001).
[43] H. Nagaoka. “The Converse Part of The Theorem for Quantum Hoeffding Bound”. Preprint, arXiv: quant-ph/0611289 (2006).
[44] H. Nagaoka and M. Hayashi. “An Information-Spectrum Approach to Classical and Quantum Hypothesis Testing for Simple
Hypotheses”. IEEE Trans. Inf. Theory 53(2): 534–549 (2007).
[45] K. Nakagawa and F. Kanaya. “On the Converse Theorem in Statistical Hypothesis Testing”. IEEE Trans. Inf. Theory 39(2): 623–628
(1993).
[46] T. Ogawa and H. Nagaoka. “Strong Converse and Stein’s Lemma in Quantum Hypothesis Testing”. IEEE Trans. Inf. Theory
46(7): 2428–2433 (2000).
[47] Y. Polyanskiy. “Saddle Point in the Minimax Converse for Channel Coding”. IEEE Trans. Inf. Theory 59(5): 2576–2595 (2013).
[48] Y. Polyanskiy, H. V. Poor, and S. Verdú. “Channel Coding Rate in the Finite Blocklength Regime”. IEEE Trans. Inf. Theory
56(5): 2307–2359 (2010).
[49] A. Rényi. “On Measures of Information and Entropy”. In Proc. 4th Berkeley Symp. Math. Stat. Probab., volume 1, pages 547–561,
Berkeley, California, USA(1961).
[50] T. Sakata and R. Sawae. “Algebraic Theory of Conditional and Sequential Conditional Tests of Three Way Contingency Tables”. Bull.
informatics Cybern. 37: 41–48, (2005).
[51] I. N. Sanov. “On the Probability of Large Deviations of Random Variables”. In Sel. Transl. Math. Stat. Probab., volume 1, pages
213–224. AMS (1961).
[52] O. Shayevitz. “On Rényi Measures and Hypothesis Testing”. In Proc. IEEE ISIT 2011, pages 894–898, (2011).
[53] R. Sibson. “Information radius”. Zeitschrift für Wahrscheinlichkeitstheorie und Verwandte Gebiete 14(2): 149–160 (1969).
[54] V. Strassen. “Asymptotische Abschätzungen in Shannons Informationstheorie”. In Trans. Third Prague Conf. Inf. Theory, pages
689–723, Prague(1962).
[55] A. Teixeira, A. Matos, and L. Antunes. “Conditional Rényi Entropies”. IEEE Trans. Inf. Theory 58(7): 4273–4277 (2012).
[56] M. Tomamichel, M. Berta, and M. Hayashi. “Relating Different Quantum Generalizations of the Conditional Rényi Entropy”. J. Math.
Phys. 55(8): 082206 (2014).
[57] M. Tomamichel and V. Y. F. Tan. “A Tight Upper Bound for the Third-Order Asymptotics for Most Discrete Memoryless Channels”.
IEEE Trans. Inf. Theory 59(11): 7041–7051 (2013).
[58] H. Tyagi and S. Watanabe. “A Bound for Multiparty Secret Key Agreement and Implications for a Problem of Secure Computing”. In
Proc. EUROCRYPT 2014, LNCS, pages 369–386, (2014).
[59] H. Tyagi and S. Watanabe. “Converses for Secret Key Agreement and Secure Computing”. Preprint, arXiv: 1404.5715 (2014).
[60] J. Unnikrishnan, D. Huang, S. P. Meyn, A. Surana, and V. V. Veeravalli. “Universal and Composite Hypothesis Testing via Mismatched
Divergence”. IEEE Trans. Inf. Theory 57(3): 1587–1603 (2011).
[61] T. van Erven and P. Harremoes. “Rény Divergence and Kullback-Leibler Divergence”. IEEE Trans. Inf. Theory 60(7): 3797–3820
(2014).
[62] S. Verdú. “Alpha-Mutual Information”. In Proc. IEEE ITA 2015, (2015).
[63] S. Watanabe and Y. Oohama. “The Optimal Use of Rate-Limited Randomness in Broadcast Channels With Confidential Messages”.
IEEE Trans. Inf. Theory 61(2): 983–995 (2015).
[64] A. A. Yushkevich. “On limit theorems connected with the concept of entropy of Markov chains”. Uspechi Mat. Nauk VIII(5): 177,
(1953).
Marco Tomamichel (M’13–SM’16) is a Senior Lecturer at the Centre for Quantum Software and Information and School of Software,
Faculty of Engineering and Information Technology at the University of Technology Sydney. He received a M.Sc. in Electrical Engineering
and Information Technology degree from ETH Zurich (Switzerland) in 2007. He then graduated with a Ph.D. in Physics at the Institute of
Theoretical Physics at ETH Zurich in 2012. Before commencing his current position, he has been a Senior Research Fellow at the Centre
for Quantum Technologies at the National University of Singapore and a Lecturer at the School of Physics at the University of Sydney.
His research interests include classical and quantum information theory with finite resources as well as applications to cryptography.
24
Masahito Hayashi (M’06–SM’13–F’17) was born in Japan in 1971. He received the B.S. degree from the Faculty of Sciences in Kyoto
University, Japan, in 1994 and the M.S. and Ph.D. degrees in Mathematics from Kyoto University, Japan, in 1996 and 1999, respectively.
He worked in Kyoto University as a Research Fellow of the Japan Society of the Promotion of Science (JSPS) from 1998 to 2000, and
worked in the Laboratory for Mathematical Neuroscience, Brain Science Institute, RIKEN from 2000 to 2003, and worked in ERATO
Quantum Computation and Information Project, Japan Science and Technology Agency (JST) as the Research Head from 2000 to 2006. He
also worked in the Superrobust Computation Project Information Science and Technology Strategic Core (21st Century COE by MEXT)
Graduate School of Information Science and Technology, The University of Tokyo as Adjunct Associate Professor from 2004 to 2007. He
worked in the Graduate School of Information Sciences, Tohoku University as Associate Professor from 2007 to 2012. In 2012, he joined the
Graduate School of Mathematics, Nagoya University as Professor. He also worked in Centre for Quantum Technologies, National University
of Singapore as Visiting Research Associate Professor from 2009 to 2012 and as Visiting Research Professor from 2012 to now. In 2011,
he received Information Theory Society Paper Award (2011) for “Information-Spectrum Approach to Second-Order Coding Rate in Channel
Coding”. In 2016, he received the Japan Academy Medal from the Japan Academy and the JSPS Prize from Japan Society for the Promotion
of Science.
In 2006, he published the book “Quantum Information: An Introduction” from Springer, whose revised version was published as “Quantum
Information Theory: Mathematical Foundation” from Graduate Texts in Physics, Springer in 2016. In 2016, he published other two books
“Group Representation for Quantum Theory” and “A Group Theoretic Approach to Quantum Information” from Springer. He is on the
Editorial Board of International Journal of Quantum Information and International Journal On Advances in Security. His research interests
include classical and quantum information theory and classical and quantum statistical inference.
V ERIFICATION
OF THE
A PPENDIX A
A XIOMS FOR E XAMPLES
IN
S ECTION III-B
A. Sibson’s identity
This appendix proves that the examples satisfy our axioms. One of the main ingredients is Sibson’s identity [53],
as presented in [10, Eq. (11)–(13)].
Lemma 29. For any distributions PXY ∈ P(X × Y), TX ∈ P(X ) and QY ∈ P(Y), and any s ∈ (0, 1) ∪ (1, ∞),
Ds (PXY kTX × QY ) = Ds (PXY kTX × Q̂sY ) + Ds (Q̂sY kQY )
where the optimal distribution Q̂sY ∈ P(Y) is given by
Q̂sY
(157)
1
PY (y)gs (PX|Y =y kTX ) s
(y) = P
1 .
s
y PY (y)gs (PX|Y =y kTX )
(158)
Thus, in particular, arg minQY ∈P(Y) Dα (PXY kTX × QY ) = {Q∗Y }.
Proof: We rewrite (157) as gs (PXY kTX ×QY ) = gs (PXY kTX × Q̂sY )·gs (Q̂sY kQY ), at which point the equality
can be verified by close inspection. The fact that Q̂sY is the unique minimizer is then a consequence of the positive
definiteness of the Rényi divergence.
B. Proof of Proposition 13
Proof: Clearly Q is compact convex and we explicitly find the optimizer using Sibson’s identity in Lemma 29.
Up to normalization it is given by
1
Q̂sY (y) ∼ PY (y)gs (PX|Y =y kTX ) s ,
(159)
and thus Axiom 1 is verified. Axiom 2 holds by definition and Axiom 3 can be verified by noting that Q̂s in (159)
×n
and TX×n are i.i.d. products.
takes on an i.i.d. product form when both PXY
Next, note that Qn is closed under permutations and convex. The universal distributions are
X
1
1
×n
n
n
n
n
(y
)
=
,
with
U
1{y n is of type λ} ,
(160)
×
U
UX
n
n
nY n = T
Y
Y
X
|Tn (Y)| |λ|
λ∈Tn (Y)
Qsym
n
∈
and thus. we find that Axiom 4 is satisfied with v(n) = |Tn (Y )| = poly(n).
as in Lemma 14. Clearly
Finally note that all the above remains true if we restrict Qn to permutation invariant or i.i.d. product distributions,
n
denoted Q′n , except that now UX
/ Q′n . However, we still have
nY n ∈
n
UX
nY n
′
×n
×n
n
×n
kUX
Ds (PXY
n Y n ) ≥ Ds (P
XY kQn ) = Ds (PXY kQn ),
×n
kQn ) is taken by a product distribution.
since additivity property guarantees that the minimum in Ds (PXY
(161)
25
C. Proof of Proposition 15
We give the proof for the case k = 1 and set X1 = X . The generalization to larger k does not require further
conceptual insights, and we will remark in a footnote where nontrivial changes are necessary.
Proof: Axiom 2 holds by definition. To verify Axiom 4 we first note that the joint permutations of X n and Y n
separate as WXn n Y n [π] = WXn n [π] × WYnn [π]. Thus, we can write
X 1
X 1
X 1
QX n WXn n [π] × QY n WYnn [π] = QX n ×
QY n WYnn [π] , (162)
QX n × QY n WXn n Y n [π] =
n!
n!
n!
π∈Sn
π∈Sn
π∈Sn
sym
where we used that QX n ∈ P
to establish the last equality. Clearly the resulting distribution lies in Qn .
n
n
n
n
n
Next, consider the universal distribution UX
n Y n = UX n × UY n with UX n and UY n given as in (40). Clearly, since
sym
Qn = P sym (X n ) × P sym (Y n ) we then find that every symmetric distribution satisfies
(X n )
n
n
QX n × QY n ≤ |Tn (X )|UX
n × |Tn (Y)|UY n
(163)
and Axiom 4 holds with v(n) = |Tn (X )||Tn (Y)| = poly(n).
For Axiom 5 we chose the following parametrization. Since P(X ) is convex subset of R|X |−1, there exists a
natural smooth parametrization Θ1 ∋ θ1 7→ QX,θ1 ∈ P(X ) where Θ1 is a convex subset of R|X |−1, and similarly
for P(Y). Combining these two parameterizations, we introduce a Θ ⊂ R|X |+|Y|−2 such that
(164)
Θ ∋ θ = (θ1 , θ2 ) 7→ QX,θ1 × QY,θ2 = QXY,θ ∈ Q .
The set Θ is evidently convex. Let us next verify the required convexity and concavity properties. First note that
the map f (x, y) = x1−s y 1−s for x, y ≥ 0 is strictly jointly concave when s ∈ ( 12 , 1) and strictly jointly convex
when s > 1. This follows, for example, from studying the Hessian matrix for f , which is
!
∂2f
∂2f
−sx−1 y (1 − s)
−s −s
∂x2
∂x∂y
(165)
= x y (1 − s)
∂2f
∂2f
(1 − s) −sxy −1
∂y∂x
∂y 2
and its determinant x−2s y −2s (1 − s)2 (2s − 1). The Hessian is negative definite for s ∈ ( 12 , 1) and positive definite
for s > 1 when x, y > 0.6 From this and the above parametrization we conclude that the Hessian of the map
Θ 7→ gs (PXY kQX × QY ) is negative definite for s ∈ ( 12 , 1) and positive definite for s > 1 in the relative interior of
Θ. Moreover, the function g(x, y) = − log xy = − log x−log y evidently has a positive definite Hessian for x, y > 0.
Hence the desired concavity and convexity properties for the maps θ 7→ gs (PXY kQXY,θ ) and θ 7→ Ds (PXY kQXY,θ )
hold. As a consequence, the minimizer is unique if it exists in the relative interior of Q (as we will show below).
Let us assume that Ds (PXY kQ) = Ds (PXY kQ̂sX × Q̂sY ) for some optimal distributions Q̂sX and Q̂sY . To verify
Axiom 3 we note that distributions are optimal only if they satisfy the self-consistency relation for a local optimum
given in Eq. (159). This implies that
1
Q̂sX (x) ∼ PX (x)gs (PY |X=x kQ̂sY ) s
1
and Q̂sY (y) ∼ PY (y)gs (PX|Y =y kQ̂sX ) s .
(166)
This solution is in the relative interior of Q and thus unique (for α > 21 ).
Furthermore, we find that the product distributions Q̂sX × Q̂sX and Q̂sY × Q̂sY satisfy the self-consistency relations
×2
kQ2 ). More precisely, we find
for the local optimal solution of Ds (PXY
1
Q̂sX (x1 )Q̂sX (x2 ) ∼ PX (x1 )PX (x2 )gs (PY |X=x1 × PY |X=x2 kQ̂sY × Q̂sY ) s ,
(167)
and vice versa. Moreover, due to the fact that the function is either convex or concave for α > 21 , we can conclude
that these product distributions are globally optimal as well. Namely
×2
×2
kQ2 ) = Ds PXY
Ds (PXY
(Q̂X × Q̂Y )×2 = 2Ds (PXY kQ) .
(168)
Applying this argument inductively yields the condition of Axiom 3.
6
For k ≥ 2 we need to consider the function f (x1 , x2 , . . . , xk , y) =
k
.
concave in all its arguments is further restricted to s > k+1
Qk
i=1
x1−s
y 1−s and the range of s where this function is jointly
i
26
D. Proof of Proposition 16
Before we commence with the proof we need to introduce some additional concepts and auxiliary results. Let
us introduce the following representation of channels, which is reminiscent of the Choi-Jamiołkowski isomorphism
in quantum information theory. Let X and Y be finite sets. For any channel QY |X ∈ P(Y|X ) we define a vector
representation Q̃XY = 1X × QY |X , where 1X is the identity vector for the Schur (element-wise) product of vectors,
i.e. 1X (x) = 1 for all x ∈ X . More concretely, the vector is given by Q̃XY (x, y) = QY |X (y|x). Note that Q̃XY is
not a probability distribution but clearly we must have
X
Q̃XY (x, y) = 1,
∀x ∈ X .
(169)
y∈Y
Using this representation we can write joint distribution after the application of the channel as
(170)
PX × QY |X = (PX × 1Y ) ◦ Q̃XY ,
where ◦ denotes the Schur product between the vectors and 1Y (y) = 1 for all y ∈ Y . Note also that the normalization
condition (169) enforces that the resulting vector is a probability distribution, and hence every vector with positive
elements that satisfies (169) corresponds to a valid channel.
Lemma 30. QY n |X n is covariant under permutations if and only if Q̃X n Y n is permutation invariant. Formally,
∀π ∈ Sn : QY n |X n WY n [π] = WX n [π]QY n |X n
⇐⇒
∀π ∈ Sn : Q̃X n Y n WX n Y n [π] = Q̃X n Y n .
Proof: The following equalities can be verified by close inspection:
Q̃X n Y n WX n Y n [π] = 1X n × QY n |X n WX n [π] × WY n [π]
= (1X n WX n [π]) × WX n [π ]QY n |X n WY n [π]
= 1X n × WX n [π −1 ]QY n |X n WY n [π] .
−1
(171)
(172)
(173)
(174)
The equivalence of the two conditions then follows from the fact that WX n [π −1 ]WX n [π] is the identity channel.
Proof of Lemma 17: Let QY n |X n be covariant under permutations. Then from Lemma 30 we learn that Q̃X n Y n
is permutation invariant and thus in particular can be written in the form
X
q̃XY (λXY )
1{(xn , y n ) is of type λXY } ,
(175)
Q̃X n Y n (xn , y n ) =
#λXY
λXY ∈Tn (X ×Y)
where q̃XY is a probability distribution over joint types λXY and #λXY denotes the number of sequences of type
λXY . Moreover, Eq. (169) enforces that for every type µX ∈ Tn (X ), and any sequence xn of type µX , we have
X
X
q̃XY (λXY ) X
1{(xn , y n ) is of type λXY } ,
(176)
Q̃X n Y n (xn , y n ) =
1=
#λ
XY
n
n
n
n
y ∈Y
y ∈Y
λXY ∈Tn (X ×Y)
|
{z
}
= #λY |X (xn )
where the number of sequences of type λXY with marginal xn , denoted #λY |X (xn ), clearly only depend on the
type µX of the marginal. Moreover, if the type of xn does not correspond to the marginal type λX of λXY then
XY
#λY |X (xn ) vanishes. Generally, we have #λY |X (xn ) = 1{µx = λx } #λ
#λX . Hence (176) simplifies to
X
q̃XY (λXY )
λXY ∈Tn (X ×Y)
λX =µX
1
=1
#λX
∀µX ∈ Tn (X ) .
(177)
A direct consequence of this condition is that q̃XY (λXY ) ≤ #λX for all λXY .
Now let us define a universal permutation covariant channel ŨX n Y n by the choice
ũXY (λXY ) :=
{κXY
#λX
∈ Tn (X × Y) : κX = λX }
(178)
27
which evidently satisfies (177). Moreover, for all permutation covariant channels with representation Q̃XY of the
form (175), we have the bound
q̃XY (λXY ) ≤ #λX ≤ {κXY ∈ Tn (X × Y) : κX = λX } ũXY (λXY ) ≤ |Tn (X × Y)|ũXY (λXY ) .
(179)
Hence also Q̃XY ≤ |Tn (X × Y)| ŨXY . The statement of the lemma then follows from the expression in (170).
Proof of Proposition 16: The set Q is clearly compact convex. Note that
X
gs (PXY Z kPY × PX|Y × QZ|Y ) =
PY (y) gs (PXZ|Y =y kPX|Y =y × QZ|Y =y ) .
(180)
y∈Y
From this we can then deduce, as in (159), that the optimal channel takes on the form
1
Q̂sZ|Y =y (z) ∼ PZ|Y =y (z)gs PX|Y =y,Z=z TX|Y =y s ,
(181)
for all y ∈ Y . As such, it is clear that Axioms 1–3 are satisfied.
It remains to verify Axiom 4. First note that Qn is closed under symmetrization. Moreover, any channel QZ n |Y n
sym
corresponding to a permutation invariant element of Qn satisfies
QZ n |Y n WZ n [π] = WY n [π]QZ n |Y n ,
∀π ∈ Sn ,
(182)
i.e. QZ n |Y n is permutation covariant. Hence, Lemma 17 applies and guarantees the existence of a sequence of
such that
universal channels {UZnn |Y n }n∈N with UZnn |Y n ∈ Qsym
n
×n
×n
× UZnn |Y n (xn , y n , z n )
× QZ n |Y n (xn , y n , z n ) ≤ v(n)PXY
PXY
∀xn ∈ X n , y n ∈ Y n , z n ∈ Z n .
(183)
E. Proof of Proposition 18
Proof: The proof proceeds similarly to the proofs of Propositions 15 and 16. Axiom 2 holds by definition and
Axiom 4 can be verified using the universal distributions
n
n
n
n
UX
n Y n Z n = UY n × U
X n |Y n × UZ n |Y n
(184)
n
n
with the universal distributions UYn n as in (40) and the universal maps UX
n |Y n and UZ n |Y n provided by Lemma 17.
Axiom 5 is verified with a construction analogous to Proposition 15 but this time we need to consider the
function (x, y, z) 7→ x1−s y 1−s z 1−s which is strictly jointly concave for s ∈ ( 23 , 1) and strictly jointly convex for
s > 1.Finally, Axiom 3 is again verified using the self-consistency relations.
| 7cs.IT
|
ZU064-05-FPR
main
28 March 2017
0:56
arXiv:1703.08683v1 [cs.PL] 25 Mar 2017
Under consideration for publication in J. Functional Programming
1
Applied Type System: An Approach to
Practical Programming with Theorem-Proving
Hongwei Xi∗
Boston University, Boston, MA 02215, USA
(e-mail: [email protected])
Abstract
The framework Pure Type System (PTS) offers a simple and general approach to designing and
formalizing type systems. However, in the presence of dependent types, there often exist certain
acute problems that make it difficult for PTS to directly accommodate many common realistic
programming features such as general recursion, recursive types, effects (e.g., exceptions, references, input/output), etc. In this paper, Applied Type System (ATS) is presented as a framework for
designing and formalizing type systems in support of practical programming with advanced types
(including dependent types). In particular, it is demonstrated that ATS can readily accommodate a
paradigm referred to as programming with theorem-proving (PwTP) in which programs and proofs
are constructed in a syntactically intertwined manner, yielding a practical approach to internalizing
constraint-solving needed during type-checking. The key salient feature of ATS lies in a complete
separation between statics, where types are formed and reasoned about, and dynamics, where programs are constructed and evaluated. With this separation, it is no longer possible for a program to
occur in a type as is otherwise allowed in PTS. The paper contains not only a formal development of
ATS but also some examples taken from ATS, a programming language with a type system rooted in
ATS, in support of employing ATS as a framework to formulate advanced type systems for practical
programming.
Contents
1
7
8
19
28
29
1 Introduction
2 Untyped λ -Calculus λdyn
3 Formal Development of ATS0
4 Formal Development of ATSpf
5 Related Work and Conclusion
References
1 Introduction
A primary motivation for developing Applied Type System (ATS) stems from an earlier attempt to support a restricted form of dependent types in practical programming (Xi, 2007).
∗ Supported in part by NSF grants no. CCR-0224244, no. CCR-0229480, and no. CCF-0702665
ZU064-05-FPR
main
2
28 March 2017
0:56
Hongwei Xi
While there is already a framework Pure Type System (PTS) (Barendregt, 1992) that
offers a simple and general approach to designing and formalizing type systems, it is well
understood that there often exist some acute problems (in the presence of dependent types)
making it difficult for PTS to accommodate many common realistic programming features.
In particular, various studies reported in the literature indicate that great efforts are often
required in order to maintain a style of pure reasoning as is advocated in PTS when features
such as general recursion (Constable & Smith, 1987), recursive types (Mendler, 1987), effects (Honsell et al., 1995), exceptions (Hayashi & Nakano, 1988) and input/output are present.
The framework ATS is formulated to allow for designing and formalizing type systems that can readily support common realistic programming features. The formulation
of ATS given in this paper is primarily based on the work reported in two previous papers (Xi, 2004; Chen & Xi, 2005) but there are some fundamental changes in terms of the
handling of proofs and proof construction. In particular, the requirement is dropped that a
proof in ATS must be represented as a normalizing lambda-term (Xi, 2008a).
In contrast to PTS, the key salient feature of ATS lies in a complete separation between
statics, where types are formed and reasoned about, from dynamics, where programs are
constructed and evaluated. This separation, with its origin in a previous study on a restricted form of dependent types developed in Dependent ML (DML) (Xi, 2007), makes it
straightforward to support dependent types in the presence of effects such as references and
exceptions. Also, with the introduction of two new (and thus somewhat unfamiliar) forms
of types: guarded types and asserting types, ATS is able to capture program invariants in
a manner that is similar to the use of pre-conditions and post-conditions (Hoare, 1969).
By now, studies have shown amply and convincingly that a variety of traditional programming paradigms (e.g., functional programming, object-oriented programming, metaprogramming, modular programming) can be directly supported in ATS without relying on
ad hoc extensions, attesting to the expressiveness of ATS. In this paper, the primary focus
of study is set on a novel programming paradigm referred to as programming with theoremproving (PwTP) and its support in ATS. In particular, a type-theoretical foundation for
PwTP is to be formally established and its correctness proven.
The notion of type equality plays a pivotal rôle in type system design. However, the importance of this rôle is often less evident in commonly studied type systems. For instance,
in the simply typed λ -calculus, two types are considered equal if and only if they are
syntactically the same; in the second-order polymorphic λ -calculus (λ2 ) (Reynolds, 1972)
and System F (Girard, 1986), two types are considered equal if and only if they are α equivalent; in the higher-order polymorphic λ -calculus (λω ), two types are considered
equal if and only if they are β η -equivalent. This situation immediately changes in ATS,
and let us see a simple example that stresses this point.
In Figure 1, the presented code implements a function in ATS (Xi, 2008b), which is a
substantial system such that its compiler alone currently consists of more than 165K lines
of code implemented in ATS itself.1 The concrete syntax used in the implementation should
be accessible to those who are familiar with Standard ML (SML) (Milner et al., 1997)).
Note that ATS is a programming language equipped with a type system rooted in ATS, and
1
Please see http://www.ats-lang.org for more details.
ZU064-05-FPR
main
28 March 2017
0:56
Journal of Functional Programming
3
fun
append
{a:type}{m,n:nat}
(
xs: list (a, m), ys: list (a, n)
) : list (a, m+n) =
case xs of
| nil() => ys (* the first clause *)
| cons(x, xs) => cons (x, append(xs, ys)) (* the second clause *)
// end of [append]
Fig. 1. List-append in ATS
the name of ATS derives from that of ATS. The type constructor list takes two arguments;
when applied to a type T and an integer I, list(T, I) forms a type for lists of length I in
which each element is of type T . Also, the two list constructors nil and cons are assigned
the following types:
nil
cons
:
:
∀a : type. () → list(a, 0)
∀a : type.∀n : nat. (a, list(a, n)) → list(a, n + 1)
So nil constructs a list of length 0, and cons takes an element and a list of length n to
form a list of length n + 1. The header of the function append indicates that append is
assigned the following type:
∀a : type.∀m : nat.∀n : nat. (list(a, m), list(a, n)) → list(a, m + n)
which simply means that append returns a list of length m + n when applied to one list of
length m and another list of length n. Note that type is a built-in sort in ATS, and a static
term of the sort type stands for a type (for dynamic terms). Also, int is a built-in sort for
integers in ATS, and nat is the subset sort {a : int | a ≥ 0} for all nonnegative integers.
When the above implementation of append is type-checked, the following two constraints are generated:
1. ∀m : nat.∀n : nat. m = 0 ⊃ n = m + n
2. ∀m : nat.∀n : nat.∀m′ : nat. m = m′ + 1 ⊃ (m′ + n) + 1 = m + n
The first constraint is generated when the first clause is type-checked, which is needed for
determining whether the types list(a, n) and list(a, m + n) are equal under the assumption that list(a, m) equals list(a, 0). Similarly, the second constraint is generated when
the second clause is type-checked, which is needed for determining whether the types
list(a, (m′ + n) + 1) and list(a, m + n) are equal under the assumption that list(a, m) equals
list(a, m′ + 1). Clearly, certain restrictions need to be imposed on the form of constraints
allowed in practice so that an effective approach can be found to perform constraintsolving. In DML, a programming language based on DML (Xi, 2007), the constraints
generated during type-checking are required to be linear inequalities on integers so that
the problem of constraint satisfaction can be turned into the problem of linear integer
programming, for which there are many highly practical solvers (albeit the problem of
linear integer programming itself is NP-complete). This is indeed a very simple design, but
ZU064-05-FPR
main
28 March 2017
4
0:56
Hongwei Xi
it can also be too restrictive, sometimes, as nonlinear constraints (e.g., ∀n : int.n ∗ n ≥ 0) are
commonly encountered in practice. Furthermore, the very nature of such a design indicates
its being inherently ad hoc.
By combining programming with theorem-proving, a fundamentally different design of
constraint-solving can provide the programmer with an option to handle nonlinear constraints through explicit proof construction. For the sake of a simpler presentation, let us
assume for this moment that even the addition function on integers cannot appear in the
constraints generated during type-checking. Under such a restriction, it is still possible to
implement a list-append function in ATS that is assigned a type capturing the invariant that
the length of the concatenation of two given lists xs and ys equals m + n if xs and ys are
of length m and n, respectively. Let us first see such an implementation given in Figure 2,
which is presented here as a motivating example for programming with theorem-proving
(PwTP).
The datatypes Z and S are declared in Figure 2 solely for representing natural numbers:
Z represents 0, and S(N) represents the successor of the natural number represented by N.
The data constructors associated with Z and S are of no use. Given a type T and another
type N, mylist(T, N) is a type for lists containing n elements of the type T , where n is
the natural number represented by N. Note that mylist is not a standard datatype (as is
supported in ML); it is a guarded recursive datatype (GRDT) (Xi et al., 2003), which is
also known as generalized algebraic datatype (GADT) (Cheney & Hinze, 2003) in Haskell
and OCaml. The datatype addrel is declared to capture the relation induced by the addition
function on natural numbers. Given types M, N, and R representing natural numbers m,
n, and r, respectively, the type addrel(M, N, R) is for a value representing some proof
of m + n = r. Note that addrel is also a GRDT or GADT. There are two constructors
addrel z and addrel s associated with addrel, which encode the following two rules:
0+n =
(m + 1) + n =
n
for every natural number n
(m + n) + 1 for every pair of natural numbers m and n
Let us now take a look at the implementation of myappend. Formally, the type assigned
to myappend can be written as follows:
∀a : type.∀m : type.∀n : type.
(mylist(a, m), mylist(a, n)) → ∃r : type. (addrel(m, n, r), mylist(a, r))
In essence, this type states the following: Given two lists of length m and n, myappend
returns a pair such that the first component of the pair is a proof showing that m + n equals
r for some natural number r and the second component is a list of length r.
Unlike append, type-checking myappend does not generate any linear constraints on
integers. As a matter of fact, myappend can be readily implemented in both Haskell and
OCaml (extended with support for generalized algebraic datatypes), where there is no
built-in support for handling linear constraints on integers. This is an example of great
significance in the sense that it demonstrates concretely an approach to allowing the programmer to write code of the nature of theorem-proving so as to simplify or even eliminate
certain constraints that need otherwise to be solved directly during type-checking. With
this approach, constraint-solving is effectively internalized, and the programmer can ac-
ZU064-05-FPR
main
28 March 2017
0:56
Journal of Functional Programming
5
datatype Z() = Z of ()
datatype S(a:type) = S of a
//
datatype
mylist(type, type) =
| {a:type}
mynil(a, Z())
| {a:type}{n:type}
mycons(a, S(n)) of (a, mylist(a, n))
//
datatype
addrel(type, type, type) =
| {n:type}
addrel_z(Z(), n, n) of ()
| {m,n:type}{r:type}
addrel_s(S(m), n, S(r)) of addrel(m, n, r)
//
fun
myappend
{a:type}
{m,n:type}
(
xs: mylist(a, m)
, ys: mylist(a, n)
) : [r:type]
(
addrel(m, n, r), mylist(a, r)
) =
(
case xs of
| mynil() => let
val pf = addrel_z() in (pf, ys)
end // end of [mynil]
| mycons(x, xs) => let
val (pf, res) = myappend(xs, ys) in (addrel_s(pf), mycons(x, res))
end // end of [mycons]
)
Fig. 2. A motivating example for PwTP in ATS
tively participate in constraint simplification, gaining a tight control in determining what
constraints should be passed to the underlying constraint-solver.
There are some major issues with the implementation given in Figure 2. Clearly, representing natural numbers as types is inadequate since there are types that do not represent
any natural numbers. More seriously, this representation turns quantification over natural
numbers (which is predicative) into quantification over types (which is impredicative),
causing unnecessary complications. Also, proof construction (that is, construction of values of types formed by addrel) needs to be actually performed at run-time, which causes
inefficiency both time-wise and memory-wise. Probably the most important issue is that
proof validity is not guaranteed. For instance, it is entirely possible to fake proof construction by making use of non-terminating functions.
ZU064-05-FPR
main
6
28 March 2017
0:56
Hongwei Xi
datasort
mynat = Z of () | S of mynat
//
datatype
mylist(type, mynat) =
| {a:type}
mynil(a, Z())
| {a:type}{n:mynat}
mycons(a, S(n)) of (a, mylist(a, n))
//
dataprop
addrel(mynat, mynat, mynat) =
| {y:mynat}
addrel_z(Z, y, y) of ()
| {x,y:mynat}{r:mynat}
addrel_s(S(x), y, S(r)) of addrel(x, y, r)
//
fun
myappend
{a:type}
{m,n:mynat}
(
xs: mylist(a, m)
, ys: mylist(a, n)
) : [r:mynat]
(
addrel(m, n, r) | mylist(a, r)
) =
(
case xs of
| mynil() => let
val pf = addrel_z() in (pf | ys)
end // end of [mynil]
| mycons(x, xs) => let
val (pf | res) = myappend(xs, ys) in (addrel_s(pf) | mycons(x, res))
end // end of [mycons]
)
Fig. 3. An example making use of PwTP in ATS
In Figure 3, another implementation of myappend is given that makes use of the support
for PwTP in ATS. Instead of representing natural numbers as types, a datasort of the
name mynat is declared and natural numbers can be represented as static terms of the
sort mynat. Also, a dataprop addrel is declared for capturing the relation induced by
the addition function on natural numbers. As a dataprop, addrel can only form types
for values representing proofs, which are erased after type-checking and thus need no
construction at run-time. In the implementation of myappend, the bar symbol (|) is used in
place of the comma symbol to separate components in tuples; the components appearing
to the left of the bar symbol are proof expressions (to be erased) and those to the right
are dynamic expressions (to be evaluated). After proof-erasure, the implementation of
myappend essentially matches that of append given in Figure 1.
ZU064-05-FPR
main
28 March 2017
0:56
Journal of Functional Programming
7
As a framework to facilitate the design and formalization of advanced type systems for
practical programming, ATS is first formulated with no support for PwTP (Xi, 2004). This
formulation is the basis for a type system referred to as ATS0 in this paper. The support for
PwTP is added into ATS in a subsequent formulation (Chen & Xi, 2005), which serves as
the basis for a type system referred to as ATSpf in this paper. However, a fundamentally
different approach is adopted in ATSpf to justify the soundness of PwTP, which essentially
translates each well-typed program in ATSpf into another well-typed one in ATS0 of the
same dynamic semantics. The identification and formalization of this approach, which is
both simpler and more general than one used previously (Chen & Xi, 2005), consists of a
major technical contribution of the paper.
It is intended that the paper should focus on the theoretical development of ATS, and
the presentation given is of a minimalist style. The organization for the rest of the paper is
given as follows. An untyped λ -calculus λdyn is first presented in Section 2 for the purpose
of introducing some basic concepts needed to formally assign dynamic (that, operational)
semantics to programs. In Section 3, a generic applied type system ATS0 is formulated and
its type-soundness established. Subsequently, ATS0 is extended to ATSpf in Section 4 with
support for PwTP, and the type-soundness of ATSpf is reduced to that of ATS0 through
a translation from well-typed programs in the former to those in the latter. Lastly, some
closely related work is discussed in Section 5 and the paper concludes.
2 Untyped λ -Calculus λdyn
The purpose of formulating λdyn , an untyped lambda-calculus extended with constants (including constant constructors and constant functions), is to set up some machinery needed
to formalize dynamic (that is, operational) semantics for programs. It is to be proven that a
well-typed program in ATS can be turned into one in λdyn through type-erasure and prooferasure while retaining its dynamic semantics, stressing the point that types and proofs in
ATS play no active rôle in the evaluation of a program. In this regard, the form of typing
studied in ATS is of Curry-style (in contrast with Church-style) (Reynolds, 1998).
There are no static terms in λdyn . The syntax for the dynamic terms in λdyn is given as
follows:
dynamic terms e
::=
x | dcx(~e) | he1 , e2 i | fst(e) | snd(e) |
lam x.e | app(e1 , e2 ) | let x = e1 in e2
where the notation ~e is for a possibly empty sequence of dynamic terms. Let dcx range
over external dynamic constants, which include both dynamic constructors dcc and dynamic functions dcf . The arguments taken by a dynamic constructor or function are often
primitive values (instead of those constructed by lam and h·, ·i) and the result returned by
it is often a primitive value as well. The meaning of various forms of dynamic terms should
become clear when the rules for evaluating them are given.
The values in λdyn are just special forms of dynamic terms, and the syntax for them is
given as follows:
values v
::=
x | dcc(~v) | hv1 , v2 i | lam x.e
ZU064-05-FPR
main
28 March 2017
0:56
8
Hongwei Xi
where ~v is for a possibly empty sequence of values. A standard approach to assigning
dynamic semantics to terms is based on the notion of evaluation contexts:
evaluation contexts E
::=
[] | dcx(v1 , . . . , vi−1 , E, ei+1 , . . . , en ) |
hE, ei | hv, Ei | app(E, e) | app(v, E) | let x = E in e
Essentially, an evaluation context E is a dynamic term in which a subterm is replaced with
a hole denoted by []. Note that only subterms at certain positions in a dynamic term can be
replaced to form valid evaluation contexts.
Definition 2.1
The redexes in λdyn and their reducts are defined as follows:
•
•
•
•
fst(hv1 , v2 i) is a redex, and its reduct is v1 .
snd(hv1 , v2 i) is a redex, and its reduct is v2 .
app(lam x.e, v) is a redex, and its reduct is e[x 7→ v].
dcf (~v) is a redex if it is defined to equal some value v; if so, its reduct is v.
Note that it may happen later that a new form of redex can have more than one reducts.
Given a dynamic term of the form E[e1 ] for some redex e1 , E[e1 ] is said to reduce to E[e2 ]
in one-step if e2 is a reduct of e1 , and this one-step reduction is denoted by E[e1 ] → E[e2 ].
Let →∗ stand for the reflexive and transitive closure of →.
Given a program (that is, a closed dynamic term) e0 in λdyn , a finite reduction sequence
starting from e0 can either lead to a value or a non-value. If a non-value cannot be further
reduced, then the non-value is said to be stuck or in a stuck form. In practice, values can
often be represented in special manners to allow various stuck forms to be detected through
checks performed at run-time. For instance, the representation of a value in a dynamically
typed language most likely contains a tag to indicate the type of the value. If it is detected
that the evaluation of a program reaches a stuck form, then the evaluation can be terminated
abnormally with a raised exception.
Detecting potential stuck forms that may occur during the evaluation of a program can
also be done statically (that is, at compiler-time). One often imposes a type discipline to
ensure the absence of various stuck forms during the evaluation of a well-typed program.
This is the line of study to be carried out in the rest of the paper.
3 Formal Development of ATS0
As a generic applied type system, ATS0 consists of a static component (statics), where
types are formed and reasoned about, and a dynamic component (dynamics), where programs are constructed and evaluated. The statics itself is a simply typed lambda-calculus
(extended with certain constants), and the types in it are called sorts so as to avoid confusion with the types for classifying dynamic terms, which are themselves static terms.
The syntax for the statics of ATS0 is given in Figure 4. Let b range over the base sorts
in ATS0 , which include at least bool for static booleans and type for types (assigned to
dynamic terms). The base sort int for static integers is not really needed for formalizing
ATS0 but it is often used in the presented examples. Let a and s range over static variables
and static terms, respectively. There may be some built-in static constants scx, which are
either static constant constructors scc or static constant functions scf. A c-sort is of the
ZU064-05-FPR
main
28 March 2017
0:56
Journal of Functional Programming
sorts
static terms
static var. ctx.
static subst.
σ
s
Σ
Θ
::=
::=
::=
::=
9
b | σ1 → σ2
a | scx(s1 , . . . , sn ) | λ a : σ .s | s1 (s2 )
0/ | Σ, a : σ
[] | Θ[a 7→ s]
Fig. 4. The syntax for the statics of ATS0
form (σ1 , . . . , σn ) ⇒ b, which can only be assigned to static constants. Note that a c-sort
is not considered a (regular) sort. Given a static constant scx, a static term scx(s1 , . . . , sn )
is of sort b if scx is assigned a c-sort (σ1 , . . . , σn ) ⇒ b for some sorts σ1 , . . . , σn and si can
be assigned the sorts σi for i = 1, . . . , n. It is allowed to write scc for scc() if there is no
risk of confusion. In ATS0 , the existence of the following static constants with the assigned
c-sorts is assumed:
true : () ⇒ bool
false : () ⇒ bool
≤ty : (type, type) ⇒ bool
∗
: (type, type) ⇒ type
→ : (type, type) ⇒ type
∧
: (bool, type) ⇒ type
⊃
: (bool, type) ⇒ type
: (σ → type) ⇒ type
∀σ
∃σ
: (σ → type) ⇒ type
Note that infix notation may be used for certain static constants. For instance, s1 → s2
stands for → (s1 , s2 ) and s1 ≤ty s2 stands for ≤ty (s1 , s2 ). In addition, ∀a : σ .s and ∃a : σ .s
stand for ∀σ (λ a : σ .s) and ∃σ (λ a : σ .s), respectively. Given a static constant constructor
scc, if the c-sort assigned to scc is (σ1 , . . . , σn ) ⇒ type for some sorts σ1 , . . . , σn , then scc is
a type constructor. For instance, ∗, →, ∧, ⊃, ∀σ and ∃σ are all type constructors. Additional
built-in base type constructors may be assumed.
Given a proposition B and a type T , B ⊃ T is a guarded type and B ∧ T is an asserting
type. Intuitively, if a value v is assigned a guarded type B ⊃ T , then v can be used only if
the guard B is satisfied; if a value v of an asserting type B ∧ T is generated at a program
point, then the assertion B holds at that point. For instance, suppose that int is a sort for
(static) integers and int is a type constructor of the sort (int) ⇒ type; given a static term
s of the sort int, int(s) is a singleton type for the integer equal to s; hence, the usual type
Int for (dynamic) integers can be defined as ∃a : int. int(a), and the type Nat for natural
numbers can be defined as ∃a : int. (a ≥ 0) ∧ int(a). Moreover, the following type is for
the (dynamic) division function on integers:
∀a1 : int.∀a2 : int. a2 6= 0 ⊃ (int(a1 ), int(a2 )) → int(a1 /a2 )
where the meaning of 6= and / should be obvious. With such a type, division by zero is
disallowed during type-checking (at compile-time). Also, suppose that bool is a type constructor of the sort (bool) ⇒ type such that for each proposition B, bool(B) is a singleton
type for the truth value equal to B. Then the usual type Bool for (dynamic) booleans can
be defined as ∃a : bool. bool(a). The following type is an interesting one:
∀a : bool. bool(a) → a ∧ 1
ZU064-05-FPR
main
10
28 March 2017
0:56
Hongwei Xi
Σ(a) = σ
(st-var)
Σ⊢a:σ
⊢ scx : (σ1 , . . . , σn ) ⇒ b Σ ⊢ s1 : σ1 · · · Σ ⊢ sn : σn
(st-scx)
Σ ⊢ scx(s1 , . . . , sn ) : b
Σ, a : σ1 ⊢ s : σ2
(st-lam)
Σ ⊢ λ a : σ1 .s : σ1 → σ2
Σ ⊢ s1 : σ1 → σ2 Σ ⊢ s2 : σ1
(st-app)
Σ ⊢ s1 (s2 ) : σ2
Fig. 5. The sorting rules for the statics of ATS0
where 1 stands for the unit type. Given a function f of this type, we can apply f to a
boolean value v of type bool(B) for some proposition B; if f (v) returns, the B must be true;
therefore f acts like dynamic assertion-checking.
For those familiar with qualified types (Jones, 1994), which underlies the type class
mechanism in Haskell, it should be noted that a qualified type cannot be regarded as a
guarded type. The simple reason is that the proof of a guard in ATS0 bears no computational
significance, that is, it cannot affect the run-time behavior of a program, while a dictionary,
which is just a proof of some predicate on types in the setting of qualified types, can and is
mostly likely to affect the run-time behavior of a program.
The standard rules for assigning sorts to static terms are given in Figure 5, where the
judgement ⊢ scx : (σ1 , . . . , σn ) ⇒ b means that the static constant scx is assumed to be of
the c-sort (σ1 , . . . , σn ) ⇒ b. Given ~s = s1 , . . . , sn and ~σ = σ1 , . . . , σn , a judgement of the
form Σ ⊢ ~s : ~σ means Σ ⊢ si : σi for i = 1, . . . , n. Let B stand for a static term that can be
assigned the sort bool (under some context Σ) and ~B a possibly empty sequence of static
boolean terms. Also, let T stand for a type (for dynamic terms), which is a static term
that can be assigned the sort type (under some context Σ). Given contexts Σ1 and Σ2 and
a substitution Θ, the judgement Σ1 ⊢ Θ : Σ2 means that Σ1 ⊢ Θ(a) : Σ2 (a) is derivable for
each a ∈ dom(Θ) = dom(Σ2 ).
Proposition 3.1
Assume Σ ⊢ s : σ is derivable. If Σ = Σ1 , Σ2 and Σ1 ⊢ Θ : Σ2 holds, then Σ1 ⊢ s[Θ] : σ is
derivable.
Proof
By structural induction on the derivation of Σ ⊢ s : σ .
Definition 3.1 (Constraints in ATS0 )
A constraint in ATS0 is of the form Σ; ~B |= B0 , where Σ ⊢ B : bool holds for each B in ~B and
Σ ⊢ B0 : bool holds as well, and the constraint relation in ATS0 is the one that determines
whether each constraint is true or false.
Each regularity rule in Figure 6 is assumed to be met, that is, the conclusion of each
regularity rule holds if all of its premisses hold, and the following regularity conditions on
≤ty are also satisfied:
1. Σ; ~B |= T ≤ty T holds for every T .
ZU064-05-FPR
main
28 March 2017
0:56
Journal of Functional Programming
11
B ∈ ~B
(reg-id)
Σ; ~B |= B
(reg-true)
Σ; ~B |= true
Σ; ~B |= false
(reg-false)
Σ; ~B |= B
Σ; ~B |= B
(reg-var-thin)
Σ, a : σ ; ~B |= B
Σ ⊢ B1 : bool Σ; ~B |= B2
(reg-bool-thin)
Σ; ~B, B1 |= B2
Σ, a : σ ; ~B |= B Σ ⊢ s : σ
(reg-subst)
Σ; ~B[a 7→ s] |= B[a 7→ s]
Σ; ~B |= B1
Σ; ~B, B1 |= B2
Σ; ~B |= B2
(reg-cut)
Fig. 6. The regularity rules for the constraint relation in ATS0
dynamic terms
e
::=
dynamic values
v
::=
dynamic var. ctx.
∆
::=
x | dcx{~s}(e1 , . . . , en ) |
he1 , e2 i | fst(e) | snd(e) | lam x. e | app(e1 , e2 ) |
⊃+ (e) | ⊃− (e) | slam a. e | sapp(e, s) |
∧(e) | let ∧(x) = e1 in e2 | hs, ei | let ha, xi = e1 in e2
x | dcc{~s}(v1 , . . . , vn ) |
hv1 , v2 i | lam x. e |⊃+ (e) | slam a. e | ∧(v) | hs, vi
0/ | ∆, x : T
dynamic subst.
Θ
::=
[] | Θ[x 7→ e]
Fig. 7. The syntax for the dynamics in ATS0
2.
3.
4.
5.
6.
7.
8.
9.
Σ; ~B |= T ≤ty T ′ and Σ; ~B |= T ′ ≤ty T ′′ implies Σ; ~B |= T ≤ty T ′′ .
Σ; ~B |= T1 ∗ T2 ≤ty T1′ ∗ T2′ implies Σ; ~B |= T1 ≤ty T1′ and Σ; ~B |= T2 ≤ty T2′ .
Σ; ~B |= T1 → T2 ≤ty T1′ → T2′ implies Σ; ~B |= T1′ ≤ty T1 and Σ; ~B |= T2 ≤ty T2′ .
Σ; ~B |= B ∧ T ≤ty B′ ∧ T ′ implies Σ; ~B, B |= B′ and Σ; ~B, B |= T ≤ty T ′ .
Σ; ~B |= B ⊃ T ≤ty B′ ⊃ T ′ implies Σ; ~B, B′ |= B and Σ; ~B, B′ |= T ≤ty T ′ .
Σ; ~B |= ∀a : σ .T ≤ty ∀a : σ .T ′ implies Σ, a : σ ; ~B |= T ≤ty T ′ .
Σ; ~B |= ∃a : σ .T ≤ty ∃a : σ .T ′ implies Σ, a : σ ; ~B |= T ≤ty T ′ .
0;
/ 0/ |= scc(T1 , . . . , Tn ) ≤ty T ′ implies T ′ = scc(T1′ , . . . , Tn′ ) for some T1′ , . . . , Tn′ .
The need for these conditions is to become clear when proofs are constructed in the
following presentation for formally establishing various meta-properties of ATS0 . For instance, the last of the above conditions can be invoked to make the claim that T ′ ≤ty T1 →
T2 implies T ′ being of the form T1′ → T2′ . Note that this condition actually implies the
consistency of the constraint relation as not every constraint is valid.
Let us now move onto the dynamic component (dynamics) of ATS0 . The syntax for
the dynamics of ATS0 is given in Figure 7. Let x range over dynamic variables and dcx
ZU064-05-FPR
main
28 March 2017
0:56
12
Hongwei Xi
dynamic constants, which include both dynamic constant constructors dcc and dynamic
constant functions dcf . Some (unfamiliar) forms of dynamic terms are to be understood
when the rules for assigning types to them are presented. Let v range over values, which
are dynamic terms of certain special forms, and ∆ range over dynamic variable contexts,
which assign types to dynamic variables.
During the formal development of ATS0 , proofs are often constructed by induction on
derivations (represented as trees). Given a judgement J, D :: J means that D is a derivation
of J, that is, the conclusion of D is J. Given a derivation D, ht(D) stands for the height of
the tree that represents D.
In ATS0 , a typing judgement is of the form Σ; ~B; ∆ ⊢ e : T , and the rules for deriving such
a judgement are given in Figure 8. Note that certain obvious side conditions associated with
some of the typing rules are omitted for the sake of brevity. For instance, the variable a is
not allowed to have free occurrences in ~B, ∆, or T when the rule (ty-∀-intr) is applied.
Given ~B = B1 , . . . , Bn , ~B ⊃ T stands for B1 ⊃ (· · · (Bn ⊃ T ) · · ·). Given ~a = a1 , . . . , an and
~σ = σ1 , . . . , σn , ∀~a : ~σ stands for the sequence of quantifiers: ∀a : σ1 . · · · ∀a : σn . A c-type
in ATS0 is of the form ∀~a : ~σ . ~B ⊃ (T1 , . . . , Tn ) ⇒ T .
The notation ⊢ dcx : ∀~a : ~σ . ~B ⊃ (T1 , . . . , Tn ) ⇒ T means that dcx is assumed to have
the c-type following it; if dcx is a constructor dcc, then T is assumed to be constructed by
some scc and dcc is said to be associated with scc. For instance, the list constructors and
the integer addition and division functions can be given the following c-types:
nil
cons
iadd
isub
imul
idiv
:
:
:
:
:
:
∀a : type. list(a, 0)
∀a : type.∀n : int. n ≥ 0 ⊃ (a, list(a, n)) → list(a, n + 1)
∀a1 : int.∀a2 : int. (int(a1 ), int(a2 )) ⇒ int(a1 + a2 )
∀a1 : int.∀a2 : int. (int(a1 ), int(a2 )) ⇒ int(a1 − a2 )
∀a1 : int.∀a2 : int. (int(a1 ), int(a2 )) ⇒ int(a1 ∗ a2 )
∀a1 : int.∀a2 : int. a2 6= 0 ⊃ (int(a1 ), int(a2 )) ⇒ int(a1 /a2)
where the type constructors int and list are type constructors of the c-sorts (int) ⇒ type
and (type, int) ⇒ type, respectively, and +, −, ∗, and / are static constant functions of the
c-sort (int, int) ⇒ int.
For a technical reason, the rule (ty-var) is to be replaced with the following one:
∆(x) = T
Σ; ~B |= T ≤ty T ′
Σ; ~B; ∆ ⊢ x : T ′
(ty-var’)
which combines (ty-var) with (ty-sub). This replacement is needed for establishing the
following lemma:
Lemma 3.1
Assume D :: Σ; ~B; ∆, x : T1 ⊢ e : T2 and Σ; ~B |= T1′ ≤ty T1 . Then there is a derivation D ′ for
the typing judgement Σ; ~B; ∆, x : T1′ ⊢ e : T2 such that ht(D ′ ) = ht(D).
Proof
The proof follows from structural induction on D immediately. The only interesting case
is the one where the last applied rule is (ty-var’), and this case can be handled by simply
merging two consecutive applications of the rule (ty-var’) into one (with the help of the
regularity condition stating that ≤ty is transitive).
ZU064-05-FPR
main
28 March 2017
0:56
Journal of Functional Programming
13
⊢ Σ; ~B; ∆ ∆(x) = T
(ty-var)
Σ; ~B; ∆ ⊢ x : T
Σ; ~B; ∆ ⊢ e : T Σ; ~B |= T ≤ty T ′
(ty-sub)
Σ; ~B; ∆ ⊢ e : T ′
⊢ dcx : ∀~a : ~σ .~B0 ⊃ (T1 , . . . , Tn ) ⇒ T
Σ; ~B |= B[~a 7→ ~s] for each B ∈ ~B0
Σ ⊢~s : ~σ
~
Σ; B; ∆ ⊢ ei : Ti [~a 7→ ~s] for i = 1, . . . , n
(ty-dcx)
Σ; ~B; ∆ ⊢ dcx{~s}(e1 , . . . , en ) : T [~a 7→ ~s]
Σ; ~B; ∆ ⊢ e1 : T1 Σ; ~B; ∆ ⊢ e2 : T2
(ty-tup)
Σ; ~B; ∆ ⊢ he1 , e2 i : T1 ∗ T2
Σ; ~B; ∆ ⊢ e : T1 ∗ T2
Σ; ~B; ∆ ⊢ e : T1 ∗ T2
(ty-fst)
(ty-snd)
~
Σ; B; ∆ ⊢ fst(e) : T1
Σ; ~B; ∆ ⊢ snd(e) : T2
Σ; ~B; ∆, x : T1 ⊢ e : T2
(ty-lam)
Σ; ~B; ∆ ⊢ lam x.e : T1 → T2
Σ; ~B; ∆ ⊢ e1 : T1 → T2 Σ; ~B; ∆ ⊢ e2 : T1
(ty-app)
Σ; ~B; ∆ ⊢ app(e1 , e2 ) : T2
Σ; ~B, B; ∆ ⊢ e : T
(ty-⊃-intr)
Σ; ~B; ∆ ⊢ ⊃+ (e) : B ⊃ T
Σ; ~B; ∆ ⊢ e : B ⊃ T Σ; ~B |= B
(ty-⊃-elim)
Σ; ~B; ∆ ⊢ ⊃− (e) : T
Σ; ~B |= B Σ; ~B; ∆ ⊢ e : T
(ty-∧-intr)
Σ; ~B; ∆ ⊢ ∧(e) : B ∧ T
Σ; ~B; ∆ ⊢ e1 : B ∧ T1 Σ; ~B, B; ∆, x : T1 ⊢ e2 : T2
(ty-∧-elim)
Σ; ~B; ∆ ⊢ let ∧ (x) = e1 in e2 : T2
Σ, a : σ ; ~B; ∆ ⊢ e : T
(ty-∀-intr)
Σ; ~B; ∆ ⊢ slam a.e : ∀a : σ .T
Σ; ~B; ∆ ⊢ e : ∀a : σ .T Σ ⊢ s : σ
(ty-∀-elim)
Σ; ~B; ∆ ⊢ sapp(e, s) : T [a 7→ s]
Σ ⊢ s : σ Σ; ~B; ∆ ⊢ e : T [a 7→ s]
(ty-∃-intr)
Σ; ~B; ∆ ⊢ hs, di : ∃a : σ .T
Σ; ~B; ∆ ⊢ e1 : ∃a : σ .T1 Σ, a : σ ; ~B; ∆, x : T1 ⊢ e2 : T2
(ty-∃-elim)
Σ; ~B; ∆ ⊢ let ha, xi = e1 in e2 : T2
Fig. 8. The typing rules for the dynamics of ATS0
Given Σ, ~B, ∆1 , ∆2 and θ , the judgement Σ; ~B; ∆1 ⊢ θ : ∆2 means that the typing judgement
Σ; ~B; ∆1 ⊢ θ (x) : ∆2 (x) is derivable for each x ∈ dom(θ ) = dom(∆2 ).
Lemma 3.2 (Substitution in ATS0 )
Assume D :: Σ; ~B; ∆ ⊢ e : T in ATS0 .
ZU064-05-FPR
main
28 March 2017
14
0:56
Hongwei Xi
1. If ~B = ~B1 , ~B2 and Σ; ~B1 |= ~B2 holds, then Σ; ~B1 ; ∆ ⊢ e : T is also derivable, where
Σ; ~B1 |= ~B2 means Σ; ~B1 |= B holds for each B ∈ ~B2 .
2. If Σ = Σ1 , Σ2 and Σ1 ⊢ Θ : Σ2 holds, then Σ1 ; ~B[Θ]; ∆[Θ] ⊢ d[Θ] : T [Θ] is also derivable.
3. If ∆ = ∆1 , ∆2 and Σ; ~B; ∆1 ⊢ θ : ∆2 is derivable, then Σ; ~B; ∆1 ⊢ d[θ ] : T is also
derivable.
Proof
By structural induction on the derivation D.
Lemma 3.3 (Canonical Forms)
Assume D :: 0;
/ 0;
/ 0/ ⊢ v : T . Then the following statements hold:
1.
2.
3.
4.
5.
6.
7.
If T
If T
If T
If T
If T
If T
If T
= T1 ∗ T2, then v is of the form hv1 , v2 i.
= T1 → T2 , then v is of the form lam x.e.
= B ∧ T0 , then v is of the form ∧(v0 ).
= B ⊃ T0 , then v is of the form ⊃+ (e).
= ∀a : σ .T0 , then v is of the form slam a.e.
= ∃a : σ .T0 , then v is of the form hs, v0 i.
= scc(~s1 ), then v is of the form dcc{~s2 }(~v) for some dcc associated with scc.
Proof
With Definition 3.1, the lemma follows from structural induction on D. If the last applied
rule in D is (ty-sub), then the proof goes through by invoking the induction hypothesis on
the immediate subderivation of D. Otherwise, the proof follows from a careful inspection
of the typing rules in Figure 8.
In order to assign (call-by-value) dynamic semantics to the dynamic terms in ATS0 , let
us introduce evaluation contexts as follows:
eval. ctx. E ::=
[] | dcx{~s}(~v, E,~e) | hE, di | hv, Ei | app(E, e) | app(v, E) |
⊃− (E) | ∧(E) | let ∧(x) = E in e | sapp(E, s) | hs, Ei | let ha, xi = E in e
Definition 3.2
The redexes and their reducts are defined as follows.
•
•
•
•
•
•
•
•
fst(hv1 , v2 i) is a redex, and its reduct is v1 .
snd(hv1 , v2 i) is a redex, and its reduct is v2 .
app(lam x.e, v) is a redex, and its reduct is e[x 7→ v].
dcf {~s}(~v) is a redex if it is defined to equal some value v; if so, its reduct is v.
⊃− (⊃+ (e)) is a redex, and its reduct is e.
sapp(slam a.e, s) is a redex, and its reduct is e[a 7→ s].
let ∧(x) = ∧(v) in e is a redex, and its reduct is e[x 7→ v].
let ha, xi = hs, vi in e is a redex, and its reduct is e[a 7→ s][x 7→ v].
Given two dynamic terms e1 and e2 such that e1 = E[e] and e2 = E[e′ ] for some redex e
and its reduct e′ , e1 is said to reduce to e2 in one step and this one-step reduction is denoted
by e1 → e2 . Let →∗ stand for the reflexive and transitive closure of →.
ZU064-05-FPR
main
28 March 2017
0:56
Journal of Functional Programming
15
It is assumed that the type assigned to each dynamic constant function dcf is appropriate,
that is, 0;
/ 0;
/ 0/ ⊢ v : T is derivable whenever 0;
/ 0;
/ 0/ ⊢ dcf {~s}(v1 , . . . , vn ) : T is derivable and
v is a reduct of dcf {~s}(v1 , . . . , vn ).
Lemma 3.4 (Inversion)
Assume D :: Σ; ~B; ∆ ⊢ e : T in ATS0 .
1. If e = he1 , e2 i, then there exists D ′ :: Σ; ~B; ∆ ⊢ e : T such that ht(D ′ ) ≤ ht(D) and the
last rule applied in D ′ is (ty-tup).
2. If e = lam x.e1 , then there exists D ′ :: Σ; ~B; ∆ ⊢ e : T such that ht(D ′ ) ≤ ht(D) and
the last applied rule in D ′ is (ty-lam).
3. If e =⊃+ (e1 ), then there exists D ′ :: Σ; ~B; ∆ ⊢ e : T such that ht(D ′ ) ≤ ht(D) and the
last rule applied in D ′ is (ty-⊃-intr).
4. If e = ∧(e1 ), then there exists D ′ :: Σ; ~B; ∆ ⊢ e : T such that ht(D ′ ) ≤ ht(D) and the
last rule applied in D ′ is (ty-∧-intr).
5. If e = slam a.e1 , then there exists D ′ :: Σ; ~B; ∆ ⊢ e : T such that ht(D ′ ) ≤ ht(D), and
the last rule applied in D ′ is (ty-∀-intr).
6. If e = hs, e1 i, then there exists D ′ :: Σ; ~B; ∆ ⊢ e : T such that ht(D ′ ) ≤ ht(D), and the
last rule applied in D ′ is (ty-∃-intr).
Proof
Let D ′ be D if D does not end with an application of the rule (ty-sub). Hence, in the rest
of the proof, it can be assumed that the last applied rule in D is (ty-sub), that is, D is of
the following form:
D1 :: Σ; ~B; ∆ ⊢ e : T ′ Σ; ~B |= T ′ ≤ty T
(ty-sub)
Σ; ~B; ∆ ⊢ e : T
Let us prove (1) by induction on ht(D). By induction hypothesis on D1 , there exists a
derivation D1′ :: Σ; ~B; ∆ ⊢ e : T ′ such that ht(D1′ ) ≤ ht(D1 ) and the last applied rule in D1′ is
(ty-tup):
′ :: Σ; ~
′ :: Σ; ~
D21
B; ∆ ⊢ e1 : T1′ D22
B; ∆ ⊢ e2 : T2′
(ty-tup)
Σ; ~B; ∆ ⊢ he1 , e2 i : T1′ ∗ T2′
where T ′ = T1′ ∗ T2′ and e = he1 , e2 i. By one of the regularity condition, T = T1 ∗ T2 for some
T1 and T2 . By another regularity condition, both Σ; ~B |= T1′ ≤ty T1 and Σ; ~B |= T2′ ≤ty T2 hold.
′ , one obtains D :: Σ; ~
By applying (ty-sub) to D21
B; ∆ ⊢ e1 : T1 . By applying (ty-sub) to
21
′
′
~
D22 , one obtains D22 :: Σ; B; ∆ ⊢ e2 : T2 . Let D be
D21 :: Σ; ~B; ∆ ⊢ e1 : T1 D22 :: Σ; ~B; ∆ ⊢ e2 : T2
(ty-tup)
Σ; ~B; ∆ ⊢ he1 , e2 i : T1 ∗ T2
and the proof for (1) is done since ht(D ′ ) = 1 + max(ht(D21 ), ht(D22 )), which equals 1 +
′ ), ht(D ′ )) = 1 + ht(D ′ ) ≤ 1 + ht(D ) = ht(D).
1 + max(ht(D21
1
22
1
Let us prove (2) by induction on ht(D). By induction hypothesis on D1 , there exists a
derivation D1′ :: Σ; ~B; ∆ ⊢ e : T ′ such that ht(D1′ ) ≤ ht(D1 ) and the last applied rule in D1′ is
(ty-lam):
D2′ :: Σ; ~B; ∆, x : T1′ ⊢ e1 : T2′
(ty-lam)
Σ; ~B; ∆ ⊢ lam x.e1 : T1′ → T2′
ZU064-05-FPR
main
28 March 2017
0:56
16
Hongwei Xi
where T ′ = T1′ → T2′ and e = lam x.e1 . By one of the regularity conditions, T = T1 → T2
for some T1 and T2 . By another regularity condiditon, both Σ; ~B |= T1 ≤ty T1′ and Σ; ~B |=
T2′ ≤ty T2 hold. Hence, by Lemma 3.1, there is a derivation D2′′ :: Σ; ~B; ∆, x : T1 ⊢ e1 : T2′
such that ht(D2′′ ) = ht(D2′ ). Let D ′ be the following derivation,
D2′′ :: Σ; ~B; ∆, x : T1 ⊢ e1 : T2′ Σ; ~B |= T2′ ≤ty T2
(ty-sub)
Σ; ~B; ∆, x : T1 ⊢ e1 : T2
(ty-lam)
Σ; ~B; ∆ ⊢ lam x.e1 : T1 → T2
and the proof for (2) is done since ht(D ′ ) = 1+1+ht(D2′′ ) = 1+1+ht(D2′ ) = 1+ht(D1′ ) ≤
1 + ht(D1 ) = ht(D).
The rest of statements (3), (4), (5), and (6) can all be proven similarly.
Theorem 3.1 (Subject Reduction in ATS0 )
Assume D :: Σ; ~B; ∆ ⊢ e : T in ATS0 and e → e′ holds. Then Σ; ~B; ∆ ⊢ e′ : T is also derivable
in ATS0 .
Proof
The proof proceeds by induction on ht(D).
• The last applied rule in D is (ty-sub):
D1 :: Σ; ~B; ∆ ⊢ e : T ′ Σ |= T ′ ≤ty T
Σ; ~B; ∆ ⊢ e : T
By induction hypothesis on D1 , D1′ :: Σ; ~B; ∆ ⊢ e′ : T ′ is derivable, and thus the
following derivation is obtained:
D1′ :: Σ; ~B; ∆ ⊢ e′ : T ′ Σ |= T ′ ≤ty T
Σ; ~B; ∆ ⊢ e′ : T
• The last applied rule in D is not (ty-sub). Assume that e = E[e0 ] and e′ = E[e′0 ],
where e0 is a redex and e′0 is a reduct of e0 . All the cases where E is not [] can be
readily handled, and some details are given as follows on the case where E = [] (that
is, e is itself a redex).
— D is of the following form:
D1 :: Σ; ~B; ∆ ⊢ hv11 , v12 i : T1 ∗ T2
(ty-fst)
Σ; ~B; ∆ ⊢ fst(hv11 , v12 i) : T1
where T = T1 and e = fst(hv11 , v12 i). By Lemma 3.4, D1 may be assumed to be
of the following form:
D21 :: Σ; ~B; ∆ ⊢ v11 : T1 D22 :: Σ; ~B; ∆ ⊢ v12 : T2
(ty-tup)
Σ; ~B; ∆ ⊢ hv11 , v12 i : T1 ∗ T2
Note that e′ = v11 , and the case concludes.
— D is of the following form:
D1 :: Σ; ~B; ∆ ⊢ lam x.e1 : T1 → T2 D2 :: Σ; ~B; ∆ ⊢ v2 : T1
(ty-app)
Σ; ~B; ∆ ⊢ app(lam x.e1 , v2 ) : T2
ZU064-05-FPR
main
28 March 2017
0:56
Journal of Functional Programming
17
where T = T2 and e = app(lam x.e1 , v2 ). By Lemma 3.4, D1 may be assumed
to be of the following form:
Σ; ~B; ∆, x : T1 ⊢ e1 : T2
Σ; ~B; ∆ ⊢ lam x.e1 : T1 → T2
By Lemma 3.2 (Substitution), Σ; ~B; ∆ ⊢ e1 [x 7→ v2 ] : T2 is derivable. Note that
e′ = e1 [x 7→ v2 ], and the case concludes.
All of the other cases can be handled similarly.
For a less involved presentation, let us assume that any well-typed closed value of the
form dcf {~s}(v1 , . . . , vn ) is a redex, that is, the dynamic constant function dcf is well-defined
at the arguments v1 , . . . , vn .
Theorem 3.2 (Progress in ATS0 )
Assume that D :: 0;
/ 0;
/ 0/ ⊢ e : T in ATS0 . Then either e is a value or e → e′ holds for some
′
dynamic term e .
Proof
With Lemma 3.3 (Canonical Forms), the proof proceeds by a straightforward structural
induction on D.
By Theorem 3.1 and Theorem 3.2, it is clear that for each closed well-typed dynamic
term e, e →∗ v holds for some value v, or there is an infinite reduction sequence starting
from e: e = e0 → e1 → e2 → · · ·. In other words, the evaluation of a well-typed program
in ATS0 either reaches a value or goes on forever (as it can never get stuck). This metaproperty of ATS0 is often referred to as its type-soundness. Per Robin Milner, a catchy
slogan for type-soundness states that a well-typed program can never go wrong.
kxk
kdcx{~s}(e1 , . . . , en )k
klam x. ek
kapp(e1 , e2 )k
k⊃+ (e)k
k⊃− (e)k
k∧(e)k
klet ∧(x) = e1 in e2 k
kslam a. ek
ksapp(e, s)k
=
=
=
=
=
=
=
=
=
=
x
dcx(ke1 k, . . . , ken k)
lam x. kek
app(ke1 k, ke2 k)
kek
kek
kek
let x =ke1 k in ke2 k
kek
kek
Fig. 9. The type-erasure function k·k on dynamic terms
After a program in ATS passes type-checking, it goes through a process referred to as
type-erasure to have the static terms inside it completely erased. In Figure 9, a function
performing type-erasure is defined, which maps each dynamic term in ATS0 to an untyped
dynamic term in λdyn .
In order to guarantee that a value in ATS0 is mapped to another value in λdyn by the
function k·k, the following syntactic restriction is needed:
ZU064-05-FPR
main
28 March 2017
0:56
18
Hongwei Xi
• Only when e is a value can the dynamic term ⊃+ (e) be formed.
• Only when e is a value can the dynamic term slam a.e be formed.
This kind of restriction is often referred to as value-form restriction.
Proposition 3.2
With the value-form restriction being imposed, kvk is a value in λdyn for every value v in
ATS0 .
Proof
By structural induction on v.
Note that it is certainly possible to have a non-value e in ATS0 whose type-erasure is a
value in λdyn . From this point on, the value-form restriction is always assumed to have
been imposed when type-erasure is performed.
Proposition 3.3
Assume that e1 is a well-typed closed dynamic term in ATS0 . If e1 → e2 holds, then either
ke1 k=ke2 k or ke1 k→ke2 k holds in λdyn .
Proof
By a careful inspection of the forms of redexes in Definition 3.2.
Proposition 3.4
Assume that e1 is a well-typed closed dynamic term in ATS0 . If ke1 k→ e′2 holds in λdyn ,
then there exists e2 such that e1 →∗ e2 holds in ATS0 and ke2 k= e′2 .
Proof
By induction on the height of the typing derivation for e1 .
By Proposition 3.3 and Proposition 3.4, it is clear that type-erasure cannot alter the
dynamic semantics of a well-typed dynamic term in ATS0 .
The formulation of ATS0 presented in this section is of a minimalist style. In particular,
the constraint relation in ATS0 is treated abstractly. In practice, if a concrete instance of
ATS0 is to be implemented, then rules need to be provided for simplifying constraints. For
instance, the following rule may be present:
Σ; ~B |= I1 = I2
Σ; ~B |= int(I1 ) ≤ty int(I2 )
With this rule, int(I1 ) ≤ty int(I2 ) can be simplified to the constraint I1 = I2 , where the
equality is on static integer terms. The following rule may also be present:
Σ; ~B |= T1 ≤ty T2 Σ; ~B |= I1 = I2
Σ; ~B |= list(T1 , I1 ) ≤ty list(T2 , I2 )
With this rule, list(T1 , I1 ) ≤ty list(T2 , I2 ) can be simplified to the two constraints T1 ≤ty T2
and I1 = I2 .
For those interested in implementing an applied type system, please find more details in
a paper on DML (Xi, 2007), which is regarded a special kind of applied type system.
ZU064-05-FPR
main
28 March 2017
0:56
Journal of Functional Programming
19
4 Formal Development of ATSpf
Let us extend ATS0 to ATSpf in this section with support for programming with theoremproving (PwTP).
A great limitation on employing ATS0 as the basis for a practical programming language lies in the very rigid handling of constraint-solving in ATS0 . One is often forced to
impose various ad hoc restrictions on the syntactic form of a constraint that can actually
be supported in practice (so as to match the capability of the underlying constraint-solver),
greatly diminishing the effectiveness of using types to capture programming invariants.
For instance, only quantifier-free constraints that can be translated into problems of linear
integer programming are allowed in the DML programming language (Xi, 2001).
With PwTP being supported in a programming language, programming and theoremproving can be combined in a syntactically intertwined manner (Chen & Xi, 2005); if a
constraint cannot be handled directly by the underlying constraint-solver, then it is possible
to simplify the constraint or even eliminate it through explicit proof construction. PwTP advocates an open style of constraint-solving by providing a means within the programming
language itself to allow the programmer to actively participate in constraint-solving. In
other words, PwTP can be viewed as a programming paradigm for internalizing constraintsolving.
≤ pr
∗
∗
→
→
∧
⊃
∀σ
∃σ
:
:
:
:
:
:
:
:
:
(prop, prop) ⇒ bool
(prop, prop) ⇒ prop
(prop, type) ⇒ type
(prop, prop) ⇒ prop
(prop, type) ⇒ type
(bool, prop) ⇒ prop
(bool, prop) ⇒ prop
(σ → prop) ⇒ prop
(σ → prop) ⇒ prop
Fig. 10. Additional static constants in ATSpf
Let us now start with the formulation of ATSpf , which extends that of ATS0 fairly lightly.
In addition to the base sorts in ATS0 , ATSpf contains another base sort prop, which is for
static terms representing types for proofs. A static term of the sort prop may be referred
to as a prop (or, sometimes, a type for proofs). Also, it is assumed that the static constants
listed in Figure 10 are included in ATSpf . Note that the symbols referring to these static
constants may be overloaded. In the following representation, P stands for a prop, T stands
for a type, and T ∗ stands for either a prop or a type.
The syntax for dynamic terms in ATSpf is essentially the same as that in ATS0 but
with a few minor changes to be mentioned as follows. Some dynamic constructs in ATS0
need to be split when they are incorporated into ATSpf . The construct he1 , e2 i for forming
tuples is split into he1 , e2 ipp , he1 , e2 ipt , and he1 , e2 itt for prop-type pairs, prop-type pairs
and type-type pairs, respectively. For instance, a prop-type pair is one where the first
component is assigned a prop and the second one a type. Note that there are no typeprop pairs. The construct lam x.e for forming lambda-abstractions is split into lampp x. e,
lampt x. e, and lamtt x. e for prop-prop functions, prop-type functions and type-type func-
ZU064-05-FPR
main
20
28 March 2017
0:56
Hongwei Xi
tions, respectively. For instance, a prop-type function is one where the argument is assigned
a prop and the body a type. The construct app(e1 , e2 ) for forming applications is split
into apppp (e1 , e2 ), apptp (e1 , e2 ), and apptt (e1 , e2 ) for prop-prop applications, type-prop
applications and type-type applications. For instance, a type-prop application is one where
the function part is assigned a type and the argument a prop. Note that there are no typeprop functions.
The dynamic variable contexts in ATSpf are defined as follows:
dynamic var. ctx. ∆
::=
0/ | ∆, x : T ∗
The regularity conditions on ≤ty needs to be extended with the following two for the
new forms of types:
3.2 Σ; ~B |= P1 ∗ T2 ≤ty P1′ ∗ T2′ implies Σ; ~B |= P1 ≤ pr P1′ and Σ; ~B |= T2 ≤ty T2′ .
4.2 Σ; ~B |= P1 → T2 ≤ty P1′ → T2′ implies Σ; ~B |= P1′ ≤ pr P1 and Σ; ~B |= T2 ≤ty T2′ .
It should be noted that there are no regularity conditions imposed on props (as it is not
expected for proofs to have any computational meaning).
There are two kinds of typing rules in ATSpf : p-typing rules and t-typing rules, where
the former is for assigning props to dynamic terms (encoding proofs) and the latter for assigning types to dynamic terms (to be evaluated). The typing rules for ATSpf are essentially
those for ATS0 listed in Figure 8 except for the following changes:
• Each occurrence of T in the rules for ATS0 needs to be replaced with T ∗ .
• The premisses of each p-typing rule (that is, one for assigning a prop to a dynamic
term) are required to be p-typing rules themselves.
As an example, let us take a look at the following rule:
Σ; ~B; ∆ ⊢ e : T Σ; ~B |= T ≤ty T ′
(ty-sub)
Σ; ~B; ∆ ⊢ e : T ′
which yields the following two valid versions:
Σ; ~B; ∆ ⊢ e : P Σ; ~B |= P ≤ pr P′
(ty-sub-p)
Σ; ~B; ∆ ⊢ e : P′
Σ; ~B; ∆ ⊢ e : T Σ; ~B |= T ≤ty T ′
(ty-sub-t)
Σ; ~B; ∆ ⊢ e : T ′
As another example, let us take a look at the following rule:
Σ; ~B; ∆ ⊢ e : T1 ∗ T2
(ty-fst)
Σ; ~B; ∆ ⊢ fst(e) : T1
which yields the following two valid versions:
Σ; ~B; ∆ ⊢ e : P1 ∗ P2
(ty-fst-pp)
Σ; ~B; ∆ ⊢ fst(e) : P1
Σ; ~B; ∆ ⊢ e : T1 ∗ T2
(ty-fst-tt)
Σ; ~B; ∆ ⊢ fst(e) : T1
ZU064-05-FPR
main
28 March 2017
0:56
Journal of Functional Programming
21
Note that there is no type of the form T1 ∗ P2 (for the sake of simplicity). The following
version is invalid:
Σ; ~B; ∆ ⊢ e : P1 ∗ T2
(ty-fst-pt)
Σ; ~B; ∆ ⊢ fst(e) : P1
because a p-typing rule cannot have any t-typing rule as its premise. Instead, the following
typing rule is introduced as the elimination rule for P1 ∗ T2 :
Σ; ~B; ∆ ⊢ e : P1 ∗ T2 Σ; ~B; ∆, x1 : P1 , x2 : T2 ⊢ e0 : T0
(ty-∗-elim-pt)
Σ; ~B; ∆ ⊢ let hx1 , x2 ipt = e in e0 : T0
As yet another example, let us take a look at the following rule:
Σ; ~B; ∆ ⊢ e1 : T1∗ → T2∗ Σ; ~B; ∆ ⊢ e2 : T1∗
(ty-app)
Σ; ~B; ∆ ⊢ app(e1 , e2 ) : T2∗
which yields the following three versions:
Σ; ~B; ∆ ⊢ e1 : P1 → P2 Σ; ~B; ∆ ⊢ e2 : P1
(ty-app-pp)
Σ; ~B; ∆ ⊢ apppp (e1 , e2 ) : P2
Σ; ~B; ∆ ⊢ e1 : P1 → T2 Σ; ~B; ∆ ⊢ e2 : P1
(ty-app-tp)
Σ; ~B; ∆ ⊢ apptp (e1 , e2 ) : T2
Σ; ~B; ∆ ⊢ e1 : T1 → T2 Σ; ~B; ∆ ⊢ e2 : T1
(ty-app-tt)
Σ; ~B; ∆ ⊢ apptt (e1 , e2 ) : T2
The first one is a p-typing rule while the other two are t-typing rules.
In ATSpf , the two sorts bool and prop are intimately related but are also fundamentally
different. Gaining a solid understanding of the relation between these two is the key to
understanding the design of ATSpf . One may see prop as an internalized version of bool.
Given a static boolean term B, its truth value is determined by a constraint-solver outside
ATSpf . Given a static term P of the sort prop, a proof of P can be constructed inside ATSpf
to attest to the validity of the boolean term encoded by P. For clarification, let us see a
simple example illustrating the relation between bool and prop in concrete terms.
dataprop
fact_p(int, int) =
| fact_p_bas(0, 1) of ()
| {n:nat}{r:int}
fact_p_ind(n+1, (n+1)*r) of fact_p(n, r)
Fig. 11. A dataprop for encoding the factorial function
In Figure 11, the dataprop fact p declared in ATS is associated with two proof constructors that are assigned the following c-types (or, more precisely, c-props):
fact p bas
fact p ind
:
:
fact p(0, 1)
∀n : nat.∀r : int. (fact p(n, r)) ⇒ fact p(n + 1, (n + 1) ∗ r)
Let fact(n) be the value of the factorial function on n, where n ranges over natural numbers.
Given a natural number n and an integer r, the prop fact p(n, r) encodes the relation
ZU064-05-FPR
main
22
28 March 2017
0:56
Hongwei Xi
stacst
fact_b : (int, int) -> bool
praxi
fact_b_bas
(
// argless
) : [fact_b(0, 1)] unit_p
praxi
fact_b_ind{n:int}{r:int}
(
// argless
) : [n >= 0 && fact_b(n, r) ->> fact_b(n+1, (n+1)*r)] unit_p
Fig. 12. A static predicate and two associated proof functions
fact(n) = r. In other words, if a proof of the prop fact p(n, r) can be constructed, then
fact(n) equals r.
In Figure 12, a static predicate fact b is introduced, which corresponds to fact p. Given
a natural number n and an integer r, fact b(n, r) simply means fact(n) = r. The two proof
functions fact b bas and fact b ind are assigned the following c-props:
fact b bas : () ⇒ fact b(0, 1) ∧ 1
fact b ind : ∀n : int.∀r : int. () ⇒ (n ≥ 0 ∧ fact b(n, r) ⊃ fact b(n + 1, (n + 1) · r)) ∧ 1
where 1 is the unit prop (instead of the unit type) that encodes the static truth value true.
Note that the keyword praxi in ATS is used to introduce proof functions that are treated as
axioms.
In Figure 13, a verified implementation of the factorial function is given in ATS. Given
a natural numbers n, f fact p returns an integer r paired with a proof of fact p(n, r) that
attests to the validity of fact(n) = r. Note that this implementation makes explicit use of
proofs. The constraints generated from type-checking the code in Figure 13 are quantifierfree, and they can be readily solved by the built-in constraint-solver (based on linear integer
programming) for ATS.
In Figure 14, another verified implementation of the factorial function is given in ATS.
Given a natural numbers, f fact b returns an integer r plus the assertion fact b(n, r) that
states fact(n) = r. This implementation does not make explicit use of proofs. Applying the
keyword $solver assert to a proof turns the prop of the proof into a static boolean term
(of the same meaning) and then adds the term as an assumption to be used for solving the
constraints generated subsequently in the same scope. For instance, the two applications of
$solver assert essentially add the following two assumptions:
fact b(0, 1)
∀n : int.∀r : int. n ≥ 0 ∧ fact b(n, r) ⊃ fact b(n + 1, (n + 1) · r)
Note that the second assumption is universally quantified. In general, solving constraints
involving quantifiers is much more difficult than those that are quantifier-free. For instance, the constraints generated from type-checking the code in Figure 14 cannot be
solved by the built-in constraint-solver for ATS. Instead, these constraints need to be
ZU064-05-FPR
main
28 March 2017
0:56
Journal of Functional Programming
fun
f_fact_p
{n:nat}
(
n: int(n)
) : [r:int]
(fact_p(n, r) | int(r)) = let
//
fun
loop
{ i:nat
| i <= n
} {r:int}
(
pf: fact_p(i, r)
| i: int(i), r: int(r)
) : [r:int] (fact_p(n, r) | int(r)) =
if i < n then
loop(fact_p_ind(pf) | i+1, (i+1)*r) else (pf | r)
// end of [if]
//
in
loop(fact_p_bas() | 0(*i*), 1(*r*))
end // end of [f_fact_p]
Fig. 13. A verified implementation of the factorial function
fun
f_fact_b
{n:nat}
(
n: int(n)
) : [r:int]
(fact_b(n, r) && int(r)) = let
//
prval() = $solver_assert(fact_b_bas)
prval() = $solver_assert(fact_b_ind)
//
fun
loop
{ i:nat | i <= n}
{ r:int | fact_b(i, r) }
(
i: int(i), r: int(r)
) : [r:int]
(fact_b(n, r) && int(r)) =
if i < n then loop(i+1, (i+1)*r) else (r)
//
in
loop(0, 1)
end // end of [f_fact_b]
Fig. 14. Another verified implementation of the factorial function
23
ZU064-05-FPR
main
28 March 2017
0:56
24
Hongwei Xi
exported so that external constraint-solvers (for instance, one based on the Z3 theoremprover (de Moura & Bjørner, 2008)) can be invoked to solve them.
By comparing these two verified implementations of the factorial function, one sees a
concrete case where PwTP (as is done in Figure 13) is employed to simplify the constraints
generated from type-checking. This kind of constraint simplification through PwTP is a
form of internalization of constraint-solving, and it can often play a pivotal rôle in practice,
especially, when there is no effective method available for solving general unsimplified
constraints.
Instead of assigning (call-by-value) dynamic semantics to the dynamic terms in ATSpf
directly, a translation often referred to as proof-erasure is to be defined that turns each
dynamic term in ATSpf into one in ATS0 of the same dynamic semantics.
Given a sort σ , its proof-erasure |σ | is the one in which every occurrence of prop in σ
is replaced with bool.
Given a static variable context Σ, its proof-erasure |Σ| is obtained from replacing each
declaration a : σ with a : |σ |.
For every static constant scx of the c-sort (σ1 , . . . , σn ) ⇒ σ , it is assumed that there
exists a corresponding scx′ of the c-sort (|σ1 |, . . . , |σn |) ⇒ |σ |; this corresponding scx′ may
be denoted by |scx|. Note that it is possible to have |scx1 | = |scx2 | for different constants
scx1 and scx2 .
Let us assume the existence of the following static constants:
∧
⊃
∀σ
∃σ
:
:
:
:
(bool, bool) ⇒ bool
(bool, bool) ⇒ bool
(σ → bool) ⇒ bool
(σ → bool) ⇒ bool
Note that the symbols referring to these static constants are all overloaded. Naturally, ∧
and ⊃ are interpreted as the boolean conjunction and boolean implication, respectively,
and ∀σ and ∃σ are interpreted as the standard universal quantification and existential
quantification, respectively. For instance, some pairs of corresponding static constants are
listed as follows:
• The boolean implication function ⊃ corresponds to the prop predicate ≤ pr .
• The boolean implication function ⊃ corresponds to the prop constructor → of the
c-sort (prop, prop) ⇒ prop.
• The boolean implication function ⊃ corresponds to the prop constructor ⊃ of the
c-sort (bool, prop) ⇒ prop.
• The boolean conjunction function ∧ corresponds to the prop constructor ∗ of the
c-sort (prop, prop) ⇒ prop.
• The boolean conjunction function ∧ corresponds to the prop constructor ∧ of the
c-sort (bool, prop) ⇒ prop.
• The type constructor ∧ of the c-sort (bool, type) ⇒ type corresponds to the type
constructor ∗ of the c-sort (prop, type) ⇒ type.
• The type constructor ⊃ of the c-sort (bool, type) ⇒ type corresponds to the type
constructor → of the c-sort (prop, type) ⇒ type.
• For each sort σ , the universal quantifier ∀σ of the sort (σ → bool) ⇒ bool corresponds to the universal quantifier ∀σ of the sort (σ → prop) ⇒ prop.
ZU064-05-FPR
main
28 March 2017
0:56
Journal of Functional Programming
|x|
|dcx{~s}(~e)|
|he1 , e2 ipt |
|he1 , e2 itt |
|fst(e)|
|snd(e)|
|let hx p , xt ipt = e1 in e2 |
|lampt x. e|
|lamtt x. e|
|apptp (e1 , e2 )|
|apptt (e1 , e2 )|
| ⊃+ (e)|
| ⊃− (e)|
| ∧(e)|
|let ∧(x) = e1 in e2 |
|slam a. e|
|sapp(e, s)|
=
=
=
=
=
=
=
=
=
=
=
=
=
=
=
=
=
25
x
dcx{|~s|}(|~e|)
∧(|e2 |)
h|e1 |, |e2 |itt
fst(|e|)
snd(|e|)
let ∧(xt ) = |e1 | in |e2 |
⊃+ (|e|)
lam x.|e|
⊃− (|e1 |)
app(|e1 |, |e2 |)
⊃+ (|e|)
⊃− (|e|)
∧(|e|)
let ∧(x) = |e1 | in |e2 |
slam a. |e|
sapp(|e|, |s|)
Fig. 15. The proof-erasure function | · | on dynamic terms
• For each sort σ , the existential quantifier ∃σ of the sort (σ → bool) ⇒ bool corresponds to the existential quantifier ∃σ of the sort (σ → prop) ⇒ prop.
For every static term s, |s| is the static term obtained from replacing in s each σ with |σ |
and each scx with |scx|.
Proposition 4.1
Assume that Σ ⊢ s : σ is derivable. Then |Σ| ⊢ |s| : |σ | is also derivable.
Proof
By induction on the sorting derivation of Σ ⊢ s : σ .
For a sequence ~B of static boolean terms, |~B| is the sequence obtained from applying | · |
to each B in ~B.
There are two functions | · | p and | · |t for mapping a given dynamic variable context ∆ to
a sequence of boolean terms and a dynamic variable context, respectively:
• |∆| p is a sequence of boolean terms ~B such that each B in ~B is |P| for some a : P
declared in Σ.
• |∆|t is a dynamic variable context such each declaration in it is of the form a : |T | for
some a : T declared in Σ.
The proof-erasure function on dynamic terms is defined in Figure 15. Clearly, given a
dynamic term e in ATSpf , |e| is a dynamic term in ATS0 if it is defined.
As the proof-erasure of ≤ pr is chosen to be the boolean implication function, it needs to
be assumed that Σ; ~B ⊢ P1 ≤ pr P2 implies |Σ|; |~B| ⊢ |P1 | ⊃ |P2 |
Lemma 4.1 (Constraint Internalization)
Assume that the typing judgment Σ; ~B; ∆ ⊢ e : P is derivable in ATSpf . Then the constraint
|Σ|; |~B|, |∆| p |= |P| holds.
Proof
ZU064-05-FPR
main
26
28 March 2017
0:56
Hongwei Xi
By structural induction on the typing derivation D of Σ; ~B; ∆ ⊢ e : P. Note that the typing
rule (ty-sub-p) is handled by the assumption that Σ; ~B ⊢ P1 ≤ pr P2 implies |Σ|; |~B| ⊢ |P1 | ⊃
|P2 | for any props P1 and P2 .
• Assume that the last applied rule in D is (ty-tup-pp):
D1 :: Σ; ~B; ∆ ⊢ e1 : P1 D2 :: Σ; ~B; ∆ ⊢ e2 : P2
(ty-tup-pp)
Σ; ~B; ∆ ⊢ he1 , e2 ipp : P1 ∗ P2
where P = P1 ∗ P2 . By induction hypothesis on D1 , |Σ|; |~B|, |∆| p |= |P1 | holds. By
induction hypothesis on D2 , |Σ|; |~B|, |∆| p |= |P2 | holds. Note that |P| = |P1 ∗ P2 | =
|P1 | ∧ |P2 |, where ∧ stands for the boolean conjunction. Therefore, |Σ|; |~B|, |∆| p |= |P|
holds.
• Assume that the last applied rule in D is either (ty-fst-pp) or (ty-snd-pp). This case
immediately follows from the fact that |P1 ∗ P2| = |P1 | ∧ |P2 | for any props P1 and P2 ,
where ∧ stands for the boolean conjunction
• Assume that the last applied rule in D is (ty-lam-pp):
D1 :: Σ; ~B; ∆, x1 : P1 ⊢ e2 : P2
(ty-lam-pp)
Σ; ~B; ∆ ⊢ lampp x1 . e2 : P1 → P2
where P = P1 → P2 . By induction hypothesis on D1 , |Σ|; |~B|, |∆| p , |P1 | |= |P2 | holds.
By the regularity rule (reg-cut), |Σ|; |~B|, |∆| p |= |P2 | holds whenever |Σ|; |~B|, |∆| p |=
|P1 | holds. Therefore, |Σ|; |~B|, |∆| p |= |P1 | ⊃ |P2| holds, where ⊃ stands for the boolean
implication. Note that |P| = |P1 | ⊃ |P2 |, and this case concludes.
• Assume that the last applied rule in D is (ty-app-pp):
D1 :: Σ; ~B; ∆ ⊢ e1 : P1 → P2 D2 :: Σ; ~B; ∆ ⊢ e2 : P1
(ty-app-pp)
Σ; ~B; ∆ ⊢ apppp (e1 , e2 ) : P2
where P = P2 . By induction hypothesis on D1 , |Σ|; |~B|, |∆| p |= |P1 | ⊃ |P2 | holds,
where ⊃ stands for the boolean implication. By induction hypothesis on D2 , the
constraint |Σ|; |~B|, |∆| p |= |P1 | holds. Therefore, the constraint |Σ|; |~B|, |∆| p |= |P2 | also
holds.
The rest of the cases can be handled similarly.
Note that a proof in ATSpf can be non-constructive as it is not expected for the proof to have
any computational meaning. In particular, one can extend the proof construction in ATSpf
with any kind of reasoning based on classical logic (e.g., double negation elimination).
If a c-type CT assigned to a dynamic (proof) constant is of the form ∀Σ.~B ⊃ (~P) ⇒ P0 ,
then it is assumed that the following constraint holds in ATS0 :
0;
/ 0/ |= ∀|Σ|.|~B| ⊃ (|~P| ⊃ |P0 |)
For instance, the c-types assigned to fact p bas and fact p ind imply the validity of
the following constraints:
0;
/ 0/ ⊢ fact b(0, 1)
0;
/ 0/ ⊢ ∀n : int.∀r : int. (n ≥ 0 ∧ fact b(n, r) ⊃ fact b(n + 1, (n + 1) · r))
ZU064-05-FPR
main
28 March 2017
0:56
Journal of Functional Programming
27
which are encoded directly into the c-types assigned to fact b bas and fact b ind.
If a c-type CT is of the form ∀Σ.~B ⊃ (~P, T1 , . . . , Tn ) ⇒ T0 , then |CT| is defined as follows:
∀|Σ|.|~B| ⊃ (|~P| ⊃ ((|T1 |, . . . , |Tn |) ⇒ |T0 |))
If a dynamic constant dcx is assigned the c-type CT in ATSpf , then it is assumed to be of
the c-type |CT| in ATS0 .
Theorem 4.1
Assume that Σ; ~B; ∆ ⊢ e : T is derivable in ATSpf . Then |Σ|; |~B|, |∆| p ; |∆|t ⊢ |e| : |T | is
derivable in ATS0 ,
Proof
By structural induction on the typing derivation D of Σ; ~B; ∆ ⊢ e : T .
• Assume that the last applied rule in D is (ty-∗-elim-pt):
D1 :: Σ; ~B; ∆ ⊢ e12 : P1 ∗ T2 D2 :: Σ; ~B; ∆, x1 : P1 , x2 : T2 ⊢ e0 : T0
(ty-∗-elim-pt)
Σ; ~B; ∆ ⊢ let hx1 , x2 ipt = e12 in e0 : T0
where e is let hx1 , x2 ipt = e12 in e0 and T = T0 . By induction hypothesis on D1 , there
exists the following derivation in ATS0 :
D1′ :: |Σ|; |~B|, |∆| p ; |∆|t ⊢ |e12 | : |P1 | ∧ |T2 |
By induction hypothesis on D2 , there exists the following derivation in ATS0 :
D2′ :: |Σ|; |~B|, |∆| p , |P1 |; |∆|t , x2 : |T2 | ⊢ |e0 | : |T0 |
Applying the rule (ty-∧-elim) to D1′ and D2′ yields the following derivation:
D ′ :: |Σ|; |~B|, |∆| p ; |∆|t ⊢ let ∧(x2 ) = |e12 | in |e0 | : |T0 |
Note that |e| equals let ∧(x2 ) = |e12 | in |e0 |, and the case concludes.
• Assume that the last applied rule in D is (ty-app-tp):
D1 :: Σ; ~B; ∆ ⊢ e1 : P1 → T2 D2 :: Σ; ~B; ∆ ⊢ e2 : P1
(ty-app-tp)
Σ; ~B; ∆ ⊢ apptp (e1 , e2 ) : T2
where e is apptp (e1 , e2 ) and T = T2 . By induction hypothesis on D1 , there exists the
following derivation in ATS0 :
D1′ :: |Σ|; |~B|, |∆| p ; |∆|t ⊢ |e1 | : |P1 | ⊃ |T2 |
Applying Lemma 4.1 to D2 yields that the constraint |Σ|; |~B|, |∆| p ; |∆|t ⊢ |P1 | is valid.
Applying the rule (ty-⊃-elim) to D1′ and the valid constraint yields the following
derivation:
|Σ|; |~B|, |∆| p ; |∆|t ⊢⊃− (|e1 |) : |T2 |
Note that |e| equals ⊃− (|e1 |), and the case concludes.
The rest of the cases can be handled similarly.
By Theorem 4.1, the proof-erasure of a program is well-typed in ATS0 if the program
itself is well-typed in ATSpf . In other words, Theorem 4.1 justifies PwTP in ATSpf as an
approach to internalizing constraint-solving through explicit proof-construction.
ZU064-05-FPR
main
28
28 March 2017
0:56
Hongwei Xi
5 Related Work and Conclusion
Constructive type theory, which was originally proposed by Martin-Löf for the purpose of
establishing a foundation for mathematics, requires pure reasoning on programs. Generalizing as well as extending Martin-Löf’s work, the framework Pure Type System (PTS)
offers a simple and general approach to designing and formalizing type systems. However,
type equality depends on program equality in the presence of dependent types, making
it highly challenging to accommodate effectful programming features as these features
often greatly complicate the definition of program equality (Constable & Smith, 1987;
Mendler, 1987; Honsell et al., 1995; Hayashi & Nakano, 1988).
The framework Applied Type System (ATS) (Xi, 2004) introduces a complete separation
between statics, where types are formed and reasoned about, and dynamics, where programs are constructed and evaluated, thus eliminating by design the need for pure reasoning
on programs in the presence of dependent types. The development of ATS primarily unifies
and also extends the previous studies on both Dependent ML (DML) (Xi & Pfenning, 1999;
Xi, 2007) and guarded recursive datatypes (Xi et al., 2003). DML enriches ML with a
restricted form of dependent datatypes, allowing for specification and inference of significantly more precise type information (when compared to ML), and guarded recursive
datatypes can be thought of as an impredicative form of dependent types in which type
indexes are themselves types. Given the similarity between these two forms of types, it is
only natural to seek a unified presentation for them. Indeed, both DML-style dependent
types and guarded recursive datatypes are accommodated in ATS.
In terms of theorem-proving, there is a fundamental difference between ATS and various
theorem-proving systems such as NuPrl (Constable et al. , 1986) (based on Martin-Löf’s
constructive type theory) and Coq (Dowek et al., 1993) (based on the calculus of construction (Coquand & Huet, 1988)). In ATS, proof construction is solely meant for constraint
simplification and proofs are not expected to contain any computational meaning. On the
other hand, proofs in NuPrl and Coq are required to be constructive as they are meant for
supporting program extraction.
The theme of combining programming with theorem-proving is also present in the programming language Ωemga (Sheard, 2004). The type system of Ωemga is largely built on
top of a notion called equality constrained types (a.k.a. phantom types (Cheney & Hinze, 2003)),
which are closely related to the notion of guarded recursive datatypes (Xi et al., 2003). In
Ωemga, there seems no strict separation between programs and proofs. In particular, proofs
need to be constructed at run-time. In addition, an approach to simulating dependent types
through the use of type classes in Haskell is given in (McBride, 2002), which is casually
related to proof construction in the design of ATS. Please also see (Chen et al., 2004) for
a critique on the practicality of simulating dependent types in Haskell.
In summary, a framework ATS is presented in this paper to facilitate the design and
formalization of type systems to support practical programming. With a complete separation between statics and dynamics, ATS removes by design the need for pure reasoning
on programs in the presence of dependent types. Additionally, ATS allows programming
and theorem-proving to be combined in a syntactically intertwined manner, providing the
programmer with an approach to internalizing constraint-solving through explicit proofconstruction. As a minimalist formulation of ATS, ATS0 is first presented and its type-
ZU064-05-FPR
main
28 March 2017
0:56
Journal of Functional Programming
29
soundness formally established. Subsequently, ATS0 is extended to ATSpf so as to support
programming with theorem-proving, and the correctness of this extension is proven based
on a translation often referred to as proof-erasure, which turns each well-typed program in
ATSpf into a corresponding well-typed program in ATS0 of the same dynamic semantics.
References
Barendregt, Hendrik Pieter. (1992). Lambda Calculi with Types. Pages 117–441 of: Abramsky,
S., Gabbay, Dov M., & Maibaum, T.S.E. (eds), Handbook of Logic in Computer Science, vol. II.
Oxford: Clarendon Press.
Chen, Chiyan, & Xi, Hongwei. 2005 (September). Combining Programming with Theorem Proving.
Pages 66–77 of: Proceedings of the Tenth ACM SIGPLAN International Conference on Functional
Programming.
Chen, Chiyan, Zhu, Dengping, & Xi, Hongwei. (2004). Implementing Cut Elimination: A Case Study
of Simulating Dependent Types in Haskell. Proceedings of the 6th International Symposium on
Practical Aspects of Declarative Languages. Dallas, TX: Springer-Verlag LNCS vol. 3057.
Cheney, James, & Hinze, Ralf. (2003). Phantom Types. Technical Report CUCIS-TR2003-1901.
Cornell University.
Constable, Robert L., & Smith, Scott Fraser. 1987 (June). Partial objects in constructive type theory.
Pages 183–193 of: Proceedings of Symposium on Logic in Computer Science.
Constable, Robert L., et al. . (1986). Implementing mathematics with the NuPrl proof development
system. Englewood Cliffs, New Jersey: Prentice-Hall.
Coquand, Thierry, & Huet, Gérard. (1988). The calculus of constructions. Information and
computation, 76(2–3), 95–120.
de Moura, Leonardo Mendonça, & Bjørner, Nikolaj. (2008). Z3: an efficient SMT solver. Pages
337–340 of: Tools and Algorithms for the Construction and Analysis of Systems, 14th International
Conference, TACAS 2008, Held as Part of the Joint European Conferences on Theory and Practice
of Software, ETAPS 2008, Budapest, Hungary, March 29-April 6, 2008. Proceedings.
Dowek, Gilles, Felty, Amy, Herbelin, Hugo, Huet, Gérard, Murthy, Chet, Parent, Catherine, PaulinMohring, Christine, & Werner, Benjamin. (1993). The Coq proof assistant user’s guide. Rapport
Technique 154. INRIA, Rocquencourt, France. Version 5.8.
Girard, Jean-Yves. (1986). The System F of variable types, fifteen years later. Theoretical computer
science, 45(2), 159–192.
Hayashi, Susumu, & Nakano, Hiroshi. (1988). PX: A computational logic. The MIT Press.
Hoare, C. A. R. (1969). An axiomatic basis for computer programming. Communications of the
acm, 12(10), 576–580 and 583.
Honsell, Furio, Mason, Ian A., Smith, Scott, & Talcott, Carolyn. (1995). A variable typed logic of
effects. Information and computation, 119(1), 55–90.
Jones, Mark P. (1994). Qualified types: Theory and practice. The Edinburgh Building, Cambridge
CB2 2RU, UK: Cambridge University Press.
McBride, Conor. (2002). Faking It. Journal of functional programming, 12(4 & 5), 375–392.
Mendler, N.P. 1987 (June). Recursive types and type constraints in second-order lambda calculus.
Pages 30–36 of: Proceedings of Symposium on Logic in Computer Science. The Computer Society
of the IEEE, Ithaca, New York.
Milner, Robin, Tofte, Mads, Harper, Robert W., & MacQueen, D. (1997). The definition of standard
ml (revised). Cambridge, Massachusetts: MIT Press.
Reynolds, John C. (1972). Definitional interpreters for higher-order programming languages. Pages
717–740 of: Proceedings of the ACM Annual Conference.
Reynolds, John C. (1998). Theories of programming languages. Cambridge University Press.
ZU064-05-FPR
main
30
28 March 2017
0:56
Hongwei Xi
Sheard, Tim. (2004). Languages of the future. Proceedings of the Onward! Track of ObjectedOriented Programming Systems, Languages, Applications (OOPSLA). Vancouver, BC: ACM
Press.
Xi, Hongwei. (2001). Dependent ML. Available at:
http://www.cs.bu.edu/~hwxi/DML/DML.html.
Xi, Hongwei. (2004). Applied Type System (extended abstract). Pages 394–408 of: post-workshop
Proceedings of TYPES 2003. Springer-Verlag LNCS 3085.
Xi, Hongwei. (2007). Dependent ML: An approach to practical programming with dependent types.
Journal of functional programming, 17(2), 215–286.
Xi, Hongwei. (2008a). ATS/LF: a type system for constructing proofs as total functional programs.
Benzmüller, Christoph, Brown, Chad, Siekmann, Jörg, & Statman, Rick (eds), Festschrift in
Honour of Peter B. Andrews on his 70th Birthday. Studies in Logic and the Foundations of
Mathematics. IFCoLog.
Xi, Hongwei. (2008b). The ATS Programming Language System. Available at:
http://www.ats-lang.org/.
Xi, Hongwei, & Pfenning, Frank. (1999). Dependent Types in Practical Programming. Pages
214–227 of: Proceedings of 26th ACM SIGPLAN Symposium on Principles of Programming
Languages. San Antonio, Texas: ACM press.
Xi, Hongwei, Chen, Chiyan, & Chen, Gang. (2003). Guarded Recursive Datatype Constructors.
Pages 224–235 of: Proceedings of the 30th ACM SIGPLAN Symposium on Principles of
Programming Languages. New Orleans, LA: ACM press.
| 6cs.PL
|
Subsets and Splits
No community queries yet
The top public SQL queries from the community will appear here once available.